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.1005891 | Dysregulated TGF-β Production Underlies the Age-Related Vulnerability to Chikungunya Virus | Chikungunya virus (CHIKV) is a re-emerging global pathogen with pandemic potential, which causes fever, rash and debilitating arthralgia. Older adults over 65 years are particularly susceptible to severe and chronic CHIKV disease (CHIKVD), accounting for >90% of all CHIKV-related deaths. There are currently no approved vaccines or antiviral treatments available to limit chronic CHIKVD. Here we show that in old mice excessive, dysregulated TGFβ production during acute infection leads to a reduced immune response and subsequent chronic disease. Humans suffering from CHIKV infection also exhibited high TGFβ levels and a pronounced age-related defect in neutralizing anti-CHIKV antibody production. In vivo reduction of TGFβ levels minimized acute joint swelling, restored neutralizing antibody production and diminished chronic joint pathology in old mice. This study identifies increased and dysregulated TGFβ secretion as one key mechanism contributing to the age-related loss of protective anti-CHIKV-immunity leading to chronic CHIKVD.
| Chikungunya virus (CHIKV) causes an acute febrile syndrome and severe acute and chronic joint disease that, for unknown reasons, especially affects older adults. CHIKV has recently reached the Americas and is threatening the Southeastern USA. At the present there is no approved vaccine or antiviral treatments against CHIKV and treatment is limited to non-steroidal anti-inflammatory drugs (NSAIDs). We have shown here in a mouse model that increased production of the immunosuppressive cytokine TGFβ contributes to the excess of CHIKV disease severity with aging. Blocking of this cytokine reduced early swelling and late joint/foot pathology, and increased production of neutralizing antibodies. Since our data demonstrate high levels of TGFβ in humans experiencing acute CHIKV infection and decreased anti-CHIKV neutralizing antibodies in older CHIKV-exposed humans, these results could pave the way to new treatments against CHIKV disease.
| Chikungunya virus (CHIKV) is a re-emerging mosquito-borne alphavirus endemic to West Africa, with outbreaks in many Asian and African countries [1], that causes a febrile illness characterized by rash and arthralgia that is often debilitating [2,3]. The distinctive severity of the joint pain causes those suffering from this virus to assume a twisted protective position, which gave chikungunya disease (CHIKVD) its name meaning “that which becomes contorted” [4]. Although some patients resolve joint pain and swelling within 10–12 days, in up to half of patients symptoms become chronic and can persist for years [5,6].
High viremia during early CHIKVD (109−12 virus copies/mL) enables transmission from person to person via mosquitoes [7]. Since its discovery, CHIKV has widened its geographic range [8], reaching the Caribbean and South America by 2013, with >1M clinical and 25,000 laboratory confirmed cases [9]. The Aedes (Ae.) species of mosquitoes that carry CHIKV reside in the U.S. (Ae. albopictus and Ae. aegypti), and Florida reported autochthonous transmission in 2014 (www.paho.org), highlighting the risk for the U.S., with its CHIKV-naïve and rapidly aging population, to become an epidemic location.
Combined data from human outbreaks and animal models have begun to reveal the pathogenesis of, and immunity against, CHIKV infection. There is a positive association between high levels of serum pro-inflammatory cytokines and CHIKV clearance in humans, monkeys, and mice [10–15]. Type I IFN (α/β) was specifically identified as a key mediator associated with CHIKV clearance [15,16], and expression of the Type I IFN receptor (IFNAR) on non-hematopoietic cells was required for survival of CHIKV [17]. Synovial macrophages are a reservoir for virus and could be involved in early joint swelling [18–20] and subsequent regulation of inflammation. CD4+ T cells appear required for the early joint swelling in mouse models [21], whereas their role and the role of CD8+ T cells in viral clearance remains controversial [21,22]. B cells infiltrate the footpads and passive transfer of antibodies (Ab) can control infection in mice [22–24]. Humans also display a robust anti-CHIKV Ab response that is believed to be protective [18,20]. However, the exact contribution of each facet of the immune response to disease presentation and resolution remains incomplete. It remains unclear (i) how the innate response coordinates the adaptive response during CHIKV infection; (ii) what the impact of cytokine, T cell, and B cell responses may be on disease severity and length; and (iii) how that may be altered in the context of specific risk factors such as advanced age. Indeed, the greater risk of persistent and severe CHIKV disease among the elderly is likely due to one or more aging-related defects in the innate and adaptive immune response.
To elucidate such potential defects, we developed a mouse model which recapitulates age-related clinical outcomes observed in CHIKV-infected elderly humans. We report that increased production of TGFβ is linked to qualitative and quantitative impairments in B and T cell responses, which fail to clear the virus. We show that anti-TGFβ Ab treatment can prevent age-related increases in CHIKV disease severity, reduce joint pathology, and improve production of neutralizing Ab. Given that TGFβ is also elevated in humans suffering from CHIKVD, we propose this pathway as a possible target in treating CHIKV infection in older adults.
To define age-related changes in anti-CHIKV immunity, we infected C57BL/6 (B6) mice with CHIKV strain SL15649. Footpad (f.p.) CHIKV inoculation in adult mice results in early biphasic foot swelling, peaking on day 3 and 8–9 and corresponding to an early, innate, and a later, adaptive, phase of the response, and resolving by d16 [21,24]. We confirmed these results (Fig 1) and extended them to old mice. Importantly, old (O) mice exhibited significantly increased swelling as compared to adult (A) during both phases of the immune response (Fig 1). In addition, in O mice peak swelling in both phases was sustained longer and at higher levels than in A animals (Fig 1). However, the onset and resolution of swelling occurred in both groups by day 16 post-infection (p.i.; Fig 1). We did not observe swelling in the non-injected or saline-injected contralateral foot and found no CHIKV-specific mortality in either age group
To examine whether increased foot swelling in O CHIKV-infected mice correlates to increased viremia, we measured viral titers in the serum, inoculated foot (Fig 2), and non-inoculated foot (S1 Fig) by plaque assay. In the blood, CHIKV was first detected on d2 p.i., (Fig 2A). By d3 p.i., O mice exhibited significantly higher CHIKV titers, suggestive of delayed viral control, but by d4, both O and A mice resolved viremia (Fig 2A), similar to findings in other viral [25]and bacterial [26] models where O mice usually manage to control systemic virus following a delay, relative to A mice. Support for the idea that O mice exhibit delayed viral control was even more remarkably illustrated by data from inoculated feet, where O mice displayed 10-fold higher viral loads than in A on d3 p.i. (Fig 2B). Despite the absence of joint swelling, CHIKV was also detectable on d3 in the contralateral, non-inoculated foot at ~1000 fold lower levels compared to the inoculated foot (S1 Fig). By d9 p.i., infectious CHIKV dropped below the limit of detection in the feet of most A mice but remained detectable in both inoculated (Fig 2B) and contralateral (S1A Fig) feet of the O animals. Viral genomes could be detected in the inoculated footpads on d60 p.i., with significantly higher viral genome copies in the O mice (Fig 2C). These results demonstrate impaired virus control with aging, consistent with data from Rhesus macaques [15]. Finally, very low levels of fluorescent infectious virus were recovered from both A and O mice at 90 days post-infection (Figs 2D and S1B) demonstrating for the first time in an animal model that replicating CHIKV persists far beyond the acute phase in the infected joints. This finding is consistent with evidence for replicating virus isolated from a single patient experiencing chronic CHIKVD [18]. While O and A mice did not show statistically significant difference in infectious viral load on d90 at this experimental power, there was a trend of higher levels in O mice, which will have to be substantiated in future experiments. While delayed viral control in O mice could suggest a link with increased early swelling, viral load did not directly correlate with, and is probably not the sole determining factor for, swelling. This follows from data showing that marked swelling was present at times where infectious viral titers were below the limit of detection for plaque assay, as is the case with most adult animals on d9 p.i. and with both age groups during chronic CHIKVD. This relationship between infectious virus, swelling and joint pathology as a function of age will, therefore, require further investigation, but data so far are consistent with prior literature suggesting the presence of host response components.
Increased foot swelling in O mice could be caused by direct cytopathic virus effects, by immunopathological actions of innate or adaptive cells or molecules (cytokines), or by a combination of both. To discern between these possibilities we analyzed maintenance and recruitment of various cells into the lymph nodes (LN) of O mice. LN from naïve O mice were visibly smaller than those in A mice (S2A Fig). Further, despite an increase in the size of the draining LN (dLN) and non-draining LN (ndLN) on d3 of CHIKV infection, O LN never reached the size of A LN at any time point (S2A Fig). Total cellularity of naïve and d3, 7, and 9 p.i. dLN (Fig 3A) was also significantly lower in O at all time points compared to A LN. This suggests an inability of the O LN to expand, recruit and/or maintain a sufficient number of cells to make up for the deficit in naïve LN, consistent with recent data [27]. The reduced LN reaction was evident in the ndLN in O mice as well, which exhibited only minimal, if any increase in total cell numbers (Fig 3A) despite the local presence of the infectious virus at these times (S1B Fig). An analysis of natural killer (NK), dendritic cells (DC) and macrophages revealed that these cells were either reduced from the beginning and/or failed to accumulate to the same levels as in A animals (S2B Fig), suggesting that none of them would be likely to account for excess foot swelling in O mice. We also found reduced CD4+, CD8+ and B cell numbers in the LN of naïve O mice and none of these populations were able to expand to equivalent numbers found in LN of A mice (S2C Fig) following CHIKV infection.
To precisely evaluate the T cell responses, we identified dominant I-Ab restricted CHIKV regions (S2D Fig), with the E22805-2820 epitope being absolutely immunodominant (S2E Fig). We found a tenfold reduction in absolute numbers of IFNγ+ CHIKV-specific CD4+ cells in O mice (Fig 3B), which would not have been revealed by percentage/frequency comparison (Fig 3C, a trend but no significant A to O difference). During our initial screen of peptide pools, we did not find differences in the CD8 IFNγ responses with aging in the spleen (S2H Fig), and subsequent preliminary analysis failed to discover differences in the LN, although cellularity of LN was sharply reduced (Fig 2). We conclude that numerically, both CD8 and CD4 responses were reduced, and, at face value, this reduction is inconsistent with the idea that these cells could mediate enhanced immunopathology in O mice.
Reduced CD4+ T cell responses in O mice, together with reduced B cell numbers in the dLN, could lead to impaired humoral responses in O mice. We found that both A and O mice produced anti-CHIKV IgM antibody by d7 p.i., yet, while IgM levels dropped in A mice, they remained significantly higher in O mice over adult on d 9 and 16 post-infection (Fig 3D), consistent with reduced efficacy in class switching in O mice [28,29]. Amounts of anti-CHIKV IgG2c Ab, the isotype considered to be most protective against CHIKV [30], did not increase to the same levels in O mice compared to A counterparts on d16 and 60, and the difference was significant at d60 (Fig 3E), following the trend of the total IgG Ab (S2F Fig), suggesting an impaired memory Ab response. An exception was anti-CHIKV IgG2b, that trended higher in O mice on d9 and was significantly elevated by d16 (S2G Fig). The IgG2b isotype is associated with a “suppressive” cytokine environment that includes production of TGFβ [31].
We next tested the neutralizing capacity of A and O serum in a plaque-reduction neutralization test (PRNT). We found that the neutralizing potency of O serum was trending lower than in A mice on d9, and that difference was statistically significant on d60 (Fig 3F). This also held true for CHIKV-infected humans, where serum from people >65y contained significantly lower neutralizing Ab titers than in those <30y (Fig 3G). Therefore, decreased amounts, suboptimal iso/allotype and reduced neutralizing potency of Ab with age all likely contribute to the increased disease severity, incomplete viral control and elevated incidence of chronic disease in the elderly.
The above age-related defects in the CD4+ T cell and the humoral responses prompted us to evaluate serum cytokine and chemokine profiles in A and O mice following CHIKV infection. While most of the cytokines and chemokines assayed by Luminex array exhibited no significant age-related differences, or exhibited differences that could not be validated by ELISA (e.g. differences in IL-10, S3 Fig), we found an early (d2 p.i.) and significant under-induction of CXCL9 in O mice, which was confirmed by ELISA (Fig 4A). CXCL9 is a proinflammatory chemokine that functions as a chemoattractant for activated lymphocytes, and its lower production could have contributed to delayed and incomplete recruitment to the dLN, an issue currently under investigation.
Moreover, O mice exhibited a significantly greater increase in TGFβ on d3 (S4A Fig) and d9 p.i. relative to A counterparts (Fig 4B), although by d30 these levels returned to baseline in both A and O mice (Fig 4B). TGFβ is a pleiotropic cytokine with diverse effects on the immune system that are incompletely understood. TGFβ operates as a switch-factor for murine antibody isotypes, inducing IgG2b, as well as for mediation of leukocyte recruitment and activation [32,33]. The increased levels of TGFβ-switched IgG2b anti-CHIKV Ab on d16 post-infection of O mice (S3G Fig) led us to hypothesize that anti-CHIKV immunity in O mice is improperly coordinated and that excessive production of TGFβ contributes to increased CHIKVD in O mice. We also found very high levels of free-active TGFβ in sera of acute CHIKV patients (Fig 4C), which validated TGFβ as a potentially relevant cytokine in CHIKVD in both humans and mice.
To test the above hypothesis, we treated A and O mice with footpad injections of anti-TGFβ Ab or isotype control on d-1, 1, 3 and 5 p.i. and demonstrated that this treatment reduced the concentration of TGFβ in serum in O mice close to or to the levels of TGFβ in A mice (S4A Fig). Reducing the serum of concentration of TGFβ did not have a direct effect on CXCL9 concentration (S4B Fig) suggesting that in O mice systemic CXCL9 may not be depressed due to elevated TGFβ. A somewhat more complex situation was seen in the case of Type I Interferon (S4C Fig), known to be required for control of early CHIKV infection [16]. O mice produce significantly less Type I Interferon than A (S4C Fig) on d2 p.i., consistent with results in old non-human primates [15]. However, that difference disappeared under TGFβ blockade, both because TGFβ blockade slightly reduced production of Type I IFN in A mice and slightly increased its production in O mice (S4C Fig). Importantly, TGFβ blockade did effectively reduce both peaks of acute foot swelling in O mice (red dashed vs. red solid line) to the levels observed in A mice (Fig 5), strongly suggesting that high levels of TGFβ contribute decisively to the age-related increase in acute CHIKVD. TGFβ Ab treatment in CHIKV infected A mice did not reduce swelling during the early, but did during the late acute phase (Fig 5, d8, black dashed vs. black solid line). It should be noted that TGFβ neutralization did not prevent swelling altogether, which could be due to the fact that systemically TGFβ was not completely neutralized in O and was only marginally reduced in A mice (S4A Fig) or could suggest the existence of other, age-independent, mediators of acute CHIKVD. Experiments assessing local levels of different cytokines, including TGFβ and Type I IFN, are in progress to further resolve this issue.
Further, TGFβ reduction/blockade did not exert a direct anti-viral effect, as determined by measuring viral titers in serum on days 1–4 (S5A Fig) and tissues on days 3 and 9 (S5B and S5C Fig), where groups with blockade did not have appreciably lower viral titers compared to control groups. TGFβ reduction/blockade also did not promote full clearance of viral genomes from the tissue (S5D Fig). This data taken together suggests that CHIKV persistence is driven by elevated TGFβ, but likely also by other factors, most notably host defense mechanisms. Genetic ablation of TGFβ signaling in specific cell subsets will be necessary to conclusively discriminate between these possibilities.
TGFβ reduction/blockade during acute infection also reduced the incidence of chronic arthritis and restored neutralizing Ab responses against CHIKV in old mice (Fig 6). When hematoxylin/eosin (H&E) stained tissue sections were evaluated for synovitis, arthritis, and metatarsal muscle inflammation on d90 p.i. using a previously described scoring system [22], we found that control-treated O mice exhibited increased frequency of chronic arthritis and/or metatarsal muscle inflammation (4 of 6 mice) as compared to A (1 of 6 mice, P = 0.07, chi-square test) (Fig 6A and 6B). Importantly, that incidence was reduced by 50% when TGFβ was blocked during acute infection in O mice (Fig 5C), whereas, neutralization of TGFβ in A mice made the late joint pathology worse (Fig 5C), suggesting that in a properly controlled response in A mice, TGFβ plays a protective role in joint infiltration and pathology.
Finally, we evaluated the neutralizing Ab capacity on d90 p.i. and found that anti-TGFβ blockade during acute infection restored CHIKV-neutralizing Ab titers to high, adult-like levels in O mice, but did not affect neutralizing Ab responses in A mice (Fig 5D). Taken together, our results identify increased, dysregulated TGFβ secretion during very early, acute infection as a key, specific mechanism contributing to the age-related loss of immune system function and increased joint pathology in the course of CHIKV infection.
CHIKV is an emerging disease with pandemic potential and pronounced acute and prolonged disability, particularly in older adults. A mouse model of A and O infection using footpad inoculation of B6 mice, described in this report, provides important clues on the basis of age-related vulnerability to CHIKVD. The disease in O animals was marked by enhanced prolonged viremia, more severe early swelling and late footpad joint and connective tissue pathology. We also present evidence that live, replication competent, CHIKV persists in the tissues of both A and O mice. This suggests that increased chronic CHIKVD with age is not due to differential viral persistence but is rather a consequence of how persistence is controlled. Importantly, we show that aged animals generated a quantitatively and qualitatively defective immune response at both innate and adaptive levels. We demonstrate that dysregulated TGFβ cytokine secretion decisively contributed to both enhanced CHIKVD and to defects in protective immunity. We further report that this dysregulation is age-specific and does not play a role in young mice. This is supported by the fact that neutralization of TGFβ in A animals did not erode their B cell response and may have been somewhat detrimental to the late joint pathology. Finally, similar signs of immune dysregulation with CHIKVD were observed in humans, including elevated TGFβ in adult and older humans and reduced neutralizing Ab titers in older humans. Our human studies were not powered or designed to conclusively assess whether in humans there are any age-related differences in TGFβ production, an issue that will have to await further studies. Also of note, we saw no sex differences in TGFβ production between male and female human subjects, and in one experiment with limited numbers of female old mice, we saw the same excess swelling and reduced neutralizing Ab responses as seen in old males However, this study chiefly studied male mice, and therefore sex differences in the susceptibility to CHIKVD with age remain to be explored.
Based on this, we propose that treating the age-related changes in the immune system (and, likely, in other systems and organs) as a simple continuum of processes known to operate in younger age could be conceptually limiting, and even erroneous in some situations. Our results are at least in part consistent with an altered state of the old immune system, where some of the rules that operate in youth no longer apply, due to dysregulated homeostasis. Results of recent studies on the maintenance of the naïve T cell pool and its diversity by us [34–36] and others [37–39] are consistent with that idea, and may suggest re-thinking of the conceptual framework within which we consider age-related changes in function.
Generation of an effective immune response requires coordinated activation of early innate and late adaptive immune responses, and any age-related changes in either of the two arms could profoundly affect the ability to fight off the virus initially or control persistent infection. With aging we found reduced levels of CXCL9, a chemo-attractant mostly secreted by macrophages, and responsible for the recruitment of lymphocytes during viral infections [40,41] which could explain the reduced numbers of lymphocytes recruited to dLN and/or infected joints. Yet that was not the only issue found in the O LN, which were smaller in size even before infection, and never reached the degree of size or cellularity measured in the A counterparts. There could be multiple reasons for this, including a well-described decline in naïve CD8+ [36,37] and CD4+ [39] numbers, and the more recently described age-related degradation of LN stromal architecture, which may render it incapable of supporting the youthful number and diversity of cell types. Indeed, recent data points to the changes in stromal architecture of the LN with age in the steady state [42] and the inability of LN to recruit and properly direct migration of lymphocytes following infection [27] where the above mentioned defect in CXCL9 production could play a role.
In addition, immunodominant CD4+ T cell epitope identification allowed us to reveal a reduced anti-CHIKV IFNγ response of old CD4+ T cells. This provides further evidence suggesting that prolonged clinical pathology observed at the site of viral infection in O mice is probably not mediated by CHIKV-specific Th1 CD4+ or CD8+ T cells, as their numbers and function are decreased in aging. Moreover, we did not observe Th17 or increased Treg cells in CHIKV infection so far Consistent with depressed Th1 immunity, we found similar initial viral titers but an age-related delay in CHIKV control in O mice. These observations closely parallel the results obtained in CD4-/- and IFNγ-/- mice [30], where defects in both cellular (CD4+ cells) and humoral (Ab) immunity contribute to impaired immunity to CHIKV. We therefore conclude that the CD4+ and B cell lymphopenia measured in the LN, decreased Th1 response of CHIKV-specific CD4+ T cells and the action of TGFβ likely contribute to the total IgG and IgG2c antibody deficiency and to reduced CHIKV-neutralizing titers on day 60 post-infection.
Of importance, the higher TGFβ levels in the O mice led not only to increased early conversion into IgG2b isotype but also established an environment conducive to swelling and tissue pathogenesis. Increased production of TGFβ in O mice during acute infection is not unique to CHIKV as it was found in the West Nile Virus infection (S6 Fig) and following Encephalitozoon cuniculi infection of old mice [43]. However, since TGFβ blockade reversed all the above phenomena in CHIKV infection, our data suggest that with aging, increased TGFβ levels likely tipped the balance away from generation of an efficient and protective immune response and towards chronic arthritis. While many details remain to be elucidated about the exact mechanistic functioning of the TGFβ axis in old mice undergoing CHIKV infection and CHIKVD, our results reported herein identify TGFβ as one key mechanism behind age-related vulnerability to CHIKVD. Further, our studies point to this cytokine and its signaling pathway as a potential target for immune intervention to remedy the pathology associated with CHIKV infection, and present preliminary validation of this target in humans.
Mouse studies were 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. Protocols were approved by the Institutional Animal Care and Use Committee at the University of Arizona (IACUC #08–102, PHS Assurance Number: A3248-01). Footpad injections were performed under isoflurane anesthesia. Euthanasia was performed by isoflurane overdose or cervical dislocation. All collection and use of samples from human subjects was approved by the Ethics Committee at The University of the West Indies and the Institutional Review Board at the University of Arizona. Informed consent was written and provided by the subject or their legal guardian.
O (18 months) and A (12 weeks) male C57BL/6 (B6) mice were purchased and/or obtained from the National Institute on Aging Rodent Resource via the Charles River Laboratories (Frederick, MD and Kingston, NY) and/or The Jackson Laboratory (Bar Harbor, ME). Mice anesthetized with isoflurane were infected subcutaneously in the footpad (f.p.) with 1000pfu of CHIKV as previously described [44]. Foot swelling was measured daily with calipers until d21 p.i. Footpad area was determined as (height x width) and expressed as increase over d0. No swelling was observed in the non-injected or saline-injected contralateral foot. All CHIKV experiments were conducted within U.S. Department of Agriculture and CDC-inspected biosafety level 3 facilities at the University of Arizona.
CHIKV strain SL15649 (Genbank accession no. GU189061) was isolated from serum of a febrile patient in Sri Lanka in 2006 and was propagated twice in Vero cells before the generation of an infectious cDNA clone, used to previously establish mouse CHIKV infection [44]. The pMH56.2 plasmids encoding SL15649 CHIKV and SL15649 CHIKV expressing mKate were generously provided by Dr. C. E. McGee and Dr. M. T. Heise [45]. Virus titer was determined by plaque assay on Vero cells [44].
Infectious viral titers were determined by a standard plaque assay on Vero cells as described in [44]. CHIKV RNA loads were measured using quantitative real time reverse transcription PCR with the following primers and probe: CHIKV-9482F 5′-GGAACGAGCAGCAACCTTTG-3′; CHIKV-9931R, 5′-ATGGTAAGAGTCTCAGACAGTTGCA-3′; and probe CHIKV-9870F, 5′-GGAATAAGGGCTTGT-3′ from viral RNA isolated as previously described [15]. Gene amplicons served as quantification standards (sensitivity, 10 to 100 copies). qRT-PCR was performed and analyzed using ABI StepOne Plus real-time PCR system (Applied Biosystems). Persistent viral infection was determined by culturing tissue lysate on C6/36 insect cells for 3 days followed by FCM analysis to detect infected cells. Samples were acquired using a BD LSR Fortessa cytometer (BD Bioscience, San Jose, CA) and analyzed by FlowJo software (Tree Star, Ashland, OR).
Accutase-treated (eBioscience, San Diego, CA) popliteal lymph nodes were disassociated over a 40uM cell strainer. Following Fc block, cells were incubated overnight in a saturating dose of mAb against CD3, CD4, CD8α, CD19, CD11b, CD11c, NK1.1 and F4/80 (eBioscience, San Diego, CA), stained with Live/Dead Yellow (Life Technologies, Grand Island, NY) and analyzed as below. Peptide stimulation was in the presence of protein transport inhibitor (eBioscience, San Diego, CA) as described [25]. Overlapping peptide pools (15mer, overlapping by 5) for the 9 proteins of CHIKV were used to determine the immunodominant regions of E2 and NSP1. Libraries of these regions (21st Century, Marlboro, MA) were used to determine individual immunodominant epitopes. Samples were acquired using a BD LSR Fortessa cytometer (BD Bioscience, San Jose, CA) and analyzed by FlowJo software (Tree Star, Ashland, OR). Cell counts were extrapolated from either a hand count on a hemocytometer or by CBC differential collected on a Hemavet LV (Drew Scientific, Waterbury, CT) instrument. The two counting methods were confirmed to be consistent.
Ab titers were assessed using a CHIKV infectious cell lysate-based enzyme-linked immunosorbent assay (ELISA). Briefly, CHIKV-infected lysate was generated by infection of primary human fibroblasts and used to coat 96 well Immulon 2 HB plates (Thermo Labsystems, Franklin, MA), with uninfected lysate used as control. Plates were blocked with PBS-0.05% Tween-20 + 5% dry nonfat milk. Serum was diluted 1:50 in the same blocking buffer, incubated for 1h at 22°C, incubated with horseradish peroxidase-labeled goat anti-mouse IgG (KPL, Gaithersburg, MD) or anti-IgM, IgG1, IgG2b or IgG2c (Southern Biotech, Birmingham, AL) and developed with 3,3′,5,5′-Tetramethylbenzidine dihydrochloride (Sigma-Aldrich, St. Louis, MO). Reaction was terminated with 1M H2SO4 and absorbance measured at 450 nm. Plaque reduction neutralization test assay was done on Vero cells and neutralization titers were determined as the serum dilution with a 90% reduction in plaques (NT90) compared to wells infected with CHIKV in the absence of serum.
Between July 2014 and April 2015, blood samples were submitted to the Department of Microbiology at the University of the West Indies to be tested for the presence of CHIKV IgM antibodies. This resulted from an enhanced fever and rash surveillance initiative by the local Ministry of Health as part of the preparedness and response plan for CHIKV introduction in the Island. Samples were tested using the Anti-CHIKV IgM human ELISA kit (Abcam, Cambridge, MA, USA). The Centers for Disease Control and Prevention (CDC) reported an overall sensitivity and specificity of 88% and 90%, respectively. For these studies, serum samples (n = 24 <30 years and n = 15 >65 years; 100μL volume) from identified IgM+ CHIKV patients were tested alongside age- and sex-matched CHIKV-naïve controls from Tucson, Arizona.
ELISAs for mouse MIG/CXCL9 (R&D Systems Inc, Minneapolis, MN), TGFβ (eBioscience, San Diego, CA), Free-active TGFβ for humans (BioLegend, San Diego, CA) and the Luminex multiplex mouse cytokine assay (Life Technologies, Inc.) were performed following manufacturer instructions.
100μg of TGFβ antibody, clone 1D11.16.8 (BioXCell, West Lebanon, NH) or IgG1 isotype control, clone MOPC-21 (BioXCell, West Lebanon, NH) were injected via f.p. route on days -1, 1, 3, and 5 p.i. in 20μl saline.
Following euthanasia, foot and ankle tissues were collected and fixed in 10% neutral buffered formalin for 24 hours, then processed and embedded into paraffin blocks. Hematoxylin and Eosin (H&E) stains were performed on 5μ sections of tissue cut from the formalin fixed, paraffin embedded (FFPE) blocks.
Data were analyzed using Prism Graph Pad software and the statistical test referenced in each figure with the exception of Figs 1 and 5. These were analyzed by mixed model, repeated measures analyses of covariance (ANOVA), with group as a between group factor and time (days post infection) as a within group factor and their interaction were used to compare differences in performance among all groups over time for analyzing outcome (increased footpad area). In a typical experiment using repeated measures, two measurements taken at adjacent times are more highly correlated than two measurements taken several time points apart; therefore, we used a first-order autoregressive (AR1) covariance structure to account for within-subject correlation. Due to the small sample size, other more complex covariance structures were not considered. Tukey-Carmer multiple comparisons correction was used to control overall type I error rate.
|
10.1371/journal.ppat.1002598 | The Core Protein of Classical Swine Fever Virus Is Dispensable for Virus Propagation In Vitro | Core protein of Flaviviridae is regarded as essential factor for nucleocapsid formation. Yet, core protein is not encoded by all isolates (GBV- A and GBV- C). Pestiviruses are a genus within the family Flaviviridae that affect cloven-hoofed animals, causing economically important diseases like classical swine fever (CSF) and bovine viral diarrhea (BVD). Recent findings describe the ability of NS3 of classical swine fever virus (CSFV) to compensate for disabling size increase of core protein (Riedel et al., 2010). NS3 is a nonstructural protein possessing protease, helicase and NTPase activity and a key player in virus replication. A role of NS3 in particle morphogenesis has also been described for other members of the Flaviviridae (Patkar et al., 2008; Ma et al., 2008). These findings raise questions about the necessity and function of core protein and the role of NS3 in particle assembly. A reverse genetic system for CSFV was employed to generate poorly growing CSFVs by modification of the core gene. After passaging, rescued viruses had acquired single amino acid substitutions (SAAS) within NS3 helicase subdomain 3. Upon introduction of these SAAS in a nonviable CSFV with deletion of almost the entire core gene (Vp447Δc), virus could be rescued. Further characterization of this virus with regard to its physical properties, morphology and behavior in cell culture did not reveal major differences between wildtype (Vp447) and Vp447Δc. Upon infection of the natural host, Vp447Δc was attenuated. Hence we conclude that core protein is not essential for particle assembly of a core-encoding member of the Flaviviridae, but important for its virulence. This raises questions about capsid structure and necessity, the role of NS3 in particle assembly and the function of core protein in general.
| Virus particles of members of the Flaviviridae consist of an inner complex of viral RNA genome and core protein that together form the nucleocapsid, and an outer lipid layer containing the viral glycoproteins. Functional analyses of core protein of the classical swine fever virus (CSFV), a pestivirus related to hepatitis C virus (HCV), led to the observation that crippling mutations or even complete deletion of the core gene were compensated by single amino acid substitutions in the helicase domain of non-structural protein 3 (NS3). NS3 is well conserved among the Flaviviridae and acts as protease and helicase. In addition to its essential role in RNA replication, NS3 apparently organizes the incorporation of RNA into budding virus particles. Characterization of core deficient CSFV particles (Vp447Δc) revealed that the lack of core had no effect with regard to thermostability, size, density, and morphology. Vp447Δc was fully attenuated in the natural host. Our results provide evidence that core protein is not essential for virus assembly. Hence, Vp447Δc might help to explain the enigmatic existence of GB viruses -A and -C, close relatives of HCV that do not encode an apparent core protein.
| The genus pestivirus, together with the genera hepacivirus, flavivirus and the newly proposed genus pegivirus [1], constitutes the family Flaviviridae. Cloven-hoofed animals are affected by pestiviruses, which cause severe diseases like classical swine fever (CSF) and bovine viral diarrhea (BVD). Pestiviruses possess a single stranded RNA genome of positive polarity with one open reading frame (orf) encoding approximately 4000 amino acids (aa). The resulting polyprotein is processed co- and posttranslationally into at least 12 viral proteins by three viral and two cellular proteases [2].
Pestiviral particles are enveloped and contain three virus-encoded glycoproteins, Erns, E1 and E2. Erns is unique for pestiviruses and is the only known viral structural protein with an uridinylate specific RNase domain belonging to the T2 RNase family [3], [4]. E1 and E2 or analogous proteins (prM, E) are encoded by all members of the Flaviviridae. Inside the virus particle, the viral genome is accompanied by a core protein. However, members of the proposed genus pegivirus, GBV- A and GBV- C [reviewed by 1], do not appear to encode a core protein. Pestiviruses encode a small, basic core protein, which, in contrast to hepaci- and flaviviruses, does not possess any predicted regular secondary structure and is intrinsically disordered [5], [6]. The pestiviral core protein has RNA chaperone activity [6] and its implicated functions are condensation of the viral RNA genome and subsequent packaging into virions. Its ability to bind RNA relies on the overall protein charge, which results in an unspecific affinity for nucleic acids [5]. The pestiviral core protein is processed at its N-terminus by the autoprotease Npro [7], whereas the C-terminus is generated by signal peptide peptidase (SPP) cleavage [8]. Recent findings revealed that deletion of basic areas of classical swine fever virus (CSFV) core protein (aa 213–231 of the viral polyprotein) results in a ten-fold reduction of virus output, whereas deletion of small, less charged stretches (aa 194–198 and aa 208–212) leads to a more than 1000-fold drop in virus output [9]. This implicates a more complex mechanism of core function in particle morphogenesis, which is not solely relying on overall protein charge. Duplication and triplication of the CSFV core protein gene as well as integration of up to 3 yellow fluorescent protein (YFP) genes between 2 core coding regions yielded replication competent viruses whose virus output was approximately 100-fold reduced in comparison to wildtype, revealing a high tolerance of core protein to size increase. We also reported the rescue of a CSFV encoding an YFP-core fusion protein by a single amino acid substitution in the NS3 helicase domain (N2256Y) [9]. This finding points to an ability of NS3 to substitute core functions.
For all members of the Flaviviridae, there is increasing evidence that nonstructural proteins are required for virus morphogenesis [reviewed by 10]. Single amino acid residues in the NS3 helicase domain of yellow fever virus (YFV) and hepatitis C virus (HCV) have been described as important for particle formation [11], [12]. Apart from NS3, p7, NS2 and NS5A have been reported as factors involved in HCV particle generation [13]–[17].
The pestiviral NS3 is a multifunctional molecule possessing protease, NTPase and helicase activity [18]–[22] and shares similarity with the analogous protein of hepaci- and flaviviruses. Its uncleaved precursor, NS2–3, has been reported to be essential for particle formation [23], [24].
In the present study, we describe the ability of CSFV NS3 to compensate for functionally compromised core proteins and even the deletion of nearly 90% of the core gene by acquisition of single codon rescue mutations in its helicase subdomain 3. These findings provide strong evidence for a major role of the NS3 helicase domain in pestiviral particle assembly and implicate questions about the function of core protein. As members of the newly proposed genus pegivirus [1] – namely GBV- A and GBV- C - do not encode an obvious core protein, we provide experimental evidence that loss of the core coding region is tolerated by another member of the Flaviviridae.
Recently, we reported that a single amino acid substitution (SAAS) (N2256Y) in the helicase domain of NS3 rescued a poorly growing CSFV construct (Vp447Yc) that encoded a core protein of which the N-terminus was fused to YFP [9]. This unexpected result prompted us to investigate spontaneously occurring revertants of a CSFV mutant in detail that initially was designed to determine requirements for core processing by signal peptide peptidase (SPP). Replacement of most of the signal peptide (aa 250–261) by a stretch of 8 leucine residues (Figure 1B) led to a poorly growing virus (4.5×103 ffu/ml) (Vp4478leu) that showed a more than 200-fold rise in titer upon passaging in SK6 cells. To identify the genomic change(s) leading to virus rescue, virus progeny was repeatedly plaque-selected. Interestingly, sequence analysis of these selected viruses did not reveal changes in the genomic sequence of the mutated core. Rescue mutations were identified by reintroducing genomic fragments (nt31–1580; nt1480–3970; nt 3900–5570; nt 5500–8590; nt8330–10510; nt 10420–12290) of the rescued viruses into the parental plasmid p4478leu. Only introduction of a genomic fragment nt 5500–8590 encoding parts of NS3-NS4B (aa 1730–2656 of the polyprotein) into the parental plasmid resulted in rescue after transfection of the respective viral genomes. Upon sequencing of this fragment one SAAS was found in each clone tested in NS3 helicase subdomain 3 (namely E2160G, N2177Y, Q2189K, P2200T and N2256D) (Figure 1B). To prove that these SAAS were indeed responsible for the rescue, the respective mutations were each engineered into the full-length cDNA construct of Vp4478leu (p4478leuE2160G, p4478leuN2177Y, p4478leuQ2189K, p4478leuP2200T, p4478leuN2256D). After transfection, the resulting viruses grew to titers exceeding 105 ffu/ml without the need for passaging (Table 1). Growth characteristics are shown for the virus growing to highest titers (Vp4478leuN2177Y) (Figure 2A). The overall titer of Vp4478leuN2177Y was about one log10 below the one of Vp447. In the background of the parental Vp447, the N2117Y substitution led to a more than 20- fold decrease of virus output (Vp447N2177Y) in comparison to Vp447 (Table 1).
To assess whether acquisition of SAAS in NS3 helicase subdomain 3 might be a general mechanism of CSFV to overcome defects in the core gene, rescue experiments with a different loss of core function mutant were attempted. An initially poorly growing CSFV (7.1×102 ffu/ml 24 h after transfection) encoding an internal deletion (aa 208–212) in the core gene (Vp447Δ208–212) (Figure 1C) was passaged in SK6 cells until an increase in virus growth was observed. Using the same approach as described above, a SAAS at position N2177H was identified. After introducing this SAAS N2177H into parental plasmid, virus titer (Vp447Δ208–212N2177H) rose to 7.9×105 ffu/ml 24 h after transfection of the respective virus genome in SK6 cells (Table 1). Apparently single amino acid substitutions in the C-terminal subdomain of the NS3 helicase compensate for functionally compromised core mutants that are compromised by N-terminal fusion to YFP (Vp447Yc), defective C-terminal processing (Vp4478leu) or an internal deletion (Vp447Δ208–212), respectively.
Core protein can be detected in lysates of cells transfected with genome of Vp447 and in pelleted virions of Vp447 (Figure 2B). Surprisingly, Western blot analysis of cell lysate and pelleted virus particles revealed that neither Vp447N2177Y nor Vp4478LeuN2177Y contained detectable levels of core protein in concentrated virus preparations. Core protein could be detected in lysates of SK6 cells transfected with genome of Vp447N2177Y, but not after transfection of genomes of Vp4478leu and Vp4478leuN2177Y.
Mutations within the NS3 helicase subdomain 3 allowed the rescue of viruses with compromised core function. To examine whether the core-coding region is dispensable altogether, almost the entire core gene (aa170–246; 77 of the 86 codons) was deleted in p447, yielding p447Δc (Figure 1D). Nine C-terminal amino acids (247–255: LEKALLAWA) were preserved as part of the signal sequence (aa 247–269) to ensure translocation of Erns into the ER lumen. While this construct lacking the core-coding region was not viable, introduction of above described SAAS in NS3 into p447Δc (p447ΔcE2160G, p447ΔcN2177H, p447ΔcN2177Y, p447ΔcQ2189K, p447ΔcP2200T, p447ΔcN2256D) led to the release of infectious virus with titers of at least 1×104 ffu/ml 24 h after electroporation of the respective transcripts (Table 1). Highest titers were observed for Vp447ΔcN2177Y and Vp447ΔcP2200T (4.0×105 and 2.3×105 ffu/ml 24 h after transfection in SK6 cells), thus being 30–50 -fold below Vp447 titer (Figure 2A). Hence, SAAS in the helicase domain of NS3 can not only compensate for functionally compromised, but even completely absent core protein. No upstream open reading frame longer than 15 codons that might provide the virus with an alternative core protein could be identified. As expected, no core protein could be detected in either cell lysate or supernatant of RNA cells transfected with Vp447ΔcN2177Y or Vp447ΔcP2200T (Figure 2B). To exclude a possible function of the C-terminal core aa 247–269 in Vp447ΔcN2177Y, they were replaced by the signal peptide of bovine CD46, a cell surface glycoprotein (Vp447ΔcN2177YCD46SP). Progeny virus production of Vp447ΔcN2177YCD46SP was slightly reduced (1×105 ffu/ml 24 h after transfection) in comparison to Vp447ΔcN2177Y. Analysis of cell lysate of Vp447ΔcN2177Y 72 h after transfection of SK6 cells did not reveal differences in the relative presence and processing of NS2–3, NS5B, Erns and E2 in comparison to wildtype (Figure S1). This suggests that cellular protein expression and polyprotein processing is neither affected by the lack of core protein nor by the presence of a SAAS in the NS3 helicase. The relative reduction of protein expression in Vp447Δc genome transfected cells results from its inability to spread. No changes in the regions surrounding the deletion of the core gene and NS3 were detected after ten passages of Vp447ΔcN2177Y in SK6 cells (data not shown). Introduction of combinations of the described amino acid exchanges in NS3 helicase of Vp447Δc showed no additive effect but rather resulted in a 10–100 fold drop in virus titer (data not shown).
The lack of a structural component of the virus particle may result in altered phenotypic properties of the virus. We therefore assessed virus infectivity, morphology, and physical stability of Vp447ΔcN2177Y compared to wildtype Vp447. The presence of viral genome in cells transfected with genomic RNA of Vp447 or Vp447ΔcN2177Y or infected with Vp447 or Vp447ΔcN2177Y was assessed by Northern blot analysis. Genomes could be detected for Vp447ΔcN2177Y (12059 nt) and Vp447 (12293 nt) (Figure 3A), but the size difference of 234 nt could not be resolved. To verify that the infectivity of Vp447ΔcN2177Y is due to proper virus particles, not secreted replication complexes, neutralization assays were performed. Incubation of Vp447ΔcN2177Y with either a monoclonal antibody against E2 (A18) or sera of one vaccinated animal (S98) and one vaccinated and subsequently CSFV infected animal (S05) neutralized infectivity in the same fashion as observed for the parental Vp447 (Figure 3B). Next, specific infectivity in the supernatant was assessed. To allow for strict discrimination between both viruses on the level of RNA, a modified Vp447ΔcN2177Y, encoding for 5 alanine residues between the Npro C-terminus and core residue 247 (Vp447Δc+5AlaN2177Y) was generated. The resulting PCR assay specifically amplified either Vp447 or Vp447Δc+5AlaN2177Y genomes (Figure S2). In cell culture, growth of Vp447Δc+5AlaN2177Y was slightly improved in comparison to Vp447ΔcN2177Y. With this approach, we determined a specific infectivity (ratio of virus genomes versus infectivity in cell culture supernatant) of 23 genomes/ffu (SD±14; n = 3) for Vp447 and 131 genomes/ffu (SD±61; n = 3) for Vp447Δc+5AlaN2177Y.
To determine density and size of Vp447 in comparison to Vp447ΔcN2177Y, equilibrium density centrifugation and size exclusion chromatography was performed. The densities of Vp447 and Vp447Δc+5AlaN2177Y were compared by separation in individual, continuous sucrose gradients (10–60%) and equilibrium centrifugation. 30 fractions of 360 µl each were harvested by bottom puncture. In repetitive experiments, infectivity peaked at a density of 1.104–1.111 g/ml for Vp447 and of 1.099–1.112 g/ml for Vp447Δc+5AlaN2177Y (Figure 4A). RNA levels, determined by virus specific real-time RT-PCR, peaked at a density of 1.10 g/ml for Vp447 and at 1.09–1.11 for Vp447Δc+5AlaN2177Y (Figure 4A). Peak E2 levels were detected from 1.10–1.14 g/ml for both viruses, but E2 was present in all fractions (Figure 4B). To avoid variations between two gradients, 106 ffu of both viruses were mixed and layered on top of the same sucrose gradient. As described above, 30 fractions of 360 µl each were harvested by bottom puncture. Again, E2 was detectable over a wide range of the gradient (1.04–1.18 g/ml sucrose) (Figure 4C) and infectivity peaked at a density of 1.105–1.113 g/ml (Figure 4D). Highest levels of Core protein were detectable at a density of 1.09–1.10 g/ml. RNA levels of either virus matched with infectivity and peaked in the same fraction (1.105 g/ml) (Figure 4D).
To address the effect of the SAAS N2177Y in Vp447Δc on particle formation, Vp447Δc+5AlaN2177 was created. 75 ml of supernatant of SK6-cells 48 h after transfection with genomes of either Vp447Δc+5AlaN2177 or Vp447Δc+5AlaN2177Y were subjected to equilibrium centrifugation (Figure S3). Highest levels of infectivity were recorded at a density of 1.117 g/ml for Vp447Δc+5AlaN2177Y and at 1.102 g/ml for Vp447Δc+5AlaN2177. Both infectivity and RNA-levels were reduced more than 400-fold in Vp447Δc+5AlaN2177 in comparison to Vp447Δc+5AlaN2177Y in all fractions tested. Overall, E2 levels were comparable between both viruses and peaked at 1.12–1.14 g/ml. However, the ratio of E2 homo- to heterodimer seemed to differ between the two viruses, as did the E2 levels at a density of 1.10 g/ml.
The nucleocapsid of Vp447 is likely composed of core protein and the viral genome but so far has not been characterized. To gain at least preliminary information about the nucleocapsid of Vp447 and whether an analogous structure exists in Vp447Δc+5AlaN2177Y, either virus was treated with a nonionic detergent (0.5% NP40) to remove the envelope prior to equilibrium centrifugation as described above. The treatment completely abrogated infectivity in the fractions recovered and viral RNA levels were reduced more than 100-fold for either virus in comparison to untreated virus. RNA levels were just above background and peak levels occurred at densities of 1.05 g/ml and 1.2 g/ml for Vp447Δc+5AlaN2177Y whereas a broad peak of genomic RNA could be detected at densities of 1.11–1.2 g/ml for Vp447 (Figure 4A). To increase precision of the analysis, both viruses were mixed, treated with 0.5% NP40 and analyzed in the same gradient. The E2 signal was shifted towards the top of the gradient (1.04–1.14 g/ml), whereas weak core signals could be detected at higher densities (1.13–1.18 g/ml) (Figure 4C). Viral genome of Vp447 was detected in highest amounts at densities of 1.14–1.2 g/ml, whereas highest levels of Vp447Δc+5AlaN2177Y genome were now observed at densities of 1.17–1.19 g/ml and 1.22 g/ml (Figure 4D). These results indicate that detergent treatment of Vp447 in fact releases nucleocapsids of higher density. This assay is complicated by the RNase activity of the structural protein Erns, which might result in degradation of the viral genome after lysis of the lipid envelope. Hence, both Vp447 (Vp447_H30K) and Vp447Δc+5AlaN2177Y (Vp447Δc+5AlaN2177Y_H30K) with an exchange of Erns residue histidine 30 to arginine, destroying the active centre of its RNase, were generated [25]. This aa exchange did not affect the amount of progeny virus produced (Figure S4, Figure S5). Both viruses were subjected to equilibrium density centrifugation to compare them with the respective parental virus. No differences were present regarding the amount and distribution of E2 (Figure S4; data for Vp447Δc+5AlaN2177Y_H30K not shown). After detergent treatment, RNA levels of Vp447 and Vp447_H30K as well as of Vp447Δc+5AlaN2177Y and Vp447Δc+5AlaN2177Y_H30K remained at low levels (Figure S4, Figure S5).
Size exclusion chromatography was performed to directly compare the Stokes diameter of Vp447 and Vp447Δc+5AlaN2177Y. For this purpose, a mixture of 108 ffu of each Vp447 and Vp447Δc+5AlaN2177Y was subjected to gel filtration using Superose 6. Infectivity was detectable in fractions 40–78. Real-time RT-PCR (as described above) differentiating Vp447 from Vp447Δc+5AlaN2177Y allowed detection of viral genomes in fractions 43–78. Peak levels of genomes of either virus were observed in fractions 59–61 and coincided with peak infectivity (Figure 5).
For electron microscopic inspection, virus was produced in SK6 cells in serum free medium, concentrated by ultracentrifugation and inspected by TEM. The identity of the virions was confirmed by immunogold (10 nm) staining with a monospecific rabbit serum against Erns (for specificity of this serum, see Figure S6). In both preparations, pleomorphic particles of about 50 nm were detectable. No morphological changes were apparent between Vp447 and Vp447ΔcN2177Y particles (Figure 6). Mean size of Vp447 particles was 51.9 nm (standard deviation 8.9 nm; n = 43) and of Vp1017 particles 50.1 nm (standard deviation 9.3 nm; n = 34). However, no exact size comparison or tomographic particle analysis was possible since required particle quantity, quality and purity was not achieved.
To address whether the absence of core protein in the virus particle affects physical stability of Vp447ΔcN2177Y, the kinetics of inactivation of Vp447 and Vp447ΔcN2177Y at 37°C and 39.5°C were determined. No major differences in thermal stability were observed between the two viruses (Figure S7). Physical stability was also assessed by freezing and thawing of defined virus preparations. After thawing, 19% of the initial virus input could be recovered for Vp447 and 13% for Vp447ΔcN2177Y (Figure S7).
CSF is a disease of pigs with strain dependent virulence. Vp447 represents a moderately virulent strain [26], causing mortality rates >50%. To assess virulence of Vp447ΔcN2177Y, a small-scale animal experiment was conducted. Two groups of two pigs each were injected intramuscularly with 5×106 TCID50 of Vp447 or Vp447ΔcN2177Y. Two days later, a sentinel pig was added to each group. Animals were evaluated according to a standard clinical scoring system [27], rectal temperature and leukocyte counts. Vp447 infected animals exhibited febrile temperatures (>40°C) on day 7–10 after infection and from day 13 after infection until the end of the experiment (Figure 7A). One Vp447 infected pig (wt2) had to be euthanized on day 21 after infection, with a clinical score of 10. The other Vp447 infected pig (wt1) had a clinical score between 2.5 and 4.5 on days 17, 18 and 21–27. Severe leukopenia (leukocyte count below 10 Giga/l), a typical symptom of CSF [reviewed with other clinical symptoms by 28], was present in wt1 and wt2 from day 4 after infection, with further declining leukocyte counts until the end of the experiment (Figure 7B). The sentinel animal (wtS) housed together with the Vp447 infected pigs developed febrile temperatures from day 14 after infection until the end of the experiment and leukopenia was present on day 21 and 28 of the experiment. Virus could be isolated from Vp447 infected animals on days 4, 7, 10 and 14 after infection (Table 2). Virus isolation was not possible from the sentinel animal on days 4, 7, 10 and 14 after infection of the other pigs. Neutralizing antibodies could not be detected in Vp447 infected animals and their sentinel on days 10, 14 and 21 after infection (Table 3). No apparent signs of disease (clinical score = 0) were observed for Vp447ΔcN2177Y infected animals (Δc1 and Δc2) and their sentinel (ΔcS) throughout the experiment. With the exception of one day of slightly elevated body temperature (Δc2 on day 8) and mild leukopenia of animal Δc1 on day 21, no fever or leukopenia were present in Vp447ΔcN2177Y infected animals (Δc1 and Δc2) and their sentinel (ΔcS). We were unable to reisolate Vp447ΔcN2177Y from sera (Table 2) and leukocytes (not shown) of infected animals on days 2, 4, 7, 10 and 14 after infection. However, viral genomes could be amplified from leukocytes until day 7 and neutralizing antibodies could be detected beginning with day 14 after infection (Table 3).
Key findings of this study are that (1) a pestivirus lacking almost the entire core coding region is viable and that (2) viability depends on single point mutations in the helicase domain of NS3. This finding questions the general assumption that a core protein is a specific and essential structural element of enveloped RNA viruses and is supported by the existence of GBV- A and GBV- C, which do not encode an obvious core protein [reviewed by 1]. Further to this, the data support a central role of the multifunctional NS3 protein in virus particle assembly.
During the characterization of different loss - of - function manipulations of the core gene of CSFV, we observed that some replicative but initially poorly growing viruses generated increased amounts of progeny virus after extended incubation periods of the transfected cells. The responsible gain-of-function mutations could not be mapped to the locus of the manipulated nucleotide sequence. Instead, single nucleotide exchanges clustered within a stretch of approximately 300 nucleotides of NS3 helicase subdomain 3, about 6000 nucleotides downstream of the core gene. The occurrence of second site mutations in NS3 upon loss of core protein function differs from results described for tick-borne encephalitis virus. In this model, the deletion of parts of the internal hydrophobic domain led to the acquisition of hydrophobic residues in the core gene itself [29].
To confirm that the observed infectivity of core deficient viruses was due to proper virus particles, Vp447 and Vp447ΔcN2177Y were compared with regard to sensitivity towards neutralizing antibodies. In both cases, infectivity was blocked by hyperimmune sera from pigs or a monoclonal antibody directed against viral E2. Differences in the stability of particles of Vp447 and Vp447ΔcN2177Y with regard to infectivity were not observed upon freezing - thawing and heat exposure. Electron micrographs of Vp447 and Vp447ΔcN2177Y were obtained from concentrated serum-free cell culture supernatants and the structures observed were immunogold labelled with a monospecific rabbit serum against Erns. This was necessary because pestivirions in general lack a characteristic morphology. No morphological differences between Vp447 and Vp447ΔcN2177Y particles were apparent. Precise determination of structure and size would require cryo EM to avoid preparation dependent artifacts and also larger numbers of particles.
With regard to particle sizes no apparent differences in Stokes diameter could be detected between Vp447 and Vp447ΔcN2177Y particles in gel filtration experiments. Both viruses eluted from the column in the same fractions. Due to difficulties in comparing different gel filtration runs, it was mandatory to separate Vp447 and Vp447ΔcN2177Y side by side. To distinguish between both viruses by real time RT-PCR, a modified Vp447ΔcN2177Y was constructed, which encodes an additional sequence of five alanines between Npro C-terminus and signal peptide (Vp447Δc+5AlaN2177Y). This construct was also employed for determination of virus density in linear sucrose gradients. After it was evident from individual gradient experiments that infectivity and genomic RNA comigrated at densities from 1.10–1.11 g/ml (Figure 4), both viruses were mixed and analyzed in the same gradient. Again RNA, E2 and infectivity accumulated at the same densities. A surprising finding was that core protein showed highest concentration at slightly lower densities than peak RNA and infectivity levels. As E2 can be detected in the supernatant of cells transfected with the genome of Vp447Δc, it was of interest to compare the suspected pseudoparticles with regard to density and genome integration to Vp447ΔcN2177Y. Hence, equal volumes of supernatant of cells either transfected with Vp447Δc+5AlaN2177 or Vp447Δc+5AlaN2177Y genomes were subjected to density gradient centrifugation. The reduction of infectivity of Vp447Δc+5AlaN2177 in comparison to Vp447Δc+5AlaN2177Y correlated with the reduction of genome levels at the densities tested, suggesting that the SAAS N2177Y is critical for the integration of the virus genome into the particles during assembly if core protein is not present. Overall, E2 levels between the two viruses were comparable both with regard to total amount in the supernatant and distribution according to density. However, the ratio of E2 homo- to heterodimer, as well as the amounts of E2 at a density of 1.1 g/ml and the density of peak infectivity differed between the two viruses, which might indicate differences in particle composition.
To determine whether core protein actually is a component of a nucleocapsid structure, the envelope of the virus particles was removed by treatment with a non-ionic detergent (Nonidet P40). NP40 treated viruses were layered on top of sucrose gradients as before and the position of infectivity, E2, core and RNA were recorded after equilibrium centrifugation. Infectivity could be abolished completely by NP40 treatment. The signal of E2 shifted towards the top of the gradient (1.04–1.14) whereas the core protein signal shifted to higher densities (1.13–1.18). Peak values of viral genomes coincided with core signal in Vp447, which might implicate the presence of a nucleocapsid like structure of higher density. For Vp447Δc+5AlaN2177Y, signals for viral genome were low, with a slight elevation at the tube bottom if the virus was separately run on a gradient. Hence, we were unable to assign the genome to a discrete density. In contrast, a slight peak of viral RNA, comparable in density to Vp447, was observed if both viruses were separated in the same gradient. One could speculate that this effect is due to a redistribution of core protein between viral genomes after detergent treatment. Overall, the amounts of RNA determined by real time RT-PCR were 102–104 lower than with intact viruses, which can be taken as evidence for RNA degradation. A comparable experiment for HCV determined only a six-fold reduction of genomic RNA after NP40 treatment [30]. A major difference between HCV and CSFV is the presence of the potent ribonuclease Erns in the virus envelope [31]. However, after mutational disruption of the RNase active centre of Erns [25], we did not observe changes in the levels of viral genome detectable in comparison to virus with intact RNase. This suggests that the analytic system itself, by employing sucrose, contains RNases, which together with the long centrifugation time (24 h), are sufficient to degrade most of the viral genomes present in the sample. To address this technical problem, improved separation methods have to be established to minimize RNA degradation. However, the relatively higher amount of viral genome detectable for Vp447 in comparison to Vp447ΔcN2177Y suggests a protective function of core protein against RNase.
The absence of core protein and thus a known proteinaceous component of the nucleocapsid questions the way how a linear viral RNA molecule of approximately 3 µm is condensed in order to fit into the virus particle of less than 50 nm diameter. Further to this, the genome has a negative charge that is partially neutralized by a usually positively charged (nucleo-) protein. Strikingly, introduction of the single amino acid substitution N2177Y into the parental Vp447 (Vp447N2177Y) reduced virus growth and abrogated the detectable incorporation of core protein into the virus particles, while at the same time the core protein accumulated intracellularly. This points to an ability of modified NS3 to counteract core particle integration, probably by modulation of core-RNA-interaction. This finding also raises the question whether NS3 might replace core in the virus particle. So far, we were unable to detect any NS3 in purified virus preparations, but we cannot exclude that a small number of molecules is packaged.
As we have no evidence for other virally encoded proteins for replacement of the missing core protein, it is conceivable that host cellular proteins, for example cytoplasmic RNA chaperones or nuclear RNA binding proteins, compensate for the lack of core protein. The association of cellular proteins with virus particles has been described for RNA and DNA viruses, like hepadnaviruses [32], rabies virus [33], filoviruses [34], respiratory syncytial virus [35] and HCV [36]. Interestingly, HSP70 or HSP90 were most often found associated with virus particles. An important task will therefore be a proteome analysis of highly purified virus particles of Vp447 and Vp447ΔcN2177Y. Epitope tagged viruses - as described for HCV [37], [38] and BVDV [39] - may be useful for such an investigation.
NS3 is functionally well conserved among members of the Flaviviridae and significant sequence conservation is apparent. It is a multifunctional protein that contains several enzymatic activities, such as serine protease, NTPase and RNA helicase [18]–[22]. Its involvement in particle assembly has been suggested for HCV [11], [13] and YFV [12], [40], [41]. The conserved helicase motifs are located in subdomains 1 and 2 of the NS3 helicase [42]. NS3 helicase subdomain 3 is the least conserved stretch in NS3 of Flaviviridae, both with regard to amino acid sequence and structure [43]. Although it is not present in all superfamily 2 helicases [44], it is essential for NS3 helicase activity. Analysis of all single aa substitutions in the putative CSFV NS3 helicase subdomain 3, which were able to rescue Vp447ΔcN2177Y, did not reveal an obvious pattern with regard to amino acid identity, charge or polarity, hence we are not able to draw conclusions about the mode of action by analysis of the sequence identities. So far, the 3D-structure of pestiviral NS3 helicase is not known and the sequence homology to HCV NS3 is too low to draw conclusions. All rescue mutations were located in regions aligning with alpha helices both in dengue virus [45] and HCV [46], [47] (Figure S8). All but one aa substitution identified were located in stretches reported to be important for NS3 helicase protein-protein-interaction and optimal replication of HCV [48]. So far, there is no mechanistic explanation how the described mutations in NS3 helicase domain 3 allow for the rescue of Vp447Δc. Structural and functional analysis of the modified NS3 proteins are needed to elucidate the gain of function in particle assembly.
Finally, the virulence of Vp447ΔcN2177Y in comparison to Vp447 was assessed in a small scale animal experiment. The parental CSFV strain used for this study causes disease in pigs with a case fatality rate of >50% [26]. While the two pigs infected with Vp447 and the sentinel housed together with these two pigs developed typical signs of CSF, the pigs infected with Vp447ΔcN2177Y and the respective sentinel animal stayed completely healthy although they were injected with the same dose of virus. Neither fever nor leukopenia was observed in pigs infected with Vp447ΔcN2177Y. Detection of genomic RNA in leukocytes up to day 7 p.i. and the appearance of CSFV neutralizing antibodies in both Vp447ΔcN2177Y infected animals beginning at day 14 suggest that a limited replication took place in the animals, despite our inability to reisolate Vp447ΔcN2177Y from serum or blood cells. This indicates that the lack of core protein leads to a strong attenuation of the virus. The sentinel pig developed no neutralizing antibodies, which can be taken as evidence that Vp447ΔcN2177Y is not or inefficiently transmitted. All this points to an important role of pestiviral core protein in vivo. Further effort will be put in the characterization of Vp447ΔcN2177Y in primary cells of its natural host to elucidate the mechanisms underlying its attenuation.
All animal work was conducted according to the legal regulations of the German Animal Welfare jurisdiction (Tierschutzgesetz). The animal experiment was subject to authorization and was recorded after approval under reference number AZ 06/1105 at the Lower Saxony State Office for consumer protection and food safety. The internal reference was V2006-6.
Sequence modifications were introduced into the core or NS3 protein of CSFV Alfort/Tübingen recombinant full length cDNA clone (p447) by site directed mutagenesis or end to end ligation, utilizing Pfu-DNA polymerase (Promega, Mannheim, Germany) (Primers are available upon request). Sequence analysis was employed to confirm the generated constructs (Quiagen, Hilden, Germany).
SK6-cells were grown in Dulbecco's modified Eagle's medium supplemented with 10% fetal calf serum at 37°C under 5% CO2. Virus cDNA was transcribed into RNA using SP6-polymerase (NEB, Frankfurt am Main, Germany) and, typically, 2.5 µg RNA were electroporated into 5×106 SK6-cells (Bio-Rad Gene Pulser). Replication was assessed 14 h after electroporation via immunohistochemistry using monoclonal antibody A18, directed against the CSFV E2 protein. Virus titer was determined in focus-forming units/ml (ffu/ml) 24 h after electroporation. For this purpose, supernatant was harvested, clarified (5 min at 3,000×g), and seeded on SK6-cells, employing 10-fold dilution steps. After 14 h, cells were fixed and stained for E2 as mentioned above. Antigen-positive foci of infected cells were counted using a Nikon Eclipse TS100 microscope and the titer was calculated. All virus titers were confirmed by multiple experiments (more than two).
For virus passaging, cell culture supernatant was harvested 72 h after electroporation of genomic RNA and clarified by centrifugation (5 min at 3,000×g). Consecutively, 2×105 SK6-cells were infected with 1 ml of supernatant of the previous passage. This procedure was repeated every 3 to 4 days along with the determination of virus titers.
Virus neutralization was tested according to [49]. Briefly, serum samples from a CSFV vaccinated (S05) and a vaccinated and infected (S98) animal, as well as cell culture supernatant containing an anti-E2 antibody (A18) and a serum of an animal neither infected nor vaccinated against CSFV were diluted 2-fold in duplicates on a 96well plate (sera were kindly provided by the Community Reference Laboratory for CSF, Hannover). Thereafter, a defined virus suspension of Vp447 was added to each well and the plate was incubated for 1 h at 37°C. Subsequently, the employed virus suspension was back titrated on the plate, a suspension of SK6-cells (3×105 cells/ml) was added to each well and the plates were incubated at 37°C for 72 h. Virus infection was detected by immunohistochemistry as described above. TCID50/ml of the employed virus suspension and ND50/ml were calculated according to [49].
Western blotting was done essentially as described by (8). Briefly, 24 h–72 h after electroporation, cells were lysed in Tris-EDTA buffer containing 2% SDS, subjected to SDS-PAGE on 7.5, 10 or 12% polyacrylamide gels using Tris-tricine buffers, and blotted to nitrocellulose. As primary antibody, mouse monoclonal antibody A18 (anti-E2), 5H4 (anti-Core), 24/16 (anti-Erns), code 4 (anti-NS3), 6B2 (anti-NS5B) or anti-β-actin antibody (A5441; Sigma-Aldrich) was utilized. Horseradish peroxidase-coupled goat anti-mouse antibody served as secondary antibody (Dianova, Hamburg, Germany). Signals were revealed using chemiluminescence (ThermoFisher, Bonn, Germany) and exposure to Kodak BioMax film.
Virus-containing supernatants were concentrated for immunoblotting by clarification for 5 min at 3,000×g, followed by pelleting of 1.2 ml in a TL100 Beckmann ultracentrifuge at 45,000 rpm for 1 h. After removal of the supernatant, the pellet was resuspended in 10 µl Tris-EDTA buffer containing 2% SDS and further processed as described for the cell lysate. Signals were quantified employing ImageJ (http://rsbweb.nih.gov/ij/index.html).
All constructs were confirmed by sequencing (Quiagen, Hilden, Germany). Revertant viruses were analyzed by sequencing after reverse transcriptase (RT)-PCR and cloning into the pGEM-T vector (Promega, Mannheim, Germany) using standard primers (oligonucleotide sequences are available upon request).
Continuous sucrose gradients (10%–60% w/v sucrose in 50 mM Tris, pH 7.4) of 11 ml were generated with a GP250 gradient programmer in conjunction with two Pharmacia P500 pumps at a flow rate of 1 ml/min. In a volume of 400 µl, 106 ffu of each Vp447 and a Vp447 with a deletion of core protein (aa 170–246 of the polyprotein) and a five alanine linker between Npro C-terminus and signal peptide (Vp447Δc+5AlaN2177Y) were layered on top of the gradient and centrifuged in a Beckman SW41 rotor at 180.000 g (32.00 rpm) for 24 h. 30 fractions of 360 µl each were collected by bottom puncture and the refractive index was determined. 30 µl of each fraction were used for titration on SK6-cells and 20 µl of two fractions pooled were subjected to Western blot analysis.
Viral RNA was purified utilizing the QuiaAmp Viral RNA kit (Quiagen, Hilden, Germany) according to the manufacturer, reverse transcribed employing the Quanti Tect Reverse Transcription kit (Quiagen, Hilden Germany) with the same reverse primer (rev: CATCCCGCGTATCTCTT) and subjected to qPCR (Quanti Tect SYBR Green PCR kit, Quiagen, Hilden, Germany) in a StepOnePlus real-time PCR system (Applied Biosystems, Darmstadt, Germany), using forward primer specific for either Vp447 (for_wt: CAAGCCACCAGAGTCCAG; fragment size 258 nt) or Vp447Δc+5AlaN2177Y (for_Δc: TGCGGCCGCAGCTCTAGA; fragment size 246 nt) and the reverse primer already employed in the reverse transcription reaction.
1×108 ffu of each Vp447 and Vp447Δc+5AlaN2177Y were pelleted at 100,000×g for 1 h in a 45Ti rotor in a Beckman L8–70 ultracentrifuge. The pellet was resuspended in 550 µl 1xTNE buffer overnight at 4°C on a shaker. The complete volume was loaded onto a Pharmacia XK16 gel chromatography column, packed with Superose 6 (prep grade, GE Healthcare, Munich, Germany) with a total volume of 138 ml (determined by dextran-blue) including the void volume of 41.5 ml (determined by 10% acetone in H2O and subsequent measurement of optical density at 280 nm). The column was calibrated employing IgM (size 21 nm), which was subsequently measured in the elution fractions by agar gel diffusion (Novartis, Marburg, Germany). The chromatography was performed at a flow rate of 6 ml/h generated by a LKB P-1 pump with 1xTNE buffer. 80 fractions of 2 ml each were collected by a LKB superfrac collector. Collector tubes were blocked with 1xTNE containing 1% BSA fraction 5 for 10 min at room temperature. RNA was prepared from the resulting fractions by QuiaAmp Viral RNA kit (Quiagen, Hilden, Germany) and analyzed for the presence of viral genome by above described real-time RT PCR for the presence of either Vp447 or Vp447Δc+5AlaN2177Y genome.
SK6 cells transfected with either Vp447 or Vp447ΔcN2177Y genome were seeded on 10 143 cm2 cell culture plates each in medium containing FCS. 18 h after transfection, the cells were washed twice with PBS and the medium was replaced by a serum free medium for MDBK cells (Sigma-Aldrich, Munich, Germany). 48 h after transfection, the supernatant was harvested and cellular debris was removed by centrifugation (5 min at 3,000×g). Subsequently, virus was pelleted at 25.000 rpm in a TI45 rotor for 8 h. Thereafter, the pellet was resuspended in PBS for 12 h at 4°C. Virus preparations were mounted on glow discharged, pioloform and carbon coated copper-rhodium grids. After saturation using 1% (w/v) bovine serum albumin (BSA) in PBS grids were transferred to droplets of the first antibody: monospecific rabbit serum anti Erns, 1∶200 in PBS, 0.5% (w/v) BSA for 1 h in a humid chamber. After 5 washing steps on droplets of PBS immune labeling was completed using goat anti-rabbit IgG conjugated to 10 nm colloidal gold (Plano, Wetzlar, Germany) 1∶25 in PBS, 0.5% (w/v) BSA. The preparation was finished by 5 washing steps on PBS followed by short incubation on distilled water and negative staining using 2% methylamine tungstate (Plano, Wetzlar, Germany). Air dried grids were examined in a Zeiss EM910 transmission electron microscope at 80 kV at an instrumental magnification of 31.500 and 50.000 and micrographs taken on Kodak SO-163 negative film.
Six weaner pigs were purchased from a commercial piggery and tested negative for infection with Pestiviruses by RT-PCR and serum neutralization test. The pigs were kept in two separately housed groups under high containment conditions. Two pigs of each group were either infected intramuscularly with 5×106 TCID50 Vp447 or Vp447 with a deletion of core amino acids 170–246 (position in the polyprotein) (Vp447ΔcN2177Y). Two days after infection, the previously separated sentinel animal was returned to each group. The animals were monitored daily for clinical signs of CSFV according to a modified clinical score developed by [27] and body temperature was recorded. The clinical score is calculated by scoring each parameter (liveliness/body tension/body shape/breathing/walking/skin/eyes+conjunctiva/appetite/defecation) from 0–3 (no signs of disease – severe signs of disease), followed by addition of all values obtained. As the animals were housed in groups in this experiment, the parameter “leftovers in feeding trough” could not be evaluated for an individual animal. EDTA blood samples were taken on days 2, 4, 7, 14, 21 and 28 after infection. The leukocyte fraction was isolated from EDTA blood by addition of 6.25% (v/v) 5% EDTA-Dextran solution, followed by sedimentation and several wash steps with PBS [49] and the leukocyte count was determined in a Neubauer chamber. Animals were euthanized because of animal welfare reasons (clinical score >20 or severe disease) during the experiment or at the end of the experiment.
|
10.1371/journal.ppat.1004595 | Different Infectivity of HIV-1 Strains Is Linked to Number of Envelope Trimers Required for Entry | HIV-1 enters target cells by virtue of envelope glycoprotein trimers that are incorporated at low density in the viral membrane. How many trimers are required to interact with target cell receptors to mediate virus entry, the HIV entry stoichiometry, still awaits clarification. Here, we provide estimates of the HIV entry stoichiometry utilizing a combined approach of experimental analyses and mathematical modeling. We demonstrate that divergent HIV strains differ in their stoichiometry of entry and require between 1 to 7 trimers, with most strains depending on 2 to 3 trimers to complete infection. Envelope modifications that perturb trimer structure lead to an increase in the entry stoichiometry, as did naturally occurring antibody or entry inhibitor escape mutations. Highlighting the physiological relevance of our findings, a high entry stoichiometry correlated with low virus infectivity and slow virus entry kinetics. The entry stoichiometry therefore directly influences HIV transmission, as trimer number requirements will dictate the infectivity of virus populations and efficacy of neutralizing antibodies. Thereby our results render consideration of stoichiometric concepts relevant for developing antibody-based vaccines and therapeutics against HIV.
| Our estimates of the HIV-1 entry stoichiometry, that is the number of envelope glycoprotein trimers needed to mediate fusion of viral and target cell membrane, close an important gap in our understanding of the HIV entry process. As we show, stoichiometric requirements for envelope trimers differ between HIV strains and steer virus entry efficacy and virus entry kinetics. Thus, the entry stoichiometry has important implications for HIV transmission, as demands on trimer numbers will dictate the infectivity of virus populations, target cell preferences and virus inactivation by trimer-targeting inhibitors and neutralizing antibodies. Beyond this, our data contribute to the general understanding of mechanisms and energetic requirements of protein-mediated membrane fusion, as HIV entry proved to follow similar stoichiometries as described for Influenza virus HA and SNARE protein mediated membrane fusion. In summary, our findings provide a relevant contribution towards a refined understanding of HIV-1 entry and pathogenesis with particular importance for ongoing efforts to generate neutralizing antibody based therapeutics and vaccines targeting the HIV-1 envelope trimer.
| To infect cells, HIV-1 virions need to fuse their membrane with the target cell membrane, a process triggered by the viral envelope (env) glycoprotein trimer [1], [2]. Due to its key function in the virus life cycle and as prime target for neutralizing antibodies and entry inhibitors, analyses of env trimer structure and function remain in the focus of current HIV vaccine and drug research [3]–[5]. Each env trimer consists of three heterodimeric protomers, composed of the non-covalently associated gp120 surface and gp41 transmembrane subunits. Binding of gp120 to the primary receptor CD4 on target cells triggers conformational changes in gp120 that expose the binding site of a co-receptor, most commonly CCR5 or CXCR4 [6]. Subsequent co-receptor binding activates the gp41 transmembrane subunits, which triggers a prototypic class I fusion process via insertion of the N-terminal fusion peptides into the target cell membrane. Refolding of the gp41 N- and C-terminal heptad repeat regions into six-helix bundles drives approximation and fusion of viral and target cell membranes [1], [7], [8].
While the HIV entry process has been defined in considerable detail, we currently lack information on the stoichiometric relations of interacting molecules. Likewise, the thermodynamic requirements of membrane fusion pore formation and pore enlargement, enabling passage of the viral core into the target cell cytoplasm, are only partially understood [9]–[11]. The energy required for the entry process is generated by structural rearrangements of the envelope trimer that follow receptor binding [7], [8], [12]. How many trimers must engage in receptor interactions (a number referred to as stoichiometry of entry) [13]–[15] in order to elicit the required energy to complete fusion has not been conclusively resolved. Whether HIV needs one or more trimers to complete entry will strongly influence virion infectivity and efficacy of neutralizing antibodies targeting the trimer. Previous studies resulted in contradicting stoichiometry estimates, suggesting that either a single trimer is sufficient for entry [13] or that between 5 to 8 trimers are required [14], [15]. In comparison, for Influenza A virus, which achieves membrane fusion through the class I fusion protein hemagglutinin (HA), postulated necessary HA trimer numbers range from 3 to 4 [16]–[18] to 8 to 9 [13]. Calculations based on the energy required for membrane fusion suggested that indeed the refolding of a single HIV envelope trimer could be sufficient to drive entry [7], [8]. Numerous lines of evidence however suggest that several env-receptor pairings are commonly involved in the HIV entry process. Electron microscopy analysis of HIV entry revealed the formation of an “entry claw” consisting of several putative env-receptor pairs [19], which is supported by biochemical analyses indicating that the number of CCR5 co-receptors needed for virus entry differs among HIV-1 isolates and requires up to 6 co-receptors [20], [21].
Precise delineation of the stoichiometry of entry, as we present it here, substantially contributes to our understanding of HIV pathogenesis by defining a viral parameter that steers virus entry capacity, potentially shapes inter- and intra-host transmission by setting requirements for host cell receptor densities, and by defining stoichiometric requirements for virion neutralization. The latter is of particular importance considering the ongoing efforts to generate neutralizing antibody based therapeutics and vaccines targeting the HIV-1 entry process [3], [4], [22].
To estimate the stoichiometry of entry (in the following referred to as T) we employed a previously described combination of experimental and modelling analyses [13]–[15]. Our strategy centers on the analysis of env pseudotyped virus stocks carrying mixed envelope trimers consisting of functional (wt) and dominant-negative mutant env, where a single dominant-negative env subunit incorporated into a trimer renders the trimer non-functional. We included envs of 11 HIV-1 strains in our analysis covering subtypes A, B and C and a range of env characteristics such as primary and lab-adapted strains, different co-receptor usage and different neutralization sensitivities (Table 1).
To derive estimates of T from mixed trimer experiments two key parameters need to be considered: the mean virion trimer numbers and the distribution of virion trimer numbers across a virion population [14], [15]. To assess virion trimer numbers, we determined p24 and gp120 content of purified virus stocks by ELISA. Although only an approximation, as also partially shed and non-functional trimers are accounted for, this analysis yielded upper limits of virion trimer content. We observed between 6 to 20 trimers per virion among the 11 strains probed (Table 1), which is in close agreement with previous estimates of HIV-1 virion trimer content [23]–[28]. While trimer incorporation into pseudotyped particles may be lower than on replication-competent virus [29], this does not preclude estimation of T as trimer content is controlled for in our mathematical analysis. Since available experimental data do not provide information on the distribution of trimers across virion populations (i.e. frequencies of virions with a given trimer number in the population), we utilized previously determined trimer number variation across virions in our modeling [26].
Key in our experimental design are dominant-negative env mutants. To obtain robust estimates of T we performed the experiments with two individual env mutations that both lead to a complete loss in entry capacity, either by a mutation of the Furin cleavage site (R508S/R511S) [13], [30], [31] or a mutation of the gp41 fusion peptide (V513E) [13], [32]. Importantly, expression levels of the mutant envs were in the range of 80 to 100% of the corresponding wt envs (Table 1), ascertaining that during mixed trimer experiments env expression levels on virions follows the ratio of wt and mutant env plasmids transfected into virus producer cells.
To assess T, mixed trimer expressing pseudovirus stocks of each strain were generated by transfecting producer cells with different ratios of mutant and wt env plasmids ranging from 0 to 100% of the dominant-negative mutant env. The resulting virus stocks were probed for infectivity on TZM-bl reporter cells and infectivity was plotted as function of mutant env content (Fig. 1A and B). Based on our model [15], differences in T result in different infectivity plots in this graphical analysis (Fig. 1A). Intriguingly, the 11 envs showed variable patterns suggesting that the strains differ in T (Fig. 1B). The mixed trimer infectivity data (Fig. 1B) and the mean virion trimer numbers (Table 1) were then used to infer T of each strain by our model. This resulted in estimates of T ranging from 1 to 7 trimers for the 11 HIV strains tested (Fig. 1C, S1 Table and S1 Fig.), with the majority of strains requiring 2 to 3 trimers for entry. Of note, the data derived with the V513E and R508S/R511S mutant returned closely matching results with identical estimates for T (n = 4) or estimates differing only by 1 (n = 6) (Fig. 1C). The only higher discrepancy was observed for the highly neutralization sensitive, T cell line adapted strain NL4-3 where the two env mutations appeared to have individual effects on the readout but both yielded estimates of T that were amongst the highest in the env panel (T = 7 with V513E and T = 4 with R508S/R511S).
To verify that the mathematical approach (in the following referred to as “basic model”) provides a valid estimate of T, we probed several alternative analyses of the data shown in Fig. 1B. These analyses incorporate previously described extensions of the basic model that account for additional parameters that could potentially influence data acquisition and analysis [15]. The model extensions showed for the majority of strains a significantly improved curve fit to the experimental data (S1 Table). However, these analyses frequently yielded highly divergent values of T and the additional model parameters included in the model extensions for the two dominant-negative mutants of the same strain (S1 Table and exemplified for the CAP88 Env in S2 Fig.). This was in stark contrast to the basic model where the two independent T estimates of each strain were with few exceptions in close agreement (Fig. 1C and S1 Fig.). As we can safely assume that the parameter estimates for the two different mutations of the same strain should be similar we can reject the model extensions. A further indication that the model extensions we probed are not valid in the context of our analysis was that the derived estimates of T were in many cases implausibly high whereas the basic model generated estimates that fit the described range of trimer levels on HIV virions. We are thus confident that the basic model we utilize for the estimation of T is valid and provides robust estimates.
The mean virion trimer number of the probed virus stocks is an important model parameter in our analysis of T and fluctuations in the mean virion trimer number may therefore influence the estimates. To test the influence of mean virion trimer number variation on our estimates of T we performed additional data analyses where instead of the measured individual trimer numbers (Table 1), identical mean trimer numbers for all 11 strains were assumed. We chose 3 values for this comparison that covered the range of trimer numbers measured across our panel: mean trimer numbers of 5 and 26 (representing the lowest and highest trimer contents measured in individual experiments) and a mean trimer number of 13, the mean of trimer numbers measured across our virus panel. Applying these trimer numbers to our data set we obtained estimates of T ranging from 1 to 17 trimers (S3A Fig.).
To determine the mean virion trimer numbers of the virus stocks we measured gp120 and p24 contents. This allows to derive virion numbers based on previously reported estimates of 1200 to 2500 p24 molecules per virion [23], [24], [33]. We chose an average estimate of 2000 p24 molecules per virion to derive the mean trimer numbers shown in Table 1. To investigate the influence of p24 assumptions on our analysis we also tested a higher p24 content estimate of 2400 molecules per virion as recently reported [33], consequently yielding 20% higher mean virion trimer numbers across all viruses. Employing these 20% higher trimer numbers in our analysis had only a modest effect on the T estimates yielding identical or slightly higher (mostly by one trimer) estimates of T (S3B Fig.). While these analyses confirm that absolute values of T vary depending on the mean virion trimer number assumed for the analysis, the differences in T among the 11 strains persisted, highlighting that they reflect qualitative entry properties of the respective envs. Hence, independent of the absolute mean trimer numbers, differences in T between viral strains can be detected by our approach.
As a further assay verification we tested the influence of target cells on our estimation of T. We reasoned that if our experimental approach truly measures the stoichiometry of entry, then the obtained data should be a sole function of the envelope trimer and not be influenced by target cell type and receptor density. We thus chose TZM-bl cells as target cells for their known reproducible performance and good signal to noise ratio in the luciferase reporter readout. Since these engineered cells overexpress the entry receptors of HIV and thus do not reflect features of physiological relevant target cells, we sought to verify that the obtained T estimates are indeed independent of the target cells used. To this end we chose two envelopes which yielded a low T estimate (primary isolate JR-FL, T = 2) and a high T estimate (lab-adapted strain SF162, T = 4 to 5) on TZM-bl cells and repeated the estimation of T on PBMC as target cells (Fig. 1D and E and S4 Fig.). As anticipated, we obtained for both viruses almost identical curves and T estimates as with the TZM-bl reporter cells, confirming that estimates of T are truly independent of the target cell type. Hence, use of TZM-bl cells for our assay setup is appropriate and the estimated T values are valid for physiologically relevant target cells of HIV-1.
The entry stoichiometry of a strain can be expected to influence virus population infectivity as strains with a low T will benefit from a higher proportion of the virus population carrying the required minimum trimer number (Fig. 2A). To directly probe the influence of T on virus infectivity we assessed the in vitro infectivity of the 11 HIV-1 strains in our panel. Of note, in the context of pseudoviruses infectivity is solely determined by the Env genes. Intriguingly, infectivity proved to be inversely correlated with T (r = −0.635, p = 0.036; Fig. 2B) indicating that strains that accomplish entry with low T are more infectious than strains with high T. Of note, we observed very divergent infectivities also for strains with very similar estimates of T (Fig. 2B). This is likely caused by different mean trimer numbers of the strains, as the mean trimer number in conjunction with T dictates virion population infectivity (Fig. 2A, C and D). For instance, amongst the viruses with T = 2 strain P3N has the highest infectivity and highest mean virion trimer number (20.3) whereas ZM214, the strain with lowest infectivity also has the lowest mean virion trimer number (6.7) measured across these viruses (Table 1). It can expected that additional factors beyond T and trimer numbers, such as propensity to shed gp120 or differential affinity for CD4, which are not covered by our analysis, may further contribute to different infectivity of the strains.
To investigate the interplay between entry stoichiometry and infectiousness of a virus population in more detail, we performed mathematical analyses of the relation between entry stoichiometry and trimer numbers per virion of a virus population. We found that indeed the entry stoichiometry steers virus population infectivity, with a higher entry stoichiometry resulting in a lower fraction of potentially infectious virions (Fig. 2C and D). Hence, the T of a strain and the therewith linked entry capacity may potentially contribute to the infectious to non-infectious particle ratio which is known to be low for HIV-1 [24].
To further explore the relation between virus infectivity and T we analyzed envs with deletions of the gp120 variable loops 1 and 2 (V1V2) and compared them to the matching wildtype envs. As we and others have previously shown, V1V2 deletion causes a dramatic reduction of virus infectivity through impairment of trimer integrity (Fig. 3A and [34]–[38]). When we probed T and compared the infectivity curves of the wt and V1V2-deleted env pairs, we observed distinct curve shifts of the V1V2-deleted envs across the majority of strains (Fig. 3B and C). Indeed, T of the V1V2-deleted envs proved significantly increased compared to the matching wt envs (Fig. 3D; mean T of 3.1 for wt envs versus mean T of 6.85 for V1V2-deleted envs; paired t test p = 0.0069). Importantly, this reduction in entry efficiency and the ensuing high estimates for T upon V1V2 deletion are not simply caused by reductions in env content of these virions, as V1V2-deleted env is expressed to similar levels on virions as the corresponding wt env (80–100% of wt, S5 Fig.). While expression levels of trimers certainly influence the estimates of T, we verified that the observed env content reduction of V1V2 deleted viruses was too low to inflict an overestimation of T (S5 Fig.) highlighting that indeed functional properties and not quantity of the respective wt and ΔV1V2 envelopes are decisive in defining T.
To further investigate the interplay between trimer numbers and T we produced pseudoviruses which expressed JR-FL wt and JR-FL ΔV1V2 with a deletion of the gp41 cytoplasmic tail (CT) as this is known to lead to an increased incorporation of trimers into virions [39], [40]. Indeed, CT deletion resulted in approximately 2-fold increased levels of trimers on virions (S6A–S6B Fig.). In support of the strong interplay between virion trimer numbers and infectivity thresholds defined by T, the infectivity of both viruses upon CT deletion was increased (S6C Fig.). Intriguingly, the increase in infectivity upon CT deletion was higher for JR-FL ΔV1V2 (9-fold) compared to JR-FL wt (2-fold), highlighting that envelopes with a reduced entry capacity, as here JR-FL ΔV1V2, benefit more if virions carry higher trimer numbers and thus meet the stoichiometric requirements for entry (S6D–S6E Fig.).
The number of trimers required for HIV entry likely influences virus infectivity in many ways. Besides determining a threshold trimer content that renders virions infectious, different T's could also manifest in different kinetics of the entry process as viruses with higher T may require more time to recruit and engage the necessary number of trimer-receptor pairings. To determine virus entry kinetics we employed a time-of-inhibitor addition experiment to derive the time required per virus strain to reach 50% of entry into target cells (Fig. 4A and S7A–S7B Fig.). Synchronized infection following spinoculation and temperature arrest in this assay setup allows assessment of entry kinetics solely as factor of envelope function post attachment to the target cells. When comparing the entry kinetics of the wt and V1V2-deleted strains we found that V1V2-deleted envs showed significantly delayed entry into target cells (Fig. 4B; mean times to 50% entry 19.9 minutes for wt envs and 37.7 minutes for V1V2-deleted envs; paired t-test p = 0.0002). As stated above, a potential explanation for this is that more time is required for V1V2-deleted strains to assemble a sufficient number of trimers in the virus-target cell contact zone to achieve entry. Indeed, we found a strong correlation between the estimated T and half-maximal entry time when all viruses, wt and V1V2-deleted strains, were analyzed (Fig. 4C; r = 0.568, p = 0.0073) but also for wt envs alone (r = 0.649, p = 0.0307), highlighting that entry stoichiometry and entry kinetics are tightly linked. The fact that we observe a significant correlation between estimated T and half-maximal entry time does however not exclude that additional processes beyond the recruitment of the necessary number of trimer-receptor pairings also influence the entry kinetics. Rates of CD4 and co-receptor binding and speed of the ensuing conformational rearrangements may differ between strains [41] and thereby contribute to the overall variance in entry kinetics. Additionally, for virions with a high T it must be considered that formation of the required number of contacts with the target cell may need longer time periods during which virions may detach again or decay before entry is completed [42].
To further explore the relationship between the entry stoichiometry and infectivity we performed additional studies with the subtype C strain CAP88 [43], which had the highest T and lowest infectivity within our panel (Fig. 1B, 1C and 2B). CAP88 is a transmitted/founder virus which carries a lysine (K) at position 160 of gp120, a site frequently targeted by neutralizing antibodies [44]. Among 4894 Env sequences deposited in the Los Alamos HIV Sequence Database asparagine (N) at position 160, as part of an N-linked glycosylation sequon, is with 93.3% the most prevalent residue at position 160 [45]. Loss of this glycosylation site is both associated with escape from PG9/PG16-like antibodies and decreased entry capacity [38], [46]–[48]. Supporting this we found that reconstitution of the N-linked glycosylation site (K160N) in CAP88 results in a 4-fold increase in virus infectivity (Fig. 5A and [38]) highlighting the importance of N160 for env functionality. To probe if the increased infectivity of the CAP88 K160N mutant may be due to changes in trimer structure and function that result in a reduction of T, we analyzed T of CAP88 wt and CAP88 K160N (S8A–S8B Fig.). Indeed, we found that the increased infectivity of CAP88 K160N is reflected by a decreased T (Fig. 5B and S8C Fig.). As this example highlights, changes in trimer structure inferred by naturally occurring mutations, possibly due to antibody escape, can result in a decreased entry capacity of the respective env, which in turn is reflected by an increase in the stoichiometry of entry.
The finding that a single point mutation in the CAP88 env could dramatically alter entry fitness and entry stoichiometry prompted us to further explore the influence of point mutations on env entry phenotype. To this end we selected three JR-FL variants mimicking resistance mutants as they may occur in vivo during neutralization escape: the JR-FL D664N escape mutant resistant to the MPER antibody 2F5, the JR-FL V549M N554D mutant which has a highly increased resistance against the entry inhibitor T-20 [49], and a JR-FL env with point mutations N332S P369L M373R and D664N rendering it resistant against the broadly neutralizing antibodies (bnAbs) PGT128, 2G12, b12 and 2F5 (S9 Fig.). While all three JR-FL variants showed similar mean virion trimer numbers as JR-FL wt, we observed differences in env infectivity with the 2F5 escape mutant infecting equally well as JR-FL wt whereas the T-20 and the multiple bnAb escape mutant env showed strongly reduced infectivity at 9% and 19% of JR-FL wt, respectively (Fig. 6A). When we compared the three escape mutant envs and JR-FL wt in the mixed trimer assays we observed a distinct curve shift for both dominant negative mutants of the T-20 escape variant (Fig. 6B), while the other three envs gave almost identical curves. Mathematical analysis of the data indeed revealed that the T-20 escape mutant requires 4 to 6 trimers for entry while all other mutants, like JR-FL wt, require 2 trimers (Fig. 6C). Hence, the T-20 escape mutant showed both, an increase in T and a loss in infectivity while the bnAb escape mutant env maintained T despite showing infectivity loss, confirming our earlier findings that viruses with the same T can still show a wide variation of infectivities (Fig. 2B). This can possibly be attributed to increased trimer decay rates or variations in CD4 and co-receptor engagement, especially since the bnAb escape mutant carried mutations in the CD4 binding site. Interestingly, the increased demand of the T-20 escape mutant for trimers during entry was also reflected in delayed entry kinetics of this env variant (Fig. 6D) [50]. These entry characteristics of the T-20 escape mutant are intriguing as the resistance mutations lie in the heptad repeat region of gp41 and interfere with six-helix bundle formation, which is a key step providing energy for membrane fusion during the entry process [51]. Thus, it is mechanistically plausible that the mutant env may require more trimers for entry to compensate for the disturbed six-helix bundle formation and generate sufficient energy to achieve membrane fusion.
Fusion of biological membranes, as required for entry of enveloped viruses, occurs in a plethora of cellular processes. In the case of HIV, fusion is executed by envelope glycoprotein trimers upon interaction with adequate receptor molecules on the target cell membrane [1], [8]. While the principle steps are known and thought to be shared across different biological systems and membrane types [7], [52], [53], the exact mechanisms and stoichiometric and thermodynamic requirements of most membrane fusion processes are not completely resolved [1], [8], [10], [11], [54], [55]. Definition of the components involved in HIV entry and the membrane fusion process is of particular interest as improved understanding of the determinants of HIV entry bears the promise to funnel the development of enhanced strategies to prevent and treat viral infections [1], [7]. The efficacy of HIV entry shapes inter- and intra-host transmission and determines the vulnerability to a range of therapeutic and preventive strategies such as neutralizing antibodies, entry inhibitors and antibody based vaccines. Considering that the stoichiometry of entry defines the number of trimers required for a virus to infect, in turn it also defines the number of trimers on a virion that need to be blocked by neutralizing antibodies. Depending on the stoichiometry of entry the quantities of antibody needed for effective neutralization can therefore vary substantially [56] (S10 Fig.).
To resolve molecular requirements of HIV membrane fusion, we explored in the present study the stoichiometry of HIV-1 entry (T), which defines the number of envelope trimers required per virion to fuse with the target cell membrane and thereby initiate infection [13]–[15]. Our estimates of T are based on a combined strategy of experimental data acquisition and mathematical modelling. We analyzed envelopes from 11 HIV-1 strains including different HIV subtypes, CCR5 and CXCR4 users, and envelopes with open (lab adapted strains) and closed (primary isolates) trimer conformation. We found that T differs substantially between individual strains with measurements ranging from 1 to 7 trimers that are required for entry. While a previous study suggested that HIV −1 isolates generally require only a single trimer for entry [13] our analysis retrieved values for T which were, with one exception (strain REJO), greater than 1, supporting the findings of alternate modelling approaches by us and others [14], [15]. Of note, only one of the two dominant-negative env mutants we probed recorded T = 1 for the strain REJO while the other mutant yielded an estimate of T = 2. Thus, while our data cannot exclude that T = 1 for some strains, based on our observations a range of different entry stoichiometries as we describe here seems more plausible.
We postulate two potential underlying causes for the variations in T we observe across strains. Our estimates are based on virion trimer content measurements by ELISA and are therefore a composite of functional and non-functional trimers present on virions. Considering this, a high estimate of T may be derived as the consequence of premature trimer inactivation through rapid trimer decay [38], [57], or a spontaneous adoption of the CD4-bound conformation [57], [58]. In both cases high estimates of T would reflect a decreased proportion of functional trimers on virions. Alternatively, a high T may be required by envelopes which have adopted a trimer conformation with a low energetic state as described for the open conformation of lab adapted strains [58],[59]. There, the lack in energy released upon trimer conformational rearrangements may be compensated by higher trimer numbers roped into the entry process. Most notably, for all three wildtype strains that yielded high estimates of T as well as the V1V2-deleted env variants, the high T was associated with a low infectivity (Fig. 2B, 3A and 3D). The latter is particularly intriguing as it supports the possibility that env deficiencies in entry can be partially overcome by higher numbers of trimers engaged during the entry process. Interesting insights also stem from the SF162/P3N env pair: P3N was isolated from a rhesus macaque after successive rapid transfer following initial challenge with SHIV-SF162 [60]. While SF162 has a high estimate of T of 4 to 5, P3N has a T of 2 and is the most infectious env in our panel (Fig. 2B). Thus, HIV (or SIV) has the potential to evolve from a less fit to a highly transmissible env in vivo. The exact mutations responsible for the different phenotypes remain to be determined; as we previously showed, the V1V2 domains of SF162 and P3N appear to play an important role in this regard [38].
A high T was also linked with slower virus entry kinetics (Fig. 4C) suggesting that engagement of multiple trimer-receptor pairings requires prolonged time periods. A similar relationship between kinetics of membrane fusion and the number of involved fusion proteins has been previously demonstrated for SNARE (Soluble NSF Attachment protein Receptor) complex mediated membrane fusion [61], [62] and Influenza virus membrane fusion [16], [17]. The interplay between HIV-1 entry stoichiometry and entry kinetics our study reveals thus underscores the general finding that the kinetics of membrane fusion processes are governed, at least in part, by the number of participating fusion proteins.
We rate the tight association of T with functional properties of the envelopes, namely entry capacity and entry kinetics, as a strong indicator of the validity of our analysis. Nevertheless, such estimates of T can only be an approximation as certain parameters which influence the estimates cannot be determined experimentally and assumptions need to be made for the mathematical analysis. In the literature different approaches towards modelling of mixed trimer experiments have been described and led to partially deviating results, highlighting the importance of validating the models and parameters used [13], [14], [63]. Virion trimer numbers are commonly estimated from gp120 and p24 ELISA data (Table 1) [23], [24]. These analyses yield values for the average envelope content of virions but do not provide information on the frequency distribution of viruses carrying different trimer numbers across a virion population. The latter is a factor that impacts on the interpretation of mixed trimer experiments [14], [15], [64]. Additionally, env content estimates by ELISA do not deliver information on env functionality, hence functional and non-functional trimers will be accounted for [65]. A further potential limitation stems from the nature of the mixed trimer experiments which require that all combinations of envelopes probed lead to a random trimer formation. Since preferential formation of homotrimers could strongly influence results obtained from mixed trimer experiments, we controlled for equal expression levels of the co-expressed env variants. In addition, previous studies from us and others indicate that related env variants indeed form randomly mixed trimers [13], [66], [67]. As outlined in our previous work, we have incorporated in our mathematical model several functions to capture these parameters involved in HIV entry and carefully verified the validity of our approach in the current study both in vitro and in silico (S1–S2 Fig., S1 Table and [15], [63], [68]). Nevertheless, this does not exclude that additional parameters beyond T contribute to the variation in entry phenotype between individual HIV-1 strains. For instance, differences in trimer stability or affinities for CD4 and co-receptors could significantly impact on virus entry efficacy without direct influence on T.
Membrane fusion via the stalk-pore mechanism [11] is a multi-step process that ultimately depends on energy provided by fusion proteins [9], [11], [52], [69]. Approximation of two membranes is followed by fusion of the two proximal membrane leaflets, forming the hemifusion stalk intermediate. Subsequent fusion of the distal membrane leaflets creates a fusion pore, which may either expand or close again depending on the forces exerted on the membranes. Viral envelope glycoproteins, such as the HIV-1 env or Influenza virus HA trimer, are metastable structures that undergo a series of conformational changes following receptor engagement which releases energy utilized in the fusion process [7], [8]. Although the energy required for hemifusion stalk formation could potentially be recovered from refolding of a single envelope glycoprotein trimer [8], [55], subsequent formation of the fusion pore and pore enlargement are thought to require higher energy levels [9], [11], [70]. Growing evidence suggest that only concerted action of several trimers leads to membrane fusion and maintenance of a fusion pore large enough to allow passage of the HIV capsid [1], [9], [10], [20]. Our estimates that, for the majority of HIV-1 primary isolates, 2 to 3 env trimers are required to mediate infection are thus in accordance with these studies on the mechanisms of membrane fusion. Interestingly, our estimates of the HIV entry stoichiometry resemble those made for Influenza A virus [16]–[18], postulated to require 3 to 4 HA trimers for entry, and vesicular membrane fusion via SNARE complexes, which generate energy during refolding at levels comparable to viral fusion proteins [71]–[73]. In high similarity to HIV and Influenza, also 1 to 3 SNARE complexes appear to be required to induce membrane fusion [61], [62], [74]. To further explore this relationship, we compared reported values of energy required for membrane fusion and energies released by fusion proteins (S11 Fig.). Membrane fusion requires an energy input of 40 to 120 kbT [9], [55], [75], [76]. Refolding of HA trimers into the six-helix bundle conformation was estimated to release 30 to 60 kbT [8], [9], [77] while SNARE complex assembly into 4-helix bundles releases an estimated 19 to 65 kbT [71]–[73], [78]. Considering that 3 to 4 HAs and 1 to 3 SNARE complexes were estimated to participate in entry, this yields total energies of 20 to 240 kbT released during the respective membrane fusion processes. In analogy, assuming a total energy of 40 to 120 kbT required for membrane fusion, our estimates that 2 to 7 trimers are required for HIV entry indicate that each trimer releases between 6 to 60 kbT during the entry process. Of note, the calculated total energies released by both HA trimers and SNARE complexes appear to be higher than the reported energy required for membrane fusion. This could potentially be due to inefficient coupling of the energy generated through protein conformational rearrangements into membrane deformation, or divergence between artificial membranes employed in biophysical experiments to measure membrane fusion energies and naturally occurring membranes containing proteins and having varying lipid composition. In sum, the strong agreement in the estimated energies across biological systems is intriguing and suggests that the overall energy requirements of membrane fusion and principles of energy elicitation by fusion proteins are closely related.
Our estimates that HIV strains typically require 2 to 7 trimers for entry are also consistent with the observations that env trimers cluster on virions and that HIV establishes a contact zone where several env-receptor pairings between virus and target cell are formed, the so-called “entry claw” [19], [28]. When defining molecular requirements for HIV entry, it is important to consider that envelope trimer activities may reach beyond solely providing the energy for membrane fusion. For example, binding of HIV trimers to their target cell receptors triggers an array of intracellular signals, which amongst other processes is thought to govern intracellular actin re-arrangements [79] and may be required to support membrane fusion and pore enlargement as previously proposed [10], [11].
In summary, our estimates of the HIV entry stoichiometry are in strong accordance with requirements found in other membrane fusion processes. Importantly, we show that capacities of individual HIV envelopes in mediating entry can vary substantially which is likely due to differences between trimers in soliciting energy required for membrane fusion. Our data strongly suggest that viruses overcome these envelope limitations by increasing the number of envelope-receptor pairings involved in the entry process. Hence, the stoichiometry of HIV entry is an important parameter steering virion infectivity and its assessment provides a relevant contribution towards a refined understanding of HIV-1 entry and pathogenesis. Knowledge obtained from the quantitative assessment of trimer-receptor interactions during HIV entry is a prerequisite in unravelling the stoichiometry of trimer interactions with target cell receptors or virus inactivation by neutralizing antibodies and entry inhibitors and thus may aid future approaches in HIV vaccine or entry inhibitor design.
293-T cells were obtained from the American Type Culture Collection (ATCC) and TZM-bl cells [80] from the NIH AIDS Research and Reference Reagent Program (NIH ARP). Both cell types were cultivated in DMEM (Gibco) containing 10% heat inactivated FCS and penicillin/streptomycin. Plasmids encoding the envelopes of strains JR-FL, SF162, NL4-3, RHPA, AC10, REJO, BG505 and ZM214 were obtained from the NIH ARP. Envelope ZA110 was described previously [67]. Envelope clone P3N [60] was a gift from Dr. Cecilia Cheng-Mayer, Aaron Diamond AIDS Research Center, New York, USA. Envelope clone CAP88 [81] was a gift from Dr. Lynn Morris, National Institute for Communicable Diseases, Johannesburg, South Africa. All envelope point mutations were generated by site-directed mutagenesis (Agilent QuikChange II XL) according to the manufacturer instructions. All point mutant envelopes were sequenced by in-house Sanger sequencing to confirm presence of the desired mutations and absence of unintended sequence changes. V1V2-deleted envelopes were previously described [67]. The Luciferase reporter HIV pseudotyping vector pNLLuc-AM was previously described [67]. T-20 [82] was purchased from Roche Pharmaceuticals.
To estimate the stoichiometry of entry we employed a previously described approach [13]. To produce HIV-1 pseudotype virus stocks expressing mixed trimers with varying ratios of functional to dominant-negative env, 293-T cells in 12 well plates (100.000 cells per well in 1 ml complete DMEM, seeded 24 h pre-transfection) were transfected with 1.5 µg pNLLuc-AM and 0.5 µg env expression plasmids, using polyethyleneimine (PEI) as transfection reagent. The ratio of functional to dominant negative env expression plasmids was varied to yield combinations with 100, 90, 70, 50, 30, 10 and 0% of functional env. Total env plasmid content was always at 0.5 µg. After overnight incubation the transfection medium was replaced with 1 ml fresh complete DMEM and virus-containing supernatants were harvested 48 h post transfection. All mixed trimer combinations of an individual virus strain were always generated in parallel, excluding influences of producer cells and transfection procedure. To determine virus infectivity, serial dilutions of virus stocks starting with 100 µl of undiluted virus supernatant were added to TZM-bl reporter cells in 96-well plates (10.000 cells per well) in DMEM medium supplemented with 10 µg/ml DEAE-Dextran. TZM-bl infection was quantified 48 h post-infection by measuring activity of the firefly luciferase reporter. For each functional to dominant negative env ratio series, the infectivity of the stock containing 100% functional (wt) env was taken as reference (100% infectivity) and the relative infectivity of the stocks with increasing percentages of dominant-negative env were calculated in relation to that infectivity value. The resulting data of relative infectivity were plotted over the fraction of dominant-negative env of each stock and the data were analyzed with mathematical models [15] as described below.
To produce HIV-1 pseudotype virus stocks for comparisons of virus infectivity and determination of virus entry kinetics, 293-T cells in T75 flasks (2.250.000 cells in 15 ml complete DMEM, seeded 24 h pre-transfection) were transfected with env and pNLLuc-AM plasmids (7.5 µg and 22.5 µg, respectively) using PEI as transfection reagent. The medium was exchanged 12 h post-transfection and virus-containing supernatants were harvested 48 h post-transfection. The supernatants were cleared by low speed centrifugation (300 g, 3 minutes), aliquoted and stored at −80°C. Virus infectivity was quantified on TZM-bl reporter cells as described above.
To determine gp120 and p24 content of HIV-1 pseudotype virus preparations, 293-T cells in T25 flasks (750.000 cells in 5 ml complete DMEM, seeded 24 h pre-transfection) were transfected with env and pNLLuc-AM plasmids (2.5 µg and 7.5 µg, respectively) using PEI as transfection reagent. The medium was exchanged 12 h post-transfection and virus-containing supernatants were harvested 48 h post-transfection. The supernatants were cleared by low speed centrifugation (300 g, 3 minutes), then ultracentrifugated (SW28 rotor, 2 h, 28.000 rpm, 4°C), the supernatant removed and viral pellets resuspended in 0.3 ml cold PBS and stored at −80°C. Virion associated p24 and gp120 antigens were quantified by ELISA as previously described [67]. Briefly, virus preparations were dissolved in 1% Empigen (Fluka Analytical, Buchs, Switzerland) and dilutions of each sample probed for gp120 and p24. Gp120 was captured on anti-gp120 D7324 (Aalto Bioreagents, Dublin, Ireland) coated immunosorbent plates and detected with biotinylated CD4-IgG2 and Streptavidin-coupled Alkaline Phosphatase (GE Healthcare, Chalfont St Giles, UK). P24 was captured on anti-p24 D7320 (Aalto Bioreagents, Dublin, Ireland) coated plated and detected using Alkaline Phosphatase-coupled antibody BC1071 (Aalto Bioreagents, Dublin, Ireland). Concentrations of gp120 and p24 in the samples were calculated in relation to standard curves obtained with gp120 and p24 proteins of known concentration. To derive virion numbers from p24 concentrations, we assumed 2000 (for the data in main manuscript and figures) or alternatively 2400 (S3B Fig.) p24 molecules per virion [23], [24], [33]. Trimers per virion were then calculated from the obtained gp120 concentration in relation to the number of virions per sample.
To determine virus entry kinetics, TZM-bl cells were seeded in 96-well plates (20.000 cells per well) in complete DMEM supplemented with 10 µg/ml DEAE-Dextran. 24 h post-seeding, cells were first cooled at 4°C for 5 minutes, then the medium was removed. HIV-1 pseudotype virus stocks adjusted to 50.000 RLU in 100 µl DMEM at 4°C were added per well and plates centrifuged for 70 minutes at 1200 g and 10°C. The low temperature was chosen to allow virus attachment during spinoculation but no entry. Following spinoculation the supernatant with unbound virus was removed and 130 µl of DMEM, pre-warmed to 37°C, were added per well to initiate infection (timepoint zero) and plates were incubated at 37°C. At defined timepoints post-infection, 20 µl of T-20 (375 µg/ml in DMEM; yielding a final assay concentration of 50 µg/ml) were added per well to stop the viral entry process. To obtain a measure for infectivity across different experiments, the wells with the last T-20 addition at 120 min after infection start were used as 100% reference infectivity value and the infectivity of all other T-20 treated wells were set in relation to it. In addition, a mock-treated well (addition of 20 µl DMEM at timepoint zero) was evaluated to assess absolute infectivity in absence of T-20.
Isolation and infection of PBMCs was performed as previously described [85]. Briefly, PBMCs were isolated from pooled buffy coat of 3 healthy blood donors, stimulated with anti-CD3 and PHA in presence of 100 U IL-2 for 2 days and then seeded in 96-well plates at a density of 100.000 cells per well in 100 µl RPMI medium supplemented with 10% FCS, antibiotics, 100 U IL-2 and 2.5 µg/ml (final assay concentration) polybrene. Serial dilutions of mixed trimer virus stocks (100 µl per well) were added to the PBMCs and cells were incubated at 37°C for 72 h before determining luciferase reporter activity.
Correlation analyses according to Pearson and multiple unpaired t-tests to derive the p-values for the comparisons of wt and V1V2-deleted envs were performed in GraphPad Prism 6.
|
10.1371/journal.pntd.0001148 | Clonal Differences between Non-Typhoidal Salmonella (NTS) Recovered from Children and Animals Living in Close Contact in The Gambia | Non-Typhoidal Salmonella (NTS) is an important cause of invasive bacterial disease and associated with mortality in Africa. However, little is known about the environmental reservoirs and predominant modes of transmission. Our study aimed to study the role of domestic animals in the transmission of NTS to humans in rural area of The Gambia.
Human NTS isolates were obtained through an active population-based case-control surveillance study designated to determine the aetiology and epidemiology of enteric infections covering 27,567 Gambian children less than five years of age in the surveillance area. Fourteen children infected with NTS were traced back to their family compounds and anal swabs collected from 210 domestic animals present in their households. Identified NTSs were serotyped and genotyped by multi-locus sequencing typing.
NTS was identified from 21/210 animal sources in the households of the 14 infected children. Chickens carried NTS more frequently than sheep and goats; 66.6%, 28.6% and 4.8% respectively. The most common NTS serovars were S. Colindale in humans (21.42%) and S. Poona in animals (14.28%). MLST on the 35 NTS revealed four new alleles and 24 sequence types (ST) of which 18 (75%) STs were novel. There was no overlap in serovars or genotypes of NTS recovered from humans or animal sources in the same household.
Our results do not support the hypothesis that humans and animals in close contact in the same household carry genotypically similar Salmonella serovars. These findings form an important baseline for future studies of transmission of NTS in humans and animals in Africa.
| Salmonellosis is a neglected tropical disease causing serious dysentery and septicaemia particularly in young infants, elderly and immunocompromised individuals such as HIV patients and associated with substantial mortality in developing countries. Salmonellosis also constitutes a major public health problem as it is considered the most widespread bacterial zoonosis of food origin throughout the world. Many epidemiological data exist from developed countries concerning transmission of Non-Typhoidal Salmonella (NTS) but few are available from developing countries. In addition few studies in sub-Saharan Africa have considered the interface between humans and their environment in relation to animals present in the household and food hygiene. This study describes the prevalence of NTS among fourteen Gambian children and 210 domestic animals living in close proximity (household) to the children in a rural setting in The Gambia. We found that the domestic animals living in the same household as patients carried different NTS serovar and genotypes; indicating that zoonotic transmission does not occur in our setting. This study provides baseline data for future studies of transmission of NTS in rural Africa.
| Non-Typhoidal Salmonella (NTS) are important causes of invasive bacterial diseases and are associated with substantial mortality. In rural Gambia, NTS was the second most important blood culture isolate after Streptococcus pneumoniae in children with invasive bacterial disease [1]. Similarly, a study in an urban hospital of The Gambia revealed that NTS represented 8.6% of the bacteraemia cases [2]. NTS also cause serious dysentery and septicaemia particularly in young infants [3], [4]. However, little is known about environmental reservoirs and predominant modes of transmission especially in the African context [5], [6]. Various sources including farm animals, pets and reptiles have been potentially implicated in the transmission of NTS between animals and humans but their direct involvement in the transmission has never been demonstrated [7]. The asymptomatic Salmonella carrier state in poultry has serious consequences on food safety and public health due to the risks of food poisoning following consumption of contaminated products. Salmonella enterica serovar Enteritidis can persist in the caecum or ovaries of chickens without triggering any clinical signs. Salmonellosis in young chickens may cause high mortality as a result of severe diarrhoea and dehydration, and include a greater risk of evolving into a carrier state in the surviving animals [8], [9], [10]. Small ruminants, such as sheep and goats, are also potential carriers of Salmonella [11], [12]. In The Gambia, like in most of the African countries, the majority of the population lives in rural areas and depends on agriculture and livestock. In these areas, the most common animals kept in the compound are chickens, sheep and goats. Therefore, farmers and their families often live in close contact with their livestock and even under the same roof and are thus at increased risk of contracting zoonotic infections, particularly children who often play on the ground. Domestic animals may thus play an important role in the maintenance and transmission of NTS to humans at the community level. Investigating the relatedness between human and animal strains of NTS could provide useful information about the epidemiology of these pathogens. Our study seeks to assess the contribution of domestic animals to the transmission of NTS to humans and to study the genetic relatedness between human and animal strains. This study will also provide useful information about the prevalence of NTS in domestic animals in rural areas of The Gambia.
The study protocol and consent form were approved by the Ethics Committees of the Joint Gambian Government/MRC Ethics Committee and the Ethics Committees of the London School of Hygiene and Tropical Medicine, UK and by the Institutional Review Board of the University of Maryland, Baltimore. For any patient eligible for the study written informed consent was obtained prior to their enrollment after the objectives and risks and benefits associated with participation. 95% of eligible subjects agreed to participate. If the participant was illiterate, a witness who was present throughout the consent procedure completed the necessary portions and signed the consent form; the parent/participant marked the consent form (either fingerprint or other notation) cases. If the person was literate, then he/she read and signed the consent form.
All animals used in this study were handled by professional veterinary staff in strict accordance with good animal practice as defined by the Gambian Government/ITC code of practice for the care and use of animals for scientific purposes. All animal work were conducted with ethical approval from both the Joint Gambian Government/ MRC Ethics Committee and the Ethics Committees of the London School of Hygiene and Tropical Medicine, UK.
Animal owners gave their written informed consent to examine and take rectal swabs from their animals and gave permission to publish questionnaire results from this study.
Human NTS was obtained through active population-based case-control surveillance between December 2007 and February 2009 which was designed to determine the aetiology and epidemiology of enteric infections in Gambian children less than five years of age as part of the Gates Enteric Multicentre Study (GEMS). The entire surveillance area (figure 1) including all the compounds was mapped under GPS coordinates. This surveillance area represented a total population of 152,393 of which 27,567 are less than five years of age. Children under five years of age who presented with severe diarrhoea (i.e., diarrhoea with dehydration, dysentery, or requiring hospitalization) within 3 days of onset of diarrhea were eligible to participate. For each enrolled child with diarrhoea, one healthy control child without diarrhoea was randomly selected from the community in which the case resided, matched to the case by age, gender, and time of presentation. After providing informed consent from the parent/guardian of each case or control a single, fresh, whole stool specimen was collected from cases and controls and cultured to detect bacteria species (Aeromonas spp., Campylobacter spp, Salmonella Typhi, NTS, Shigella spp, Vibrio spp, diarrheagenic E. coli strains), viral (Rotavirus, Adenovirus, Astrovirus, Norovirus, Sapovirus,) and protozoa (Cryptosporidium spp Entamoeba histolytica, Giardia lamblia).
During the surveillance period, 495 diarrhoea cases were identified and NTS were isolated from eight patients and six from community healthy controls. The 14 children were enrolled in this study. and traced back to their family compounds and five apparently healthy animals per species (chicken, sheep and goat) residing in the same household as the child were randomly enrolled within a week of isolating NTS from humans.
Anal swabs were collected from 210 household contact animals and 21 NTS strains were isolated from the faeces.
Stool specimens were transported in buffered glycerol saline (BGS) to the laboratory and processed within 6 hours of collection. Stools were plated on Xylose Lactose Desoxycholate (XLD) and MacConkey (MAC) agar and incubated at 36°C for 24 hours. Suspected non lactose fermenter colonies were subjected to biochemical reactions using Analytical Profile Index 20 Enteric (API 20E) according to manufactures' instructions (BioMerieux SA, REF 20 100/20 160). Serotyping was done by slide agglutination using Salmonella polyvalent and monovalent O and H antisera (Diagnostic Pasteur, Paris, France) according to the Kauffmann-White classification scheme [13].
Antimicrobial susceptibility tests were performed on Muller-Hinton agar (Oxoid, USA) using the agar diffusion method with the Bio-Rad discs (Marne-La-Coquette, France) according to the guidelines of the Antibiogram Committee of the French Society for Microbiology (CA-SFM) [14]. Strains were tested with 22 antimicrobial disks (Bio-Rad): amoxicillin (25 mg), amoxicillin (20 mg) plus clavulanic acid (10 mg), ticarcillin (75 mg), cephalotin (30 mg), cefoxitin (30 mg), cefotaxime (30 mg), ceftazidime (30 mg), tobramycin (10 mg), amikacin (30 mg), nalidixic acid (30 mg), pefloxacin (5 mg), norfloxacin (10 mg), trimethoprim (1.2 mg) plus sulfamethoxazole (23.75 mg), tetracycline (30 mg), chloramphenicol (30 mg), gentamicin (10 mg), trimethoprim (300 mg), ciprofloxacin (5 mg), spectinomycin (100 mg), streptomycin (10 mg), sulfonamides (200 mg), and nitrofurantoin (300 mg). Diameters of the inhibition zones were measured with OSIRIS version 3.6xE2.0 (Bio-Rad), and results were interpreted as susceptible, intermediate, or resistant according to the recommendations of the CA-SFM [14]. All these tests were carried out at the MRC microbiology laboratory which is enrolled in the external quality assurance programme of the United Kingdom National External Quality Assessment Scheme [15].
MLST was performed on the 35 Salmonella isolates as previously described [3]. The seven genes targeted were aroC, dnaN, hemD, hisD, purE, and thrA. Amplification of all genes was carried out in a 25 µl reaction mixture of the following items: 10xBuffer with 1.5 mm MgCl2 (2.5 µl); 2 mM dNTP'S (0.5 µl); 12.5 mM forward primer (1 µl); 12.5 mM reverse primer (1 µl); 5 U/µl Qiagen Hotstart Taq Polymerase (0.25 µl); Template (cell lysate) (2 µl) and 17.75 µl sterile DNA free water. PCR cycling conditions were as follows: 10 min at 94°C, followed by 32 cycles of 94°C for 1 min, 55°C for 1 min and 72°C for 1 min, and a final extension at 72°C for 5 min. 2 µl aliquots of PCR products were separated on 1% agarose gel electrophoresis, and visualized with ethidium bromide staining and UV illumination, and using a gel documentation system. PCR products were purified using Qiagen kit (Qiagen). Sequencing was done on both strands with BigDye Terminator Cycle Sequencing kit (Applied Biosystems, UK). The labelled fragments were separated by size using 3130xl Genetic Analyser (Applied Biosystems, UK). Sequences were edited and complementary sense antisense fragments were aligned using the Laser Gene DNA star 7.1 software. Finally, the sequences were submitted to the MLST database website [16] and assigned to existing or novel allele or sequence type numbers defined by the database.
Tests of association were done using Fisher's exact test in Stata 11 (StataCorp. 2009. Stata Statistical Software: Release 11. College Station, TX: StataCorp LP). Stata provides one-sided p-values only for Fisher's exact test unless the table is 2×2 and results with p-values of less than 0.05 for the one-sided test were considered statistically significant. The parameters were grouped for the purpose of the analysis. The secondary diagnosis was categorized into three groups for the Fisher's exact test: group 1, children co-diagnosed with another disease other than malaria; group 2, children co-diagnosed with malaria and group 3, children who were not co-diagnosed with any other disease. The age was also categorized into 2 groups: less than or equal to 18 months (the average age), and more than 18 months. The mapping of the case locations was done using Arc Gis 9.3 software. To perform the cluster analysis of the serovars, MLST data were analysed with Bionumerics software (version 4.0; Applied Maths, Sint-Martens-Latem, Belgium). Analysis using a hierarchic unweighed pair group method (UPGMA) with averaging was used to generate a dendrogram describing the relationship among Salmonella serovars (figure 2).
A diversity of serovars was found in both the human and animal population. None of the compounds showed similar serovars in both humans and animals (table 1). Nevertheless, one serovar namely S. Moualine was simultaneously found in a diarrheic child and in a chicken but in different compounds: C1 in Banico Allunhare and C8 in Koina (table 1, figure 1). MLST revealed a single locus variant at sucA between those 2 strains (table 1). In compounds 9 (C09), 10 (C10), 11 (C11), 12 (C12), 13 (C13) and 14 (C14), no Salmonella was isolated from animals (Table 1). The most prevalent serovars were S. Colindale in the human population (21.42%) and S. Poona in the animal population (14.28%).
The proportion of Salmonella isolated was higher in the chicken population than in other species: 66.6%, 28.6% and 4.8% in chickens, sheep and goats, respectively. In the same compound, a diversity of Salmonella serovars were circulating at animal level especially in chickens; the number varying from 2 to 3 different serovars: in compound 1 (C1), S. Moualine, S. Tornow and S. Poona; in compound 6 (C6), S. Schwarzengrund and S. Stanleyville and in compound 7 (C7), S. Offa, S. Give and S. Poona (table 1).
The mean age of children enrolled in the study was (to the nearest integer) 18 months and the age varied between 9 and 26 months (table 2). All children who presented with diarrhoea except one were secondarily diagnosed with another disease such as malaria (4 children) or other disease symptoms including fever, or cough (3 children) (table 2). There was a significant association (p-value<0.01) between expressing clinical signs of salmonellosis, i.e. diarrhoea and being co-diagnosed with a secondary disease (table 3). Age was not associated with the expression of clinical signs of salmonellosis (p-value = 0.16).
All serovars were fully susceptible to all antibiotics tested except one of each the following serovars: Salmonella enteritica serovar Poona, Salmonella enteritica serovar Johannesburg, Salmonella enteritica serovar Chile and Salmonella enteritica serovar Colindale which were resistant to streptomycin.
Four new alleles were discovered: hemD (152), hemD (153), hisD (256) and purE (226) and eighteen novel sequence types (ST) (table 1). Similar serovars exhibited the same allelic profiles, except S. Moualine which had two different allelic profiles for the serovars isolated from humans and animals (table 1).
The seven housekeeping genes were concatenated for all isolates and the UPGMA tree was constructed (figure 2). All Salmonella genotypes had at least 80% similarity and the majority varied between 99% and 100%. Salmonella genotypes causing diarrhoea in children were always clustered with animal genotypes. Both Salmonella enteritica serovar Moualine isolated from a chicken and a child was clustered (figure 2).
The serovar diversity of NTS within the human and animal population was high showing that Salmonella is carried by both humans and animals in the community. The serovars were also widely geographically distributed in our surveillance area located in a typical African rural setting in The Gambia. We showed that several Salmonella serovars were circulating in the chicken population within the same compound or household; such as compounds 1 (C1), 6 (C6) and 7 (C7) respectively in the following 3 villages: Baniko Allunhare, Bagadagy and Misra Ba Mariama (figure 1). The higher proportion of Salmonella serovars in the chicken population compared to the goats and sheep is not surprising as it is known that chicken is the most important reservoir of NTS and thus thought to be the major source of transmission to humans [17]. Salmonella can persist in the chicken cecum or ovaries without triggering clinical signs in the host. Salmonellosis in young chickens may cause high mortality as a result of severe diarrhoea and dehydration, and creates a greater risk of evolving into a carrier state in the animals which survive [8], [9], [10]. The asymptomatic Salmonella carrier state in poultry has serious consequences for food safety and public health due to the risks of food poisoning following consumption of contaminated products. Small ruminants, such as sheep and goats, are also potential carriers of Salmonella [11], [12]. In Ethiopia, a study indicated that Salmonella is common in apparently healthy slaughtered sheep and goats. It also showed the presence of a wide range of Salmonella serovars in sheep and goats, which are of veterinary and public health significance [18]. The high rate of NTS clones circulating in the same compound could lead to mixed infection or carriage within the chicken population. This situation could result in extensive genetic diversity and variability due to frequent intraspecific recombination as it occurs with Helicobacter pylori [19]. This could have as consequence a wider range of clones and thus more difficulties to control NTS infections at animal level.
The diversity of serovars that we observed in this study is different from what we previously reported [3] where Salmonella enterica serovar Enteritidis was the most common serovar (80.6%) followed by Salmonella enterica serovar Typhimurium (8.0%) among NTS isolated from children with pneumonia and/or septicemia patients. It appears that the epidemiology of NTS is changing in The Gambia or the serovars detected might be site or disease specific, i.e. gastroenteritis vs. systemic infections.
MLST provides the best phylogenetic-relationship inference for the Salmonella genus [20]. Therefore, it may be invaluable for determination of the relationship among various Salmonella strains and serovars [21]. As expected, the similarity matrix (figure 2) of the serovars revealed close genetic relationship (>80% and the majority between 99 to 100%) between human and animal serovars, showing the genetic homogeneity of Salmonella. A study done in Senegal has also revealed a high degree of similarity among Salmonella enteritica serovar Brancaster and Salmonella enteritica serovar Enteritidis serovars from poultry and from humans by the use of PFGE techniques [22], but direct evidence of Salmonella transmission from poultry to humans could not be provided.
The high degree of similarity between human and animal serovars supports the theory that Salmonella clones are stable [23]. The genetic tree has also revealed that all lineages contained isolates of mixed origin (human and animal). From the present data, there is therefore no indication of clonal groups or lineages that are adapted to any specific host. These findings support the conclusions of other authors who used the same techniques (MLST) with Salmonella from human and veterinary sources in Denmark [24] or a different technique like Pulsed Field Gel Electrophoresis (PFGE) with isolates from humans and animals, food or the environment in close contact with humans, which was the case in Kenya [25]. Like in the Kenya study [25], we also showed that there was no relatedness between NTS genotypes from humans and those from animals in close contact to humans, this other potential sources of transmission such as environmental or the human-to-human transmission need to be examined.
A statistical significant association (P<0.05) was observed between children expressing clinical signs of salmonellosis (diarrhoea) and co-diagnosis with malaria (table 3). The association between salmonellosis and immunocompromising diseases is well known in the African context. NTS infections are usually associated with opportunistic infections especially in immuno-compromised patients, e.g. HIV-infected adults [26], [27] or with other diseases like malaria or anemia [1], [5], [28], [29]. Children especially those less than 3 years old are the age-group at risk of expressing clinical signs of salmonellosis [1], [26]. The contradiction between our observation and those of most other authors is certainly due to the small sample size and the small range of age groups in our study. All individuals were in fact less than three years old. Malaria has long been suspected to increase the risk of invasive NTS infection and might contribute to the seasonality of NTS disease, although the mechanism underlying the association between malaria and NTS is only partially understood [6].
All serovars were susceptible to most commonly used antibiotics for the treatment of clinical infections in The Gambia such as amoxicillin, amoxicillin plus clavulanic acid, trimethoprim plus sulfamethoxazole, tetracycline, streptomycin and chloramphenicol and also to cephalosporins of the third generation which are considered as the drugs of choice for invasive Salmonella infections in humans. This result is in contrast to previous studies done in urban [1], [2] and rural [3] areas in The Gambia where Salmonella strains expressed multi-resistance to several commonly used antibiotics. This could be explained in part by the fact that the NTS isolated from those studies were from invasive cases (pneumonia and sepsis) whereas this study focused on non-invasive cases (diarrhea). In addition, in those studies Salmonella enterica serovar Enteritidis was the most common serovar followed by Salmonella enterica serovar Typhimurium. While these serovars were not detected in this study as a result of temporal trends of childhood NTS infection in The Gambia {Mackenzie, 2010 #33}. There is lack of resistance of serovars to antibiotics in rural areas even to those commonly used in the hospitals to treat bacterial infections. This is due to the fact that in rural areas, NTS infections are not treated because patients are often not conducted to the hospital due to poor access to medical centers, lack of transport facilities and of financial resources. Our findings suggest that these drugs remain suitable for the treatment of salmonellosis in humans and animals. However, we have to interpret these results with caution because the sample size in our study was small.
Our study showed that the use of serotyping data combined with MLST and phylogenetic analysis can provide important information about the epidemiology of NTS in humans and animals. However, our results do not support the hypothesis that humans and animals in close contact in the same household carry genotypically similar Salmonella serovars. Nevertheless these findings have stirred up the problem of the transmission of NTS in African context and suggest that poultry may play an important part in the epidemiology of Salmonella infections of this condition. A better control of malaria may lead to a reduction in the incidence of invasive NTS disease in The Gambia. Multidrug resistance has not yet been a problem in human and animal NTS isolates in this area of the country. Thus, commonly available drugs may still be used for the treatment of NTS infections in rural Gambia. Nevertheless, public authorities must be alert to detect any change in the behavior of Salmonella towards antibiotics, with a view of establishing appropriate control measures for use of these drugs in humans and animals.
|
10.1371/journal.pbio.0050290 | Systematic In Vivo Analysis of the Intrinsic Determinants of Amyloid β Pathogenicity | Protein aggregation into amyloid fibrils and protofibrillar aggregates is associated with a number of the most common neurodegenerative diseases. We have established, using a computational approach, that knowledge of the primary sequences of proteins is sufficient to predict their in vitro aggregation propensities. Here we demonstrate, using rational mutagenesis of the Aβ42 peptide based on such computational predictions of aggregation propensity, the existence of a strong correlation between the propensity of Aβ42 to form protofibrils and its effect on neuronal dysfunction and degeneration in a Drosophila model of Alzheimer disease. Our findings provide a quantitative description of the molecular basis for the pathogenicity of Aβ and link directly and systematically the intrinsic properties of biomolecules, predicted in silico and confirmed in vitro, to pathogenic events taking place in a living organism.
| A wide range of diseases, including diabetes and common brain diseases of old age, are characterised by the deposition of protein in the affected tissues. Alzheimer disease, the most common neurodegenerative disorder, is caused by the aggregation and deposition of a peptide called Aβ in the brain. We have previously developed a computational procedure that predicts a particular peptide or protein's speed of aggregation in the test tube. Our goal was to test whether the speed of aggregate formation that we observe in the test tube is directly linked to the brain toxicity that we see in our fruit fly model of Alzheimer disease. We made flies that produce each of 17 variant forms of Aβ and show that the toxicity of each variant is closely linked to the tendency of each variant to form small soluble aggregates. Our computational procedure has previously been shown to be applicable to a wide range of different proteins and diseases, and so this demonstration that it can predict toxicity in an animal model system has implications for many areas of disease-related research.
| A wide range of proteins has been found to convert into extracellular amyloid fibrils, or amyloid-like intracellular inclusions, under physiological conditions [1,2]. Such proteins have largely been identified through their association with disease, although a number have been found to have beneficial physiological functions in organisms including, amongst others, bacteria [3], yeast [4], and humans [5]. Indeed, the ability to aggregate and assemble into amyloid-like fibrils has emerged as a common, and perhaps fundamental, property of polypeptide chains [1,6,7]. This discovery has stimulated extensive biophysical and mutational analysis of the underlying molecular determinants of amyloid fibril formation. These studies have resulted in the derivation of general models, based on physicochemical parameters, that both rationalise and predict the propensity of proteins to convert from their soluble forms into intractable amyloid aggregates in vitro [8–10].
The misfolding and aggregation of proteins in vivo, however, differ from similar processes taking place under in vitro experimental conditions, in that they occur in complex cellular environments containing a host of factors that are known to modulate protein aggregation and protect against any subsequent toxicity [11]. This difference between in vitro and in vivo experimental conditions represents a significant barrier to the development of a molecular understanding of protein aggregation in living systems and its consequences for disease. In this paper we describe the results of an approach designed to bridge this divide by expressing a range of mutational variants of Aβ42 in a Drosophila model of Alzheimer disease [12] and correlating their influence on the longevity and behaviour of the flies with their underlying physicochemical characteristics.
The expression of the Aβ42 peptide (coupled to a secretion signal peptide) in the central nervous system of Drosophila melanogaster results in both intracellular and extracellular deposition of Aβ42, along with neuronal dysfunction, revealed by abnormal locomotor behaviour and reduced longevity [12–14]. Learning and memory deficits are also observed in flies expressing Aβ42 and to a lesser extent in those expressing Aβ40. Importantly, the severity of the cognitive deficits is closely correlated with the magnitude of the locomotor and longevity phenotypes [14]. Our system, as with other recently developed invertebrate models of neurodegenerative disease, therefore produces clear, quantitative phenotypes that allow rapid and statistically robust assessments of the effects of mutations [15,16]. Using an algorithm described previously [8,10] we computed the intrinsic aggregation propensities (Zagg) of all 798 possible single point mutations of the Aβ42 peptide and also of the more toxic E22G Aβ42 peptide. A total of 17 mutational variants, with a wide range of aggregation propensities (Table 1), were then expressed throughout the central nervous system of Drosophila melanogaster, and their effects were compared to those of wild-type (WT) and E22G Aβ42 expression. The longevity of multiple lines of flies (n = 4–6 independent lines) for each variant was compared to that of flies expressing the WT or E22G Aβ42 peptide. This pooling of data from multiple independent lines for each Aβ42 mutant studied serves as a control for the potential variation in expression levels between transgene insertion sites. In addition, the locomotor ability of a representative selection of the Aβ42-variant-expressing flies was assessed to provide a measure of the early effects of the peptides on neuronal dysfunction. Examples of the results of this analysis are shown for four of the variants studied (Figure 1).
Flies expressing the WT Aβ42 peptide have a median survival of 24 ± 1 d; flies expressing the E22G Aβ42 peptide associated with familial Alzheimer disease have a median survival of only 8 ± 1 d. In contrast, some of the peptide variants are less harmful. For example, flies expressing F20E Aβ42 have a median survival of 29 ± 1 d (Figure 1A), and flies expressing I31E/E22G Aβ42 peptide have a median survival of 27 ± 1 d (Figure 1B), representing substantial increases in longevity compared to WT and E22G Aβ42 flies. Furthermore, the longevity of these variants is comparable to that of flies expressing the Aβ40 peptide (median survival = 30 ± 1 d; Figure S1A and S1B), which has been previously shown to be non-toxic when expressed both in transgenic flies [12,13] and in transgenic mice [17]. F20E Aβ42 and I31E/E22G Aβ42 flies also have very significantly improved locomotor ability compared to WT and E22G Aβ42 flies (Figure 1C and 1D; Videos S1 and S2) and are comparable in locomotor performance to flies expressing the Aβ40 peptide (Figure S1C and S1D). We also analysed a range of Aβ42 variants that were more harmful than the WT peptide; for example, flies expressing the E11G or M35F variants of the Aβ42 peptide have significantly shorter lifespans than WT Aβ42 flies (median survival = 21 ± 1 and 15 ± 1 d, respectively; Figure 1E and 1F).
Quantitative analysis of all 17 Aβ variants studied reveals a highly statistically significant correlation between the propensity of a variant to aggregate (Zagg) and its effect on the survival of the flies (Stox) (Figure 2A; r = 0.75, p = 0.001). A significant correlation is also observed when we analyse the relationship between the predicted aggregation propensity (Zagg) of a representative selection of Aβ variants and their effects on mobility or locomotor performance (Mtox) (Figure 2B; r = 0.65, p = 0.009). We have also verified that correlations exist between the measured aggregation rates (Kagg) and both Stox and Mtox for a representative selection of the Aβ42 variants, as we would expect from our predictions (Figure 3).
Whilst our analysis reveals a significant relationship between the aggregation propensity of Aβ42 and its effects on neuronal integrity in vivo, it has also uncovered a small number of variants that do not conform to this trend, most notably the I31E/E22G Aβ42 peptide. In order to determine the significance of such divergent behaviour for the origins of Aβ42 pathogenicity, we selected one peptide whose effects on the longevity and mobility of the flies is well predicted by its Zagg (F20E) and one whose effects did not correlate with its Zagg (I31E/E22G) and performed further analysis of their aggregation in vitro and in vivo.
The F20E mutation is predicted to reduce significantly the propensity of the Aβ42 peptide to aggregate (Table 1). Indeed, when we measure the rate of aggregation using thioflavin T fluorescence we find that F20E Aβ42 does aggregate significantly more slowly in vitro than the WT Aβ42 peptide (t1/2 = 44 and 11 min, respectively; Figure 4A), in good accord with our predictions.
The in vivo aggregation of the F20E Aβ42 peptide is also significantly reduced compared to that of the WT Aβ42 peptide. Anti-Aβ42 immunohistochemistry using a C-terminal-specific antibody that binds an epitope (Aβ residues 35–42) [18] that does not include the residues being studied here, reveals progressive accumulation of Aβ42 in the brains of WT-Aβ42-expressing flies from 10 d of age, with extensive deposition being evident by day 20 (Figure 4B). In contrast to this behaviour, flies expressing F20E Aβ42 show no signs of Aβ42 deposition at day 20 (Figure 4C). Quantitative reverse transcription polymerase chain reaction (qRT-PCR) analysis of Aβ42 transcription levels was also carried out on WT Aβ42 and two independent lines of F20E Aβ42 fly brains to ensure that the reduced deposition and toxicity of the F20E Aβ42 peptide was not due to coincidentally lower transcription levels. In fact, the F20E Aβ42 transgene was transcribed at slightly higher levels than WT Aβ42 (Figure S2) in both lines tested.
That the F20E Aβ42 peptide does not form in vivo deposits, despite being able to form amyloid fibrils in vitro (albeit significantly more slowly than WT Aβ42) suggests that the F20E mutation reduces the aggregation propensity of Aβ42 sufficiently to allow cellular clearance mechanisms such as proteases (e.g., neprilysin) [13] to prevent its accumulation in vivo. We conclude, therefore, that the increased longevity and locomotor performance of F20E Aβ42 flies are indeed attributable to a measurable reduction in the aggregation propensity of this peptide in vivo, as predicted by our analysis.
In the case of the I31E/E22G Aβ42 variant there appears to be no correlation between its predicted aggregation propensity (which is very similar to that of the highly pathogenic E22G Aβ42 peptide; Table 1) and its effects on longevity and locomotor behaviour in the fly (Figure 1B and 1D). However, studies of the I31E/E22G and E22G Aβ42 peptides in vitro show that, as predicted by our algorithm, they aggregate at very similar rates (t1/2 = 7 and 4 min, respectively; Figure 4D). Furthermore, anti-Aβ42 immunohistochemistry reveals similar levels of deposition in the brains of both E22G- and I31E/E22G-Aβ42-expressing flies at 8 d of age (Figure 4E and 4F) that cannot be accounted for by variations in transcription level as measured by qRT-PCR (Figure S2). Together these observations confirm that our predictions of aggregation propensity are accurate for these peptides in vivo as well as in vitro. To determine the consequences of peptide deposition on the integrity of the brain, we looked for the presence of vacuoles, which are a well-documented sign of neurodegeneration [19]. Despite comparable levels of deposition, the vacuoles seen in the brains of E22G-Aβ42-expressing flies are entirely absent from the brains of I31E/E22G-Aβ42-expressing flies. In this case, therefore, the relationship between the presence of Aβ42 deposits and the functional and anatomical integrity of the brain does not appear to hold.
This observation is reminiscent of the finding that there are cases in which the presence of Aβ plaques in the brains of elderly humans, and indeed in transgenic mouse models of Alzheimer disease, does not correlate with cognitive ability [20,21]. It has been proposed, in explanation of this finding, that the neuronal dysfunction and degeneration historically attributed to the presence of Aβ amyloid fibrils in the brains of patients with Alzheimer disease may in fact be caused by the concomitant presence of prefibrillar aggregates [22–24]. With this in mind, the unexpected in vivo effects of variants such as the I31E/E22G Aβ42 peptide prompted us to develop a second algorithm (see Materials and Methods) by analysing a set of data for which the rates of formation of protofibrils containing β-sheet structure have been reported [8]. This algorithm is able to predict the propensity of other polypeptides to form protofibrils. Whilst there are a few Aβ42 variants (including I31E/E22G Aβ42) whose global aggregation propensities (Zagg) do not correlate well with their in vivo effects on neuronal dysfunction (Figure 2), we find that the predicted propensities of these variants to form protofibrillar aggregates (Ztox) correlate very strongly with their in vivo effects (Stox, r = 0.83, p < 0.00001; Mtox, r = 0.75, p = 0.001; Figure 5).
We propose, therefore, that the effects of all Aβ42 variants in the flies can be directly attributed to their effects on the intrinsic propensities to form deleterious protofibrillar aggregates. It is extremely interesting in this regard that a comparison, using electron microscopy, of the morphology of E22G and I31E/E22G Aβ42 aggregates formed under identical conditions reveals the presence of a significant quantity of protofibrils in the former and only well-defined fibrils in the latter (Figure S4). Furthermore, we propose that it is possible to predict accurately in silico the in vivo effects of the Aβ42 peptide from a knowledge only of the intrinsic physicochemical properties of its constituent amino acids. We believe that this approach to understanding the determinants of protein misfolding in vivo will be applicable to many other diseases as we have demonstrated previously that the physicochemical parameters that determine the aggregation propensity of Aβ also determine the aggregation behaviour of a wide range of both disease- and non-disease-related proteins [10,25].
It is also remarkable that, despite the fact that the intrinsic aggregation propensities of typical protein sequences vary by at least five orders of magnitude [25], we have been able to achieve profound alterations in the pathogenic effects of Aβ42 by increasing or decreasing its propensity to aggregate by less than 15%. This result suggests that proteins implicated in misfolding diseases are likely to be extremely close to the limit of their solubility under normal physiological conditions [26], and consequently the small alterations in their concentration, environment, or sequence, such as occur with genetic mutations [27] or with increasing age [23], are likely to be the fundamental origin of these highly debilitating and increasingly common conditions [28].
In conclusion, we have presented accurate, quantitative measurements of the relationships between the manifestations of neuronal dysfunction in a complex organism, such as locomotor deficits and reduced lifespan, and the fundamental physicochemical factors that determine the propensity of the Aβ42 peptide to aggregate into protofibrils. These results provide compelling evidence that, despite the presence within the cell of multiple regulatory mechanisms such as molecular chaperones and degradation systems [29], it is the intrinsic, sequence-dependent propensity of the Aβ42 peptide to aggregate to form protofibrillar aggregates that is the primary determinant of its pathological behaviour in living systems.
Mutant Aβ42 expression constructs were produced by site-directed mutagenesis of the WT Aβ42 sequence in the pMT vector (Invitrogen) and were subcloned into the pUAST vector. Transgenic Drosophila expressing the desired Aβ42 variants were generated according to the procedures described by Crowther et al. [12].
All survival assays were carried out as described previously [12]. Survival curves were calculated using Kaplan–Meier statistics, and differences between them analysed using the log rank method. All survival times in the text are given as median ± standard error of the median. For previously characterised control lines expressing either WT or E22G Aβ42, the survival of one representative line was measured. For each novel mutational variant of Aβ42, between four and six independent lines were analysed (n = 100 for each line) in order to control for variability in expression levels between individual lines due to the varying location of transgene insertion. The effect of a mutational variant on survival (Stox) was calculated by comparing the survival time of the flies in which it was expressed (Smut) to the survival of Aβ40-expressing flies (Smax) used as a negative control in the same experiment: Stox = (Smax − Smut)/Smax.
The locomotor ability of the flies was assessed in a 45-s negative geotaxis assay. Flies were placed in a plastic 25-ml pipette and knocked to the bottom of the pipette. The number reaching the top of the pipette (above the 25-ml line) and the number remaining at the bottom (below the 2-ml line) after 45 s was measured. The mobility index was calculated as (ntop− nbottom + ntotal)/2ntotal. Two representative lines were tested for each novel mutant Aβ42 and one line for each previously characterised control (WT Aβ42 and E22G Aβ42). Three independent groups of 15 flies each were tested three times at each time point for each line. Differences between genotypes were analysed by ANOVA. The effect of each mutational variant on locomotor performance (Mtox) was calculated by fitting the decline in mobility index over time to a straight line and then estimating the time at which each mutant line of flies had declined to a mobility index of 0.5.
Immunohistochemistry analysis was performed as described previously [12] on single representative lines for each genotype using the G2–11 anti-Aβ42 antibody (The Genetics Company). Representative lines of F20E- and I31E/E22G-Aβ42-expressing flies were chosen to have median survivals within 1 d of the combined median survival determined for each genotype.
The propensity to form amyloid aggregates (Zagg) was calculated using an approach described previously [10]. Briefly, for a given protein, Zagg is obtained by averaging the propensities that are above zero in the aggregation profile. All the propensities are normalised into a variable that has an average of zero and a standard deviation that equals one (the normalisation is made using the propensities of a set of random sequences). In a profile there can be residues with a propensity larger than one, but these peaks are usually sparse and their contribution is diluted upon averaging. Consequently, sequences with an overall Zagg score larger than one are very rare. In order to calculate the propensity for forming protofibrillar aggregates (Ztox), we developed a method based on an equation containing the same physicochemical contributions used to calculate the propensity for fibrillar aggregation, but with specific weights determined using a set of experimentally determined protofibrillar aggregation rates for the protein acylphosphatase [30]. A Web server for calculating Zagg and Ztox is available at http://rd.plos.org/10.1371_journal.pbio.0050290_01.
All peptides were dissolved in trifluoroacetic acid and sonicated for 30 s on ice. The trifluoroacetic acid was removed by lyophilization and the peptides were then dissolved in 1,1,1,3,3,3-hexafluoro-2-propanol and divided into aliquots that were dried by rotary evaporation at room temperature. The amount of peptide in the aliquots was determined by quantitative amino acid analysis.
The peptides were dissolved at a concentration of 30 μM in 50 mM NaH2PO4 (pH 7.4) and incubated at 29 °C with continuous agitation. At regular time intervals, 5 μl of the peptide solution was removed and added to 100 μl of 20 μM thioflavin T in 50 mM Gly-NaOH (pH 8.5). Fluorescence intensity was measured at 440 nm excitation and 480 nm emission using BMG FLUOstar OPTIMA. The rate of aggregation (k) was determined by fitting the plot of fluorescence intensity versus time to a single exponential function y = q + Ae(−kt) [30], and t1/2 was calculated using t1/2 = ln2/k.
Twenty flies expressing each variant of Aβ42 were collected at day 0 (i.e., on the day of eclosion) for each transgenic line to be analysed. The flies were then anaesthetised and decapitated, and the heads were collected and snap frozen in liquid N2. Total RNA was extracted from the fly heads using the Qiagen RNeasy mini kit with on-column genomic DNA digestion using DNAse 1. The concentration of total RNA purified for each line was measured using a NanoDrop spectrophotometer. One microgram of RNA was then subjected to reverse transcription using the Promega Reverse Transcription System with oligo dT primers. qRT-PCR was performed using a BioRad iCycler and Absolute QPCR SYBR Green Fluorescein Mix (ABgene). Each sample was analysed in triplicate and with both target gene (Aβ42) and control gene (RP49) primers in parallel. The primers for the Aβ42 PCR were directed to the 5′ end of the signal secretion peptide sequence and the 3′ end of the Aβ coding sequence: forward, GCATTCGTGAATTCATGGCGAGCAAAGT; reverse, TACTTCTAGATCCTCGAGTTACGCAATCAC. The RP49 primers were designed across an intron to avoid amplifying any residual genomic DNA contamination: forward, ATGACCATCCGCCCAGCATCAGG; reverse, ATCTCGCCGCAGTAAACG. Relative expression levels were calculated using the Livak method.
|
10.1371/journal.pcbi.1001104 | Context-Dependent Encoding of Fear and Extinction Memories in a Large-Scale Network Model of the Basal Amygdala | The basal nucleus of the amygdala (BA) is involved in the formation of context-dependent conditioned fear and extinction memories. To understand the underlying neural mechanisms we developed a large-scale neuron network model of the BA, composed of excitatory and inhibitory leaky-integrate-and-fire neurons. Excitatory BA neurons received conditioned stimulus (CS)-related input from the adjacent lateral nucleus (LA) and contextual input from the hippocampus or medial prefrontal cortex (mPFC). We implemented a plasticity mechanism according to which CS and contextual synapses were potentiated if CS and contextual inputs temporally coincided on the afferents of the excitatory neurons. Our simulations revealed a differential recruitment of two distinct subpopulations of BA neurons during conditioning and extinction, mimicking the activation of experimentally observed cell populations. We propose that these two subgroups encode contextual specificity of fear and extinction memories, respectively. Mutual competition between them, mediated by feedback inhibition and driven by contextual inputs, regulates the activity in the central amygdala (CEA) thereby controlling amygdala output and fear behavior. The model makes multiple testable predictions that may advance our understanding of fear and extinction memories.
| The amygdaloid complex is one of the key brain structures involved in fear-related processes. A typical way to study neural correlates of fear expression (e.g. freezing response) in the amygdala is to perform a fear conditioning paradigm, which yields a conditioned fear response. This response can be reversed by another procedure called fear extinction. Thanks to the experimental approaches to date we have some understanding about the putative roles of specific subnuclei within the amygdala in the formation of these fear and extinction memories. Here, we complement the experimental studies by providing a computational model that addresses the question of how fear and extinction memories are encoded in the amygdala, and specifically, in the basal nucleus (BA). We propose a specific neural mechanism to explain how the BA may integrate information about a salient, conditioned stimulus and the environment, thereby enabling it to switch the state of the animal from low to high fear and vice versa. We also provide possible explanations for various other behavioral findings, such as the recovery of fear after it had been extinguished (renewal). Finally, we make specific, experimentally testable predictions that need to be addressed in future work.
| In classical fear conditioning an animal learns to associate an initially neutral stimulus (the conditioned stimulus, CS) with an aversive stimulus (the unconditioned stimulus, US) after paired exposure to the CS and the US. Subsequent repeated non-reinforced presentations of the CS alone result in a decline of the conditioned response, a process called fear extinction [1]. Fear extinction is a highly context-dependent process: the conditioned fear response returns when the animal is exposed to an extinguished CS outside the extinction context [2], [3].
Studies over the last decades have identified the amygdaloid complex as a key brain structure involved in both fear conditioning and extinction [4]–[6]. In the lateral nucleus of the amygdala (LA), signals carrying information about the CS and the US converge onto the same neurons where they become associated through activity-dependent plasticity mechanisms [7]–[9]. The LA can directly or indirectly influence activity in the central nucleus (CEA) [10], the major output nucleus of the amygdala that can trigger fear responses via its projections to the hypothalamus and to the brainstem [11]. The basal nucleus of the amygdala (BA) has been suggested to play an important role in contextual fear conditioning [12], [13], cued fear conditioning [14], fear extinction [15]–[17] and context-dependent fear renewal [17].
Recently, two distinct fear and extinction specific neuronal sub-populations in the BA have been identified [17]. The balance of activity between fear and extinction neurons was correlated with states of high and low fear, respectively. Moreover, pharmacological inactivation of the BA blocked the acquisition of fear extinction and context-dependent fear renewal, suggesting that BA fear and extinction neurons may underlie the induction of behavioral changes and contribute to the formation of fear and extinction memories.
These findings raise the question of what the potential mechanisms underlying the differential activation of these two neuronal sub-populations are. Here, we used a modeling approach based on in vivo physiological data to address this specific question and to draw more general conclusions on potential neural mechanisms involved in fear and extinction memories in the BA. In vivo stimulation of identified fear and extinction neurons revealed that the two neuronal populations receive differential functional input from the hippocampus and from the medial prefrontal cortex (mPFC) [17]. This finding could reflect anatomical specificity of inputs and/or selective functional plasticity of non-specific inputs. Independently of these two possibilities, in our model, we assume that anatomically and/or functionally distinct inputs from the hippocampus or the mPFC modulate the activity of BA fear and extinction neurons in a context-specific manner. That is, sub-populations of BA neurons are innervated by hippocampus/mPFC efferents that represent the current context. In addition, all BA neurons receive inputs from US/CS responsive LA neurons during conditioning and extinction. Those sub-populations of BA neurons that receive simultaneous LA and context-specific inputs become responsive during conditioning or extinction and, thus, emulate the “fear” and “extinction” neurons reported by Herry et al. [17]. Activation of BA neurons per se, however, is not sufficient to cause or prevent a behavioral response, but the selective activation of BA neurons conveys important information about the context-CS relation to the CEA. Although we do not model here the CEA, we stipulate that context-dependent BA activity provides an instructive signal to CEA neurons. In the CEA, it is likely that conditioning [18] and possibly extinction learning-induced changes act upon this signal in order to activate or suppress a fear response. If more experimental data, sufficient to constrain the possible parameter space, become available, then our present model of the BA could be extended to study the impact of context-dependent BA activity on learning-induced changes in the CEA as well.
We test the plausibility of context-dependent activation of BA neurons in two different approaches: first, in an abstract firing rate model; second, in a more realistic spiking neuron network (SNN) model of the BA. Based on the results of our model we provide plausible explanations for several experimental observations in fear and extinction learning and make specific, experimentally testable predictions.
The description of the evolution of the firing rates of BA neurons during fear conditioning and extinction reported by [17] provide certain simple, yet important, indications on the underlying dynamics in the BA network:
To test the feasibility of the above observations and their inferences in explaining the emergence of fear and extinction neurons in BA, we first studied the dynamics of a mean-field (or firing rate) model of the BA. Subsequently, we constructed a spiking neuron network (SNN) model to examine our hypotheses and their implications under more realistic conditions.
The mean-field model of BA consisted of two neuron populations, A and B, described by Wilson-Cowan type rate dynamics [19] (Fig. 2A). Both populations were identical in their properties (Eqs. 1–2) and received both CS input and non-specific background input. There is ample experimental evidence that in different contexts, different sets of hippocampal neurons (e.g. in CA1) are active [20]–[22]. Thus, to mimic context-specific inputs - either directly from hippocampus or indirectly via the mPFC or other brain structures such as entorhinal cortex - we provided population A with additional input reflecting , and likewise, population B with additional input reflecting . Populations A and B were mutually interconnected with inhibitory synapses. The system of differential equations describing the activity of the populations A and B is as follows:
(1)
(2)where .
The evolution of the connection strengths is given by (3)
(4)
Here, represents the connection strength from population (or external input) Y to population X, is the time constant governing the dynamics of population X, kX is the maximum firing rate of population X, and rX captures the refractoriness of neurons in X. The transfer function S is a sigmoid function, integrating all inputs to population X in a non-linear fashion and producing a bounded output rate. The parameters p and θ of the sigmoid function determine the steepness and the position of its maximum slope, respectively. The term η(t), with zero mean, reflects the stochastic input to the two populations, mimicking the background activity in the BA.
Equations 3 and 4 describe the dynamics of the connection strengths of the CS afferents onto populations A and B respectively. These weights were increased in an additive way whenever the respective CS and CTX inputs were present simultaneously and remained constant otherwise. The parameters aA and aB specify the learning rates (see also Eqs. 6–8).
We simulated fear conditioning and extinction by applying CS input to both populations in the form of short pulses of 50 ms duration each, based on the experimental design used in [17]. Contextual input was provided continuously. Note that we did not make any explicit distinction between the unconditioned stimulus (US) and conditioned stimulus (CS). Instead, we assumed that during conditioning, neurons in the LA initially responded to the US and eventually to the CS, while continuing to respond to the CS during extinction [23]. The output of these LA neurons was then fed downstream to the BA. In addition, US or CS inputs from the thalamus or the primary sensory cortex may directly target BA neurons [24]. In our model, we represented those inputs, independently of their origin, as CS-US in the conditioning context and CS in the extinction context.
For the description of the SNN we adopted the good model description practice proposed by [25], which provides guidelines for a standardized way of describing complex neural networks. We share the authors' belief that such model description facilitates reproducibility and direct comparisons between models. Within this framework, we organized the description in different subsections, complemented by additional information on the model parameters. This collected information is presented in an easily accessible, tabular form in the Supplementary Materials (Table S1).
Our choice to use leaky-integrate-and-fire (LIF) neurons was motivated by four major arguments: (i) multiple combinations of sub-cellular parameters can result in the same network state [26]; (ii) even simple neuron models such as LIF with minor modifications are sufficient to reproduce complex in vivo spike patterns [27]; (iii) realistically-sized large scale networks of LIF neurons can now be simulated with the currently available simulation technology [28]; this is hardly possible for similarly large networks built of detailed compartmental models and, finally, (iv) the extent to which sub-cellular properties of individual neurons influence the global network dynamics is presently not clear. Most importantly, however, here we are interested in understanding the key network level properties of the BA which play a critical role in the formation of fear and extinction memories. For this purpose, the LIF neurons, although they are reduced models of a biological cell, provide an adequate level of biophysical realism, sufficient to identify these key network properties.
We modeled the BA as a random recurrent network, consisting of excitatory (EXC) and inhibitory (INH) neurons [24], [29]. A total number of 4000 neurons corresponds roughly to 10% of all neurons in the rat BA [30]. The schematic diagram of the network is shown in Fig. 2B. Each connection from a pre- to a post-synaptic neuron had an assigned probability, the value of which depended on the types of pre- and postsynaptic neurons involved (EXC and INH, respectively): , , , and . Thus, each EXC neuron received on average excitatory and inhibitory connections. Likewise, each INH neuron received excitatory and inhibitory connections. Neurons were allowed to form recurrent connections to themselves. For the simulations shown in the last figure, we systematically varied the connection probability of the recurrent inhibition from 0.1 to 1.0.
EXC and INH neurons received inputs encoding information on the CS. Similarly to the rate model, these inputs represented initial responses of LA neurons to combined CS and US presentations, later only to the CS. They might also reflect more peripheral, thalamic or cortical responses to CS-US. A fraction of BA EXC neurons (20%, randomly chosen) received inputs representing CS and . Similarly, another 20% of BA EXC neurons received inputs representing CS and . Thus, similar to the rate model, we assumed that BA EXC neurons receive contextual information directly from the HPC (or entorhinal cortex) and/or via the mPFC. Crucially, CS-US and contextual inputs converged onto the same neurons [8]. Furthermore, EXC and INH neurons received unspecific background inputs (BKG), representing activity originating in other areas, either within or outside the amygdaloid complex. The BKG inputs accounted for the baseline spiking activity of EXC and INH neurons at <1 Hz and 10–15 Hz, respectively [24].
The exact temporal and spatial patterns of the spiking inputs to the BA are not known. Here, we used independent Poisson spike generators with different firing rates to produce the specific inputs. Contextual and BKG inputs provided a tonic drive to BA neurons. By contrast, the CS input had a short duration of 50 ms, based on the experimental design used in [17]. All external inputs formed excitatory synapses onto their target neurons.
Neurons were modeled as leaky-integrate-and-fire (LIF) neurons. The subthreshold dynamics of each LIF neuron were governed by the following equation(5)
A spike was generated whenever the membrane potential crossed a predefined static threshold θ in upgoing direction. The potential was then reset to a value Ek and clamped for tref ms before the synaptic integration started again (Table S1F). Neurons made either excitatory or inhibitory connections onto their postsynaptic targets via conductance-based synapses [31]–[33].
The synapses of all connections were non-modifiable, except those providing CS and contextual input to EXC neurons. These latter, plastic synapses were modified according to the following phenomenological rule:
(6)
(7)
(8)
Note that three variables were used: the synaptic weight w and the auxiliary variables c and h. Each time a presynaptic neuron fired, the value of c increased by a fixed amount. Afterward, this value relaxed towards zero. Thus, variable c acted as a synaptic tag, encoding the recent activity in the synapse receiving CS input. Likewise, variable h encoded information about recent activity in neighboring synapses receiving contextual input.
At the offset of each CS presentation, the variables c and h were probed in the synapses of all EXC neurons and the strength of each synapse was modified accordingly. The synaptic strengths before and after the update are denoted by w− and w+, respectively. If CS and contextual inputs at the same neuron coincided within a temporal window of ∼100 ms, then both synapses were strengthened [34]. By contrast, if only one of the inputs was present, both synapses were weakened (Eq. 6). This decrease of synaptic strength was based on studies reporting that synapses in LA, which had been strengthened during fear conditioning, depotentiated after extinction training [35], [36]. We assumed a similar mechanism to hold for the BA. This type of bidirectional plasticity rule implemented in our model is similar to the BCM rule [37], the “calcium-control hypothesis” [38]–[40] and the ABS rule [41], [42]. Common in all these rules is the specification that the level of postsynaptic Ca2+ determines the direction of plasticity (for review see [43]). A large increase in Ca2+ causes LTP, whereas a moderate increase results in LTD. Low levels of Ca2+ do not cause any modification at all. We essentially incorporated this bidirectional induction of plasticity in our rule using fixed thresholds (Fig. 3C), rather than sliding ones, as is the case e.g. in the BCM rule. The parameters a1 and a2 denote the learning rates for potentiation and depotentiation of the synapses, respectively.
Ca2+ influx depends on NMDA receptor activation and sufficient postsynaptic depolarization. The latter can be caused by coincident presynaptic input or by a backpropagating action potential (BAP). However, in our model, a BAP was not required. That is, we assumed that if the total presynaptic firing rates were high enough, they could cause sufficient depolarization to unblock NMDA receptors. This assumption is supported by experimental evidence showing that a BAP is neither necessary nor sufficient for synaptic plasticity [44], [45].
Note that this plasticity rule is also compatible with changes induced purely in the presynaptic terminal. In fact, experimental evidence suggests that presynaptic induction, completely independent of postsynaptic activity, occurs in the LA [46]. Thus, the plasticity rule implemented in our model incorporates both changes that are dependent on post-synaptic depolarization, but not postsynaptic spiking, and changes that are presynaptic and entirely independent of post-synaptic depolarization or spiking.
Because in our model the presynaptic spiking was caused by CS and contextual inputs, their total activity encoded in the variables c and h, respectively, determined the direction of plasticity. Thus, both c and h functioned as eligibility traces for synaptic modification [34], [47]. They could be interpreted as describing any relatively slow process associated with the effects of Ca2+, e.g. autophosphorylation of CaMK-II [39], [48].
The terms and in the update rule were introduced to provide upper and lower bounds to the synaptic weights, such that they did not increase or decrease indefinitely. They also controlled the step-size with which synapses were modified: the closer a weight was to wmax (wmin) the smaller were its increments (decrements).
The parameter m represented the action of neuromodulators released during fear conditioning and extinction. It is known that many neuromodulators target the BA [5], possibly affecting synaptic plasticity in a complex way. Among the possible candidates are norepinephrine (NE) [49]–[52], dopamine (DA) [53], [54] and opioids [5]. Here, however, lacking more detailed experimental data, we cannot be more specific about which exact neuromodulators are involved and how they interact. Fortunately, this lack of knowledge does not pose a problem for the plasticity rule we propose, because it is general enough to accommodate any combination of neuromodulators that may turn out to be involved in BA fear processing.
The dynamics of the mean-field model were simulated in MATLAB. The SNN simulations were written in python (http://www.python.org), using the PyNN interface [55, http://neuralensemble.org/trac/PyNN] to the NEST simulation environment [56, http://www.nest-initiative.org].
Fig. 4 shows the response of the mean-field model, i.e. the firing rate model, of BA during fear conditioning and extinction. To simulate fear conditioning in , we stimulated the population A five times with CS, US and inputs (Eqs. 1,2). This resulted in a progressive strengthening of CS synapses onto population A () (Fig. 4C), accompanied by a corresponding increase in the response of population A (Fig. 4A). To simulate fear extinction training in , we stimulated population A with CS input and population B with CS and input six times to mimic a different context. Now, in , the synaptic strength of the CS input synapses () onto population B progressively got stronger, whereas remained unchanged (Fig. 4D). The slow increase in the response of population B resulted in a small decrease in the response of population A, due to the recurrent inhibition. When the strength of became larger than (Fig. 4D), the activity of population B dominated and, hence, the response of population A was suppressed (Fig. 4B).
The differential activation of two neuronal sub-populations in two different contexts can be interpreted as fear (population A) and extinction (population B) neurons as observed in [17]. This is purely a functional characterization of the two sub-populations, which are identical otherwise. That is, we used exactly the same parameters for both sub-populations and the differential activation results solely from differences in contextual inputs they receive. Thus, the two populations were not different in terms of their intrinsic properties. Of course, cases where the two subpopulations do have different properties can be easily accommodated in the model resulting in an enhancement of the differential activation. To be consistent with [17], we used the terms fear and extinction neurons to refer to those subpopulations that are active in and respectively.
Note that we did not include any component that imitates behavioral output, i.e. freezing. Instead, we assume, in agreement with experimental findings [17], that high activity of fear neurons directly corresponds to a high level of freezing whereas high activity of extinction neurons and low activity of fear neurons corresponds to low levels of freezing.
Although a simple firing rate model was able to account for the dynamic emergence of fear and extinction neurons, such mean-field models have only limited explanatory and predictive power. For instance, they assume uncorrelated activity in the underlying neuronal populations and, thus, cannot be used to predict any correlations in firing rate or spike timing that may emerge in the network. In addition, these models cannot be used to predict the spike patterns of individual neurons. Thus, to understand the dynamics of the BA network beyond average firing rates only, we simulated a biologically realistic large-scale network composed of spiking neurons. Again, fear conditioning and extinction were simulated by applying five CS-US presentations in and six CS presentations in respectively. In the two different environments tonic contextual input was provided to EXC neurons (cf. Models).
The results of the simulation are presented in Fig. 5. Initially, all EXC neurons spiked at very low firing rates. Presentations of the CS-US led to a steady increase in the firing rates of one sub-population (fear neurons) within the EXC population, which peaked at the end of conditioning (Figs. 5A, E amber dots). The increase in activity of fear neurons was a direct consequence of the potentiation of CS and contextual inputs onto fear neurons (Figs. 5 G, I; amber triangles). In , the fear neurons still responded with high firing rates upon the first CS presentation, even though they did not receive contextual inputs (Figs. 5A, F). With further CS presentations, however, synapses became potentiated (Eq. 6, Figs. 5H, J; cyan dots), causing a steady increase in the firing rate of the second sub-population of neurons (extinction neurons) (Figs. 5A, F; cyan dots). The increased recurrent inhibition in the network then caused a decrease in the activity of the fear neurons (Figs. 5A–C, F). At the end of extinction, the population rate of the extinction neurons peaked, whereas the firing rate of the fear neurons had returned to the initial, pre-conditioning values. The reduction of fear neurons activity was further facilitated by small depotentiation of CS and contextual input synapses onto the fear neurons (Eq. 6, Figs. 4H, J; amber triangles). Note that depotentiation of CS synapses onto extinction neurons also occurred during conditioning (Fig. 5G) as described by the learning rule. By contrast, CTX synapses were not decreased during conditioning, because their initial values were close to the lower bound (w−) (Fig. 5I). During conditioning and extinction the baseline firing rates increased as well (Fig. 5A). This increase was induced by the strengthening of the contextual inputs (Figs. 5H, I), providing an explanation for contextual conditioning. However, because only a small percentage of neurons exhibited this increase in firing rates, this could make it difficult to measure it experimentally. This fact reveals a key advantage of network models which allow for simultaneously sampling a large number of neurons. Based on this, predictions can be inferred which otherwise would not have been possible.
Note that, again, the assignment of BA EXC neurons in fear and extinction sub-populations is purely a functional one. That is, neurons were characterized post-experiment as fear or extinction cells depending on whether they responded to the CS after conditioning or after extinction training respectively. In particular, they were not predetermined in terms of their intrinsic properties and the two sub-populations resulted solely from the differences in the contextual inputs they received.
Also, it is important to emphasize that whereas the population rates of fear and extinction neurons increased gradually during conditioning and extinction training respectively, this was not the case for individual neurons. Instead, they changed their state quite abruptly from non-responding to responding (Fig. 6A). The further the training advanced, the more neurons started to respond. Hence, the gradual increase in population rates (Figs. 5E, F) reflects the growing recruitment of responding neurons, rather than a gradual increase of single neuron activity itself (Fig. 6). The responsive neurons fired maximally two spikes per CS presentation. The baseline firing rates for the inhibitory population were normally distributed with a mean of 10 Hz, whereas the CS-evoked rates shifted their distribution towards a mean of 20 Hz. This is consistent with the neuronal firing patterns in vivo reported by [17].
Although we performed our main simulations using separated contextual inputs to distinct neuronal subpopulations within the BA (cf. Models), this is not a necessary requirement of the model. In fact, performing simulations with varying amounts of contextual input overlap showed that fear and extinction neurons still existed as distinct populations, even when contextual inputs had an overlap of around 50% (Fig. S1). In addition, the simulations revealed the existence of a third sub-populations of neurons. These were the neurons receiving inputs in both contexts and, thus, were active during both fear conditioning and fear extinction (so called persistent neurons). Note that, similar to the case of fear and extinction neurons, the characterization of cells as “persistent” is functional and denotes the fact that these neurons were responding to the CS during both conditioning and extinction. Moreover, these neurons had much stronger CS and CTX synapses, which resulted in higher firing rates. This observation of the model is consistent with the experimental data [17], suggesting that conditioning and extinction are not affected by overlapping inputs, unless the overlap is high (>50%).
Following extinction training in , presentations of the CS in the original fear conditioning context () resulted in context-dependent renewal (ABA renewal) of conditioned fear responses [2]. This renewal phenomenon points at two important aspects of possible neural mechanisms underlying fear extinction: (i) extinction is mainly new learning and only partly unlearning of previously acquired fear memories ([57]; see also Discussion), (ii) extinction learning is context-dependent.
We simulated ABA renewal by changing context at the end of extinction (Fig. 7). This resulted in a sudden switch of activity between fear and extinction neuronal subpopulations. That is, although the activity of extinction neurons was high after extinction learning, the contextual switch caused the activation of fear neurons and a significant drop in the extinction neurons activity. These results are in complete accordance with the experimental findings reported by [17].
It is important to note that this rapid activity switch is purely a network phenomenon and not an effect of synaptic plasticity, as the change is much too fast for the plasticity mechanisms to act. We illustrate this point by depicting the average membrane potentials of 100 randomly selected fear and extinction neurons (Figs. 5D, 7D; amber and cyan traces respectively). It is evident that in either context there was a clear difference between the membrane potentials of the two cell populations, stemming from the fact that one of the populations continuously received a higher excitatory drive due to the additional contextual input. Switching contexts led to a corresponding instantaneous switch in the assignment of the contextual input and, hence, in opposite shifts in the average membrane potentials of the two sub-populations, which was immediately reflected in corresponding shifts in the firing rates.
We also modeled the case where the renewal context was different from both the conditioning and the extinction context (ABC renewal). The results of the simulations revealed that if after extinction training the CS was presented in a third, different , fear neurons became rapidly active again and suppressed extinction neurons (Fig. S2 middle). However, our model also indicated that the absolute response of fear neurons - and thus the magnitude of the fear response- would be weaker than in the ABA case. The reason is that in CTX synapses had not been strengthened during the conditioning procedure. This provides an account for the experimentally observed ABC renewal [58], [59] explaining why ABC renewal may occur in the first place and also why the effect may be weaker compared to ABA renewal.
Moreover, our simulations also suggested that massive extinction (extinction over-training) in can abolish ABC renewal, because depotentiation of CS and afferents onto BA neurons yield less excitatory input to these neurons. Extinction over-training can also impair ABA renewal, although to a lesser extent (Fig. S2 right). The reason that ABA renewal is more robust and ABC renewal more vulnerable to massive extinction stems from the fact that in the latter case not only CS synapses onto fear neurons are weakened, but also potentiated CTX synapses are entirely missing. These findings are in agreement with and provide a possible explanation for the experimentally observed effects of massive extinction [60].
Although we did not focus on extinction of contextual fear, it is important to note that our model also accounts for this specific conditioning phenomenon. Indeed, the plasticity rule dictates that in the absence of the CS synaptic weights will decay. That is, CTX synapses, which had been strengthened during conditioning in and encode contextual fear, will depotentiate in the same context if the CS is not present. This will yield decreased fear neuron activity and, thus, extinction of contextual fear. Note that within the framework of our model, this form of extinction is truly unlearning and not masking of contextual fear memories.
The experimentally reported connection probabilities from excitatory to inhibitory neurons as well as among inhibitory neurons in the BA are around 0.5 [61]. This is a much higher value than the ones we used in our initial simulations (Figs. 5–7, Table S1E). To test the effects of such higher connectivity, we performed additional simulations adopting the experimentally reported values for the connection probabilities. The qualitative behavior of the model did not change (data not shown). However, a new aspect in the network dynamics emerged. High frequency oscillations - typically in the gamma range (30–80Hz) - occurred throughout the simulation in both excitatory and inhibitory populations. These oscillations were present already in the ongoing activity patterns and CS-US presentation enhanced them even further (Fig. 8A). They resulted from the high shared connectivity and, hence, large amount of shared inputs that caused correlated spiking in the neurons. The oscillation frequency was determined by synaptic time constants and delays in the network.
Gamma oscillations in networks of excitatory and inhibitory neurons have been reported in many experiments [62]–[67] and discussed in multiple theoretical studies [68]–[75]. Moreover, several studies have reported gamma oscillations in the amygdala under certain conditions, e.g. in anesthetized animals [76], in slow wave-sleep [77], in the presence of reward predicting stimuli [78] and in paradigms involving consolidation of emotional memories [79]. Therefore, there is at least partial experimental and theoretical support for the gamma range oscillations observed here in high connectivity BA network simulations.
Yet, in networks with high mutual connectivity between excitatory and inhibitory neurons and within inhibitory neuron populations such as in the BA, oscillations should be a prevailing feature and should, therefore, be readily identifiable in vivo recordings under all conditions and not only in the special cases mentioned above. It is, thus, possible that certain mechanisms operate in the BA that could dampen gamma oscillations (but see Discussion). We, therefore, used our network model to investigate this issue in further simulations by exploring the parameter space of the network properties that could quench oscillations. Two mechanisms proved to be effective in reducing the power of gamma oscillations. The first one was the introduction of heterogeneity in the inhibitory population [80], [81]. This approach was motivated by experimental data showing that interneurons in the BA exhibit a large diversity in terms of their morphological and electrophysiological properties [24], similar to interneurons in the cortex [82] and hippocampus [83]. In the latter case, the diversity was expressed in a wide range of values for synaptic rise times, reversal potentials, response latencies etc. In a preliminary study [84], we introduced heterogeneity in one of the neuronal properties in our model, the spiking threshold, by drawing values from a bimodal distribution with peaks at −35 mV and −28 mV. This corresponds to the experimentally measured threshold values of two subclasses of parvalbumin-expressing interneurons in the BA: the fast-spiking (FS) and the delayed-firing (DF) interneurons [24], [61]. In such heterogeneous networks, oscillations were indeed reduced, but not totally eliminated [84].
A second, more effective way to reduce the network oscillations was to decrease the synaptic delays between inhibitory neurons (Fig. 8B). First, we studied the oscillations dynamics for different connection probabilities in a network of homogeneous neurons, with synaptic delays drawn from a uniform distribution (1–2 ms). In such networks, increasing connectivity (>0.2) enhanced the oscillations and synchrony to their maximum (Fig. 8B, green solid lines). Only for very weak synapses (1 nS), that is, when the network was mainly driven by external inputs, increasing the connectivity did not add to the oscillations (Fig. 8B, gray solid line). Increasing the width of the synaptic delay distribution did not reduce the synchrony and oscillations in high-connectivity networks (data not shown, Vlachos et al. in prep.). However, choosing short delays from a narrow uniform distribution (0.2–1 ms) considerably reduced the oscillations, up to connection probabilities of 0.4 (Fig. 8B, green and gray dashed lines). Thus, in a recurrent network, smaller delays have a powerful effect in reducing synchrony and oscillations. This finding is in agreement with a previous numerical study [69] and also with more recent analytical approaches [74], [75], [85]. At first sight, synaptic delays less than 1 ms might appear unrealistically small. However, delays as short as 0.5 ms have been reported among inhibitory neurons in the hippocampus [66]. Moreover, the delays between inhibitory neurons in the BA have been reported to be around 0.7 ms [61], or even smaller (Lüthi, unpublished data). Therefore these short delays, might indeed account for the lack of gamma band oscillations observed under baseline conditions in experimental recordings.
Because inhibition plays a critical role in our model, we tested the effects of partial inactivation of inhibitory neurons. For this, we performed two additional sets of simulations, in which, during acquisition of extinction, we deactivated 50% and 90% of INH neurons, respectively. The results are shown in Fig. 8C. As expected, with reduced inhibition the activity of both fear and extinction neurons increased. The increase of activity of the latter population was more pronounced, due to the fact that it received additional excitatory drive from contextual inputs in . This suggests that blockage of inhibitory activity should lead to enhanced, context-specific extinction. This is consistent with the finding that GABA blockage enhances extinction of contextual freezing [86]. However, there is a potential caveat here. Activity of both fear and extinction neurons is increased upon blockage of inhibition and it is not clear how downstream structures, specifically CEA neurons, would respond to this. If the relative difference between fear and extinction neuron activity matters, then extinction should be facilitated by impaired inhibition. If, by contrast, the ratio between fear and extinction neuron activity is more relevant, then extinction might be impaired. Note that these two possibilities apply to both blocking of inhibition during acquisition of extinction training and blocking of inhibition during expression of fear extinction.
Because contextual input is one of the key aspects in our model we tested how removal of these inputs would affect the behavior of the network. The simulations yielded two different results depending on the exact time point of removal of contextual inputs. When contextual inputs were removed after fear conditioning, fear neurons remained active during extinction training and no extinction neurons emerged (Fig. S3; left). This result is a direct consequence of our synaptic learning rule, because strengthening of synapses requires temporal overlap of CS and CTX inputs. Note that although fear neurons remained active, their firing rates were reduced, because they now lacked contextual input. Thus, our model suggests that lesions of hippocampal or prefrontal areas after fear conditioning may result in impaired extinction. This conclusion is supported by experimental evidence [87].
By contrast, when contextual inputs were removed after fear extinction, activity of neither fear nor extinction neurons was sufficiently strong to suppress the other neuron group (Fig. S3; right). That is, because the decisive contextual input was lacking and, thus, both groups were simultaneously active, although to a lesser degree than in case either group was active alone. The behavioral consequences of these results are beyond the scope of our model, because here we did not model any downstream structures such as the central amygdala that presumably further process output from fear and extinction neurons. Thus, at present, we can only speculate that lesions of hippocampal or prefrontal areas after extinction training may result in impaired renewal, because fear neuron activity will be both decreased and also counteracted by simultaneous extinction neuron activity. In fact, experimental evidence supports this conclusion [88]. However, it is important to point out a subtle difference between our model and certain lesion experiments. In our model, removing CTX input means that the BA network does not receive any contextual input at all. By contrast, in some experiments in which the hippocampus had been lesioned or inactivated, contextual information may still have been accessible, because the context in which the CS was presented was still decisive for the behavioral outcome [89], [90].
Our model enables us to make a number of specific predictions that can be tested experimentally:
Here we presented for the first time a large-scale network model of the BA addressing the question how contextual inputs may shape the activity of distinct sub-populations of BA neurons. Although we started from a very specific experimental data-set [17], we implemented a network model that has more far-reaching implications. That is, the results of the simulations together with the model architecture provide, non-trivial and experimentally testable new insights into potential neural mechanisms underlying cued and contextual fear conditioning and extinction, ABA and ABC renewal, and extinction over-training. In addition, a specific and important function of inhibition is sketched as a mechanism that could enable mutual competition between fear and extinction memories. These results allow us to provide a synthesis of several experimental findings and to propose a role for the BA as a nucleus that integrates information about the CS and the context. This brings it into the position to provide a context-dependent instruction to downstream structures, enabling the switching of states from low to high fear and vice versa.
Specifically, we propose that context modulates neuronal activity within the BA, resulting in the formation of associations between CS, US and context in this nucleus. During fear conditioning the representation signals danger and causes a high fear state. During extinction, the newly formed representation signals safety and suppresses the fear state. Back in the conditioning context, the initial representation dominates again (renewal). Thus, as far as neural mechanisms within the BA are concerned, conditioning and extinction could be understood as mutual competition between different representations of fear and safety. Partial unlearning or erasure may also occur, although to a limited degree.
Memories are assumed to be stored in a distributed manner in the brain [92]. Consistent with this view, fear-related memories may also be distributed among different nuclei within the amygdala and brain regions connected to it [10]. Our model suggests that context-related features of these distributed fear memories are represented in the BA. Thus, inactivation of the BA would impair context-related aspects of fear and extinction memories, whereas non-contextual features, represented in LA or CEA, would remain unaffected [17].
One core feature of our model is that contextual inputs are gated to the BA. In this framework, the precise origin of these inputs does not matter; as long as the BA neurons receive differential inputs in two different contexts, the model behavior remains unaltered. However, there are strong indications from anatomical [29], [93], [94] and physiological [93] studies that the HPC is a major source of contextual information to the BA. In addition, a previous report showed context-dependent modulation of neuronal activity in the LA [95]. By designing our model to have contextual input directly influencing the activity of excitatory neurons in the BA, we have essentially postulated a similar mechanism for this subnucleus. This assumption is further supported by the finding that fear neurons show orthodromic responses to HPC stimulation [17].
A second source of contextual input may be the mPFC. There is anatomical evidence that the mPFC projects to the BA [29]. Moreover, [17] reported that mPFC stimulation induces orthodromic responses in identified extinction neurons. Here, we suggest that part of the information conveyed by these projections might be contextual. This assumption is based on evidence reporting extinction-related induction of LTP on hippocampus-mPFC afferents [96]. In our model both fear and extinction neurons receive context-specific information either directly from hippocampus or indirectly via the mPFC. This may also explain the ambiguous results that the hippocampus may or may not interact with the mPFC during extinction learning [97].
The context-specific modulation of activity in the BA presented here provides a general framework that can explain experimental findings on the involvement of the hippocampus in the acquisition, encoding, and context-dependent retrieval of both conditioning [13], [98], [99] and extinction memories [3], [87]. Future refinements of the model, in combination with new experimental data are necessary for a better understanding of the detailed interactions between hippocampus, mPFC and amygdala.
We showed that high connectivity between excitatory and inhibitory and within inhibitory neuron populations results in robust oscillations in the gamma range, characterized by high activity correlation among neurons. The main cause of these oscillations was the high degree of shared inputs among neurons as a result of the dense connectivity. We suggested two different, biologically plausible ways to reduce these oscillations: by either introducing heterogeneity in neuron properties and/or by reducing synaptic delays to sub-millisecond time scales. Yet another way would be to have synapses exhibit a certain transmission failure rate [100], [101], resulting in activity dependent reduction of the effective connectivity. However, we do not wish to imply that gamma oscillations do not exist in the BA. In fact, as noted earlier, gamma oscillations have been reported in the amygdala under various conditions [76]–[79]. Here, we want to emphasize the point that in networks with high connectivity, gamma range oscillations are a salient feature of the network dynamics. Therefore, they should be visible even in the ongoing activity, unless suppressing mechanisms, such as those elaborated here, are in effect.
Several suggestions for a specific role of gamma oscillations have been made in the past. For instance, it has been proposed that in the cortex or the hippocampus oscillations might contribute to temporal encoding [102], sensory binding [103], attentional selection [104] and memory formation or retrieval [105], [106]. It is currently unclear whether these hypotheses also apply to the amygdala. Oscillations in lower frequency ranges (delta and theta) have also been reported. For example, increased theta oscillations - that synchronized with hippocampal theta activity - were shown to be related to conditioned freezing [107], [108], whereas delta oscillations have been implicated in gating aversive stimuli [109]. Gamma oscillations, on the other hand, have been suggested to facilitate interactions between the amygdala and connected structures [78], [110].
Here, because we modeled only the BA, we cannot give any informed predictions about how gamma oscillations may affect those various interactions. Moreover, in our current model, we have used plasticity only in the input connections and those are not affected by oscillatory activity in the recurrent network. However, before addressing the effects of gamma oscillation on the dynamics of the BA network, it is of key importance to resolve experimentally whether gamma oscillations are indeed present in BA activity and, if so, under which conditions.
A well-known behavioral phenomenon is conditioned inhibition, referring to the ability of a second CS (CS−) to suppress the conditioned response, after it has been paired several times with the first CS (CS+) in the absence of a US [57], [111]. It is possible that the CS−, referred to as conditioned inhibitor, employs similar mechanisms to those described in our model to suppress the conditioned response. That is, neural subpopulations in the BA encoding the CS− might, similar to extinction neurons, use local inhibitory circuits to suppress fear neuron activity. Future work is needed to explore further this interesting line of reasoning.
Our model accounts for experimental paradigms that use a different extinction context from the conditioning one, but not for those in which fear conditioning and extinction occur in the same context. For instance, if conditioning and extinction both occur in , then only those neurons that receive inputs in this context will be active. Thus, downstream structures will not be able to differentiate between fear conditioning and extinction training solely from spiking activity in the BA. It is evident that performing conditioning and extinction in the same context per se increases ambiguity about the meaning of the context. Thus, it is likely that circuits within the BA alone are not sufficient to solve this computational problem. Both, a detailed description of neural activity during this type of extinction and a more detailed analysis of interactions between the BA and downstream structures are required to address this behavioral phenomenon.
Although a wealth of experimental studies exist on the amygdala and its role in fear conditioning and extinction, computational or theoretical approaches to study amygdala function are largely lacking. Most of the previous theoretical studies involve symbolic models [112], [113], mainly based on the Rescorla-Wagner rule [114]. These models have their merit in describing behavioral findings such as generalization, blocking etc. However, since these models treat the amygdala as a “black-box”, it is not within their scope to account for neuroanatomical or electrophysiological data, therefore providing little insight into the underlying neuronal mechanisms involved. Despite these apparent differences, it is still possible to draw some parallels to symbolic models. For instance, in our model, potentiation of synapses occurs only if CS and CTX inputs temporally overlap. This is similar to the SOP model, where US and CS have to coincide for strengthening of associations to take place [115], [116].
Connectionist or parallel-distributed (PDP) models of fear related processes go one step further than symbolic models by introducing networks composed of multiple, mutually connected computational units. One such model was successful in capturing certain features observed in fear conditioning studies [117]. Its main limitation, however, is the fact that it does not take into account the different substructures within the amygdala, nor do the computational units used in the model map to any biophysically realistic counterparts.
Fortunately, the computational power presently available allows us to improve these models and to overcome many of their limitations. The model presented here is to our knowledge the first large-scale spiking neuron network model that investigates the mechanisms of fear conditioning and extinction within the amygdala using biologically realistic neurons in adequate detail. The model closest to this is a compartmental model introduced by [118] to investigate the function of the LA in fear conditioning and extinction. However, [118] used a small network composed of only eight two-compartment neurons and focused on role of the kinetics of multiple ionic currents in fear conditioning and extinction. By contrast, we modeled the BA using a large network of 4000 LIF neurons, which enabled us to identify the network level interactions involved in the formation of fear and extinction memories.
The present model provides a plausible explanation for the neural mechanisms underlying fear conditioning and extinction within the BA. We did not address the question of how the neural activity within the BA impacts on downstream structures, such as CEA or mPFC. We neither attempted to model the interactions between hippocampus and mPFC in conditioning and extinction, which would require additional experimental data to constrain the possible models. Given these restrictions, we provided a plausible mechanism of how contextual inputs may affect the activity of distinct neuronal subpopulations in the BA, enabling them to control downstream structures such as the CEA. We proposed that context-related aspects of fear and extinction memories are partially stored in the BA and that they provide a context-dependent instruction for the triggering or blocking of the fear-response. In addition, we showed how extinction training may mask previously acquired fear memories and, thus, provided an account for renewal. Finally, our model, next to yielding several interesting predictions discussed above, raises the important question of how downstream structures such as the CEA or mPFC discriminate the activity of the distinct neuronal subpopulations within the BA. Is this problem solved purely on an anatomical level, e.g. by differential projections of the BA subpopulations to specific target neurons? Or do specific features in the activity of the BA subpopulations, e.g. the statistical structure of pairwise or higher-order correlations, also play a role, providing downstream networks with a mechanism to distinguish between them? These questions need to be addressed in future work combining experimental and theoretical approaches.
|
10.1371/journal.pntd.0001858 | Leishmania mexicana Induces Limited Recruitment and Activation of Monocytes and Monocyte-Derived Dendritic Cells Early during Infection | While C57BL/6 mice infected in the ear with L. major mount a vigorous Th1 response and resolve their lesions, the Th1 response in C57BL/6 mice infected with L. mexicana is more limited, resulting in chronic, non-healing lesions. The aim of this study was to determine if the limited immune response following infection with L. mexicana is related to a deficiency in the ability of monocyte-derived dendritic cells (mo-DCs) to prime a sufficient Th1 response. To address this issue we compared the early immune response following L. mexicana infection with that seen in L. major infected mice. Our data show that fewer monocytes are recruited to the lesions of L. mexicana infected mice as compared to mice infected with L. major. Moreover, monocytes that differentiate into mo-DCs in L. mexicana lesions produced less iNOS and migrated less efficiently to the draining lymph node as compared to those from L. major infected mice. Treatment of L. mexicana infected mice with α-IL-10R antibody resulted in increased recruitment of monocytes to the lesion along with greater production of IFN-γ and iNOS. Additionally, injection of DCs into the ear at the time of infection with L. mexicana also led to a more robust Th1 response. Taken together, these data suggest that during L. mexicana infection reduced recruitment, activation and subsequent migration of monocytes and mo-DCs to the draining lymph nodes may result in the insufficient priming of a Th1 response.
| Leishmaniasis, caused by protozoan parasites belonging to the genus, Leishmania, exhibits clinical symptoms ranging from mild cutaneous lesions to more severe cutaneous or visceral disease. Here, we focus on L. major and L. mexicana, two species that lead to self-resolving and chronic cutaneous lesions, respectively. A strong Th1 response is necessary for resolution of disease following L. major infection. However, L. mexicana infection induces a limited Th1 response resulting in chronic disease. Monocyte-derived dendritic cells are believed to be important in priming the Th1 response during L. major infection, and therefore in this study we evaluated whether there are quantitative and/or qualitative differences in monocyte-derived dendritic cells following L. mexicana infection. We found that fewer monocytes were recruited to the lesions of L. mexicana infected mice as compared to mice infected with L. major. In addition, there were fewer iNOS producing monocyte-derived dendritic cells in the lesions of L. mexicana infected mice and less migration of monocyte-derived dendritic cells to the draining lymph node. Manipulations that allow for increased monocytes in the lesions of L. mexicana infected mice also resulted in a more robust Th1 response. Thus, these findings provide a mechanistic basis for the limited Th1 response observed during L. mexicana infection and also offer a better understanding of the important role that monocytes play during infection with Leishmania.
| Infection with Leishmania results in a variety of outcomes, depending on the parasite species and immune response mounted by the host [1]. Murine disease models resemble human disease, with some infections being self-healing and others chronic. Resolution of leishmaniasis requires the production of IFN-γ by Th1 cells; the absence of a strong Th1 response results in chronic disease with non-healing lesions [2], [3]. Th1-mediated protection is promoted by IFN-γ-induced production of nitric oxide (NO) in infected cells, which ultimately leads to parasite killing [2], [3]. In C57BL/6 mice, infection with L. major results in a strong Th1 response with self-resolving lesions, in contrast, L. mexicana lesions fail to resolve [4]. The chronic nature of L. mexicana lesions is most likely due to their inability to stimulate an effective Th1 response [4], [5], [6]. Similarly, L. amazonensis fails to induce a strong Th1 response and leads to chronic lesions in mice [7], [8], [9], [10]. However, the immune mechanisms limiting Th1 responses following either L. mexicana or L. amazonensis infection are not yet fully defined.
Several demonstrations that infection with L. mexicana suppresses IL-12 production by macrophages and dendritic cells (DCs) [11], [12], [13] suggested that failure to produce IL-12 may limit the Th1 response, resulting in the observed susceptibility to L. mexicana [14], [15], [16]. However, we found that administration of IL-12 failed to promote disease resolution, suggesting that the inability of L. mexicana mice to resolve their infection is not solely dependent upon lack of IL-12 [17]. Therefore, we hypothesized that a more generalized deficit in DC function may contribute to the chronic lesions that develop following L. mexicana infection.
Monocyte-derived DCs (mo-DCs) play an important role in the development of protective immunity [18], [19], [20], [21]. Mo-DCs differentiate from inflammatory monocytes (CD11b+, Ly6C+, CCR2+ and CX3CR1lo) recruited to sites of inflammation. Once activated, mo-DCs produce inducible nitric oxide synthase (iNOS) [22]. Indeed, mo-DCs appear to be the major producers of iNOS during L. major infection [23] and are therefore likely essential for reducing the parasite burden.
In addition to iNOS production, mo-DCs contribute to immunity following infection with L. major by migrating to draining lymph nodes (dLN) where they stimulate antigen-specific Th1 T cell responses [24]. Moreover, we recently found that L. major-activated DCs induce lymph node hypertrophy, which promotes additional recruitment of naïve T cells into the lymph node to enhance the protective response [25]. Taken together, these data indicate that mo-DCs play an important role in the development of protective immunity to. L. major, and that a deficit in their recruitment or activation might limit a protective Th1 response.
In the present study, we investigated whether the meager Th1 response observed during L. mexicana infection is due to limitations in the: 1) recruitment of monocytes from the blood to the site of infection; 2) differentiation of monocytes into iNOS-producing mo-DCs; and/or 3) migration to the draining lymph node. We found that monocyte recruitment to the site of infection was reduced in L. mexicana infected mice compared to L. major infected mice. Moreover, while monocytes in L. mexicana lesions upregulated expression of CD11c, they produced significantly less iNOS and migrated less efficiently to the draining lymph node relative to monocytes in L. major infected mice. Following treatment of L. mexicana infected mice with α-IL-10R antibody, there was increased recruitment of monocytes to the lesion, as well as increased production of IFN-γ and iNOS. Additionally, when DCs were injected into the ear at the time of infection with L. mexicana, there was a more robust Th1 response. These data imply that the poor Th1 response observed during L. mexicana infection results from both reduced monocyte recruitment to the lesions, and a relative deficit in their differentiation into functional effector populations.
All animal studies were carried out in compliance with the guidelines of the Institutional Animal Care and Use Committee (IACUC) of the University of Pennsylvania and in accordance with the recommendations in the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health. The animal protocol was approved by the IACUC of the University of Pennsylvania, Philadelphia PA.
Female C57BL/6 (B6) and B6-Ly5.2/Cr (CD45.1) mice were purchased from the National Cancer Institute (Fredricksburg, MD). Animals were maintained and experiments were carried out in a specific pathogen-free environment. L. major V1 parasites (MHOM/IL/80/Friedlin) or L. mexicana parasites (MNYC/BZ/62/M379) were grown until stationary phase in Schneider's Drosophila medium (Gibco, Grand Island, NY) supplemented with 20% heat-inactivated FBS (Gibco) and 2 mM l-glutamine (Sigma). Metacyclic promastigotes were isolated by density gradient [26]. For infection of mice, 1×105 metacyclic parasites were injected into the ear.
For flow cytometry, cells were isolated from ears or draining lymph nodes. Dermal ear sheets were separated and incubated in Liberase TL (Roche, Indianapolis, IN) for 1 hr at 37°C. Ears and draining lymph nodes were made into single cell suspensions and washed with PBS. Fixable Aqua dye (Invitrogen, Carlsbad, CA) was added to assess cell viability. Cells were then incubated with Block (CD16/32, inactivated mouse sera and Rat IgG) followed by fluorochrome-conjugated antibodies for surface markers (CD11c, CD11b, CD45.1, CD45.2, MHCII (e-Bioscience, San Diego, CA), and Ly6C (BD Pharmingen, San Diego, CA). Intracellular staining was performed for iNOS using an unconjugated anti-iNOS/NOS II rabbit polyclonal IgG (Millipore, Temecula, CA) followed by flourochrome-conjugated donkey-anti-rabbit IgG (e-Bioscience). Briefly, surface-stained cells were fixed in PBS with 2% paraformaldehyde and then permeabilized with 0.2% saponin in FACS staining buffer (PBS containing 0.1% BSA). Cells were fixed by using 2% paraformaldehyde and samples were acquired on a FACS Canto flow cytometer (BD Pharmingen). Analysis was performed using FlowJo software (Tree Star, Ashland, OR).
Bone marrow was harvested from tibias and femurs of naïve mice. Following lysis of red blood cells with ACK lysis buffer (Lonza, Walkersville, MD), Miltenyi MACS columns were used to purify monocytes. Briefly, anti-Ly6G-biotin antibody and biotin microbeads (Miltenyi, Auburn, MD) were added to the cells, which were then placed over a LS column. Ly6G+ cells on the column were discarded and anti-PE CD11b (e-Bioscience) and PE microbeads were added to the flow through, which was placed over a 2nd LS column. CD11b+ cells attached to the column were washed off. Monocytes were enriched to 50% as evaluated by flow cytometry. 1×106 total cells from the final column were transferred into mice by intradermal injection into the lesions.
Mice were anesthetized using ketamine/xylazine and the ventral side of each ear was painted with FITC isomer (Sigma, St. Louis, MO). FITC (8 mg/mL) was dissolved in an equal volume of acetone and dibutyl phthalate (Sigma) and 25 µL of the mixture was applied to the skin. Migration of FITC+ cells was assessed in the draining lymph node 48 hours following application.
Mice were injected intraperitoneally with 500 µg of α-IL-10R antibody (1B1.3A, BioXcell, West Lebanon, NH) one day prior to intradermal infections of 1×105 L. mexicana metacyclics in both ears. Mice were subsequently treated with 500 µg of α-IL-10R antibody on day 3 and then 250 µg every 3 days until the final harvest at 2 weeks post-infection.
Supernatants from draining lymph node cultures stimulated with L. mexicana freeze-thaw antigen for 72 hrs were collected and assayed by sandwich ELISA using paired monoclonal antibody to detect IFN-γ.
DCs were generated as previously described [27]. Briefly, bone marrow cells from C57BL/6 mice were isolated from femurs and tibias of mice by syringe flushing. Bone marrow cells were counted and seeded into 6-well plates at 5×105 cells/mL in 3 mLs of media - RPMI 1640 (Gibco) supplemented with 10% heat-inactivated FBS (Gibco), 2 mM glutamine (Sigma), 50 µM 2-ME (Gibco), 100 U/mL penicillin (Sigma), 100 µg/mL streptomycin (Sigma) and 20 ng/mL GM-CSF (Peprotech, Rocky Hill, NJ) per well. Cells were maintained at 37°C with 5% CO2 and fed on days 3, 6 and 8 with 3 mLs of fresh media. Cells were harvested on day 10 and injected into the ear of C57BL/6 mice at the time of infection. Briefly, 1×106 DCs and 1×105 L. mexicana metacyclics were mixed immediately prior to injecting into the ear.
Statistical significance was determined using unpaired, two-tailed Student's t test. Results with a p value ≤0.05 were considered significant.
The development of a protective response following L. major infection is associated with the recruitment of monocytes into the lesion, which are believed to differentiate into mo-DCs (defined as CD11bhi, CD11c+, Ly6C+) to prime a strong Th1 response [24]. Since L. mexicana infection promotes chronic, non-healing lesions and a minimal Th1 response, we hypothesized that fewer monocytes would be recruited to lesions following infection with L. mexicana compared to L. major. To test this, we infected C57BL/6 mice with either L. major or L. mexicana parasites and assessed the cellular composition of the lesions at 3 and 14 days post-infection. Expression of CD11b, a subunit of αMβ2 (also known as Mac-1 and CR3), was used to detect infiltrating leukocyte populations, including monocytes, macrophages, and granulocytes [28]. At 3 days post-infection, there was a significant increase in the percentage of CD11bhi cells in dermal lesions from L. major infected mice compared to normal skin (Fig. 1A). In contrast, no increase in CD11bhi cells was observed in lesions from L. mexicana infected mice. Moreover, the difference in percentage of CD11bhi cells between L. major and L. mexicana lesions was still evident two weeks after infection (Fig. 1A). In contrast, neutrophil (CD11bhi Ly6G+) frequency increased equally in lesions of both L. major and L. mexicana infected mice by day 14 as compared to normal skin (data not shown). Consistent with the observed alterations in CD11bhi cells, there was an increase in inflammatory monocytes (CD11bhi CD11c− Ly6C+) in the lesions from 3-day and 14-day L. major infected mice as compared with normal skin (Fig. 1B) while no such increase was observed following L. mexicana infection (Fig. 1B). By day 14, mo-DCs (CD11bhi CD11c+ Ly6C+) were evident in lesions of both L. major and L. mexicana infected mice, however, mo-DCs were preferentially represented in L. major lesions (Fig. 1C). Taken together, these results suggest that L. mexicana fails to promote the recruitment of monocytes, reducing the number of cells available for subsequent differentiation into mo-DCs capable of controlling parasite numbers.
Although there were fewer monocytes recruited to lesions from L. mexicana infected mice compared with those from L. major infected mice, the ratio of monocytes to mo-DCs was similar (Fig. 2). In these experiments we did not determine if the CD11c+ Ly6C+ cells were derived from monocytes, but based on previous findings [24], this is our assumption. However, mo-DCs in the lesions from L. mexicana infected mice expressed significantly less iNOS compared with mo-DCs from L. major lesions. Thus, while approximately 20% of the mo-DCs in lesions from L. major infected mice were iNOS+, only 3% were iNOS+ in lesions from L. mexicana infected mice (Fig. 3A). Similarly, the number of iNOS-producing mo-DCs was significantly reduced in lesions from mice infected with L. mexicana (Fig. 3B). CD11bhi CD11c− Ly6C+, inflammatory monocytes, did not make iNOS in either L. major or L. mexicana infected mice (data not shown). These data demonstrate that there are fewer iNOS-producing mo-DCs in L. mexicana infected mice, potentially contributing to the inability of these mice to resolve their infection.
In addition to killing parasites at the site of infection through iNOS-dependent mechanisms, mo-DCs also migrate to dLNs where they orchestrate the developing immune response through antigen presentation and regulation of cytokine production [29]. Recently, we have also shown that L. major-activated DCs promote lymph node hypertrophy following infection [25] and the impaired lymph node expansion following L. mexicana infection [30] led us to investigate if a reduction in mo-DCs migration to the draining lymph node during L. mexicana infection limits the Th1 response and impairs lymph node expansion.
To evaluate the ability of DCs to migrate from the site of infection, C57BL/6 mice were infected in the ear with L. major or L. mexicana and two weeks post-infection the ears were FITC painted. After 48 hours we compared the FITC+ DCs (CD11c+ MHCIIhi) in the draining lymph node from naïve, L. major infected or L. mexicana infected mice. Notably, there were significantly more FITC+ DCs in L. major infected mice when compared to either naïve or L. mexicana infected mice. In contrast, there was no difference in the number of FITC+ DCs between naïve and L. mexicana infected mice (Fig. 4), indicating that mo-DCs migration to the dLN is compromised in L. mexicana infected mice.
We next wanted to determine if the microenvironment within L. mexicana lesions actively inhibited DC migration. Therefore, we injected the same number of CD45 disparate monocytes into L. major or L. mexicana lesions and evaluated their migration to the draining lymph node. Figure 5A shows an equivalent number of CD11b+ CD45.1+ cells in the ear of L. major or L. mexicana infected mice approximately 18 hours following monocyte transfer. Interestingly, the expression of Ly6C on the donor monocytes was lower in L. major infected mice as compared to L. mexicana infected mice (Fig. 5B). As downregulation of Ly6C is associated with activation of mo-DCs [24], [31], these data suggest that mo-DCs in L. mexicana infected mice do not differentiate as efficiently as mo-DCs from L. major infected mice. Even more strikingly, there is a dramatic increase in both the frequency and absolute number of transferred cells in the draining lymph node of L. major infected as compared to L. mexicana infected mice (Fig. 5C and D), indicating that L. mexicana infection does not increase mo-DC trafficking to dLNs. However, we cannot exclude the possibility that there may be a difference in retention in the dLN of L. major versus L. mexicana infected mice. Together, these data indicate that a lack of mo-DCs migration from the site of L. mexicana infection to the draining lymph node may prevent T cell priming and impair lymph node expansion, precluding the induction of a protective Th1 response and resulting in the development of chronic disease.
IL-10 has been described as having anti-inflammatory effects during infection by inhibiting cytokine production and antigen presentation [32], however, more recently it was shown that IL-10 also limits recruitment of CD11b+ Ly6C+ monocytes following T. brucei infection [33]. Since IL-10−/− mice infected with L. mexicana resolve their lesions [6], we wanted to investigate whether blocking interaction of IL-10 with its receptor would lead to increased monocyte recruitment. We infected C57BL/6 mice as before with L. mexicana and treated one group with α-IL-10R antibody. We evaluated monocyte recruitment to lesions on days 7 and 14 following infection and found that there was a greater percentage and number of monocytes recruited to L. mexicana lesions in mice treated with α-IL-10R (Fig. 6A). Similarly, the percentage and number of mo-DCs in the lesions of L. mexicana infected mice was also significantly increased when IL-10R was blocked (Fig. 6A). Moreover, there were increased levels of IFN-γ in the draining lymph nodes (Fig. 6B), as well as a greater percentage and number of iNOS-producing mo-DCs in the lesions of L. mexicana infected mice treated with α-IL-10R (Fig. 6C). These data suggest that IL-10 is a key factor contributing to the limited number of monocytes observed during L. mexicana infection since blocking the interaction of IL-10 with its receptor results in a dramatic increase in monocytes and mo-DCs in the lesion. Surprisingly, we did not see a difference in the parasite burden in treated and untreated L. mexicana infected mice at this early time point, in spite of the fact that we have previously shown that IL-10−/− mice eventually resolve their L. mexicana lesions [6]. Our assumption is that the effect on parasite burden is simply delayed and will develop later.
The previous experiment, where L. mexicana infected mice were treated with α-IL-10R antibody, suggests that the increase in mo-DCs in the lesion may result in the priming of an improved Th1 response. Here, we test whether there is a correlation between increased numbers of DCs in the lesion and a more robust Th1 response. We injected DCs into the ear at the time of infection with L. mexicana and we compared the Th1 response 14 days post-infection to L. mexicana infected mice receiving no DCs. As predicted, L. mexicana infected mice receiving DCs produced greater levels of IFN-γ (Fig. 7A), and had a greater percentage and number of iNOS-producing DCs (Fig. 7B). Moreover, the impaired lymph node expansion that occurs during infection with L. mexicana was overcome in mice that received DCs (Fig. 7C). Taken together, these data suggest that the limited Th1 response observed in L. mexicana infected mice can be overcome if a greater number of DCs can be established in the lesion.
Infection of C57BL/6 mice with either L. major or L. mexicana results in cutaneous lesions. However, while L. major-induced lesions heal, those induced by L. mexicana infection do not. The chronicity of L. mexicana infections is attributable to the limited Th1 response mounted by the host to the parasite [5], [6], [34]. Since the development of a Th1 response in leishmaniasis depends upon IL-12 production by DCs [35], [36], [37], [38], it was originally thought that L. mexicana fails to induce a healing response due to its inability to stimulate IL-12 production [11], [12], [14], [16]. However, the limited Th1 response in L. mexicana infected mice is not reversed by treatment with rIL-12 [17], suggesting that there is a more generalized impairment in DC function.
Differentiation of mo-DCs from inflammatory monocytes at the site of infection plays an essential role in immune protection in a number of infectious diseases [19], [21]. Monocytes are recruited to L. major infected skin [23], [24], [39] and mo-DCs are thought to be essential for the induction of the Th1 response in L. major infection [24], suggesting that limitations in monocyte recruitment and differentiation (or both) may lead to chronic disease following L. mexicana infection. In support, our current studies demonstrate that fewer monocytes are recruited during infection with L. mexicana when compared to L. major, and the consequent reduction in differentiated mo-DCs present in L. mexicana lesions likely compromises generation of a protective Th1 response. In addition, reduced monocyte recruitment and the observed decrease in iNOS expression will limit the killing capacity of these cells [39], presumably leading over time to increased parasite burden. Together, the limited Th1 response and enhanced parasite burden could promote the chronic exacerbated disease observed following L. mexicana infection.
The importance of monocyte recruitment in limiting the progression of infectious diseases has been most clearly demonstrated in CCR2 deficient (CCR2−/−) mice. CCR2 is a chemokine receptor expressed on inflammatory monocytes that mediates monocyte chemotaxis. In CCR2−/− mice, Ly6Chi monocytes accumulate in the bone marrow due to their inability to emigrate from this site. Limited recruitment of monocytes to the site of infection likely contributes to the enhanced susceptibility of CCR2−/− mice to Listeria infection [40]. In addition, following oral Toxoplasma gondii infection of CCR2−/− mice, monocytes fail to be recruited to the illeum, allowing for uncontrolled parasite growth. However, adoptive transfer of CCR2-expressing monocytes into T. gondii infected CCR2−/− mice protected them from this otherwise lethal infection [41]. CCR2−/− mice infected with L. major are also more susceptible to infection due to an attenuated Th1 response [42]. Interestingly, treatment of CCR2−/− mice with rIL-12 is able to reverse the susceptibility to L. major infection [43]. Since mo-DCs have been described as the major producers of IL-12 during L. major infection [24], these data support our hypothesis that compromised recruitment of monocytes to the lesion influences the development of a Th1 response in L. mexicana infected mice.
As IL-10 has been shown to limit the recruitment of CD11b+ Ly6C+ monocytes during infection with T. brucei [33] and we have previously shown that IL-10−/− mice infected with L. mexicana resolve their lesions [6], we hypothesized that monocyte recruitment following L. mexicana infection is impacted by IL-10 production at the lesion site. In fact, we showed that by blocking IL-10R, there was increased recruitment of CD11bhi Ly6C+ monocytes to L. mexicana infected lesions. Moreover, L. mexicana infected mice treated with α-IL-10R produced significantly more iNOS and IFN-γ than L. mexicana infected C57BL/6 mice. As during T. brucei infection [33], it is likely that production of IL-10 in L. mexicana-induced lesions may work on several levels. IL-10 could lead to decreased levels of CCL2, which would explain the limited recruitment of monocytes into the lesions. IL-10 is also capable of dampening Th1 responses, which would result in lower levels of iNOS and IFN-γ. Therefore, these data provide a mechanism as to why there is limited recruitment of monocytes to the lesion during infection with L. mexicana.
Finally, while DCs are clearly needed to prime T cells in the draining lymph node, they also promote lymph node hypertrophy. We previously demonstrated that lymph node hypertrophy is associated with the protective response to L. major infection [30] and have more recently revealed that L. major-activated DCs are responsible for lymph node expansion [25]. During infection with L. mexicana, lymph node hypertrophy is greatly reduced, potentially limiting the immune response [30]. Here we have used two methods to track migration of mo-DCs from the lesion to the draining lymph node; one method marked endogenous mo-DCs in the lesion and the other utilized injection of CD45 disparate monocytes directly into the lesion. While fewer endogenous mo-DCs from the L. mexicana lesion migrated to the dLN as compared to L. major, this may have been due to the relatively low numbers of monocytes initially present within the lesions of L. mexicana infected mice. To address this problem, we injected equal numbers of monocytes into L. major or L. mexicana lesions, and found that there was still a deficit in the migration of mo-DCs to the dLN from L. mexicana lesions. An inability of mo-DCs to migrate to the dLN could prevent both antigen-specific responses, as well as mo-DC-driven lymph node hypertrophy, providing a potential explanation for the reduction in lymph node size in L. mexicana infected mice. Interestingly, if DCs are injected into the ear at the same time of infection with L. mexicana, mice have significantly larger lymph nodes and are able to mount a more robust Th1 response compared to mice that did not receive DCs. These data clearly demonstrate that mo-DCs are important in initiating an appropriate immune response against Leishmania and that the limited recruitment of monocytes observed during L. mexicana infection could lead to the chronic nature of the disease.
In summary, we have demonstrated that 1) fewer monocytes are recruited to lesion during infection with L. mexicana as compared to L. major, 2) fewer iNOS producing mo-DCs are present in the lesions of L. mexicana infected mice 3) fewer mo-DCs migrate to the dLN node during L. mexicana infection, 4) blocking IL-10R leads to increased monocyte recruitment and a more robust Th1 response during L. mexicana infection, and 5) injection of DCs into the ear at the time of infection with L. mexicana also leads to increased levels of iNOS and IFN-γ. Together, these findings provide a mechanistic basis for the limited Th1 response, and lack of lymph node hypertrophy observed in L. mexicana infected mice and offer a better understanding of the important role that monocytes play during infection with Leishmania.
|
10.1371/journal.pntd.0002818 | Projected Future Distributions of Vectors of Trypanosoma cruzi in North America under Climate Change Scenarios | Chagas disease kills approximately 45 thousand people annually and affects 10 million people in Latin America and the southern United States. The parasite that causes the disease, Trypanosoma cruzi, can be transmitted by insects of the family Reduviidae, subfamily Triatominae. Any study that attempts to evaluate risk for Chagas disease must focus on the ecology and biogeography of these vectors. Expected distributional shifts of vector species due to climate change are likely to alter spatial patterns of risk of Chagas disease, presumably through northward expansion of high risk areas in North America.
We forecast the future (2050) distributions in North America of Triatoma gerstaeckeri and T. sanguisuga, two of the most common triatomine species and important vectors of Trypanosoma cruzi in the southern United States. Our aim was to analyze how climate change might affect the future shift of Chagas disease in North America using a maximum entropy algorithm to predict changes in suitable habitat based on vector occurrence points and predictive environmental variables. Projections based on three different general circulation models (CCCMA, CSIRO, and HADCM3) and two IPCC scenarios (A2 and B2) were analyzed. Twenty models were developed for each case and evaluated via cross-validation. The final model averages result from all twenty of these models. All models had AUC >0.90, which indicates that the models are robust. Our results predict a potential northern shift in the distribution of T. gerstaeckeri and a northern and southern distributional shift of T. sanguisuga from its current range due to climate change.
The results of this study provide baseline information for monitoring the northward shift of potential risk from Chagas disease in the face of climate change.
| Chagas disease kills thousands of people annually. Triatomine insects (family Reduviidae, sub-family Triatominae), can be potential vectors of the parasite (Trypanosoma cruzi) that causes the disease. There are often no symptoms until cardiac and digestive system dysfunction (possibly including heart failure) after 10 to 30 years of infection. Climate change can shift the distribution of triatomine insects, favoring the spread of the disease to non-original areas. We used distributional information on the most commonly found triatomine species and the most important vectors of Trypanosoma cruzi in South Texas and North Mexico (T. gerstaeckeri and T. sanguisuga), and explanatory climatic variables to forecast the potential distribution of the insects in the year 2050. We used two different scenarios of climate change and three different general circulation models. Our results showed that the triatomine species studied will likely shift their distribution northwards in the future. There is thus a need to monitor areas that are not currently endemic for Chagas disease but may potentially be affected in the future due to climate change.
| Climate change has been implicated in shifts of the geographic distribution of many species[1], enabling some taxa to increase their distributions into northern latitudes [1], [2]. Thus, changes in climate can potentially alter the spatial range of vector-borne diseases through shifts in geographical distributions of their vectors [3], [4], [5]. Despite some positive developments such as better access to clean drinking water, lower exposure to insect vectors, and higher-quality housing, the projected changes in climate over the next decades may exacerbate infectious disease incidence even in developed regions such as North America [6]. Habitat changes, alterations in water storage and irrigation habits, pollution, development of insecticide and drug resistance, globalization, tourism and travel are additional factors that may help to aggravate this threat [4].
The southern United States is highly vulnerable to outbreaks of vector-borne diseases due to many factors, including poor housing conditions, suboptimal drainage, lack of electricity in some areas, the presence of feral dogs, and human migration [7], [8], [9]. Moreover, that some southern states, such as Texas, share a legacy of neglected tropical diseases (NTDs [9]) with Mexico, increases the urgency of the development and deployment of active surveillance programs necessary for optimal management and control of vector-borne diseases including Chagas disease [7], [9] and leishmaniasis [5].
Chagas disease is a zoonosis caused by Trypanosoma cruzi, a flagellated protozoan parasite. Trypanosoma cruzi is transferred from mammalian reservoirs (e.g., Neotoma woodrats) to humans through a triatomine vector [7]. These vectors are insects from the family Reduviidae, sub-family Triatominae [7], [10]. Trypanosoma cruzi is most characteristically transmitted by infected feces of triatomines entering the human bloodstream. However, it can also be transmitted through blood transfusion, organ transplants and ingestion of infected food; congenital parasite transmission has also been demonstrated [7]. After contamination with the parasite, Chagas disease develops from an acute phase (period during which the parasites can be found easily in the blood) followed by an asymptomatic period of varying length; this stage is called the indeterminate phase. During the indeterminate phase, the parasites disappear from the blood. A chronic phase can be followed after 5 to 40 years, and ∼30% of infected people develop the disease [11], [12].
Chagas disease kills approximately 45,000 people annually [13] and affects 10 million people in several countries of Latin America [14]. In the United States around 300,000 individuals could be infected with T. cruzi, causing a considerable disease burden [15]. Several factors might influence the geographical distribution of Trypanosoma cruzi vectors and reservoirs (e.g., historical presence, the existence of barriers and dispersal capabilities), but anthropogenic factors play a fundamental role in the spread of the disease (e.g., through habitat changes, globalization, and travel [4]). The geographical distribution of Chagas disease has increased beyond regions of endemic occurrence during the last half-century and is now considered a worldwide problem [10].
Species distribution models (SDMs) based on machine-learning algorithms and Geographic Information Systems (GIS) platforms have been used to predict areas of potential distribution of Trypanosoma cruzi vectors [7], [16], [17], [18], [19]. These analyses typically show that climatic factors significantly influence the potential geographic distributions of vector (and reservoir) species. Additionally, temperature may have a particularly strong influence on the behavior of triatomine species [20], [21]. For instance, temperatures exceeding 30°C combined with low humidity,cause insects toincrease their feeding rate to avoid dehydration. In addition, in domestic life cycles, when indoor temperatures increase, the insects may develop shorter life cycles and higher population densities [20]. High temperatures can also speed up the development of T. cruzi in vectors [22].
In this paper, we forecast the future (2050) distribution in North America of Triatoma gerstaeckeri and T. sanguisuga, two of the most commonly found triatomine species and important vectors in the southern United States [7]. Triatoma gerstaeckeri is one of the most widely distributed Triatoma species in Texas [7], occurring mainly in the southern areas of the state. It is also found in New Mexico and in northeast Mexico [7]. Triatoma gerstaeckeri is more frequently found in economically poorly-developed areas; though it is naturally found in sylvan environments, it is able to disperse to human dwellings [23]. Triatoma sanguisuga can be found in several environments similar to T. gerstaeckeri, including domestic surroundings [24]. Triatoma sanguisuga has been found in several states across United States including Alabama, Arizona, Florida, Georgia, Kansas, Kentucky, Louisiana, Maryland, Mississippi, Missouri, New Jersey, New Mexico, North Carolina, Ohio, Oklahoma, Pennsylvania, South Carolina, Tennessee, Texas, and Virginia [24]. The species has also been found near the Canadian border in Illinois and Indiana [20]. We used geographic information (longitude/latitude distributional data) (Tables S1 and S2) and explanatory climatic variables (temperature, precipitation, etc., Table 1) to produce Species Distribution Models (SDMs) using a maximum entropy algorithm. Current SDMs were projected to 2050 using three different Global General Circulation models (the Canadian Centre for Climate Modelling and Analysis (CCCMA), the Commonwealth Scientific and Industrial Research Organization (CSIRO) and the Hadley Centre for Climate Change (HADCM3). We used two scenarios A2A and B2A from the International Panel on Climate Change [1]. Our aim was to analyze how climate change might affect the future spread of Chagas disease in North America.
For modeling purposes, geographic data (i.e., longitude and latitude) were gathered from data bases from museum collections, voluntary collectors, and through field work by members of our team in South Texas. For the original field work reported here, insects were collected either from public lands or donated by the owners of private lands. As a pilot study, field work was conducted in one sylvatic area, “La Sal del Rey”, Texas (26° 31′ N and 98° 03′ W), on 8 July 2011. We did not collect insects in domestic areas, we only included the La Sal del Rey locality in the model construction. To collect the insects, we used suspended dark ultraviolet light traps with a white background sheet and baited with carbon dioxide from dry ice. All geographic localities for both species are reported in Supplemental files (Tables S1 and S2). Following the methodology of Sarkar et al. [7], only post-1980 records with an estimated error of <1.0 km were used; these choices ensured compatibility between the resolution of the occurrence data and the spatial and temporal resolution of the environmental layers.
The study area includes the continental portions of Mexico and the United States and was delimited in the south by the 14°55′S line of latitude and to the north by the 49° 38′N line of latitude, continued by the lines −66° 97′E boundary and −124° 71′W. It was divided into 14 520 497 cells with an average area of 1.03 km2 (SD = 0.27). This ensured the enclosure of all points used in the analysis.
Present and projected future potential distributions for the target species were computed using presence records for the species (longitude/latitude) and with climatic parameters as exploratory variables, using a maximum entropy algorithm incorporated in the Maxent software package [11], [25]. Maxent predicts probability values (thresholds) from 0 (least suitable) to 1 (most suitable) of habitat suitability over the study area [11], [25]. We used Maxent Version 3.3.3k (http://www.cs.princeton.edu/~schapire/maxent/) with the default modeling parameters (convergence threshold = 105, maximum iterations = 500, regularization value β = auto) [26]. Climatic variables were selected from the 19 WorldClim variables [27] available at WorldClim database. Following Sarkar et al. [7], four climatic variables were eliminated from the analysis since these variables have presumed artifactual discontinuities for Texas (mean temperatures of the wettest quarter, driest quarter, warmest quarter, and coldest quarter; Table 1). These climatic variables have a resolution of approximately 1×1 km2 (more accurately, 30 arc-seconds). Twenty models were developed and evaluated via cross-validation per species. The final model presented is the average of the replicates. Model results were processed and visualized using ArcGIS 10.
For the future climate projections we used three GCMs: the Canadian Centre for Climate Modelling and Analysis (CCCMA), the Commonwealth Scientific and Industrial Research Organization (CSIRO) and the Hadley Centre for Climate Change (HADCM3). We used two scenarios of climate change, A2A and B2A, from the International Panel on Climate Change (IPCC 2007). Both scenarios assume a more heterogeneous world and are oriented toward regionalization. The A2A scenario assumes an increase in population, economic development, regionally oriented and per capita economic growth and technological change that is more fragmented than the scenario B2A. The focus of this scenario is more economic. On the other hand, the B2A scenario describes a world in which the emphasis is on local solutions to economic, social and environmental sustainability. It assumes a constant increase of population, but at a rate lower than A2A and intermediate levels of economic development as well. This scenario is oriented towards environmental protection and social equity.
We calculated the Area Under the Curve (AUC) of Receiver Operating Characteristic plots (ROC); [28] to evaluate the models by cross-validation of the 20 replicates using the training and test data as described above. Receiver Operating Characteristic is a threshold–independent measure that evaluates the sensitivity (probability that the model produces a positive result in a positive locality) versus the specificity (probability that the model produces a negative result in a negative locality) of a model when presented with new data. A ROC plot is obtained by plotting the sensitivity on the y–axis versus one minus specificity for all available decision thresholds on the x–axis. The theoretically perfect result is AUC = 1, whereas a test performing no better than random yields AUC = 0.5. The AUC was calculated internally by Maxent. The final AUC is the average AUC for all maps.
The averaged habitat suitability spatial distributions were converted into binary maps for further analysis using two thresholds: a “minimum training presence threshold” and a 0.5 habitat suitability threshold. A “minimum training presence threshold” is a threshold in which at least one known presence for the target species was found; therefore it guarantees that all presences are predicted as suitable [29]. Shifts on suitable habitat were calculated in km2. Percentage of change in suitable habitat comparing present and future projections was calculated using the formula ((future gain - future loss)*100)/present area.
A total of 84 unique geo-referenced localities, i.e., one locality per cell, were used to develop models of present and future suitable habitat for Triatoma gerstaeckeri and 24 for T. sanguisuga (Tables S1 and S2). Table 2 shows AUC values. For T. gertaeckeri the averages AUC were 0.9857 (SD = 0.0015) and 0.9738 (SD = 0.0279) for training and testing data, respectively; for T. sanguisuga the corresponding numbers were 0.9680 (SD = 0.0026) and 0.9323 (SD = 0.0982). Figures 1 and 2 show models of present and future distributions for both species.
Models of future distribution for the suitable habitat of T. gerstaeckeri show a shift to northern areas in USA, with projected suitable habitat in Michigan and in New York (Fig 1B-E). However, distributional shifts northward showed marked differences in habitat suitability between different climate change models and scenarios. For example, CCCMA-A2A and CCCMA-B2A models showed wide regions of unsuitable habitat between extant distributions and future northward shifts (Fig 1B–C). Conversely, CSIRO-A2A and CSIRO-ABA models showed contiguous suitable habitat between extant distribution and future northward shifts (Fig 1D–E). No shifts were observed between extant and future distributions with HADCM3_A2A and HADCM3_B2A models (Fig 1F–G)
Increases in future suitable habitat can be also observed for T. sanguisuga through the northeast and northwest of the USA. In all models, north-east shifts showed contiguous habitat suitability. This was not the case for future northwest shifts, where regions of unsuitable habitat were observed between extant and future shifts, except for the CCCMA-A2A model (Fig 2B). In just one model, CCCMA-A2A, the suitable habitat for this vector extended to Florida (Fig 2B). For this species, a shift of suitable habitat to South Texas (Lower Rio Grande Valley) and North Mexico in the State of Tamaulipas is observed using the HADCM3 model for both A2A and B2A (Fig. 2F–G) scenarios of the IPCC, while the CCCMA and CSIRO models (Fig. 2B–E) showing lower suitability habitat compared with the model of present distribution for this region (South Texas-northern Mexico) (Fig. 2A)
For both triatomine species, the variable that contributed the most to the distribution of the species was annual mean temperature (Figs. 1-H and 2-H). The minimum training presence threshold value for T. gerstaeckeri was 0.017 and for T. sanguisuga 0.068. For T. gerstaeckeri, the 0.5 threshold predicted loss on suitable habitat in 2050 compared with the minimum presence threshold for climatic change scenarios, A2A and B2A, and the three general circulation models (CCCMA, CSIRO, and HADCM3) (Table 3). For T. sanguisuga, both thresholds predicted an expansion of the suitable habitat by 2050 (Table 3).
For both species, Triatoma gerstaeckeri and T. sanguisuga, our SDMs predicted that there may be range shifts as result of climate change. Species distribution models for T. gerstaeckeri [30] and other triatomine species of North America [31] have been developed previously to this paper, but these models were constructed with a coarser spatial resolution (e.g. >1 km2). The influence of climatic change has been previously addressed by other authors with a consideration of three triatomine species (T. lecticularia, T. protacta, and T. sanguisuga) [30]. However, our analysis is the first attempt to model future distribution of suitable habitat for Triatoma gersteckeri and T. sanguisuga performed with the knowledge that all specimens were professionally identified and all locations for the species were explicitly reviewed for accuracy in their geography and method of recording (GPS coordinates with >1m error) and with a finer spatial resolution (1 km2). In addition, the cross validation and the low standard deviations in the model evaluations show no sampling biases attributed to the heterogeneity in the source of data and insect collection protocols. That is, models were neither strengthened nor weakened by the inclusion or exclusion of localities chosen based on this information.
Our results support [32] the conclusion that an increase in temperature is correlated with a potential increase of Chagas disease risk, defined as shifts in suitable habitat of T. gerstaeckeri and T. sanguisuga in the United States. Future distribution models showed marked differences for both triatomine species with important consequences for predicting Chagas disease risk. Overall, future distributions for T. gerstaeckeri showed wide discontinuous regions of suitable habitat between extant distributions and north-east shifts in the US. Thus, future north-east shifts of T. gerstaeckeri will depend heavily on natural abilities of this triatomine to disperse across wide regions of unsuitable habitat or to be transported by humans, except for CSIRO-A2A model showing more contiguous suitable habitat (Fig 1D). Two models, HADCM3_A2A, and HADCM3_B2A, did not predict northward shifts of this triatomine to Michigan and New York (Fig 1F–G).
Predicted north-east shifts of T. sanguisuga suggest contiguous suitable habitat, facilitating potential dispersal of this species to Michigan and New York (Fig 2A–G). Thus, T. sanguisuga is the target species most likely to be a threat of spreading Chagas disease in the north-eastern US, although this species is not considered an efficient vector for transmitting the parasite to humans [33]. Conversely, a different Chagas disease risk resulted for future shifts in northwest US. For both triatomine species, north-west shifts included wide areas of discontinuous suitable habitat between extant and future distributions (excluding T. sanguisuga in the CCCMA-A2A model). Thus, future shifts necessarily require high dispersal abilities for both triatomine species to represent a Chagas disease risk in north-west US. Other similar studies have identified important future shifts in north-east United States for other vector-borne diseases such as leishmaniasis [5]. Future distributional shifts of vector species can help to forecast expected number of human individuals potentially exposed to infectious diseases under climate change scenarios.
In addition to climate, several other factors not considered in this analysis could influence the distribution of the insects both under present circumstances and future ones. These factors can be biological (i.e., species interactions: competition, parasitism and trophic interactions), historical (e.g., barriers and speciation process), geographic (capabilities of dispersion, accessible regions for dispersal, evolutionary capacity of species' populations to adapt to new conditions), and/or anthropogenic[34], [35]. However, climatic variables (abiotic factors) are frequently used to estimate species' distributions [36], [37] since climate can limit distributions directly by affecting growth or survival (e.g., lower and upper lethal temperatures), and indirectly via interacting species (e.g., food sources, pathogens, competitors, or predators). Additionally, mechanism-based analysis have shown that temperature might have a strong influence on the behavior of triatomine species [20], [21], increasing their feeding rate when temperature increases and humidity is low, or by developing shorter life cycles and higher population densities [20]. High temperatures can also speed up the development of T. cruzi in vectors [22]. Therefore, as seen in our results, changes in temperature and precipitation based on the different climate change scenarios and general circulatory models can positively influence the spread of triatomine species to non-original distribution in North America.
Any study that attempts to evaluate the risk for Chagas disease should focus on the ecology and biogeography of triatomine vectors and reservoir species (e.g., woodrats), as well as the incidence of the parasite that causes the disease, Trypanosoma cruzi [7]. There is currently research to develop a vaccine for Chagas disease [9], but this is not available yet and drug treatments have limited efficacy. Chagas disease is controlled by using insecticides and improvements in housing, but such publicly organized programs do not exist in the United States, partly due to lack of information regarding human cases, vector-parasite incidence, and reservoirs of the disease. Studies that can provide baseline data for addressing these critical concerns should combine field work, molecular analysis (e.g., examining blood meals of triatomines) and ecological modeling techniques to assess the potential for Chagas disease at a fine-geographic scale (e.g., areas at most risk for Chagas disease; see [38]) are encouraged. Findings from that work can be used to advise health program managers in their efforts to control or prevent transmission of Chagas disease effectively and provide a cost-effective method of predicting locations of high transmission risk of this disease, particularly in light of the economic burden that Chagas disease might represent (similar or higher than other diseases such as rotavirus, cervical cancer, or Lyme disease [39]).
Although we acknowledge several important shortcomings discussed below, our study emphasizes one issue that has not been previously considered: the importance of climate change in the transmission of T. cruzi.
The transmission of T. cruzi includes several vectors and hosts in domestic, peri-domestic, and sylvatic cycles. Trypanosoma. cruzi has three infective forms capable of infecting its host, and currently 6 DTUs (discrete typing units) are recognized in the taxon. These DTUs establish with mammalian hosts peculiar interactions in distinct time-space scales. Thus, the transmission of T. cruzi is a complex system for its non-linearity, unpredictability and also for being multivariable.
Ideally, the potential distribution of most hosts should be included in the modeling exercises. We know relatively little about which mammal species are confirmed hosts of T. cruzi. To include simply a large list of mammals into the modeling approach without the certainty of being confirmed hosts of this parasite will add confusion into our understanding of this crucial biotic interaction. More studies are needed to produce a comprehensive list of confirmed hosts for T. cruzi as well as time-space scales for the operative interactions of hosts, vectors, and parasites. Novel modeling techniques developed to provide a predictive list of potential hosts for other emerging diseases, such as leishmaniasis [40], can be applied for T. cruzi.
Landscape and ecotypic scenarios under climate change are also needed to refine distribution shifts of species at finer spatial scales. This information should be associated with data on the salient features of landscape diversity, roles of extant members of regional mammalian faunas, local cultural, social and economic diversity, as well as the land use practices. This information will provide a more comprehensive understanding of the complexity in the transmission of T. cruzi.
|
10.1371/journal.pntd.0002961 | Insights into Embryo Defenses of the Invasive Apple Snail Pomacea canaliculata: Egg Mass Ingestion Affects Rat Intestine Morphology and Growth | The spread of the invasive snail Pomacea canaliculata is expanding the rat lungworm disease beyond its native range. Their toxic eggs have virtually no predators and unusual defenses including a neurotoxic lectin and a proteinase inhibitor, presumably advertised by a warning coloration. We explored the effect of egg perivitellin fluid (PVF) ingestion on the rat small intestine morphology and physiology.
Through a combination of biochemical, histochemical, histopathological, scanning electron microscopy, cell culture and feeding experiments, we analyzed intestinal morphology, growth rate, hemaglutinating activity, cytotoxicity and cell proliferation after oral administration of PVF to rats. PVF adversely affects small intestine metabolism and morphology and consequently the standard growth rate, presumably by lectin-like proteins, as suggested by PVF hemaglutinating activity and its cytotoxic effect on Caco-2 cell culture. Short-term effects of ingested PVF were studied in growing rats. PVF-supplemented diet induced the appearance of shorter and wider villi as well as fused villi. This was associated with changes in glycoconjugate expression, increased cell proliferation at crypt base, and hypertrophic mucosal growth. This resulted in a decreased absorptive surface after 3 days of treatment and a diminished rat growth rate that reverted to normal after the fourth day of treatment. Longer exposure to PVF induced a time-dependent lengthening of the small intestine while switching to a control diet restored intestine length and morphology after 4 days.
Ingestion of PVF rapidly limits the ability of potential predators to absorb nutrients by inducing large, reversible changes in intestinal morphology and growth rate. The occurrence of toxins that affect intestinal morphology and absorption is a strategy against predation not recognized among animals before. Remarkably, this defense is rather similar to the toxic effect of plant antipredator strategies. This defense mechanism may explain the near absence of predators of apple snail eggs.
| Filled with nutritious substances to nourish the embryos, eggs of most animals are often the targets of pathogens and predators. An exception are the eggs of Pomacea canaliculata –known as the apple snail– which have hardly any predators. This freshwater snail is a serious aquatic crop pest in several continents, listed among the 100 worst invasive species. It is the host of a roundworm responsible for the rat lungworm disease causing human eosinophilic meningitis. The spread of this emerging infectious disease has been associated with the expansion of apple snails. They lay eggs above water level in bright pink-reddish masses, presumably a warning coloration. Indeed, eggs have chemical defenses, including neurotoxic and antinutritive proteins. The authors found that the ingestion of egg extracts adversely affects rat small intestine inducing large, reversible changes in the intestinal wall that limits the ability to absorb egg nutrients causing a diminished growth rate. Apple snail eggs are the first animal known to deter predators by this mechanism, but remarkably this defense is rather similar to the toxic effect of plant seeds proteins. These overlapping egg defenses that predators have not managed to overcome yet may partially explain the reproductive success of P. canaliculata.
| The invasive apple snail Pomacea canaliculata (Lamarck, 1822) (Architaenioglossa, Ampullariidae) has become a serious aquatic crop pest in Asia and a vector of the rat lungworm Angiostrongylus cantonensis that causes human eosinophilic meningitis, a potentially fatal disease considered an emerging infectious disease. Unfortunately angiostrongyliasis (rat lungworm disease) continues to be reported in new regions beyond its native range which has been associated with the expansion of this snail [1]; [2]. P. canaliculata is the only freshwater snail listed among the 100 worst invasive species worldwide [3]. Their successful establishment in invaded areas may be related, among other factors, to their high fecundity, and the unusual characteristics of their eggs that increase the risk of the expansion of the disease. Females of P. canaliculata deposit hundreds of bright pink-reddish egg masses, each containing 30–300 eggs [4]; [5]. These egg clutches are remarkable in three respects: they are cemented outside the water, they are brightly colored and have virtually no predators, presumably because they have unusual defenses against predation [5]–[8]. Though filled with a perivitellin fluid (PVF) containing large amounts of carbohydrates and storage proteins (called perivitellins), these toxic eggs have no predators reported in their original South American range and only one in the newly colonized habitats in SE Asia: the fire ant Solenopsis geminata (Fabricius, 1804). The presence of these egg defenses [6]; [8]; [9] would explain the behavior of the snail kite Rostrhamus sociabilis (Vieillot, 1817) and Norway rat Rattus norvegicus (Berkenhout, 1769) that invariably discard the gland that synthesizes the egg defenses when predating on adult female P. canaliculata [10]–[14].
Work in the last two decades identified the perivitellins PcOvo and PcPV2 in the egg defenses against predation [6]; [8]; [13]; [15]–[17]. These are the most abundant perivitellins stored in large quantities in the PVF (57.0% and 7.5% of egg total protein for PcOvo and PcPV2, respectively) [15]. Both are resistant to proteolysis reaching the intestine in a biologically active conformation [6]; [8]. PcPV2 is a neurotoxic storage lectin with a strong lethal effect on selected neurons within the spinal cord of mice [9]; [18]. It is a novel combination of a tachylectin-like subunit with a membrane attack complex/perforin (MACPF)-like subunit, not reported in animals before [8]; PcOvo, on the other hand, is a storage carotenoprotein that provides the conspicuously reddish coloration of the clutches which presumably advertises to visual-hunting predators the presence of egg defenses (aposematic warning) [19]. In addition, PcOvo is a proteinase inhibitor limiting the ability of predators to digest egg nutrients. In fact, oral administration of purified PcOvo to rats significantly diminished rat growth rate presumably by a dual mechanism: the inhibition of trypsin activity (antidigestive role) and the resistance of the inhibitor to digestion by gut enzymes (antinutritive) [6]; [20]–[22]. A recent proteomic analysis of P. canaliculata PVF identified a small amount of over 50 other proteins, including two F-type lectins, many proteins involved in innate immunity in other mollusks and some with potential roles against insects and fungi [23].
As the epithelial cells along the digestive tract of animals are fully exposed to food contents, they are possible target sites for defense proteins. In this regard, plants have evolved a wide array of toxic dietary lectins that interact with the membrane glycoproteins of the luminal side of the gut of higher animals having an important role in plant defenses against predation [24]. There are, however, no reports in animals of such a defense mechanism [25].
With the aim to further understand the role of egg defenses of a host of the lungworm disease, in the present work we studied the effect of P. canaliculata PVF on the small intestine of rats. Through a combination of biochemical, histopathological, cell culture and feeding experiments, we provide evidence that oral administration of apple snail PVF adversely affects rat small intestine metabolism and morphology and consequently rat growth rate, presumably by proteins displaying lectin-like activity. This overall effect has not been found in other animals, but it is remarkably similar to that for plant seed lectins on the gastrointestinal tract of rats and other vertebrates.
All studies performed with animals were carried out in accordance with the Guide for the Care and Use of Laboratory Animals [26] and were approved by the “Comité Institucional de Cuidado y Uso de Animales de Experimentación” of the School of Medicine, UNLP (Assurance No. P08-01-2013).
Egg masses of P. canaliculata were collected either from females raised in our laboratory or taken from the wild in streams or ponds near La Plata city, Province of Buenos Aires, Argentina, between November and March of consecutive reproductive seasons. Only egg masses with embryos developed to no more than the morula stage were employed. Embryo development was checked microscopically in each egg mass as described elsewhere [16].
All experiments with rats were performed using male Wistar rats from the Animal Facility of the School of Medicine of the National University of La Plata (UNLP), Argentina. Rats came from a colony started with the strain WKAHlHok (Hokkaido University, Japan). Six-week-old animals weighing 180±2 g at the start of the experiments were housed in cages with 12 h day-night cycle, temperature of 22±1°C and relative humidity of 45–60%.
Fertilized eggs were repeatedly rinsed with ice cold 20 mM Tris-HCl, pH 6.8, containing a protease inhibitor cocktail (Sigma Chemicals, St. Louis) and homogenised in a Potter type homogeniser (Thomas Sci., Swedesvoro, NJ). Ratio of buffer: sample was kept 5∶1 v/w. The crude homogenates were then sonicated for 15 sec and centrifuged sequentially at 10,000×g for 30 min and at 100,000×g for 60 min. The pellet was discarded and the supernatant comprising the egg PVF was equilibrated in 50 mM phosphate buffer pH 7.4 using a centrifugal filter device of 50 kDa molecular weight cut off (Millipore Corporation, MA) to eliminate potentially interfering compounds. Total protein concentration of the PVF (13.3 g/L) was measured by the method of Lowry et al. [27].
Cylindrical tissue samples of the small intestine were post fixed in 10% neutral formaldehyde for 24 h at room temperature and then embedded in paraffin wax. Representative 5–7 µm sections were stained with haematoxylin and eosin for histological examination of general morphology. In addition, periodic acid Schiff (PAS) staining was performed to highlight carbohydrate distribution and goblet cells.
Fifty properly oriented villi and crypts from duodenum were selected at random from each animal and their length and width measured to calculate mucosal absorptive surface area following the method of Kisielinsky [29] whose results have no significant differences compared with the Harris method, widely used in rats [30]. The method considers a geometric mucosal unit of a cylindrical villous with rounded tip surrounded by cylindrical crypts. It assumes that the whole mucosa is an iteration of this unit, and the surface area can be calculated with mean values of structures that define the mucosal unit: villus length, villus width, and crypt width. Thus, the mucosal-to-serosal amplification ratio M was calculated considering these 3 variables, as follows:
Small intestine sections were assayed by immunohistochemistry (IHC) to evaluate cellular proliferation using a primary monoclonal mouse against the proliferating cellular nuclear antigen (PCNA) as a proliferation marker (Dako, Clon PC10). The antibody was diluted in 0.1% BSA in phosphate buffer and incubated overnight at 4°C. PCNA is a nuclear acid protein which functions as δ DNA polymerase helper. In the presence of PCNA and a replication C factor, δ DNA polymerase starts the synthesis of DNA and the progression of the cellular cycle.
Samples were incubated overnight at 4°C as mentioned above, and visualized using the LSAB kit (Dako Cytomation Lab, Carpinteria, USA) detection system which is based on a modified labeled avidin-biotin (LAB) technique in which a biotinylated secondary antibody forms a complex with peroxidase-conjugated streptavidin molecules. In short, after incubation with the appropriate primary antibody, a sequential 10 min incubation with an anti-mouse biotinylated antibody and peroxidase-labelled streptavidin is performed. Then staining is completed by incubation with 3,3′diaminobenzidine tetrahydrochloride (DAB) and H2O2. Positively stained cells showed a golden, dark-brown color. All sections were counterstained with Maeyer haematoxilyn before analysis. Primary antibody was replaced by normal mouse antiserum in control sections.
Small intestine sections were assayed with seven lectins (Table 1) (Lectin Biotinylated BK 1000 Kit, Vector Laboratories Inc., Carpinteria, CA, USA) namely: Con A (Concanavalia ensiformis), DBA (Dolichos biflorus), SBA (Glycine max), PNA (Arachis hypogaea), RCA-I (Ricinus communis-I), UEA-I (Ulex europaeus-I) and WGA (Triticum vulgaris) to reveal possible changes of the glycosylation pattern.
In short, paraffin sections were deparaffinized with xylene dehydrated with 100% alcohol twice, 10 min each, and then endogenous peroxidase activity was quenched by incubating 5 min with hydrogen peroxide in methanol 0.3–3.0%.
They were then hydrated, washed in phosphate-buffered saline, and incubated with biotinylated lectins overnight. Then sections were washed with PBS, followed by 10-min incubation with streptavidin-HRP (streptavidin conjugated to horseradish peroxidase in PBS containing stabilizing protein and anti-microbial agents (Vector Laboratories Inc., USA). Finally the bound lectins were visualized by incubation during 4–10 min with a buffered Tris-HCl solution (0.05 M, pH = 6.0) containing 0.02% 3,3′-diamino-benzidine tetrahydrochloride (DAB) and 0.05% H2O2 (DAB; Dako, Carpinteria, USA). Positively-stained cells were demonstrated by a dark golden brown coloration. The sections were counterstained with Maeyer haematoxilyn.
After 2-hour fixation in 2% (v/v) glutharaldehyde, samples were dehydrated in graded series of ethanol. Then ethanol was replaced by liquid carbon dioxide and samples were dried by critical point in a CP-30 (Balzers). Samples were gold metalized in a JEOL Fine Ion Sputter, JCF-1100. Observations and photomicrographs were obtained with a JEOL JSM 6360 LV SEM (Jeol Technics Ltd., Tokyo, Japan) at the Service of Electron Microscopy, Facultad de Ciencias Naturales y Museo, Universidad Nacional de La Plata, Argentina.
Horse, goat, rabbit and rat erythrocytes were obtained from the animal facilities at the University of La Plata (UNLP). Blood samples were obtained by venous puncture and collected in sterile Elsever's solution (100 mM glucose, 20 mM NaCl, and 30 mM sodium citrate, pH 7.2) (Sigma Chemicals, St. Louis). Prior to use, erythrocytes were washed by centrifugation at 1500 g for 10 min in TBS buffer (20 mM Tris, 150 mM NaCl, pH 7.4). This procedure was repeated several times until the supernatant remained clear. Hemagglutinating activity was assayed in microtiter U plates (Greiner Bio One, Germany) by incubating a two-fold serial dilution of PVF (6 mg/mL) in TBS with 2% erythrocyte suspension in TBS at 37°C for 2 h. Results were expressed as the inverse of the last dilution showing visible hemagglutinating activity by naked eye.
Human colorectal adenocarcinoma cells (Caco-2) were cultured in Dulbecco's modified Eagle's medium (DMEM) (4.5 g/liter D-glucose) supplemented with 10% newborn calf serum, penicillin (10 U/mL), streptomycin (10 µg/mL), amino acids and vitamins (Life Technologies-Invitrogen). Cells were cultured at 37°C in a humidified atmosphere of 5% CO2. Culture medium was replaced every 2 days and subcultured by trypsinization when 95% confluent. Passages 60 through 65 were used for the experiments. Prior to each experiment, the viability of the cells was determined by trypan blue exclusion. Viability of every cell preparation exceeded 90% as determined by counting the stained cells.
The cytotoxic effect of the PVF on Caco-2 cells was evaluated using the 3-(4,5-dimethythiazol-2-yl)-2,5-diphenyl tetrazolium bromide (MTT) assay [31]. Cells were seeded in 200 µL of culture medium on 48-well plates at densities that ensured approximately 90% confluency after 24 h. Once cell cultures reached the desired confluence, 50 µl/well of a serial dilution of PVF (6 mg/mL) in PBS were added and incubated at 37°C for 24 h. Control wells were prepared with 50 µL/well of PBS. After treatments, culture medium was removed and cells were incubated with fresh medium containing 0.5 g/L of MTT at 37°C for 1 h. Plates were then centrifuged, the supernatant discarded and the cells were washed three times with PBS. Finally the cell monolayers were extracted with 200 µL/well of DMSO and the absorbance of each well recorded at 540 nm with background substraction at 640 nm in a microplate reader Multimode Detector DTX-880 (Beckman Coulter, Inc., CA, USA). Cell viability was expressed as control percentage [31].
%Viability = (OD treated cells/OD control cells) ×100
Data collected from all experiments were analyzed individually by either t test (histology) or ANOVA (bioassays) using Instat v.3.05 (Graphpad Software Inc.). Where significant differences between samples occurred, a post-hoc Tukey's HSD test was performed to identify the differing means. Results were considered significant at the 5% level.
GenBank accession numbers for PcOvo subunits: JQ818215, JQ818216 and JQ818217; GenBank accession numbers for PcPV2 subunits: JX155861 and JX155862.
During the first 3 days of treatment with PVF, treated rats showed a significantly lower standard growth rate than the control ones (Fig. 1). This effect on growth rate disappeared after the fourth day of treatment and animals began to grow at the same rate as control groups. Daily food ingestion was similar in control and treated rats along the experimental period (results not shown).
Oral administration of PVF for 10 days increased the mean intestinal length of the rats though a tendency was already evident after a 4-day treatment (Fig. 2). Within four days of switching the 10-day treated animals to a control diet, the total length of the small intestine returned to control values (Fig. 2). Crypt dimensions and general morphology of intestine were virtually restored to normal.
At day 4, samples from control animals showed the characteristic tall, finger-like villi, whereas villi from treated animals showed significantly less height and were wider with some proliferation in the basal zone of the epithelia. In certain areas of the epithelium of treated animals, altered villi with a double, fused or “tongue” shape, displaying a bridge pattern were observed by SEM and light microscopy (Fig. 3). PAS staining was moderate on the glycocalyx of villi and crypt enterocytes while the mucin of goblet cells showed a strong stain in both control and treated samples. Mucose epithelia from treated animals showed an increased number of goblet cells (Fig. 4 A,B).
SEM analysis of control and treated animals confirmed the remarkable differences on the length and width of the villi of treated animals (Fig. 3). The increased amount of mucus on the mucosa in the treated animals was also observed with this technique, as well as areas displaying conical, dome-shaped mucosal elevations which seem to connect two villi in a bridge-epithelial pattern (Fig. 3).
PCNA labeling showed moderate immunostaining in the basal areas of the epithelium from controls, while a stronger staining was evident on the treated animals. (Fig. 4 C,D arrows).
Intestinal epithelial cells were studied using a set of 7 lectins, of which PNA and SBA produced the most remarkable results (Fig. 4). PNA was strongly positive on the supranuclear region of enterocytes of control animals, while in treated animals its binding was observed not only in this region (strong staining) but also in the whole enterocyte (light staining) (Fig. 4, E,F arrows). Besides, SBA lectin binding was strong on the glycocalyx of the apical zone of the enterocytes of treated rats in comparison to the moderate staining in control group, indicating SBA-binding glycans were more expressed on enterocytes exposed to snail egg PVF (Fig. 4 G,H arrows).
When the effect of PVF on rat small intestine absorptive surface was quantified on histological sections, a significant decrease of the 4-day treated animals was observed while if the ingestion is continued for 8 days, the absorptive surface reverted to normal (Table 2). When rats were exposed to PcOvo, the small intestine did not show significant changes in absorptive surface for up to 8 days (Table 2) and villi morphology was normal.
The egg PVF of P. canaliculata showed hemagglutinating activity against horse red blood cells up to a protein dilution of 0.15 mg/mL, indicating the presence of active lectins. Moreover, a moderate agglutinating activity against rabbit and rat red blood cells was also observed at 0.6 mg/mL of PVF protein concentration (Fig. 5).
The MTT assay showed that PVF displays cytotoxic activity on Caco-2 cell monolayers in a dose-dependent manner. A very significant reduction of cell viability to only 6.6±0.6% in PVF-treated monolayers as compared to control ones was observed at a PVF protein concentration of 0.6 mg/mL (Fig. 6).
The ingestion of apple snail PVF severely affects the gastrointestinal tract rapidly causing a decrease in growth rate. Shortly after feeding a diet containing PVF the rat intestinal morphology undergoes a dramatic change. This included shorter and wider villi and fusion of villi by epithelial bridging, which might be related to the ability of the epithelial cells to stretch in order to cover denuded areas [32]. The observed enlargement of both villous and crypt thickness in treated animals was associated with the presence of hyperplasic crypts and hypertrophic mucosal growth changes. The notable increase in enterocyte proliferation and the presence of immature enterocytes in the crypts suggest increased mitotic activity in treated animals. This in vivo effect was further supported by the analysis of PVF cytotoxicity toward differentiated intestinal cells which indicates the presence of toxins somehow damaging these enterocytes. This damage in turn would induce the proliferative response observed at the crypt. Thus, the ingestion of PVF seems to interfere with gut and systemic metabolism, inducing hyperplasia and hypertrophy of the small intestine and alterations in organ function. Despite this effect on the gut being well established for plant toxic lectins [33] it has not yet been reported for the ingestion of animal proteins.
The enterocyte proliferation was also associated with changes of the glycosylation pattern revealed by the differential binding of the plant lectins PNA and SBA. PNA binds to the supranuclear portion of enterocytes where Golgi apparatus is located. It has been reported in humans orally intoxicated with PNA that the perturbation of cell kinetics and the more rapid cell migration and turnover of enterocytes may be reflected as synthesis of incomplete nascent glycoproteins, and expressed by altered PNA binding patterns [34]. This is also a well-known effect caused in rats intoxicated by other plant lectins that, as metabolic signals, can radically alter the pattern of glycosylation of the gut epithelium and thus further amplify their potent physiological effects [33]; [35]. These similarities between the effects of PVF and plant lectins lead us to look for lectin hemagglutinating activity in the PVF, which was found positive for some mammalian erythrocytes. This agrees with the recent identification of two putative lectins in a proteomic study of P. canaliculata PVF [23]. As mentioned before, one of these lectins, PcPV2, is the second most abundant egg protein. A functional study performed after its ingestion by rats showed that PcPV2 has the ability to withstand protease digestion, displaying structural stability within the pH range of the gastrointestinal tract of rats. Moreover, this toxic lectin binds to the glycocalyx of rat enterocytes in vivo and to Caco-2 cells in culture [8]. This interaction is also in agreement with the high cytotoxic effect of the snail PVF on Caco-2 cells observed in this study.
These properties are concurrent with those of many plant lectins which are resistant to mammalian gastrointestinal digestion and their toxicity is mainly attributed to the binding to the glycan surface of the small intestine epithelial cells, which leads to interferences with the digestion and anatomical abnormalities [35]–[38].
Besides lectins, PVF also contains the proteinase inhibitor PcOvo. When the effect of a PVF-containing diet on rat growth rate (Fig. 1) is compared with that of a PcOvo-containing diet [6], a larger decrease of rat growth rate was observed with PVF, indicating there are more defensive compounds acting synergistically. In addition, a PVF-supplemented diet, unlike a PcOvo-supplemented one, diminished intestinal absorptive surface. A literature survey reveals no information on animals in this regard but again, a reduction of the absorptive surface area was reported after the administration of diets containing plant lectins to rats, causing malabsorption of nutrients [33]; [39]. As a whole, the decrease on rat growth rate and changes in intestine morphology and absorptive surface caused by PVF ingestion together with the reported ability of PcPV2 toxic lectin to bind intestinal cells were rather similar to the effect observed on rodents fed with diets containing plant lectins strongly suggesting that PVF lectins may be involved in the observed effect of snail toxic eggs on the gut of the rat.
If the ingestion of PVF is continued, the rat growth rate becomes indistinguishable from that in control rats indicating an adaptation overcoming the antinutritional effect. Changes in the length of small intestine are often related with the difficulty in digesting the food. Greater length increases the transit time, thus maximizing digestion [40]. The adaptation to the PVF involved a time-dependent increase of the small intestine length, clearly observed after 10-day treatment. Similar effects were also observed in rats 3 days after administering diets containing phytohemagluttinin (PHA) from red kidney beans and other plant lectins [41]; [42]. However, those studies have shown that PHA-treatment of rats resulted in pancreas growth [42]; [43]. No such effect was observed in the current study (results not shown). In addition, the increase in mucous secretion suggests another adaptation allowing the isolation and protection of the intestinal surface from the toxic proteins. The change in length was virtually reverted 4 days after the elimination of the toxins from the diet along with the recovery of the normal tissue morphology. It is worth recalling that the mucosa of the small intestine is lined with epithelium that has the shortest turnover rate of any tissue in the body and in about 3 days' time the entire surface is covered with new cells [44]. Although there is no report of other animal lectins causing this effect, a fast remodeling of intestine by reversible effects on anatomy and morphology are known in rats and pigs administered diets containing plant lectins [41]; [45].
Resting eggs are particularly vulnerable, since they are most attractive to potential parasites and predators and may lack an active defense system (because of their inactive metabolic state). Apple snails seem to have evolved passive defense systems to protect their developing embryos; the preferential accumulation of large quantities of lectins, and protease inhibitors is certainly indicative of that strategy. Moreover, it is believed that the main antinutrients responsible for reducing the nutritional value of many plant seeds are a combination of lectin and trypsin inhibitors [46]. Similarly, in apple snail eggs these two types of proteins may be also the main factors responsible for this effect. This further highlights the previously reported similarities between apple snail egg and plant seed embryo defenses [6]; [8]. In a broader view, the overall effect of P. canaliculata PVF on rats bears many similarities with the effect of plant dietary lectins not only against mammals but also birds, insects and nematodes, preventing these predators from digesting and incorporating nutrients from the tissues consumed [47]–[49]. Unlike plants, P. canaliculata advertises its defenses by a conspicuous coloration of the egg masses. Eggs indeed seem to have a large number of defensive proteins against predation, such as other protease inhibitors, chitinases, glycanases, lectins and antifungal proteins, as the analysis of the PVF proteome revealed [23]. Interestingly, all of these defensive proteins are also present in many plant seeds. It is possible that the combined effect of these defensive perivitellins -some targeting the digestive system while others aiming at other organs- may be an evolutionary adaptation. Although these defenses may not completely protect an egg from consumption, they may very well confer an advantage that increases its fitness helping to explain the virtual absence of egg predators. With more than 80,000 species, gastropods are the second largest class of animals after insects. It is therefore not surprising that a better understanding of gastropod egg biochemical defenses, little studied to date, is unveiling novel strategies not previously recognized among animals. In this regard, this study provides insights on the unique defenses against predators of a snail egg that are advertised by conspicuous coloration, and suggests that the acquisition of this protection may have conferred a survival advantage. This places apple snail eggs in the “winning side” of the predator-prey arms race.
In this work we demonstrate that the oral administration of apple snail egg PVF promotes alterations in rat growth rate and small intestine morphophysiology for short periods, whereas prolonged exposure to the toxic PVF induces an adaptation overcoming the antinutritonal effects. This defense has not been reported in animals before, but resembles those well established for plant seeds.
The severe effects of PVF on digestive tract adds another line of defense to the previously reported suite of biochemical defenses of apple snail eggs. This study helps to explain the near absence of predators and their successful establishment in invaded areas.
|
10.1371/journal.pntd.0002078 | Electrocardiographic Abnormalities in Trypanosoma cruzi Seropositive and Seronegative Former Blood Donors | Blood donor screening leads to large numbers of new diagnoses of Trypanosoma cruzi infection, with most donors in the asymptomatic chronic indeterminate form. Information on electrocardiogram (ECG) findings in infected blood donors is lacking and may help in counseling and recognizing those with more severe disease.
To assess the frequency of ECG abnormalities in T.cruzi seropositive relative to seronegative blood donors, and to recognize ECG abnormalities associated with left ventricular dysfunction.
The study retrospectively enrolled 499 seropositive blood donors in São Paulo and Montes Claros, Brazil, and 483 seronegative control donors matched by site, gender, age, and year of blood donation. All subjects underwent a health clinical evaluation, ECG, and echocardiogram (Echo). ECG and Echo were reviewed blindly by centralized reading centers. Left ventricular (LV) dysfunction was defined as LV ejection fraction (EF)<0.50%.
Right bundle branch block and left anterior fascicular block, isolated or in association, were more frequently found in seropositive cases (p<0.0001). Both QRS and QTc duration were associated with LVEF values (correlation coefficients −0.159,p<0.0003, and −0.142,p = 0.002) and showed a moderate accuracy in the detection of reduced LVEF (area under the ROC curve: 0.778 and 0.790, both p<0.0001). Several ECG abnormalities were more commonly found in seropositive donors with depressed LVEF, including rhythm disorders (frequent supraventricular ectopic beats, atrial fibrillation or flutter and pacemaker), intraventricular blocks (right bundle branch block and left anterior fascicular block) and ischemic abnormalities (possible old myocardial infarction and major and minor ST abnormalities). ECG was sensitive (92%) for recognition of seropositive donors with depressed LVEF and had a high negative predictive value (99%) for ruling out LV dysfunction.
ECG abnormalities are more frequent in seropositive than in seronegative blood donors. Several ECG abnormalities may help the recognition of seropositive cases with reduced LVEF who warrant careful follow-up and treatment.
| Chagas disease (ChD), caused by the protozoa Trypanosoma cruzi, is endemic in most Latin America countries and may be transmitted via blood transfusions. Cardiac disease is a major feature of chronically infected patients and may be lethal. Universal blood bank screening for ChD has been established in most Latin American countries, as well as in non-endemic countries with large immigrant populations, including the United States, Canada, Spain and Portugal. Blood donor screening leads to large numbers of new diagnoses of chronic T. cruzi infection. Counseling these individuals should address the recognition of those with more severe disease that deserve to be rigorously evaluated by experienced cardiologists and treated more promptly. The electrocardiogram is an important exam that can help in the recognition of cardiac disease and the evaluation of prognosis in ChD patients, but its role in blood donors has not been studied. The authors describe some electrocardiographic abnormalities that are typical of the infected blood donors, as well ECG abnormalities that help in the identification of those with severe cardiac involvement. These results may guide the evaluations of patients with incidentally detected T. cruzi infection from blood bank testing or public health screening.
| Chagas disease (ChD), caused by a flagellate protozoon, Trypanosoma cruzi (T. cruzi), is a major health problem in Latin America, where more than 8 million persons are infected [1], [2]. Chronic cardiopathy is the most important and severe manifestation of human Chagas disease, eventually affecting approximately 20% to 40% of those in the chronic phase of the disease [1], [2]. Due to migratory movements, ChD is now a world-wide challenge, since hundreds of thousands of chronically infected persons are now living not only in T.cruzi endemic countries but also in developed countries, mainly in Europe and the United States and Canada [3], [4]. Since one of the mechanisms of transmission of the disease is via blood transfusions, universal blood bank screening for ChD has been established in most endemic countries, as part of South American regional initiatives of elimination of transmission of the disease [5], [6]. Non-endemic countries with large immigrant populations including the United States, Canada, Spain and Portugal, have also begun to institute interventions to prevent blood-borne T. cruzi transmission [7]. Blood donor antibody screening results in large numbers of new diagnoses of chronic T. cruzi infection, most of them in the asymptomatic, indeterminate form of infection [8], [9]. Counseling these individuals should address the recognition of those with more severe disease that deserve to be rigorously evaluated by experienced cardiologists and treated more promptly.
Electrocardiogram (ECG), one of the most important tests in evaluation of ChD, is used to define the clinical stage of the disease with potential prognostic implications [10]. Most ECG studies of newly diagnosed ChD patients were performed decades ago, generally in patients identified by community or hospital based sampling; information on ECG findings in seropositive blood donors is lacking as is data relative to matched seronegative controls evaluated in parallel [11]–[17]. Additionally, most studies of ECG findings are not accompanied by systematic results that use core-lab reading of Echo and ECG results and codification by internationally accepted criteria, as the Minnesota Code for ECG findings [18].
As part of the National Heart, Lung and Blood Institute (NHLBI) Retrovirus Epidemiological Donor Study-II (REDS-II), we developed a study to evaluate the prevalence of ECG abnormalities in seropositive blood donors and to recognize typical ECG abnormalities associated with left ventricular dysfunction, the most important prognostic marker in ChD.
This study was conducted from July 2008 to October 2010, part of a retrospective cohort study in which T. cruzi seropositive blood donors identified by blood bank screening and well-matched seronegative donors. Using blood donation records from 1998–2002, we enrolled 500 T.cruzi seropositive subjects (250 from the city of São Paulo and 250 from the city of Montes Claros in the State of Minas Gerais) and 500 seronegative donors from the same time period, as previously described [19]. Recruited individuals underwent a health questionnaire, a medical evaluation, fasting blood sample collection for lipid profile, glucose analysis and NT-pro brain natriuretic peptide (NT-proBNP), an ECG and an Echo. Results of ECG and Echo were reviewed blindly by centralized reading centers. T cruzi antibody status was confirmed on plasma samples collected at the time of enrollment.
The study follows the Declaration of Helsinki of the Ethical Principles for Medical Research Involving Human Subjects, was approved by the Brazilian National Ethical Committee (CONEP # 1312/2006) and all subjects gave written informed consent.
In 1998–2002, Fundação Pro-Sangue in São Paulo performed T. cruzi antibody screening using three serological methods: ELISA, hemaglutination and immunofluorescence for Chagas [20]. For the purpose of this study, we included as seropositive cases donors who were positive by all three assays at the time of donation and on a separate sample obtained at the time of counseling, generally one to four months after the donation. In Montes Claros, Hemominas screened all donations with two serological assays: ELISA and hemaglutination. All donors reactive to both assays at the time of donation and counseling were considered eligible for this study. Follow-up samples from all enrolled subjects were retested for T. cruzi antibody under code at the REDS-II Central Laboratory in the US using an FDA-approved assay manufactured by Ortho Diagnostics [21]. All 483 control donors tested negative on their follow-up samples collected at the time of enrollment and clinical assessments for this study. Of the 499 case donors who originally screened as seropositive at the time donation in 1998–2002, serum collected in 2008–2010 from 498 tested repeat reactive, while the samples from one donor tested just below the assay cutoff consistent with slowly progressive seroreversion [21].
A face to face T. cruzi risk factor and health history questionnaire was administered by a trained nurse in each site. The questionnaire collected detailed information regarding demographics, cities of residence, physical activity, medical history, exposure to T. cruzi, previous ChD diagnoses, cardiac and GI symptoms, past medical history and medication history. All subjects received a physical examination by a non-blinded physician with recording of height, weight, blood pressure, heart rate and physical exam findings.
Resting 12-lead ECG were recorded using the same model of machine at both sites (General Electric MAC 1200 electrocardiograph; GE Healthcare, Waukesha, WI) using standardized procedures. All ECGs were processed blindly by the central ECG laboratory (Epidemiological Cardiology Research Center, Wake Forest University, Winston-Salem, NC), where they were visually inspected for technical errors and inadequate quality and processed with the 2001 version of the GE Marquette 12-SL program. ECGs were analyzed electronically, with manual over-reading by trained cardiologists to ensure quality control. ECGs were classified by Minnesota code criteria [18] using variables that were derived from the median complex of the Marquette measurement matrix. In this study, major and minor ECG abnormalities were defined as previously established [22], modified to include ECG abnormalities typical of Chagas cardiomyopathy with prognostic significance, as frequent supraventricular or ventricular premature beats [10]. Old myocardial infarction (MI) on ECG was defined by the presence of major Q wave abnormalities (MC 1.1.x or 1.2.x) or minor Q waves abnormalities with ST segment or T-wave abnormalities (1.3.x and [4.1.x, 4.2, 5.1, or 5.2]) [18].
Echocardiographic studies were performed using Sequoia 512 ultrasound machine (Acuson, Mountain View, CA, USA) at São Paulo site and GE Vivid3 (GE Healthcare, Waukesha, WI) at Montes Claros site. Cardiac measurements were performed according to the guidelines of the American Society of Echocardiography [23]. Studies were recorded in digital format and all measurements were performed on digital loops using a Digisonics offline analysis station (version 3.2 software, Digisonics, Houston, Tex) at the Cardiovascular Branch, Echocardiography Laboratory, National Heart, Lung, and Blood Institute, Bethesda, Maryland. LV ejection fraction (LVEF) was calculated based on modified form of Simpson's biplane method [23] and, if atrial fibrillation is present, LVEF is estimated by a visual method. For this study, LV systolic dysfunction was defined as LVEF<0.50.
Statistical analyses were conducted using SAS 9.2 and SPSS 18. Distributions of data were examined for normality by using Kolmogorov-Smirnov tests. Continuous variables were expressed as median [interquartile range (IQR)] and differences between seropositive and seronegative donors were compared using Wilcoxon-Mann-Whitney test, since these variables were not normally distributed. Categorical variables were summarized as counts and percentages and differences were compared using the Chi-square test or Fisher exact test. Association between quantitative ECG variables and LVEF was evaluated by Spearman correlation coefficient (rs). A p-value<0.05 was considered significant.
Receiver Operator Characteristic (ROC) curves were plotted in order to evaluate the accuracy of ECG measurements in detecting reduced LVEF and the area under the curve (AUC) was calculated. Sensibility, specificity and positive and negative predictive values of abnormal ECG, wide QRS duration (≥120 ms) and long QTc interval (>440 ms) were calculated with 95% confidence intervals. These cut points were selected since they are well-established in cardiology practice.
The study sample consisted of 499 seropositive and 488 seronegative donors. Seropositive had a higher proportion of non-white skin color and lower weight, body mass index, total cholesterol levels and pulse heart rate, as well as higher NT-proBNP values (table 1). LV systolic dysfunction was more commonly found in seropositive cases. The groups were comparable with regard to age, gender or other major cardiovascular risk factors (table 1).
Quantitative variables are shown in table 2. All ECG intervals were longer in seropositive donors, and both heart rate and HRV indexes had lower values. ECG abnormalities were similar in seropositive and seronegative donors, except for a higher frequency in seropositive subjects of right bundle branch block and left anterior fascicular block, isolated (16% vs. 2%, p<0.001 for RBBB and 15% vs. 2%, p<0.001 for LAFB) or in association (4% vs. 0, p<0.001); and a higher frequency of left ventricular hypertrophy in seronegative subjects (<1% vs. 1%, p = 0.004).
Seropositive cases showed more abnormal ECGs (51% vs. 32%, p<0.001) than seronegative donors due to a higher prevalence of major (26% vs. 9%) ECG abnormalities. They also presented a higher number of major ECG abnormalities per tracing when compared to seronegative donors: 20% vs. 7% had one and 6% vs. 2% had 2 or more ECG abnormalities, respectively (p<0.001).
LVEF was reduced in 36 out of 497 seropositive subjects with available data (prevalence of LV systolic dysfunction of 7.2%); in two patients LVEF measurement was not obtained due to technical reasons. Most seropositive subjects had LV systolic dysfunction considered mild (LVEF ranging between 40 and 49%, n = 17, 3.4%) or moderate (LVEF from 30–39%, n = 11, 2.2% of total); only 8 (1.6%) showed markedly depresses LVEF (<30%). Seropositive blood donors with and without LV dysfunction had comparable demographic and medical characteristics, although NT-proBNP levels were higher in those with LVEF below 50% (45 [25–76] vs351 [109–789], p<0.001).Seropositive donors with LV dysfunction showed longer PR, QRS and corrected QT intervals/durations (Figure 1), although both the heart rate and HRV indexes were not different between groups (data not shown). Both QRS and QTc duration were associated with LVEF values (rs: −0.159, p<0.0003, rs: −0.142, p: 0.002), and showed moderate accuracy in the detection of reduced LVEF (ROC AUC: 0.778 and 0.790, both p<0.0001, Figure 2). None of the other quantitative ECG variables showed significant correlation with measured LVEF.
Several ECG abnormalities were more commonly found in seropositive donors with depressed LVEF (table 3), including rhythm disorders (frequent supraventricular ectopic beats, atrial fibrillation or flutter and pacemaker), intraventricular blocks (right bundle branch block and left anterior fascicular block) and ischemic abnormalities (old myocardial infarction and major and minor ST abnormalities).
Almost all seropositive donors (33/36) with depressed LVEF showed at least one major or minor ECG abnormality. Only three subjects presented with normal ECG and abnormal LVEF: 0.46. 0.45 and 0.40, respectively; this last subject is a hypertensive patient with typical features of hypertensive cardiomyopathy in the Echo study who was not classified as having Chagas cardiomyopathy by the expert adjudication process employed in the parent study [19]. Seropositive donors with depressed LVEF had a higher prevalence of ECG abnormalities (69% vs. 23%, p<0.001) and a higher number of major ECG abnormalities per tracing when compared to seropositive donors with normal LVEF (39% vs. 19% with one and 31% vs. 4% with 2 or more abnormalities, p<0.001)
The diagnostic performance of selected ECG abnormalities was evaluated (table 4). An abnormal ECG (with minor or major abnormalities) is a sensitive marker for recognition of seropositive donors with depressed LVEF, with a negative predictive value of 99% (96–100%).
In this analysis of ECG findings from a large, controlled study focused on former blood donors with and without T. cruzi seropositivity, the frequency of major ECG abnormalities was higher in seropositive donors than in seronegative subjects; right bundle branch block and/or left anterior were more commonly found in seropositive than in well-matched seronegative donors. When considering only seropositive donors, those with LV dysfunction rarely presented with a normal ECG, having one or more major ECG abnormalities, including rhythm disturbances, intraventricular blocks and ischemic abnormalities. The presence of either major or minor ECG abnormalities is therefore a sensitive marker of the presence of LV dysfunction in ChD, and, equally important, the absence of ECG abnormalities has high negative predictive value. Consequently, we recommend that initial clinical evaluation of seropositive blood donors (and probably seropositive subjects identified through other population based screening programs) can be limited to ECG with more expensive echocardiography performed on patients with abnormal ECG and/or clinical findings suggesting of ChD.
Our findings suggest that seropositive blood donors have a similar profile to community ChD populations in terms of the prevalence and type of ECG abnormalities [12], [14], [15], [24], albeit with significantly lower rates and severity of abnormalities than in hospital and clinic-based series [11], [16], [16], [25]–[28]. However, the mean age of blood donors was higher than those studied in early community-based series and the prevalence of concurrent chronic conditions in blood donors was relatively high (24% hypertension, 5% diabetes), in contrast to some studies in which those conditions were excluded [16], [28]. Considering this higher frequency of ECG abnormalities in seronegative donors, right bundle branch block and/or left anterior fascicular block was the only finding more frequently found in seropositive subjects in relation to well-matched seronegative controls. This was also observed by Williams-Blandero et al. [14], in one of the most recent of those community-based studies, with a similar age profile. Indeed, since the interruption of vector-mediated transmission has been achieved in many Latin American countries [2], the age of T. cruzi infected subjects is increasing, and ChD is now a public health problem among older individuals in previously endemic regions [29].
LV systolic dysfunction, defined as reduced LVEF (<0.50), is a major marker of higher risk of death in ChD [30]. LV dysfunction, generally mild or moderate, occurred in a minority of cases in this sample (7.2%), reflecting the low risk profile of the seropositive donor population studied (the donors had to be clinically asymptomatic to give blood donations in 1998–2002, 8–12 years prior to the rigorous assessment including ECG and Echo that was the basis of the current analysis). In contrast, Ribeiro et al. [31] found a prevalence of more severe LV dysfunction (defined as LVEF≤0.40) of 9.1% and 14.5% in two samples from a Brazilian Outpatient Clinic that is a regional reference center for blood banks and primary care units, and Salles et al. [28] observed LV dysfunction in 109 out of 738 patients (14.8%) at another Brazilian Reference Outpatient Clinic.
In this study, several ECG abnormalities typical of ChD were predictive for LV dysfunction among seropositive donors. Right bundle branch block, frequently combined with left anterior fascicular block, is the most characteristic ECG abnormality in ChD [10] and is associated with higher risk of death in longitudinal studies [13], [32]. ChD patients with pacemakers have lower LVEF comparative to pacemaker patients without ChD [33]. In contrast to what we found in this study, frequent ventricular ectopic beats have been repeatedly related to LVEF depression in ChD [26], [34] and the observed association of supraventricular ectopic beats with LV dysfunction has not been reported before. Since frequent supraventricular ectopic beats can precede the development of atrial fibrillation, we hypothesize that higher left atrial pressure and volume secondary to LV systolic dysfunction may lead to frequent atrial ectopic beats and, after years, to atrial fibrillation. Atrial fibrillation is a late abnormality in the natural history of ChD, related to other ECG abnormalities, LV dysfunction and higher risk of death [12], [35]. Pathological q waves, ischemic ST-T abnormalities and abnormal T waves have also been reported to be markers of risk in Chagas cardiopathy [27], [36]. Prolonged QRS duration and QTc interval were both related to LV dysfunction [27], [37] and to worse prognosis [27], [38]. Since prolonged excitation time in ventricular conduction defects may induce secondary prolongation of the QT interval [39], the significance of prolonged QTc interval in seropositive donors should be interpreted with caution, considering that no correction for QRS duration was made in this study. Those with more than one ECG abnormality are at greater risk of having LV dysfunction, as reported in other studies [26], [40].
Heart rate variability indexes SDNN and RMSSD, calculated from standard 10-seconds ECG tracing, were reduced in seropositive donors but these findings were not correlated to the presence of LV systolic dysfunction. Both SDNN and RMSSD indexes from a 10-second ECG are markers of parasympathetic modulation of the sinus node [41]. In some studies in other settings [42]–[44], 10-s HRV indexes were predictive of the risk of death. Impairment of cardiac vagal modulation has been consistently reported in ChD [45]–[47] and occurs early in the evolution of the disease, preceding LV systolic dysfunction [48]. However, the association of vagal dysautonomia with LVEF and with prognosis in ChD is still controversial [46], [49].
Data from our study suggest that the ECG can be useful to guide the management of seropositive blood donors: an abnormal ECG is a sensitive marker of LV dysfunction, while a normal ECG carries a high negative predictive value. Indeed, a normal ECG is an established marker of excellent prognosis in medium-term follow-up of T. cruzi seropositive subjects [12], [13], [16]. The diagnostic performance of a single ECG interval measurement, such as a normal QRS duration or corrected QT interval, is not good, as previously reported [37].
The main strength of this analysis is that it is based on findings from a large controlled, rigorously conducted study, with central and blinded reading of both ECG and Echo results, and with classification using an internationally accepted ECG code, the Minnesota Code. This is, to the best of our knowledge, the only study with these features in a population of seropositive and matched seronegative blood donors. Because these blood donors were healthy at study baseline, the incidence of pathology will be lower than in prevalent patient cohorts. On the other hand, because patients with common chronic conditions, including hypertension and diabetes, were not excluded, it allows the generalization of findings to other seropositive blood donor populations in endemic and non-endemic countries. The main limitations are related to the cross-sectional status of the current analysis, with no information on the prognostic role of the observed abnormalities. We also lacked baseline measurements on the cohort to assure the absence of cardiac pathology, although all were healthy enough to donate blood. Moreover, blood donors from non-endemic countries may have different epidemiological profiles due to country of origin, socio-economic status and time since acquisition of infection prior to detection as seropositive donors, as well as possible differences in disease progression depending on the strain of T. cruzi.
In conclusion, we identified several ECG abnormalities that are predictive of LV dysfunction in ChD. Due to the study setting involving previously healthy seropositive donors who developed incident ChD, these study findings may be extrapolated to other low-risk populations. In particular, the results may guide the evaluations of patients with incidentally detected T. cruzi seropositivity from blood bank testing in endemic and increasingly in non-endemic countries, and from public health screening in endemic countries.
|
10.1371/journal.pcbi.1000365 | Divisive Gain Modulation with Dynamic Stimuli in Integrate-and-Fire Neurons | The modulation of the sensitivity, or gain, of neural responses to input is an important component of neural computation. It has been shown that divisive gain modulation of neural responses can result from a stochastic shunting from balanced (mixed excitation and inhibition) background activity. This gain control scheme was developed and explored with static inputs, where the membrane and spike train statistics were stationary in time. However, input statistics, such as the firing rates of pre-synaptic neurons, are often dynamic, varying on timescales comparable to typical membrane time constants. Using a population density approach for integrate-and-fire neurons with dynamic and temporally rich inputs, we find that the same fluctuation-induced divisive gain modulation is operative for dynamic inputs driving nonequilibrium responses. Moreover, the degree of divisive scaling of the dynamic response is quantitatively the same as the steady-state responses—thus, gain modulation via balanced conductance fluctuations generalizes in a straight-forward way to a dynamic setting.
| Many neural computations, including sensory and motor processing, require neurons to control their sensitivity (often termed ‘gain’) to stimuli. One common form of gain manipulation is divisive gain control, where the neural response to a specific stimulus is simply scaled by a constant. Most previous theoretical and experimental work on divisive gain control have assumed input statistics to be constant in time. However, realistic inputs can be highly time-varying, often with time-varying statistics, and divisive gain control remains to be extended to these cases. A widespread mechanism for divisive gain control for static inputs is through an increase in stimulus independent membrane fluctuations. We address the question of whether this divisive gain control scheme is indeed operative for time-varying inputs. Using simplified spiking neuron models, we employ accurate theoretical methods to estimate the dynamic neural response. We find that gain control via membrane fluctuations does indeed extend to the time-varying regime, and moreover, the degree of divisive scaling does not depend on the timescales of the driving input. This significantly increases the relevance of this form of divisive gain control for neural computations where input statistics change in time, as expected during normal sensory and motor behavior.
| Gain modulation (or gain control) is an adjustment of the input-output response of neurons, and is widely observed during neural processing [1]. Gaze direction sets the response gain in primary visual [2], posterior parietal cortex [3], and auditory brainstem [4]. In specific species, gain control mechanisms produce an invariance of receptive field properties [5] and orientation selectivity [6] to changes in overall stimulus contrast. Higher cognitive processes, such as attention, modulate the response gain of cells in primary visual cortex [7], as well as in V4 [8]. Finally, it has recently been shown that gain control schemes are needed to control behavior in invertebrates [9]. Despite the clear importance of gain control in a variety of neural computations, the biophysical mechanisms that support specific gain control mechanisms have been elusive [10]–[20].
Noise induced phenomena in nonlinear systems are a rich avenue of study [21], with recent interest on the impact of fluctuations on excitable systems, such as neurons [22]. Chance et al., Doiron et al., and Hô & Destexhe [11],[16],[23] all report that an increase in the fluctuations of background conductance inputs results in a decrease of the overall gain of the transfer between a static driving input and the mean output firing rate. In particular, if the balance between background excitation and inhibition is carefully controlled [11], then the gain control is purely divisive (or multiplicative). This means that an increase in conductance fluctuations acts to scale the transfer function over a large range of input by a simple constant multiplier (<1). Related work has further explored the impact of fluctuations on spike response [13]–[15],[20],[24],[25], with a the manipulation of the neural transfer function by background fluctuations being a central focus.
These studies address the gain control of a transfer function where the signal is either static or statistically stationary and the neural output is the time averaged firing rate. However, many neural coding tasks involve the processing of time-varying, high frequnecy stimuli. In these situations neural response are often transient, and a quasi-static approximation of input-output transfer fails to capture the actual spike response. For example, in the rodent vibrissa sensory [26], auditory [27]–[29], and electrosensory systems [30] stimuli and responses modulate on the order of a few milliseconds, i.e., on the order of, or even faster, than typical membrane time constants of neurons. Even in the visual system, where the relevant timescales of natural scenes are much slower, the response precision of thalamic neurons is at the millisecond level, and standard static transfer function analysis fails to capture neural response [31], yet contrast induced gain control persists [32]. In this study, we address the question of whether the fluctuation induced gain control mechanism explored for static transfer [11],[16],[23] can be operative for dynamic stimuli as well.
Any theoretical treatment of this problem requires 1) a framework accurately capturing the time varying spike response owing to time varying input statistics (e.g. temporally inhomogeneous input and output statistics), and 2) sufficient biophysical detail to incorporate conductance based synaptic inputs within spike creation. A useful tool for incorporating these two features into neuron models is the population density method [33]–[39]. In particular, Nykamp & Tranchina [35] have developed a simple one-dimensional population density method of conductance based leaky integrate-and-fire models (LIF). The one-dimensional version of the population density method allow us to easily study the firing rate responses to dynamic stimuli in the conductance based formalism of Chance et al. [11]. Minor differences in our proposed model and their dynamic clamp experiments to mimic conductance based inputs are presented in the discussion section. We first show that divisive gain modulation of the steady-state responses only hold for low output firing rates, in particular, where neurons are in the classical subthreshold regime. Second, when restricted to this regime we find the transient responses to dynamic stimuli, which can differ greatly from the quasi-static equilibrium response, also exhibit divisive gain modulation via fluctuation background conductances with the same scaling factor as computed in the static case. Thus, the divisive gain modulation proposed by Chance et al., Doiron et al., and Hô & Destexhe [11],[16],[23] generalizes to the dynamic situation in a very natural way.
We consider a leaky integrate-and-fire neuron (LIF) driven by a pre-synaptic population of excitatory (e) and inhibitory (i) cells. The neuron's voltage change is given by a random differential equation:(1)Dividing by the leakage conductance yields:(2)where is the membrane time constant, is the random size (for simplicity, chosen from the same distribution) of the excitatory/inhibitory synaptic event. The arrival times of both excitatory and inhibitory synaptic inputs are governed by modulated Poisson processes with mean rates , respectively. Throughout, is the resting membrane voltage, is the excitatory while is the inhibitory reversal potential. When the neuron's voltage crosses , a spike is recorded and the neuron enters a refractory period for a fixed time of , after which, its voltage is reset to . Consequently, the neuron's voltage varies between (). Throughout this paper, we will set , , , , , , and in accordance with estimates from experimental measures. We choose the average value of the random variables so that the neuron's voltage change (from ) is ±0.5 mV [40]. The random variable has a parabolic distribution function with finite support: for and 0 otherwise. It is convenient to define a new random variable , because upon receiving an excitatory synaptic event, the neuron's voltage will increase by (see Nykamp & Tranchina [35] for a derivation). The neuron's voltage will decrease in a similar way upon receiving an inhibitory event: . Thus, satisfy: and .
We decompose the pre-synaptic input into a time-inhomogeneous ‘driver’ term , and time-homogeneous background terms :(3)The background synaptic activity is balanced [11],[40], namely are chosen so that, in the absence of the driver input (), the random target voltage will have mean equal to the resting potential: . This will be true if:(4)
Of interest are the output threshold crossing times, and we estimate response statistics by combining the responses from trials where the arrival times are statistically independent across trials yet share the same generating intensities . The instantaneous firing rate of the neuron is defined as(5)where is the threshold crossing recorded during trial . Throughout the paper we are interested in the relationship between the driver and the response , and specifically how the balanced activity can modulate the relationship.
Figure 1 is a schematic diagram of the representative leaky integrate-and-fire neuron from the population receiving the combination of driver and balanced inputs. For the sake of exposition, we focus on three different intensities of balanced background inputs: to be 1100 s−1, 1400 s−1, and 1900 s−1 (with corresponding , 1361.24 s−1, and 1847.40 s−1, respectively), which we respectively label low (black), medium (red), and high (blue). Chance et al. [11] modified the background level by various rate factors and labeled the regimes 1X, 2X, etc., which is slightly different than our convention of low, medium, and high. However, the resulting steady-state input/output curves (Fig. 2) are similar to those in Chance et al. [11]. Also, our results below hold equally well for many other sets of balanced background activity. For a particular background intensity (low in this case) with random excitatory drive , the output is random (see spike raster plots). As the balanced background activity is increased, the variability in the voltage also increases (Fig. 1B).
A Monte Carlo simulation of Equation (2) would be computationally expensive to ensure an accurate result. In many studies only qualitative effects are reported, and thus quantitative accuracy is not at a premium. However, in our study the accuracy demands are large, as we will quantitatively compare the time dependent for various levels of background intensities. To overcome the errors inherent in finite data from Monte Carlo simulations we use population density methods [35], known to give very accurate estimates of (Fig. 1E) for the idealized neural models described by Equations (2)–(5).
In the population density method, neurons with similar biophysical properties are grouped together, and the evolution of a density function is considered. In brief, describes the voltage probability density over many statistically independent neurons. Integrating the density over a region in state space gives the probability that a neuron randomly chosen from the population will be in that region of state space:(6)
Let denote the probability current; a signed quantity with the convention that positive/negative is the probability per unit time of crossing from below/above. The evolution of is governed by a continuity equation [35]:(7)We separate the probability current into three distinct terms:The first term, represents the deterministic leak to rest in the absence of synaptic events. The second and third terms, , model the excitatory and inhibitory synaptic input driving the population. Mathematically we have:where is the complementary cumulative distribution function: , for . With the chosen distribution for , the functions above are (setting ):Finally, the instantaneous firing rate, , is the flux of probability current through from below:(8)The firing dynamics are implemented by an absorbing boundary condition at spike threshold, , and the source term in Equation (7), modeling membrane reset after a refractory delay. The population average firing rate by the population density method (see Eqs (6)–(8)) is a computationally efficient way of capturing , compared to computationally expensive Monte Carlo simulations.
Gain modulation is typically studied in the equilibrium regime [10],[11],[16], where the driver input is constant in time and the response denotes the equilibrium firing rate as a function of input. A divisive gain modulation for satisfies(9)where is the response of the population with some background activity ‘j’ (for our purposes j = 1 is low, j = 2 is medium, and j = 3 is high background activity) and is a scalar. To measure the divisive gain modulation of a response we fit the scaled response curves to (low) by minimizing the mean-squared error :(10)Finding the that minimizes is easily obtained by orthogonally projecting onto the subspace spanned by :(11)Because we want the largest possible range for divisive gain modulation, we steadily increase the maximum of until the error in (10) becomes significant (i.e., the scaled curves no longer lie on top of each other). Let that maximum value be .
The nonequilbirum response to a time-varying input is given by Equation (8). We extend divisive gain modulation to the nonequilbrium setting, with the analogous description:(12)(13)(14)The degree of divisive gain modulation in the nonequilibrium setting is determined by how well the time-varying responses scale with one another over a range of . Thus, the scalar is now computed from integrals over , compared to integrals over for the equilibrium case (11).
We compute the equilibrium input/output relationship, , using the same framework as Chance et al. [11] (section: Integrate-and-fire Neuron), in hopes of first reproducing their results. Using balanced excitation and inhibition to mimic background synaptic activity, we compute for different fixed values of the driving input (Fig. 2). To obtain computationally accurate results in reasonable time we employ population density methods (section: Population Density Approach) to estimate the response . Divisive gain modulation via increases of the background rates occurs in the low firing rate region (boxed region of Fig. 2A). In this regime the neuron response is dominated by input fluctuations rather than any intrinsic spike rhythmicity, thereby replicating the high variability observed in the spike responses in cortical networks [40]. Throughout we refer to this as the fluctuation driven regime. In the fluctuation dominated regime the responses can be scaled, in the sense of Equations (10)–(11), to nearly quantitatively match one another (Fig. 2B). This rescaling of the response by background fluctuations qualitatively matches the results presented in [11].
In contrast, for very high output firing rates divisive gain modulation does not occur. The responses are nearly linear with very similar slopes (Fig. 2A), showing only a background activity induced translation of the response (often termed subtractive gain modulation [10]). This region corresponds to a regime where input fluctuations have limited impact and the neuron response is predominately determined by the mean value of the input rates, and we refer to this as drift dominated regime. Fluctuation induced divisive gain control restricted to low firing rates is consistent with [16], where simulations of a large-scale compartmental neuron model were used. The insensitively of to input fluctuations at large has also been recorded in pyramidal cells and fast-spiking interneurons [41],[42]. However, the exact where gain manipulation changes from divisive to subtractive (as increases) is difficult to compute and is often model specific [14]. Indeed, there are neurons where the influence of noise persists at high firing rates, such as in layer 5 of rat medial prefrontal cortex [24], however, the biophysical mechanisms that support this effect are absent in the standard LIF model.
In summary, population density methods (section: Population Density Approach) can replicate fluctuation-induced divisive gain modulation of the equilibrium response at low firing rates, previously observed in: simple integrate-and-fire models [14]–[16], simulations of biophysical realistic cell models [16],[20], as well as simulated conductance experiments in vitro [11].
We study the influence of background fluctuations on the nonequlibrium response to a highly time-varying excitatory drive. We choose an input rate consisting of sums of sinusoids with various amplitudes, phases, and frequencies to mimic ‘rich’ time varying stimuli (for an example see Fig. 3A). This produces an inhomogeneous Poisson process driver input , resulting in a non-stationary in time stochastic driving current. The response inherits the non-staionarities of and is temporally modulated (Fig. 3B, black curve). Even though the stimulus results in a rather narrow range of response firing rates , it has adequately rich temporal modulation to produce output firing rates that are different than the quasi-static response (Fig. 3B, brown curve), obtained by setting .
The main result of our study is that fluctuation-induced divisive gain modulation is robust for low to moderate output firing rates in response to dynamic stimuli, despite the complicated dynamics of the leaky integrate-and-fire neuron in the nonequilibrium regime (i.e ). To demonstrate we compute the nonequilibrium responses (, , and ), for the three levels of background activity used for the equilibrium case (low, medium, and high). For larger background activity the overall response is reduced, observed here since for all (Fig. 3C). We compute the dynamical analogue of , (see Equations (11) and (14)) and the scaled response , which quantitatively matches the base response (Fig. 3D). This mimics the results for the equilibrium case (compare Figs. 2B and 3C and 3D). It is, a priori, unexpected that the dynamic response (with timescale and refractory period ) should scale in the same way as the equilibrium response .
Previously, Holt & Koch [10] showed that an increase in membrane shunting without a change in input fluctuation causes a translational shift, rather than division of the equilibrium response curves, which was also verified by Chance et al. [11]. To verify that a pure shunting change cannot result in divisive gain modulation of the nonequilibrium responses (Fig. 3D), we fix the background fluctuation level and driver , but increase the deterministic leakage conductance (Equation 1) to mimic different background synaptic activity (conductance) levels. Equivalently, is replaced with a scaled version: , which has the same mean conductance in the absence of driving input as (). In simulations where the background activity is set by deterministic leak rather than by synaptic conductance fluctuations, the neurons had negligible firing rates because they were unlikely to fire by random chance and did not scale in a divisive manner (not shown). For exposition, we set all of the background fluctuation levels to that of low and vary , and hence , to mimic deterministic effects of changing background activity so that there are less fluctuations, but still some amount to induce background firing. The unscaled responses (Fig. 4A) were scaled via a least squares fit (Equation (14)). Not surprisingly, the responses do not scale in a divisive manner (Fig. 4B). Thus, divisive gain modulation in the nonequilibrium regime critically depends on changing the background fluctuation levels. We remark that Chance et al. were in the high conductance state when they verified this whereas our regime has less overall conductance.
When the dynamic stimuli are increased so that resulting output firing rates are larger, the neurons no longer exhibit divisive gain modulation. Increasing the overall intensity of the driving input (compare Fig. 5A with Fig. 3A) yields firing rates that are an order of magnitude larger (compare Fig. 5B with Fig. 3C). Increasing the overall background activity reduces the overall response magnitude (Fig. 5B), similar to what is observed in both the equilibrium and nonequlibrium regimes. However, when the response curves are scaled by computed for the low rate case pure divisive gain modulation is not observed for the high rate response. There is no trivial (a time independent ) or natural way to scale the output firing rate curves so that they lie on top of each other. Since divisive gain modulation does not hold in the equilibrium setting for high output firing rates (drift dominated regime), one would expect that it does not hold in the nonequilibrium state. However, both the equilibrium and nonequilibrium states are quite different and we present the failure of fluctuation induced division for the sake of completeness. It is interesting to note that for periods of time when the output firing rates are low, divisive gain modulation appears evident, likely owing to a transient excursion into the fluctuation driven regime.
In our model, when the driving input rate is low, the population of neurons rarely fire action potentials (i.e., low spontaneous activity). The firing rates in our simulations in this state range from nearly 0 to 3 s−1, depending on the background level of activity. Although extracellular recordings in the cortex suggest the neurons can fire spontaneously at rates larger than 2 s−1 [43], such experiments are usually biased towards active neurons. Extracellular recordings by [44] that were unbiased towards responsive neurons suggest that many neurons have low spontaneous firing rates and that only a small fraction of neurons respond ‘well’ to stimuli in unanesthetized animals; this fact was also discussed in [43]. Moreover, calcium imaging experiments of awake and anesthetized rats in layer 2/3 of the cortex show that many neurons have resting firing rates less than 1 s−1 [45]. The actual firing rate of neurons in the resting state is a contentious issue, but our results hold for many parameter regimes (see Fig. 6).
Divisive gain modulation with dynamic stimuli is robust in the subthreshold regime (Fig. 6). To illustrate this point, the response with low background level to a time-varying driver input and the response to the same driver input with a second level of background activity are computed. We plot the logarithm of the area (or error , see Equation (13)) between the time-dependent response scaled by the equilibrium scale factor and the response:A logarthmic scale was used (Fig. 6) to better highlight the variety of values. The driver input is scaled as follows:where is the driver input used previously (see Fig 3A), is the scaling parameter, and is a parameter that insures in positive and not too small. Notice corresponds to the driver input in Fig. 3A, and with corresponds to the driver input in Fig. 5A. The vertical dot-dashed line in black corresponds to 0 error because it is the reference background curve for a given . The two points marked by stars (*) in Fig. 6A at and , 1900 s−1 correspond to the difference in area between the curves in Fig. 3D, which is quite small. In fact, for a large region of parameter space, divisive gain modulation holds (any patch that is orange to blue in Fig. 6A). The two black circles (•) in Fig. 6A at and , 1900 s−1 correspond to the difference in area between the curves in Fig. 5B. With larger values, the neurons are in the drift dominated regime, and divisive gain modulation no longer holds, as expected (red regions in Fig. 6A).
The average (unscaled) time-dependent response of the neurons with the same parameters and driver inputs as Fig. 6A are plotted in Fig. 6B on a logarithmic scale:The three stars (*) at are the average firing rates of the unscaled responses in Fig. 3C, and the three black circles (•) at are the average firing rates of the unscaled responses in Fig. 5A (bottom panel). The average firing rate gives a qualitative idea of how large or small the response is as are varied. For example, the average firing rate at the low level is about 2 s−1 but the firing rate response can be quite low and as high as 15 s−1 (see Fig. 3B). Thus, divisive gain modulation holds for many parameters in a variety of subthreshold regimes.
A gain control scheme will be effective in unpredictable environments if it is quantitatively insensitive to the timescales of the input, or in other words the degree to which the response is scaled should not depend on the spectral content of the signal. For fluctuation induced gain control we then require that the scaling factor associated with a specific background state would need to be independent of the temporal frequencies in the driver input . To test this we compare the optimal scaling factor between two dynamical responses (each with a distinct balanced background state) where the synaptic driving input is:Notice the specified varies from 0 to , so that synaptic input rates are non-negative. Let us denote by for two given background rates driven by sinusoidal input with frequency in Hz (here is not the conventional radian frequency). When is in a low range the differences between and are negligible over a wide range of (Fig. 7). This result is robust for a range of background states (Fig. 7A–D). The quantitative match between the divisive scaling of equilibrium and nonequlibrium responses extends to more complicated temporal modulations of the driving input (Fig. 3A). Specifically, we find that for the results shown previously ( and in Figs. 2 and 3). Thus, fluctuation induced gain control is quantitatively insensitive to the timescales of the driving input.
To better describe the mechanism underlying fluctuation-induced divisive gain control in the nonequlibrium, we focus on a weak time modulation of the input drive and compute the linear frequency response [34],[46],[47]. The frequency response function gives the first order temporal modulation of output firing rate assuming the synaptic driving input consists of a large constant component and a small time-varying component:Here we have set to be some fixed driver synaptic input rate, making the overall time independent excitatory input , while the inhibitory input is still . In total we then have the equilibrium state defined by the triplet , and the time dependent component of the driver simply . Assuming we approximate:(15)The time modulation of the response is characterized by , indicating how large or small the first order response is to time-varying input of frequency and amplitude . is a complex number with a modulus and phase : .
Our earlier results (Fig. 3) show that for the same driver input and different background inputs that for some scaling factor . However, we know that in limit the equilibrium response also satisfies (Fig. 2). Combining these two results, and neglecting the terms in Equation (15), predicts that where is a background state. Satisfyingly, when neurons are in the fluctuation-dominated regime does indeed multiplicatively scale in the same quantitative manner for different levels of balanced background synaptic input scale (Fig. 8A and 8B). The phase component is the same for all and tested (insert Fig. 8A) and hence can not change the response for different . Thus from the quantitative scaling match of both and we expect fluctuation-induced divisive gain control to extend to weak inputs.
We remark that the near exact scaling of in the high frequency range () is important; if the multiplicative scaling was only true in the flat region of () then fluctuation-induced divisive gain control would fail in the nonequlibrium, i.e. when the quasi-static approximation fails. To be more specific, , if we neglect the fluctuations given by the driver Poisson process. Thus the scaling of for is completely explained by the scaling of the gain of . However, multiplicative scaling for ensures that fluctuation induced gain control will extend into the nonequilibrium regime, even though the quasi-static approximation fails.
When the neurons are in the drift dominated regime, the frequency responses does not scale in a multiplicative manner because there are resonant peaks at integer multiples of the steady-state firing rate [34],[48],[49], and these resonant peaks occur at different frequencies for various balanced background synaptic activity (see Fig. 5B). Thus, divisive gain modulation with dynamic stimuli cannot possibly occur. The frequency responses in the drift dominated regime are scalar multiples of each other up to 10 Hz, where there appears to be divisive gain modulation with the same equilibrium scaling factors (see Fig. 5B). As explained in the previous paragraph, frequency response for is equal to the frequency response for . However, the multiplicative scaling breaks down for the same range where the quasi-static approximation breaks down, meaning that for drift dominated responses any fluctuation induced gain control in the equilibrium regime will not transfer to the nonequilibrium response.
Chance et al. [11] described a mechanism by which divisive gain modulation results from a balanced, fluctuating background synaptic activity which both shunts and linearizes the membrane to spike transfer. The response is a scaled version of a baseline condition , and the dividing factor is independent of the driver intensity . However, many stimuli induce input and output statistics which vary on the timescale of neural integration [26],[27],[30],[31]. Extending fluctuation induced gain control to accurately divide the response to these inputs is not automatic, as the spike-reset and refractory dynamics significantly shape the response in the nonequilibrium regime to be significantly different than the quasi-static approximation. However, our results show that the fluctuation induced gain control does extend to the nonequilibrium regime, increasing the potential utility of this form of gain control in neural processing. Furthermore, establishing the independence of the scaling term from the timescale of the driver greatly simplifies the circuitry required to implement gain control. In its simplest scenario, the gain of the response is set by the background rates which maintain their scaling effect despite processing unpredictable environments where inputs statistics can vary dramatically.
The analysis of the time dependent response for weak signals showed how a scaling of is inherited from an equivalent scaling of and the response function by fluctuating background conductances. The response to an input of arbitrary strength and spectrum can be written using the Volterra expansion [50]:where is the inverse Fourier transform of the response function described in Equation (15). Fluctuation induced gain control extends well into the nonlinear regime, evidenced by the empirical agreement in regimes where varies significantly about (Fig. 3D). In this case, the influence of the higher order terms in the Volterra expansion are likely important. We conjecture that, within the fluctuation dominated regime, each response function , meaning that the multiplicative scaling extends, response function-by-response function, analogously into the nonlinear regime. This scenario is opposed to the one where each term exhibits scaling with distinct terms, yet the sum of terms somehow scales with , forcing agreement with our results where has large temporal variance (Fig. 3D). In principle computing is quite difficult, however, if this scaling is correct then the influence of the stochastic background on , in the fluctuation driven regime, becomes straightforward.
Divisive gain control is a central tool in many neural computations [1], yet robust biophysical mechanisms that produce gain control are elusive [10]–[13],[16]. Our work gives further evidence that using background fluctuations as a mechanism to scale responses is a surprisingly stable mechanism operable for a variety of input statistics. Fluctuation induced effects on the equilibrium state transfer different from divisive gain control have been reported [24],[25]. Notably, [24] have shown that the firing response of pyramidal cells in layer 5 is sensitive to fluctuations at high rates, where the mean current no longer determines the spike rate. The mechanisms responsible are not present in the standard LIF model, however, modifications could possibly be made to model these effects and a population density equation could, in principle, be derived. These models would require more state variables and/or equations and in general are not computationally tractable without some reduction or approximation. Sophisticated methods for other neuron models have been developed [51]–[53]. Extending gain control to nonequilibrium responses to a larger class of models is currently an open avenue of research.
The LIF model we have used is an approximation to the dynamic clamp experiments of Chance et al. [11]. One difference is that our model does not have temporal correlations in the synaptic conductances, while there are temporal correlations in the experiments even though Chance et al. average over time (and trials) to obtain the firing rate. Also, we are using a simple yet biophysical spiking neuron model, where the level of background activity determines the variance of the background voltage (see Fig. 1B), consistent with the observations that membrane potential variability changes with the internal brain state [54]. In Chance et al. the variance of background voltage was the same for all background fluctuation levels, ensuring that the variability of the output firing rate is constant. Despite these differences, our results suggest that fluctuation induced divisive gain modulation is viable with dynamic stimuli.
The population density equations (6)–(8) that characterize the LIF model contain a partial differential-integral equation that is difficult to analyze. Our model is more general than white noise models that have an advection/diffusion density equation (e.g, Fokker-Planck equation) because it allows for large voltage changes upon receiving synaptic input events. However, the simulations shown in this paper are in the regime where the diffusion approximation is good. If the voltage change upon receiving synaptic events (excitatory or inhibitory) is assumed to be small, a good diffusion approximation of (6)–(8) is obtained by replacing with in the integrals in the probability current terms and (see Text S1 and Figure S1). With large voltage changes, a similar approximation can be obtained by a re-scaling of the equation around the (deterministic) mean. However, a direct comparison to the Fokker-Planck equation with white noise conductances still must be done numerically because of the conductance-based input (Text S1 outlines the Fokker-Planck approximation to the full density equation and Figure S1 shows the magnitude of the advection/diffusion coefficients). Moreover, the analytic formulas obtained with advection/diffusion equations are often computed numerically and usually assume at least a quasi-static approximation. With Poisson current injection however, a closed form Fokker-Planck approximation is obtained with drift and diffusion coefficients that can be written exactly in terms of voltage, input rates, and the statistics of . An analytical explanation of the robust scaling of the firing rate responses that is observed remains elusive yet is conceivable because of the many analytical results obtained for density equations of a variety of neuron models [51],[52]. However, we remark that even in the equilibrium regime an analytic explanation of divisive gain modulation via conductance fluctuations is difficult to obtain [14].
|
10.1371/journal.pntd.0002335 | Identification of Immunogenic Salmonella enterica Serotype Typhi Antigens Expressed in Chronic Biliary Carriers of S. Typhi in Kathmandu, Nepal | Salmonella enterica serotype Typhi can colonize and persist in the biliary tract of infected individuals, resulting in a state of asymptomatic chronic carriage. Chronic carriers may act as persistent reservoirs of infection within a community and may introduce infection to susceptible individuals and new communities. Little is known about the interaction between the host and pathogen in the biliary tract of chronic carriers, and there is currently no reliable diagnostic assay to identify asymptomatic S. Typhi carriage.
To study host-pathogen interactions in the biliary tract during S. Typhi carriage, we applied an immunoscreening technique called in vivo-induced antigen technology (IVIAT), to identify potential biomarkers unique to carriers. IVIAT identifies humorally immunogenic bacterial antigens expressed uniquely in the in vivo environment, and we hypothesized that S. Typhi surviving in the biliary tract of humans may express a distinct antigenic profile. Thirteen S. Typhi antigens that were immunoreactive in carriers, but not in healthy individuals from a typhoid endemic area, were identified. The identified antigens included a number of putative membrane proteins, lipoproteins, and hemolysin-related proteins. YncE (STY1479), an uncharacterized protein with an ATP-binding motif, gave prominent responses in our screen. The response to YncE in patients whose biliary tract contained S. Typhi was compared to responses in patients whose biliary tract did not contain S. Typhi, patients with acute typhoid fever, and healthy controls residing in a typhoid endemic area. Seven of 10 (70%) chronic carriers, 0 of 8 bile culture-negative controls (0%), 0 of 8 healthy Bangladeshis (0%), and 1 of 8 (12.5%) Bangladeshis with acute typhoid fever had detectable anti-YncE IgG in blood. IgA responses were also present.
Further evaluation of YncE and other antigens identified by IVIAT could lead to the development of improved diagnostic assays to identify asymptomatic S. Typhi carriers.
| Salmonella enterica serotype Typhi is the cause of typhoid fever and infects over 21 million individuals and causes 200,000 deaths each year. With adequate treatment, most patients recover from their acute stage of illness and clear infection. However, a small percentage of S. Typhi infected individuals develop a chronic but asymptomatic infection in the biliary tract that can persist for decades. Since S. Typhi is a human-restricted pathogen, chronic carriers may act as reservoirs of infection. Correctly identifying and treating asymptomatic chronic carriers could be critical for ultimate control of typhoid fever. Using an immunoscreening technique called in vivo-induced antigen technology (IVIAT), we have identified potential biomarkers unique to S. Typhi chronic carriers. Further evaluation of these antigens could lead to the development of improved diagnostic assays to detect asymptomatic S. Typhi carriers in typhoid endemic zones, and to an improved understanding of the pathogenesis of S. Typhi in the chronic carrier state.
| Salmonella enterica serovars Typhi (S. Typhi) and Paratyphi A (S. Paratyphi A) are human-specific pathogens, and the predominant cause of enteric (typhoid) fever globally. Enteric fever affects over 21 million people each year, resulting in 200,000 deaths [1]. Infection with S. Typhi and S. Paratyphi A usually begins with ingestion of contaminated water or food. The pathogens invade the gastrointestinal mucosa, translocate to the lymphoid follicles where they survive and replicate within macrophages, and then disseminate via the bloodstream to the liver, spleen, intestinal lymph nodes, bone marrow, and gallbladder [2]. With adequate treatment, most patients recover from their acute stage of illness and clear infection. However, a small percentage of S. Typhi (and S. Paratyphi A) infected individuals develop a chronic, but apparently asymptomatic, infection in the biliary tract that can persist for decades [3]–[6]. The likelihood of this is not known, but it is estimated that chronic carriage can complicate perhaps 1–3% of acute infections [7].
Since S. Typhi and S. Paratyphi A are human-restricted pathogens, chronic carriers may act as reservoirs of infection within a community. They contribute to the transmission cycle through the intermittent shedding of bacteria in feces (especially in areas of low transmission [8]) and may act as vehicles for introducing S. Typhi and S. Paratyphi A into previously uninfected communities. Therefore, correctly identifying and treating asymptomatic chronic carriers is critical for the long-term control of enteric fever. Currently, there is no reliable diagnostic assay to identify asymptomatic S. Typhi and S. Paratyphi A carriage. Bacterial stool culture has been used, yet is challenging due to the expense and logistics of obtaining multiple samples from patients, since shedding is typically low level and intermittent [5]. Measurement of antibody responses to the S. Typhi capsular Vi antigen has been previously described as a potential method to detect chronic S. Typhi carriers [7]. In laboratory settings, IgG to the Vi antigen has been shown to have a sensitivity of 75% and specificity of >95% and has proven to complement other strategies in outbreak investigations [7], [9]–[11]. However, its role in detecting asymptomatic carriers in a general endemic-zone population is unclear. In Chile, anti-Vi antibody responses had a sensitivity of 75% and specificity of 92%–97% for S. Typhi carriage; however, due to a low prevalence rate of carriage in the general population, its positive predictive value was only 8–17% [12]. In Vietnam, a large community-based survey for anti-Vi antibodies demonstrated a 3% positivity rate in the population; however, S. Typhi was never detected in the stool of individuals identified by such anti-Vi screening [13].
Understanding the mechanisms involved in development and persistence of the carrier state may facilitate development of improved diagnostic assays and therapeutic approaches for S. Typhi carriage. Currently, little is known about host-pathogen interactions in the biliary tract of chronic human carriers. Much of what is known about biliary carriage has been extrapolated from in vitro and murine studies with S. Typhimurium, which causes an enteric fever-like illness in mice [5]. From these animal studies and a complimentary study in humans, we know that gallstones facilitate S. Typhi carriage [5]. In the presence of bile, the bacterium regulates the expression of genes that allow it to colonize and persist in the gallbladder through formation of biofilms that mediate resistance against host defenses [14], [15]. There are likely other niches of persistent infection outside of the gallbladder, including the biliary tree, liver, and mesenteric lymph nodes. This is suggested by the observation that although cholecystectomy increases cure rates, it does not always result in clearance of the pathogen in humans [16]. In a murine model of Salmonella chronic infection, S. Typhimurium infection in Slc11a1 (Nramp1) wild-type mice demonstrated that the most common site of persistent infection was in hemophagocytic macrophages within mesenteric lymph nodes [2], [17], [18].
To advance our understanding of Salmonella pathogenesis of the chronic carrier state, and identify potential biomarkers unique to S. Typhi chronic carriers, we applied an immunoscreening technique called in vivo-induced antigen technology (IVIAT) [19]–[21]. IVIAT identifies humorally immunogenic bacterial antigens expressed in vivo and not in bacteria grown in standard laboratory conditions. We hypothesized that S. Typhi surviving in the biliary tract of humans may express a proteomic profile distinct from that expressed in bacteria grown using standard in vitro conditions or during acute infection.
This study was approved by the human studies committees of the involved research institutions: Massachusetts General Hospital, International Centre of Diarrheal Disease Research, Bangladesh (icddr,b), Patan Hospital, The Nepal Health Research Council, and the Oxford Tropical Research Ethics Committee. The study was conducted according to the principles expressed in the Declaration of Helsinki/Belmont Report, and informed written consent was obtained from adult participants and from guardians of children prior to study participation.
Salmonella enterica serotype Typhi strain CT18 [22] was obtained from the Salmonella Genetic Stock Centre (Calgary, Alberta, Canada). Genomic DNA from this strain was used to construct a genomic inducible expression library in host strain Escherichia coli strain BL21(DE3). Bacterial strains were grown in Luria-Bertani (LB) media (with 50 µg/ml kanamycin for clones containing pET30 constructs) and maintained at −80°C in LB broth containing 15% glycerol.
Individuals undergoing elective cholecystectomy in Kathmandu, Nepal were enrolled. At the time of cholecystectomy, a venous blood sample was stored and a bile sample was taken for microbiologic analysis as previously described [6]. Patients were categorized as (1) S. Typhi carriers if their bile culture was positive for S. Typhi; (2) S. Paratyphi A carriers if their bile culture was positive for S. Paratyphi A, or (3) cholecystectomy controls if their bile cultures were negative for any organism. Sera samples were also obtained from the following groups: (1) healthy Bangladeshi residents of Dhaka (a typhoid endemic area) enrolled at the International Centre for Diarrhoeal Disease Research, Bangladesh (icddr,b); and (2) acute (day 0–3) and convalescent sera (day 14–28) of Bangladeshi patients who presented to icddr,b with S. Typhi bacteremia [23]–[25].
Genomic DNA was purified from S. Typhi strain CT18 using a Genomic DNA Isolation kit (Qiagen, Valencia, Ca), sheared using a Covaris sonicatior (Woburn, Ma) optimized to generate 0.5–1.5 kb DNA fragments, and resulting fragments were gel purified using the Qiagen Qiaquick Gel Extraction kit. After terminal overhangs were removed using End-It DNA end-repair kit (Epicenter Biotechnologies, Madison, WI), the blunt-end products were ligated into pET-30c vectors (Novagen, San Diego, CA) that had been digested with EcoRV and treated with calf intestinal alkaline phosphatase. The library was electroporated into E. coli DH5α and bacteria were plated onto selective LB media containing kanamycin. After overnight incubation at 37°C, the plates were scraped and the plasmid DNA from collected colonies was recovered using Qiagen Miniprep kit. EcoRI and KpnI digestion was performed on a random sample of plasmids, and an insertion frequency greater than 80% and insert size between 500 to 1500 bp was verified. The plasmid DNA mixture was electroporated into E. coli BL21 (DE3), and collected colonies were stored in LB broth containing 15% glycerol.
Convalescent sera of 5 patients with bile cultures positive for S. Typhi were pooled, and adsorbed with in vitro grown S. Typhi strain CT18 and E. coli BL21 (DE3) [19]. Immunoblot techniques were used as previously described [19]. Briefly, the genomic library was plated on LB plates containing kanamycin to obtain a colony density of approximately 500 to 1000 clones per plate. After overnight incubation at 37°C, the resultant colonies were lifted off the plate using nitrocellulose membranes, and then the membranes were placed on LB media containing kanamycin and 1 mM isopropyl-β-D-thiogalactopyranoside for 4 hours at 37°C to induce transcription of insert DNA. Membranes were exposed to chloroform-soaked blotting paper to lyse bacteria, blocked for 1 hr using 5% milk in PBS with 0.25% Tween-20 (PBS/Tween), washed five times in PBS/Tween, and then incubated overnight with adsorbed sera at 1∶10,000 dilution. After membranes were washed 3 times with PBS/Tween, immunoreactive clones were detected using anti-human IgG conjugated to horseradish peroxidase (MP Biomedicals/Cappel, Aurora, OH) at a 1∶20,000 dilution, and immunoblots were developed with an enhanced chemiluminescence (ECL) kit (Amersham, Piscataway, NJ). Reactive clones were recovered from the master plates and saved as frozen glycerol stocks.
To confirm immunoreactive clones, secondary screening was performed comparing IgG immunoreactivity of the clones against E. coli BL21DE3 with an empty pET30c vector. Inserts of confirmed clones were sequenced to identify gene insert. Constructs designed to express the full length native protein were generated by amplifying the entire ORF of identified genes by PCR, and cloning these amplicons into pET30c as NdeI and NotI inserts. Immunoreactivity of these full ORF clones was compared to E. coli BL21DE3 with an empty pET30c vector. To assess immunoreactivity of identified antigens among the pertinent general population, immunoreactive clones were also screened using pooled sera of individuals living in a typhoid endemic area (Bangladesh). These sera were pre-adsorbed against in vitro grown E. coli BL21DE3, as described above, to reduce background reactivity against the host strain.
Functional classifications of identified proteins were assigned using published articles and available protein information resources, including J. Craig Venter Institute annotations (http://cmr.jcvi.org/tigr-scripts/CMR/CmrHomePage.cgi) and Pfam 26.0 (http://pfam.sanger.ac.uk/).
YncE (STY1479) was PCR-amplified from S. Typhi strain CT18 and the product was cloned into Gateway vector pDONR221 using BP reaction kit according to manufacturer's instructions (Invitrogen). The full length sequence was verified and transferred from pDONR221 into the Gateway expression vector pDEST17 using LR reaction kit (Invitrogen) generating pDEST17His6-yncE. The reaction product was transformed first into E. coli DH5α, and then the recovered plasmid was transformed into the expression strain BL21AI. To overproduce His6-YncE, E. coli BL21AI (pDEST17His6-yncE) was grown in 250 mL LB broth containing ampicillin at 37°C until OD600 0.6, and then expression of his6-yncE was induced by the addition of L(+) arabinose (0.2%). After 4 hours, the pellet was harvested by centrifugation, and the cells were lysed by sonication after resuspension in 15 mL lysis buffer (50 mM Tris Hcl, 5% glycerol, 0.1 M NaCl pH 8) containing 100 ug/ml lysozyme. Following centrifugation, the pellet was washed in lysis buffer with and without 1% Triton X-100, and the pellet was resuspended in 10 mL of 8 M urea, 50 mM NaH2PO4 and 300 mM NaCl (pH 7.4). His6-YncE was purified by HisPur Cobalt Resin (ThermoScientific, Rockford, Il) under denaturing conditions per the manufacturer's instructions. His6-YncE was then refolded by dialysis into 25 mM Tris-HCL 0.15 M NaCl, pH 8.0 using decreasing concentrations of urea. Product purity was assessed by polyacrylamide gel electrophoresis and Coomassie staining, and product identity was assessed by Mass spectrometry analysis. Protein concentration was determined via Coomassie (Bradford) Protein Assay Kit (ThermoScientific, Rockford, Il).
To further characterize immunoreactivity of the antigen with the most prominent immunoreactivity in our initial screening, anti-YncE (STY1479) IgG and IgA responses were measured in the sera of 10 S. Typhi carriers, 3 S. Paratyphi A carriers, 8 patients at acute (day 0–3) and convalescent phase (day 14–28) of typhoid fever with confirmed S. Typhi bacteremia, 8 Nepalese controls undergoing elective cholecystectomy with negative bile cultures, and 8 healthy Bangladeshis, in duplicate. Plates were coated with 100 ng/well of YncE and then sera were added at a 1∶200 dilution. Bound antibody was detected with anti-human IgG or IgA conjugated with horseradish peroxidase (Jackson Laboratories, Bar Harbor, ME) at a 1∶1000 dilution, and peroxidase activity was measured with the substrate 2,2-azinobis (ethylbenzthiazolinesulfonic acid). To compare across plates, kinetic readings (mAb/sec) of samples were averaged, divided by kinetic readings (mAb/sec) of an in-house pooled standard (pooled sera of five S. Typhi carriers confirmed by biliary culture), and then multiplied by 100. Results were expressed as units (U). The Mann-Whitney U test was used to compare differences between groups.
For evaluation of anti-Vi IgG and IgA responses, ELISA plates were coated with 200 ng/well of Vi antigen (Sanofi Pasteur, Lyon, France). The above sera were applied at a 1∶100 dilution, and bound antibody was detected with anti-human IgG and IgA conjugated with horseradish peroxidase at a 1∶1000 dilution. Peroxidase activity was measured with the substrate 2,2-azinobis (ethylbenzthiazolinesulfonic acid). To compare anti-Vi responses across plates, duplicate kinetic readings of samples were averaged, divided by average kinetic readings of an in-house pooled standard, and then multiplied by 100, as described above. Results were expressed as units (U). Differences between groups were assessed using the Mann-Whitney U test.
In the primary screen of over 120,000 clones, 565 clones were identified as immunogenic; 210 were confirmed by secondary screening. Sequence analysis of these inserts (many of which carried multiple potentially expressible ORFs) revealed 268 genes of interest with over 20% of genes identified multiple times, supporting validity of their identification and saturation of library screening. We subsequently sub-cloned the full coding sequences of 235 genes into individual expression plasmids, and identified 56 proteins with prominent IgG immunoreactivity using S. Typhi carrier sera, comparing immunoreactivity of expression clones to a clone containing an empty vector (Supplementary Table S1). Forty-eight of the identified genes are encoded on the chromosome of S. Typhi, 5 are encoded on the drug resistance plasmid pHCM1, and 3 on cryptic plasmid pHCM2. The most highly represented functional groups included proteins of unknown function and those involved in transport and binding, synthesis or salvage of ribonucleotides, and energy metabolism.
To assess the degree of immunoreactivity of antigens identified by IVIAT within the pertinent endemic-zone population, we screened the 56 immunoreactive clones against pooled sera of individuals living in a S. Typhi endemic area (Bangladeshi residents of Dhaka) [26]. Of these 56 proteins, 13 proteins had more prominent immunoreactivity when screened with sera of S. Typhi carriers compared to sera of healthy Bangladeshis. These 13 proteins included a number of putative membrane proteins, lipoproteins, and hemolysin-related proteins (Table 1). YncE, a possible ATP- binding protein, had the overall highest differential immunoreactivity compared to healthy endemic-zone control sera in our immunoblot assay.
To further characterize whether the immunoreactivity to YncE in S. Typhi carriers was specific, we also evaluated the immunoreactivity to YncE using sera of 5 groups of individuals: (1) S. Typhi carriers, (2) patients at the acute and convalescent phase of typhoid fever, (3) S. Paratyphi A carriers, (4) individuals who underwent cholecystectomy in Nepal whose bile cultures were negative for any pathogen, and (5) healthy controls from a typhoid endemic area (Dhaka, Bangladesh). We found significantly higher IgG immunoreactivity to YncE in S. Typhi carriers compared to bile culture-negative patients (p = 0.0205), healthy Bangladeshis (p = 0.0005), and patients at the acute and convalescent phases of typhoid infection (p = 0.0044 and p = 0.0266, respectively); there was a trend toward statistical significance when compared to S. Paratyphi A carriers (p = 0.21) (Figure 1A). Of the 10 S. Typhi carriers, 7 (70%) had an anti-YncE IgG response (Units >100). None of 8 bile culture negative controls (0%), 0 of 8 healthy Bangladeshis (0%), 0 of 3 S. Paratyphi A carriers (0%) and 1 of 8 (12.5%) Bangladeshis at the acute and convalescent phase of S. Typhi had an anti-YncE IgG response. Thus, in our small subset of patients, using a cut-off value of >100 Units (U), anti-YncE IgG had a sensitivity of 70%, and specificity of 100% when using endemic zone healthy individuals and cholecystectomy patients without detectable S. Typhi as controls. The specificity decreased to 95% if we included patients with acute typhoid fever.
Although, the values did not reach statistical significance, S. Typhi carriers also had a higher IgA immunoreactivity to YncE compared to our two control groups: bile culture-negative patients (p = 0.2370) and healthy Bangladeshis (p = 0.2031) (Figure 1B). There was no significant difference between the IgA immunoreactivity to YncE in S. Typhi carriers in comparison to patients convalescing from acute typhoid infection or S. Paratyphi A carriers.
Since immune responses to S. Typhi Vi antigen have been the best characterized diagnostic method for identifying S. Typhi carriers to date, we also assessed the anti-Vi IgG and IgA responses in the same cohort of patients. We found significantly higher IgG immunoreactivity to Vi antigen in S. Typhi carriers compared to PTA carriers (p = 0.0070), bile culture negative controls (p = 0.0343), healthy Bangladeshis (p = 0.0021), and patients at the acute phase of typhoid infection (p = 0.0155). (Figure 2A). There was a trend toward statistical significance when the immunoreactivity of S. Typhi carriers to Vi antigen was compared to patients at the convalescent phase of typhoid infection (p = 0.0830) (Figure 2A). In our evaluation of IgA anti-Vi responses, we did find a significant difference in the immunoreactivity of S. Typhi carriers compared to healthy Bangladeshis (p = 0.0155), and patients convalescing from acute typhoid infection (p = 0.0266) (Figure 2B). There was no significant difference in immune responses between S. Typhi carriers and bile-culture negative patients, S. Paratyphi A carriers, or patients at the acute phase of typhoid infection. The sensitivity for anti-Vi IgG and IgA was 40% (cutoff value >1250 U) and 50% (cutoff value >1250 U), respectively. The specificity was 100% for IgG irrespective of controls. For IgA, the specificity was 97% when using endemic zone healthy individuals and cholecystectomy patients without detectable S. Typhi as controls. The specificity was 94% if patients with acute typhoid fever were included in the analysis.
In our analysis, using a cut-off value of >100 U of anti-YncE IgG and/or >1250 U anti-Vi IgA, we could identify 8 out of 10 S. Typhi carriers. There was no added benefit seen when pairing anti-YncE responses with anti-Vi IgG.
In our immunoscreen using IVIAT, we were able to identify 56 immunogenic S. Typhi proteins using the sera of S. Typhi carriers. Of these, 13 had higher immunoreactivity when screened with S. Typhi carrier sera compared to sera of endemic zone residents. These proteins represent a working list of candidate diagnostic biomarkers of asymptomatic S. Typhi carriage and their analysis may further our understanding of survival adaptations of S. Typhi in chronic carriers.
Human epidemiologic studies as well as murine models of S. Typhi carriage suggest that gallstones facilitate the development of the chronic carrier state [5]. In support of this, we identified SirA in our IVIAT screen, which is part of the two-component response regulator SirA-BarA [27]. In S. Typhimurium, this regulator plays a role in the down-regulation of genes involved in invasion (i.e. Salmonella Pathogenicity Island-1) when the bacterium is in the presence of bile [28], and mutations in sirA result in decreased biofilm formation on plastic surfaces [28]. The role SirA may play in human or murine Salmonella carriage, or why a cytoplasmic regulatory protein generated a humoral response, has yet to be characterized. Other proteins identified in the IVIAT screen may also affect carriage in the presence of gallstones.
Although S. Typhi may persist in the gallbladder in association with gallstones [3], S. Typhi likely has other niches of infection, including the gallbladder epithelium, biliary tree, and in macrophages of mesenteric lymph nodes [2], [3], [5], [16]–[18]. Proteins identified in our screen may play a role in persistence of S. Typhi within host cells or the stringent environment of bile. For instance, YejE is a putative permease that is thought to be a component of a putative ABC transporter system. YejE plays a role in survival within epithelial cells and in antimicrobial peptide resistance [29]. In both S. Typhi and S. Typhimurium, yejE expression is upregulated inside host macrophages [30], [31]. PduG is a protein encoded within the pdu operon that is part of the coenzyme B12-dependent 1,2-propranediol utilization pathway [32]. This operon is upregulated during acute S. Typhi and S. Paratyphi A infection in humans [24], [33], and may be associated with use of alternative carbon sources in the nutrient-limited environment of the Salmonella-containing vacuole within host cells [32]. We also identified PurH and XapB, which are proteins involved in purine biosynthesis and acquisition, respectively, by functional classification. In S. Typhimurium, PurH is associated with virulence [34], and we have previously shown that genes involved in purine synthesis are upregulated during acute typhoid infection in humans [24]. CorC is a hemolysin-related protein involved in magnesium and cobalt efflux, and is part of the CorA transporter system containing CorA-D [35]. CorA, with associated proteins, is required for efflux of Mg2+ [35]. CorA is required for S. Typhimurium virulence [36], and corA is expressed by S. Typhi during acute human infection [24]. However, while some information is known regarding the above mentioned Salmonella carrier-specific antigens, their potential role in carriage is presently unclear.
The majority of the genes identified by IVIAT encode for proteins with putative or unknown function. For example, STY2386 is an uncharacterized lipoprotein found uniquely in Salmonella. STY1364 is a hypothetical periplasmic protein in S. Typhi and S. Paratyphi A, and is rarely found in other Salmonella spp. STY1364 belongs to the structural classification of bacterial enterotoxins and is a subtilase cytotoxin subunit B-like protein. We previously identified STY1364 in S. Typhi infected patients using a separate immunoscreening technology (immunoaffinity proteomic-based technology, IPT) [23].
In our screening, YncE (STY1479) was the most immunoreactive antigen identified, and we thus focused our more detailed analysis of immunoreactivity on this antigen. YncE has a putative N-terminal signal sequence suggestive of export, with ATP and DNA-binding domains. yncE is present in a number of Salmonella spp., and has orthologs in a number of other Gram-negative enteric organisms, including Escherichia coli, Citrobacter spp, and Shigella spp. In E. coli, YncE is secreted into the periplasm via the Sec-dependent pathway [37], and its expression is induced under iron-restricted conditions when repression by the Fur protein is relieved [38]. Its role in the pathogenesis of Salmonella infection has yet to be characterized. However, our results suggest that it may be involved in long-term persistence of the bacterium in chronic carriers.
In our analysis, we show that S. Typhi carriers have an IgG response to YncE that is not present in bile culture-negative controls in Nepal or healthy controls in Bangladesh. Although we did not reach statistical significance in this small pilot study, a similar trend was seen for IgA as well. One patient convalescing from acute typhoid infection had a detectable IgG anti-YncE response, and another had an IgA response. This may suggest that anti-YncE responses occur during acute disease; however, it should be noted that we do not know the current or future carrier status of the acute typhoid patients, and an elevated level of YncE during an episode of typhoid fever may represent an acute on chronic infection, or may be a marker of future progression to the chronic carrier state.
All of the identified genes except three (xapB and the two genes encoded on the cryptic plasmid pHCM2) are present in the genome of S. Paratyphi A sequenced strains ATCC 9150 and AKU 12601 based on <60% nucleotide identity. It is interesting then, that we did not see an IgG or IgA immune response to YncE in S. Paratyphi A carriers. This finding may suggest that S. Typhi and S. Paratyphi A use different strategies to persist in chronic carriers, that expression of YncE may be distinct in these two organisms, or that our study did not have sufficient power to examine this, as it included only three S. Paratyphi A carriers.
Despite this, in our small cohort of patients, measurement of anti-YncE IgG responses did appear to be both relatively sensitive and specific for identifying asymptomatic chronic S. Typhi carriers. Further studies will be needed to evaluate the diagnostic capabilities of anti-YncE responses in a larger and different cohort of patients. Of note, if such studies demonstrate higher anti-YncE IgA levels in S. Typhi carriers than in control groups, that information could support consideration of a salivary diagnostic to facilitate community-based screening for carriage.
The other antigens identified in our IVIAT analysis may also be useful diagnostic biomarkers of S. Typhi carriage, and the sensitivity of carrier detection may be improved when responses against these or anti-Vi responses are paired with responses to YncE. For example, in our analysis, using a cut-off value of >100 U of anti YncE IgG and/or >1250 U anti-Vi IgA, we could identify 8 out 10 S. Typhi carriers. There was no added benefit seen when pairing anti-YncE responses with anti-Vi IgG. Another potential pairing could include a marker of biliary tract inflammation such as elevated bilirubin values, since S. Typhi carriage is often associated with chronic inflammation of the gallbladder [5]. We did not assess this parameter in this study.
Our study has a number of limitations. First, the number of patients involved in our study is small, although it should be noted that our analysis is the largest study involving immunoproteomic screening and pilot confirmation of the carriage state that includes appropriate control groups. A second limitation is that IVIAT identifies proteins that are uniquely expressed in vivo compared to standard in vitro culturing, and that also induce an antibody response. Proteins that induce cellular responses and/or that are expressed both in vivo and in vitro may also play a role in the pathogenesis of chronic carriage and serve as useful biomarkers for asymptomatic carriage. In addition, altering in vitro culturing conditions may also change the expression profile of S. Typhi, thereby changing the comparison groups. In addition, IVIAT does not identify non-protein antigens that may also be useful in diagnostic assays. However, despite these limitations, we have used IVIAT to identify a subset of immunoreactive antigens in S. Typhi carriers, including YncE. Further evaluation of YncE and other identified antigens could lead to the development of improved diagnostic assays to detect asymptomatic S. Typhi carriers in typhoid endemic zones, and analysis of YncE, along with other identified antigens, could lead to an improved understanding of host-pathogen interactions during chronic carriage of S. Typhi in humans.
|
10.1371/journal.ppat.1003738 | Viral Escape from HIV-1 Neutralizing Antibodies Drives Increased Plasma Neutralization Breadth through Sequential Recognition of Multiple Epitopes and Immunotypes | Identifying the targets of broadly neutralizing antibodies to HIV-1 and understanding how these antibodies develop remain important goals in the quest to rationally develop an HIV-1 vaccine. We previously identified a participant in the CAPRISA Acute Infection Cohort (CAP257) whose plasma neutralized 84% of heterologous viruses. In this study we showed that breadth in CAP257 was largely due to the sequential, transient appearance of three distinct broadly neutralizing antibody specificities spanning the first 4.5 years of infection. The first specificity targeted an epitope in the V2 region of gp120 that was also recognized by strain-specific antibodies 7 weeks earlier. Specificity for the autologous virus was determined largely by a rare N167 antigenic variant of V2, with viral escape to the more common D167 immunotype coinciding with the development of the first wave of broadly neutralizing antibodies. Escape from these broadly neutralizing V2 antibodies through deletion of the glycan at N160 was associated with exposure of an epitope in the CD4 binding site that became the target for a second wave of broadly neutralizing antibodies. Neutralization by these CD4 binding site antibodies was almost entirely dependent on the glycan at position N276. Early viral escape mutations in the CD4 binding site drove an increase in wave two neutralization breadth, as this second wave of heterologous neutralization matured to recognize multiple immunotypes within this site. The third wave targeted a quaternary epitope that did not overlap any of the four known sites of vulnerability on the HIV-1 envelope and remains undefined. Altogether this study showed that the human immune system is capable of generating multiple broadly neutralizing antibodies in response to a constantly evolving viral population that exposes new targets as a consequence of escape from earlier neutralizing antibodies.
| Four sites of vulnerability for broadly neutralizing antibodies to HIV-1 have been identified thus far. How these broadly reactive antibodies arise, and the host-pathogen interactions that drive the affinity maturation necessary for neutralization breadth are poorly understood. This study details the sequential development of three distinct broadly neutralizing antibody responses within a single HIV-1 infected individual over 4.5 years of infection. We show how escape from the first wave of antibodies targeting V2 exposed a second site that was the stimulus for a new wave of glycan dependent broadly neutralizing antibodies against the CD4 binding site. These data highlight how antibody evolution in response to viral escape mutations served to broaden the host immune response to these two epitopes. Finally, we document a third wave of neutralization that targets an undefined epitope that did not appear to overlap with the four known sites of vulnerability on the HIV-1 envelope. These data support the design of templates for sequential immunization strategies aimed at increasing neutralization breadth through the recognition of multiple epitopes and their immunotypes.
| Neutralizing antibodies are the principal correlate of protection for most preventative vaccines. Designing suitable vaccine immunogens to elicit these types of antibodies has been relatively simple for conserved pathogens such as smallpox and other DNA viruses. For more diverse pathogens like HIV-1, the neutralizing antibodies elicited by vaccination or during natural infection are largely strain-specific and therefore would not be protective against globally circulating viral variants [1]–[5]. The HIV-1 envelope glycoprotein spikes mediate viral entry and are the sole targets for neutralizing antibodies. The spikes are trimeric, made up of three non-covalently associated gp41-gp120 heterodimers, each with a conserved core that mediates infection of CD4+ T-cells. Functionally conserved sites are protected by extensive glycosylation, and large solvent exposed hypervariable structures (the V1–V5 loops, and the α2-helix in C3) [6]. All HIV-1 infected individuals develop strain-specific neutralizing antibodies which target these sequence variable regions, but only a quarter develop broadly neutralizing antibodies [7]–[11], which will likely be needed for a preventative HIV-1 vaccine. To engineer an envelope immunogen that can specifically elicit these antibodies, the HIV-1 vaccine research field has adopted a strategy based largely on rational design: identifying the targets for these broadly cross-reactive antibodies, and elucidating the pathways that promoted their development.
Plasma mapping strategies and the isolation of monoclonal antibodies have defined four major targets for broadly neutralizing antibodies on the HIV-1 glycoprotein [7]–[10], [12]–[20]. The CD4 binding site (CD4bs) of gp120 and the membrane proximal external region (MPER) of gp41 are glycan independent epitopes, while the V1/V2 sub-domain and the co-receptor/V3 site on gp120 are sites of vulnerability for glycan binding antibodies (predominantly at positions N156/N160 and N301/N332 respectively) [14]–[16]. Both CD4bs antibodies and co-receptor/V3 antibodies bind well to monomeric gp120, while MPER antibodies bind to a linear peptide in gp41. This makes it possible to adsorb out their neutralization activity from plasma with various recombinant proteins. In contrast the epitope for V2 antibodies (such as PG9/16) consists of two anti-parallel β-sheets (B- and C- strands) of a Greek key motif, and the glycans therein, that is preferentially formed on the native trimer. This region is critically important for the gp120-gp120 interactions that stabilize the envelope glycoprotein spike in its unliganded conformation, and therefore cannot be readily adsorbed [16], [21]. Various sub-epitopes within each of these four major sites of vulnerability have also been identified through subtle differences in the mechanism of neutralization [22]. For instance antibodies targeting the CD4bs can be sub-divided into two groups: those that are sensitive to the D368A and/or E370A mutations in the CD4 binding loop α3 (such as VRC01); and those that are dependent on amino acids D474, M475, and/or R476 in α5 termed CD4bs/DMR (such as HJ16) [23]–[26]. Despite this detailed knowledge, epitope mapping strategies have failed to identify the neutralization targets in a subset of plasma samples [7]–[9], [17], [19], [27], [28]. The antibodies mediating breadth in these samples could target sub-epitopes within one of the four sites of vulnerability or they may target entirely novel epitopes.
In the CAPRISA 002 Acute Infection Cohort, we previously identified seven individuals with broadly neutralizing antibodies. In five cases we were able to map the plasma antibody specificities to known epitopes (two targeted N332, two the V2 epitope, and one the MPER) [7]. In this study we have focused on one of the individuals (CAP257) for which the target was undefined. Heterologous and autologous neutralization data as well as viral sequences from longitudinal samples were used to identify the epitopes for CAP257 broadly neutralizing antibodies. We showed that heterologous neutralization in CAP257 was conferred by three distinct, sequentially occurring antibody waves, two of which were mapped to epitopes in V2 and the CD4bs respectively. While individuals with more than one broadly neutralizing antibody specificity have been previously identified [29]–[31], there is little information on how the dynamic relationship between host and pathogen contributed to the development of antibodies targeting multiple epitopes. We have shown previously that escape from strain-specific neutralizing antibodies can drive the formation of epitopes for broadly neutralizing antibodies [32]. Here we found that viral escape from broadly neutralizing antibodies targeting V2 promoted the development of a second broadly neutralizing antibody response targeting a glycan dependent epitope in the CD4bs. We also identified early escape mutations from both the V2 and CD4bs antibodies that drove an increase in the neutralization breadth of CAP257 plasma. These findings have implications for the design of HIV-1 vaccine antigens and sequential immunization strategies.
We have previously described the development of neutralization breadth in CAP257 using longitudinal plasma samples from HIV-1 seroconversion to three years post-infection (p.i.) [7]. Here, we extended this analysis until the start of anti-retroviral therapy at four and a half years p.i. (Figure 1A). Longitudinal plasma was tested against the autologous CAP257 virus amplified from the earliest available time point (7 weeks p.i.), the subtype C consensus sequence (ConC) [33], 4 Tier 1b viruses, and 39 Tier 2 viruses [34]. Autologous neutralizing antibodies appeared by 14 weeks of infection with a peak titer at two years of 1∶6,754. This was followed by the neutralization of heterologous viruses 30 weeks after infection. CAP257 neutralized 84% of the heterologous viruses at three years (174 weeks) with neutralization breadth of 100% against subtype A (6/6 viruses), 96% against subtype C (25/26 viruses, including ConC), and 50% against subtype B (6/12 viruses). The titers of these broadly neutralizing antibodies peaked and waned in three separate waves.
The first wave of neutralization breadth (typified by CAP63) peaked at 67 weeks p.i. with a maximum titer of 1∶1,493 and exclusively neutralized subtype C viruses (Figures 1A and 1B – red curves). Wave 1 titers dropped to as low as 1∶145 by 149 weeks of infection. As this early heterologous neutralization began to wane, CAP257 plasma gained the capacity to neutralize additional subtype C viruses as well as several subtype A and B viruses (Figures 1A and 1B – blue and green curves). This second wave (typified by Q842) peaked at 122 weeks p.i. with titers as high as 1∶8,565 against RHPA that dropped to 1∶1,254 by 213 weeks of infection. Finally, a third wave of heterologous neutralization (represented by Du156) appeared by 149 weeks p.i. and peaked at 213 weeks p.i. (Figure 1B – brown curve). This third wave was also largely subtype C specific. These data suggested that the neutralization breadth of CAP257 plasma was mediated by at least three distinct antibody specificities.
To identify the targets of each of the three broadly neutralizing antibody specificities in CAP257 plasma we first assessed whether they targeted epitopes in monomeric gp120 (such as the CD4bs or N301/N332 glycans). We adsorbed out the gp120 binding antibodies in plasma samples from the peak neutralizing activity of each wave using recombinant ConC gp120 coupled to tosyl-activated magnetic beads (Figure 1C). The adsorbed plasma was compared with untreated plasma for activity against a heterologous virus neutralized by each wave. Neutralization by wave 1 (67 weeks p.i.) and wave 3 (213 weeks p.i.) was not affected by adsorption with monomeric gp120 (Figure 1C – red and brown curves), however neutralization by wave 2 antibodies (122 weeks p.i.) could be partially adsorbed with the ConC gp120 protein (Figure 1C – blue curve). The neutralizing activity of wave 2 could also be equally adsorbed with a core gp120 lacking the hypervariable loops V1/V2 and V3 (Figure 1C – green curve). These data supported our hypothesis that CAP257 heterologous neutralization was mediated by more than one neutralizing antibody specificity, two of which were largely subtype C specific and targeted an epitope not present on monomeric gp120, and a third whose epitope in gp120 was more conserved across clades and did not require the hypervariable loops V1/V2 or V3.
The inability of gp120 to adsorb out wave 1 neutralization suggested these antibodies might recognize the trimer specific epitope in V1/V2 defined by PG9/16 [16]. Therefore, we performed mapping studies using ConC, which was neutralized by all three waves of neutralizing antibodies (Figure 2A – red curve). Seven mutations in the V2 region (F159A, N160A, R166A, K168A, K169E, K171A, and I181A) each abrogated wave 1 neutralization, but did not significantly affect the titers of waves 2 or 3 (Figure 2A – purple curves). In contrast, a D167N mutation resulted in enhanced neutralization by wave 1 antibodies, but did not significantly affect the titers of waves 2 or 3 (Figure 2A – orange curve). A second mutation (L165A) also resulted in significant neutralization enhancement at all the time points tested, including those preceding breadth (Figure 2A – grey curve), suggesting that this mutation resulted in general neutralization sensitivity. Overall, these data indicated that wave 1 antibodies (but not waves 2 or 3) targeted residues in the V2 region.
To define whether the V2 epitope recognized by CAP257 plasma antibodies overlapped with that of known broadly neutralizing antibodies to this site, we tested the sensitivity of PG9/16, CH01-04, and PGT145 to the same ConC V2 mutations described above and compared them to CAP257 neutralizing antibodies at the peak of wave 1 activity, 67 weeks p.i. (Figure 2B). Of the seven mutations that abrogated CAP257 neutralization only two (N160A and K169E) resulted in complete resistance to all the antibodies tested, consistent with previous data [35], [36]. Neither the monoclonal antibodies nor CAP257 wave 1 antibodies were sensitive to deletion of the N156 glycan (through the S158A mutation) in ConC. Lastly, mutations at two hydrophobic amino acids in V2 (F159A and I181A) that do not form part of the PG9 epitope as defined by the crystal structure [37], had a significant effect on the neutralization of monoclonal antibodies targeting V2, and CAP257 wave 1 antibodies.
To define escape from wave 1 neutralizing antibodies, we examined sequences from the V2 region of CAP257 over time. Using single genome amplification (SGA) we obtained 125 full envelope sequences from twelve time points between 7 and 213 weeks p.i., and focused on the N160 glycan and the cationic C-strand in V2 that are the targets of wave 1 antibodies (Figure 3A). Interestingly, the earliest virus (7 weeks p.i.) had an asparagine at position 167. This N167 residue is rare, occurring in only 5.6% (196 of 3,478) of sequences in the Los Alamos National Laboratory (LANL) HIV sequence database. By the time of the earliest detectable heterologous neutralization (30 weeks p.i., maximum titer of 1∶49) mutations in sites forming part of the wave 1 V2 epitope were already apparent in 6/14 autologous sequences at positions R166, K169, and Q170 (Figure 3A). Of the remaining eight sequences, six exhibited other mutations either in the N160 glycosylation sequon or the V1/V2 C-strand. This rapid selection pressure in the C-strand of V1/V2 was sometimes an N167D mutation (4/14 autologous sequences) that was unlikely to be selected for by wave 1 broadly neutralizing antibodies, as all the heterologous viruses neutralized by wave 1 had a D167 residue. Since the V1/V2 region is a common target of strain-specific neutralizing responses [38]–[41], these data suggested the possibility of an earlier neutralizing response targeting N167 in V2 that preceded the development of broadly neutralizing antibodies. Wave 1 mapping data (Figure 2A) further supported this possibility because the reverse D167N mutation enhanced the neutralization of ConC by wave 1 antibodies only, and resulted in earlier neutralization kinetics (Figure 2A – orange curve).
To test this we selected an envelope from 174 weeks p.i. (CAP257 3 yr) that was completely resistant to CAP257 neutralizing antibodies (consistent with ongoing neutralization escape), and back-mutated the V1/V2 region to match the earliest sequence from 7 weeks p.i. (Figure 3B). The neutralization sensitivity of the back-mutated virus, CAP257 3 yr(V1/V2s), was then compared to the parental CAP257 3 yr virus using longitudinal plasma samples. In contrast with the resistant CAP257 3 yr virus, the back-mutated V1/V2 virus (CAP257 3 yr(V1/V2s)) became sensitive to neutralization at 23 weeks p.i. (Figure 3B – black curve). This suggested the emergence of a strain-specific V1/V2 response 7 weeks prior to the development of wave 1 broadly neutralizing antibodies at 30 weeks p.i.
To establish whether these strain-specific V1/V2 neutralizing antibodies targeted the same epitope as wave 1 broadly neutralizing antibodies, we introduced selected escape mutations (N167D, N160D/S and K169E) into the sensitive CAP257 3 yr(V1/V2s) back-mutated envelope. Introduction of the N167D mutation (Figure 3B – orange curve), a common V2 change at 30 weeks p.i., shifted the timing of autologous V1/V2 neutralization to overlap with the emergence of wave 1 broad neutralization. The introduction of N160D/S mutations that deleted the N160 glycan (Figure 3B – purple curves), further shifted autologous neutralization to overlap with the emergence of wave 2 neutralizing antibodies. Finally, when the K169E mutation was introduced (Figure 3B – pink curve), autologous V1/V2 neutralization titers were completely abrogated. As the N160A and K169E mutations in ConC also completely abrogated neutralization by CAP257 wave 1 broadly neutralizing antibodies (Figure 2B) these data suggest that the strain-specific V2 neutralizing antibodies in CAP257 plasma targeted the same site of vulnerability that was later targeted by wave 1 broadly neutralizing antibodies. However the strain-specific V2 antibodies recognized the rare N167 immunotype of V2 present in the CAP257 infecting virus. Following an early N167D escape mutation at this site, V1/V2 neutralization became N167 independent, allowing recognition of the more common D167 immunotype. This switch in the fine specificity of CAP257 V2 antibodies correlated with the emergence of broadly neutralizing wave 1 antibodies.
Wave 1 broadly neutralizing antibodies were completely dependent on the glycan at N160 (Figure 2A). Viral escape from wave 1 neutralizing antibodies by deletion of this glycan first occurred at 54 weeks p.i. (in 37.5% of the sequences), immediately prior to the development of wave 2 neutralizing antibodies (Figure 3A – pie charts). This escape pathway persisted at 93 weeks p.i. (in 30% of sequences), but by 122 weeks p.i., at the peak of wave 2 activity, alternative escape pathways existed, and all sequences contained the N160 glycan. The transient nature of this highly effective escape pathway suggested that deletion of the N160 glycan had a deleterious effect on the virus.
A potential mechanism for this came from the observation that deleting the N160 glycan (critical to wave 1 neutralization) in the CAP257 3 yr(V1V2s) virus conferred slight neutralization sensitivity coinciding temporally with wave 2 (Figure 3B – purple curves). These data suggested that mutations at N160 exposed the wave 2 epitope. To examine whether loss of the N160 glycan enhanced CAP257 wave 2 neutralization we selected two viruses, Q842 and RHPA, which were neutralized at high titer by wave 2 but resistant to wave 1, and deleted the N160 glycan in each. The effect of these N160K mutations was assessed longitudinally using CAP257 plasma. While the timing of RHPA N160K and Q842 N160K neutralization by CAP257 was not altered compared to the wild-type viruses, these mutant viruses were neutralized 2–8 fold more potently by wave 2 antibodies (Figure 3C – purple curves). A similar 2 fold increase in titer was shown at the peak of wave 2 activity when the N160 glycan was deleted in ConC (Figure 2A – purple closed circles). As deletion of the N160 glycan in the autologous virus occurred prior to the development of wave 2 neutralizing antibodies, these data suggest that this particular escape pathway from wave 1 neutralizing antibodies may have contributed to the development of wave 2 antibodies, possibly by better exposing the epitope. Therefore, after the development of wave 2, the K169E mutation that also allowed escape from wave 1 broadly neutralizing antibodies, but did not enhance wave 2 neutralization, was preferentially selected over deletion of the N160 glycan (Figure 3A).
The V1/V2 sub-domain of gp120 plays an important role in shielding the envelope from neutralization. More specifically, several modifications in V1/V2 (particularly at N-glycosylation sites) have been implicated in either shielding or exposing the CD4bs to neutralizing antibodies [42]–[51]. As escape from wave 1 at N160 enhanced neutralization by wave 2, these data suggested the CD4bs as the target for wave 2 antibodies. This was supported by the adsorption data (Figure 1C), showing that wave 2 neutralizing antibodies bound the core gp120 protein. Therefore we examined envelope sequences from 191 weeks p.i. for selection pressure in the conserved CD4bs (Figure 4). Three mutations in the D-loop (N276D/S, T278A/K and N279D) and one at the base of V5 in the β23 sheet of C4 (R456W) dominated the viral population at this time point. Both the D-loop and V5 have been previously implicated in resistance to CD4bs antibodies [52]–[54]. To assess whether these CD4bs mutations mediated resistance to wave 2 neutralization, we introduced the three most common mutations (T278A, N279D and R456W) simultaneously into two heterologous viruses (Q842 and RHPA) neutralized by wave 2 but not by wave 1 (Figure 1A), and tested them against plasma from the peak of wave 2 activity (122 weeks p.i.). The mutants were at least 20 fold more resistant to neutralization at this time point than the wild-type viruses, confirming the role of CD4bs mutations in escape from wave 2 antibodies (Figure 5A).
To further characterize the epitope targeted by wave 2 antibodies, we assessed the dependence of wave 2 binding on the D368 residue in α3 (critical for VRC01-like antibodies), and residues 474/475/476 in α5 (critical for HJ16-like antibodies) by adsorption studies. The D368R or D474A/M475A/R476A mutations were separately introduced into ConC gp120, and compared to the wild-type protein for their ability to adsorb out the neutralizing activity against Q842 and RHPA at peak wave 2 titers. Both mutant gp120s (Figure 5B – yellow and brown bars) adsorbed out a significant fraction of the neutralizing activity against Q842 and RHPA, equivalent to that adsorbed by wild-type gp120 (Figure 5B – white bars). These data suggested that CAP257 antibody binding was not dependent on these residues in the α3 and α5 helices. CAP257 neutralization could not be adsorbed with the RSC3 protein used to isolate VRC01, which binds weakly to the HJ16 class of CD4bs antibodies [55].
Of the three changes identified above as mediating escape from wave 2 antibodies, N279 and R456 make contact with CD4, while T278 forms part of an adjacent glycosylation sequon [6]. This glycosylation sequon is conserved in 96% (n = 3,475) of envelope sequences in the LANL HIV sequence database. Its deletion via N276D/S or T278A/K mutations in later viruses (Figure 4) for wave 2 escape was therefore striking, and suggested a possible role for glycan recognition by CAP257 wave 2 antibodies. To examine wave 2 glycan binding, we expressed an RHPA core gp120 in GnTI(−/−) 293S cells, which allowed for deglycosylation of the protein using Endo-H, and assessed the ability of both glycosylated and deglycosylated proteins to adsorb out wave 2 neutralization activity. While neutralizing activity against RHPA was efficiently adsorbed with the glycosylated RHPA gp120 core (Figure 5C – white bars), the deglycosylated protein only adsorbed out a fraction of that activity (Figure 5C – purple bars) confirming the importance of glycans in wave 2 antibody binding. Similarly HJ16 was not adsorbed by the deglycosylated protein suggesting glycan dependence. In contrast, VRC01 was adsorbed equally effectively by both proteins.
To assess in more detail the role of the N276 glycan in wave 2 neutralization, we generated four mutants in RHPA, comparing the effects of two conservative mutations (N276Q, T278S) with the effects of two alanine substitutions (N276A, T278A) in the N276 glycosylation sequon (Figure 5D – boxed in purple). Both alanine mutations significantly affected wave 2 peak titers by 18 and 30 fold respectively. The N276Q mutation which deleted the glycan but retained the amino acid properties at position 276 also affected wave 2 neutralization by 21 fold (a similar effect to the N276A mutation), while the T278S mutation that retained the N276 glycan had no effect on neutralization. These data suggested that sensitivity to wave 2 neutralization was largely dependent on the glycan at position 276, rather than the N276 amino acid side chain. We also assessed the effect of the remaining two autologous mutations in the CD4bs (N279D and R456W) identified above (Figure 5D). The N279D mutation alone had a relatively small 2 fold effect on neutralization, suggesting only a minor role in escape from wave 2. When an alanine was substituted at position 279 instead, wave 2 neutralization was enhanced. The R456W mutation had a more significant 10 fold effect on CAP257 neutralization, but was still less effective than the glycan deleting mutations at positions 276/278 which were the major escape mutations.
We next compared the epitope for CAP257 wave 2 neutralizing antibodies with that of HJ16 (CD4bs/DMR) and VRC01 (CD4bs). Both monoclonal antibodies were profoundly affected by the R456W mutation (514 and 30 fold respectively). Like CAP257, HJ16 neutralization was significantly dependent on the glycan at position 276 with glycan deleting mutations (N276Q/A and T278A) resulting in a 74–106 fold increase in IC50 (Figure 5D). This dependence on the N276 glycan distinguished both CAP257 and HJ16 from VRC01, for which neutralization was slightly enhanced (2–3 fold) when the glycan was removed. This effect on VRC01 is consistent with previous studies showing that deleting the N276 glycan exposes the CD4 binding site to neutralization by VRC01 or b12 [52], [56]. Introducing all three CAP257 escape mutations identified above therefore had a compensatory effect on VRC01 resistance (8 fold compared to 30 fold effect for R456W alone), but completely abrogated HJ16 neutralization, confirming the similarities between HJ16 and CAP257 plasma. Despite these overall similarities, some differences were apparent between CAP257 wave 2 antibodies and HJ16, such as the preference of HJ16 for the threonine at position 278 and the asparagine at position 279. Unlike CAP257, the N279A mutation did not enhance HJ16 neutralization but rather resulted in a 4 fold reduction in titer. These data may suggest minor contacts between HJ16 and the amino acid side chain at position 279.The N279A mutation also significantly affected VRC01 neutralization (76 fold), consistent with either asparagine or aspartic acid residues at position 279 being directly contacted by W100B in the CDR3-H3 of VRC01 [52], [57].
To clarify the role of glycan binding we tested three CD4bs neutralizing antibodies (VRC01, b12, and HJ16) in ELISA for binding to either the glycosylated or deglycosylated RHPA gp120 core proteins (Figure 5E). The neutralizing antibody 2G12 has a well-defined glycan epitope and was used as a positive control [58]–[62], while CAP88-3468L (a V3 binding antibody) served as a negative control [63]. Deglycosylation significantly affected the binding of both 2G12 and HJ16 (Figure 5E – blue and purple curves), but did not significantly affect binding of either VRC01 or b12 (Figure 5E – green and yellow curves) to the RHPA gp120 core. These data confirmed the glycan binding properties of HJ16, and suggest that both CAP257 wave 2 neutralizing antibodies and HJ16 have a glycan dependent mechanism of neutralization at the CD4bs.
While the simultaneous introduction of T278A, N279D, and R456W mutations into heterologous viruses Q842 and RHPA made them resistant to wave 2 neutralization at 122 weeks p.i. (Figure 5A), when each mutation was introduced individually into RHPA the results were varied (2–30 fold reductions in titer) with no single mutation resulting in complete escape (Figure 5D). These data suggested that escape from wave 2 required a combination of all three mutations. Longitudinal sequence analysis showed that the N279D mutation emerged first at 93 weeks in 5% of the population, followed by N276D/S glycan deleting mutations at 122 weeks in 36% of the population, and then by substitution of the R456 residue with a bulky amino acid side chain (H, Y, or W) at 161 weeks p.i. in 17% of the population (Figure 4).
To better characterize their contributions to escape from wave 2 antibodies, each mutation (T278A, N279D, and R456W) was introduced separately into Q842 and RHPA and compared to the wild-type viruses. The N279D mutation (Figure 6 – orange curves) had a significant effect only at the beginning of wave 2 activity against Q842, shifting the earliest heterologous neutralization of Q842 from 80 weeks to 93 weeks p.i. (with a 5 fold effect on titers at 107 weeks p.i.). The mutation also affected the titers against RHPA by 2–3 fold. This suggested an initial dependence on N279, with later wave 2 antibodies being less vulnerable to mutations at this residue. This reduced dependence on N279 coincided with the emergence of the N279D mutation at 93 weeks p.i. (Figure 6 – orange dotted lines) providing a mechanism for maturation of the antibody response. The T278A mutation (Figure 6 – purple curves) that deleted the N276 glycan conferred almost complete resistance in Q842, and shifted the earliest neutralization of RHPA from 67 to 122 weeks p.i. The R456W mutation (Figure 6 – pink curves) did not result in a further right shift in the neutralization curves for RHPA relative to the T278A mutation, but did affect the neutralization kinetics of Q842 relative to the wild-type or N279D mutant viruses. As with the N279D mutation, these changes in the fine specificity of the maturing antibody response reflected the emergence of either T278A or R456W mutations in longitudinal autologous sequences (Figure 6 – purple and pink dotted lines respectively).
These data are consistent with accumulating resistance to wave 2, and suggest that after each successive round of escape, new antibody variants emerged that were able to neutralize first the N279D mutant, and then later autologous viruses with additional polymorphisms at positions 276, 278, or 456. We wished to determine whether the ability to neutralize escaped viral variants correlated with increased wave 2 neutralization breadth. Heterologous viruses neutralized by wave 2 were divided into two groups, those neutralized at 67 weeks p.i. (early wave 2 neutralization), and those first neutralized at 93 weeks p.i. (late wave 2 neutralization) after the emergence of initial wave 2 escape mutations (Figure 7A). Viruses also neutralized by wave 1 were omitted as the overlapping titers confounded this analysis. Inspection of the envelope sequences (particularly in the D-loop) showed that all of the viruses neutralized by early wave 2 antibodies had the N279 immunotype (Figure 7B – boxed in orange). In contrast 44% of viruses neutralized by later wave 2 antibodies had the D279 immunotype. Furthermore, of the viruses neutralized by later wave 2 antibodies, one (Q259) lacked the N276 glycan and four others also had additional non-conservative mutations at positions 273–275 in the N-terminus of the D-loop (Figure 7B – boxed in blue) that may also affect early wave 2 neutralization. These data suggest that wave 2 escape mutations guided maturation of the CD4bs response, enabling later wave 2 antibodies to neutralize additional heterologous viruses and ultimately resulting in the increased neutralization breadth of CAP257 plasma.
Like wave 1, wave 3 neutralization could not be adsorbed with monomeric gp120, suggesting that these antibodies targeted a quaternary epitope, or an epitope in gp41. To assess whether wave 3 was a distinct antibody specificity, or a re-emergence of wave 1 antibodies, we selected a virus (Du156) that was sensitive to waves 1 and 3, and less sensitive to wave 2 (Figure 8A – red curve). Resistance to wave 1 (V2) and wave 2 (CD4bs) neutralizing antibodies was established by introducing the N160K and T278A mutations identified above. The resulting virus (Du156 N160K/T278A) remained sensitive to wave 3 neutralization only (Figure 8A – brown curve), suggesting that wave 3 differed from wave 1 (which was completely abrogated by the N160K mutation). To confirm this, we introduced additional V2 mutations (R166A, K168A, K169E, K171A) known to abolish wave 1 neutralization, into the Du156 double mutant. None of these mutations significantly affected the titers of wave 3 neutralizing antibodies (Figure 8A – yellow curves), confirming that the epitope for wave 3 antibodies did not overlap with the wave 1 V2 epitope. An N332A mutation was also introduced to confirm that wave 3 antibodies did not target this glycan.
To test whether CAP257 wave 3 neutralizing antibodies targeted the MPER region of gp41, we coupled MPER peptides to magnetic beads and used them to adsorb out MPER specific binding antibodies in CAP257 plasma. The adsorption of MPER binding antibodies (Figure 8B) did not affect neutralization of Du156 or the double mutant Du156 N160K/T278A (with wave 1 and 2 resistance mutations) when compared to untreated plasma (Figure 8C). Thus, while we cannot exclude the possibility that wave 3 antibodies target V2 or gp41, these neutralizing antibodies appear to target an epitope distinct from any of the four known sites of vulnerability in the HIV-1 envelope.
We hypothesized that the selection pressure exerted by these broadly neutralizing antibodies would impact on the overall neutralization phenotype of CAP257 viruses over time. Therefore we tested the sensitivity of envelopes from 7, 30, 54, 93, and 174 weeks p.i. to broadly neutralizing monoclonal antibodies targeting the four major epitopes (Figure 9). The 7 week clone from CAP257 was sensitive to neutralization by anti-V2 antibodies, but following the development of wave 1 V2 neutralizing antibodies, CAP257 viruses became more resistant to PG9 and PGT145 neutralization. The 7 week clone was also highly sensitive to HJ16 (0.02 µg/mL) and VRC01 (1.79 µg/mL), however clones isolated after the emergence of wave 2 neutralization were increasingly resistant to neutralization by antibodies targeting the CD4bs. CAP257 viruses from all the time points selected were sensitive to neutralization by antibodies targeting the MPER or N301/N332, neither of which was targeted by broadly neutralizing antibodies in CAP257 plasma. These data confirm that CAP257 developed broadly neutralizing antibodies targeting the CD4bs and V2. Furthermore, the finding that CAP257 viruses remained sensitive to neutralization by antibodies targeting N301/N332 and the MPER supports our hypothesis that wave 3 antibodies target a novel epitope on the HIV-1 envelope.
A preventative HIV-1 vaccine remains the most likely way to end the HIV pandemic, but current envelope immunogens have so far failed to elicit broadly neutralizing antibodies. Nonetheless, the development of cross-reactive antibodies in approximately a quarter of HIV-1 infected individuals has confirmed that the human immune system can make such antibodies. Much emphasis has been placed on mapping the targets for these broadly neutralizing antibodies in an attempt to define viral vulnerabilities for immunogen design. Here we analyzed CAP257 heterologous neutralization over a 4.5 year period, describing the sequential evolution of three distinct broadly neutralizing antibody specificities within a single HIV-1 subtype C infected individual. We further showed how early viral evolution in the context of broadly reactive antibodies may profoundly shape the maturing antibody response towards enhanced neutralization breadth, in a process that may inform immunogen design. These data have been summarized in Figure 10.
The CAP257 autologous virus efficiently escaped all three specificities. As a consequence the antigenic stimulus for these broadly neutralizing antibodies declined, and antibody titers dropped at least ten fold within a three year period. The waxing and waning of the broadly neutralizing specificities in CAP257 confounded our previous attempts to map the targets at 174 weeks p.i. [7], when the titers of the three waves overlapped significantly. As most mapping studies are cross-sectional, the number of individuals who mount multiple broadly neutralizing antibody responses may therefore be underestimated, and may make up a significant proportion of those plasma samples that remain undefined.
Nonetheless we were able to finely map 2 of the 3 specificities in this study and showed that they targeted known sites of vulnerability on the HIV-1 envelope. Wave 1 antibodies targeted the site defined by PG9/16 and were completely dependent on N160 and K169, consistent with previous data describing PG9/16 dependence on the N160 glycan and the positively charged amino acids in the C-strand of V1/V2 [16], [37]. In general the epitope for wave 1 antibodies showed a larger footprint in V2 when compared to the epitopes of other monoclonal antibodies targeting this site, but behaved most similarly to PGT145. However none of the antibodies or plasma tested was sensitive to removal of the N156 glycan. While the crystal structure of PG9 with V2 showed interactions with the N156 glycan [37], the effect of deleting the N156 glycan is variable [16], [35], [36]. This effect might be explained by recent data suggesting that PG9 recognizes two N160 glycans (from two adjacent gp120 monomers) but only one N156 glycan [64]. The requirement for a lysine at position 169 explains the subtype C specificity of CAP257 wave 1 antibodies, as this residue is less common in subtypes A and B [35].
Wave 2 antibodies targeted a known site of vulnerability, the CD4bs, but these antibodies had an unusual glycan dependent mechanism of neutralization. CAP257 wave 2 and HJ16 neutralization were both highly dependent on interactions with the N276 glycan. N276 is also the recently described target of the broadly neutralizing antibody 8ANC195, but this antibody does not appear to interact with the CD4bs [65]. While glycan dependence for neutralization has not previously been described for CD4bs antibodies, including HJ16, Balla-Jhagjhoorsingh et al. reported that resistance to HJ16 involved an N276D mutation (deleting the glycan) with a hundred fold drop in titer [66], providing support for our observations. The glycan dependence of both HJ16 and CAP257 wave 2 antibodies suggests that they target a similar sub-epitope of the CD4bs that may be better defined as an N276 glycan dependent class of neutralizing antibodies, which is distinct from the VRC01 class. Recently it was shown that related variants of VRC01 do bind the glycan at N276 [67], however this glycan is not a major determinant of neutralization sensitivity to VRC01 [52], [56], [57], [68]. Rather, N276 has been described as a protective shield for the CD4bs, and deleting this glycan enhances the neutralization of CD4bs antibodies VRC01 and b12 [52], [56]. Removing N276 from gp120 also enabled binding to the predicted germline antibody for VRC01, which otherwise did not bind to gp120, and has been suggested as a modification for candidate vaccine immunogens [69], [70]. However, the glycan shield is increasingly recognized as a major site of vulnerability on the HIV-1 envelope [15], [16], and as with the glycans at N156/N160 and N301/N332, conservation of the N276 glycan bordering the CD4bs may make it a promising target for vaccine design.
Characterization of viral escape from CAP257 CD4bs antibodies indicated that deletion of the N276 glycan alone did not confer complete resistance. Escape required accumulating mutations in the CD4bs site, consistent with the functional conservation of this epitope. In addition to deletion of the N276 glycan, CAP257 escape occurred through a R456W mutation that also significantly affected neutralization by VRC01 (30 fold) and HJ16 (514 fold). This mutation likely contributed towards the evolution of VRC01 resistant virus by 174 weeks. W456 is extremely rare, occurring in only 0.78% (27 of 3,481) of sequences in the LANL HIV-1 sequence database. A crystal complex for HJ16 is not available, however the structure of VRC01 bound to its epitope showed that this antibody does not make significant contact with the R456 side chain in gp120, but rather hydrogen bonds with the R456 backbone carbonyl group. This suggests that the R456W mutation provides an indirect mechanism for resistance. The highly conserved R456 side chain can make hydrogen bonds with backbone carbonyl groups of amino acids at position 277 and 278 in the D-loop, as well as hydrogen bonds with E466 side chain in β24, C-terminal to V5 (Figure 11). Loss of these bonds and localized conformational changes to accommodate a bulky tryptophan residue may destabilize this critical component of the CD4bs epitope [52], [56], [71].
This study adds to data showing that the immune system can target multiple conserved epitopes [29]–[31]. It is striking that in three of these four studies, antibodies targeted both V2 and the CD4bs (donors CH219, AC053, and CAP257), suggesting an association between these two epitopes. Indeed, there is a well-documented relationship between V1/V2 and the CD4bs. The V1/V2 region protects the receptor binding sites from neutralization [42]–[51], and also interacts with V3 at the trimerization domain to hold the CD4bs in its pre-liganded conformation [21]. The crystal structures of monoclonal antibodies PG9, CH58, and CH59 bound to their epitopes in V2 show that the conformation of the V1/V2 sub-domain may vary significantly, but the factors that govern these conformational states are not known [37], [72]. In CAP257, deletion of the N160 glycan increased exposure of the CD4bs. It is possible that certain immunotypes of the V1/V2 epitope, such as the rare mutations at N160 or D167 described here, shifted the equilibrium of V1/V2 toward conformations that better exposed the CD4bs to neutralizing antibodies. This is supported by previous observations that introduction of the N160K and D167N mutations simultaneously into JR-FL resulted in a >50 fold increase in neutralization sensitivity to CD4bs antibodies [73]. CAP257 was infected with the less common N167 variant, and therefore following escape from V2 wave 1 antibodies a similar conformational state (D/S160, N167) would have been presented to the immune system. Our data suggest that this escape pattern further exposed the glycan dependent CD4bs sub-epitope. Although vaccination with the K160 and/or N167 immunotype may improve antibody responses to the CD4bs site, antibodies induced to the V2 epitope would be relatively strain-specific, like monoclonal antibody 2909 that recognizes the K160 immunotype and is therefore specific for SF162 [74]. In CAP257, switching from N167 to the more common D167 residue resulted in escape from the strain-specific response to V2, and this coincided with the development of a much broader response targeting the same epitope.
Sequential immunization may be a useful strategy to promote the broadening of the B-cell response. Recently Murphy et al. showed that two light chain variants paired with a single heavy chain of a strain-specific neutralizing antibody differentially neutralized early autologous envelopes [75]. While evolution of that strain-specific epitope did not affect the development of broadly neutralizing antibodies in this individual, the data supports the possibility that viral evolution might facilitate the neutralization of amino acid variants within a given epitope. Similarly, we have previously shown in an individual who developed broadly neutralizing antibodies to the V2 region that viral escape drove a maturation of the antibody response towards recognizing multiple V2 variants [35]. Here we show that emergence of an aspartic acid at position 279 preceded a broadening of the B-cell response to the CD4bs. The N279 and D279 amino acid variants of the CD4bs are equally common among sequences in the LANL HIV-1 database (50% and 48% respectively, n = 3,479), and preferential neutralization of either immunotype would halve the neutralization breadth of an antibody. CAP257 wave 2 neutralizing antibodies and HJ16 were somewhat sensitive to the N279D change. However the resistance of D279 containing viruses to CAP257 antibodies was rapidly lost after the emergence of the N279D escape mutation. Therefore, like the N167D mutation in V2 described for wave 1, this change in the fine specificity of CAP257 antibodies coincided with an increase in the neutralization breadth of the CD4bs response. We hypothesize that position N279 was non-critical for antibody binding, and despite temporarily facilitating neutralization escape, the N279D mutation then promoted affinity maturation by reducing the dependence of CAP257 antibodies on this amino acid. This adds to recent data suggesting a major role for viral evolution in the development of neutralization breadth to the CD4bs [76]. These data support the possibility that a sequential immunization strategy would enhance neutralization breadth by systematically presenting common variants in a given epitope. Such residues would have to be non-critical to antibody binding allowing for the evolution of higher affinity variants that would in turn recognize multiple immunotypes.
Overall these data highlight how interactions between the host immune system and viral escape mutations shaped the development of broadly neutralizing antibodies. The escape pathways identified here that led to the increased breadth of neutralization for both V2 and CD4bs antibodies provide potential pathways for generating broadly neutralizing antibodies. Further defining these pathways through the isolation of monoclonal antibodies will provide valuable insight into how these types of antibodies could be elicited using sequential immunization in a vaccine setting.
The CAPRISA Acute Infection study received ethical approval from the Universities of KwaZulu-Natal (E013/04), Cape Town (025/2004), and the Witwatersrand (MM040202). CAP257 provided written informed consent for study participation.
The CAPRISA Acute Infection cohort is comprised of women at high risk of HIV-1 infection in Kwa-Zulu Natal, South Africa [77]. Here we studied one individual (CAP257) from seven weeks p.i. through four and a half years of infection, until she started anti-retroviral therapy. During this time she had an average viral load of 60,784 copies/mL and an average CD4 count of 498 cells/µL. Plasma samples collected at 30 time points were used in this study.
The amplification of envelope genes from single HIV-1 RNA genomes has been previously described [78]. Viral RNA was isolated from CAP257 plasma using a Viral RNA Extraction Kit (QIAGEN), and cDNA was synthesized with Superscript III Reverse Transcriptase (Invitrogen). The reaction product was treated with RNase H (Invitrogen) and envelope genes were amplified by nested PCR with Platinum Taq (Invitrogen). Amplicons were purified with a PCR Clean-up Kit (QIAGEN) and single genome amplification confirmed by DNA sequencing using the ABI PRISM Big Dye Terminator Cycle Sequencing Ready Reaction kit (Applied Biosystems) and an ABI 3100 automated genetic analyzer. The full-length env sequences were assembled and edited using Sequencher v.4.5 (Genecodes) and Bioedit v.7.0.5.3.
The TZM-bl cell line engineered from CXCR4-positive HeLa cells to express CD4, CCR5, and a firefly luciferase reporter gene (under control of the HIV-1 LTR) was obtained from the NIH AIDS Research and Reference Reagent Program, Division of AIDS, NIAID, NIH (developed by Dr. John C. Kappes, and Dr. Xiaoyun Wu [79], [80]). The 293T cell line was obtained from Dr George Shaw (University of Alabama, Birmingham, AL). Cells were cultured at 37°C, 5% CO2 in DMEM containing 10% heat-inactivated fetal bovine serum (Gibco BRL Life Technologies) with 50 ug/ml gentamicin (Sigma) and disrupted at confluency by treatment with 0.25% trypsin in 1 mM EDTA (Sigma).
Selected envelope sequences were re-amplified from first round nested PCR products (described above) with PfuUltra II (Stratagene), purified by a Gel Extraction Kit (QIAGEN) and cloned into pcDNA3.1 (Invitrogen). The envelope plasmids were co-transfected into 293T cells using FuGENE 6 (Roche) with the pSG3ΔEnv backbone (obtained from the NIH AIDS Research and Reference Reagent Program, Division of AIDS, NIAID, NIH). Cultures were incubated for 48 hours to produce Env-pseudotyped viral stocks that were filtered through 0.45 µm and frozen in DMEM supplemented with 20% FBS. Mutant envelopes were generated with the QuikChange Lightning Kit (Stratagene) and confirmed by DNA sequencing (as above).
The TZM-bl neutralization assay has been described previously [4], [81]. It measures a reduction in relative light units generated by a single round of infection in TZM-bl cells with Env-pseudotyped viruses after pre-incubation with monoclonal antibodies or a plasma sample of interest. Samples were serially diluted 1∶3 and the ID50 calculated as the dilution at which the infection was reduced by 50%.
Plasmids encoding Histidine tagged recombinant envelope proteins were transfected into 293T cells using polyethylenimine 25 kDa (Polysciences). Recombinant proteins were expressed and purified as previously described [9]. Aliquots of 400 µg each were coupled to MyOne Tosyl-activated magnetic Dynabeads (Invitrogen) at 37°C pH 9.5 overnight, and then blocked with 0.5% BSA in 0.05% Tween20 PBS overnight at 37°C. Protein coupled beads were incubated with 200 µL of plasma (diluted 1∶20) for two hours at 37°C, then the beads were removed magnetically and the remaining plasma assessed for binding and neutralizing antibodies using ELISA and neutralization assays respectively.
Protein antigens were coated at 4 µg/mL onto high binding 96-well ELISA plates (Corning) overnight at 4°C. All subsequent steps were carried out in 5% fat-free milk, 0.05% Tween20 in PBS for 1 hour at 37°C. The plates were blocked and then probed with serial dilutions of the adsorbed plasma or specific monoclonal antibodies, biotinylated goat anti-human polyclonal antibodies (KPL), and an anti-biotin monoclonal conjugated to HRP (Calbiochem). Antigen-antibody complexes were detected by incubating with 100 µL 1-Step Ultra TMB-ELISA (Thermoscientific) for five minutes and then the reaction was stopped with 25 µL 1 M H2SO4. Absorbance was read at 450 nm on a VERSAmax tunable microplate reader (Molecular devices).
Plasmid encoding the RHPA gp120 core was transfected with polyethylenimine-MAX 40 kDa (Polysciences) into GnTI(−/−) 293S cells and purified using two step lectin chromatography and Ni-NTA affinity chromatography. Protein was assessed for purity and conformation by SDS-PAGE and ELISA. 1 mg of the RHPA core gp120 was deglycosylated overnight at 37°C in 500 mM sodium chloride, 100 mM sodium acetate pH 5.5 with 0.5 U of Endo-H. Glycans were removed through buffer exchange into PBS using Vivaspin 20 mL concentrators (Sartorius stedim). Deglycosylation was confirmed by SDS-PAGE and sandwich ELISA (described above) using lectins as a capture protein.
|
10.1371/journal.pntd.0003449 | Severity of Old World Cutaneous Leishmaniasis Is Influenced by Previous Exposure to Sandfly Bites in Saudi Arabia | The sandfly Phlebotomus papatasi is the vector of Leishmania major, the main causative agent of Old World cutaneous leishmaniasis (CL) in Saudi Arabia. Sandflies inject saliva while feeding and the salivary protein PpSP32 was previously shown to be a biomarker for bite exposure. Here we used recombinant PpSP32 to evaluate human exposure to Ph. papatasi bites, and study the association between antibody response to saliva and CL in endemic areas in Saudi Arabia.
In this observational study, anti-PpSP32 antibodies, as indicators of exposure to sandfly bites, were measured in sera from healthy individuals and patients from endemic regions in Saudi Arabia with active and cured CL. Ph. papatasi was identified as the primary CL vector in the study area. Anti-PpSP32 antibody levels were significantly higher in CL patients presenting active infections from all geographical regions compared to CL cured and healthy individuals. Furthermore, higher anti-PpSP32 antibody levels correlated with the prevalence and type of CL lesions (nodular vs. papular) observed in patients, especially non-local construction workers.
Our findings suggest a possible correlation between the type of immunity generated by the exposure to sandfly bites and disease outcome.
| Leishmania is transmitted by the bite of infected female sandflies. When a sandfly bites a vertebrate host, it injects a cocktail of salivary proteins meant to facilitate blood feeding. The constant exposure to sandfly bites in endemic areas triggers a humoral response against the major antigenic components in the saliva. These antibodies can be then exploited to measure exposure to vector sandflies, which is useful for surveillance in leishmaniasis control programmes. In Saudi Arabia, cutaneous leishmaniasis (CL) is mainly transmitted by the Phlebotomus papatasi sandfly. Here we study the recognition of the main antigenic salivary protein from Ph. papatasi, PpSP32, in leishmaniasis patients and healthy individuals from three CL endemic areas in Saudi Arabia. Anti-PpSP32 antibody levels were significantly higher in CL patients presenting active infections from all geographical regions compared to the CL-cured and healthy individuals. Furthermore, higher anti-PpSP32 antibody levels correlated with the prevalence and type of CL lesions observed in patients. Our results suggest that previous long-term exposure to sandfly saliva can have a role in modulating the severity of leishmaniasis infection, resulting in a milder form of the disease.
| Cutaneous leishmaniasis (CL) in Saudi Arabia is an increasing public health problem due to rapid urbanization, intensive agriculture and human migration [1]. Zoonotic CL (ZCL) is the most prevalent form of leishmaniasis in the country, which is caused by Leishmania major and transmitted by the sandfly Phlebotomus papatasi. Leishmania tropica on the other hand is exclusively endemic to the South Western region [2], where it is transmitted by Ph. sergenti and causes anthroponotic CL (ACL).
The saliva that sandflies inject into their vertebrate host impairs the haemostatic and inflammatory systems allowing the insects to efficiently take a blood meal [3]. These salivary components were also shown to promote or inhibit the development of Leishmania in the vertebrate host [4]. Increased sandfly-host contact translates into an increased risk of being infected. Repeated exposure to sandfly bites produces antibodies against its salivary components in the host, providing an indirect measure of exposure to vectors [5]. The presence of IgG antibodies against Ph. papatasi saliva has been associated with a higher risk of being infected with L. major [4,6]. The transient nature of the antibody response to sandfly bites [6–10] allows for the study of temporal changes in transmission risk and the efficacy of vector control programmes [11].
Biomarkers used to evaluate sandfly exposure need to be species-specific in order to differentiate between antibody responses to vector and non-vector species, or between sandflies and other blood-feeding insects including mosquitoes. The sandfly salivary protein PpSP32 has been described as a 30 kDa immunodominant target of the host antibody response against Ph. papatasi saliva [12,13], and was highly specific when tested against individuals living in a region with high prevalence of Ph. perniciosus. Additionally, expression of the PpSP32 salivary transcript is not influenced by age or diet of the sandfly [14]. B-cell epitope prediction analysis showed six epitopes were identical between the Tunisian PpSP32 and the PpSP32 protein deposited in GenBank (Israeli strain), indicating it is a good candidate to assess biting exposure in different ZCL foci [13]. Furthermore, the production of rPpSP32, a recombinant form of the Ph. papatasi PpSP32 protein, overcomes the difficulty of obtaining large quantities of salivary glands, and facilitates the use of salivary biomarkers for large scale epidemiological studies in endemic areas.
To better understand the correlation between sandfly biting exposure and leishmaniasis infection, we determined the level of exposure to Ph. papatasi bites in individuals from several CL endemic areas in Saudi Arabia by measuring the levels of anti-PpSP32 antibodies present in the sera of patients and healthy volunteers.
The study was approved by the Liverpool School of Tropical Medicine Ethics Committee UK (12.03RS). All participants provided written informed consent for the collection of blood samples and subsequent analyses. All research was conducted according to Declaration of Helsinki principles.
Peripheral blood samples were obtained from 411 individuals (106 females and 305 males, aged 18–60 years, median of 36 years) living in two ZCL (Al Ahsa and Al Madinah) and one ACL (Asir) endemic areas in Saudi Arabia (S1 Table). Study sites were chosen to include areas were patients would be exposed to the bite of Ph. papatasi (ZCL transmission) or Ph. sergenti (ACL transmission) (Al Salem et al, 2014. Submitted) to test the specificity of the biomarker. Samples were collected during the months of April and December 2012. Cases were diagnosed through parasitological confirmation of Leishmania by a trained clinician, and the infecting Leishmania species was confirmed in patients with both active and cured infections (through clinical history). Clinical cure was signified by successful re-epithelialisation of the lesion(s) after treatment.
Donor sera were classified as healthy (no history of leishmaniasis infection), ZCL (L. major) or ACL (L. tropica) patients with either active or cured CL. An additional 80 serum samples of patients with active infection from Al Ahsa were used for the analysis of local versus non-local exposure; although these were likely to be infections with L. major, they are unconfirmed and therefore considered separately. We used sera from five United Kingdom residents as non-endemic controls. These healthy volunteer donors have no history of leishmaniasis or travelling to sandfly endemic areas.
Mammalian VR-2001 plasmid coding the PpSP32 protein with 6 histidine tag was sent to the Protein Expression Laboratory at the Frederick National Laboratory for Cancer Research (Frederick, Maryland). Expression was performed by transfecting HEK-293F cells. The supernatant was collected after 72 hours, filtered and concentrated from 1 litre to 300 ml using an Amicon concentrator device (Millipore, Billerica, MA, USA) in the presence of NaCl 500mM. The volume was returned to 1 litre at a final concentration of 10 mM Tris, pH 8.0. The expressed protein was purified by an HPLC system (DIONEX, CA, USA) using two 5 ml HiTrap Chelating HP columns (GE Healthcare, Buckinghamshire, UK) in tandem and charged with 0.1 M NiSO4. The protein was detected at 280 nm and eluted by an imidazole gradient as described by Teixeira et al. [15]. Eluted proteins were collected every minute in a 96-well microtiter plate using a Foxy 200 fraction collector (Teledyne ISCO, Lincoln, NE, USA). Fractions corresponding to eluted proteins peaks were selected and run on a NuPage Bis-Tris 4–12% Gel (Novex, Life Technologies, Carlsbad, CA, USA) with MES running buffer under reducing conditions as per manufacturer’s instructions. Appropriate fractions, as determined by molecular weight were pooled and concentrated to 1 ml using an Amicon Ultra Centrifugal Filter (Millipore, Billerica, MA, USA). Protein concentration was measured using a NanoDrop ND-1000 (Thermo Scientific, Waltham, MA, USA) spectrophotometer at 280 nm and calculated using the extinction coefficient of the protein.
Exposure to sandfly bites was measured through the levels of anti-PpSP32 IgG antibodies in the sera of participants. Anti-PpSP32 antibodies were measured by ELISA (Enzyme-Linked Immunosorbent Assay), as described by Marzouki et al. [13] with some modifications. Briefly, microtiter plates (Thermo-Scientific) were coated overnight with 50 μl of PpSP32 (2 mg/ml = 0.1 mg/well) in 0.1M carbonate buffer (pH 9.6). Plates were blocked with PBS-BSA at 37°C for one hour and then washed several times with PBS. Diluted sera (1:200) were added to the plates and incubated at 37°C for 2 hours. After washing, plates were incubated with anti-human IgG peroxidase-conjugated antibody (1:10000) (Jackson ImmunoResearch, Suffolk, UK) for one hour at 37°C. Antibody binding was visualized using the substrate, 3,3′,5,5′ tetramethylbenzidine (Biolegend, San Diego, CA, USA), and absorbance was read at 450 nm on a Fluorostar Omega microplate reader (BMG Labtech, Ortenberg, Germany). Each serum was tested in triplicate. Wells without serum were used as negative controls.
To determine the relative abundance of vector species in each of the endemic areas, sandfly collection was conducted between March and November of 2012. Adult sandflies were collected using CDC light traps placed from 6:00pm to 6:00am in the peridomicile of houses, including sheds harboring domestic animals such as chickens and rabbits. Sticky traps were used to capture sandflies in rodent burrows. Sandflies were preserved in 70% alcohol and identified to species [16].
Software ArcGIS 10 (ESRI, Redlands CA) was used to show the presence of vector species.
The Kruskal-Wallis test was used to compare sets of groups. GraphPad Prism Software 5 was used for all data analysis. Statistical significance was considered as P<0.05.
In the regions of Al Ahsa and Al Madinah, ∼99% of sandflies were identified as Ph. papatasi, with the additional presence of a few Ph. bergeroti (∼1%) in Al Madinah (Table 1). The Southern region of Asir showed the highest diversity of vector species; Ph. sergenti was the most abundant (21%), followed by Ph. bergeroti (10%). Although Sergentomyia species (of non-medical importance) represented only a small percentage (∼1%) in Al Ahsa and Al Madinah, they constituted over half of the specimens identified in Asir (67%). The predominant presence of Ph. papatasi in both Al Ahsa and Al Madinah, and of Ph. sergenti in Asir, is in agreement with the prevalence of infections caused by L. major and L. tropica, respectively (Fig. 1).
We found that the levels of anti-PpSP32 antibodies in the sera of healthy individuals from Saudi Arabia were significantly higher (P≤0.01) (S1 Fig.) when compared to unexposed individuals from the UK. This indicates the biomarker is successfully recognized by Saudi individuals, and furthermore agrees with the expected level of exposure to sandflies in CL-endemic areas.
When we compared healthy individuals from the two ZCL endemic regions studied, there was a significantly higher level of anti-PpSP32 antibodies in Al Ahsa compared to Al Madinah (Fig. 2A). To test for a possible correlation between exposure to sandfly bites and leishmaniasis infection, we compared healthy and infected individuals. In both Al Ahsa (Fig. 2B) and Al Madinah (Fig. 2C), patients with an active infection (CL) showed significantly higher levels of anti-PpSP32 antibodies compared to healthy residents (P<0.001). Overall, comparing the groups from both Al Ahsa and Al Madinah, the levels of anti-PpSP32 in Al Ahsa individuals appear to be higher than those from Al Madinah, suggesting that Al Ahsa populations are more exposed to Ph. papatasi bites.
In individuals from the region of Asir (endemic for ACL L. tropica infections), both the healthy and cured groups showed very low levels of anti-PpSP32 antibodies (Fig. 3), which agrees with the near absence of Ph. papatasi from this region (Table 1). Unexpectedly, the levels of anti-PpSP32 antibodies were significantly higher (P<0.01) in individuals with an active L. tropica infection, compared to healthy residents and cured patients (Fig. 3). Sequence alignment of the Ph. papatasi PpSP32 [17] and the PpSP32-like protein from Ph. sergenti [18] confirmed a significant level of similarity between these homologous proteins (S2 Fig.), suggesting cross-reactivity. Although these patients were Saudi residents and their migration is uncommon, we cannot discard either the possibility that these individuals might have been exposed to Ph. papatasi bites while traveling outside this area, or the presence of Ph. papatasi in low numbers. In both cases, the anti-PpSP32 levels may reflect a low exposure to this sandfly species.
To test for a correlation between exposure to sandfly bites and the clinical presentations of L. major infection in human patients, we compared the levels of anti-PpSP32 antibodies in patients presenting nodular, papular or ulcerated-nodular lesions. Of the three, nodular lesions and then papular are the least severe; both of these lesion types can progress to the more severe ulcerated-nodular form. ZCL patients from Al Madinah with nodular and ulcerated nodular type lesions have higher levels of anti-PpSP32 than those with papular type lesions (Fig. 4A), but a statistical difference was only observed between papular and nodular lesions (P<0.01). There were no significant differences in anti-PpSP32 levels between different types of lesions in Al Ahsa patients (Fig. 4B).
We also looked at the levels of anti-PpSP32 in ZCL patients according to the lesion characteristics. Lesion size was classified as being either 10–15mm or >15mm. Patients from Al Ahsa with large lesions >15mm had significantly higher antibody levels (P<0.01) than individuals with lesions between 10–15mm (Fig. 4C). This difference was not observed in Al Madinah. Additionally, when we compared the patients with different lesion numbers (< 3 or > 3 lesions) (Fig. 4D), no significant differences in antibody levels were found within each region. However, the same figure shows the difference in anti-PpSP32 levels was significant, with higher levels in Al Ahsa than Al Madinah.
In Al Ahsa, we found that non-local patients (visiting labour) had significantly higher levels (P<0.001) of anti-PpSP32 compared to the local residents (Fig. 5A). Interestingly, nearly three quarters of the non-local patients developed more than three lesions compared to only 40% in the local group (Fig. 5B). Although such differences did not correlate with the levels of anti-PpSP32 (S3 Fig.), patients from the visiting labour group presented in general a higher number of lesions compared to the residents (S2 Table).
Antibodies to sandfly saliva can be used to indicate disease risk in endemic areas [4,6,12,19], and the development of biomarkers for this purpose depends on the discovery of highly conserved yet species-specific molecules. SP32-like proteins are unique to sandflies and occur in all species studied to date [18]. Among these, PpSP32 is a highly immunogenic protein isolated from the saliva of Ph. papatasi that serves as a biomarker for vector exposure [13]. Data obtained from a CL-endemic area in Tunisia showed that the human antibody response to PpSP32 is representative of the humoral response against whole salivary gland extract [6]. Here, we used a recombinant form of this protein to evaluate the level of exposure to sandfly saliva in three endemic areas in Saudi Arabia. Our results show that the severity of human CL pathology appears to be influenced by previous exposure to sandfly bites.
The migration of non-immune people into leishmaniasis endemic areas has been well documented to affect groups such as civilian workers and military personnel [20,21], resulting in leishmaniasis outbreaks [22]. Evaluation of biting exposure can be useful for assessing disease risk of such populations in Saudi Arabia. The higher serum levels of anti-saliva antibodies in the visiting workers compared to the long-term residents of Al Ahsa suggest the migrant population is highly exposed to sandfly bites and less immune to CL. Residents have a lower (but continuous and long-term) exposure to bites, which might induce desensitisation (tolerance) to sandfly saliva, thus explaining their lower antibody levels compared to the non-locals. This desensitization after long term exposure has been previously observed in mice models [23]. Moreover, the residents seem to suffer less severe leishmaniasis lesions. Exposure to uninfected bites of Ph. papatasi has been shown to be protective against L. major in mice [24] and whether the same level of protection is conferred to humans in CL-endemic areas remains to be determined. Non-locals typically work and dwell closer to sandfly habitats like the burrows of rodents (reservoirs of disease) and are consequently plagued by biting sandflies. Previously unexposed to this level of biting, they showed a more intense antibody response over a shorter period of time. The high exposure to sandfly bites might increase susceptibility to infection and severe clinical outcomes, as nearly three quarters of them developed multiple lesions. Other factors such as genetic background can also influence susceptibility to disease [25]; however, this is unlikely in this situation as the visitors originate from eight different countries, mainly from Middle East, Southern Asia and Africa.
Interestingly, CL patients from both ZCL regions (Al Ahsa and Al Madinah) exhibited even higher levels of anti-PpSP32 antibodies compared to healthy residents from their respective areas. Marzouki et al. [6] previously investigated this relationship using whole salivary gland extract and associated the significantly higher antibody levels in ZCL patients with increased risk of developing CL. This difference was also reported for ACL [12], where exposure to Ph. sergenti bites was evaluated in both healthy individuals and patients with L. tropica. Similarly, ACL patients produced a significantly higher IgG response compared to healthy people from the same area, likewise supporting the relationship between exposure and leishmaniasis infection. B-cell clonal expansion, which increases production of non-specific antibodies in some parasitic infections [26], could be an alternative explanation to an increased antibody response in CL patients; however, this has only been reported in visceralizing forms of leishmaniasis [27,28].
Our research identified the sandfly species inhabiting the three CL endemic areas in order to complement the data obtained on bite exposure. In agreement with the anti-PpSP32 levels in patient sera, the majority of sandflies found in Al Ahsa and Al Madinah were identified as Ph. papatasi. Other sandfly species identified belong to the Sergentomyia genus, whose members rarely bite humans (they are mostly zoophilic) and have been shown to be refractory to Leishmania species pathogenic to humans [29] Ph. papatasi accounts for most, if not all, of the bites sustained by individuals in the ZCL areas. This was further supported by finding significant levels of anti-PpSP32 antibodies in healthy donors of these regions compared to UK control sera. However, anti-PpSP32 antibodies were significantly higher in Al Ahsa, suggesting a higher exposure to Ph. papatasi in this region.
Unexpectedly, we found that sera of L. tropica patients from the Southwest region of Asir (where Ph. sergenti is the predominant CL vector) also recognized PpSP32, although levels were much lower compared to ZCL patients. This could be due to a cross reaction with salivary proteins from Ph. sergenti. In fact, there is a high degree of similarity (52%) between Ph. sergenti SP32-like protein and Ph. papatasi SP32. In mice exposed to Ph. sergenti bites, a partial cross-reactivity to Ph. papatasi whole salivary gland homogenate was reported [12,30]. A similar level of cross-reactivity could also be present between salivary proteins from Ph. papatasi and Ph. bergeroti [31] (the second most abundant species in Asir).
Is there a correlation between CL clinical forms and exposure to sandfly bites? We detected higher levels of anti-PpSP32 antibodies in patients with nodular-type lesions compared to those with papular lesions in Al Madinah, but not in Al Ahsa. This differential response could be attributed to a) the genetic background of the infected patients, b) a cumulative exposure to sandfly bites or c) the parasite strains found in each area. It would be interesting to further study how the interaction between these factors affects the immune responses to salivary proteins and disease pathology.
The immune response elicited by sandfly salivary proteins and how it modulates the Leishmania infection, varies depending on the vector species and vertebrate host [32]. Some reports have shown that sandfly saliva is able to preferentially trigger a protective Type I delayed-type hypersensitivity response [33–35]. In animal models a Th1 response to salivary proteins is correlated with protection against CL, and immunization with single proteins from sandfly saliva conferred protection against a L. major infection when animals were challenged with infectious Ph. papatasi bites [35–37]. On the other hand, a Th2 response (and antibodies to salivary proteins) correlates with higher susceptibility and in some cases exacerbation of the disease [38,39]. Furthermore, individuals living in a CL endemic region of Tunisia, where the main vector is Ph. papatasi, developed a mixed response with a dominance of Type II immunity [40]. It may possible that subjects that develop antibodies (in a Th2 environment) to PpSP32 (and perhaps other salivary proteins) may be more susceptible to CL. It would be relevant to characterize the immune response(s) in individuals with different clinical presentations and from different geographical locations.
In summary, the use of recombinant salivary proteins can help us understand the impacts of natural exposure to sandflies in leishmaniasis endemic areas [3]. Our results provide insights into the relationship between the human antibody response to sandfly saliva and development of cutaneous leishmaniasis in different transmission contexts. In addition, they support the use of biomarkers as epidemiological tools to improve the surveillance of human-vector contact and disease transmission.
Protein accession numbers (NCBI): Phlebotomus papatasi SP32 GI:449060662, Phlebotomus sergenti SP44: GI:299829437
|
10.1371/journal.pntd.0003541 | Low-level Laser Therapy to the Mouse Femur Enhances the Fungicidal Response of Neutrophils against Paracoccidioides brasiliensis | Neutrophils (PMN) play a central role in host defense against the neglected fungal infection paracoccidioidomycosis (PCM), which is caused by the dimorphic fungus Paracoccidioides brasiliensis (Pb). PCM is of major importance, especially in Latin America, and its treatment relies on the use of antifungal drugs. However, the course of treatment is lengthy, leading to side effects and even development of fungal resistance. The goal of the study was to use low-level laser therapy (LLLT) to stimulate PMN to fight Pb in vivo. Swiss mice with subcutaneous air pouches were inoculated with a virulent strain of Pb or fungal cell wall components (Zymosan), and then received LLLT (780 nm; 50 mW; 12.5 J/cm2; 30 seconds per point, giving a total energy of 0.5 J per point) on alternate days at two points on each hind leg. The aim was to reach the bone marrow in the femur with light. Non-irradiated animals were used as controls. The number and viability of the PMN that migrated to the inoculation site was assessed, as well as their ability to synthesize proteins, produce reactive oxygen species (ROS) and their fungicidal activity. The highly pure PMN populations obtained after 10 days of infection were also subsequently cultured in the presence of Pb for trials of protein production, evaluation of mitochondrial activity, ROS production and quantification of viable fungi growth. PMN from mice that received LLLT were more active metabolically, had higher fungicidal activity against Pb in vivo and also in vitro. The kinetics of neutrophil protein production also correlated with a more activated state. LLLT may be a safe and non-invasive approach to deal with PCM infection.
| PCM triggers a typical granulomatous inflammatory reaction with PMN playing a major role; these inflammatory cells are crucial in the initial stages of PCM, participating in the innate immune reaction and also directing the acquired immune response in the later stages. In some PCM patients, these immune mechanisms are insufficient to eradicate the infection, and need to be boosted with antifungal drugs that have to be administered for long periods and can show serious side-effects. We aimed to develop a novel and safe way to activate PMN through low-level laser irradiation of the bone marrow in the mouse femoral medulla. LLLT increased PMN viability and activation, shown by a significantly greater production of protein and ROS, as well as a higher fungicidal capacity; PMN even retained their higher metabolic activity and fungicidal ability after a second exposure to the pathogenic fungus in vitro. This is the first time that LLLT has been shown to increase the immune response against a fungal infection, and could be a promising and safe technique to be used with antifungal drugs in PCM.
| Paracoccidioides brasiliensis (Pb) is a non-sexual thermodimorphic fungus that exists in either a mycelium or a yeast form; the latter being pathogenic to humans and can cause an important and neglected systemic infection called paracoccidioidomycosis (PCM). The likelihood of infection and its severity depends on the amount of inhaled fungi as well as the immunological status of the individual [1]. Patients with immune suppression or defects in immune cell activation are more susceptible to PCM [2,3].
PCM presents as a primary acute infection that is later transformed to a chronic phase. However, regardless of the stage of the disease, inflammatory cells play a central role in fighting Pb, particularly the neutrophils or polymorphonuclear cells (PMN) [4]. Besides the production of several direct antimicrobial factors, PMN may also secrete cytokines, chemokines and growth factors [5] that promote the host response against the infection. PMN are not only critical for the innate immune response, but can also help the adaptive immune response by interacting with B lymphocytes [6], T cells [7] and dendritic cells [8].
Previous studies have reported prominent neutrophilic infiltrates in paracoccidioidomycotic lesions in experimental animal models such as hamsters [9,10], rats [11] and also in tissue samples from patients [12]. Along with macrophages and plasmocytes, PMN are conspicuous in PCM granulomatous lesions and lead to altered morphology of the nearby fungal cells [13]. The immunological defense against fungi relies on the interaction between specific components of the fungal cell (pattern-associated molecular patterns or PAMPs) and pattern recognition receptors (PRRs) on host phagocytes. Through the binding of PAMPs to PRRs, a signaling cascade is initiated leading to release of pro- and anti-inflammatory cytokines linked to phagocytosis and intracellular fungal cell killing [14].
PMN can also help to eradicate pathogens via phagocytosis and the generation of reactive oxygen species (ROS) during the respiratory burst [15]. Nevertheless, despite the crucial role of inflammatory cells, they are usually not sufficient to entirely eliminate the Pb on their own, and patients usually need additional antifungal drug therapy [16]. Itraconazole, for instance, is effective in treatment of PCM, although its use may allow the relapse of the disease several months after discontinuation of the drug therapy [17].
Antifungal medication can also lead to diverse side-effects including dizziness, headaches, epigastric pain [16] and, more importantly, to the development of drug-resistance in the targeted microorganisms [18]. Therefore we asked whether there could be a novel way of activating PMN through the safe and non-invasive technique of low-level laser therapy (LLLT). LLLT uses non-thermal and non-ionizing light irradiation that has been successfully used for acceleration of healing as well as reduction of pain and inflammation [19–21].
Although LLLT may often work as an anti-inflammatory modality [22], it can, depending on the parameters, also trigger the activation of immune cells [23,24] and the activation of pro-inflammatory pathways [25]. While the activation of PMN by LLLT is not a completely novel process and has been reported in vitro [26,27], the use of LLLT to help the organism to combat PCM is a new idea; thus, we aimed to assess the fungicidal capacity of PMN after LLLT by characterizing these cells on secretory protein levels, mitochondrial activity and ROS discharge following a first and second exposure to Pb.
This research was carried out in accordance with the ethical principles required for animal experimentation and was approved by the Ethics Committee on Animal Research of the Federal University of Alfenas, under the protocol registration No. 477/2012. The animal procedures were conducted in accordance with the guidelines with animal care and use committee at Brazil`s National Council for the Control of Animal Experimentation.
Swiss outbred female mice were kept in controlled temperature rooms and fed with sterile food and distilled water ad libitum. The animals were kept under a 12 light/12 dark cycle, and it was ensured that personnel did not enter the mouse facilities during the dark cycle.
Isolates of the highly virulent Paracococcidioides brasiliensis Pb18 strain [28] were grown in semi-solid culture of Fava Netto [29], with the culture medium replaced routinely every 7 days. A polysaccharide preparation known as Zymosan, derived from cell walls of the yeast Saccharomyces cerevisiae and containing β-D-glucan was commercially obtained (Sigma-Aldrich, St. Louis, MO, USA).
Pb18 cells or Zymosan were washed with sterile 0.9% saline solution and centrifuged (5810R Centrifuge, Eppendorf, NY, USA) 3X at 1300g. A fungal suspension containing 5x107 yeast cells/ml was measured using a cell viability count after staining by the vital dye Janus Green B [30] and a hemocytometer. Zymosan particles were directly counted by hemocytometer.
At 6 weeks of age and weighing approximately 25g, the animals received an “air pouch” as described by Harmsen and Havell in 1990 [31] and modified by Meloni-Bruneri et al. in 1996 [32]. An air pouch was produced in the dorsal region of mice by a subcutaneous injection of 2 ml of air; then, 0.1 ml of either the fungal suspension, Zymosan or saline was subsequently injected in the same region. It was previously shown by our group that P. brasiliensis elicits a marked neutrophil recruitment in vivo after air-pouch inoculation of the virulent Pb18 in mice; the mechanism behind this cell recruitment is probably due to chemotactic factors produced by the fungi and injured tissue [32]. In order to show that the PMN recruitment was truly invoked by the fungal cells or its derivatives and not by the air-pouch procedure itself, two additional groups were created and consisted of saline solution inoculation either followed or not by LLLT.
The animals were divided into four groups, namely, group 1: animals infected with Pb18 and light irradiated; group 2: animals infected with Pb18 but not irradiated; group 3: animals inoculated with Zymosan and light irradiated; and group 4: animals inoculated with Zymosan and not irradiated.
LLLT was performed on two points on each hind leg; the laser device used was a Twin flex laser (MMO, São Carlos, SP—Brazil) with a spot size of 0.04 cm2. The laser parameters were: continuous wave near-infrared light (780nm) to deliver 12.5 J/cm2 with a 50 mW total power; the total energy was 0.5 J per point (30 seconds per point). Our goal was to reach the bone marrow of the femoral bones, where the process of blood cell formation, known as hematopoiesis, including neutrophils is originated [33]. LLLT was performed on alternate days, with the animals first irradiated immediately after infection and last just before the neutrophil collection. In that way, the animals were irradiated on day 0 (infection or inoculation); day 2; day 4; day 6; day 8; and day 10 (collection of PMN); thus, 6 irradiations were performed.
PMN were collected 10 days after the infection or the inoculation of the mice. The animals were anesthetized with a lethal dose (0.5 ml of a 10% ketamine hydrochloride and 2% Xylazine solution); after a skin flap procedure was performed, the cells were collected and placed in sterile tubes with the help of a sterile glass Pasteur pipette and were subsequently dissociated by pipetting. The cells were then transferred and stored in Falcon tubes containing RPMI (Sigma-Aldrich, St. Louis, MO, USA) with 10% Fetal Bovine Serum (FBS—Sigma) and were kept refrigerated (2–6°C) to be used for the subsequent experiments described below. The cells were quantified using a hemocytometer and the cell viability was assessed with 0.2% Trypan blue (Sigma).
For the fungal co-culture experiment with PMN, the refrigerated cells were centrifuged at 1780g and washed once before suspension in 15 ml of RPMI; then, the cells were quantified in a hemocytometer and viability was assessed with Trypan blue. The final concentration was adjusted to 106 PMN/ml. Pb cells were 3X washed with sterile 0.9% saline and centrifuged at 1300g and re-suspended in RPMI with 10% FBS. The concentration of the suspensions was adjusted according to the concentration of the obtained phagocytic cells, so that the cultures remained in a proportion of 1 Pb to 25 PMN to be further utilized for the evaluation of PMN metabolic activity, ROS quantification and quantification of viable fungi. Cells were counted in an hemocytometer and the Pb viability was determined by the staining with Janus Green B vital dye [30].
After adjusting the PMN suspension (106 PMN/ml), and the Pb fungal suspension (4x104 cells/ml) to provide the co-cultivation mixture (1ml of each suspension), which was added to 12 well plates (Corning, New York, USA), the plates were incubated at 5% CO2 and 37°C for 2, 6 and 18 hours. After incubation, the cells were centrifuged at 1780g and the PMN pellets had their viability assessed by 0.2% Trypan Blue staining.
In a 96 well plate (Corning) we added 100 μl of a 106 Pb18/ml suspension and 100 μl of a 5x106 PMN/ml suspension maintaining a ratio of 1:5 (Pb:PMN). The experiment was performed in triplicate. After 2 hours of incubation (5% CO2 and 37°C) we added 20 μl of MTT (Sigma) to the wells. The plate was further incubated for 4 hours. The supernatant was removed, leaving only the pellet at the bottom of each well. Then, 200μL of DMSO (Sigma) was added to each well and the plate was read in a microplate reader at 540nm (Anthos Zenyth 200, Biochrom, Cambridge, UK).
The BCA method (Sigma) allows colorimetric detection and quantification of the total level of protein in a solution. This method combines the reduction of Cu2+ to Cu+ by protein in an alkaline medium (the Biuret reaction) with highly sensitive and selective colorimetric detection of the Cu+ ion using a reagent containing bicinchoninic acid [34]. The assays were performed in triplicate and the optical densities were measured in a microplate reader (Biochrom) at a wavelength of 560 nm. The results were expressed in mg of protein/ml, comparing the optical density with a standard curve containing known concentrations of bovine serum albumin (BSA—Sigma). The calibration curve was made with a BSA solution of 10 μg/ml at 6 different protein concentrations: 10; 5; 2.5; 1.25; 0.67 and 0.33 μg/ml. The total protein concentration of each sample was calculated by pipetting 50μl of previously disrupted cells (ultrasonic method) along with 200μl of BCA. All samples were pipetted in triplicate and the results corresponded to the mean of the values obtained after blank (RPMI medium) subtraction for PMN cultured in vitro or co-cultivated with Pb18.
The quantification of reactive oxygen species produced by the PMN oxidative “burst” was carried out by the luminol chemiluminescence assay. PMN were obtained from the experimental groups and adjusted to a suspension of 1x106 PMN cells/ml; for the co-cultivation experiments PMN cells were adjusted to the proportion of 1 Pb to 25 PMN (Pb concentration 4x104 cells/ml, PMN 1x106 cells/ml). Luminol (Sigma) was used as the substrate for this assay; 135 μl of the PMN suspension was added into a cuvette along with 30 μl of luminol; followed, for the co-cultivation experimental groups, by 135μl of the Pb18 suspension. A luminometer (Glomax 20/20 Luminometer, Promega, USA) was used to measure the chemiluminescence signal over 30 minutes. Positive (PMA—phorbol myristate acetate, Santa-Cruz, Brazil) and negative (DPI—diphenyleneiodonium, Sigma) controls were employed.
The material collected from the subcutaneous air-pouches was immediately centrifuged at 1780g (5810R Centrifuge, Eppendorf, NY, USA). The pellets were re-suspended in 100μl PBS, and spread on Petri dishes with the aid of a sterile Drigalski spreader. Similarly, after centrifugation at 1780g, 100μl of PMN/Pb mixed suspensions obtained after 2 hours of co-cultivation were spread on Petri dishes. The experiments were performed in triplicate. The fungal growth on plates was allowed to take place over a period of 12 days, when a paintbrush marker was used to highlight the colonies. The culture medium used in this procedure was BHI agar (HiMedia Laboratories, India) supplemented with 1% glucose, 30% growth factor mixture produced by the fungus itself and 10% FBS, as described by Singer-Vermes et al. in 1992 [35].
The results were analyzed using the Shapiro-Wilk normality test and were all considered to have a normal distribution. Groups were compared using a Student`s T test with the level of significance set at 5%. The software used for the analyses was Graph-Pad Prism 6 (GraphPad Software, Inc; La Jolla, CA 92037, USA).
The animals inoculated with saline showed no neutrophils at the site of infection even after 10 days (S1 Fig.), which clearly showed that neither the air-pouch procedure alone nor the laser irradiation alone was responsible for the PMN recruitment. The PMN produced by the inflammatory stimuli (either Pb18 infection or Zymosan inoculation) were harvested from the subcutaneous air-pouches (Fig. 1), and whilst the total number of PMN recruited to these air pouches was significantly diminished (p = 0.0001) when LLLT was used after the Pb infection, the number of PMN was significantly increased (p = 0.0001) when LLLT was used after mice were inoculated with Zymosan (Fig. 2).
Interestingly, the kinetic study of PMN cell viability showed that LLLT was able to sustain a more viable population of neutrophils for the 18-hour time course both after Pb infection (Fig. 3A) at 6 hours (p = 0.0278) and also after Zymosan inoculation (Fig. 3B) at 2 hours (p = 0.0274). There was no statistical significant difference between the viability of the PMN from irradiated or non-irradiated mice after co-cultivation with Pb for up to 18 hours, though the viability of the irradiated cells was kept at high levels, similarly to the non-irradiated PMN (Figs. 3C and 3D).
After being co-cultivated with Pb18 the PMN recruited by either the Pb infection or Zymosan inoculation showed a significantly higher mitochondrial activity (p = 0.0029 and p = 0.0004, respectively) if they had been previously light irradiated in vivo (Fig. 4). In addition, the Zymosan irradiated group had a significantly higher mitochondrial activity (p = 0.0012) than the Pb irradiated group, while the non-irradiated Zymosan group also had a significantly higher mitochondrial activity (p = 0.0001) than the non-irradiated Pb group (Fig. 4).
Protein production was significantly enhanced with LLLT at earlier evaluation periods when compared to the non-irradiated groups (p = 0.0001 and p = 0.009 for Pb and Zymosan recruited PMN, respectively). The kinetics of protein production illustrates an intriguing crescent behavior for non-irradiated/Pb stimulated PMN (p = 0.002) and an opposite decaying curve for the LLLT/Pb neutrophils (p = 0.001); this decaying curve was also obtained with the highly activated PMN from the irradiated Z groups (Fig. 5). Likewise, after Pb co-cultivation, the kinetic production of proteins by irradiated PMN (Pb and Zymosan recruited) underwent decreasing curves that were distinct from the growing curves produced by the non-irradiated groups; this led to very distant values between the non-irradiated and the irradiated groups at 18 hours of co-cultivation (p = 0.002) (Fig. 5). In summary, after 2 hours the Pb or Zymosan recruited PMN were significantly more metabolically active than their non-irradiated counterpart (p = 0.0001 and p = 0.009, respectively); in addition, after 18 hours of co-culture the Pb-recruited PMN that did not receive LLLT were significantly more active than the Pb irradiated group (p = 0.002); the Zymosan-recruited group also showed an initial disparity between irradiated and non-irradiated groups (p = 0.0043) when co-cultivated with the Pb; this disparity was neutralized after 18 hours of the co-culture (Fig. 5).
As seen in Fig. 6A, a significantly higher amount of ROS production, as measured by chemiluminescence, was seen with PMN from LLLT treated mice for both Pb and Zymosan groups (p = 0.0425 and p = 0.0325, respectively). In the co-cultivated groups, the light irradiated PMN consistently produced a significantly higher amount of ROS than their non-irradiated counterparts (p = 0.0356 and p = 0.0325 for the Pb and Zymosan recruited PMN, respectively) (Fig. 6B). The non-irradiated Pb PMN also produced more ROS than the Zymosan non-irradiated PMN after co-cultivation (p = 0.0406) (Fig. 6B).
LLLT treatment of mice was able to induce a higher fungicidal capacity in PMN cells, which was indirectly shown by a significantly lower number of Pb colonies growing from material isolated from the air pouches when evaluated after a 12-day growth period (p = 0.0002) (Fig. 7A). Moreover, LLLT was able to induce a significantly higher fungicidal capacity in PMN recruited by either Pb or Zymosan after 7 (p = 0.0369 and p = 0.0232, respectively) (Fig. 7B) or even after 12 days (p = 0.0193 and p = 0.0492, respectively) (Fig. 7C) of co-cultivation with Pb. Nevertheless, none of the groups was able to totally inhibit the growth of the fungi.
LLLT treatment of the mouse femurs elicited a more active PMN population that could better deal with the Pb infection. This was warranted by the characterization of protein levels, mitochondrial activity and ROS assessment that all together showed that PMN from light-irradiated mice were more metabolically active and also produced more ROS, thus being more fungicidal in the actual lesion, and more fungicidal even after a later ex vivo re-exposure to the fungus.
PMN from infected patients may inactivate Pb [36] and are considered important as immune attack cells that contribute to the host response against this fungal infection, especially in the early stages of PCM [37]. Nevertheless, neutrophil functions such as fungal killing require activation by cytokines and other elements of the immune system [37]; IFN-γ, TNF-α, granulocyte-macrophage colony-stimulating factor (GM-CSF) and IL-15 are some examples of factors that can activate human neutrophils to carry out increased fungicidal activity by a mechanism dependent on production of ROS such as H2O2 and superoxide anion [38,39].
The more severe outcome of a Pb infection after depletion of PMNs has been previously published [40]; susceptible neutrophil-depleted mice displayed uncontrolled inflammatory responses, while normal resistant mice produced well-balanced Th1/Th2 responses [41] and thus were able to better clear out the fungi [40]. Moreover, normal resistant mice were stronger against the Pb infection and had highly activated PMN, in contrast to less activated neutrophils and macrophages seen in susceptible mice [32,42].
Furthermore, defects in PMN activation also correlate with the lack of fungicidal activity; a mutation in the CD40L gene and the lack of CD40L expression by activated T cells are examples of this [2]. An efficient Th1 immune response characterized by sufficient IFN-γ production and the satisfactory activation of phagocytic cells is required to eradicate the Pb infection [3]. CD40L-deficient patients showed a T cell response that yielded lower IFN-γ and higher IL-4 and IL-5 production, which led to a higher susceptibility to PCM infection [2].
Since the PMN from healthy human subjects can readily phagocytose the Pb cells [43], it has been proposed that neutrophil deficiencies must be present in PCM patients, especially related to the capacity of PMN to phagocytose and destroy this fungus [44–47]. Some previous studies suggested that only neutrophils from pre-sensitized mice could inactivate Pb in vitro [45,48]. The concept of activating PMN to produce an improvement in their fungicidal capacity is not completely new. Previously, the enhanced candidacidal activity of PMN and the improved ability of PMN to kill Blastomyces dermatitidis in vitro was achieved with an intraperitoneal injection of homologous antigen in B. dermatitidis-immune mice [27]. Stimulation of sensitized spleen cells with specific antigen can also be helpful to activate PMN for improved microbicidal activity [26].
We aimed to activate the “Pb-fighting PMN” with a new method never before utilized for this purpose, namely LLLT delivered to the bone marrow of the femurs of mice. This technique should be completely safe, as it has no known contra-indications and is quick and easy to perform. LLLT is a non-ionizing, non-thermal type of radiation that is known to improve tissue healing [20,21,49], whilst activating several signaling pathways related to cell proliferation, survival, repair and regeneration [19,50–54].
Antifungal drugs may lead to side-effects [16] and may induce development of drug-resistance, which in the case of Pb is primarily mediated by increased melanization [18]. However there is no report related to the use of LLLT producing any development of resistance in microorganisms. Moreover, antifungal drugs such as itraconazole have been implicated in producing relapses in a percentage of patients treated for PCM; 50% of these recurrences occurred after 36 months after discontinuing antifungal treatment [17], which is undeniably a considerable interval.
Both the subcutaneous infection with Pb and inoculation of Zymosan trigger a marked neutrophil response [55]. Zymosan consists of a mixture of fungal cell wall and intracellular components, amongst which beta-glucans are the most important and elicits many inflammatory responses, such as the production of ROS and cytokines that are involved in phagocytosis of microorganisms by neutrophils and macrophages [56]. Although Zymosan alone did not induce neutrophils as much as the Pb itself, LLLT produced bone marrow stimulation that was translated into a higher migration of activated PMN also in the Zymosan group.
The PMN of the Zymosan group, either irradiated or non-irradiated, were always more metabolically active than the Pb-exposed cells when co-cultivated with Pb. The neutrophils that were facing the Pb for the first time showed a higher mitochondrial activity than the cells that were re-exposed to this fungus. Interestingly, Zymosan did provide a more controlled (and thus more beneficial) host immune response than Pb, probably due to its recognition by TLR-2 and dectin-1 receptors leading to production of IL-10 [57]. The presence of activated neutrophils even in the Zymosan-irradiated groups showed that LLLT can serve as a tool to activate the bone marrow to produce an improved host defense especially against pathogens that require a rapid attack by the innate immune system. LLLT did not elevate the number of migrated PMN after Pb infection, although the neutrophils that were recruited by Pb alone were not as effective as the ones that were also light stimulated.
Alpha-1–3 glucan is the major cell-wall component of the yeast phase of Pb, therefore, Zymosan and Pb share the presence of glucans in common. Glucans are recognized by the c-type lectin, dectin1, as well as by CR3 complement receptor and lactosylceramide [56]. Since we did not add complement to the co-culture, we can surmise that the recognition of Pb by the neutrophils during the in vitro experiment was by a non-opsonic process. We cannot however, rule out the presence of complement in the exudate at the site of the subcutaneous air pouch. In addition, the air pouch technique was elected over the intraperitoneal [58], intra-oral [59] or intratracheal [60] routes due to its potential to raise a wide pool of almost pure neutrophils and yet localized and controlled Pb infection [32]; in that way, LLLT was securely delivered to the bone marrow and not to fungal cells, guaranteeing that the Pb killing was in fact due to neutrophil roles and not the laser acting directly upon the fungi.
In wound healing studies, LLLT has been observed to stimulate inflammation in some circumstances [23,24], which would be consonant with our goal of combating infections. Conversely, LLLT is often used to reduce the inflammatory response, and dampen down pro-inflammatory signaling [19,61–63], which was clearly not our focus with the present study. In fact, studies utilizing different models of acute inflammation have presented a declined edema formation and diminished neutrophil influx after LLLT [64–66].
According to these aforementioned studies, we showed that the quantity of the recruited neutrophils was diminished with LLLT, which could have been interpreted as an anti-inflammatory response; however, the level of activation of these cells was significantly improved with the use of LLLT, which was indeed applied to the bone marrow and not to the actual air-pouch. Interestingly, it has been postulated that LLLT may be potentially pro-inflammatory in the absence of antioxidants, while it can act as an anti-inflammatory stimulus when in the presence of sufficient antioxidants [67]. Moreover, neutrophils from patients with PCM are functionally deficient against suspensions of live Pb [44,45,68]; these neutrophils degenerate during the process of phagocytosis [4]. Thus, even though the neutrophil influx was higher in the non-irradiated group, this infiltrate was less efficient than the light-stimulated PMN.
The stimulation of bone marrow by LLLT seems to require an additional stimulus by the fungal cells, since the animals inoculated only with saline solution did not show increased recruitment of neutrophils after laser stimulation. Thus, the LLLT mechanism in our study could be described as biomodulatory [67] rather than pro-inflammatory, as if the PMN were primed to respond better against the invasion by fungal cells. LLLT typically leads to an increase in mitochondrial activity [69], and consequent induction of the cell-cycle with the synthesis or release of growth factors, interleukins, cytokines etc [70].
The higher mitochondrial activity seen in the LLLT group could be correlated with the protein production of these cells, while the kinetics of protein production was different among the irradiated and non-irradiated groups. Two hours after PMN extraction the neutrophils of the LLLT group were highly activated and showed a tendency to decrease until they were poorly activated after 18 hours. By contrast, the non-irradiated PMN started as less activated and began to produce higher quantities of proteins as time passed by. This same shaped curves (decreasing for illuminated PMN and increasing for non-illuminated PMN) were obtained for both Pb and Zymosan treated groups after they were extracted from the air-pouches and even after they were placed in contact with Pb in vitro.
The neutrophils from the mice that did not receive LLLT only achieved the same level of initial activation of the LLLT group after they were cultivated for 18 hours; or co-cultivated for 6 hours along with Pb. For the Zymosan group, not even after 18 hours of culture or co-cultivation did the non-illuminated neutrophils achieve the same degree of protein production as the LLLT-group cells. Even the cells that were facing Pb for the first time (Zymosan group co-cultivated with Pb) were more capable of dealing with this pathogenic fungus once they had been activated by the light.
The appropriate activation of phagocytic cells and particularly the production of ROS by nicotinamide adenine dinucleotide phosphate oxidase are important for the control of fungal infections [2,14]. Our present results show that Pb is able to activate the oxidative burst of neutrophils and that these cells are efficient in killing Pb, confirming earlier data from our group that showed that PMNs from resistant mice are more efficient in killing Pb than PMNs from susceptible mice [32]. In addition, the outbred Swiss mice utilized here were shown to be resistant since their survival rate after Pb infection was similar to that of A/J or A/Sn resistant mouse strains [71]; accordingly, LLLT stimulated even further the “already more efficient” [32] PMN from resistant mice.
The PMN from light-irradiated mice produced more ROS than their respective control groups, whether they were recruited through Pb or Zymosan inoculation. Rodrigues et al. [38] activated normal human neutrophils in vitro by using cytokines (IFN-γ, TNF-α, GM-CSF), thereby increasing their fungicidal activity against Pb, and showed the participation of ROS in this process. The same group also showed the suppressive effect of IL-10 in the same process [3].
We could also establish a good correlation between the ROS production and the fungicidal activity of PMN; CFU counting demonstrated that the material from irradiated mice had less viable Pb cells, after 12 days growth on solid media. Furthermore, PMN from light-irradiated mice that were re-exposed to Pb retained their higher fungicidal activity. Moreover, even the PMN from the LLLT-Zymosan group that underwent an initial contact with Pb in vitro were able to substantially impair the growth of Pb.
In the literature concerning co-cultivation studies between PMN and Pb, there is a report of a fungistatic (not a fungicidal effect), and only after a long incubation time with Pb (72 hours). Although PMN treated with IFN-γ did show better killing abilities (not against all studied strains), tumor necrosis factor-α and IL-8 did not improve PMN antifungal capacity [47]. In our study LLLT appeared to be an effective approach since it did enhance the fungicidal capacity of PMN after co-cultivation. It should be noted that LLLT was delivered to the mouse femur to activate the bone marrow, not to the actual PMN in vitro. The effect of the LLLT enabled the recruited PMN to fight the highly virulent Pb18 strain [28] both in vivo and in co-culture; nevertheless, the subcutaneous air pouch route utilized herein does not represent the natural course of PCM within the patients (inhaled fungal cells), so the results of this study may not be overestimated.
PMN activation through LLLT to the bone marrow led to a higher cell activity that correlated with two main effects: enhancement of innate immunity, shown by the higher yield of ROS and inhibition of Pb CFU in the lesion; and possible stimulation of acquired immune response shown by the increased yield of proteins seen in the LLLT groups. Finally, it is worth mentioning that although LLLT could be an effective and totally safe technique to activate fungicidal neutrophils, it was still not enough to eradicate the PCM; as previously stated, the phagocytic activity of PMN is considered not sufficient to entirely kill Pb [4]. Further study is warranted to see if different LLLT parameters, different sites of mouse irradiation or even distinct Pb infection routes could produce even better results from this promising technique.
|
10.1371/journal.pntd.0003758 | High Rates of Asymptomatic, Sub-microscopic Plasmodium vivax Infection and Disappearing Plasmodium falciparum Malaria in an Area of Low Transmission in Solomon Islands | Solomon Islands is intensifying national efforts to achieve malaria elimination. A long history of indoor spraying with residual insecticides, combined recently with distribution of long lasting insecticidal nets and artemether-lumefantrine therapy, has been implemented in Solomon Islands. The impact of these interventions on local endemicity of Plasmodium spp. is unknown.
In 2012, a cross-sectional survey of 3501 residents of all ages was conducted in Ngella, Central Islands Province, Solomon Islands. Prevalence of Plasmodium falciparum, P. vivax, P. ovale and P. malariae was assessed by quantitative PCR (qPCR) and light microscopy (LM). Presence of gametocytes was determined by reverse transcription quantitative PCR (RT-qPCR).
By qPCR, 468 Plasmodium spp. infections were detected (prevalence = 13.4%; 463 P. vivax, five mixed P. falciparum/P. vivax, no P. ovale or P. malariae) versus 130 by LM (prevalence = 3.7%; 126 P. vivax, three P. falciparum and one P. falciparum/P. vivax). The prevalence of P. vivax infection varied significantly among villages (range 3.0–38.5%, p<0.001) and across age groups (5.3–25.9%, p<0.001). Of 468 P. vivax infections, 72.9% were sub-microscopic, 84.5% afebrile and 60.0% were both sub-microscopic and afebrile. Local residency, low education level of the household head and living in a household with at least one other P. vivax infected individual increased the risk of P. vivax infection. Overall, 23.5% of P. vivax infections had concurrent gametocytaemia. Of all P. vivax positive samples, 29.2% were polyclonal by MS16 and msp1F3 genotyping. All five P. falciparum infections were detected in residents of the same village, carried the same msp2 allele and four were positive for P. falciparum gametocytes.
P. vivax infection remains endemic in Ngella, with the majority of cases afebrile and below the detection limit of LM. P. falciparum has nearly disappeared, but the risk of re-introductions and outbreaks due to travel to nearby islands with higher malaria endemicity remains.
| Solomon Islands, an island nation in the Southwest Pacific that has seen dramatic reductions in malaria transmission over the past 20 years, is aiming for malaria elimination. There is an increasing recognition that a substantial reservoir of asymptomatic and often sub-microscopic Plasmodium spp. infections exists even in low transmission settings. However, the potential role for these infections in sustaining transmission and the difference in response of the two most common malaria parasites, P. vivax and P. falciparum, to intensified control remains unclear. In May-June 2012, we therefore performed a cross-sectional survey of 3501 residents of all ages of Ngella, a low transmission area in Central Islands Province, to assess the prevalence of P. vivax and P. falciparum infection, determine the proportion of sub-microscopic and afebrile infections and evaluate whether gametocytaemic, and thus potentially infectious, individuals are present. Our survey showed a marked epidemiological contrast between P. vivax and P. falciparum. Although prevalence varied significantly among different regions of Ngella, P. vivax remains firmly endemic, with high rates of sub-microscopic, afebrile and genetically diverse infections. The presence of gametocytes among both sub-microscopic and microscopy positive, asymptomatic infections indicates that these infections contribute significantly to sustaining P. vivax transmission. P. falciparum, on the other hand, appears to be more amenable to control interventions. Only five P. falciparum infected individuals were detected, and all were residents of the same village. These infections carried the same msp2 clone. This difference highlights the larger challenge of eliminating P. vivax compared to P. falciparum in areas where they are co-endemic. In particular, the challenge posed by the presence of a large reservoir of silent P. vivax infections will need to be addressed if control of this parasite is to be accelerated and elimination achieved.
| Nations in the Southwest Pacific have endured considerable malaria transmission, with the highest Plasmodium falciparum burden outside the African continent and possibly the highest Plasmodium vivax transmission in the world [1]. Historically, transmission has ranged from hyperendemic areas in West Papua (Indonesia) and Papua New Guinea [2] to high and moderate transmission in Solomon Islands and Vanuatu [3], which are the southwestern boundary of global malaria transmission. Intensified control over the last 20 years has resulted in remarkable declines in malaria transmission in this region [3,4], reviving the agenda of elimination. However, it is in these countries where outstanding progress towards elimination has been made, that more knowledge is needed if the vision of malaria elimination is to be realized, such as reliable prevalence estimates, role of low-density, asymptomatic carriers and determinants of transmission maintenance.
In Solomon Islands, the incidence of clinical malaria cases diagnosed by light microscopy (LM) dropped by 90% from 442/1000 population in 1992 [5] to 44/1000 population in 2012 [6]. These drops in incidence are similar to those achieved by the Malaria Eradication Program in Solomon Islands (1970–1975) [7]. National statistics based on passive surveillance indicate that 65% of clinical malaria cases in 2012 were attributable to P. falciparum, 33% to P. vivax and 2% to mixed P. falciparum/P. vivax. Conversely, active case detection surveys indicate that P. vivax is the predominant species in the general population [6]. Current malaria transmission appears to be focal, ranging from moderate to high levels in Honiara City (96/1000) and Guadalcanal (64/1000) to very low in Temotu (10.8/1000) and Isabel provinces (1.2/1000).
Temotu and Isabel are the only two provinces in which pilot elimination agenda has been proposed to be actively pursued, having resulted in more intensive control activities and interventions including stratification, active case detection, and the earlier roll out of control activities (e.g. rapid diagnostic tests, RDTs and indoor residual spraying) than the rest of the country [3]. These provinces are also the only areas of Solomon Islands with recent surveys in which both LM and PCR-based diagnoses of Plasmodium spp. infections were performed [8,9]. In 2008, a parasite prevalence of 2.7% by LM was found in Temotu, with P. vivax accounting for 82.5% of infections. Only 5.5% of these infections were associated with febrile illness. Among a subset of 1,748 samples, which included LM positive, febrile and 10% of LM negative participants, an additional 63 P. falciparum, 23 P. vivax and 10 mixed P. falciparum/P. vivax infections were detected by PCR, indicating a 6.5% prevalence of sub-microscopic infections. Even lower levels of infection were reported in Isabel in 2009: 1 of 8,554 participants had a LM-detectable P. falciparum infection (0.01%). In a random subset of 2001 participants, PCR identified an additional 13 (0.55%) P. vivax infections.
PCR consistently detects at least twice as many infections as LM [10]. Numerous studies have confirmed that sub-microscopic infections are a common feature of malaria endemic areas, spanning all age groups and involving both P. falciparum and P. vivax [11–13]. Although these sub-microscopic infections are rarely associated with febrile illness, they have been shown to be efficient gametocyte producers [14–19] and thus constitute a source of ongoing transmission [10].
Given the lack of data from other areas of Solomon Islands, it is currently unknown whether the pattern of asymptomatic, low-density infection carriage identified in Temotu and Isabel [8,9] is unique to these elimination provinces. In addition, whereas these earlier surveys detected a large burden of sub-microscopic infections, they did not determine if these infections were also gametocytaemic and therefore did not assess their potential contribution to transmission. Therefore, we conducted in May-June 2012 a household-based, cross-sectional survey in Ngella, Central Islands Province to determine how common low-density, asymptomatic infections are in communities where transmission is mesoendemic and whether these infections are gametocyte producers and hence, potential contributors to local transmission. This survey is the first epidemiological description of malaria in Ngella since the 1970–1975 Malaria Eradication Program [7] and the only one in Solomon Islands to employ highly sensitive molecular diagnosis for the detection of both blood-stage parasites and gametocytes.
This study was approved by The Walter and Eliza Hall Institute Human Research Ethics Committee (HREC number 12/01) and the Solomon Islands National Health Research Ethics Committee (HRC12/022). The informed consent process recognized the community and cultural values of Solomon Islands. Following consultation with and approval by community leaders, community meetings were held to explain the aims, risks and potential benefits of the study. Individual informed consent was obtained from all participants or the parent or legal guardian of children<18 years of age. At the point of collection, all samples were de-identified.
Ngella, previously known as the Florida Islands, consists of 3 islands, Anchor, Big Ngella and Small Ngella, located approximately 27 miles north of Guadalcanal and 50 miles southwest of Malaita (Fig 1). Along with Tulaghi, Savo, Russel and Buenavista Islands it forms part of the Central Islands Province (Fig 1). Despite their proximity, the three islands of Ngella have diverse geographical characteristics: Anchor Island is characterized by less dense rainforest and sandier soil. Big Ngella is heavily forested, although commercial deforestation is common, and smaller villages are encountered in the Bay area around Tulagi, the provincial capital. The more remote northern villages of Big and Small Ngella and those on the southern coast are larger. The communities of the Utuha Channel lay in an extensive mangrove system and are smaller in size. There is minimal seasonal variation in temperature and despite a northwesterly monsoon from November-April, the distinction between wet and dry season is not pronounced. The most recent census estimates 26,051 inhabitants (approximately 60% of these reside in Ngella), 49% females and a median age of 19.9 years [20]. There is significant migration between Ngella and other malaria endemic areas, in particular Honiara (Guadalcanal) and Malaita provinces. These provinces are well connected to Ngella by a popular ferry service and numerous private, unscheduled motorized boat trips. The Ngella population is serviced by a hospital in Tulagi, six rural health sub-centres and ten nurse aid posts. National malaria statistics describe Ngella as mesoendemic, with a reported Annual Parasite Index [21] of 46.1/1000 in 2012, P. falciparum being the main cause of malaria cases [6].
Overall API for Solomon Islands indicates that there were two transmission peaks in 2012 for the months of February and October. As elsewhere in the country, long lasting insecticidal nets and indoor residual spraying are the mainstay of malaria control in Ngella. Cases are diagnosed by LM or RDT and treatment with artemether-lumefantrine has been introduced nationally in 2008.
The last malaria epidemiological report of Ngella [7] described it as ‘the most malarious group in all Solomon Islands” and the “most difficult from which to clear malaria”. Malariometric surveys preceding the Malaria Eradication Program (March 1965—January 1970, unpublished World Health Organisation Field Reports (reviewed in [8]) identified a combined parasite rate of 69.6% and a spleen rate of 69.3% in the 2–9 years age group. In the same surveys, villages on the North coast had spleen rates in the 80% range and qualified for the hyperendemic classification [7], whereas the villages in the Bay area and South coast were noted to have had spleen rates in the 30–50% range [7].
A representative population sample was obtained with a household-based sampling strategy of villages in 5 distinct geographical regions (Fig 1B, Anchor Island, North Coast, Bay, South Coast and Channel). The survey included 3501 individuals of all ages ≥6 months residing in 874 households in 19 randomly selected communities. The households were enumerated and geo-positioned and demographic information of the male and female heads of the household collected. Enumeration, but not geopositioning, was achieved for the households in the villages of the South Coast. The timing of the survey was approximately 4 weeks after the peak of the wet season.
Following consent and enrolment from each participant, a short clinical assessment was conducted (including tympanic temperature, history of fever during the previous 48 hours, history of malaria in the last 2 weeks and spleen size in children 2–9 years old) and demographic information collected (age, sex, residency status and history of travel, bed net use). A febrile participant was defined as having a tympanic temperature ≥38.0°C and/or a history of febrile illness in the past 48 hours. Study participants who reported being ill at the time of the survey were diagnosed by RDT (Access Bio, CareStart, USA) and treated if positive with artemether/lumefantrine, as per the national treatment guidelines. Where available the participant’s health records were checked for recent anti-malarial treatment and applicable information recorded.
A 250 μL finger prick blood sample was collected into EDTA-Microtainer tubes (Becton Dickinson, NJ, USA). 50 μL were immediately stabilized in 250 μL RNAProtect (Qiagen, Germany) for RNA studies and stored at ice pack cooling conditions until their transport to a centralized field laboratory. Thick and thin films were prepared for determination of microscopic malaria infection. Haemoglobin measurement was performed with Hemocue HB 301 analyzer. A measurement below 11g/dL was classified as anaemia. Upon return to the centralized laboratory, the RNAProtect fractions were frozen immediately. The remaining 200 μL of whole blood was separated into red blood cells pellets and plasma and promptly frozen.
Giemsa stained blood films were examined under x1000 power. One hundred fields of view were examined before calling a sample “no parasites seen”. When a parasite was observed, counts of both white cells and parasites were commenced, and continued until 300 white cells had been counted. The parasite count was then calculated, based on an assumed white cell count of 8,000 white cells/ μL. However if no further parasites were observed, the process of scanning to a total of 100 fields of view was completed. When only 1 parasite had been observed in 100 fields of view, an assumed count at the notional lower limit of detection of 10 parasites/μL was applied, based on a further assumption of an average of 8 white cells per field of view. All slides were stained within 24 hours at the regional malaria laboratory and read by experienced microscopists, all of whom had completed WHO quality assurance courses. All LM positive slides as well as the slides from all PCR positive / LM negative plus 10% of LM & PCR negative slides were re-read by an Australian Level 1 expert microscopist that was blinded to the PCR results. None of the 10% LM negative slides were found to be positive by the expert microscopist. In case of discrepancies between the two microscopy reads, the read of the expert microscopist was considered final.
Genomic DNA (gDNA) was isolated from red blood cell pellets (100 μL, corresponding to 200 μL whole blood) using FavorPrep 96-well Genomic DNA kit (Favorgen, Taiwan). DNA was eluted in 200 μL elution buffer and stored at -20°C. The RNA isolation procedure from whole blood in RNAProtect cell reagent has been described elsewhere [22], the only exception being an increased elution volume of 60μL of RNase-free water. Due to problems with storage of RNAProtect samples in the field, the quality of the RNA was tested using an RT-qPCR for the human beta globin transcript [23]. This revealed a 10x lower total human RNA concentration than in samples from a comparable study in Papua New Guinea [24]. RNA samples were therefore concentrated 10-fold using a CentriVap Concentrator (Labconco, United States) before testing for the presence of gametocytes.
All 3501 DNA samples were first screened using a genus-specific qPCR targeting a conserved region of the 18S rRNA gene [22]. Singleplex species-specific P. falciparum and P. vivax Taqman qPCRs and a duplex P. malariae/P. ovale qPCR, targeting species-specific regions of 18S rRNA gene, were used to identify species as described previously [22,25]. Prevalence values reported in this study include only those infections confirmed by the species-specific qPCR Taqman assays. Each detection experiment carried a dilution series of plasmids containing the target sequence of each PCR (104, 103, 102, 101, 5, 100 copies/μL), in duplicate, and were used to determine standard curves and therefore estimate parasite densities (reported as 18S rRNA gene copy numbers/μL). All assays were run in 384-well plate format on the Roche LightCycler480 platform. Those infections detected by qPCR, but not by LM, were defined as sub-microscopic infections. P. falciparum and P. vivax samples that were positive by species-specific Taqman qPCR were examined for presence of gametocytes using RT-qPCRs targeting the pfs25 and pvs25 orthologues, which are expressed only in mature gametocytes, as described previously [22]. All gametocyte assays were also run in 384-well plate format on the Roche LightCycler480 platform.
All samples that were P. falciparum or P. vivax positive were genotyped to determine the multiplicity of infection (MOI) using highly diverse size-polymorphic molecular markers msp2 for P. falciparum and msp1F3 and MS16 for P. vivax, respectively. PCR and capillary electrophoresis were performed with slight modifications to the published protocols [26,27]. Genotyping data was analyzed as described previously [26,27].
Study data were collected and managed using REDCap electronic data capture tools hosted at the Walter and Eliza Hall Institute [28]. Analyses were done using the STATA12 statistical software package (College Station, TX). Differences in participant characteristics at enrolment and prevalence differences among geographical areas and groups of individuals were assessed using Chi-square (χ2) or Fisher’s exact tests. Differences in median ages and median household size were explored with quantile regression. Univariable and multivariable logistic regression were used for associations of P. vivax infection and exposure variables. Associations with P. vivax parasite density were investigated in simple and multivariable linear regression models on only those subjects who tested positive to qPCR diagnosis. Poisson regression analyses were utilized to explore associations between multiplicity of infection and exposure variables.
A total of 3501 Ngella residents across 874 households were surveyed. The gender and age profiles of the participants were representative of the Central Islands Province population, with 52.5% females and a predominance of younger individuals (median age 18 years). The age distribution was as follows: <2 years, 4.7%; 2–4 years, 10.6%; 5–9 years, 14.8%; 10–14 years, 14.3%; 15–19 years, 7.5%; 20–39 years, 27.0%;>40 years, 21.3%. The majority of participants (95.2%) resided in the village for ≥ 2 months. Of 447 participants who spent at least one night outside their village of residence in the last month, 69.1% travelled within Central Islands Province. 73.3% of participants reported having slept under a long lasting insecticidal net the night before and 56.4% owned a bednet for longer than 24 months. Of all households, 84.5% of households reported to have been sprayed with insecticide, and 70.4% of household heads spoke English. Of all participants, 687 (19.4%) had a history of fever in the previous two days, 685 (19.7%) reported feeling unwell/sick at the time of survey and 23.3% had a haemoglobin measurement <11g/dL. No participant aged 2–9 years of age was found to have an enlarged spleen. A detailed description of demographic and clinical characteristics by geographical region is given in S1 Table.
Overall, 130 individuals (3.7%) had Plasmodium spp. parasites detectable by LM: 126 P. vivax, three P. falciparum mono-infections and one P. vivax/P. falciparum mixed infection. No infections with P. malariae or P. ovale were observed. The prevalence of P. vivax infection varied significantly by geographical region (p<0.001) (Fig 1B) and was lowest in the South Coast and Anchor regions (0.8%), followed by Channel (3.0%) and Bay (3.5%) and North Coast (11.7%). P. vivax prevalence showed strong age trends and peaked in adolescents 10–15 years of age (8.6%, p<0.001) (Fig 2A).
Overall, 468 participants (13.4%) had qPCR-detectable infections: 463 were P. vivax mono-infections and five were mixed P. falciparum/P. vivax infections (0.14%). The 126 P. vivax infections and the one mixed infection by LM were confirmed by qPCR. Overall, 72.9% of P. vivax infections were sub-microscopic. In two of the catchments, Kelarekeha and Vuturua (Fig 1B), only sub-microscopic infections were observed among 59 and 156 individuals surveyed, respectively. P. vivax qPCR prevalence displayed spatial heterogeneity among the five geographical areas and the 19 catchments, varying from 3.0–38.5%. Prevalence by qPCR was highest in villages on the North Coast (25.0–38.5%) and lowest on Anchor Island (3.0–5.5%) (Fig 1B).
Of 874 households sampled across Ngella, 559 had no infected members, 210 had only one infected member and 105 had two or more infected members. There was no association between household size and probability of being infected (p = 0.550). Not taking into account any other variables, there was an increased risk of being infected if at least one other member of the household was infected (OR = 2.59, p<0.001, CI95[2.13, 3.16]).
P. vivax prevalence was age-dependent (p<0.001), lowest in<2 years (3.0%, n = 166) and peaking in the 10–14 year old age group (24.3%, n = 499) (Fig 2A). Prevalence of infection did not differ significantly between male and female participants (p>0.650). Participants who were residents (lived in the village ≥ 2 months) were more frequently infected with P. vivax than non-residents (infected residents: 13.7% vs. infected non-residents 8.3%, p = 0.045). Once residency status was taken into account, recent travel (defined as spending at least 1 night away from the village of residence in the last month) was not associated with a difference in infection risk (p = 0.300). Those living in a household where the household head speaks English, a proxy for education level, were infected less frequently (12.0%) than those living in a household where the head does not speak English (18.5%, p<0.001). There was a moderate increase in risk of P. vivax infection in those who reported not having slept under a net the night before compared to net users (users: 12.6% vs. non-users: 15.6%, p = 0.022).
The majority of P. vivax-infected individuals (84.6%) neither reported febrile symptoms (defined as history of fever or measured fever at survey) nor feeling ill (85.4%). Six of the 26 participants that had a measured fever at the time of the survey (tympanic temperature ≥38°C, 18.8%) were infected with P. vivax. Compared to uninfected participants, those with a P. vivax infection were less likely to report having had febrile symptoms in the previous two days (uninfected 20.0% vs. infected 15.2%, p = 0.014) or report feeling unwell at the time of survey (uninfected 20.5% vs. infected 14.6%, p = 0.003). A total of 280 P. vivax infections (60.0%) were both asymptomatic and sub-microscopic. There were no significant differences in the proportion of asymptomatic P. vivax infections between different age groups (p> 0.200) and regions (p> 0.240). Of 468 P. vivax-infected individuals, 19.6% had a haemoglobin<11 g/dL compared to 23.9% of uninfected individuals (p = 0.045).
Age was the strongest independent association with P. vivax. infection, peaking in 10–14 year olds (Table 1, AOR = 8.41, p<0.001) and remaining relatively steady for subjects aged 15 years and above (Table 1, 15–19 years AOR = 5.03; 20–39 years AOR = 4.75; ≥ 40 years AOR = 4.40; p<0.001). The reference group in this analysis was composed of children aged <2 years. Being a local resident, region of residency and living in a household with at least 1 additional infected member were all associated with an excess risk of infection. Significant protective factors included an English-speaking household head and reporting feeling unwell at the time of the survey. Detailed results are given in Table 1.
Parasite densities by LM count were low (estimated geometric mean 24.3 parasites/μL, CI95 [19.2, 30.8]) (Fig 2B). Of 130 LM-detectable infections, 55% (n = 70), were at the assumed limit of practical detection, i.e. approximately 10 parasites/μL. Similarly, P. vivax parasite densities by qPCR were also low (estimated geometric mean 18S DNA copy numbers of 4.6/μL, CI95 [3.9, 5.4]) (Fig 2B). In LM and qPCR positive infections, parasite density by LM (parasites/μL) and by qPCR (18S DNA copies/μL) were correlated (n = 130, R2 = 0.76, p<0.001). Factors predicting parasite densities by qPCR included age, history of fever in the preceding two days, anaemia (haemoglobin<11 g/dL) and geographical region of sampling. Parasite densities were highest among individuals aged <2 years and decreased in older age groups (p<0.001). Detailed results of associations with P. vivax density are given in Table 2.
Genotyping results for markers msp1F3 and/or MS16 were obtained from 349 P. vivax positive samples. Both markers were highly diverse; 15 msp1F3 alleles and 43 MS16 alleles were detected (Fig 3), resulting in an expected heterozygosity (HE) of 0.834 for msp1F3 and 0.937 for MS16. Out of 349 samples, 102 (29.2%) carried multi-clonal infections. MOI (combined msp1F3 and MS16), defined as the concurrent infections per individual, ranged from 1 to 4; mean MOI was 1.36. Mean MOI did not differ significantly among age groups (Fig 2D, p = 0.774) or among geographical regions (p = 0.610). P. vivax-infected individuals living in a household with at least one other infected individual had moderately higher mean MOI (1.56) than those who were the sole infected person of the household, but evidence for an association was moderate to weak (mean MOI = 1.42, p = 0.086). Mean MOI was positively associated with qPCR parasite density (p = 0.007).
In 110 of 468 (23.5%) P. vivax infections, gametocytes were detected by the pvs25 RT-qPCR, resulting in a population gametocyte prevalence of 3.14%. More gametocytes were detected in LM-positive (41.5%) than sub-microscopic infections (16.6%, p<0.001). The proportion gametocyte positive (defined as percentage of gametocytaemic P. vivax carriers) was highest in the 2–9 year olds (39.8%, n = 98) and decreasing sharply in the 10–14 year olds (20.7%, n = 121, p>0.001). The lowest proportion gametocyte positive was found among the P. vivax carriers in the 15–19 years and>40 years age groups (15.8%, n = 38 and 15.9%, n = 88, respectively) (Fig 2C).
By qPCR, P. falciparum was detected in 5 individuals, all of whom were co-infected with P. vivax. Of these, only four infections were detectable by LM, one as a co-infection with P. vivax and three as P. falciparum mono-infections. In two of the LM mono-infections and the mixed infection, only P. falciparum gametocytes were observed on the blood smear. The range of parasite densities, by qPCR, was 7.35–364 copy numbers/μL and in LM positive samples the densities ranged from 20 to 1430 parasites/μL. The low number of P. falciparum infections precluded analyses with densities for this species. The presence of gametocytes in the four LM-positive infections was confirmed by pfs25 RT-qPCR. No gametocytes were detected in the one sub-microscopic P. falciparum infection by RT-qPCR.
All five P. falciparum-infected individuals resided in the same village (Halavo, circled in Fig 1B) and ranged in age from 3 to 60 years. The oldest had a history of fever in the preceding two days. Two of the carriers had a haemoglobin measurement <11g/dL. None of the individuals reported to have slept outside the village in the previous month. All five P. falciparum infections were monoclonal and carried the same msp2 genotype of the Fc27 subtype.
Solomon Islands has achieved a remarkable 90% reduction in malaria incidence over the last two decades as a result of scaled-up malaria control interventions [6] and is now intensifying its efforts towards malaria elimination [3]. The present study is the first to undertake sensitive molecular diagnosis at this scale in Solomon Islands and the first large epidemiological description of malaria in Ngella since the Malaria Eradication Program (1970–1975). Our findings illustrate a striking distinction between the epidemiology of P. falciparum and P. vivax in Ngella.
High prevalence (13.4% by qPCR) and genetic diversity, as well as an increased risk for local residents and evidence of potential within-household transmission indicate considerable levels of endemic P. vivax transmission. There was significant variation of P. vivax transmission in different regions of Ngella, with the highest prevalence found on the remote North Coast (25.0–38.5%), which prior to the Malaria Eradication spraying operations was described as holoendemic and having an environment highly favourable to the mosquito [7]. The lowest rates of P. vivax infection were observed on Anchor (3.9% by qPCR), where 15 years ago a community-based initiative eliminated a substantial number of breeding sites through environmental management [Lodo, personal communication]. It is therefore likely that the presence of suitable larval habitats and vector abundance may be key factors influencing P. vivax transmission on Ngella.
It remains unclear whether autochthonous P. falciparum transmission remains in Ngella or parasites are being re-introduced by incoming travelers or returning residents from areas with higher P. falciparum burden, such Guadalcanal or Malaita provinces. In this survey, only five P. falciparum cases were identified in the village of Halavo (Fig 1B). As all five infections carried the same msp2 allele and four were gametocytaemic, a small local outbreak following recent re-introduction seems more likely. This is reminiscent of the situation in epidemic-prone areas of the Papua New Guinea highlands, where a clonal P. falciparum epidemic on a background of endemic, low level P. vivax transmission has been reported [29]. P. falciparum populations in neighbouring Guadalcanal province were in fact found to be of low genetic diversity [30,31]. Based on case statistics at the local health facilities, 30% of malaria cases detected in Central Islands Province are caused by P. falciparum [6] indicating that either importation of P. falciparum parasites is common or that low levels of endemic P. falciparum transmission may remain in some parts of Ngella. Further studies are therefore required to ascertain the absence of endemic P. falciparum transmission in this area of Solomon Islands and whether the cases found are the result of inter-island travel.
The current situation of malaria in Ngella (i.e. 3.7% prevalence by LM, clear P. vivax dominance and absence of enlarged spleens in children 2–9yrs of age) is a consequence of the dramatic reduction in malaria transmission achieved throughout Solomon Islands in the last 20 years [6]. This change is similar to that encountered at the end of 1974, after approximately 5 years of twice-yearly Malaria Eradication Program spraying. Then, prevalence in 2–9 year olds had dropped from pre-spraying rates of 60% to 1.4% and P. vivax became predominant [7]. Similar shifts in malaria epidemiology were also observed in the elimination provinces of Temotu [9] and Isabel [8]. In Temotu, P. falciparum accounted for 17.5% of infections in population survey conducted in 2008 [9], but by 2012, the national program’s surveillance system reported only P. vivax cases from both Temotu and Isabel [6]. This shift in the relative importance of P. falciparum and P. vivax are not unique to Solomon Islands and have been reported after periods of sustained malaria control from other settings where P. falciparum and P. vivax occur sympatrically, such as the Amazon [21,32], Central America [4] and Thailand [33].
As in other endemic settings [12,13,34], P. vivax infections were of low density and PCR found three times more infections than LM. The majority of infections were not accompanied by febrile symptoms or anaemia. On the contrary, participants who reported feeling unwell or febrile were less likely to be infected with P. vivax. While this significantly lower level of febrile symptoms in P. vivax carriers is likely to be an artifact of the large samples size it does indicate that P. vivax is not a common cause of fever in Ngella. Whereas asymptomatic P. vivax infections have been commonly found in areas of high transmission [12,35,36], the advent of molecular diagnosis has revealed that even at low transmission the majority of infections in cross-sectional surveys are symptomless [11,37,38], including in the previous surveys in Temotu [9] and Isabel [8] where 97.1% and 92.9% of P. vivax infected individuals infections were asymptomatic, respectively.
Both the presence of P. vivax infections and their level of parasitaemia were found to be strongly age-dependent, albeit in different ways: while P. vivax parasite densities decreased with age, prevalence of P. vivax infections rose throughout childhood and only started dropping in adolescents and adults. These contrasting patterns are most likely due to local mosquito biting behavior and acquisition of immunity. Anopheles farauti, the only coastal malaria vector in Solomon Islands, is biting predominantly in the early evening (i.e. before 10pm) and outdoors [39], when small children tend to be indoors but older ones still active. The increase in prevalence during childhood is thus likely to represent an increase in exposure to infective bites. At all levels of transmission, immunity to P. vivax tends to be more rapidly acquired than that to P. falciparum [40]. Thus, the strong reduction in prevalence and parasite densities with increasing age in Ngella indicate that P. vivax transmission there remains sufficiently high for relatively rapid acquisition of clinical and anti-parasite immunity.
Despite very low overall parasite densities, gametocytes were detected in almost a quarter of all P. vivax infections (in 41.5% of LM-positive infections and 16.6% of sub-microscopic infections). Given issues with RNA quality, it is likely that the gametocytaemic reservoir in Ngella was underestimated in our survey and the true prevalence of gametocytes is higher, especially in the sub-microscopic group. Given the rapid and ongoing production of P. vivax gametocytes, most if not all, blood stage infections could harbor concurrent gametocytes [41]. Whilst sub-patent P. falciparum infections have been shown to infect up to 43.5% of mosquitoes [17,19], the role of sub-microscopic P. vivax gametocyte carriage in sustaining transmission is poorly understood. The capacity of sub-microscopic P. vivax infections to infect mosquitoes has been established in studies from Thailand [18,42,43], Sri Lanka [44], Peru [45] and malaria therapy settings [14,15], but at varying proportions and with weak associations of gametocyte density. Although sub-patent infections may infect fewer mosquitoes, their higher prevalence in endemic settings may mean that the net transmission potential of low-density infections is higher. In Ngella, asymptomatic, sub-microscopic infections of adolescents and adults may thus be an important source of local transmission.
These considerations may constitute a significant challenge to the success of the Solomon Islands malaria control program. The national malaria surveillance system, based on passive case detection and irregular mass blood surveys, only employs traditional microscopy diagnosis. This diagnostic test may not only underestimate the true burden of malaria in the Solomon Islands but also lack the means to detect and attack a substantial part of the P. vivax transmission reservoir. Despite outstanding gains in the last two decades, traditional tools of the Solomon Islands malaria control program may therefore have reached their effectiveness in the face of a large and silent reservoir of P. vivax infection.
Our observation that people living in a household with another P. vivax infected individual is a noteworthy finding. Not only does it indicate likely within-household transmission, but also highlights that reactive case detection strategies [46–48] and focal mass drug administration [34] might be appropriately applied in Solomon Islands. In the Southwest Pacific, MDA campaigns that included primaquine to target the undetectable liver stage parasites have previously been successful in interrupting P. vivax transmission on Aneytium Island in Vanuatu [49] and Nissan Island in Papua New Guinea [50]. Combining automated registration of observed cases and rapid identification of transmission foci (e.g. in a spatial decision support system) [51] with reactive mass-screen and treat (MSAT) or with focal, household-based mass drug administration [52,53] should therefore be evaluated as possible additional malaria elimination tools in Solomon Islands and neighbouring Vanuatu. All interventions will be most efficacious if they include routine administration of primaquine to all P. vivax infected individuals. This will however require addressing the challenges posed by the potential primaquine toxicity in G6PD deficient individuals.
|
10.1371/journal.ppat.1004405 | Insights into Vibrio cholerae Intestinal Colonization from Monitoring Fluorescently Labeled Bacteria | Vibrio cholerae, the agent of cholera, is a motile non-invasive pathogen that colonizes the small intestine (SI). Most of our knowledge of the processes required for V. cholerae intestinal colonization is derived from enumeration of wt and mutant V. cholerae recovered from orogastrically infected infant mice. There is limited knowledge of the distribution of V. cholerae within the SI, particularly its localization along the villous axis, or of the bacterial and host factors that account for this distribution. Here, using confocal and intravital two-photon microscopy to monitor the localization of fluorescently tagged V. cholerae strains, we uncovered unexpected and previously unrecognized features of V. cholerae intestinal colonization. Direct visualization of the pathogen within the intestine revealed that the majority of V. cholerae microcolonies attached to the intestinal epithelium arise from single cells, and that there are notable regiospecific aspects to V. cholerae localization and factors required for colonization. In the proximal SI, V. cholerae reside exclusively within the developing intestinal crypts, but they are not restricted to the crypts in the more distal SI. Unexpectedly, V. cholerae motility proved to be a regiospecific colonization factor that is critical for colonization of the proximal, but not the distal, SI. Furthermore, neither motility nor chemotaxis were required for proper V. cholerae distribution along the villous axis or in crypts, suggesting that yet undefined processes enable the pathogen to find its niches outside the intestinal lumen. Finally, our observations suggest that host mucins are a key factor limiting V. cholerae intestinal colonization, particularly in the proximal SI where there appears to be a more abundant mucus layer. Collectively, our findings demonstrate the potent capacity of direct pathogen visualization during infection to deepen our understanding of host pathogen interactions.
| Vibrio cholerae is a highly motile bacterium that causes the diarrheal disease cholera. Despite our extensive knowledge of the genes and processes that enable this non-invasive pathogen to colonize the small intestine, there is limited knowledge of the pathogen's fine localization within the intestine. Here, we used fluorescence microscopy-based techniques to directly monitor where and how fluorescent V. cholerae localize along intestinal villi in infected infant mice. This approach enabled us to uncover previously unappreciated features of V. cholerae intestinal colonization. We found that most V. cholerae microcolonies appear to arise from single cells attached to the epithelium. Unexpectedly, we observed considerable differences between V. cholerae fine localization in different parts of the small intestine and found that V. cholerae motility exerts a regiospecific influence on colonization. The abundance of intestinal mucins appears to be an important factor explaining at least some of the regiospecific aspects of V. cholerae intestinal localization. Overall, our findings suggest that direct observation of fluorescent pathogens during infection, coupled with genetic and/or pharmacologic manipulations of pathogen and host processes, adds a valuable depth to understanding of host-pathogen interactions.
| Cholera, a severe and potentially fatal diarrheal disease, is caused by ingestion of food or water contaminated with the highly motile gram-negative rod Vibrio cholerae. Although the disease has been recognized for centuries, cholera still causes significant morbidity and mortality in several parts of the developing world, and it is an ongoing threat to public health in regions where access to clean water and adequate sanitation is limited [1]. For example, since the accidental introduction of V. cholerae to Haiti following a 2010 earthquake, cholera has already sickened ∼700,000 and killed more than 8,500 (http://www.mspp.gouv.ht/). V. cholerae is a non-invasive pathogen that colonizes the mucosal surface of the small intestine (SI). The majority of V. cholerae, including strains of the El Tor biotype within the O1 serogroup – the cause of the ongoing seventh pandemic of cholera - do not induce damage to host tissue; instead, mortality is principally due to the extreme dehydration that ensues from disease-associated diarrhea.
Analyses of V. cholerae infections in several animal models of disease, as well as in human volunteers, have enabled identification of numerous factors that contribute to bacterial colonization and disease. A key element is V. cholerae's production of cholera toxin, an ADP-ribosylating toxin that accounts for cholera's hallmark secretory diarrhea [2]. The toxin is not directly required for bacterial colonization of mammalian hosts [3]; however, due to the profuse diarrhea it induces, the toxin is thought to promote bacterial dissemination to new hosts. Cholera pathogenesis is also dependent upon V. cholerae's production of a type IV pilus, TCP, whose expression is co-regulated with cholera toxin [4], [5]. TCP is essential for V. cholerae to colonize the SI; it promotes bacterial aggregation and microcolony formation, and may also facilitate V. cholerae's adhesion to the mucosal surface and protect V. cholerae from antimicrobial agents in the intestine [6]. Additional genes and processes that are critical for V. cholerae survival and growth in vivo include LPS O-antigen, transport systems, such as RND efflux pumps [7], and metabolic processes, including biosynthesis of certain amino acids [8] [9] [10] (reviewed in [11]). Many of these have been identified in studies of suckling mice orogastrically infected with V. cholerae, a disease model that was developed more than 40 years ago.
Processes and apparati that modulate V. cholerae motility also influence intestinal colonization by this pathogen. Early studies showed that non-motile V. cholerae mutants had reduced virulence, and it was proposed that motility could enable the pathogen to penetrate the mucus barrier covering the epithelium [12]. More recently, targeted mutations that inactivate V. cholerae's single polar flagellum have also been shown to inhibit intestinal colonization [13]. Flagellum-based motility may enable the pathogen to reach preferred niches within the intestine; however, only its effect on net bacterial accumulation within the intestine has been investigated. Flagellum-based motility is also necessary for V. cholerae chemotaxis, but chemotaxis and motility mutants have distinct phenotypes in vivo. Of V. cholerae's 3 clusters of genes that encode chemotaxis-related proteins, only genes in cluster 2 have been found to be required for chemotaxis in vitro [14] [15]. Unexpectedly, cluster 2 mutants exhibit enhanced intestinal colonization in infant mice, particularly but not exclusively in the proximal intestine [16] [13] [17]. In particular, hypercolonization is associated with non-chemotactic V. cholerae mutants that exhibit counter clockwise-biased flagellar rotation, which results in longer stretches of smooth swimming and greater net movement, while mutants with clockwise-biased flagellar rotation reverse their swimming direction more often and exhibit attenuated colonization [17]. It has been proposed that chemotaxis facilitates movement toward the pathogen's preferential site of colonization in the distal half of the SI [17], [18]. The niches colonized by chemotaxis-deficient strains have not been identified.
Host factors and processes are also thought to modulate V. cholerae's capacity to colonize the SI, although there have been far fewer studies of these than of bacterial attributes. The acidic pH of the stomach is thought to kill most V. cholerae before the pathogen reaches the SI. Within the SI, mechanical and physical barriers include motility, which propels ingested and secreted material (e.g. mucus) toward the distal intestine, the mucus layer, which covers and protects the epithelial surface, and immune effectors (e.g. cryptidins), all of which are thought to limit V. cholerae colonization [19] [20]. The main component of the single layer of mucus that covers the small intestine is the mucin MUC-2, a large and highly glycosylated protein secreted by goblet cells [20]. The mucus layer is a highly viscous and complex structure, due in part to the disulphide crosslinks that form between mucin monomers [21]. Additional mucins that (unlike the mucus layer) are anchored to the epithelial cell membrane constitute the glycocalyx, another important protective barrier for the epithelium.
To date, most analyses of V. cholerae colonization and pathogenesis have not included analyses of the distribution of this pathogen within the SI or the bacterial and host factors that account for it. Enumeration of colony forming units (cfu) recoverable from different regions of the suckling mouse intestine has revealed that the proximal third of the SI harbors 40–100 fold less bacteria than the middle and distal regions [22]; however, this disparity has not been explained. Furthermore, with the exception of work monitoring fluorescently labeled V. cholerae in rabbit ligated ileal loops, which bypass the pathogen's ordinary route into the intestine [23], there is scant knowledge of how V. cholerae localizes along the villous axis in different regions of the SI. Here, we used confocal and two-photon microscopy to analyze the fine localization of fluorescent V. cholerae in different regions of the SI. Our observations suggest that most V. cholerae microcolonies arise from single cells attached to the epithelium. Unexpectedly, there are differences in V. cholerae localization in different regions of the SI. Notably, in the proximal SI, bacteria reside exclusively within the developing intestinal crypts. Furthermore, there are regiospecific requirements for motility in V. cholerae colonization; motility is critical for colonization of the proximal, but not the distal SI. Unexpectedly, neither motility nor chemotaxis were required for proper V. cholerae distribution along the villous axis, suggesting that yet undefined processes enable the pathogen to find its niches in the intervillous space. Additionally, our findings suggest that host mucins are a key inhibitor of V. cholerae colonization, particularly in the proximal SI.
In order to visualize V. cholerae within intestinal tissue from infected infant mice, we orogastrically inoculated animals with fluorescent derivatives of C6706, a 7th pandemic El Tor O1 V. cholerae isolate. One strain (VcRed) constitutively expresses a codon-optimized gene encoding the red fluorescent protein tdTomato (tdT), while a comparable C6706 derivative constitutively produces GFPmut3 (VcGreen) [24]. The growth of VcRed and VcGreen was indistinguishable from that of C6706, both in LB cultures and in the small intestines (SI) of suckling mice, as assessed by competition assays (Figure S1). These data suggest that VcRed and VcGreen can be used as reliable reporters of V. cholerae localization during infection of infant mice.
For localization studies, equal mixtures of VcRed and VcGreen were inoculated into suckling mice. In most cases, infection was allowed to proceed for ∼24 hr, as this yields maximal colonization; however, some experiments were terminated at 8 or 16 hr, to explore earlier stages of the infection process. At each end point, the small intestines were divided into three equal parts, and total bacterial load and distribution were monitored by plating intestinal homogenates and by confocal microscopy respectively. After only 8 hr, bacteria were difficult to visualize, particularly within the proximal SI, although analyses of cfu confirmed that they were present throughout the intestine (Figure S2A). Microcolonies were not yet evident 8 hr PI (Figure S2B), and we suspect that the majority of V. cholerae were not yet attached to intestinal tissue this early during infection. Even at 16 hr PI, only a few small microcolonies were evident (Figure S2B); for this reason, we focused our localization analyses on the 24 PI time point. Consistent with previous analyses of cfu in both infant mice and infant rabbits [22], [24], at 24 hr post-infection (PI) V. cholerae were most abundant within the medial and distal thirds of the intestine, and ∼20–100-fold less abundant within the proximal third of the SI (Figure 1B). However, confocal microscopy images revealed striking and previously unrecognized features of V. cholerae intestinal colonization. First, we observed that V. cholerae microcolonies on the intestinal epithelium are nearly always uniformly red or green (Figure 1CEG and Figures S2, S3, S4), strongly suggesting that the cells in microcolonies are clonal, i.e., that microcolonies arise from a single attached bacterium and do not trap or recruit unattached bacteria as they expand.
We also detected notable differences between the distribution of microcolonies along the intestinal villi in the proximal vs the medial and distal SI segments. Unexpectedly, in the proximal SI, V. cholerae microcolonies were almost exclusively (>90%) located at the base of the villi, within the forming crypts (Figs. 1CD), whose development is initiated during the first postnatal week [25]. In contrast, microcolonies in the medial and distal SI, which were more numerous, were predominantly detected in the bottom halves of the ∼300 µm long villi, but only ∼30% were located at the base of the villi (Figure 1C–I). The predilection for microcolony formation at the bases of villi was not anticipated, since crypts are known to produce antimicrobial products, such as cryptidins [26]. However, such crypt-protecting defenses may not be present in the 5 day old mice used here. Notably, a majority of colonies observed on the sides of the villi appeared to occupy crevices within the intestinal epithelium, although a precise frequency was not determined (Figure 1EG, white arrowheads). Preferential localization of V. cholerae at the bases of villi and in crevices likely shelters the organism from peristaltic forces that would propel the pathogen towards the distal intestine.
Our observations that microcolonies are largely clonal and have distinct localization in the proximal SI vs. the medial and distal SI were confirmed using intravital two-photon microscopy. In contrast to the confocal microscopy-based imaging, which requires dissection and processing (i.e. fixation and washing) of SI segments, intravital microscopy is performed using intact tissue, and thus is less likely to perturb pathogen localization. For our experiments, segments of small intestines of anesthetized infected or mock-infected suckling mice were exteriorized from the peritoneal cavity and placed on a microscope stage, and intestinal contents were visualized from the exterior of the tissue (Figure 2A). With this protocol, we could image microcolonies and tissue structure as far as ∼150 µm from the intestinal wall (Figure 2BC), which permits analysis from the serosa through the basal half of the villi, but not into the intestinal lumen. Twenty four hr after inoculation of infant mice with VcGreen and VcRed, small monoclonal colonies of either VcRed or VcGreen were detected only in crypts in the proximal small intestine; larger colonies were observed at the bases and along the bottom third of villi in the medial and distal segments of the small intestine (Figure 2C). These observations closely mirror the findings obtained with confocal microscopy, and thus provide support for the idea that V. cholerae microcolonies have distinct distributions in different segments of the intestine.
We also imaged explanted SI segments from VcGreen infected animals with two-photon microscopy. The explants (which were not opened en face) were mounted in a saline/lubricant gel imaging chamber that enables enhanced visualization of the intervillous space. In this setting, we were able to detect individual VcGreen cells moving through the intervillous spaces and occasionally contacting the large attached microcolonies that were particularly prominent in these images (Figure 2D and Videos S1, S2). Although the movement of VcGreen cells may reflect external convective forces rather than intrinsic bacterial motility, these images suggest that it may be possible to analyze the interactions of single tagged V. cholerae cells with each other and with the epithelium in future studies.
Although luminal (unattached) bacteria cannot be monitored using two-photon microscopy, luminal V. cholerae were observed in the medial and distal segments of the SI using confocal microscopy. These bacteria were often present as large clonal (all green or all red) aggregates, but mixed populations of VcRed and VcGreen were observed as well (Figure 3A). In the distal SI, clonal microcolonies were detected on the surface of digesta, suggesting that V. cholerae may adhere to and grow upon luminal contents (Figure 3A).
We also visualized tissue sections from mice inoculated with a single marked strain (either VcGreen or the cholera toxin-deficient mutant, ΔctxAB-GFP) that were stained with wheat-germ agglutinin (WGA), a lectin that binds to terminal N-acetyl-D-glucosamine and sialic acid residues on sugar chains [27]. WGA allows visualization of the highly glycosylated mucins in the glycocalyx that lines the epithelial brush border surface and that constitute intestinal mucus. Luminal V. cholerae colonies were often embedded in a WGA-rich matrix (Figure 3B). As was previously seen in infant rabbits infected with VcGreen, these clumps are reminiscent of the V. cholerae/mucus aggregates found in the ‘rice-water’ stool of cholera patients. Interestingly, in the rabbit model, luminal mucus accumulates in response to cholera toxin, which induces release of mucins from intestinal goblet cells [24]; however, in infant mice, the luminal WGA-reactive material was also present in uninfected control mice. Furthermore, and, in contrast to observations in V. cholerae-infected infant rabbits, no obvious difference between the amounts of luminal WGA-reactive material was observed in mice infected with VcGreen vs its colonization proficient but toxin-deficient ΔctxAB counterpart. (Figure 3B). Thus, in infant mice, the WGA-stained matrix in which V. cholerae is embedded does not appear to be derived from mucins released by goblet cells in response to cholera toxin.
To begin to understand the determinants of V. cholerae localization within the SI, we investigated the impact of disrupting bacterial or host processes that might contribute to bacterial localization, including bacterial motility and chemotaxis and the host mucus layer. In previous analyses, enumeration of V. cholerae in homogenates of the entire suckling mouse SI revealed that motility-deficient V. cholerae strains have a reduced capacity to colonize [12], [13], [29], [30], perhaps because flagellar-based motility enables the pathogen to reach particular intraintestinal sites; however, with the exception of one early study using undefined non-motile V. cholerae mutants [31], the impact of flagellar-based motility upon bacterial localization within the intestine has not been reported. Therefore, we carried out in vivo competition assays using VcRed and a GFP-marked ΔflaA V. cholerae mutant, which lacks the major flagellin subunit and does not produce a flagellum [32], [33]. As found in previous studies, the non-flagellated strain displayed a colonization defect, but notably, the effect of the mutation was not uniform across the small intestine. Instead, colonization was reduced (relative to the wt strain) by ∼1000-fold and 500-fold in the proximal and medial SI segments, but unimpaired in the distal segment (Figure 5A). To exclude the possibility that the flagellum might promote colonization via mechanisms independent of motility, such as enhancing adhesion, a GFP-marked non-motile but flagellated strain lacking the MotB component of the flagellum motor (ΔmotB-GFP) [32] was also tested in in vivo competition assays. Similar to the ΔflaA mutant, the ΔmotB mutant was markedly defective at colonizing the proximal and medial segments of the small intestine, but it also exhibited a modest colonization defect (5-fold) in the distal SI (Figure 5A). The similar phenotypes of the ΔflaA and ΔmotB mutants are consistent with the idea that the colonization defect of the ΔflaA mutant is due to its motility deficiency. To our knowledge, flagellar motility is the first V. cholerae attribute shown to be required for colonization of only a subset of intestinal sites. Typically, colonization factors are required throughout the intestine, as we observed for a TCP-deficient mutant (ΔtcpA), which exhibits highly compromised colonization in all SI segments (Figure 5A). Our data suggests that flagellar-based motility is critical for V. cholerae's ability to reach and/or be maintained in the proximal ∼2/3 of the SI, but that it is relatively unimportant for infection of the distal third of the SI.
The distribution of the non-motile strains was also assessed using confocal microscopy. Consistent with the findings from the plating assays discussed above, neither GFP-marked ΔflaA or ΔmotB V. cholerae were visible in the proximal SI (Figure S5), and colonies were rare in the middle SI as well. Surprisingly, the absence of flagellar-based motility did not dramatically alter the distribution of V. cholerae in the distal SI; ΔflaA and ΔmotB colonies were detected both at the base of villi and at lateral positions (Figure 5BC), suggesting that V. cholerae cells do not depend on flagellar-based motility to penetrate into intervillous spaces in the distal SI, as has previously been proposed [16]. Since functional flagella are also required for chemotaxis, these data also suggest that V. cholerae does not depend upon chemotaxis to penetrate into the intervillous spaces within the distal SI of infant mice.
We performed similar analyses of the colonization and intestinal distribution of V. cholerae lacking various chemotaxis genes. V. cholerae contains 3 gene clusters that encode homologues of chemotactic proteins, one of which (cluster 2) is known to be required for chemotaxis in vitro [14] [13], [15]. Inactivation of particular cluster 2 genes can lead to enhanced colonization of the infant mouse intestine, especially but not exclusively in the proximal SI [13], [17]. Roles for chemotaxis clusters 1 and 3 have not yet been defined. Consistent with previous observations of hypercolonization by a mutant lacking cheY3 or cheA2 (components of cluster 2) [13], [17], we found that a V. cholerae strain harboring a deletion of the entire set of cluster 2 genes (Δche2) out-competed the wt strain ∼100× and ∼10× in the proximal and medial SI segments respectively (Figure 6A). In contrast, colonization by a mutant lacking the other 2 clusters (Δche13) did not differ from that of the wt strain (Figure 6A). A triple mutant harboring deletions of all 3 putative chemotaxis clusters (Δche123) exhibited hypercolonization indistinguishable from the Δche2 mutant (Figure 6A), providing further evidence that the products of clusters 1 and 3 do not contribute to colonization, even in a secondary role. Notably, the hyper-colonization phenotype of the Δche2 mutant was disrupted by inactivation of motB, suggesting that bacterial motility is required for hypercolonization, even though the motility of the Δche2 mutant is undirected (Figure S6). The Δche2ΔmotB mutant exhibited a colonization defect similar to the ΔmotB strain in all parts of the SI (Figure 6A).
To further assess the importance of chemotaxis in promoting V. cholerae's capacity to navigate into and through the intervillous spaces, we monitored the distribution of Δche2-GFP microcolonies along the villous axis in the different SI segments. Notably, the fine localization of Δche2-GFP strain was very similar to that of wt V. cholerae in all intestinal segments, despite the markedly increased number of cfu in some segments. Like VcGreen and VcRed, in the proximal SI, nearly all Δche2-GFP microcolonies were found at the bases of villi, though they were found with much higher numbers than the chemotaxis-proficient bacteria (Figure 6B). No notable differences in the sizes of Δche2 and WT microcolonies were observed, suggesting that the hypercolonization of Δche2 is likely explained by the elevated number of crypts occupied by this mutant (although this remains a small fraction of crypts overall). In the medial and distal SI segments, Δche2-GFP was found at the base of villi and along the lower third of villus surfaces, as also was observed for VcRed (Figure 6B, note in the medial segment that Δche2-GFP significantly outcompetes VcRed). Thus, our results indicate that V. cholerae's only known functional chemotaxis cluster does not guide its fine localization in the small intestine, and counter the long-standing hypothesis that V. cholerae chemotaxis directs the organism toward the crypts [16], [34]. Additionally, our results suggest that hypercolonization by the Δche2-GFP strain does not reflect occupancy of a novel niche within the proximal SI; instead, in the absence of chemotaxis, V. cholerae simply establishes microcolonies within a higher percentage of proximal SI crypts than are occupied by wt bacteria.
To investigate whether the more abundant mucus layer in the proximal SI contributes to the relative resistance of this region to V. cholerae colonization, we treated mice with the mucolytic agent N-acetyl-L-cysteine (NAC), which is thought to disrupt the disulfide bonds between mucins [35]. Six hours after NAC was introduced by gavage into infant mice, there was marked reduction in WGA staining on the surface of intestinal villi (Figure 4A, NAC); in addition, this treatment appeared to partially disrupt and disorganize the mucus layer detected with PAS staining of Carnoy's fixed samples (Figure 4C). The effects of NAC were reversible, and by 24 hr after NAC treatment, staining was restored to pre-treatment intensity (Figure S7). Notably, pre-treatment of mice with NAC 30 minutes before V. cholerae inoculation increased colonization of all SI segments, but particularly the proximal SI. Nearly 150-fold more V. cholerae CFU were recovered from the proximal SI of NAC treated mice than from control (PBS-treated) animals (Figure 7A). Furthermore, confocal imaging revealed V. cholerae along the villi as well as at the base of villi in the proximal SI of NAC treated mice, rather than solely within crypts (Figure 7B). Increased colonization was also detected for the medial and distal SI (∼10× and ∼6×, respectively; Figure 7A). Overall, NAC treatment largely abolished differential colonization of SI regions, suggesting that mucus is a key factor in countering intestinal colonization by V. cholerae. NAC is also known to function as an antioxidant, and is possible that NAC also promotes bacterial growth by reducing the level of reactive oxygen species (ROS) in the intestinal lumen; however, NAC appears to be most potent against intracellular ROS [36].
Additional experiments with NAC treated mice suggest that the inability of the motility deficient ΔmotB V. cholerae mutant to penetrate the mucus barrier accounts for a significant portion of this strain's colonization deficiency. In both single infection and competitive infection experiments, the capacity of the ΔmotB strain to colonize the intestines of untreated mice is lower than that of the wt by ∼1 to several orders of magnitude, with the largest deficiency seen in the proximal intestine (Figure 4 and Figure 7AC). However, NAC treatment promoted colonization by the ΔmotB mutant, particularly in the proximal intestine, resulting in much less marked attenuation compared to the wt strain (Figure 7AC). Thus, although directed (i.e., chemotaxis-based) movement is not required for establishment of an infection, the bacterial ability to propel itself through, or escape from, mucus, seems to play a significant role, at least in the proximal intestine, where the mucus layer appears to be most prominent.
Using confocal and two-photon microscopy to detect fluorescent V. cholerae in the suckling mouse intestine, we have obtained new insights regarding where and how this pathogen grows in the host, as well as the bacterial and host processes that modulate colonization. Direct visualization of the pathogen within the intestine suggests that the majority of V. cholerae microcolonies observed on the intestinal epithelium arise from single attached cells; expansion of such colonies likely accounts for a significant proportion of V. cholerae proliferation within the host environment. Visualization of the pathogen also uncovered unexpected and striking differences between the fine localization of V. cholerae microcolonies within distinct regions of the SI. Notably, microcolonies were found almost exclusively in the developing crypts in the proximal intestine but at the bases and along the bottom third of villi in the distal 2/3 of the SI. The predilection of V. cholerae to occupy the crypts, the lower parts of villi and crevices within villi, likely provides a means for the pathogen to avoid the propulsive force of intestinal motility, which directs ingested material and secreted fluid and mucus toward the distal intestine. Residency in crypts may particularly protect bacteria against being shed from the epithelial surface. Host mucus seems to be a key factor that limits V. cholerae intestinal colonization, particularly in the proximal SI where there appears to be a thicker mucus layer.
Surprisingly, V. cholerae motility proved to have a regiospecific influence on intestinal colonization. Nonmotile mutants failed to colonize the proximal SI but were not compromised in their capacity to colonize the distal SI, where their distribution was similar to that of wt V. cholerae. It is possible that motility is required to penetrate the mucus layer, as originally proposed by Guentzel et al decades ago [31], since NAC treatment partially alleviated the colonization defect of the motility-deficient motB V. cholerae mutant. However, since NAC treatment also augments colonization by wt bacteria, it is likely that mucus imposes a barrier to colonization by wt V. cholerae as well. The relative lack of mucus in the distal SI may at least in part explain why the motility-deficient strains retained the capacity to colonize this part of the SI, and may also contribute to the preferential colonization of this region by wt V. cholerae. However, it is important to note that despite the impact of motility on the gross distribution of V. cholerae in the SI, motility is dispensable for the pathogen's proper fine localization in the distal SI. These results raise the possibility that flagellar motility enables V. cholerae dissemination throughout the lumen of the small intestine, but that additional (non-flagellum based) processes control its penetration into the intervillous space. Such processes could include peristalsis, mucus structure/organization and the distribution of (currently unknown) host targets of V. cholerae adhesions. In addition, V. cholerae has been reported to possess flagellum-independent motility on surfaces [37], and it has been proposed that flagellum-independent motility may aide V. cholerae migration through intestinal mucus [38].
Our findings counters the long-standing hypothesis, developed more than 30 years ago in pioneering studies by Freter, that chemotaxis facilitates V. cholerae penetration deeper into the intestinal mucosa and intervillous space, and that such penetration results in bacterial killing, due to the presence of unknown antimicrobial factors [16]. We demonstrate that although chemotaxis-deficient V. cholerae has an enhanced capacity to colonize the upper SI, its fine localization in both the upper and lower SI is equivalent to that of wt V. cholerae. The abundant nonchemotactic V. cholerae detected in the upper SI reside entirely within the crypts, clearly demonstrating that V. cholerae does not need chemotaxis to penetrate into the deepest zones of this tissue. Thus, like motility, chemotaxis appears to play a more prominent role in the overall distribution of V. cholerae within the intestine than in its fine localization within intestinal segments. As noted by Butler and Camilli [17], the tendency of non-chemotactic mutants to be biased towards straight swimming may help them to enter new intestinal sites and may contribute to their colonization phenotype. Indeed, such altered swimming could potentially have more impact than an inability to respond to either positive or negative chemotactic stimuli. Consistent with this possibility, we observed that the hypercolonization associated with the Δche2 mutation is dependent upon flagellar motility; a Δche2 motB mutant did not exhibit hypercolonization.
Both the distribution of host glycans and the effects of NAC treatment support the idea that host mucins restrict V. cholerae localization along the SI as well as along the villous axis. NAC treatment rendered the proximal SI much more permissive to V. cholerae colonization; it enabled the pathogen to occupy new sites along the villous axis in this intestinal region. Intestinal mucins are thought to constitute a key host defense against a variety of enteric pathogens [21], and many commensals and pathogens, including V. cholerae, produce enzymes (e.g the ToxR-regulated TagA mucinase [39]) that cleave sugars from or the peptide backbone of mucins. Although host mucus likely serves as a physical barrier between V. cholerae and intestinal tissue that limits infection, it is also likely to be an important source of energy for V. cholerae and other enteric pathogens that can digest its carbohydrate components. Previous studies have already revealed that a V. cholerae sialidase promotes robust V. cholerae colonization [40], and we observed that V. cholerae in the intestinal lumen is often associated with intestinal mucus. It should be possible to use fluorescence microscopy-based approaches along with genetically engineered mice (e.g., mutants unable to glycosylate the principal secreted mucin, MUC-2) and wt and mutant V. cholerae to further characterize the interplay between host mucins and this pathogen.
Finally, our observations of SI segments with intravital two-photon microscopy, a technique that does not perturb host tissues, corroborated our findings with confocal microscopy, which requires tissue processing. Like the confocal images, the two-photon images revealed that V. cholerae microcolonies are primarily monoclonal and showed differences between the fine localization of V. cholerae along the villous axis in different parts of the SI. To our knowledge, these observations represent the first application of intravital microscopy to imaging an orogastrically inoculated enteric pathogen in an intact intestine. Previous intravital imaging of enteric pathogens have relied on surgical exposure of the intestinal lumen and have primarily focused on interactions of pathogens with dendritic cells/macrophages (e.g. [41] [42]). Our findings suggest that it should be possible to use intravital microscopy to monitor host-pathogen and potentially pathogen-pathogen and pathogen-commensal interactions that occur on intestinal epithelial surfaces in real time.
All V. cholerae strains used in this study are streptomycin-resistant derivatives of C6706, a 1991 El Tor O1 Peruvian clinical isolate. The ΔflaA, ΔmotB, ΔtcpA, and ΔctxAB strains have been described previously [24], [32]. The chemotaxis operon deletion strains Δche2 (strain SR28, Δvc2059-vc2065), Δche13 (strain SR31, Δvc1394-1406 (che1), Δvca1088-vca1096 (che2)) and Δche123 (strain SR33, Δvc1394-1406 (che1), Δvc2059-vc2065 (che2), Δvca1088-vca1096 (che3)) were created by allelic exchange as described in [43], [44]. GFP-labeled strains, which constitutively express GFPmut3 under the control of the lac promoter, were generated by introducing the suicide vector pJZ111 (a kind gift of Dr. Jun Zhu) into the lacZ locus as described [24]. A derivative of pJZ111 (pYM50) was generated by inserting a V. cholerae codon-optimized version of the tdTomato gene (Genscript) in place of the GFPmut3 gene. This plasmid was used to generate the strain VcRed, which constitutively expresses tdTomato.
5-day old CD-1 mice were intragastrically inoculated as described [22]. For in vivo competition assays, 1∶1 mixtures of a GFP-labeled strain and VcRed were inoculated into each mouse (∼2×105 cells/mouse). After 24 h, unless otherwise noted, animals were euthanized and their small intestines removed and divided into three parts of equal length (proximal, medial and distal, ∼3.5 cm each); the central 1 cm segment of each part was removed, homogenized in LB and plated. For in vitro competition assays, 5 mL of LB containing streptomycin (200 µg/mL) were inoculated with 10 µL of the in vivo inoculum and grown at 30°C for 24 h. Serial dilutions were then plated. The number of CFUs of the GFP-labeled strain were determined by scanning the plates using a fluorescent image analyzer (Fujifilm FLA-5100). The ratio between GFP-labeled and VcRed CFUs was calculated and normalized by the ratio in the inoculum to determine the competitive index (CI). For single infection assays, ∼2.105 cells were inoculated into each mouse and after 24 h, the SI segments were prepared and processed as described above.
Statistical analyses were performing with Prism (GraphPad).
This study was performed in strict accordance with the recommendations in the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health. All animal protocols were reviewed and approved by the Harvard Medical Area Standing Committee on Animals (protocol #04316).
Tissue from a subset of mice used in infection studies was analyzed via confocal microscopy (n = 3 per assay). Mice were inoculated with VcRed and/or a GFP-labeled strain as described above. Tissue samples from the proximal, medial, and distal intestine were fixed in PBS with 2% paraformaldehyde for two hours at room temperature (RT), placed in PBS with 30% sucrose for two hours at RT, mounted in tissue freezing medium (EMS), snap-frozen in dry ice-cold 2-metylbutane and sectioned (10 µm). Initially, bacteria labeled with GFP were visualized via direct detection of the fluorescent protein; however, however, we found that these signals were less stable than those obtained via immunodetection of GFP, and so most images presented here were generated via immunostaining. No difference was detected between bacterial localization observed with the two approaches. For staining, frozen sections were washed in PBS for 5–15 minutes at RT, blocked in blocking buffer (1% BSA, 5% normal donkey serum in PBS) for 1 hour at RT, stained with a primary anti-GFP antibody (Abcam, ab13970) 1/1000 in blocking buffer with 0.2% tween20 for 1 hour at RT, washed three times in PBS, stained with a FITC-coupled secondary antibody (Abcam, ab6873) 1/1000 in blocking buffer with 0.2% tween20 for 1 hour at RT, washed three times in PBS, counterstained with DAPI (1 µg/mL) and in some cases with phalloidin-alexa fluor 647 or wheat germ agglutinin (WGA)-alexa fluor 633 1/1000 (Life Technologies) for 20 min at RT and washed twice in PBS. Slides were mounted in fluorsave (calbiochem) and observed under an Olympus FluoView confocal microscope using a 20× objective or a Nikon Perfect Focus spinning disc confocal microscope. Multiple images were collected per section. Distances separating microcolonies from the base of the villi were measured using the imaging software Imaris.
Mice were anesthetized with ketamine, xylazine, and acepromazine and placed in a supine position on a temperature-controlled heating pad. An ∼1.2 cm vertical incision was made along the midline of the abdomen through the skin and peritoneal membrane to expose the peritoneal cavity. A 1 cm loop of small intestine (proximal, medial, or distal segment) was carefully exteriorized through the peritoneum using cotton-tipped applicators to avoid tissue damage, and lightly immobilized with tissue-adhesive glue onto a heated stage. For intravital imaging, the intestinal loop was not opened along the antimesenteric border but rather left intact for the duration of the imaging procedure. Importantly, this approach best-preserved the physiology of the small intestine, including maintaining intact blood and lymphatic flow. The intestinal loop was kept hydrated by overlaying a mixture of saline/lubricant gel, and covered by a glass coverslip. Mice were given Hoechst 33342 (Sigma; 10 mg/kg i.v.) for nuclear staining in vivo, or Qtracker-655 non-targeted quantum dots (Invitrogen; 0.2 uM i.v.) to label the vasculature in vivo. In some experiments, segments of the small intestine were occluded at either end with sutures, and then surgically removed and imaged as an explant in a heated imaging chamber containing a mixture of saline/lubricant gel and covered by a glass coverslip. Time-lapse or static imaging was performed using an Ultima Two-Photon Microscope (Prairie Technologies). Two-photon excitation and second-harmonic signals were generated using a Tsunami Ti:sapphire laser with a 10-W MilleniaXs pump laser (Spectra-Physics), and outfitted with a 20× (0.95NA Olympus) water immersion objective. Two-photon excitation wavelength was tuned to 880–950 nm for optimal fluorescence excitation of fluorescent V. cholerae. Emitted light and second-harmonic signals were detected through 450/50-nm, 525/50-nm, 590/50-nm, and 665/65-nm bandpass filters for four-color imaging. Image sequences were transformed into volume-rendered z-stacks with Volocity software (Improvision) or Imaris (Bitplane).
A 100 mg/mL N-acetyl-L-cysteine (NAC) solution was prepared fresh in PBS and its pH adjusted to 7.3 with NaOH. 2 mg/g of the NAC solution or an equivalent volume of PBS (mock) was administered by gavage to 5-day old CD-1 mice.
Tissue samples were fixed in freshly made Carnoy's fixative (60% ethanol, 30% chloroform 10% acetic acid) for one hour at room temperature, washed in 70% ethanol and stored in 70% ethanol until further processing. Samples were embedded in paraffin, sectioned and stained with periodic acid-Schiff (PAS) at the Dana Farber/Harvard Cancer Center Rodent Histology Core.
|
10.1371/journal.ppat.1006320 | Localization of adenovirus morphogenesis players, together with visualization of assembly intermediates and failed products, favor a model where assembly and packaging occur concurrently at the periphery of the replication center | Adenovirus (AdV) morphogenesis is a complex process, many aspects of which remain unclear. In particular, it is not settled where in the nucleus assembly and packaging occur, and whether these processes occur in a sequential or a concerted manner. Here we use immunofluorescence and immunoelectron microscopy (immunoEM) to trace packaging factors and structural proteins at late times post infection by either wildtype virus or a delayed packaging mutant. We show that representatives of all assembly factors are present in the previously recognized peripheral replicative zone, which therefore is the AdV assembly factory. Assembly intermediates and abortive products observed in this region favor a concurrent assembly and packaging model comprising two pathways, one for capsid proteins and another one for core components. Only when both pathways are coupled by correct interaction between packaging proteins and the genome is the viral particle produced. Decoupling generates accumulation of empty capsids and unpackaged cores.
| Viruses assemble in particular locations inside infected cells where newly replicated genomes and capsids proteins meet, called viral factories. Virus genomes are packaged inside capsids by one of two mechanisms: concerted, where a protein shell is built around the genome, or sequential, where the genome is pumped into a preformed empty shell. Although adenoviruses have been studied for more than 60 years, these two basic aspects of their biology had not been elucidated. In this work, we address these two questions by determining the location in the cell nucleus where adenovirus assembly and packaging factors meet, describing failed assembly products containing unpackaged genomes, and showing for the first time images of adenovirus capsid fragments in the process of engulfing the viral DNA. Adenoviruses can cause serious clinical problems in immunosuppressed individuals, but can also be harnessed to turn from a pathogen into a useful therapeutic tool. Understanding their biology is crucial to succeed in curing adenovirus infections, and to repurpose the virus to our advantage.
| AdV virions consist of a 95 nm, icosahedral pseudoT = 25 protein shell enclosing a non-icosahedral DNA-protein core. Each capsid facet has 12 trimers of the major coat protein, hexon. A pentamer of penton base associated with a fiber trimer sits at each vertex. The most studied AdV, human AdV type 5 (Ad5), incorporates also four different minor coat proteins: IIIa, VI, VIII and IX. The AdV genome, a linear double stranded DNA molecule (~36 Kbp in Ad5), is tightly packed together with histone-like, virus-encoded proteins: core polypeptides V, VII and μ. The core also contains the terminal protein (TP) and the maturation protease (AVP) [1–3]. AdV assembly occurs in the nucleus, where hexon and penton, together with the minor coat proteins and the packaging protein L1 52/55 kDa, assemble into empty capsids. Viral genomes and core proteins are inserted into these capsids to yield noninfectious, immature particles [4]. These contain the precursor version of several capsid (pIIIa, pVI, pVIII) and core (pVII, pμ, pTP) proteins, as well as L1 52/55 kDa. Mature virions are produced upon cleavage of these precursors by AVP [5].
The Ad5 infectious cycle is completed in ~36 hours, with progeny virions appearing at 24 hours post infection (hpi). Tracing of AdV nucleic acids in infected cells has revealed where genome replication takes place [6–12]. Viral DNA is first detected (8 hpi) at the so-called early replicative sites (ERS), expanding away from nuclear domains 10 (ND10s) and containing viral ssDNA, dsDNA, and replicative activity [13]. Later on (~17 hpi), ERS evolve into two differentiated regions: the ssDNA accumulation site (DAS), harboring replicative intermediates and intermittent replicative activity, and the surrounding peripheral replicative zone (PRZ) where viral dsDNA is accumulated and there is continuous replicative activity (Fig 1A). Capsid proteins have been detected in a variety of AdV-induced nuclear structures, such as electron-clear inclusions (hexon, penton and polypeptide IX), electron-dense inclusions and compact rings (IVa2), or protein crystals (penton and fiber) [14–17]. However, in spite of the large amount of experimental work summarized above, it is still not clear where in the nucleus AdV assembly occurs.
The connection between capsid assembly and genome packaging is another poorly understood aspect in AdV morphogenesis. Two models have been proposed: sequential and concerted. In the sequential model, based on the dsDNA bacteriophage packaging mechanism, a motor complex would transfer the genome into a preformed capsid, using energy derived from ATP hydrolysis [18]. In the concerted model, the capsid proteins would assemble around the chromatin-like structure formed by the genome and condensing core proteins, similarly to what is thought to happen in polyomaviruses [19].
Evidence suggesting that AdV assembly is coupled to DNA synthesis would favor the concerted model. Inhibition of DNA synthesis, but not of viral proteins, results in reduced virus assembly, despite the presence of viral DNA accumulated previously to the inhibition; only DNA being synthesized is packaged into mature virions, and a thermo-conditional mutant in the AdV ssDNA binding protein (DBP) did not produce virus particles [20, 21]. DBP and ssDNA were found close to virions in cell sections, and interactions of DBP with packaging proteins have been reported [22, 23].
AdV packaging begins from the left end of the genome, where the specific packaging sequence (Ψ) is located [24]. Several virus-encoded proteins are required for genome packaging to occur: IVa2, L1 52/55 kDa, L4 22 kDa, L4 33 kDa, and IIIa. Production of empty capsids by thermo-sensitive or deletion mutants demonstrates that these proteins are required for DNA packaging but not for capsid assembly [25–30]. IVa2 and L1 52/55kDa interact during the course of AdV infection, and bind to Ψ in vivo independently of each other [31, 32]. Both L4 22 kDa and IVa2 bind to Ψ in vitro, and are required to recruit L1 52/55 kDa in vivo [27, 33, 34].
In genome-less AdV particles, L1 52/55 kDa forms a disordered shell beneath the icosahedral capsid, with preferential location under the vertex [35]. Interaction of L1 52/55 kDa with polypeptide IIIa, a component of the icosahedral shell located beneath the pentons, determines packaging specificity [3, 36]. Polypeptide IIIa interacts with L1 52/55 kDa in vitro and with Ψ in vivo, indicating how the genome may be tethered to the capsid during assembly. L1 52/55 kDa is released from the viral particle by proteolytic maturation, which leads to loss of interaction of this protein with itself, core and capsid proteins [35, 37]. The available evidence suggests that L1 52/55 kDa mediates the stable association between the viral DNA and the empty capsid to produce a full particle. Immature particles contain full length L1 52/55 kDa and are unable to release their genome, which stays attached to capsid fragments even under harsh in vitro disruption treatments [38].
Polypeptide IVa2 is thought to be present at a single vertex in the virion [39], and has Walker A and B motifs associated with ATP hydrolysis [40]. IVa2 binds ATP, and an intact Walker box is required for virion production [41], but only weak ATPase activity has been reported [42]. All these findings support the idea of IVa2 acting as the packaging motor, or a part of it, favoring the sequential assembly and packaging model. A difficulty that this model has to circumvent arises from the fact that AdV genomes are bound to cellular and/or viral proteins throughout the infectious cycle [43–47]. It is not clear how dsDNA bound to proteins would be translocated by a motor, or how the proteins would first be removed, then penetrate the capsid to reassociate with the packaged DNA. Packaging of the protein-bound AdV genome via an ATP driven portal would require a portal structure or mechanism different from those currently known in other viruses, or a nucleoprotein remodeling activity. Further complication is posed by the observation that the genome bound to immature core proteins is a highly compact, ~70 nm sphere [38, 48–50]. Proteins IVa2 and L1 52/55K are found both in empty capsids and bound to Ψ, suggesting that they may be present in two separate pools: a capsid-associated pool poised to receive viral DNA for encapsidation, and a second pool bound to Ψ to promote interaction between viral DNA and capsid components [36].
Another pillar supporting the sequential model is the routine appearance in AdV purifications of low density particles. These particles are considered procapsids (precursors to mature virions), because in pulse-chase experiments they appear earlier, contain protein precursors and no genome, or only fragments [4, 51–55]. The presence of different lengths of packaged DNA has also been taken as evidence for the sequential packaging model. However, the actual origin of these DNA fragments is not well understood, and an alternative origin as replication artifacts has been proposed [55, 56]. Additionally, light particles do not progress into mature virions, suggesting that they are not assembly intermediates but defective assembly products [57, 58]. This possibility is supported by the recent molecular characterization of incomplete particles produced by an Ad5 variant with delayed packaging, Ad5/FC31, showing that low density particles had started, but failed to complete packaging [35].
Ad5/FC31 was generated by insertion of the Φ31 recombinase target sequences, attB/attP, flanking Ψ [59]. In cells not expressing the recombinase, Ad5/FC31 viral protein and DNA production levels are similar to those of control virus, but the mutant produces negligible amounts of mature virions at 36 hpi, reaching virus yields similar (10-fold lower) than the control at only 56 hpi [35, 59]. Electrophoretic mobility shift assays suggested that nuclear proteins bound to attB interfered with correct interaction between packaging proteins and Ψ. As a result, packaging would be hindered until the interfering proteins are depleted [60].
Here we use immunofluorescence and immunoelectron microscopy to investigate the location in the cell of AdV packaging factors at late times post infection, and determine the location where genome encapsidation occurs. The labeling patterns of these factors, and the nuclear modifications induced by Ad5 and the delayed packaging mutant Ad5/FC31, are compared to obtain new information on the connection between assembly and packaging.
To start defining the nuclear region where AdV packaging happens, viral genomes and packaging factors IVa2 and L1 52/55 kDa were localized by immunofluorescence microscopy in Ad5 wt or Ad5/FC31 infected cells at late times post infection. Previous work had shown a large divergence in mature virus production between control virus and Ad5/FC31 at 36 hpi [59]. Additionally, an electron microscopy (EM) survey of Epon-embedded infected cells from 24 to 56 hpi indicated that the nuclear modifications induced by both viruses during the first 24 hours were highly similar [61], while some noticeable differences (discussed later on) were found at 36 hpi and peaked at 48 hpi. Therefore, we carried out the experiments described here at 36 or 48 hpi.
Taking into account that at 18 hpi cellular DNA replication no longer occurs (S1 Fig) [62], two doses of BrdU were supplied to infected cells at 18 and 25 hpi, to ensure that all viral DNA synthesized at late times post-infection was labeled. BrdU presented a diffuse ring pattern, and more DNA (more labeled cells) was detected in Ad5/FC31 than in wt infections, consistent with the mutant normal replication but deficient packaging phenotype [60] (Fig 1B and 1C). Measurements in orthogonal views indicated that the labeled rings were in fact ellipsoids with maximum and minimum axes 6.8 ± 1.8 μm and 5.0 ± 1.2 μm (n = 40). Since this label pattern is similar to that previously reported for earlier (20–24 hpi) replication foci [9, 12], we reasoned that the diffuse ring BrdU label corresponded to the PRZ, and the unlabeled area inside the ring could correspond to the DAS (Fig 1A). In double labeling experiments, DBP was found adjacent to or surrounded by BrdU label (Fig 1D). Regions labeled for DBP but not for BrdU would correspond to ssDNA synthesized at early times post-infection, before addition of the first BrdU dose. These results indicate that BrdU labeling is revealing AdV replication centers at late times post-infection.
To investigate where packaging proteins and viral genomes meet, double labeling assays for BrdU and L1 52/55kDa, or BrdU and IVa2 (S1 Text), were carried out. Label for IVa2 was weak, perhaps correlating with the low copy number of the protein in the virions, or with limited antibody reactivity. L1 52/55kDa was detected at the periphery of replication centers labeled with BrdU (Fig 1E), where both signals intermingled. There was no label for L1 52/55 kDa in the BrdU-unlabeled areas corresponding to the DAS. The presence of both L1 52/55 kDa and viral genomes in the PRZ suggests that DNA packaging occurs in this area.
To obtain more detail on the possible packaging site of AdV, EM was used. First, sections from Epon-embedded cells were analyzed for regions in the nucleus that could correspond to PRZs. Previous studies [10, 11, 23] indicated that PRZs are ring-shaped, moderately electron-dense regions, surrounding electron-clear areas corresponding to the DAS, and containing electron-opaque granules (EOGs) and viral particles. Regions corresponding to these characteristics, and in a size range compatible with the immunofluorescence observations described above, were identified in both Ad5 wt and Ad5/FC31 infected cells (S2 Fig). EOGs were often found in contact with loose electron-dense material, which was tentatively called DNA bundle because of its texture (S2 Fig).
To corroborate the identity of the possible PRZs, infected cells were treated with BrdU and processed for immunoEM by freeze substitution (FS), to preserve both structure and immunoreactivity. In FS samples, the possible PRZs and DAS regions were identified on the basis of their electron-density and the presence of EOGs and viral particles (Fig 2A and 2E). Label for BrdU was specifically found in the possible PRZ area, confirming its identity (Fig 2B, 2F and 2I), and was frequently associated to full particles (virions) (Fig 2B and 2G), EOGs (Fig 2D and 2H) and the loose electron-dense material (“bundles”, Fig 2B, 2C, 2F and 2G), confirming that they contain viral DNA. The different electron-density levels of these structures are suggestive of different degrees of DNA condensation, from most relaxed (bundles) to most condensed (virions and EOGs). The BrdU signal in EOGs indicates the presence of viral DNA, and not only RNA as reported by other studies [63]. EOGs had various sizes but they were generally larger than viral particles. No significant BrdU label was observed in other nuclear regions or in the cytosol. The electron-clear area proposed to be the DAS showed weak label (Fig 2B and 2F), in agreement with the immunofluorescence results and previous reports indicating low replicative activity in this region. Labeling with anti-DBP antibody was unsuccessful, suggesting that even in FS conditions the reactivity of the DBP epitopes was not preserved.
Next, the localization of packaging proteins was analyzed. Scattered label for L1 52/55 kDa was observed throughout the infected nuclei, including the PRZ (Fig 3A, 3E and 3I). However, very few gold particles were observed in the DAS, supporting the specificity of the label (Fig 3I and S1 Table. L1 52/55 kDa signal in the PRZ was usually associated to the electron-dense, BrdU-positive features present in this area (EOGs and bundles, Fig 3). As expected, L1 52/55 kDa was detected in viral particles, particularly in those with lower electron density indicating incomplete packaging (Fig 3B and 3F). Label in viral particles often presented an arch pattern consistent with a shell of this protein inside the capsid [35]. Arch patterns were also found in EOGs (Fig 3C and 3G), suggesting the formation of an L1 52/55 kDa shell on the electron dense material they contain. These observations are consistent with the presence of L1 52/55 kDa in two pools, one binding to the viral DNA (bundles and EOGs) and another binding to capsid proteins (electron-clear capsids) [36]. Interestingly, groups of gold particles also forming little arches were frequently found near the PRZ (Fig 3D and 3H), suggesting L1 52/55 kDa shell fragments on their way to assemble with viral genomes or capsid proteins. This interpretation is consistent with the previously reported homo-oligomerization capacity of L1 52/55 kDa [37, 64].
Label for packaging protein IVa2 was weak, as previously observed in immunofluorescence (S1 Text). Nevertheless, signal for IVa2 was present in the PRZ on electron-dense material (bundles and some EOGs) (S4 Fig and S1 Table). Summarizing, immunoEM confirmed that AdV genomes and packaging proteins are present in the PRZ. Genomes were present in the form of loose bundles and EOGs that by their electron density, texture and label for BrdU and packaging proteins are consistent with condensed viral genomes. PRZs also contained viral particles. These results suggest that the PRZ could be the location in the nucleus where AdV genome encapsidation takes place.
To further assess the hypothesis that AdV assembly occurs at the PRZ, the presence of core (VII) and capsid (fiber) proteins was analyzed. Label for VII was exclusively observed in the PRZ, frequently in EOGs (Fig 4) and in lower amounts in DNA bundles (Fig 4F), corroborating the idea that they contain viral DNA condensed to different degrees by core proteins. Protein VII was also detected in viral particles (Fig 4C).
Analyzing capsid protein localization in AdV infected cells is not straightforward, since they are produced in large excess [65]. Antibodies against fiber labeled protein crystals (Fig 5A and 5C) and viral particles (Fig 5B and 5D), as expected. Fiber was also detected in the PRZ (Fig 5E and 5F), in particular in EOGs and bundles, but only a weak signal was observed in the DAS, supporting the specificity of the label (Fig 5G and S1 Table). The presence of core and capsid proteins, together with viral DNA and packaging factors, as well as viral particles, indicates that the PRZ is the AdV assembly site, i.e. the AdV assembly factory.
After determining that representatives of all AdV morphogenesis players (genome, packaging factors, core and capsid proteins) were present in the PRZ, we addressed the question of where exactly in this region was assembly taking place. Both EOGs and bundles were positive for all tested assembly factors, and both viral particles and EOGs were often found at the DNA bundle periphery, suggesting a topological relation (Fig 6A). Detailed observation of FS samples revealed half capsids engulfing DNA condensations protruding from the bundles (Fig 6B), indicating that these are most likely the assembly sites. These assembly intermediates were extremely infrequent in Ad5 wt infections, and very hard to find in Ad5/FC31: only 8 (Ad5 wt) and 25 (Ad5/FC31) such intermediates were found in a dataset consisting of 19 cells and a total scanned area of over 1600 μm2 for each virus. This observation indicates that AdV assembly is a highly cooperative process.
Exhaustive examination of all FS samples yielded a possible sequence of events in AdV assembly. On the one hand, capsid fragments are assembled containing L1 52/55 kDa (Fig 6C, early stage for capsids). On the other, L1 52/55 kDa also appears at the periphery of DNA bundles, often in small electron-dense protrusions suggesting the condensing action of core proteins (Fig 6C, early stage for core). We propose that these protrusions are nascent viral cores, containing one of the two L1 52/55 kDa pools (the one bound to the packaging sequence), and serving as the recruitment spot for the other L1 52/55 kDa pool (the one bound to capsid fragments). Incoming capsid fragments assemble around the coalescing core (Fig 6C, intermediate stage 1), and gradually grow (Fig 6C, intermediate stage 2) until the complete particle is formed (Fig 6C, late stage 1) and finally detaches from the DNA bundle (Fig 6C, late stage 2). Although EM images from cell sections are static snapshots, the observation for the first time of AdV capsids assembling around nascent cores is strongly suggestive of a concerted rather than sequential assembly and packaging mechanism.
Empty capsids generated by AdV mutants with packaging defects are failed assembly products made by capsid components [25–29, 35]. However, the fate of unused core components in packaging defective mutants is not known. We show here that EOGs arise from viral DNA bundles, and are labeled for structural and packaging factors, but adopt variable shapes and sizes different from those expected for a viral particle (Fig 6A). EOGs can therefore be interpreted as a different class of failed assembly events: cores whose association with capsid fragments was unsuccessful. In agreement with this hypothesis, we observed that EOGs were more abundant in Ad5/FC31 than in wt infections (Figs 3A and 3E; 4B and 4E; S5 Fig and S1 Table).
Apart from the amount of EOGs and light particles, the most remarkable difference found between Ad5/FC31 and wt infected cells was a new structure consisting of dots with very high electron density embedded in a high electron-dense background, that we named “speckled body” (SB) due to its appearance (Fig 7). The possibility that SBs were compact nucleoli was ruled out, because the speckles (70 ± 11 nm, n = 50) are larger than nucleoli dots (ribosomes, 25–30 nm) (Fig 7E), and similar in size to viral particles (Fig 7A, insets).
SBs were observed in Ad5/FC31 infected cells from 36 hpi, and their presence was most noticeable at 48 hpi, when they had a lobular, loose organization (Fig 7A), while at later times they appeared more compact and circular (Fig 7B and 7D). Extensive search revealed that SBs were also present in Ad5 wt infected cells, but their occurrence was extremely rare: only 2 SBs were found in a sample of 36 Ad5 wt infected cells, while 14 SBs were found in 45 Ad5/FC31 infected cells, giving a 5.5% vs 31.1% probability of finding a SB in an infected cell. Because of their size and texture, the speckles in the SBs are reminiscent of viral cores (Fig 7A, insets) [48]. SBs were often located adjacent to the PRZ (Fig 7A and 7B). We therefore hypothesized that SBs could be PRZ regions containing viral condensed genomes that had not been packaged due to the Ad5/FC31 mutation.
To assess the hypothesis that SBs contain condensed AdV genomes, the presence of two core components (DNA and protein VII) was tested. In initial immunolabeling experiments, no signal was observed for BrdU (viral DNA), and the signal for VII was low (Fig 7F). Because there was a possibility that the VII epitopes were masked by the tight complex between the condensing protein and DNA, sections were treated with DNase before incubation with the anti-VII antibody. This treatment increased the signal for VII (Fig 7G and 7H), indicating not only that SBs contain VII, but also that they contain DNA, and therefore viral cores. The lack of BrdU signal would indicate that the DNA in SBs was produced in the first 18 hpi, prior to incorporation of BrdU. Label for L1 52/55 kDa and fiber in SBs was comparable to that in PRZs (Figs 3I and 5G, and S1 Table). Some SBs had a ring shape with an electron-clear center reminiscent of the DAS (Fig 7C), suggesting that they could be early collapsed PRZs. However, no viral particles (neither empty nor full) were found in SBs. We conclude therefore that SBs are early PRZs, where capsid to core recruitment started, but failed due to the Ad5/FC31 mutation that interferes with the interaction between packaging proteins and Ψ. As a result, assembly failed to proceed.
AdV packaging (genome, IVa2, L1 52/55k) and assembly (core and capsid proteins) factors meet at the PRZ, where virus particles are also found (Fig 7I). Importantly, viral genomes and core protein VII were exclusively localized at the PRZ. These observations indicate that the PRZ is the AdV assembly factory, and not only the DNA replication zone as previously described [10, 11]. This localization of the AdV assembly site is consistent with previous evidence indicating that replication and assembly are coupled, and therefore should happen in the same place [20]. Our results may seem contradictory with previous studies concluding that packaging protein L1 52/55 kDa is not present at replication centers [12, 64]. The use of different labeling probes (for DBP or nucleolar proteins vs viral genomes) or strategies (pulse vs prolonged deoxyuridine incubation, cell extraction vs FS) may be responsible for this discrepancy.
EM images of assembly intermediates in infected cells that show capsids growing around cores provide new evidence to support the concerted assembly and packaging model. Putting together previous evidence [35–37] and the results presented here, we propose the following model for AdV morphogenesis (Fig 7J). On the one hand, packaging proteins bind to Ψ in the DNA bundles produced during genome replication. At the periphery of the bundles, viral genomes start to condense by the action of core proteins. On the other hand, in areas close to the PRZ, full length L1 52/55k would bind to IIIa in icosahedral shell fragments. The two pools of L1 52/55 kDa (in capsid fragments and nascent cores) interact and act as a Velcro to recruit and tether the condensing core to the nascent capsid. Capsid growth proceeds by addition of capsomers or other fragments around the core, while simultaneously L1 52/55k is cleaved by AVP and removed from the particle before capsid closure and disengagement from the DNA bundle. Changes in Ψ or the packaging proteins that impair their interaction with each other would hinder these processes at different points, resulting in genome-less particles with different degrees of maturation cleavages and abandoned cores that would eventually coalesce into EOGs and SBs. Like empty particles, EOGs and SBs are more abundant in Ad5/FC31 than in Ad5 wt infections.
Our observations and proposed assembly model are in remarkable agreement with recent computational predictions on the assembly of protein shells around a simultaneously coalescing multiparticulate cargo, which in this case would be the AdV core organized in nucleosome-like DNA-protein units [1, 66]. A delicate balance between shell-shell, cargo-cargo, and shell-cargo interactions is required for correct assembly of full particles (Fig 7I). Mutations that impair the network of interactions established by packaging proteins with both capsid and genome upset this balance and lead to abortive assembly products. In particular, the simulations in [66] predict that when shell-core interactions are diminished, only a few cargo molecules will be captured, leading to nearly empty shells. Notably, these empty shells can be completely closed, explaining a puzzling occurrence previously observed for the genome-less particles generated by Ad5/FC31 [35]. The simulations in [66] also show that in this type of concurrent mechanism, assembly of full shells is remarkably fast (two orders of magnitude faster than assembly of empty shells). Our results indicate that, when capsid-core assembly coupling is correct, AdV morphogenesis is a fast, highly cooperative process, as indicated by the scarcity of assembly intermediates in both virus variants studied, but most markedly in Ad5 wt. The Ad5/FC31 mutation changes shell-core interactions allowing us to catch assembly intermediates in a process otherwise too fast to be experimentally observed.
Because AdV has a dsDNA genome, encodes a putative packaging ATPase (IVa2), and produces large amounts of empty capsids, it seemed plausible that its assembly and packaging would take place in a sequential way similar to that of tailed phages. However, the organization of an AdV virion is quite different from that of a tailed phage. Instead, AdV belongs to a large group of viruses (the PRD1-adenovirus lineage) that use β-barrels orthogonal to the capsid surface to assemble icosahedral capsids with diameters ranging from ~0.06 to ~1 μm. The structural similarity found in the viral particles raised the question of a common ancestry for these viruses [67], which might also be reflected in aspects of their assembly mechanism.
Assembly and packaging in the smallest PRD1-AdV lineage viruses seem to follow different strategies depending on their particular genome type [68]. Bacteriophage PRD1 has a multi-protein channel spanning both capsid and membrane at a unique vertex, through which the linear genome is encapsidated into preformed capsid-membrane particles using power provided by its packaging ATPase [69, 70]. Conversely, the membrane-containing marine bacteriophage PM2 has a highly supercoiled circular genome, which would topologically complicate the motor driven DNA translocation. It has been proposed that interactions between the condensed PM2 genome and scaffolding membrane proteins start membrane bending and nucleate coat protein recruitment, coupling membrane morphogenesis, genome encapsidation, and capsid assembly [71]. The abundant coverage of the linear AdV genome by condensing proteins, producing a particularly tight complex in the immature particle [1, 38, 48–50, 72], would also pose a considerable topological complication for DNA translocation, which might be better resolved by a simultaneous assembly and packaging strategy.
Details consistent with the assembly and packaging mechanism proposed for PM2, which also bear striking similarities to our findings on AdV assembly, have been reported for some of the most complex PRD1-AdV lineage members. The Mimivirus factory originates from replication centers, as shown here for AdV [73]. Structural proteins are recruited to the viral membrane to create a first icosahedral vertex from which the capsid will grow. The open membrane, which holds the viral DNA, is progressively coated by capsid proteins to form an icosahedral particle, similarly to our observation of AdV capsid fragments engulfing the condensing core. A similar mechanism has been proposed for African swine fever virus (ASFV) [74]. Bacteriophage PM2, ASFV, and Mimivirus, all contain a membrane inside the icosahedral shell. In AdV, there is no internal membrane, but packaging protein L1 52/55 kDa may play a similar role to the membrane proteins in these viruses, helping to nucleate and keep together capsids fragments and growing cores during assembly.
Numerous questions remain to be answered to fully understand the AdV assembly mechanism. One of them is whether at least an initial, however imperfect, contact between capsid and core (via packaging proteins) would be necessary for capsids to start assembling. The existence of numerous packaging protein mutants that produce large amounts of empty capsids would indicate that this is not the case. However, a DBP thermo-conditional mutant with a 3-fold reduction in genome replication did not produce any kind of particles [21], and Ad5/FC31 incomplete particles had at least initiated packaging [75]. Since a variety of proteins (L1 52/55 kDa, IVa2, L4 33 kDa, IIIa) have a role in encapsidation, it is possible that deletion of only one of them still allows the initial stages of packaging, and therefore capsid assembly, to occur. The role of the packaging sequence Ψ is also intriguing. No empty particles were produced when part of the packaging sequence was deleted (mentioned in [64]), but an Ad5 construct containing loxP sites flanking the packaging sequence does accumulate empty particles when propagated in Cre-expressing cells [27, 28]. This observation does not rule out the concerted model, because deficient Cre function may result in incomplete cleavage of Ψ, still allowing an initial interaction between capsid and core components; also, it has been observed in the development of helper systems for gutless AdV production that residual helper virus can package genomes lacking Ψ [76]. Systematic molecular and structural characterization of assembly products from the various mutants would be required to assess this point.
Another unanswered question is whether the first capsid fragment to be recruited to the core would be chosen at random, or be defined by the presence of some “special feature”, similar to the stargate in Mimivirus [73]. Part of this special feature may be the presence of the putative packaging ATPase IVa2 [39]. It is an outstanding mystery what the role of the IVa2 ATPase function would be, if encapsidation does not happen by translocation through a portal. While packaging ATPases clustering in the HerA/FtsK superfamily are a common feature in PRD1-AdV lineage members [77], their exact function is not clear yet, and may vary in each particular virus. Additionally, IVa2 is an outlier, making up its own separate branch in the ATPase classification. In Mimivirus, it has been proposed that the ATPase may have a segregation-like role in disentangling individual genomes from the replication factory [78]. A similar role would be conceivable for IVa2 in our model.
In summary, we present here evidence that defines the location of the AdV assembly factory, and favors a concerted assembly and packaging mechanism for AdV morphogenesis. Detailed understanding of AdV assembly has been hindered by the lack of an in vitro, cell-free assembly system. The work presented here, together with the recently described isolation of functional replication centers [79], may constitute the basis for advancing towards such a system and conclusively resolve the dynamics of AdV morphogenesis.
HEK293 cells (ATCC CRL-1573) were cultured at 37°C in Dulbecco’s modified Eagle’s medium (DMEM, Sigma Cat# D6429) supplemented with 2% fetal bovine serum (FBS, Biological Industries Cat# 04-001-1A), 10 units-10 μg/ml penicillin-streptomycin (Sigma Cat# P4333), 0.05 mg/ml gentamicin (Sigma Cat# G1397), 4 mM L-Glutamine (MERCK Cat# 3520) and 1X non-essential amino acid solution (Sigma Cat# M7145).
Control wild type virus was the Ad5 variant Ad5GL, where the E1 region has been deleted and substituted by GFP and firefly luciferase genes [80]. The Ad5/FC31 variant has an attB/attP insertion flanking Ψ and a GFP cassette following Ψ [59]. Virus propagation and purification was carried out as described [35].
The primary antibodies used were: rat anti BrdU monoclonal abcam #ab6326; rat anti Ad5 pVII serum [81]; mouse anti Ad2 IVa2 serum [82]; mouse anti Ad5 DBP monoclonal [83]; rabbit anti Ad5 L1 52/55 kDa serum [84]; and rabbit anti Ad5 fiber serum [85].
The secondary antibodies used were: for immunofluorescence assays, Alexa Fluor594 Goat Anti-Rat (Invitrogen #A-11007), Alexa Fluor555 Goat Anti-Rat (Invitrogen #A-21434), Pacific Blue Goat Anti-Mouse (Invitrogen #P-31582) and Pacific Blue Goat Anti-Rabbit (Invitrogen #P-10994). For immunoEM, 10 nm gold conjugated gold anti-rabbit (BB International #EM-GFAR10), 10 nm gold conjugated anti-mouse (EM-GFAF10), and 15 nm gold conjugated anti-rat (EM-GAT15).
HEK293 cells grown in p100 culture plates to 70% confluence were infected with Ad5/FC31 or Ad5 wt with a multiplicity of infection (MOI) of 5. At the desired time post infection, the medium was removed and the cells were fixed with 2% glutaraldehyde and 1% tannic acid in 0.4 M HEPES pH 7.2 for 1.5 h at room temperature. Embedding in Epon was carried out as previously described [86]. Ultrathin sections (~70 nm) were collected on Formvar-coated nickel grids, stained with saturated uranyl acetate and lead citrate as described previously [86] and examined in a JEOL JEM 1230 transmission electron microscope at 100kV. Mock infected cells were processed in the same manner as a control.
Monolayer HEK293 cells grown in cover glasses were infected with Ad5 wt or Ad5/FC31 with MOI = 50. The infection was synchronized by incubating the cells for 30 min at 4°C and then 30 min at 37°C. Then, the inoculums were removed and medium was added. For BrdU labeling, after 18 h at 37°C the medium was changed by medium containing 25 μg/ml BrdU (5-Bromo-2’-deoxyuridine, Sigma Cat#B5002-1G), followed by another change at 25 hpi. Incubation with BrdU proceeded at 37°C. After 36 hpi, the medium was removed and 4% paraformaldehyde in PBS was added to the cells during 10 min. After 3 rinses with PBS, cover glasses were incubated with a mixture of 0.5% saponin and 10% FBS in PBS for 10 min. Samples were incubated with primary antibody in 0.5% saponin and 2% FBS in PBS during 45 min. After three more rinses, incubation with secondary antibodies diluted in 0.5% saponin and 2% FBS in PBS was carried out in darkness. After a final rinse with PBS, cover glasses were mounted on glass slides using ProLong (Invitrogen Cat# P36930) (4 μl drops). The antifade reagent was allowed to dry overnight before sample observation. All incubations were carried out at room temperature. Images were taken using a confocal multispectral Leica TCS SP5 system. Negative controls consisted of incubations without primary antibody, and immunolabeling of mock infected cells. In double labeling experiments, absence of cross-reactivity was verified by incubations with only one of the primary antibodies and the secondary antibody corresponding to the other.
The following modifications to the protocol described above were applied for anti-BrdU labeling: fixed samples were washed three times with saponin 1% in PBS (3x5 min), then subjected to DNA denaturing treatment: 1N HCl during 10 min at 4°C, followed by 2N HCl during 10 min at room temperature, and finally 20 min at 37°C. To neutralize, borate buffer (4 g NaOH; 23.5 g boric acid to 500 ml in milliQ water, pH 8.2) was added for 12 min at room temperature, and the protocol continued with the block-permeabilization step described above, but with 1% instead of 0.5% saponin. For double labeling assays, the primary antibodies were used together in the same incubation, except for the case of double labeling for IVa2 and L1 52/55 kDa. In this case, samples were incubated with the mouse anti-IVa2 serum and fixed with 4% paraformaldehyde for 5 min before incubating with the rabbit anti-L1 52/55 kDa serum. Immunofluorescence image analyses were carried out with Image J [87].
BrdU labeling of newly synthesized DNA was carried out as described for immunofluorescence, with HEK293 cells grown in p100 culture plates instead of cover glasses. Infected and control cells were fixed with 4% paraformaldehyde in PBS. After rinsing three times with PBS, glycerol was added drop by drop to 15% final concentration. After 15 min at 4°C, glycerol was increased to 30% and 15 min later the cells were harvested, pelleted and frozen by plunge freezing in liquid ethane using a Leica CPC plunger. Freeze substitution was carried out in a Leica EM automatic freeze-substitution system (AFS). Samples were maintained in 0.5% uranyl acetate in dry methanol for 60 h at -90°C, with several changes of dehydrating solution. Then, the temperature was raised in a controlled manner to reach -40°C after 7 h and maintained to the end of the procedure. Samples were rinsed with dry methanol, and infiltrated with growing concentrations of Lowicryl HM20 in methanol for 24 h. Polymerization was carried out by UV irradiation for 48 h at -40°C, and then 48 h at 20°C. Ultrathin sections were obtained as described in the conventional electron microscopy section.
For immunoEM assays, grids carrying freeze-substituted ultrathin sections were placed on TBG (30 mM Tris-HCl pH 8, 150 mM NaCl, 0.1% BSA and 1% gelatin) drops for 10 min, incubated with the primary antibody in TBG for 30 min, washed 3 times with PBS, floated on 4 TBG drops (5 min per drop) and incubated in gold-conjugated secondary antibodies diluted in TBG for 30 min. Then, the grids were washed 3 times with PBS and milli-Q water, and stained with saturated uranyl acetate [86]. All incubations were carried out at room temperature. Negative controls consisted of incubations without primary antibody, and immunolabeling of mock infected cells.
For anti-BrdU labeling, an additional step was required. Sections were treated with 0.2 mg/ml proteinase K (Roche Cat# 3115879) for 15 min at 37°C, then washed with milli-Q water and denatured with 2N HCl for 25 min. After several (~4) rinses in milli-Q water, the protocol continued with the TBG incubation. To unmask protein VII, sections were floated on three drops of DNase buffer (10 mM Tris-HCl pH 8.2, 10 mM NaCl, 5 mM MgCl2) (5 min per drop), then incubated with 50 μg/ml DNAse (Sigma Cat# D5025) for 1 hour at 37°C, rinsed in milli-Q water and transferred to TBG to start the immunogold labeling protocol. Incubation with the anti-VII primary antibody was performed overnight at 4°C.
To quantify the occurrence of structural features (EOGs, SBs, assembly intermediates) or gold labels in Epon embedded or FS samples, sections were scanned at low magnification to locate cells with replication centers, which were then imaged at high magnification in as many micrographs as needed to cover all the visible part of the nucleus. The area corresponding to the different nuclear regions (PRZ, DAS, rest of the nucleus) was measured using ImageJ [87]. Gold labels in viral particles were not included in the quantification. Data are presented as box plots where center lines show the medians; box limits indicate the 25th and 75th percentiles; and whiskers extend 1.5 times the interquartile range from the 25th and 75th percentiles. Details of statistical analyses are summarized in S1 Table.
|
10.1371/journal.pntd.0006851 | Caprine humoral response to Burkholderia pseudomallei antigens during acute melioidosis from aerosol exposure | Burkholderia pseudomallei causes melioidosis, a common source of pneumonia and sepsis in Southeast Asia and Northern Australia that results in high mortality rates. A caprine melioidosis model of aerosol infection that leads to a systemic infection has the potential to characterize the humoral immune response. This could help identify immunogenic proteins for new diagnostics and vaccine candidates. Outbred goats may more accurately mimic human infection, in contrast to the inbred mouse models used to date. B. pseudomallei infection was delivered as an intratracheal aerosol. Antigenic protein profiling was generated from the infecting strain MSHR511. Humoral immune responses were analyzed by ELISA and western blot, and the antigenic proteins were identified by mass spectrometry. Throughout the course of the infection the assay results demonstrated a much greater humoral response with IgG antibodies, in both breadth and quantity, compared to IgM antibodies. Pre-infection sera showed multiple immunogenic proteins already reactive for IgG (7–20) and IgM (0–12) in most of the goats despite no previous exposure to B. pseudomallei. After infection, the number of IgG reactive proteins showed a marked increase as the disease progressed. Early stage infection (day 7) showed immune reaction to chaperone proteins (GroEL, EF-Tu, and DnaK). These three proteins were detected in all serum samples after infection, with GroEL immunogenically dominant. Seven common reactive antigens were selected for further analysis using ELISA. The heat shock protein GroEL1 elicited the strongest goat antibody immune response compared to the other six antigens. Most of the six antigens showed the peak IgM reactivity at day 14, whereas the IgG reactivity increased further as the disease progressed. An overall MSHR511 proteomic comparison between the goat model and human sera showed that many immune reactive proteins are common between humans and goats with melioidosis.
| B. pseudomallei infection, the causative agent of melioidosis, results in severe disseminated or localized infections. A systemic study of the humoral immune response to B. pseudomallei infection using the B. pseudomallei aerosol caprine model would help understand the detectable antigenic proteins as the infection progresses. To study the immune response, IgG and IgM antibody responses to whole cell lysate proteins were identified and analyzed. Antigenic carbohydrates were also studied. From the results, this study suggests that the caprine humoral immune response to aerosolized B. pseudomallei has similarities to human melioidosis and may facilitate the analysis of the temporal antibody responses. In addition, commonly detected immunogenic proteins may be used as biomarkers for the future point of care (POC) diagnostics.
| Burkholderia pseudomallei is a Gram-negative, non-spore forming, aerobic, and motile bacillus [1] and the etiological agent of melioidosis. This disease has emerged as a significant public health threat in Southeast Asia and Northern Australia [2]. Both B. pseudomallei and its close relative, B. mallei, the cause of glanders, are classified by the Centers for Disease Control and Prevention as category B bioterrorism agents [3]. In Thailand, B. pseudomallei is widely distributed in water and wet soils, such as rice paddies [4, 5]. In the two highly endemic regions of Northern Australia and Northeast Thailand, B. pseudomallei is responsible for a melioidosis fatality rate of around 10% and 40%, respectively [2, 5, 6]. There is also emerging evidence that melioidosis is endemic in Central and South America [7, 8], southern regions of China [9, 10] and India [11, 12]. The global distribution of B. pseudomallei endemicity has been linked to anthropogenic dispersal, both ancient and more recent [13]. Furthermore increased travel [14], and soldiers returning from endemic countries have led to many cases in non-endemic regions such as the USA and Europe [14, 15]. Limmathurotsalkul et al reported an evidence-based estimate of B. pseudomallei global distribution across the tropics; 46 countries were identified as suitable for melioidosis and with environmental suitability for the persistence [16]. Detection of cases outside endemic countries is also now helped by an increased awareness of melioidosis by clinicians worldwide [2, 16]. The main routes of B. pseudomallei infection are dermal inoculation, inhalation and ingestion [17–19]. Melioidosis clinical presentation spans from pneumonia (50% of all cases) [18, 20], and sepsis often leading to septic shock often with multiple abscesses in internal organs such as spleen, liver, kidney and prostate [21], to chronic abscesses in the skin without sepsis [18, 22].
Melioidosis can be successfully treated if diagnosed early and correctly. However, confirmed evidence of melioidosis infection in a patient currently relies on isolation and identification of B. pseudomallei from culture, often requiring use of selective medium [23, 24]. The bacteriological method takes a minimum of 24–48 hours, making it too slow to guide early treatment, which is particularly problematic for severe sepsis with its high mortality rate [2, 24–26]. To improve the diagnosis of melioidosis, a number of techniques have been attempted, such as antigen detection in specimens, antibody detection, molecular and rapid culture techniques [27]. A number of serological tests for antibody detection have also been developed for possible early diagnosis of melioidosis, viz. indirect hemagglutination (IHA), protein microarray, immunofluorescence (IFA), enzyme linked immunosorbent assay (ELISA), Flow lateral analysis [2, 28–31]. However, there is still a need to improve the test specificity and sensitivity. Some of the problems could be associated with selection of biomarker candidate antigens, resulting in false positive results. Therefore, a sensitive, specific and rapid test is still needed for diagnosis of melioidosis, especially for those presenting with severe sepsis [6, 24]. Recombinant proteins as diagnostic antigens for serology may offer advantages over IHA, which is currently the only serological assay which used. The current IHA tests have low sensitivity early in infection plus poor specificity in endemic regions due to prior background exposure; i.e. false positives are not uncommon [2, 32].
In the last 10 years, melioidosis and B. pseudomallei have attracted increased global attention and the related research has likewise increased considerably. However, there are large gaps in our knowledge of the pathogenesis of B. pseudomallei infection and the host immune response [6]. The comparative studies of an animal model and human infection can provide significant insight on the understanding of pathogenesis and immune response. Through the systemic study of the immune response and antigenic protein detection, potential candidates for the diagnostics of the infection in the early stage of disease process can be found. The purpose of this study was to analyze the humoral immune response of the caprine model that was aerosol-challenged with B. pseudomallei, mimicking inhalational melioidosis in humans. Melioidosis is common in goats living in melioidosis-endemic locations and the disease in goats has many parallels to human melioidosis [33]. We hypothesize that goats will have a complex and an immunological response to multiple antigen during melioidosis and that this immune response will be similar to melioidosis as humans. In this present study, we analyzed antibody reactive proteins and determined quantitative humoral antibody response to melioidosis as measured by western blotting and whole cell lysate ELISA. Antibody generation by the humoral response is a key component of understanding the immune response and a foundation of the potential biomarkers and vaccine development. Understanding the progression of immune response through an animal model could give insight into how the host reacts to B. pseudomallei infection and what antigens contribute to immune reactivity. The research also can provide a very detailed information on the antigens that induce an immunological response during infection.
Whole cell lysate (WCL) generated from the B. pseudomallei infection strain (MSHR511) was used as the antigenic material to characterize the humoral response in goat sera for 1, 4, and 5 days before infection (pre-challenge) and for days 7, 14 and 21 after infection (post-challenge). The strain of B. pseudomallei MSHR511 was isolated from a goat that had melioidosis. There was an outbreak of melioidosis in goats on a farm outside Darwin, Northern Territory, Australian [33, 34].
This study was performed in strict accordance with the recommendations in the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health, and every effort was made to minimize suffering. The protocol was approved by the Animal Care and Use Committee of Colorado State University (approval 11-2414A). The handling and use of Select Agents was approved by Colorado State University, Northern Arizona University, and monitored for dual use research of concern potential. The strain of B. pseudomallei MSHR511 was used for the challenge. The bacterium was grown in Muller-Hinton (MH) broth as described in Soffler et al. [35]. The bacterium was harvested in mid-log phase and diluted to 1 X 104 CFU/ml as final concentration. The bacteria suspension was then delivered as an intratracheal aerosol [35]. Healthy Nubian goats without clinical evidence of disease were obtained in Colorado, USA through a private sale and were acclimatized for 1 week before being infected with B. pseudomallei under anesthesia as described previously (Table 1) [35]. The goats were prescreened and were healthy, but prior exposure to bacteria via disease or just the microbiome is to be expected. The study documents a normal situation that will be encountered in animal models and human clinical studies. The goats were monitored by rectal temperature and complete blood counts pre- and post-infection [35]. Pre-infection sera were available from 8 goats (4 males and 4 females), which limited the number of goats used in the study. At the different time points of day 7, 14 and 21 post infection, 2–3 goats were euthanized and sera were collected with the exception of goat no. 16 planned for day 21 but became moribund. The serum from the goat euthanized on day 16 was included with the day 14 for calculations and analysis. The ethics approval for human sera data used at this study was obtained through the Human Research Ethics Committee of the Northern Territory Department of Health and Menzies School of Health Research [36]
B. pseudomallei strain MSHR511 was used in this study. The bacterial strain was grown on minimal media (BD) supplemented with casamino acids and glucose agar plates at 37°C for 35–48 hr. Bacterial gene expression pattern and phenotypes are affected by the growth medium [37]. The minimal media and rich medium were tested for the study. The minimal media with casamino acids and glucose produced more detectable antigenic proteins (data was not added in this manuscript). The minimal media is known to upregulate Type VI secretion system proteins, which are known virulence factors in B. pseudomallei and might generate an antibody response [38]. After incubation, single colonies of bacteria were scrapped and suspended in phosphate buffered saline (PBS) solution, pH 7.4 to give a turbidity reading of 1.0–1.2 at OD 600 nm.
WCL proteins were surveyed for immunogenic reactivity using sera from B. pseudomallei infected goats by 2-DE western blots. The bacterial cells suspended in PBS buffer, pH 7.4 were washed and centrifuged twice at 16,000 xg for 3 min at 4°C to pellet the cells. The cell pellets were resuspended in lysis buffer (50 mM KH2PO4, 400 mM NaCl, 100 mM KCl, 0.5% Triton X-100, and 10 mM imidazole), pH 7.4. The bacterial cells were lysed by a freeze and thaw technique using liquid nitrogen and 42°C heat block, respectively, repeated three times. Whole cell lysate proteins were separated by centrifugation at 18,000 xg for 15 min at 4°C. After separation, lysis buffer was exchanged with Tris-HCl buffer, pH 7.8 by centrifugation using microcentrifuge tubes. Protein concentration was determined using the Bradford technique [39] with bovine serum albumen (BSA) as a standard. WCL protein was enzyme treated to remove nucleic acid and precipitated using 15% trichloroacetic acid (TCA) in acetone and centrifuged at 18,000 xg for 18 min at 4°C to purify the proteins. Protein pellets after purification were dissolved in rehydration buffer containing 7 M urea, 2 M thiourea, 1.3% CHAPS, 30 mM DTT, 0.5% NP40, and 0.25% IPG ZOOM carrier ampholyte of pH 4–7.
For the CPS purification, broth in 2 L baffled Erlenmeyer flasks was inoculated with B. pseudomallei RR2683 and incubated overnight at 37°C with shaking (200 rpm). Cell pellets were obtained by centrifugation and extracted using a modified hot aqueous-phenol procedure [40]. Purified CPS antigens were then obtained essentially as previously described [41]. 1026B, the parent strain of B. pseudomallei is a type A LPS producing strain [42]. Bp82 is an excluded select agent since it is avirulent [43] and is an easier source to isolate OPS from.
For the OPS extraction and purification, Intron LPS extraction kit reagents were used (Intron biotechnology, South Korea). The cells were collected from agar plates inoculated with B. pseudomallei strain Bp82 and incubated 36 ~ 48 hr. at 37°C. Cells were collected from the plates and transferred to PBS buffer. Cell pellets were obtained by centrifugation 1300 x g. The cells were lysed using lysis buffer, chloroform, and purification as per manufactures instructions. The precipitated LPS was pelleted and suspended in water and sterility was confirmed by plating on 5% sheep blood in Tryptic soy agar (Hardy Diagnostics, Santa Maria, CA). The LPS was then further purified using proteinase K and 70% ethanol wash and drying. B. pseudomallei RR2683 is an OPS-deficient derivative (ΔrmlD) of the select agent excluded strain Bp82. By using this strain, we can safely and cost-effectively purify CPS from B. pseudomallei without the requirement for BSL-3 containment.
The antibody-reacted proteins from matched silver stained gels were identified by mass spectrometry. The strongly reactive immunogenic proteins were cloned, expressed and used in ELISA assays. 2-DE analyses were conducted using isoelectric focusing (IEF) of whole cell lysate proteins on immobilized pH gradient (IPG) gel strips (7 cm, pH 4–7 NL) for the first dimension according to Rabilloud [44] and Gorg et al. [45]. IPG strips were passively rehydrated with proteins dissolved in 165μl of rehydration buffer solution (Invitrogen, Carlsbad, CA) containing 100μg protein. IEF was focused on an electrophoresis apparatus (Xcell6, LifeTech, Carlsbad, CA) for a total of 8000V·hr. Focused proteins in the IPG strips were reduced and alkylated for the second dimensional electrophoresis by reduction and alkylation buffers for two consecutive 20 min incubations in 100mM Tris-HCl, pH 8.8 containing 5M urea, 800mM thiourea and 4% SDS, alternately, dithiothreitol (DTT) and iodoacetamide (IAA), each with a concentration of 130mM. The second dimension was separated on 4–20% Tris-Glycine gradient SDS-PAGE gels (Novex gels; Invitrogen, Carlsbad, CA). The gels were run at a constant voltage of 110V for 90 min and visualized using silver staining (Shevchenko et al., 1996). The images were captured using the UVP gel documentation system (UVP, Upland, CA). Image analysis was performed using Melanie (GeneBio, Geneva, Switzerland).
Proteins on 2-DE gels were transferred onto nitrocellulose membrane (Immun-Blot, 0.2 μm) using a dry transfer technique (iBlot, Dry Blotting System; Invitrogen, Carlsbad, CA) for 10 min at a constant current of 25 mA. The membrane was blocked for 1h in PBS containing 1.5% skim milk. The membranes were incubated with goat serum at a dilution of 1:1000 in PBS containing 1.5% skim milk for 1h, and then washed three times in PBS. Next, a secondary antibody of donkey anti-goat IgG or IgM conjugated with horseradish peroxidase (abcam, Cambridge, MA) was added for immunoglobulin G or M detection. Both of the secondary antibodies were diluted with 1/1000. Protein spots were visualized using a chromogenic substrate, 3, 3ʹ-diaminobenzidine (DAB) solution. The human sera used in this study were treated similarly to the goat sera analysis except for the use of a goat anti-human IgG secondary antibody (Promega, Madison, WI) with 1/1000 dilution and goat anti-human IgM (abcam, Cambridge, MA) with 1/1000 dilution.
Immunostained protein spots were matched with their corresponding protein spots on silver stained 2-DE gels using Melanie software (Melanie version 7.0.6 software; GeneBio, Geneva, Switzerland). The matched protein spots on silver stained 2-DE gels were excised and destained in 0.02% sodium thiosulfate and 0.5% potassium ferricyanide solution [46]. Gel pieces were washed, dried in 50% acetonitrile, reduced and alkylated in buffer (100 mM Tris-HCl, pH 8.8 with 5 M urea, 0.8 M thiourea and 4% SDS) containing 10 mM DTT followed by 100 mM IAA. Proteins were digested overnight in a digestion buffer (50 mM NH4HCO3, containing 1 mM CaCl2) and 12.5 ng/ml trypsin (Promega, Madison, WI) at 37°C. The enzyme treated peptides were extracted using 5% formic acid and 50% acetonitrile. The extraction of the digested peptides was facilitated by vortexing followed by sonication each for 30 minutes.
Mass spectrometry analysis was performed using a 4800 Plus Matrix Assisted Laser Desorption Ionization Time of Flight (MALDI-TOF) analyzer (AB Sciex, Toronto, Canada) with a 400Å Anchorchip TM target plate (AB Sciex, Toronto, Canada). Recrystallized α-hydroxycinnamic acid (1 mg/ml) in acetone was diluted 1:2 with ethanol and 1 μl was mixed with 0.5 μl of peptides and crystallized on the target. Spectra were analyzed and proteins were identified in a 4000 series Explorer V3.5.3 and Protein Pilot V4.0 software (AB Sciex, Toronto, Canada). Peptide mass fingerprints were searched against RAST annotated protein database. One missed cleavage per peptide was allowed, and the fragment ion mass tolerance window was set to 100ppm.
Using MS peptide sequence results, “immunogenic” proteins were identified with the help of several software programs available online. Protein identity was performed using BLAST (www.ncbi.nlm.nih.gov) against our laboratory created database and NCBI database was used for inferred bioinformatics information of the identified proteins. Theoretical molecular weight and pI values were taken from NCBI database and calculated using Compute pI/Mw tool on ExPASy website (http://web.expasy.org/compute_pi/).
Primers were designed based on the 5’ and 3’ ends of the gene sequences using online software Integrated DNA Technology. Oligonucleotides were generally between 18–24 bases (S1 Table) with a melting temperature of between ≥ 54–60°C. PCR was by Pfx DNA polymerase (Invitrogen) [47] to produce blunt-ended PCR products suitable for cloning in pcDNA 3.1 (Invitrogen). PCR reactions consisted of 0.3 mM of each dNTP, 1 mM of MgSO4, 3x PCR enhancer solution, 0.3 μM of forward and reverse primers and 0.05 ng/μl template DNA using 1.25U/well Pfx DNA polymerase in a final volume of 20 μl. The PCR amplicons were quantified and analyzed by nanodrop spectroscopy and/or gel electrophoresis and working concentrations of DNA made for use in the ligations reactions. The cloned gene inserts were confirmed by Sanger DNA sequencing [48] and transformed into Escherichia coli (strain BL21-DE3).
After overnight dialysis, the supernatant was separated from cell debris by centrifugation at 10°C at 10,000 rpm for 30 min. The recombinant proteins were collected as soluble protein. The soluble protein fraction was separated by fast protein liquid chromatography system (BioLogic, Bio-Rad Laboratories, Hercules, CA) using nickel affinity chromatography. The dialyzed protein solution was applied to a 20 ml nickel nitrilotriacetic acid (Ni-NTA) column and eluted with dialysis buffer containing 500 mM imidazole. The purification of His-tagged recombinant proteins was based on the His-tag protocol [52]. The purification of His-tagged recombinant proteins was monitored by SDS-PAGE and confirmed by western blotting using HisG monoclonal antibody, mouse HRP conjugate (Invitrogen). Protein concentration was estimated by Pierce BCA method with BSA as the standard (ThermoFisher Scientific, Grand Island, NY).
To understand the humoral response to B. pseudomallei in an experimental goat model for melioidosis, we developed the following ELISA assays to quantify the antibody concentration to the whole cell lysate and each individual protein and cell wall antigens. Five of these were protein antigens (dihydrolipoamide dehydrogenase of pyruvate dehydrogenase complex, PDHD; thiol peroxidase, TPX; alkyl hydroperoxide reductase subunit C-like protein, AhpC2; enolase, Eno; heat shock protein 60 family chaperone, GroEL1) and two polysaccharides (capsular polysaccharide, CPS and type A O-polysaccharide, OPS A) known to be immunogenic in B. pseudomallei (S2 Table) [36]. The polysaccharides were selected because B. pseudomallei produces both CPS and OPS A, which are implicated in B. pseudomallei virulence [6, 53, 54].
The 96-well immuno plates (Microfluor 2; Fisher Scientific, Pittsburgh, PA) were first coated with individual antigens, namely, recombinant proteins (250ng/well), OPS A (2000 ng/well), or CPS (125 ng/well) in PBS coating buffer (137 mM NaCl, 2.7 mM KCl, 10 mM Na2HPO4, 2 mM KH2PO4, pH 7.2) (Fisher Scientific, catalog no. BP3991) and left overnight at 4°C. After overnight coating, wells were washed 4 times with 200 μl PBS. The 96 well plates were flicked between the washes to remove all the PBS from the wells and for the fourth wash the plates were smacked onto a stack of paper towels. After washing, the wells were blocked with blocking buffer solution containing 1% (v/v) BSA in PBS for 2 hr at room temperature (RT). After 2 hr incubation the wells were washed 4 times with PBS containing 0.05% (v/v) Tween-20 (PBS/T) and three different goat sera dilutions (1/500, 1/1000, and 1/1500) in blocking solution added to the wells in a volume of 100 μl per well. Following the 2 hr incubation with the primary antibody (sera from B. pseudomallei infected goats), the wells were again washed 4 times with PBS/T to remove unbound goat serum antibody. A secondary antibody of donkey anti-goat IgG or IgM conjugated with horseradish peroxidase (Santa Cruz Biotechnology, Dallas, TX) was added for immunoglobulin G or M detection. Both of the secondary antibodies were diluted with 1/5000. Wells were washed 4 times with PBS/T and the enzyme reaction was detected by adding 100 μl substrate (Amplex Red reagent, Life Tech) for a defined time at RT. The fluorescence of the wells was read using a BioTek plate reader at 530/25 excitation and 590/35 emission wavelengths.
Data analysis was performed using Prism 6 statistics software (GraphPad software Inc, La Jolla, CA). All data in in quantitative histogram figures (Figs 1, 4 and 6 and S2 Fig) are presented as mean averages with error bars representing the standard deviation. The sample size was based upon ethical considerations from the previously reported study design [35] and capitalized upon existing samples. The spot counts in Figs 3, 4 and 5 is limited by the power of 2D-PAGE to resolve every protein and by duplicate protein locations created by posttranslational modification. Interpretation of these data needs to recognize that individual proteins can generate multiple spots and that individual spots may be composites of multiple proteins. Within these limitations, comparative analyses over time and among animals can still identify trends in the immunological response.
We initially set out to evaluate the general antibody response to B. pseudomallei during an aerosol infection of goats. To do this we determined the relative number of the immune reactive antigenic protein spots as an indicator of the diversity of the antigens the immune response generated to and the quantity of the reactive antibodies. This involved ELISA, 2D gel electrophoresis, and 2D western blot analyses using a whole cell lysate of the infecting strain (MSHR511). Both ELISA and western blot results indicated a much greater humoral response in both breadth and quantity of IgG antibodies compared to that of IgM antibodies throughout the course of infection (Fig 1). Goat IgG antibodies assayed using ELISA on WCL from B. pseudomallei MSHR511 showed a rapid increase, rising from an average concentration of 11,348 μg/ml on day 7 to 26,666 μg/ml on day 14 and reaching a maximum of 35,088 μg/ml by day 21. The concentration of IgM goat antibodies increased over time as well, from a mean of 4,026 μg/ml on day 7 to 8,565 μg/ml on day 14 and reaching 9,590 μg/ml by day 21. The total amount of goat humoral IgM antibody was 2-4-fold less than goat IgG antibodies for all three-time points after challenge with B. pseudomallei. In the western blots, the relative number of IgG-reactive protein spots detected after being probed with goat sera mirrored the increase observed using ELISAs for days 7, 14 and 21 (Fig 1). However, the number of antigenic WCL protein spots on IgM western blots remained relatively the same across all time points, with an average of 18 on Day 7, 19 on Day 14, and 12 antigens on day 21. The difference of the IgM immune response between ELISA and 2D western blot might be explained by the particular antigens used by the two methods to detect antibodies. ELISA used whole cell lysate, which included both proteins and carbohydrates, while the western blot used only the whole cell lysate proteins, although this would include glycoproteins. CPS and OPS are carbohydrate antigens known to trigger strong IgM reaction and would only be detected by the WCL ELISA. Overall, the B. pseudomallei aerosol challenge induced an increasing and diverse antibody response over the disease progress in the goats.
We characterized the IgG and IgM antibody responses of 8 individual goats’ serum samples using 2D electrophoresis and western blots. Our 2DE reference maps prepared from MSHR511 WCL allowed us to detect nearly 600 bacterial protein spots by silver stain and to match those spots with antigenic spots detected by IgG and IgM western blots (S1 Fig), thus providing a large population of proteins to investigate. High numbers of immunogenic protein spots were detected (marked in blue), denoting a highly diverse humoral response to bacterial proteins. Immunogenic protein spots detected on the western blots showing strong immunoreactivity against IgG increased in intensity over the 21 days of infection. In contrast, immunogenic protein spots for IgM showed low intensity and remained that way for the entire infection period (Fig 2). Even 2D western blots probed with pre-infection sera showed multiple immunogenic protein spots for IgG (0–20 spots) and IgM (0–12) in most goats (Figs 3 and 4). This suggests that the humoral antibody response was previously primed to recognize these or similar antigens from prior bacterial exposure. After infecting the goats with aerosolized B. pseudomallei, the number of immunogenic protein spots producing goat IgG showed a marked increase over day 7, 14, and 21. The average number of IgG spots shows an increase from 29 protein spots at day 7 to 74 on day 14 and 129 protein spots on day 21 (Fig 3).
The number of IgM immunogenic protein spots was much lower compared to IgG. At the pre-infection time point, a mean of about 6 IgM antigenic protein spots were detected but all were very faint (Fig 2). The number of IgM immunogenic protein spots increased 3-fold by day 7 (mean of 19 spots), but then remained stable for most goats with a mean of ~25 IgM immunogenic protein spots on day 14 and 21 immunogenic protein spots on day 21 (Fig 4). This result was in agreement with that total IgM response to MSHR511 WCL (above) grew weaker over the 21 day time interval.
Over the time course of this challenge study, we found that a subset of proteins raised a consistently strong immunogenic response at each time point for all individual goats, especially for the IgM response: heat shock 60 family protein GroEL, elongation factor Tu, ATP synthase beta chain, and heat shock protein DnaK. Many antigenic proteins were observed to remain immunogenic after the initial antibody response was seen. However, there were also many antigenic proteins observed at only a single time point per goat. Considering all of the antigens detected for IgG and IgM, there were 44 antigenic proteins found at all-time points post challenge, 14 of which were also observed in pre-infection sera (Fig 3). There were 6 protein spots that showed immunoreactivity to both of IgG and IgM. Of all of the antigens for IgG, 38 protein spots were detected at all of the post-infection time points, making proteins producing IgG antibody as the most common (38 out of 44 proteins). Eight of these antigens were found in the pre-infection sera (Fig 3 and S4 Table). On the other hand, IgM had the least number of antigenic proteins spots present at all-time points with 19 out of 44 proteins and only 5 antigenic proteins immunoreactive in the pre-infection sera (Fig 3). The total WCL reactive IgM was stable with the majority of antigens (19) present throughout Day 7–14.
To place our results in a broader context, we compared the goat immune response in this study with antigens previously described in human melioidosis patients [36]. B. pseudomallei infection leads to strong immune responses in both goats and humans, and many antigens elicited a comparable antibody response. According to these data, there are 98 antigenic proteins detected in both of human and goat immune responses, which is more than half the number of antigenic proteins in the human immune response. Considering the total amount of antigens, 282 antibody reactive antigens detected with goat IgG and IgM, the common 98 antigens out of 282 is still higher number in comparison with 135 total detected human antigenic proteins (Fig 5). The difference in the number of the detected antigenic proteins might be caused by the various biological facets including anatomy, pathophysiology, and genetics, resulting in the different disease progress due to dissimilar route, dose, and immune state prior to infection in/between human infection and animal model. The strain differences might also contribute the antigenic proteins difference. Sela et al. (2018) research demonstrated that the variation of strains in a species of Staphylococcus produced different T helper response, resulting in wide variability in the acute adaptive immune response [55]. So, the highly common antigenic proteins detected in both human and animal melioidosis imply that this goat model study could be an appropriate animal model to understand disease progress and humoral immune response with mimicking human disease condition.
Antibodies to a subset of identified B. pseudomallei antigens were found in pre-infection goat sera. We compared these 17 antigens from MSHR511 against a subset of common Gram-negative bacteria that could possibly be the source of a pre-challenge infection, including Pseudomonas aeruginosa, Escherichia coli, Campylobacter jejuni, and Mycobacterium tuberculosis (Table 2). None of these 17 B. pseudomallei MSHR511 proteins had an amino acid identity above 83% when compared to their homologues in the selected species of Gram-negative bacteria. Amino acid identity was highest between MSHR511 and P. aeruginosa, with 13 antigens in the range of 55–83% and four with much lower similarity (33–40%). Eight proteins from all of the four bacteria evaluated had a high identity of > 50% compared to B. pseudomallei: ATP-dependent chaperone protein (ClpB), elongation factor G2 (EFG2), chaperone protein (DnaK), heat shock protein 60 family chaperone (GroEL1), ATP synthase alpha chain (Atp), ATP synthase beta chain (AtpD1), Translation elongation factor Tu (EF-Tu), and enolase (Eno) (Table 2). The remaining 9 proteins out of 17 were the least similar to B. pseudomallei proteins across the four Gram-negative bacteria, with one exception, viz. C. jejuni, S-adenosylmethionine synthetase having < 50% amino acid sequence identity to the B. pseudomallei protein. C. jejuni had the lowest number of proteins with similarity to B. pseudomallei, with only eight proteins out of 17 having > 50% amino acid sequence identity to B. pseudomallei proteins. Whether the < 50% amino acid identity of B. pseudomallei to C. jejuni means reduced immune reaction to the C. jejuni protein antigens and its protein epitopes is unknown. Interestingly, amino acid similarity for GroEL was moderately high across all species (58–73%) and may provide insight into the source of the pre-challenge antibody responses we observed.
We compared antigenic spots detected using pre-infection sera for antibody immune reactivity over time. We found that antigenic protein spots detected before the aerosol challenge were also detected at time points after the B. pseudomallei aerosol challenge. The chaperone and cell division related proteins in particular were detected pre- and post-challenge with high signal intensity (S4 Table).
We chose to more fully investigate the antibody response to six proteins that induced an antibody response prior to B. pseudomallei infection. Using 2D western analysis and spot area (a relative measure of immune reactive intensity), each antibody response appeared to peak at a specific time point after the aerosol infection. Goat sera probed against the IgG western blots showed that the antigenic spot areas for each protein antigen (GroEL, EF-Tu, AtpD1, DnaK, AtoC, and Eno) generally increased through day 21 (S2 Fig). This paralleled the pattern observed for the total IgG response to WCL (above). In addition, the antibody response to these antigens occurred by seven days after infection. In contrast, IgM antibody responses for these six antigens typically showed an early peak in spot area on day 7, which then decreased slowly throughout the remainder of the challenge study. The day 7 peak was particularly strong for GroEL and EF-Tu, and both of GroEL and EF-Tu, and AtoC induced a magnitude of response on par with the IgG antibody at that time point. Interestingly, none of the IgM response for these six antigens matched the overall pattern observed in the total IgM count (i.e., none increased from day 7 to day 14). This pattern indicates that the IgM response to other antigenic proteins must generally grow stronger over infection and may represent new antibody responses.
We used ELISA assays to more thoroughly investigate seven antigens (PDHD, TPX, AhpC2, Eno, GroEL, CPS and OPS A) that were immunogenic in this goat study and a previous human melioidosis study including 4 patients (Fig 6). The overall IgG reactivity against these antigens increased over the time. The amount of IgG antibody detected on day 14 for 4 antigens (PDHD, AhpC2, Eno and CPS) ranged from 110 to 133 μg/ml, with the remainder antigens having < 48.9 μg/ml (Fig 6). The exception was GroEL. IgG response specific to GroEL were at a concentration of 330 μg/ml on day 7 and reached up to 5579 μg/ml by day 21 (Fig 6). The concentration of IgM on the other hand, increased mostly from day 14 for CPS with 28 μg/ml to 65 μg/ml for GroEL antigen. The 3 antigens showing high levels of antigen-specific IgG antibodies compared to IgM by day 21 were AhpC2, Eno and GroEL. In contrast, the antigens; PDHD, TPX, CPS and OPS A on day 21 had elicited fairly high concentrations of IgM-specific antibodies compared to that of IgG antibodies (Fig 6). The percentage of individual antigens for the goat antibody immune response was calculated relative to the total immune response of the 7 selected antigens (S3 Fig). IgM antibody responses constituted the highest percent of the immune response with the exception of GroEL antibody immune response, which had very similar percentages for both IgG and IgM antibodies for days 7 and 14 and a significantly high level of IgG goat antibody response for day 21 (S3 Fig). Because GroEL showed the highest immune response for both IgG and IgM antibodies, we compared the individual antigens antibodies immune response to GroEL (S3 Fig). The heat shock protein GroEL elicited the strongest goat antibody immune response compared to the other six antigens (PDHD, TPX, AhpC2, Eno, CPS and OPS A) measured in this study, specifically for IgG. This response may be due to prior exposure of the goats to GroEL protein found in other Gram-negative bacteria which may possess conserved amino acid sequences similar to B. pseudomallei GroEL. For goat IgM antibody immune response, three antigens, CPS, GroEL and OPS A showed the strongest immune responses for day 7. The other antigens (PDHD, TPX, AhpC2 and Eno) produced similar goat IgM antibody responses for day 7. However, for day 14 the proportion of goat IgM antibody immune response was very similar for all the 7 antigens (S3 Fig). But by day 21, CPS showed the highest goat IgM antibody response compared to the other antigens; GroEL1, PDHD, TPX, AhpC2, Eno and OPS A. There was not much change in IgM antibody response for both GroEL and OPS A from day 14 to day 21; while the antibodies to the other antigens remained at similar levels for day 14 and 21.
The overall humoral immune response of goats showed an increase of antibody intensity and immune diversity during the infection progression, especially for IgG. In contrast, IgM reaction showed only a slight increase in antibody intensity and the immune diversity decreased after 14 days post challenge. The overall immune response intensity increase of IgG from the earliest stages of infection implies that there are pre-existing immune conditions for many of the antigens including individual proteins tested in this research. The IgG immune response was greater than the intensity for IgM reactivity. IgM is the primary adaptive immune response to infection while IgG usually develops in the later stage of the infection course after class switching from IgM. However, our investigation showed that the overall IgG humoral response was stronger than IgM at day 7 (initial assay date after goat aerosol infection). This stronger intensity of the IgG response may be explained by protein epitope ability to induce a strong IgG antibody response to protein epitopes of pre-existing immunity such as memory B-cells, which could be produced by cross-reactive epitopes before the bacterium strain challenge. The pre-existing immune memory B cells could produce rapid antibody right after the challenge (or natural infection) [56].
The immune diversity results as determined by western blot analysis showed a slightly different pattern from the overall humoral response determined by ELISA. The count of IgM immunoreactive proteins showed a similarity to IgG reactive protein spots on day 7 and decreased with the decreased intensity of IgM reactive protein spots over the infection progress (Fig 2 and S2 Fig). At day 14 onward, we detected more IgG reactive proteins than IgM reactive ones (Figs 1 and 2). The IgG antibody demonstrated more unique immunoblot protein spots from day 14 to 21, whereas, IgM reacted with more protein spots on day 7 compared to IgG (Fig 3).
The antibody isotype switch happens between day 7 and day 14 when an antigen(s) persists [57, 58]. Even though there are more detected IgM reactive protein spots on day 7, the intensity of humoral immune response was stronger to IgG. The strong IgG response at the early infection stage like day 7 may arise from the memory cells and prior existing health conditions, which produce faster and stronger humoral immune response with the current infection. The immunogenic proteins sequence comparisons of the selected bacteria showed low similarity to the selected immunogenic proteins of the infecting strains, which could produce memory cells for the fast and strong immune response. However, there is the potential that conserved sequence or structure epitopes of the selected immunogenic proteins between the selected bacteria and the infecting strain or other antigens cause fast and strong IgG response among the unexamined antigens even though there was no noticeable clinical presentation before the challenge. The T helper cells contribute to the antibody class switch and affinity maturation for stronger immunoreactivity. Other researchers have reported similar results of elevated IgG and IgM post infection with B. mallei, a closely related species to B. pseudomallei [59].
We found 44 antigenic proteins commonly detected in all of the individual goats, and 8 immunogenic proteins detected in pre-infection sera for IgG and IgM using western blot. Those immunogenic proteins are a small portion out of the total 282 immunogenic proteins identified in this research (S4 Table). The commonly detected immunogenic proteins were decidedly fewer than the specific immunogenic proteins only detected in an individual or several goats. These results demonstrate how variable individual immune responses are to even infection of the same infecting strain with the same conditions. The individual animals may vary in their prior exposures to other bacterial infections. In addition, these are outbred animals with different genetics, perhaps creating different immune responses because of the polymorphic differences of MHC alleles. However, commonly detected immunogenic proteins with the different timelines for antibody characterization generate the antibody response to the same or similar epitopes despite the host variability to infection. These common immunogenic proteins are mainly immunodominant proteins (S4 Table) [60]. These common immunogenic proteins give the insight of how the host generally reacts to the infection and it also demonstrates those proteins’ potential as general diagnostic antigens for bacterial infection. Through the observation of common antigenic proteins detected between goats and human sera, we prove that the goat model is an appropriate model to understand the human infection progress with the antigens producing humoral immune reaction. Therefore, the detected antigenic proteins from this study could be used for the further analysis of melioidosis in aspect of detectable antibodies and their immunogenic reaction.
By investigating and identifying highly immunogenic proteins, we found that five proteins (PDHD, TPX, GroEL, AhpC2, and Eno) and two polysaccharides (CPS, OPS A) that showed high immunoreactivity to melioidosis patient sera against four B. pseudomallei strains in our previous study [36]. Those highly immunogenic proteins were also commonly detected proteins using sera from the eight different goat individuals. The western blot and ELISA results for the selected antigens showed a strong ELISA signal for IgG and IgM antibodies from day 14 onwards. Specifically, as for the proteins, GroEL, AhpC2, and Eno induced a stronger IgG response, while the IgM response was strong for all five antigenic proteins (Fig 6). This is probably due to antigenic protein epitopes inducing a stronger IgG antibody response with class switching after an initial immune response [61].
Polysaccharides, namely CPS and OPS A were assessed in this study. CPS demonstrated a strong ELISA signal for IgG and IgM antibodies from day 14 onwards. While OPS A antigen demonstrated a good IgM antibody response from day 7. CPS and OPS A are thymus-independent antigens known to activate B-cells to elicit low-affinity IgM antibodies [56]. Therefore the presence of IgM up to 21 days in goat sera can likely be explained by the persistence of polysaccharide antigens for an extended period in the lymphoid tissues as observed during Yersinia pestis infection, continually stimulating newly maturing B-cells to produce IgM antibodies [57]. Other researchers [28, 32, 62] also studied CPS and OPS as biomarkers for the melioidosis diagnostics using the high seropositivity of two antigens. The high seroreactivity is in agreement with this study. The researches demonstrated high seropositivity against melioidosis patient sera, resulting in high diagnostics sensitivity and specificity for melioidosis. Therefore, the CPS and OPS A antigens are potential of biomarkers to diagnose melioidosis.
As is evident from the results, there were differences in the levels of goat antibody titers expressed against the seven highly immunoractive antigens, which may suggest possible differences in the amounts of each protein and polysaccharide antigen produced by B. pseudomallei or differences in antigen immunogenicity. Vasu et al. [61] has shown that the primary antibody response to B. pseudomallei protein and polysaccharide antigens in melioidosis patients was IgG, subclasses IgG1 and IgG2 antibodies, suggesting a Th1 antibody response in both septicemic and non-septicemic melioidosis cases.
As the most immunoreactive antigen, GroEL is one of the immunodominant antigenic proteins, which gave a strong immune intensity even at the day 7 (Fig 6). This result could be explained by the presence of memory B cells from a prior non-melioidosis infection with a related organism and may cause a strong immune response rapidly right after the host encountered the antigenic proteins. Amemiya et al. [63] using ELISA assays reported a 10-fold increase of IgG antibodies against the heat shock protein (hsp), GroEL. This is in agreement with our results, where GroEL specific goat antibody titer was ~12.8-fold higher than it was to AhpC2 protein; the second highest goat IgG antibody titer to GroEL (Fig 6).
Because of this strong IgG and IgM antibody immune response to GroEL, it was decided to examine this result further. First, GroEL family of proteins is reported to contain epitopes that are highly conserved from prokaryotes to humans [64]. GroEL-like proteins are reported to be immunodominant antigens from infectious pathogens, such as Mycobacterium leprae, M. tuberculosis, Coxiella burnettii and Legionella pneumophila [64]. During infection, a pathogen undergoes selective pressure which increase microbial heat shock protein synthesis to withstand the harsh environment inside the host [65]. This reason is why hsps are major antigens in infectious agents that induce a strong innate and cellular immune response [66].
Most of the pre-infection sera showed some IgM and IgG immune reactivity to B. pseudomallei proteins. Most of the antigenic proteins identified using pre-infection sera were immunodominant proteins [36], and were mostly faint spots of low intensity. Strongly reactive proteins to IgG from the pre-infection sera set were detected, which were mostly immunodominant proteins and also detected at post-infection sera (S4 Table). The goat sera showing pre-primed immune reactivity showed fewer detected antigenic proteins, thus exhibiting less immune diversity (Fig 4). The results imply that the pre-primed response and immunodominant antigens impede developing a broader antibody response against multiple epitopes of the strain. Those antigens may be cross-reactive with proteins from prior bacteria that the goat came in contact with.
As observed with other diseases, potential serological cross-reactivity to proteins of closely related pathogens can produce the immunological memory of the B-cells and cause cross-reactivity in immunoassay, presenting a diagnostic challenge. This cross-reactivity was investigated by comparing the amino acid sequences of seventeen antigenic proteins with the same proteins found in a selected number of Gram-negative bacteria, viz. P. aeruginosa, E. coli, C. jejuni and M. tuberculosis (Table 2). The percent identity of B. pseudomallei proteins to the above four Gram-negative bacteria ranged from 5% for PDHD and ETFα of C. jejuni, FtsA and FliC of M. tuberculosis to 83% for EF-Tu of P. aeruginosa. The genus, Pseudomonas was where B. pseudomallei was classified before Burkholderia was proposed as a new genus [67, 68]. The overall sequence comparison results showed the low similarity of the selected proteins of the chosen bacteria even though the investigated proteins are commonly detected proteins among the bacteria. The low sequence similarity might indicate that the studied proteins do not contribute to the early stage of immediate antibody response with high humoral response. However, there is potential that there are still cross-reactive epitopes or unstudied antigens causing cross reactive humoral response at the early challenge stage. The faint IgG and IgM immunoreactive protein spots detected in this study support the postulate. Overall, the selected proteins even showed cross reactivity in this research but the intensity and the sequence similarity were very low. Thus, the selected proteins could be the potential biomarkers of B. pseudomallei infection (Table 2). The investigated antigens from this study and the melioidosis patient cases [36] showed the detection biomarker potentiality for this infection in the aspect of their detection frequencies in the progress of the infection and commonness in the detection of different individuals.
In conclusion, this study characterized the overall antibody response of IgG and IgM antibody response, delineating the diversity of immunogenic proteins generated within the host after B. pseudomallei aerosol challenge. There was a detectable immune response from the early stage of the infection and there are antigens eliciting strong signal intensity for either or both of IgG and IgM. Of the antigens detected, there were 44 commonly detected antigens among the eight individual goats (S4 Table). The variation of the immune response was demonstrated against many of the detected antigens among the goats and during the infection progression. This study also involved expression and purification of five recombinant proteins and 2 polysaccharides detected to be immunogenic using sera from B. pseudomallei infected goats and also from human patient sera. Of the seven antigens assayed, AhpC2, Eno and GroEL had a stronger IgG response, while CPS, OPS and TPX showed the stronger the IgM response. Even though we tested the seven potential antigens for their immunoreactivity and the potential as diagnostics biomarkers, we detected additional antigenic proteins with potential as diagnostics targets showing high detection frequency among the goats. Further ELISA assay evaluation of these antigens is needed to determine if they are an improvement over the IHA assay sera and in clinical settings.
|
10.1371/journal.pgen.1007719 | MutLγ promotes repeat expansion in a Fragile X mouse model while EXO1 is protective | The Fragile X-related disorders (FXDs) are Repeat Expansion Diseases resulting from an expansion of a CGG-repeat tract at the 5’ end of the FMR1 gene. The mechanism responsible for this unusual mutation is not fully understood. We have previously shown that mismatch repair (MMR) complexes, MSH2/MSH3 (MutSβ) and MSH2/MSH6 (MutSα), together with Polβ, a DNA polymerase important for base excision repair (BER), are important for expansions in a mouse model of these disorders. Here we show that MLH1/MLH3 (MutLγ), a protein complex that can act downstream of MutSβ in MMR, is also required for all germ line and somatic expansions. However, exonuclease I (EXO1), which acts downstream of MutL proteins in MMR, is not required. In fact, a null mutation in Exo1 results in more extensive germ line and somatic expansions than is seen in Exo1+/+ animals. Furthermore, mice homozygous for a point mutation (D173A) in Exo1 that eliminates its nuclease activity but retains its native conformation, shows a level of expansion that is intermediate between Exo1+/+ and Exo1-/- animals. Thus, our data suggests that expansion of the FX repeat in this mouse model occurs via a MutLγ-dependent, EXO1-independent pathway, with EXO1 protecting against expansion both in a nuclease-dependent and a nuclease-independent manner. Our data thus have implications for the expansion mechanism and add to our understanding of the genetic factors that may be modifiers of expansion risk in humans.
| The Fragile X-related disorders arise from expansion of a tandem repeat or microsatellite consisting of CGG-repeat units. The expansion mutation is not well understood, but our previous data suggests that MutSα and MutSβ, mismatch repair (MMR) proteins that normally protect the genome against microsatellite instability, are actually responsible for these mutations in a knockin mouse model of these disorders. In this manuscript we describe the role in expansion of two proteins that act downstream of the MutS proteins in MMR, MutLγ and EXO1. Our data suggests that expansion occurs via a MutLγ-dependent, EXO1-independent pathway, with EXO1 playing both a nuclease-dependent and a nuclease-independent role in preventing expansions.
| The 5’ end of the human FMR1 gene (MIM* 309550) contains an unstable CGG/CCG-repeat tract. This instability shows a strong expansion bias, with alleles having 55–200 repeats, known as Premutation (PM) alleles, being as much as 10 times more likely to expand than contract [1]. The likelihood of expansion increases with increasing repeat number [1]. PM alleles confer risk of a neurodegenerative condition known as Fragile X associated tremor/ataxia syndrome (FXTAS; MIM# 300623) and a form of female infertility known as Fragile X-associated primary ovarian insufficiency (FXPOI; MIM# 300624) [2]. Expansion is seen in somatic cells and in the germline, where it can produce alleles with >200 repeats. Such full mutation (FM) alleles result in Fragile X syndrome (FXS; MIM# 300624), a disorder whose major symptoms include intellectual disability (ID) and autistic behaviors [2]. Collectively these three clinical consequences of CGG/CCG-repeat expansion in the FMR1 gene constitute the Fragile X-spectrum disorders or Fragile X-related disorders (FXDs). These disorders belong to a larger group of genetic disorders known as the Repeat Expansion Diseases, that all result from an expansion of a tandem repeat in a disease-specific gene. However, whether these diseases share a common expansion mechanism is unclear.
Most models for repeat expansion invoke hairpin loop-outs formed by the individual strands of the expansion-prone repeat tract as intermediates in the expansion process [3–7]. In principle, these loop-outs could form during any time the DNA was unpaired, including during replication, repair synthesis or transcription. Loop-out formation on one strand could lead to the formation of loop-outs on the complementary strand since perfect realignment of the two strands would be blocked. These “double loop-outs” may resemble the Holliday Junctions (HJ) formed during meiosis. We have shown that the FX-repeat loop-outs are bound by MSH2/MSH6 (MutSα) and MSH2/MSH3 (MutSβ), the 2 complexes involved in lesion recognition in mismatch repair (MMR) in mammals [8]. We have also shown that MutSβ is required for almost all expansions in a knock-in FXD mouse model, with MutSα contributing significantly to the MutSβ-dependent expansions [8–12]. Since MutSβ is less much abundant than MutSα in our mouse strain background [8], it suggests that some unique property of MutSβ is important for expansion.
It has been suggested that simple incorporation of the loop-outs could result in expansions [3]. This could occur via a second DNA synthesis step that uses the loop-out as a template. This could involve a non-canonical MMR pathway [13, 14] and some expansion models invoke MutSβ-dependent nick-directed excision of one or both strands as the first step in this process [5, 15–17]. MLH1/PMS2 (MutLα) or MLH1/MLH3 (MutLγ) normally act downstream of the MutS proteins to coordinate excision in MMR. MutSβ binding to repeat-containing loop-outs can trigger MutLα cleavage that can occur on either strand [14]. Such cleavage could provide the nick(s) necessary for excision to take place [13, 14, 18]. Since EXO1 normally acts downstream of the MutL proteins in mismatch excision and is the only exonuclease thus far implicated in MMR, EXO1 may be the protein responsible. However, it is also possible that, instead of such a loop incorporation step, the loop-outs are channeled to a different repair pathway that ultimately leads to expansions.
To address events occurring downstream of MutSβ in the expansion process we decided to test the effect of a null mutation in Mlh3 in our mouse model since MutLγ interacts preferentially with MutSβ [19, 20]. MutLγ is also known to be required for expansions in a mouse model of Huntington Disease, a CAG-repeat expansion disorder [21]. In contrast, the more abundant MutLα complex either plays a smaller role in expansion in other model systems [16], or is protective [22, 23]. We also tested the effect of an Exo1 null mutation [24] and a point mutation in the EXO1 nuclease catalytic site. The point mutation prevents EXO1 acting in MMR but does not affect its ability to act in a structural capacity in meiosis where it is required for the proper orientation of cleavage of Holliday Junctions (HJs), a step that involves MutLγ, but not MutLα.
We show here that MutLγ is required for all germ line and somatic expansions in the FXD mouse. However, rather than promote expansion, we found that EXO1 protects against it. It does so in two distinct ways, one that is dependent on its nuclease activity and one that is not. This has interesting implications for the expansion mechanism.
In order to assess the role of Mlh3 in somatic expansions we compared the repeat PCR profiles in different organs of 6 months-old Mlh3+/+, Mlh3+/- and Mlh3-/- male mice that had inherited alleles with 150–160 repeats and determined the average repeat number added to the expanded alleles as an indicator of the extent of expansion. With the exception of heart, an organ that shows no postnatal expansion, expansion was less extensive in the organs of a Mlh3+/- male than in the organs of the Mlh3+/+ male, while in the Mlh3-/- male, no evidence of expansion was seen in any of the tissues tested (Fig 1A). When the repeat number added to the expanded allele in each organ from multiple animals was averaged, the effect of the loss of one or both Mlh3 alleles was found to be highly significant for all organs (Fig 1B; p < 0.0001). The number of repeats added was significantly lower in Mlh3+/- males than in Mlh3+/+ males in all expansion-prone organs and in Mlh3-/- males the average number of repeats added was <0.5 repeat for all organs, a result that falls within the margin of error of the assay. Since females show much less extensive expansions than males [25], we examined the effect of the loss of Mlh3 in females at 12 months of age. Even at this age expansions in some organs are too small for differences between Mlh3+/+ and Mlh3-/- mice to reach statistical significance. However, while expansions are clearly seen in the ovary, brain, liver and tail of Mlh3+/+ females, no expansions were seen in Mlh3-/- females in any organ (Fig 1C) and the difference between the extent of expansion in the brains and livers of these animals was large enough to be statistically significant. Thus, we conclude that Mlh3 is required for all somatic expansions in both males and females.
This effect is not mediated via an effect on the levels of MutLα since, with the possible exception of testis, the loss of MLH3 does not affect the levels of either MLH1 or PMS2, the constituents of the MutLα complex (S1 Fig). In the case of testis, PMS2 levels were elevated, while MLH1 levels were unaltered. This would be consistent with the idea that when MLH3 levels decrease, more MLH1 is available to form a heterodimer with PMS2. This would reduce PMS2 degradation, analogous to what is seen with constituents of the MutS complexes [26–28]. This effect may be limited to testis since MLH3 is normally present at ~60-fold lower levels than MutLα in somatic cells [29].
Since Mlh3-/- mice are sterile because of a defect in crossing over in meiosis [30], we could not monitor the incidence of germ line transmission of expanded alleles in these animals. However, the testis of these animals shows no evidence of expansion (Fig 1A and 1B). We have previously demonstrated that expansions are limited to premeiotic stages of gametogenesis (S2 Fig) [31]. In WT animals a bimodal distribution of repeat sizes is seen in the testis with the smaller peak corresponding to unexpanded alleles and the larger peak to the expanded alleles [31]. A single peak of the same size as the expanded allele seen in testis can be seen in the primary spermatocytes and the size of this allele does not change in more mature gametes [31]. While Mlh3-/- male mice do not have secondary spermatocytes or mature sperm, they do have primary spermatocytes [30]. Thus, any expansions in Mlh3-/- mice should be apparent as a second peak in the testis repeat PCR profile. Since the testis PCR profile lacks a second peak and is, in fact, indistinguishable from the heart PCR profile, we conclude that in addition to being required for somatic expansion, Mlh3 is also required for germ line expansion in males. Mlh3+/- males also show a significantly smaller number of repeats added to the PM allele in the testis, consistent with fewer repeats having been added to their gametes.
Small pool PCR from 3-month-old Mlh3+/- males shows clearly that they have fewer expanded alleles in their gametes than Mlh3+/+ males of the same age (Fig 2A). Thus, our data demonstrate that MLH3 is required for all germ line and somatic expansion in male FXD mice and that even heterozygosity for the null allele causes a significant decline in the extent of both germ line and somatic expansions.
The progeny of Mlh3+/- mothers showed a small reduction in the proportion of expanded alleles compared to the progeny of age-matched Mlh3+/+ mothers (Fig 2B). Furthermore, the average size of the expansions in the progeny of Mlh3+/- mothers was also significantly smaller than those in the progeny of Mlh3+/+ mothers (Fig 2C). We have previously shown that expansions continue to accumulate on previously expanded alleles as animals age [12, 31, 32]. Thus, the smaller size of the transmitted alleles from Mlh3+/- mothers would be consistent with them having undergone fewer rounds of expansion. Thus, as with males, even the loss of a single functional Mlh3 allele is enough to reduce the extent of germ line expansion in females.
To evaluate a role of EXO1 in expansion we tested the effect of both an Exo1 null mutation and a D173A mutation. The D173A mutation is located in the active site of EXO1 and thus abolishes its hydrolytic activity. However, X-ray crystallography and in vitro biochemistry indicates that the protein retains its native conformation and DNA binding affinity [33–35] and it is expressed at similar levels as WT EXO1 in our mouse strain background (Chahwan et. al., manuscript in preparation). Exo1 null mice, like Mlh3-/- mice, are defective in MMR and are sterile because they are unable to complete crossing-over during meiosis and thus make no mature gametes [24]. Mice homozygous for the mutant allele (Exo1A/A) are MMR defective, but fertile consistent with the idea that Exo1 plays a structural role in meiosis [36].
When we examined the number of repeats added to the PM allele in testes, tail, brain, liver, and kidney, organs we have examined in previous studies [8, 10, 11, 37], large differences between Exo1+/+ and Exo1-/- male mice was only seen in the testis (Fig 3A and 3B). A failure to see large differences in the tail, kidney, liver and brain of male mice is consistent with the fact that EXO1 not highly expressed outside of the testis [38, 39]. However, it is known that the small intestine shows an increased mutation rate in the absence of EXO1 [40]. We therefore decided to also test this organ for expansions. Alleles in the small intestine of Exo1-/- mice gained roughly twice as many repeats as the Exo1+/+ mice (Fig 3A and S3 Fig). The Exo1A/A mice also showed the gain of significantly more repeats than Exo1+/+ mice but, as in testis, the number of repeats gained was fewer than in the Exo1-/- mice. The differential effect of the null and D173A mutation in small intestine suggests that EXO1 plays both a nuclease-independent and a nuclease-dependent role in reducing somatic expansions in this tissue. The failure to see large changes in other somatic tissue of males may reflect the relatively low level of expression of EXO1 in these tissues. In WT females, the somatic expansion frequency is much lower than it is in males [25, 37]. This makes it difficult to accurately determine the mean expansion size or the somatic instability index. Expansions in different females are also much more variable, due in part to the fact that expansion only occurs when the repeat is on the active X chromosome [25]. Since X chromosome inactivation is a stochastic process, female mice show a wide variation in the fraction of expanded alleles that are on the active X [25]. However, while direct comparisons are difficult, expansions, in general, do seem to be more extensive in Exo1-/- and Exo1A/A females than in Exo1+/+ females (S4 Fig, panel B).
The testes of Exo1-/- mice lack mature gametes that make up ~95% of the testicular cells. They thus produce a repeat PCR profile that differs from what is seen in Exo1+/+ mice and Exo1A/A mice (Fig 3B). The presence of a peak corresponding to the original allele size reflects the fact that the somatic cells of the testes [41], where the repeat does not expand, constitutes a greater fraction of the testicular cells in Exo1-/- mice than they do in Exo1+/+ and Exo1A/A mice. Nevertheless, it is clear that the residual gametes showed an average gain of 24 repeats in Exo1-/- mice compared to 14 repeats in Exo1+/+ mice and thus that expansion in the gametes was more extensive in Exo1-/- mice. This was consistent with the gain of repeats seen in purified primary spermatocytes purified from Exo1-/- mice (S2 Fig). As in the small intestine, Exo1A/A mice had gained considerably more repeats in the testis than the Exo1+/+ mice, but significantly fewer repeats than the Exo1-/- mice. This data suggests that, as with somatic expansions, EXO1 protects against germ line expansions in both a nuclease-dependent and nuclease-independent manner.
We verified the effect of losing EXO1 exonuclease activity on germ line expansion by comparing the proportion of expanded alleles transmitted from Exo1A/A sires and dams. Consistent with our interpretation of the data from testes, Exo1A/A males had significantly more progeny with expansions than Exo1+/+ mice (Fig 3C). No significant differences in the proportion of expanded alleles were seen on maternal transmission (Fig 4A). This is likely because the fraction of expanded alleles was already so high in this population. However, the progeny of Exo1A/A mothers had significantly larger alleles than the progeny of Exo1+/+ mothers (Fig 4B). This is consistent with our previous demonstration that expansions continue to accumulate with time on previously expanded alleles [31] and suggests that the loss of EXO1’s nuclease activity alone is sufficient to significantly increase expansions. Failure to see an effect of EXO1 in ovary may reflect the fact that oocytes represent a relatively small fraction of the cells of this organ making any specific effect in the gamete difficult to discern. Nonetheless, our data supports the contention that EXO1 protects the genome against intergenerational repeat expansion in both males and females.
We have shown that MLH3, a component of the MutL complex, MutLγ, is required for both somatic and germ line expansions in both males and females in the FXD mouse model (Figs 1 and 2). The effect of the loss of MutLγ on germ line expansion has not previously been reported, although a similar dependence on MutLγ for somatic expansions was seen in a mouse model of Huntington Disease [21] and a cell culture model of GAA-repeat expansion [22]. The role of MutLγ, coupled with the requirement for MutSβ, increases the likelihood that a similar basic mechanism accounts for all expansions in these disorders, despite differences with respect to which tissues are expansion prone and the contribution of MutSα to expansions in the FXD mouse model [8] and in FRDA iPSCs [42], but not in models of other Repeat Expansion Disorders [43–45]. We have also shown that not only is EXO1 not required for expansion, but it is actually protective, reducing the extent of both germ line and somatic expansions (Figs 3 and 4). We also showed that Exo1A/A animals have significantly more expansions than Exo1+/+ mice, but significantly fewer expansions than Exo1-/- mice. Given that the available data suggest that the D173A protein has a similar overall structure [33–35] and stability as the WT protein (Chahwan et al., manuscript in preparation), our data are consistent with the hypothesis that EXO1 reduces expansion in at least two different ways, one that is dependent on its nuclease activity and one that is not.
The fact that EXO1 is not required for expansions, but is actually protective, suggests that models that propose a role for excision of the strand opposite loopouts by enzymes like EXO1 (e.g., [5, 15–17]), may not account for expansions in the FXD mouse. Furthermore, while it is formally possible that EXO1 provides protection against expansions that occur via a MutS/MutLγ-independent pathway that is not seen when EXO1 is present, the most parsimonious explanation of our data is that EXO1 is acting as it usually does downstream of the MutS/MutL proteins, where it may be functioning, at least in part, to repair the repeat tract in a manner consistent with its normal role in MMR. Experiments are underway to test this hypothesis.
Since EXO1 is protective, it may be that expansions arise via the Exo1-independent sub-pathway of MMR [46, 47]. This pathway has been suggested to involve strand displacement or excision by one or more other nucleases [46]. It has been suggested that Artemis, FAN1 and/or MRE11 may be the nucleases involved [47]. However, since EXO1 is protective, it is unclear how these other nucleases would act to promote expansions and in fact, we have recently demonstrated that FAN1 is also protective [41]. Interestingly, the protective effect of FAN1 and EXO1 seem to be complementary, with EXO1’s effect being most apparent in the gonads (this manuscript) and FAN1’s effect being strongest in somatic tissue [41].
While our data shows that expansion proceeds via an EXO1-independent pathway, the fact that EXO1 can protect against expansion in a nuclease-independent way provides an important clue as to the events downstream of MutLγ-binding in the expansion process. MutLγ is much less abundant than MutLα in mammalian cells and unlike MutLα, MutLγ only plays a minor role in MMR. However, MutLγ is essential for processing of Holliday Junctions (HJ) during meiosis, a process in which MutLα plays no role. Thus, MutLγ processing of an intermediate that resembles a HJ may account for the specific requirement for MutLγ in expansions. EXO1 plays an important structural role in facilitating the proper orientation of MutLγ cleavage of HJs during meiosis [36]. We speculate that MutLγ processing of a HJ-like intermediate in the absence of EXO1 would generate staggered double strand breaks that could then be processed to generate expansions. A simple model for such a process is illustrated in S4 Fig. In the presence of EXO1, MutLγ processing may result in products that are processed in a way that does not generate expansions.
Human Genome-Wide Association Studies (GWAS) show that genetic factors previously implicated in mouse models of the Repeat Expansion Diseases, including MSH3 and FAN1, are important disease modifiers for a variety of these disorders [48–50]. This would suggest that expansions in mice and humans share a common mechanism. In light of that, our data demonstrating that EXO1, like FAN1, protects against expansions, suggests that EXO1 variants may also be important modifiers of disease risk.
MutLγ has been implicated in chromosome breakage/fragility of CAG-repeats in yeast [51]. This is consistent with a role of MutLγ in the generation of double-strand breaks, a role we suggest MutLγ plays in repeat expansion in the FXD mouse. Mutations in the putative endonuclease domain of Mlh3 results in meiotic defects [52] and the preferential knockdown of the Mlh3 isoform that contains the nuclease domain reduces expansion in a tissue culture model of GAA-repeat expansion [22]. Thus, the genetic data support a nucleolytic role for MutLγ in repeat expansion. However, no specific MutLγ cleavage has, as yet, been demonstrated on synthetic substrates [19]. Thus, further work is needed in order to be able to test this model and to better understand the events responsible for repeat expansion.
All reagents were from Sigma-Aldrich (St. Louis, MO) unless otherwise specified. Primers were from Life Technologies (Grand Island, NY). Capillary electrophoresis of fluorescently labeled PCR products was carried out by the Roy J Carver Biotechnology Center, University of Illinois (Urbana, IL).
The generation of the FXD mice was described previously [53]. The mice with null mutations in Mlh3 were obtained from Paula Cohen (Cornell University, Ithaca, NY) [30]. The generation of Exo1-/- mice has been previously described [24]. A full description of the generation of Exo1A/A mice will be published elsewhere (Chahwan et al, manuscript in preparation). Briefly, the mice were generated by CRISPR/Cas9 mediated gene editing of C57BL/6 zygotes with a guide RNA targeting the area containing Exo1 codon D173 in exon 6 (5’-ACUCUGACCUCCUCGCAUUUGG-3’) and donor DNA carrying the desired mutation and 1.0 kb homologous arms on each side (illustrated in S5 Fig). The resulting offspring were genotyped by PCR and Sanger sequencing to identify founders carrying the EXO1 D173A mutation (S5 Fig). The founders (F0) were mated with wild type C57BL6 mice to produce F1 heterozygotes carrying the D173A mutation. F1 mice were backcrossed to C57B6 mice (4-6x). The mutant protein was expressed a level similar to wild type EXO1 protein (Chahwan et al, manuscript in preparation). All mice were on a C57BL/6 background. Mice were maintained in accordance with the guidelines of the NIDDK Animal Care and Use Committee, who approved this research (ASP-K021-LMCB-15) and in accordance with the Guide for the Care and Use of Laboratory Animals (NIH publication no. 85–23, revised 1996). Euthanasia was carried out using compressed CO2 followed by cervical dislocation.
DNA from mouse tails at 3-week-old for genotyping was extracted using KAPA Mouse Genotyping Kit (KAPA Biosystems, Wilmington, MA). DNA from sperm was extracted as previous described [10, 41]. Briefly, sperm were collected by centrifugation at 500 g for 5 minutes then resuspended in 300 μL ATL buffer (Qiagen, Hilden, Germany) containing 0.55 mg/mL Proteinase K and 30 mM DTT and incubated overnight at 55°C. The samples were then mixed with 90 μL of 5 M NaCl and centrifuged at 13,000 g for 10 minutes. The supernatant was transferred to a new tube, mixed with 390 μL ethanol and placed at −20°C for 1 hour. The DNA was then pelleted by centrifugation at 13,000 g for 10 minutes, washed with 70% ethanol and dissolved in TE buffer at 55°C for 15 minutes. DNA was isolated from the organs of 6-month old male mice using a Maxwell 16 Mouse Tail DNA Purification kit (Promega, Madison, WI) according to the manufacturer’s instructions. A 5 cm region of the jejunum starting 10 cm downstream of stomach was used as the small intestinal sample and the DNA was isolated from this segment as described above for intact organs.
Genotyping of Mlh3 and Exo1 null mice was carried out with KAPA mouse genotyping kit (KAPA Biosystems, Wilmington, MA) according to the manufacturer’s instructions with primers Mlh3A (also known as Primer 12265) (5’-GGCCTCTTCGCTATTACGC-3’)/Mlh3B (Primer 16984) (5’-AAGCCAGTGTCTGCCACTCC-3’) primer pair to detect the mutant Mlh3 allele and Mlh3B/Mlh3C (Primer 16985) (5’-CCCACCTTCTCTACATCGTC-3’) to detect the WT Mlh3 allele (as described at https://www2.jax.org/protocolsdb/f?p=116:5:0::NO:5:P5_MASTER_PROTOCOL_ID,P5_JRS_CODE:24295,018845). The Exo1A (5’-CTCTTGTCTGGGCTGATATGC-3’)/Exo1B (5’-ATGGCGTGCGTGATGTTGATA-3’) primer pair was used to detect the WT Exo1 allele and Exo1C (5’-AGGAGTAGAAGTGGCGCGAAGG-3’)/Exo1B to detect the mutant Exo1 null allele. Exo1 D173A genotyping was carried out using an Amplification‐Refractory Mutation System (ARMS)-based assay [54] that we designed (illustrated in S5 Fig). Briefly, the tetra-primer pairs for this assay were designed using BatchPrimer3 (https://wheat.pw.usda.gov/demos/BatchPrimer3/). The primers Exo1-ARMS-AF1 (5’-GAAATGGCTTTTGGAAAGTTTTGTTCGC-3’)/Exo1-ARMS-AR2 (5’-CTTCTTACAGCCAAATGCGAGGAAGG-3’) were using for the mutated Exo1 A (GCC) allele and the primers Exo1-ARMS-DF2 (5’-CAGGCTGTCATCACAGAGGACTCCGA-3’)/ Exo1-ARMS-DR1 (5’-CCAAACTCCAAAGGATAAAACCAAGCCC-3’) were using for the WT Exo1 D (GAC) allele. The primer bases shown in bold are allele specific. The underlined bases in the primer are mismatches introduced to improve the specificity of the assay. The PCR mix contained 10 ng DNA, 1x PCR buffer, 0.2 mM dNTPs, 0.5 μM of each primer and 1 U JumpStart REDTaq, and the PCR parameters were 1x 96°C for 3 minutes, 8x (94°C for 30 seconds, 72–65°C (-1°C/cycle) for 30 seconds, 72°C for 1 minutes), followed by 27x (94°C for 30 seconds, 65°C for 30 seconds, 72°C for 1 minutes), and ending with 72°C for 10 minutes. Fmr1 PM allele genotyping and repeat size analysis was carried out using a fluorescent PCR assay and FAM-labelled FraxM4 (FAM-5’-CTTGAGGCCCAGCCGCCGTCGGCC-3’) and FraxM5 (5’-CGGGGGGCGTGCGGTAACGGCCCAA-3’) primer pair as described previously [31]. Small pool PCR was used to analyze sperm DNA as previously described [10]. The PCR reactions were resolved by capillary electrophoresis on an ABI Genetic Analyzer [37]. The resultant fsa file was then displayed using a custom R script [55] that is available on request. For intergenerational (germ line) transmissions the number of alleles that were larger, smaller or the same size as the parental allele were then scored.
The Somatic Instability Index (SII) typically used to quantify somatic expansions [57] is sensitive to the ratio of cells containing expanded alleles to the cells lacking expanded alleles. Therefore, it is not suitable for comparing somatic expansions in the testes of wildtype and either Mlh3-/- or Exo1-/- mice, since the Mlh3-/- and Exo1-/- mice lack mature gametes that constitute the bulk of cells present in the wildtype testes. It is also unreliable in organs like small intestine, where cells of the mucosal layer, which show a high degree of expansions, are easily lost during isolation. We therefore used slight modification of a previously described approach to quantitate somatic expansions [56]. We have previously shown that heart shows no postnatal expansions [37]. The peak seen in the repeat PCR profile from heart thus reflects the original inherited allele size. The PCR profiles for other organs show either a single peak that is larger than the peak seen in heart or two peaks, one corresponding to the original allele and the second corresponding to the expanded allele. The size of the expanded allele and the original allele are obtained from the repeat PCR profiles and the difference between them divided by 3 to obtain the repeat number added to the expanded allele. This measure correlates well with the SII for tissues where the SII is appropriate to use.
The extent of expansion in different tissues was compared in animals of 3 different genotypes using a Jonckheere-Terpstra test with permutation-based exact inference and a Hommel procedure to adjust for multiplicity. Pairwise comparisons of organ-tissue and genotype combinations whose differences remained significant at a nominal level (10%) were then carried out using Mann-Whitney U tests (exact inferences), again with a Hommel procedure to adjust for multiple comparisons. These calculations were carried out using R version 3.2 (exactci, COMPoissonReg packages), SAS version 9.4, and StatXact version 8 (cran.r-project.org; www.sas.com; www.cytel.com). Fisher’s exact test for the comparison of the number of intergenerational expansions relative to alleles that did not expand were carried out using the GraphPad QuickCalcs website (http://www.graphpad.com/quickcalcs). Mann-Whitney U tests for comparisons of two sample groups using normal-approximation inferences were carried out using the Vassarstats website (vassarstats.net).
|
10.1371/journal.pbio.1000443 | A New Type of Proton Coordination in an F1Fo-ATP Synthase Rotor Ring | We solved the crystal structure of a novel type of c-ring isolated from Bacillus pseudofirmus OF4 at 2.5 Å, revealing a cylinder with a tridecameric stoichiometry, a central pore, and an overall shape that is distinct from those reported thus far. Within the groove of two neighboring c-subunits, the conserved glutamate of the outer helix shares the proton with a bound water molecule which itself is coordinated by three other amino acids of outer helices. Although none of the inner helices contributes to ion binding and the glutamate has no other hydrogen bonding partner than the water oxygen, the site remains in a stable, ion-locked conformation that represents the functional state present at the c-ring/membrane interface during rotation. This structure reveals a new, third type of ion coordination in ATP synthases. It appears in the ion binding site of an alkaliphile in which it represents a finely tuned adaptation of the proton affinity during the reaction cycle.
| Like the wind turbines that generate electricity, the F1Fo-ATP synthases are natural “ion turbines” each made up of a stator and a rotor that turns, when driven by a flow of ions, to generate the cell's energy supply of ATP. The Fo motor rotates by reversible binding and release of coupling ions that flow down the electrochemical ion gradient across the cytoplasmic cell membrane (in the case of bacteria) or intracellular organelle membranes (in the case of eukaryotic cells). Here, we present the structure of a rotor (c-)ring from a Bacillus species (B. pseudofirmus OF4) determined at high-resolution by X-ray crystallography. This bacterium prefers alkaline environments where the concentration of protons (H+) is lower outside than inside the cell – the inverse of the situation usually found in organisms that prefer neutral or acidic environments. The amino acid sequence of the protein subunits in this rotor, nevertheless, has features common to an important group of ATP synthases in organisms from bacteria to man. The structure reveals a new type of ion binding in which a protonated glutamate residue in the protein associates with a water molecule. This finding raises the possibility considered by Nobel laureate Paul Boyer several decades ago that a hydronium ion (a protonated water molecule, H3O+), rather than a proton alone, might be the coupling species that energizes ATP synthesis. Also, it demonstrates the finely tuned adaptation of ATP synthase rotor rings and their ion-binding sites to the specific requirements of different organisms.
| Most living cells depend upon the adenosine triphosphate (ATP) generated by F1Fo-ATP synthases that are energized by a proton- or a sodium-motive force (pmf, smf). These multi-subunit enzymes contain a cytoplasmic F1 catalytic domain (subunits α3β3γδε) that is connected with a membrane-embedded Fo domain (ab2c10–15 in bacteria) by a central (γε) and peripheral (b2δ) stalk. Energetically down-hill ion translocation across the membrane through the Fo complex is mediated by successive interactions between the stator a-subunit and a rotor ring (c-ring). Translocation involves ion binding to an unoccupied c-subunit, rotation, and subsequent ion release. The c-ring is attached to the γε stalk subunits so that c-ring rotation causes rotation of the stalk. The inherently elastic [1] and asymmetric γ-subunit extends into the α3β3 headpiece [2] and by rotation [3] induces conformational changes [4] in the catalytic β-subunits, which results in ATP synthesis.
In the Na+-binding c11 ring from Ilyobacter tartaricus [5] and the H+-binding c15 ring from Spirulina platensis [6], the translocated ions are bound within the groove of two adjacent c-subunits in a coordination network including a conserved carboxylate (Glu). In both cases, the ion is further coordinated by a precise network of residues, several of which are common to both organisms (Figure 1). The ion specificity of these two systems is determined by several factors including the geometry and distances of the ion coordination network, and a water molecule [7] providing a coordination site for Na+. The ion binding specificity of ATP synthases in various cells is adapted to the physiological requirements of the organism, and the different binding motifs observed presumably reflect these adaptations. The range of the ion-binding motif includes variations from complete Na+-binding signatures to c-subunits where the conserved carboxylate (E/D) of the C-terminal helix is the only residue that can be predicted with confidence to play a role in ion coordination (e.g., in Escherichia coli or Homo sapiens, Figure 1). On this basis, we here assign the name “E/D-only” to the c-subunits of this sub-class of proton-coupled ATP synthases.
Alkaliphilic Bacillus species are among the bacteria having proton-coupled ATP synthases [8] with E/D-only c-subunits (Figure 1). The extreme alkaliphile Bacillus pseudofirmus OF4 grows by oxidative phosphorylation with cytoplasmic pH values maintained 1.5–2.3 pH units below the high external pH (up to 11) of the medium [9]. The existence of this reversed ΔpH poses a major thermodynamic problem, with which these cells must cope. Among a variety of adaptive strategies to resolve the energetic problem [10], some special adaptations of the ATP synthase itself have evolved: latent ATPase activity [9],[11], a-subunit modification [12],[13], and in particular, specific adaptations of the c-subunit sequence [14] resulting in a large c-ring width with more c-subunits [15]. The adaptations to alkaliphilic conditions in an ATP synthase rotor, with a widely found but structurally uncharacterized E/D-only motif, made its c-ring an attractive candidate for an X-ray diffraction study.
Three-dimensional crystals of the c-ring from native Bacillus pseudofirmus OF4 F1Fo-ATP synthase were obtained and diffracted to 2.5 Å (Table 1). An asymmetric unit contains one complete c-ring, forming crystal contacts with three neighboring, laterally translated c-rings including two loop-to-loop contacts and one loop-to-C-terminus contact. One c-ring is formed by 13 identical c-subunits with a central pore (Figure 2A). Because of the unusually short N- and C-termini typical of Bacillus species (Figure 1), it does not extend into the periplasmic space but instead ends at the periplasmic membrane surface, whereas on the cytoplasmic side, the protrusion out of the membrane is comparable to that observed in other c-rings (Figure S1). In contrast with the distinctly hour-glass shape of the c11, K10, and c15 rotor rings [5],[6],[16], this c13 ring resembles a “tulip beer glass” appearance, in which the only slightly curved periplasmic side transitions at about the depth of the ion-binding site into a more significantly flared cytoplasmic side (Figure 2B). The c13 ring has an outer diameter of 63 Å on the cytoplasmic side, 54 Å on the periplasmic side, and 52 Å in the middle. Each c-subunit consists of two α-helices that are connected by a short cytoplasmic loop (RQPE, residues 34 to 37). The N-terminal α-helices (residues 1 to 33) form an inner ring, surrounded by an outer ring of the C-terminal α-helices (residues 38 to 69) in staggered position. While the inner helices are remarkably straight (Figure 3A) and only slightly tilted by ∼5° toward the c-ring axis (as seen from the periplasm), the outer helices are more curved and have a kink in the middle of the helix at the key carboxylate residue (Glu54). Above that kink, toward the cytoplasmic loop, the outer helices are bent ∼30° to the membrane vertical and form a convex surface shape on the outside of the ring. Toward the periplasmic side the helices remain more straight but still tilt ∼10° against the membrane vertical (Figure 3). The outer surface of the c13 ring (Figure S2A) is encircled by a large hydrophobic region forming membrane borders at the level of Phe47 on the cytoplasmic side and Phe69 at the very end of the ring on the periplasmic side with a height of 34 Å. This region is the contact region of the c13 ring with the hydrophobic part of a membrane, in good agreement with previously described membrane borders for other c- (or K-) rings [5],[6],[16]. Except for the cytoplasmic end, the surface in the inner pore of the c13 ring is overall hydrophobic and binds detergent molecules (Figure S2B), which apparently replace phospholipids. A second leaflet of phospholipids/detergent molecules at the periplasmic side of the c-ring could only be discerned partially from the electron densities, but its existence can be inferred. In support of this notion are data showing a plug of phospholipids [17] in the I. tartaricus c11 ring and also atomic force microscopy topographs from Bacillus sp. strain TA2.A1 c13 ring [15].
Special amino acid sequence motifs are found in the c-subunits of alkaliphilic Bacillus species (Figure 1). An altered glycine motif in the inner helices is shown to affect the structure and biochemistry of c-rings [14],[15]. The 63 Å diameter of the c13 ring width (Figure 2A) is larger than expected if compared with the diameters of the c11 and c15 c-rings [5],[6], 50 Å and 65 Å, respectively, since it is closer to the diameter of the c15 ring. The more relaxed positioning of the helices in the alkaliphile ring is caused by the structural impact of the alanine motif AxAxAxA that replaces the glycine motif GxGxGxG most often found in the first helix coding region of c-subunits. Similarly, replacement of two glycines with serines also accounts for the larger than anticipated c-ring diameter observed in the ATP synthase of Bacillus TA2.A1 [15]. On the outer helices of the c13 ring from an alkaliphile, two prolines located one helix turn below and above the ion-binding glutamate can be identified in a motif (PxxExxP) [14] in which the first proline, Pro51, is specific for alkaliphiles [13]. Both prolines break the regular α-helix hydrogen bonding pattern and cause helix bends; these two motifs of this c-ring are important factors that have an impact on the c13 diameter and ion binding as well as the overall “tulip beer glass” shape of the complex (Figure S1).
Ion binding in all c-rings includes a conserved outer helix Glu (or Asp). In the c13 ring of B. pseudofirmus OF4, this residue (Glu54) is located ∼6 Å above the middle of the membrane (Figure 3) toward the cytoplasmic side. At the pH (4.5) used in the cryo-protection buffer, Glu54 is protonated. Two outer helices from neighboring c-subunits form an ion binding site (Figure 3C). During structure refinement a sphere-shaped density in 2Fobs-Fcalc as well as in the omit map remained unassigned. In close proximity of this density center (2.8-3.2 Å), four atoms were identified. One of these belongs to the side chain carboxyl oxygen (Oε2) of Glu54, whereas the three others originate from the backbone carbonyls of Leu52 and Ala53 and from the backbone nitrogen of Val56. The observed distances and the arrangement of the four associated hydrogen donor/acceptor sites around this density are in almost pyramidal arrangement and the hydrogen atom positions lie on a plane. Such an arrangement resembles internal protonated water molecules (hydronium ion) [18] in other proteins.
Electron densities of certain cations (e.g., Na+) or oxygen atoms from water molecules are similar and difficult to distinguish by X-ray crystallography at the given resolution. However, several lines of evidence indicate the density seen in the binding pocket of this c13 ring should be interpreted as water rather than Na+. The hydrogen-bonding distance and angle geometry for the ligands are in the typical range for water molecules in proteins, as they are also found, for example, in carboxypeptidase (Figure S3) or in the protonated water cluster of bacteriorhodopsin [19]. In contrast, the mean distances for Na+ coordination such as those in the Na+-binding c11 ring [7] are significantly shorter (∼2.3 Å) [20] than observed in the c13 ring. For direct experimental evidence we used NCD-4, a fluorescent analogue of the ATP synthase inhibitor DCCD, which is known to react covalently with protonated glutamates/aspartates in c-subunits [21]. Figure 4 illustrates the time-dependent labeling of detergent-solubilized c13 ring at pH 6.0. Consistent with dependence of labeling on protonation of the carboxylate, an increase of pH to 9.0 immediately and significantly decelerates the reaction. This observation has also been reported by others [22],[23] and a control experiment showed that the labeling by NCD-4 continued to increase linearly when the labeling time was extended to 9,000 s without an alkaline pH shift (Figure S4). The marked deceleration upon imposition of an alkaline pH shift also resembles the DCCD labeling pattern of the proton-coupled c15 ring as a function of pH (unpublished data). Most importantly, the presence of 200 mM Na+ shows no dramatic influence on the labeling kinetics and addition of 200 mM K+ leads to comparable effects to Na+. Whereas the presence of salts, either Na+ or K+, presumably causes minor changes on, for example, detergent micelles, water availability, or fluorophore quenching [24], the lack of a major Na+-specific effect on NCD-4 binding rates contrasts sharply with the immediate and strong Na+-protective effect of a much lower Na+ concentration (15 mM) on the Na+-binding c11 ring [25]. This key evidence for H+ rather than Na+ binding of the B. pseudofirmus OF4 c13 ring is fully consistent with biochemical studies on the pmf- (but not smf-) dependent B. pseudofirmus OF4 cells [26] and its H+-dependent ATP synthase [9]. The result furthermore suggests that the ion binding site is highly selective for H+ over Na+ essentially under any (physiological) condition with a minimum concentration excess of ∼108 Na+ (200 mM) over H+ (pH 9.0). Taken together, the findings indicate that the density observed in the binding site of the c13 ring of the B. pseudofirmus OF4 ATP synthase is an oxygen atom with four valences. The data show that the proton must be located in between the atom positions of Glu54(Oε1/2) and the water oxygen (O).
In contrast to other rotor ring structures [5],[6],[16], no residues from the inner helices contribute to ion coordination in the c13 ring (Figure 3B and Figure S1B). In the c11 and c15 rings a proline (Pro28 and Pro25, respectively) is involved in kinking the inner helices (not shown), thereby allowing the hydrogen bonding of the glutamine (Gln32 and Gln29, respectively) with the glutamate on the outer helix. This proline and glutamine are replaced by a glycine and valine, respectively, in the c13 ring. Consequently, the inner helices form a complete α-helical hydrogen bonding pattern, retain a straight shape, and cannot hydrogen bond with the ion-binding glutamate.
A second notable difference of the ion binding site in the c13 ring as compared with the c11 and c15 rings is visible at the second oxygen of the glutamate (Glu54 Oε1). Whereas in c11 and c15 this oxygen forms a hydrogen bond with a tyrosine from the adjacent outer helix, such an interaction is missing in the c13 ring (Figure 3C and Figure S1B). The hydrogen bonding network of Glu54 in the c13 ring is therefore reduced to one bond only compared to the others. The additional freedom allows more rotameric flexibility of the glutamate carboxylate. This property becomes impressively visible in an overlay of 13 single c-subunits taken from one asymmetric unit (Figure 5). Notably, although the carboxyl group appears in different rotameric states in the crystal structure, the distance of the closest Glu54 oxygen to the water oxygen remains in bonding distance (2.6–3.1 Å) across all binding sites of the c13 ring.
The subtle but important differences in the H-bonding network geometry allow a fine-tuning of the pKa of the carboxylic acid [27] and serve to optimize the required solvation energy [28] which is necessary to unlock the site and allow ion release and reloading during the ion translocation mechanism in the Fo complex (Figure 6 and Text S1). Fine tuning of these parameters is of crucial importance within the a/c-ring interface, where the rotor binding sites pass a more hydrophilic environment [29] (and J. D. Faraldo-Gómez, personal communication with TM) that is somewhat unique because of the adaptations in both the a- and c-subunits of the alkaliphile [13],[14],[15]. By contrast, while the ion-binding site is in contact with the hydrophobic barrier of the lipid phase, during the long rotation cycle of the rotor ring, the glutamate is expected to be neutralized. The structural data suggest that under these conditions the ion binding site of the c13 ring remains in the ion-locked state, much in analogy with the Na+- and H+-locked states in the c11 and c15 ring, respectively [5],[6].
The amino acid residues involved in the coordination of the water are exactly at the same positions of the c13 ring as those that coordinate the water in the Na+-binding c11 ring (Figure 1 and Figure S1B) [7]. This commonality of binding pattern underlines the evolutionary and functionally conserved relationship between the pmf- and smf-driven systems. The smf-driven ATP synthases have been suggested to be evolutionary pioneers in the establishment of the modern ATP synthases [30]. If this hypothesis is correct, the E/D-only type such as that seen in the B. pseudofirmus OF4 c13 ring and in the mixed type (c15 ring) primarily found in light-driven systems (chloroplasts, cyanobacteria) could be derivatives of the c11 basic structure from an evolutionary point of view. Rather small differences in the amino acid sequences of the c-subunits apparently account for the different ion binding types, which are phenotypically manifested in the differently coupled and differently environmentally adapted ATP synthases.
Hydrogen atoms have a very weak X-ray diffracting power and their electron density often does not match with the exact position of the nucleus. Therefore, from a crystallographic point of view, at current resolution, the data do not allow the distinction between a protonated glutamate associating with a water molecule and a hydronium ion as a separate species. The scenario of a hydronium ion as a possible coupling ion species in F1Fo-ATP synthases was proposed for consideration by Boyer more than 20 years ago [31]. Later, experimental differences in the pH-dependent inhibition kinetics of Na+- and H+-ATP synthases were interpreted to be in support of this hypothesis [32], but recent high-resolution structure data on the cyanobacterial [6] and chloroplast [33] c-rings clearly conflict with this as a general hypothesis. The possibility raised by the structural data presented here that this E/D-only type of c-ring may ultimately conform to Boyer's suggestion awaits further experimental (and/or quantum mechanical) analyses.
This work shows a new type of proton coordination in an F1Fo-ATP synthase rotor ring. An additional electron density within the protonated ion binding site corresponds to a water molecule (but not Na+). It is evident that the coordination network of the water itself, in analogy with the stable water coordination network in the Na+-binding c11 ring, is a stabilizing and therefore a structural part of this c-ring. The presence of the water has been shown to enhance the Na+-binding affinity in the Na+-binding c11 ring [7]. Given this observation we propose that the water in the c13 ring binding pocket also enhances the proton affinity. High affinity rotor binding sites are of central importance for all ATP synthases but are especially important for ATP synthases of bacteria that grow in alkaline environments [13],[14].
Some of the novel details of this c-ring are likely to be specific to Bacillus species growing at high pH [34], especially those differences in shape that relate directly to alkaliphile-specific motifs. Perhaps the novel manner in which a water participates in proton binding is also a consequence of adaptation of the ATP synthase to alkaliphily. Further structural analysis of c-rings from the large groups of non-alkaliphilic species harboring the E/D-only motif is necessary to clarify the precise role of such water molecules in the ion translocation process. It may reveal the presence of this ion binding type throughout a broader subset of H+-coupled rotors, where it influences both ion affinity and selectivity during torque generation in the Fo motors of the H+-dependent F-type ATP synthases, and possibly also for the ion-driven motors known from V-type and A-type ATPases/synthases.
The ATP synthase was purified from B. pseudofirmus OF4 in which a six histidine tag was inserted after the N-terminal methionine in the chromosomal gene encoding the β-subunit of the ATP synthase. The complex was extracted from everted vesicles with 1% β-dodecyl maltoside in the presence of 3 mg/ml soybean asolectin and purified by affinity chromatography on NiNTA agarose. The isolation of the c-ring was carried out according to [25]. To improve purity of the sample, the c-ring was concentrated by ultrafiltration with an Amicon tube with a molecular weight cut-off of 10’000 (Millipore GmbH, Schwalbach, Germany), incubated with 1.5% Foscholine-12 (w/v) at 45°C for 10 min, and run on a sucrosegradient [15],[35]. The c-ring containing fractions were concentrated by Hydroxyapatite (BioGelHT, Bio-Rad, Munich, Germany) [36] and dialyzed for 12 h (10 mM Tris/HCl pH 8.0) at 4°C. The c-ring was further concentrated using polyethylene glycol (PEG) [37] to a final concentration of 2.5 mg/ml (bicinchoninic acid assay, Pierce, Rockford, IL, USA). Crystals were grown by vapor diffusion in hanging drops at 18°C to a size of approx. 200×100×100 µm3. The c-ring sample was supplied with 1% (w/v) of β-undecyl maltoside and mixed with crystallization buffer (0.1 M sodium acetate, pH 4.3) and 20% PEG 400 (v/v). Before flash-freezing in liquid nitrogen, the rod shaped clear crystals were transferred for 2 min into a buffer containing 30% PEG 400 (v/v), 0.1 M sodium acetate pH 4.5, and 0.05% β-dodecyl maltoside (w/v).
A 60 µl sample (0.2 mg/ml) of purified c13 from Bacillus pseudofirmus OF4 in 12.5 mM MES-HCl (pH 6.0) buffer and 0.05% β-dodecyl maltoside (w/v) was used. Continuous increase of fluorescence was recorded with an F-4500 Hitachi Fluorescence Spectrophotometer (λex = 342 nm, λem = 465 nm). The reaction was started by the addition of 0.6 µl of NCD-4 (Invitrogen Inc.) from a 10 mM stock solution in dimethylformamide. After 2,000 s, the rate of reaction was greatly reduced by addition of 11 µl of 1 M Tris/HCl pH 9.0. The time required for the addition of these compounds (NCD-4 and Tris buffer) was approximately 5 s in both cases.
Data to 2.5 Å resolution were collected from a single crystal at the Max-Planck beamline X10SA (PX-II) at the Swiss Light Source (SLS, Villigen, Schwitzerland) and processed using the XDS package (Kabsch, 1993). The structure was determined by molecular replacement using PHASER (McCoy, 2007) with two bundles of six subunits from the structure of the c15 ring from Spirulina platensis [6] as search model. Model bias was removed by density modification and solvent flattening with RESOLVE [38]. Iterative cycles of model building and refinement were performed using COOT [39] and phenix.refine of the PHENIX package [40], respectively. During refinement, no non-crystallographic symmetry operation was applied. The refinement resulted in electron density maps that were unambiguously interpretable and after chain fitting the Ramachandran plot shows no outliers. Figures were generated using Povscript [41], POV-ray (http://www.povray.org), and Pymol [42]. Electrostatic potential distribution was generated using Pymol [42].
|
10.1371/journal.pgen.1002684 | Insulin Signaling Mediates Sexual Attractiveness in Drosophila | Sexually attractive characteristics are often thought to reflect an individual's condition or reproductive potential, but the underlying molecular mechanisms through which they do so are generally unknown. Insulin/insulin-like growth factor signaling (IIS) is known to modulate aging, reproduction, and stress resistance in several species and to contribute to variability of these traits in natural populations. Here we show that IIS determines sexual attractiveness in Drosophila through transcriptional regulation of genes involved in the production of cuticular hydrocarbons (CHC), many of which function as pheromones. Using traditional gas chromatography/mass spectrometry (GC/MS) together with newly introduced laser desorption/ionization orthogonal time-of-flight mass spectrometry (LDI-MS) we establish that CHC profiles are significantly affected by genetic manipulations that target IIS. Manipulations that reduce IIS also reduce attractiveness, while females with increased IIS are significantly more attractive than wild-type animals. IIS effects on attractiveness are mediated by changes in CHC profiles. Insulin signaling influences CHC through pathways that are likely independent of dFOXO and that may involve the nutrient-sensing Target of Rapamycin (TOR) pathway. These results suggest that the activity of conserved molecular regulators of longevity and reproductive output may manifest in different species as external characteristics that are perceived as honest indicators of fitness potential.
| In nature, a myriad of specialized traits have evolved that are used for intraspecific communication and mate choice. We postulated that certain traits may have evolved to be attractive by virtue of their accurate representation of molecular pathways that are critical for determining evolutionary fitness. Insulin signaling (IIS) is one such pathway. It has been shown to modulate aging, reproduction, and stress resistance in several species and to contribute to variability of these traits in natural populations. We therefore asked whether IIS affected key sexual characteristics and overall attractiveness in the fruit fly Drosophila melanogaster. We found that IIS regulates cuticular hydrocarbons (the key pheromones in flies), that reduced IIS also reduced attractiveness, and that flies with increased IIS were significantly more attractive than wild-type animals. Further experiments revealed that these effects may also be influenced by a second conserved nutrient-sensitive pathway, the TOR pathway. We suggest that natural selection may have favored a plethora of species-specific sexual characteristics because they accurately represent a small number of influential pathways that determine longevity and reproductive output across taxa. In other words, it may be that, whether fly or human, beauty is more than skin-deep.
| Organismal fitness is influenced by social interactions, which drive sexual selection and individual attractiveness. In nature, a myriad of specialized signals and cues are used for intraspecific communication and mate choice, and many attractiveness traits are known to reflect an individual's health and reproductive value. These indicator traits are presumed to be reliable because they are either costly to produce/maintain and therefore difficult to fake [1] or because they are subject to direct physiological constraints [2]. Regardless of their nature, effective quality indicators must be an honest reflection of an individual's reproductive potential [3], [4] and as such, must be linked at the molecular level to the key fitness parameters —longevity and reproductive output— that they represent. However, very few studies have identified specific molecular relationships that link attractive traits to the pathways that influence overall health and individual fitness (reviewed in [5]).
In Drosophila melanogaster, attractive traits include cuticular hydrocarbons (CHC), which are long-chain lipids deposited on the insect cuticle [6]. Their presumed ancestral function is desiccation resistance [7], but they also play major roles in insect social communication, species recognition, and as sex pheromones [8]–[10]. In Drosophila, manipulation of certain neuropeptide and endocrine systems, such as dopamine or juvenile hormone [11], affect CHC profiles, but the biological function of these alterations in CHC are unclear. At the molecular level, several genes have been implicated in CHC synthesis [12]–[15], but there is little information about the mechanisms that regulate their expression.
Insulin/insulin-like growth factor signaling (IIS) is an evolutionarily conserved pathway that influences animal development, metabolism, longevity, and fecundity [16], [17]. Reduced IIS generally extends lifespan, but it is normally accompanied by reduced reproduction [18], [19]. Conversely, increasing insulin signaling results in increased body weight and fecundity [20]. Pleiotropic effects like these are not uncommon, and they likely represent underlying trade-offs associated with the plasticity through which organisms alter their life history characteristics in response to environmental conditions to maximize individual fitness [16], [18], [19]. For example, animals with reduced insulin signaling are more likely to survive periods of acute stress or prolonged malnutrition, but they are readily outcompeted when nutrients are replete [21]. Standing genetic variation is also known to influence basal transcript levels of IIS pathway genes in flies [22], [23] and in humans [24] with potentially long-term effects on phenotypic condition (e.g. obesity in humans, [25]), and developmentally-determined traits (e.g. beetle horns, [26]). IIS is therefore likely to be an important mechanism through which many different organisms respond to variable environmental conditions to maximize fitness [21].
We hypothesized that certain attractive traits might represent conspicuous extensions of molecular pathways that are critical for determining fitness. Because fitness components are strongly influenced by shifts in resource allocation in response to changing environmental conditions, we reasoned that the chooser/assessor will be most interested in the immediate physiological state of a potential mate and that relevant pathways are likely to be master regulators of resource allocation. The IIS pathway was an obvious candidate to test.
To test our hypothesis we focused our initial experiments on CHC profiles in female flies carrying a loss of function mutation in the insulin receptor substrate, chico. chico mutant females have attenuated insulin signaling, and they are small, long-lived, and sterile [27]. We reasoned that studying female profiles would provide a clearer picture of the links between IIS and attractiveness because female attractiveness, unlike male, is less influenced by behavior and because the effects of IIS manipulation on lifespan and reproductive output are better understood, phenotypically and genetically, in females [28], [29]. In nature, male choice is important in many species [30], [31], including Drosophila, where mating opportunities are constrained by allocation of time and energy into courtship and ejaculate production [32].
Gas chromatography/mass spectrometry (GC/MS) and laser desorption/ionization orthogonal time-of-flight mass spectrometry (LDI-MS) were used to generate comprehensive CHC profiles in chico mutant and control flies sampled at four different ages (6, 23, 37 and 48 days post-eclosion) [33]–[35]. chico flies exhibited significant differences in the levels of most compounds (23/26 compounds in the GC/MS and 5/12 compounds in the LDI-MS analysis) (Figure 1). The number of differences was substantially greater than the expected number based on chance alone (1.3 differences for α = 0.05). Furthermore, of the 23 differences that were significant based on individual tests, 20 remained significant after a Holm-Bonferroni correction for multiple testing (7,11-heptacosadiene [7,11-HD], C26:2, and C24:0 did not achieve the modified threshold). Age-dependent changes in CHC profiles corresponded well with previous studies [36], and we were surprised to observe that these patterns were largely unaffected by chico mutation, despite a significant extension of their lifespan [27]. Only one CHC exhibited a statistically significant interaction between genotype and age (7-heptacosene, 7-H), suggesting that the majority of age-dependent CHC changes were independent of the mutation in chico (Figure 1).
To confirm that the observed phenotypes in chico mutants were indeed due to modulation of IIS, we measured changes in CHC profiles following manipulation of other components of the pathway. InR is the single insulin receptor in Drosophila, which binds insulin-like peptides and leads to activation of Akt kinase [37]. Pten phosphatase antagonizes IIS [38]. To avoid the developmental consequences associated with IIS manipulation, we employed the Geneswitch system (driven by a ubiquitous tubulin promoter in response to the drug RU486) together with UAS-AktRNAi, UAS-Pten, or UAS-InR to target transgene expression to adult flies. Comparisons were then made between adult females that experienced transgene expression following exposure to RU486 and control animals of the same genotype that were not exposed to the drug.
Down-regulation of IIS through expression of UAS-AktRNAi or UAS-Pten phenocopied the effects of chico mutation. Changes in CHC caused by the chico mutation and the two transgenic manipulations were highly positively correlated (Figure S1), and consistent changes were observed for several individual CHC. We observed reductions of 7-tricosene (7-T), n-tricosane (nC23), 9-pentacosene (9-P), 7,11-pentacosadiene (7,11-PD in GC/MS and C25H48 in LDI-MS), and 7-pentacosene (7-P). The levels of 2-methylhexacosane (2-MeC26), 5,9-heptacosadiene (5,9-HD) and 7,11-nonacosadiene (7,11-ND in GC/MS and C29H56 in LDI-MS) were increased (Figure S2, Table S1).
Activation of IIS through overexpression of InR produced effects on CHC profiles that were generally the converse of those generated by IIS knock-down. There was a highly significant negative correlation between CHC changes in chico mutant flies and InR over-expressing animals (Figure S1C), with overexpressing females exhibiting greater levels of 7-T, 9-P, 7,11-PD, and 7-P and reduced levels of 2-MeC26, 5,9-HD and 7,11-ND (Figure S2, Table S1). We note that RU486 alone had no significant effects on CHC profiles (Figure S3A). Together these data show that modulation of IIS is capable of both increasing and decreasing the representation of specific CHC from the levels observed in wild-type animals.
Having established that alterations in IIS impact CHC profiles, we next asked whether these changes affect sexual attractiveness. chico mutant flies were not studied in this context because of their small size [39]. We instead began by examining female attractiveness in Akt knockdown flies by assessing male preference to decapitated females in a two-choice courtship assay using live observation and video tracking. We found that wild-type Canton-S males spend significantly less courtship time with GeneSwitch>UAS-AktRNAi females exposed to RU486 (thus expressing the RNAi) compared to females not exposed to the drug (Figure 2A). Inhibition of IIS by overexpression of Pten also decreased female attractiveness, while activation of the pathway through InR overexpression increased attractiveness (Figure 2A). Control animals in these experiments are genetically identical but have not been exposed to the drug RU486, which induces transgene expression and itself has no effect on attractiveness (Figure S3C).
To confirm that preferences were based on chemical cues, CHC transfer experiments were conducted. We “perfumed” same-age oenocyte-less flies, which do not produce CHC [40], with either CHC from control flies or flies in which IIS was manipulated. We then tested male preference and found that males preferred oenocyte-less females perfumed with CHC from animals that overexpress InR over those covered with CHC from their corresponding control animals (Figure 2B). By design all characteristics except transferred CHC were effectively identical between perfumed oenocyteless females, demonstrating that CHCs are responsible for IIS-dependent increases in female attractiveness in our 2-choice assays. Conversely, experiments using UAS-AktRNAi resulted in reduced preference for oenocyte-less flies perfumed with CHC from Akt knockdown animals compared to CHC drawn from their controls (Figure 2C). The AktRNAi perfuming results were consistently more variable than those obtained using transgenic animals directly, and a strong trend was consistently observed (Figure 2C). However, when courtship assays were performed in the dark to exclude potential involvement of visual cues, a strong preference for control females remained, and when Geneswitch - UAS-AktRNAi transgenic animals either fed or not fed RU486 were perfumed with CHC from wild-type Canton-S females, their differences in attractiveness were masked (Figure 2D). These data further support the notion that differences in CHC are responsible for the differences in attractiveness. Consistent with its effects on female CHC profiles, therefore, modulation of the IIS pathway can both increase or decrease the attractiveness of wild-type females.
Our data reveal unexpected complexities by which individual CHC affect attractiveness. Several compounds are known to stimulate male courtship behaviors, including 7-P, 9-P, 7,11-HD and 7,11-ND [6], . While 7-P and 9-P levels were decreased following reduction in IIS, which is consistent with their reduced attractiveness, we did not observe significant changes in the levels of 7,11-HD. More surprising was that 7,11-ND, which is thought to promote male courtship, was increased following reduced IIS. Incidentally, an increase in 7,11-ND levels was recently observed in aging flies, which also resulted in reduced attractiveness [36]. It is possible that potent and unidentified pheromones are playing a large role in our observed effects. Candidates include C27H54O2, which is strongly promoted by IIS, and 2-MeC26, which is reduced.
Attractiveness may instead be determined by global properties of CHC profiles rather than by the additive contribution of select compounds. chico mutant flies and flies overexpressing Pten had relatively more CHC with longer carbon chains and fewer CHC with shorter chain lengths (Figure 3A). Expression of AktRNAi produced similar changes (P = 0.08, data not presented). RU486 alone had no systematic effect on CHC profiles of a control strain (Figure S3B). Aging has also been reported to result in increased longer-chained CHC [36], and the recurring similarities between reduced IIS and aging led us to examine this relationship more closely. Principle component analysis was used to distill changes in CHC across the profile into a small set of uncorrelated components and visually summarize their relationships. Based on the first two principle components (accounting for 57% of the variation), CHC profiles of young chico mutant flies resembled those of old control flies (Figure 3B). Aging impacted the components equally in both genotypes. Therefore, aging and IIS appear to act in parallel to shift the distribution of CHC in favor of those with longer carbon chains, which may reduce attractiveness. The similarities between young chico females and old control females may be reflective of their reduced reproduction.
To explore the molecular mechanisms through which IIS modulates CHC profiles, we measured expression of genes involved in CHC synthesis. Given that reduced IIS increased the representation of longer-chained CHC (Figure 3A), we predicted that these manipulations would result in increased expression of eloF, which is female-specific and involved in long-chain hydrocarbon synthesis [13]. Indeed, we found that mRNA levels of eloF were significantly elevated in manipulations that reduced IIS, including chico mutation, AktRNAi, and overexpression of Pten (Figure 4). Overexpression of InR had the opposite effect. Expression of desat2, which acts to produce 5,9-dienes, was significantly increased by reduction of IIS. Similar trends were observed for expression of desat1, which is required for the production of many alkenes [14], [43], and desatF, which introduces a second double bond to form female-specific dienes [12]. Together, these data suggest that IIS modulates CHC profiles at least in part through transcriptional regulation of the genes involved in their synthesis.
Four lines of evidence suggest that the effects of IIS on CHC expression are largely independent of its canonical transcription factor target, dfoxo. First, reduced IIS leads to activation of dFOXO, but overexpression of dfoxo had a negligible effect on overall CHC profiles (Table S1). Second, there was no significant correlation between changes observed in chico mutant flies and flies overexpressing dfoxo (Figure S1D). Third, unlike all of our other IIS manipulations, there was no effect of dfoxo overexpression on compound chain length (Figure 3C). Forth, CHC regulatory gene expression changes that were observed in chico mutant animals largely persisted in chico; dfoxow24 double mutants (Figure S4).
Because of its emerging importance in the biology of aging we asked whether the modulation of the target of rapamycin (TOR) pathway might be involved in the effects of IIS on CHC profiles. These two pathways are known to interact. Many studies have implicated insulin signaling as an important regulator of TOR activity [44], [45], and TOR signaling can activate IIS intracellularly through phosphorylation of Akt [46]. We found that suppression of TOR signaling through transgenic overexpression of a dominant negative TOR (UAS-TORTED) [47] resulted in CHC changes that were strongly positively correlated, but smaller in magnitude, to those induced by chico mutation (Figure S1E). There was also a significant effect of down-regulation of TOR signaling on the relative levels of CHC with greater chain length (Figure 3C). Together, our data suggest the hypothesis that alterations in IIS affect pheromone production and sexual attractiveness through mechanisms that are independent of dfoxo but involve the nutrient-sensing TOR pathway. Future studies focusing on specific TOR pathway modulators, such as S6K or 4E-BP, will be insightful in this regard. Finally, it may be interesting to examine the effect of juvenile hormone, which has been shown to influence fly CHC and is regulated by the IIS and TOR pathways, as potentially involved in the preferences that we report [11]. It has been linked to sexual attractiveness in other insect species.
We have found that key attractive traits in Drosophila melanogaster females, specifically cuticular pheromones (a.k.a., cuticular hydrocarbons, or CHC), along with gene expression of CHC synthesis enzymes and attractiveness of females, robustly respond to genetic manipulations of the IIS pathway. Based on these data, we suggest that CHC are readily detectable manifestations of IIS pathway activity and that they are used as agents of choice because they provide individuals with information about the reproductive potential—in accordance with environmental conditions—of a possible mate.
Why might CHC profiles be the indicators of IIS activity in flies? A putative ancestral function of CHC in insects is prevention of water loss and resistance to desiccation. Flies may actively increase CHC production, specifically heavy-chain CHC, to protect against stressful environments, as in the case of reduced IIS. Alternatively, it may be that alterations in CHC are pleiotropic side-effects of IIS targeted to other physiological traits. For example, IIS may regulate triglyceride levels by modulating the expression of desat1, which has an important function in lipid metabolism [48]. Functions for desat2 in starvation, cold resistance, and desiccation resistance have also been suggested [49]. Recent work has also shown that IIS influences female remating rate through unknown mechanisms likely related to metabolism, suggesting an additional link between this pathway and individual fitness [50].
Regardless of whether CHC production is a bona fide target of IIS, our data support a model whereby CHC profiles constitute reliable physiological indices of molecular pathways that determine fitness (Figure 5). Such indicator traits are honest, therefore, not because they are costly to produce but because their expression is tightly linked to the activity of these underlying major molecular pathways. Cheaters would therefore suffer from altering IIS to change CHC through pleiotropic effects on physiology, which would bring them out of line with existing environmental conditions and reduce individual fitness. We suggest that many sexually attractive characteristics, including those unique to individual species, may convey a universal aspect of beauty by accurately representing the molecular activity of a small number of highly conserved pathways that influence longevity and reproductive output across taxa. It will be interesting to determine whether IIS and possibly TOR signaling also impact attractiveness in other species, such as nematodes, mice, or humans, where the activities of these pathways have important health consequences.
Canton-S, w1118, and UAS-GFP was obtained from the Bloomington Stock Center. chico mutant flies and UAS-dFoxo flies were provided by M. Tatar [51] and L. Partridge [27], respectively. chico and their respective control flies are maintained contemporaneously in the same population and segregation of chico alleles is maintained by propagation of heterozygotes (normal-size, cinnabar). Segregating genotypes among sibs were identified as: ch1/+ normal-size, cinnabar; ch1/ch1, dwarf, cinnabar; +/+, normal-size, apricot [51]. The dfoxow24 strain was obtained from K. Weber [27], [52] and was subsequently backcrossed to a w1118 control strain for over 20 generations. This strain lacks four of five Foxo isoforms and has reduced expression of the fifth (dFoxoA). It is therefore expected to be a strongly hypomorphic allele. UAS-AktRNAi was purchased from the VDRC stock center. UAS-Pten/CyO was provided by S. Leevers, and UAS-InR was obtained from B. Edgar [53]. TOR dominant negative (UAS-TORTED) flies were obtained from the Bloomington Stock Center [47]. Oenocyte-less flies were created from the progeny of the cross of ‘+; PromE(800)-Gal4, tubP-Gal80ts; +’ to ‘+; UAS-StingerII, UAS-hid/CyO; +’; both strains were provided by J. Levine. tublin5-GeneSwitch flies were made by cloning the promoter of alphatubulin into the pSwitch2 vector.
The generation of oenocyte-less flies, which are largely devoid of CHC, followed published protocols [40]. Briefly, the progeny of the cross of “+; PromE(800)-Gal4, tubP-Gal80ts; +” to “+; UAS-StingerII, UAS-hid/CyO; +” were maintained at 18°C until eclosion. Following emergence, adult were kept at 25°C for at least 24 h. Then flies were subjected to three overnight heat treatments at 30°C (on days 2, 3 and 4) and left to recover for at least 24 h. GFP fluorescence was checked to confirm oenocyte ablation.
For all experiments, larvae were cultured in cornmeal-sugar-yeast “larval” media, and virgin adults were collected shortly after eclosion. For Canton-S, chico, dfoxow24 and chico; dfoxow24 double mutants (and control), flies were kept on 10% sugar/yeast (SY) food. All other mutants made by crossing tublin5-GeneSwitch flies to specific UAS-lines (AktRNAi, Pten, InR, dfoxo and TORTED) were placed into 10% SY food with RU486 (200 µM) to activate transgene expression (treatment) or with vehicle only (80% ethanol) (control) for 10–15 days before experiments. All flies were maintained at 25°C and 60% relative humidity in a 12∶12 h light∶dark cycle. Fresh food was provided every 2 or 3 days. Detailed media recipes can be found in Poon el al. [54].
Independent procedures were applied to collect aged flies (chico and control) for examing CHC in GC/MS and LDI-MS. For GC/MS, a large cohort of each genotype was established by collecting virgin females into vials following eclosion. CHC samples were extracted from these cohorts every 2–3 weeks. In contrast to the GC/MS analysis, multiple, independent cohorts were established for LDI-MS measurement every 2–3 weeks, and all flies were sampled on the same day for CHC analysis.
Total RNA was extracted from 10 virgin females at 10–15 days of age by Trizol (Invitrogen). Extracted RNA was treated with 1 U DNAse I (Invitrogen) and reverse transcribed into cDNA by Superscript III First-Strand Synthesis (Invitrogen) using oligo-dT primers. For each RNA extraction, five replicate RT-PCR reactions were performed using an ABIPrism 7000 and RT2 SYBR Green/Rox PCR Master Mix (SA Biosciences) with specific primers. The quantitative levels were normalized to an endogenous control rp49, calculated by the ΔΔCT method, and presented as fold-change of mutant to wildtype in expression levels. The results for CHC synthesis genes (eloF, desat1, desat2 and desatF) were based on at least three, independent RNA extractions. The following primers were used: desat1F (TGCCGATTGCTTGCTTCAT), desat1R (TTCACCCCAGGCGTACATG), desat2F (GGTGGTGCTTCCAGCTAAACA), desat2R (GGCGATTTCCGAATTTATGG), desatFF (TCCGTGTGGGTGAGGGATA), desatFR (AGCTCGGCGCTCTTGTAGTC), eloFF (CCATTATTCTGCTCCACTGTACCA), eloFR (GTCTGTTGACCGCGCAGTT), Rp49F (ACTCAATGGATACTGCCAG) and Rp49R (CAAGGTGTCCCACTAATGCAT).
For GC/MS and LDI-MS data, pairwise comparisons between IIS mutants and control flies were examined by two-factor ANOVA. Statistical analysis and data presentation (see Figure 1) used CHC values after transformation to the natural log scale, where it was determined that model residuals were sufficiently normally distributed and independent of fitted values. A single potential outlier was present for each of four individual CHC. After removal and reanalysis, all four compounds retained their significance, and P-values were substantially reduced in all cases. For consistency, therefore, we report the conservative P-values from ANOVA using all data. Data from only one compound in the GC data (7-H) exhibited a significant genotype×age interaction (P = 0.004). Standard least-squares regression was used to determine the correlation between chico and other IIS mutants (Figure S1) and the correlation between carbon chain length and the percent change of normalized intensity in IIS mutant from control (Figure 3). It should be noted that these P values may be liberal because, without detailed knowledge of the biochemical pathways of all CHC, we can not rule out that the levels of some CHC may be correlated. chico data represent the genotype main effect derived using data from all ages, while data from other genotypes represent replicate measures obtained at roughly two weeks of age. Principle component analysis (PCA) on correlations followed by ANCOVA was used to visualize the effect of aging and chico mutation (Figure 3). PCA was done using 72 CHC samples from transgenic flies manipulated for different components of the insulin signaling pathway and their appropriate controls. PC1 was responsible for 44% of the variation and is represented by positively loading C21–26 CHC (8 CHC have factor loadings of >0.8, and another 4 CHC have factor loading of >0.6) and negatively loading 7-H (−0.680) and 7,11-ND (−0.735). PC2 explains 13% of variation and is represented by three positively loading C25 compounds, and 2MeC28. For both courtship assay and video analysis (Figure 2), a Wilcoxon signed rank test was applied to test the null hypothesis of no preference (no difference from 50%). For quantitative PCR, a z-test was applied to test the null hypothesis of no change in expression level. Analyses were performed using JMP 8.0.1 and R 2.13.0.
To avoid biasing results due to timing and positioning in behavioral trials (timing of decapitation and placement in choice chambers, position relative to the light source), females from different experimental treatments (RU+ and RU−) were alternated in space and time. When several replicates were perfomed, they were pooled together and significance values were determined by a permutation procedure whereby treatment labels were randomized among flies within a specific replicate. For each of 30,000 randomizations, an attractiveness value was calculated, and the 30,000 values were then pooled together to create the null distribution. One- or two-sided p-values were then determined by integrating appropriate tails of the null distribution that were more extreme than the observed attractiveness value.
|
10.1371/journal.ppat.1004205 | A Central Role for Carbon-Overflow Pathways in the Modulation of Bacterial Cell Death | Similar to developmental programs in eukaryotes, the death of a subpopulation of cells is thought to benefit bacterial biofilm development. However mechanisms that mediate a tight control over cell death are not clearly understood at the population level. Here we reveal that CidR dependent pyruvate oxidase (CidC) and α-acetolactate synthase/decarboxylase (AlsSD) overflow metabolic pathways, which are active during staphylococcal biofilm development, modulate cell death to achieve optimal biofilm biomass. Whereas acetate derived from CidC activity potentiates cell death in cells by a mechanism dependent on intracellular acidification and respiratory inhibition, AlsSD activity effectively counters CidC action by diverting carbon flux towards neutral rather than acidic byproducts and consuming intracellular protons in the process. Furthermore, the physiological features that accompany metabolic activation of cell death bears remarkable similarities to hallmarks of eukaryotic programmed cell death, including the generation of reactive oxygen species and DNA damage. Finally, we demonstrate that the metabolic modulation of cell death not only affects biofilm development but also biofilm-dependent disease outcomes. Given the ubiquity of such carbon overflow pathways in diverse bacterial species, we propose that the metabolic control of cell death may be a fundamental feature of prokaryotic development.
| Many bacterial species including the pathogen Staphylococcus aureus are capable of adhering to surfaces and forming complex communities called biofilms. This mode of growth can be particularly challenging from an infection control standpoint, as they are often refractory to antibiotics and host immune system. Although developmental processes underlying biofilm formation are not entirely clear, recent evidence suggests that cell death of a subpopulation is crucial for its maturation. In this study we provide insight regarding the metabolic pathways that control cell death and demonstrate that acetate, a by-product of glucose catabolism, potentiates a form of cell death that exhibits physiological and biochemical hallmarks of apoptosis in eukaryotic organisms. Finally, we demonstrate that altering the ability of metabolic pathways that regulate acetate mediated cell death in S. aureus affects the outcome of biofilm-related diseases, such as infective endocarditis.
| The balanced progression of cell division and apoptotic events is a classic hallmark of eukaryotic development [1]. Intriguingly, a similar homeostatic control of cell death, lysis and proliferation is predicted to benefit the development of adherent multicellular bacterial assemblages (called biofilms) by providing nutrients and critical biofilm building matrix components like extracellular DNA (eDNA) [2], [3]. Consistent with this assumption, recent investigations have revealed that bacteria, like eukaryotes not only harbor elaborate regulatory systems that modulate cell death, but also display biochemical and physiological hallmarks characteristic of programmed cell death (PCD) [4], [5], [6].
The molecular components that mediate cell death in S. aureus are regulated, in part, by the LysR-type transcriptional regulator, CidR [7] and include a set of membrane bound proteins, CidA and CidB, whose functions are predicted to be analogous to the Bcl-2 family of apoptotic modulators in eukaryotes [2], [3]. However, less clear are the mechanistic contributions of other members of the CidR regulon in cell death, specifically those enzymes that are active during overflow metabolism, pyruvate oxidase (CidC) and α-acetolactate synthase/decarboxylase (AlsSD) [8], [9]. Given that these enzymes are the only additional members of the CidR regulon and that multiple physiological signals that directly affect both central metabolism and cell senescence coordinate their expression [10], we predicted an intricate role for these proteins in the physiology of cell death.
Here, we report that both CidC and AlsSD carbon-overflow pathways contribute to staphylococcal cell death. Our results demonstrate that cell death is potentiated by acetate, a major weak acid byproduct of glucose catabolism, whose levels are antithetically modulated by CidC and AlsSD activities. We also report that the physiological features accompanying staphylococcal cell death resemble eukaryotic PCD (apoptosis) wherein cell death is associated with respiratory dysfunction, increased ROS production and DNA damage. Finally, we demonstrate a role for staphylococcal PCD in biofilm development and pathogenesis.
Multiple studies have linked the uptake and metabolic fate of glucose to the regulation of PCD during eukaryotic development [11]. To determine whether such correlations are broadly conserved in bacteria, the effects of glucose on staphylococcal cell death were assessed over a period of five days by monitoring the colony forming units (cfu/ml) of wild-type cells grown aerobically in rich media (tryptic soy broth, TSB) containing either 14 mM or 35 mM glucose. Although there appeared to be no significant difference in the viable cell counts after 24 h of growth in either type of media, subsequent stationary phase survival of wild-type cells was dependent on initial glucose concentrations wherein S. aureus grown in TSB-35 mM glucose displayed a steep decline in viability (∼7 log10 difference) compared to the modest decline (∼1.2 log10) observed for cells grown in TSB-14 mM glucose over the same period of time (120 h) (Fig. 1A). These results indicate that growth in excess glucose reduced the survival of S. aureus in stationary phase.
To ascertain whether excess glucose-mediated staphylococcal cell death bore similarities to PCD, we explored the physiological status of the dying population by flow cytometry after 72 h of growth and compared them to a 24 h reference point when cells were relatively healthy based on viable counts. Respiratory potential was estimated using the cell permeable redox dye, cyano-2,3-ditolyl tetrazolium chloride (CTC). Reduction of CTC into a red insoluble fluorescent formazan that accumulates intracellularly is achieved by dehydrogenases of the electron transport chain (ETC) that are expressed within actively respiring bacterial populations [12]. In addition to CTC, cells were co-stained with 3′-(p-hydroxyphenyl) fluoroscein (HPF), a cell permeable fluorescent reporter that has widely been used to detect levels of highly reactive oxygen species like hydroxyl radicals (OH•) [13]. CTC staining of S. aureus grown in TSB-14 mM glucose revealed a healthy respiring sub-population (∼34%) at 24 h and a relatively smaller population of cells (13%) undergoing oxidative stress (Fig. 1B, Table S1). By 72 h the respiring population under these very same conditions increased to 79% (Fig. 1B, Table S1). This is expected as dependency on the TCA cycle and oxidative phosphorylation for cellular energetic needs increases in stationary phase, upon exhaustion of glucose from the media. Interestingly, S. aureus grown in TSB-35 mM glucose revealed an even larger population (∼61%) that reduced CTC at 24 h when compared to 14 mM glucose (Fig. 1B, Table S1). The increased reduction of CTC under these conditions could not have resulted from a corresponding increase in the rate of cellular respiration, as the ability of these cells to consume oxygen as a terminal electron acceptor had significantly decreased (Fig. 1C). These observations suggest that aerobic growth in excess glucose not only results in the inhibition of respiration, but may also promote the promiscuous transfer of electrons to alternate acceptors like CTC, due to a bottleneck in the ETC.
The transfer of electrons via a functional ETC has previously been proposed to ameliorate oxidative stress by curtailing the single electron reduction of oxygen to superoxide radicals (O2•−), a precursor of the highly reactive hydroxyl radical (OH•) [14]. Hence, we argued that the decreased functionality of ETC observed for cells grown in excess glucose may eventually promote the production of ROS. Although we did not observe OH• at 24 h of growth, we detected a dramatic increase of HPF stained cells by 72 h of growth in TSB-35 mM glucose but not in TSB-14 mM glucose (Fig. 1B, Table S1). The temporal production of ROS was confirmed by electron paramagnetic resonance (EPR) spectroscopic analysis of samples incubated with the spin probe, 1-hydroxy-methoxycarnonyl-2,2,5,5-tetramethyl-pyrrolidine hydrochloride (CM-H). Cells grown in TSB-35 mM glucose exhibited approximately 3-fold increase in EPR peak amplitude by 72 h relative to those grown in TSB-14 mM glucose (Fig. 1D). To determine the chemical nature and relative levels of various ROS produced, we incubated samples with either superoxide dismutase (SOD; O2•− scavenger) or dimethyl thiourea (DMTU; OH• scavenger) prior to the addition of CM-H. This approach revealed that cells undergoing cell death produced both superoxide and hydroxyl radicals (Fig. S1).
An abundance of cellular ROS mediates several types of DNA damage, including single and double stranded breaks that lead to DNA fragmentation [15]. We performed TUNEL (terminal deoxynucleotidyl transferase dUTP nick end labeling) assays on S. aureus undergoing oxidative stress to estimate the population of cells undergoing DNA fragmentation by flow cytometry. Consistent with the temporal pattern of ROS production observed earlier, we detected a sub-population of cells with fragmented DNA (TUNEL positive) by 72 h of growth under excess glucose conditions (Fig. 1E). Notably, only minimal TUNEL staining (Fig. 1E) was detected after 72 h for cultures supplemented with 14 mM glucose. Collectively, these observations suggest that cell death resulting from growth under excess glucose exhibits multiple hallmarks of eukaryotic PCD. Interestingly, the phenotypic hallmarks of PCD were not restricted to growth of S. aureus under glucose rich conditions alone, but were also observed when grown in the presence of excess fructose, mannitol and sucrose suggesting a strong association between carbon catabolism and cell death (Fig. S2).
How does excess glucose mediate cell death? It is well known that S. aureus cultured in excess glucose undergoes CcpA-mediated catabolite repression [16]. This ensures that acetate accumulates in the media as a byproduct of glucose catabolism. However, once glucose is completely exhausted from the media, the TCA cycle is progressively relieved of CcpA repression and excreted acetate is oxidized to generate energy required for subsequent growth [16]. Growth of S. aureus in TSB-14 mM glucose displayed such a classic diauxie, where glucose was consumed within 5 h and subsequent growth was dependent on the consumption of acetate by 9 h (Fig. 2A–C). The temporal dynamics of acetate levels in the media were also reflected in the pH shift of the culture supernatant from 7.2 to 5.5 during acetate accumulation and from 5.5 to 7.2 during its depletion (Fig. 2C, 2D). As observed previously [17], growth of S. aureus in TSB-35 mM glucose did not display the expected diauxic shift (Fig. 2A). Although excess glucose was consumed within 9 h, acetate remained unutilized and the pH of the culture was maintained at 4.6 (Fig. 2B–D).
Based on these observations, we hypothesized that cells grown in excess glucose would eventually be inhibited by high concentrations of acetate and low pH. The growth inhibitory effects of weak organic acids like acetate are largely dependent on pH [18]. As the extracellular pH nears the pKa of acetate (pKa = 4.76), the protonated (uncharged) membrane permeant form of the acid (CH3COOH) replaces its corresponding ionic forms (CH3COO−; H+), thus allowing the former species to passively breach bacterial membranes and dissociate within their relatively neutral cytoplasm [18]. If left unchecked, such an event may lead to growth inhibition and lethality through cytoplasmic acidification [18]. Given that the pKa of acetate is easily met during growth in excess glucose, it seemed plausible that this acidic metabolite may represent a physiological trigger for cell death. To test whether acetate was capable of inhibiting S. aureus growth under acidic conditions, we buffered TSB at pH 4.8 using 30 mM HOMOPIPES (homopiperazine-N,N′-bis-2-(ethanesulfonic acid) and challenged cultures with the sodium salt of acetate. As expected, acetate inhibited the growth of S. aureus under these conditions, but neither acidic pH alone nor addition of an equimolar concentration of sodium chloride (50 mM) inhibited growth to the same extent as sodium acetate (Fig. 2E). Further, the addition of a non-metabolizable weak acid, benzoate (pKa = 4.2) was as toxic as acetate under low pH (Fig. 2E). These observations suggest that acetate mediated growth inhibition is a direct consequence of intracellular acidification and not due to secondary metabolites resulting from intracellular catabolism of acetate.
To confirm that cell death is dependent on the weak acid properties of acetate, we grew S. aureus in TSB-35 mM glucose that was buffered to a pH of 7.3 with 50 mM MOPS (3-(N-morpholino) propanesulfonic acid). We reasoned that although cells would utilize excess glucose to generate acetate, the relatively neutral pH of the medium would allow it to remain in the ionic state and prevent it from permeating and acidifying the interior of cells. Indeed as predicted, S. aureus under these conditions did not undergo cell death despite its growth in excess glucose (Fig. 2F). Remarkably, we also observed a dramatic reduction in the generation of ROS (Fig. 2G–H, Table S1), decreased DNA damage (Fig. 2I) and comparable rates of respiration relative to wild-type (Fig. 2J) under these conditions, suggesting that excess glucose per se was not responsible for the phenotypes associated with cell death. Rather these observations collectively demonstrate that acetate, a metabolic byproduct of glucose catabolism triggered these phenotypes under acidic pH.
Although acetate is primarily produced in S. aureus by the phosphotransacetylase (Pta)-acetate kinase (AckA) pathway, this metabolic pathway is unlikely to be directly involved in cell death as its activity is evident even during growth in 14 mM glucose, a condition where cell death is not triggered [19]. Additionally, disruption of this pathway surprisingly enhanced the rate of cell death during growth despite a decrease in acetate production [19]. This led us to reason that acetate-dependent cell death must be controlled by an alternate pathway, such as CidC. The cidC gene encodes a pyruvate oxidase that directly converts pyruvate to acetate and carbon dioxide and its expression is partly under the control of the regulator, CidR, whose activity is up-regulated in response to excess glucose [7], [19]. As the alsSD metabolic operon that results in the conversion of pyruvate to acetoin is also co-regulated by CidR, we hypothesized that both these pathways may modulate acetate-dependent cell death by competition for their common substrate, pyruvate, under conditions of excess glucose (Fig. 3A).
To test this hypothesis, we initially determined the levels of acetate and acetoin in 24 h culture supernatants of both ΔcidC and ΔalsSD mutants relative to WT. Compared to the wild-type strain, the ΔcidC mutant grown in TSB-35 mM glucose accumulated less acetate and relatively higher levels of acetoin (Fig. 3B, 3C). Conversely the ΔalsSD mutant excreted an excess of acetate (Fig. 3B) suggesting that both pathways competitively displaced pyruvate. We then tested the effects of cidC and alsSD pathways on cell death. In agreement with earlier studies [17], [20], mutation of either of these pathways resulted in opposing survival trends in stationary phase. Accordingly, a metabolic block in CidC activity (ΔcidC) enhanced stationary phase survival, while that of AlsSD (ΔalsSD) resulted in an increased rate of cell death compared to the wild-type strain (Fig. 3D). Consistent with the increased survival of the ΔcidC mutant and in contrast to WT and the ΔalsSD mutant, HPF-CTC double staining of 72 h cultures revealed a healthy population of respiring cells that exhibited low levels of ROS (Fig. 3E, Table S1), a phenotype that was also confirmed by EPR spectroscopy (Fig. 3F). In addition, flow cytometry detected fewer TUNEL-positive cells in the ΔcidC mutant, suggesting decreased DNA damage in these cells (Fig. 3G). Indeed, the cell death phenotypes associated with both ΔcidC and ΔalsSD mutants could be complemented in trans (Fig. S3) confirming their role in modulating cell death. Taken together, these data support the hypothesis that both CidC and AlsSD pathways modulate cell death by controlling flux through the pyruvate node.
Surprisingly, the ΔalsSD mutant generated ROS and exhibited DNA damage at levels similar to the wild-type strain (Fig. 3E–G, Table S1) despite an increase in loss of viability (Fig. 3D). This raised the possibility that in addition to affecting excreted acetate levels, there may be additional mechanisms by which the ΔalsSD mutant regulates cell death. To test this hypothesis we constructed a double ΔcidCΔalsSD mutant. We reasoned that if regulation of extracellular acetate by substrate competition was the primary mechanism by which AlsSD modulated cell death, then the ΔcidCΔalsSD double mutant would phenocopy the ΔcidC mutant. However, our results clearly demonstrate that the ΔcidCΔalsSD double mutant exhibited increased loss of viability (Fig. 3D), increased generation of ROS (Fig. 3E–F, Table S1) and excessive levels of DNA damage (Fig. 3G) than the ΔcidC mutant despite excreting comparable levels of acetate (Fig. 3B). Such an effect may occur if mutation of alsSD caused cells to be more susceptible to lower concentrations of weak acids in the culture media. Indeed, growth of ΔalsSD mutants challenged with acetate, lactate or pyruvate was more easily inhibited than wild-type (Fig. S4).
We tested two plausible hypotheses to explain the observed hyper-susceptibility of ΔalsSD mutants to weak acid stress. First, we argued that the end product of AlsSD catabolism, acetoin, itself may be necessary to withstand weak acid stress as it is known to contribute to the maintenance of cellular redox status upon being converted to butanediol or serve as a carbon source during stationary phase of growth. To test this hypothesis we subjected the ΔalsSD mutant to pyruvic acid stress and asked whether supplementation of excess acetoin in culture could rescue the pyruvate-mediated growth inhibition of this mutant. Our results demonstrate that acetoin could not restore pyruvate-mediated growth inhibition of the ΔalsSD mutant (Fig. S5). Furthermore, transformation of the ΔalsSD mutant with a plasmid bearing the alsS gene alone (in the absence of its cognate partner, alsD) under the control of its native promoter was able to rescue this mutant from pyruvic acid-mediated stress to growth rates comparable to the parental control, thus excluding any role for acetoin in promoting weak acid resistance (Fig. 4A).
We next hypothesized that AlsSD might play an active and crucial role in detoxifying intracellular acidification that accrues from the deprotonation of weak organic acids in the relatively neutral bacterial cytoplasm [21]. In such a scenario, intracellular protons would be consumed during multiple stages of decarboxylation, first of pyruvate into acetolactate catalyzed by AlsS, followed by that of the acetolactate intermediate into acetoin by AlsD, both leading to a gradual alkalization of the cytoplasm during weak acid stress. Direct evidence confirming a role for AlsSD in regulating intracellular pH was obtained by using cells loaded with the fluorescent pH probe 5 (and 6-)-carboxyfluorescein succinimidyl ester (cFSE). Resting cells that were suspended in potassium phosphate buffer (pH 4.5) maintained a slightly acidic interior (pHinternal of 5.9; Fig. 4B, top), resulting in a transmembrane pH gradient (ΔpH = pHinternal- pHexternal) of approximately 1.4 units. Addition of pyruvate under these conditions initiated a pH gradient (ΔpH) decay across the membrane that was exacerbated in the ΔalsSD and ΔcidC ΔalsSD backgrounds compared to either the parental or ΔcidC strains (Fig. 4B, bottom). Although the pH gradient decay recovered and stabilized over time, the rate and magnitude of the pHi recovery in different strains appeared to be dependent on the activity of AlsSD (Fig. 4B). In the presence of pyruvate, both the parental control and the ΔcidC mutant displayed comparable recovery rates of (23.83±1.9)×10−3 min−1 and (27.45±5.5)×10−3 min−1, respectively, and reached a pHi comparable to those of control cells (untreated resting cells) within 20 minutes (Fig. 4B, 4C). In contrast, the pHi of both the ΔalsSD and ΔcidCΔalsSD mutants had stabilized approximately 0.2–0.3 units below that of the pyruvate treated parental control and exhibited significantly decreased pHi recovery rates (P<0.05) of (11.66±2.1)×10−3 min−1 and (5.24±3.8)×10−3 min−1, respectively leading to incomplete recovery from acidic stress even after 60 minutes (Fig. 4B, 4C). These data demonstrate a role for the enzymatic activity of AlsSD in countering weak acid mediated intracellular acidification of the bacterial cytoplasm.
In eukaryotes, there is increasing evidence that ROS plays a key role in mediating PCD [22], [23]. Given that the physiological induction of ROS in S. aureus is dependent on the accrual of extracellular acetate, we next asked whether cell death is a direct consequence of acetic acid-mediated intracellular acidification or is an indirect result of oxidative stress. To this end we devised a strategy to determine the contribution of intracellular acidification on triggering cell death, independent of the ROS generated after 72 h of growth under aerobic conditions.
Wild-type S. aureus was aerobically grown for 24 h in TSB-35 mM glucose, followed by a sudden shift to anaerobic conditions. While this process ensured as much acidification as aerobically grown cultures, it conveniently eliminated ROS (due to the absence of oxygen). Although it is plausible that a sudden transition of cultures to anaerobic conditions may induce cell death independent of acidic stress, we controlled for this possibility by performing a similar experiment with wild-type S. aureus grown in TSB-35 mM glucose buffered to a pH of 7.3 with 50 mM MOPS. Cell viabilities monitored over a 120 h period showed only minimal loss of viability for neutrally buffered cultures, suggesting that cell death following anaerobic transition was dependent on culture acidification, similar to aerobically grown cells (Fig. 5A). Most importantly, unbuffered cultures that were shifted to anaerobic conditions displayed a partial restoration of viability compared to their corresponding aerobic cultures (Fig. 5A). Similarly a partial rescue was also observed when well-aerated cultures grown for 24 h were left standing without further agitation to minimize aeration (Fig. S6). Together, these data suggest a contributory role for oxidative stress in cell death.
Consistent with trends in cell viability, HPF-CTC staining of 72 h unbuffered cultures confirmed the absence of ROS following transition to anaerobiosis and the presence of a respiring population in contrast to that observed for the same time period under aerobic conditions (compare Fig. 5B with Fig. 1B, Table S1). Surprisingly, there was also a dramatic reduction of TUNEL positive cells following the shift to anaerobiosis (Fig. 5C), suggesting that ROS rather than intracellular acidification played a crucial role in DNA damage.
Cell death in staphylococcal biofilms is often spatially restricted to developing microcolonies [24]. Given that the CidR regulon is also actively expressed in microcolonies [24], we predicted that cell death may be modulated by both CidC and AlsSD activities and further contribute to the structural and developmental integrity of the maturing biofilm. To test these hypotheses, we assayed the ability of the ΔcidC and ΔalsSD mutants to form biofilms, relative to the wild-type strain on glass surfaces exposed to a continuous flow of nutrients. Wild type biofilms appeared as a confluent biomass frequently interspersed with microcolonies that differentiated from the primary biofilm mat. As previously noted, live/dead cell staining of wild-type biofilms confirmed that dead cell populations were predominantly localized within microcolonies (Fig. 6A). However compared to the wild-type strain, COMSTAT analysis of ΔcidC biofilms revealed significantly decreased total biofilm biomass and thickness, indicative of developmental defects during biofilm formation (Fig. 6B, 6C). Additionally, the roughness coefficients (a measure of the biofilm architectural heterogeneity) of the ΔcidC mutant biofilms were significantly lower than those of wild-type (Fig. 6D), a phenotype that was also consistent with the decreased ability of the ΔcidC mutant to differentiate into microcolonies. Finally, the ΔcidC biofilm revealed significantly less dead cell biomass compared to the parental strain, strongly suggestive of its involvement in promoting cell death in biofilms (Fig. S7). Similar to the ΔcidC biofilm, COMSTAT analysis of biofilms formed by the ΔalsSD mutant also exhibited decreased biofilm biomass and thickness compared to the wild-type strain (Fig. 6B, 6C). However it is unlikely that the observed decrease in biomass of one-day old ΔalsSD biofilms was due to an early developmental defect as they were able to differentiate into microcolonies and attain similar roughness coefficients to that of its isogenic wild-type strain (Fig. 6D). Rather, these defects are consistent with increased sensitivity of the AlsSD mutant to weak acids and low pH environments of biofilm microcolonies. Collectively, these observations suggest that the activities of both CidC and AlsSD regulate cell death at the population level to achieve optimal biomass and structural integrity during biofilm development.
Since the ability of bacteria to develop as biofilms on heart valves constitutes the root cause of infective endocarditis, we speculated that staphylococcal cell death may also contribute towards pathogenesis in a rabbit model of infective endocarditis. This model not only provides an estimate of the organism's biofilm forming capability in vivo but also allows for the simultaneous assessment of embolization (dissemination of the bacterial vegetation to secondary sites due to blood flow associated shear forces in the heart). To test this hypothesis we induced left-sided endocarditis in rabbits and infected them with wild-type, ΔcidC and ΔalsSD mutant strains. At 48 h post-infection, the bacterial burden in the primary vegetation (heart valves), heart tissue, kidney and blood were determined. Consistent with being the primary site of biofilm infection, the heart valves exhibited the maximum bacterial burden (∼108 cfu/gm of tissue) among the various tissues harvested (Fig. 6E). Relative to the wild-type strain, both ΔcidC and ΔalsSD mutants had similar bacterial loads at this site (Fig. 6E) suggesting that these metabolic pathways did not affect the growth of these strains or their ability to colonize the primary infection sites in vivo. Interestingly however, the ΔcidC mutant displayed significantly decreased bacterial burdens in the blood and other secondary sites of infection including heart tissue (excluding valves) and kidneys (Fig. 6E). As bacterial colonization of these secondary sites primarily results from infectious emboli originating from the heart valve, it may be argued that the cell-death associated CidC pathway may play a role in metastasis of the valvular vegetation.
In the present study, we demonstrate that both CidC and AlsSD pathways, that have traditionally been considered metabolic routes for excess carbon flow, antithetically modulate staphylococcal cell death by regulating the levels of excreted acetic acid. Exercising such metabolic control over cell death affords S. aureus a means to modulate biofilm development and possibly disperse and colonize alternate sites during the course of biofilm-associated infections.
How does acetate potentiate cell death? Our results reveal that both intracellular acidification and ROS generation may play a role in acetate dependent cell death. Although intracellular acidification can result from other fermentative metabolites like D- or L-lactate, we were unable to detect excretion of L-lactate and observed only small differences in the excretion of D-lactate in the ΔcidC and ΔalsSD mutants relative to WT (Fig. S8). Furthermore, given that the pKa of D-lactate (pKa = 3.86) is lower than acetate and its levels in culture supernatants were minute (∼35 fold less than acetate) we reasoned that it is unlikely to have a similar effect to that of acetic acid on cell death during aerobic growth. At the molecular level, it is not clear how acetate initiates ROS production. The evidence presented in this study suggests that acetate may contribute to a bottle-neck in electron transport by reducing the functionality of the respiratory chain. This could catalyze the promiscuous reduction of molecular oxygen and result in the production of ROS. Consistent with this argument, we have confirmed increased levels of superoxide and hydroxyl radicals following growth of S. aureus in 35 mM glucose, a condition that leads to acetate stress. Given that the pKa of superoxide anion is 4.88, it is very likely that this species freely traverses the cytoplasmic membrane and mediates oxidative damage during acetic acid stress. In addition to ROS, cytoplasmic acidification due to acetate influx itself can be a significant cause of cell death. Evidence for this conclusion arises from the observation that both metabolizable (acetate, lactate and pyruvate) and nonmetabolizable (benzoate) weak acids inhibit S. aureus growth. Such inhibition can result from acid catalyzed intracellular protein unfolding and aggregation. In support of this conclusion, a recent transcriptomic analysis of S. aureus challenged with the weak acid, lactate, revealed various clp genes (including clpB, clpC and clpP) involved in protein folding and recycling to be strongly up regulated [25].
Extending these findings to biofilms, we argue that the acidic pH microenvironments of biofilm microcolonies [26], [27] may spatially bias these biological structures as sites of respiratory inhibition and cell death. Weak acid metabolic byproducts like acetate and lactate are thought to accumulate within the biofilm interior, primarily due to diffusion limits resulting from reduced fluid flow and accumulation of biofilm matrix components [28]. We propose a model of staphylococcal PCD wherein the acidic pH within microcolonies activates expression of the CidR regulon [29] in a subpopulation of cells. This subsequently would lead to a feed-forward loop in which toxic acetate levels are reached through CidC activity. Ultimately cell death would ensue when macromolecular repair mechanisms are exhausted and cells within the biofilm are overwhelmed by the damaging effects of acetate (Fig. S9). To prevent a disproportionate number of cells from undergoing cell death, S. aureus co-expresses the AlsSD pathway along with CidC. Enzymatic activity of acetolactate synthase (AlsS) results in the condensation of two molecules of pyruvate to acetolactate and thereafter to acetoin by acetolactatae decarboxylase (AlsD). This effectively minimizes the carbon diverted to the generation of toxic acetate through the CidC pathway. More importantly evidence presented here also shows that the AlsSD pathway consumes protons from the cytoplasm and helps maintain pH homeostasis similar to the enterococci and lactobacilli [21], [30]. We contend that the ability of AlsSD to antagonize CidC activity effectively results in the modulation of intracellular pH and constitutes a robust mechanism to limit cell death and optimize biomass within the microenvironment of a microcolony.
The pathways that generate acetate vary among organisms. Whereas most bacteria, including E. coli and S. aureus generate acetate under aerobic conditions through the Pta-AckA and the CidC pathways, most eukaryotes (yeasts and mammals) lack these enzymes. Instead yeasts produce acetate as a natural byproduct of ethanol fermentation from acetaldehyde using acetaldehyde dehydrogenase [31]. Mammalian cells rarely generate significant quantities of acetate. But under certain conditions, acetate is produced by the enzymatic hydrolysis of acetyl-CoA in the cytoplasm [32]. Irrespective of the diversity in production routes, multiple studies have demonstrated that acetate itself can act as a potent inducer of PCD in yeasts and mammalian cells [33], [34]. Consistent with these studies, we demonstrate multiple hallmarks of eukaryotic PCD including respiratory dysfunction, generation of ROS and DNA fragmentation to be conserved during acetate mediated cell death in S. aureus. Further, similar to eukaryotic PCD [1], acetate-mediated cell death functions in a developmental context and appears to be crucial for optimal staphylococcal biofilm development.
It is noteworthy that two different activities of pyruvate oxidase (CidC, also annotated as PoxB in E. coli and SpxB in S. pneumoniae) have been described previously [35], [36]. In E. coli and S. aureus, this enzyme catalyses the decarboxylation of pyruvate into acetic acid and carbon dioxide, whereas acetyl phosphate and hydrogen peroxide are the predominant products of a similar reaction in L. plantarum and S. pneumoniae. Remarkably, both these reactions appear to induce stationary phase cell death in bacteria with either acetate or hydrogen peroxide as principal determinants of cell death [37]. Additionally, similar to S. aureus, cell death due to pyruvate oxidase activity is associated with apoptotic hallmarks in S. pneumoniae [37]. These observations appear to clearly mark pyruvate oxidase activity as a suicidal marker in bacteria.
Finally, what are the biological implications of regulating PCD? We used a well-established rabbit model of infective endocarditis to assess the effects of altering PCD on in vivo biofilm development. S. aureus injected intravenously in rabbits is rapidly cleared from the blood within the first 30 minutes leaving only minute residual amounts to linger over longer periods of time [38]. However any injury to the heart valves marks a preferred site for bacterial colonization and eventual development into a biofilm (vegetation). The pathological progression of infective endocarditis subsequently involves embolization of bacterial vegetations to alternate sites including surrounding heart tissues and other peripheral organs like the brain and kidneys [39]. This process not only poses a constant seeding source of infection but also hinders the normal functioning of peripheral organs and is often associated with a high degree of mortality [39]. Our investigations failed to reveal a colonization defect of the ΔcidC and ΔalsSD mutants on heart valves relative to the wild-type strain. However we observed a significant decrease in ΔcidC burdens in the blood, heart and kidneys after 48 h of infection. Although not conclusive, these findings strongly suggest that the ΔcidC mutant had lower rates of dissemination to secondary infection sites in vivo. Alternately, it is also possible that the ΔcidC mutant exhibits tissue specific fitness and survival defects in vivo. Either way, these findings suggest that alterations to the activity of cell-death associated metabolic pathways during biofilm development could affect staphylococcal pathogenesis.
In conclusion, the activation of pathways that generate metabolic acids from glucose during carbon-overflow and oxygen replete conditions have long been considered paradoxical in bacteria that are capable of undergoing oxidative phosphorylation, as it results in low energy yields, potentially toxic acid by-products and activation of cell death pathways [40]. Based on the current study we propose that the extracellular accumulation of metabolic acids is a developmental strategy that bacteria undertake to initiate cell death, a necessary precursor to optimal biofilm development. The initiation of staphylococcal cell death by intracellular acidification bears some striking resemblance to eukaryotic PCD. For instance, the dimerization and insertion of the pro-apoptotic modulator, Bax, into the membrane is thought to be triggered by intracellular acidification of eukaryotic cells [41] just prior to the release of cytochrome c into the cytoplasm. In this regard, it is possible that membrane oligomerization of CidAB and LrgAB (functional analogs of Bax and Bcl-2 in S. aureus, respectively) may also be initiated following intracellular acidification. Additionally, cytoplasmic acidification in eukaryotes also activates caspases, essential components of the apoptotic pathway [42]. Collectively, these observations are suggestive of a conserved role for glycolysis-mediated intracellular pH regulation in the modulation of PCD in eukaryotes and prokaryotes.
Animal experiments were conducted in compliance with a protocol (# 12-048-08-FC) approved by the Institutional Animal Care and Use Committee (IACUC). The University of Nebraska Medical Center is accredited by the Association of for Assessment and Accreditation of Laboratory Animal Care International (AALAC). In addition, all animals at the University of Nebraska Medical Center are maintained in accordance with the Animal Welfare Act and the DHHS “Guide for the Care and Use of Laboratory Animals.”
Strains and plasmids used in this study are listed in Table S2. The ΔcidCΔalsSD double mutant was created by moving the ΔcidC::erm allele from KB1058 into the ΔalsSD mutant (UAMS-1489) using bacteriophage Φ11-mediated transduction. In addition to growth on selective antibiotic media, the ΔcidC transductants were confirmed phenotypically and by PCR using the following primer pairs: cidC UP (5′-CACATGCATTTGGCACAGCT-3′) and cidC DN (5′-TGCTCATGCCTGCATTACCA-3′). The plasmid, pVCT2, containing the alsS gene with its native promoter was constructed by amplifying an approximately 2-kb region from the UAMS-1 genome using the primers, alsS-comp-F (5′-GATCGAGCTCTCCCTTATAATCACTCCCTTCA-3′) and alsS-comp-R (5′-AGTCTCTAGATGTGCCTAATGTACCATGTTG-3′), and inserting the resulting DNA fragment into the Sac1 and Xba1 sites of the shuttle vector, pLI50 [43]. Similarly, plasmid pVCT3 (containing cidC gene with its native promoter) was amplified from a ΔcidAB double mutant using primers, cidC comp-F (5′-GATCGAATTCACTCATTATTTGTGATTCCTCA-3′) and cidC comp-R (5′-AGTCGTCGACCAATTCAGTACAATCATTTGTG-3′). The resulting amplification product was inserted into the EcoR1 and Sal1 sites of pLI50. Subsequently, both pVCT2 and pVCT3 were transformed into RN4220 and transduced into the ΔalsSD and ΔcidC mutants respectively, using bacteriophage ϕ11 for phenotypic complementation.
E. coli cultures were grown in Luria Bertani (LB) broth. S. aureus cultures were grown in trypticase soy broth (TSB) supplemented with 35 mM glucose (unless specified otherwise). Bacterial cultures were aerobically grown at 37°C in either Erlenmeyer flasks fitted with bug stoppers to minimize evaporation during long-term growth or in 96-well flat, clear bottom micro-titer plates. For anaerobic growth, cultures were supplemented with cysteine (0.5 mg/ml) and 10 mM nitrate (or fumarate) and agitated at 250 rpm in an anaerobic hut. When necessary, antibiotics were added to cultures as follows: ampicillin (100 µg/ml); erythromycin (5 µg/ml); tetracycline (10 µg/ml); and chloramphenicol (10 µg/ml).
All analyses were performed using 1- and 3- day old stationary phase cultures of S. aureus on a BD LSRII flow cytometer (Beckton and Dickinson, San Jose, California). Cell samples were washed twice and diluted to a final concentration of 107 cells per ml in PBS. Cells were stained for 30 min with 5-cyano-2,3-ditolyl tetrazolium chloride (CTC, 5 mM) and 3-(p-hydroxyphenyl) fluorescein (HPF, 15 µM) followed by FACS analyses at a flow rate of ∼1000 cells per second. A total of 10000 events were collected for each sample. Bacteria were discriminated from background using a combination of forward scattered ligt (FSC) and side scattered light (SSC). Samples were excited at 488 nm using an argon laser and HPF emission was detected at 530±30 nm (with a 505 nm long-pass mirror) whereas CTC emission was detected at 695±40 nm (with a 685 nm long-pass mirror). Raw data were analyzed using the FlowJo software.
Quantitative assessment of DNA fragmentation was performed using the ApoDirect kit (BD bioscience). Samples were collected at the appropriate time points and fixed in 1% paraformaldehyde for 30 minutes. Cells were then washed twice in PBS, resuspended in 70% ethanol and stored at −20°C. Subsequent labeling of 3-OH ends of fragmented DNA was performed according to the manufacturer's instructions. Flow cytometry to detect TUNEL positive cells was performed as previously described [4].
Aliquots from stationary phase cultures (1- and 3 days) were withdrawn and resuspended to an OD600 of 10 units in 1 ml KDD buffer (Krebs-HEPES buffer, pH 7.4; 99 mM NaCl, 4.69 mM KCl, 2.5 mM CaCl2, 1.2 mM MgSO4, 25 mM NaHCO3, 1.03 mM KH2PO4, 5.6 mM D-glucose, 20 mM HEPES, 5 µM DETC and 25 µM deferoxamine). The resuspended culture aliquots were then incubated with 200 µM of cell-permeable ROS sensitive spin probe 1-hydroxy-3-methoxycarbonyl-2,2,5,5-tetramethylpyrrolidine (CMH; Noxygen Science Transfer and Diagnostics, Elzach, Germany) for 15 minutes at room temperature prior to analysis using a Bruker e-scan EPR spectrometer with the following settings: field sweep width, 60.0 gauss; microwave frequency, 9.75 kHz; microwave power, 21.90 mW; modulation amplitude, 2.37 gauss; conversion time, 10.24 ms; time constant, 40.96 ms. To identify the nature of ROS produced, cells resuspended in KDD buffer were incubated with either 400 units of superoxide dismutase (SOD; O2•− scavenger) or cell permeable dimethyl thiourea (20 mM DMTU; OH• scavenger) prior to the addition of CMH.
S. aureus was cultured at 37°C in TSB supplemented with either 14 or 35 mM glucose and aerated at 250 rpm with a flask-to-medium ratio of 10∶1 for 24 h. Cultures were subsequently diluted to an OD600 of 0.1 in fresh TSB (14 mM glucose). Oxygen consumption rates were determined for a period of 30 minutes at 37°C by using a MitoXpress oxygen-sensitive probe (Luxcel Biosciences) according to the manufacturer's instructions. The data were normalized to the corresponding OD600 units.
For these analyses, bacterial growth was allowed to proceed at 37°C and 200 rpm in BugStopper-sealed flasks containing TSB (35 mM glucose) in a 1∶10, flask to volume ratio. Preliminary experiments suggested that the assayed metabolite by-products were not significantly consumed following exhaustion of glucose from the media for up to 24 h. Therefore metabolite excretion profiles were determined from culture supernatants that were harvested at 24 h of growth. Acetate and glucose from culture supernatants were measured using commercial kits (R-Biopharm, Marshall, MI), according to the manufacturer's instructions. Acetoin was measured as previously described [44].
Overnight grown (16 to 18-hr) S. aureus cultures were resuspended to an OD600 of 0.06 in TSB (35 mM glucose). Bacterial suspensions were dispensed into 96-well microtiter plates and grown for 24 h at 37°C in a Tecan infinite 200 spectrophotometer under maximum aeration. The absorbance signals (OD600) were recorded every 30 minutes for the entire period of growth. For various experiments bacteria were challenged with the following compounds (final concentrations): acetic acid (30 mM), lactic acid (40 mM), pyruvic acid (30 mM) and acetoin (10 mM).
Intracellular pH was determined as previously described [30] with the minor modifications. Briefly, bacterial cells were grown in TSB (35 mM glucose) and 1 mL was harvested by centrifugation upon reaching an OD600 of 2. Cells were washed twice with an equal volume of 10 mM potassium phosphate buffer (pH 7) and resuspended in an equal volume of the same buffer. To load cells with the intracellular pH probe, 10 µl of 1 mM CFDA SE (5-(and 6)- carboxyfluorescein diacetate succinimidyl ester) was added to the suspension and incubated for 15 minutes at 30°C. Excess dye was removed by incubating the cells for 15 minutes at 30°C in potassium phosphate buffer (pH 7) containing 10 mM glucose. Cells were subsequently washed twice in the same buffer and finally resuspended in 50 mM potassium phosphate buffer (pH 4.5). Labeled cells were kept on ice until use.
To measure intracellular pH, 100 µl of the labeled cell suspensions were introduced into a 96-well flat bottom, black polystyrene plate (COSTAR 3916). Fluorescence was measured using a Tecan Infinite 200 spectrofluorimeter with excitation and emission wavelength set at 490 nm and 525 nm, respectively. Fluorescence emission units were converted to pH units using a standard calibration curve derived from labeled cells whose internal pH was equilibrated to the external pH (in citric acid buffers) ranging from pH 4 to 8, by the addition of 1 mM valinomycin and 1 mM nigericin.
Biofilms were grown in either FC280 or FC285 flow-cell systems (Biosurfaces Technology Inc, Bozeman, MT) and were analyzed by CLSM as described previously [45]. Briefly, biofilms stained with SYTO-9 (1.3 µM final concentration) and TOTO-3 (2 µM final concentration) fluorophores were excited with the 488 nm and 633 nm lasers respectively, and the emissions were collected using a 525±25 nm and 680±30 nm band-pass filter. For pictorial representation, the biofilms were imaged using an Achroplan 40×0.8 n.a. water dipping objective and for COMSTAT image analysis, images were acquired using a 20×1.2 n.a. dry objective to achieve a larger biofilm surface area for statistical purposes. Regions of interest within the biofilms were selected from similar areas within each flow-cell chamber and each confocal experiment was repeated a minimum of three times. Biofilm architecture was characterized using the COMSTAT software and measures of total biomass, average thickness, maximum height and roughness coefficients were determined [46]. Images were rendered using Imaris software (Bitplane, Saint Paul, MN).
Experimental endocarditis on the aortic valve of female New Zealand White rabbits (3 kg) were carried out as previously described with minor modifications [47]. Briefly, rabbits were anesthetized by intramuscular injection of ketamine hydrochloride (35–50 mg/kg), xylazine (2.5–6 mg/kg) and atropine (0.005–0.01 mg/kg) cocktail. An incision was made dextrolateral to the trachea and a polyethylene catheter (Becton Dickinson, MD) was then introduced into the left ventricle via the right carotid artery to produce sterile thrombotic endocarditis, and the skin incision sutured. To induce bacterial endocarditis, animals were intravenously challenged with 1 ml inocula (105 cfu/ml) via the marginal ear vein 24 h after catheterization. The animals were challenged with either the wild-type or mutant strains (ΔcidC and ΔalsSD) and subsequently (48 h post-infection) euthanized by lethal injection of a solution containing sodium pentobarbital (200 mg/kg). Bacterial loads from various tissue homogenates were determined by serial dilutions on THB agar plates.
|
10.1371/journal.pgen.1000330 | Evolution of Regulatory Sequences in 12 Drosophila Species | Characterization of the evolutionary constraints acting on cis-regulatory sequences is crucial to comparative genomics and provides key insights on the evolution of organismal diversity. We study the relationships among orthologous cis-regulatory modules (CRMs) in 12 Drosophila species, especially with respect to the evolution of transcription factor binding sites, and report statistical evidence in favor of key evolutionary hypotheses. Binding sites are found to have position-specific substitution rates. However, the selective forces at different positions of a site do not act independently, and the evidence suggests that constraints on sites are often based on their exact binding affinities. Binding site loss is seen to conform to a molecular clock hypothesis. The rate of site loss is transcription factor–specific and depends on the strength of binding and, in some cases, the presence of other binding sites in close proximity. Our analysis is based on a novel computational method for aligning orthologous CRMs on a tree, which rigorously accounts for alignment uncertainties and exploits binding site predictions through a unified probabilistic framework. Finally, we report weak purifying selection on short deletions, providing important clues about overall spatial constraints on CRMs. Our results present a complex picture of regulatory sequence evolution, with substantial plasticity that depends on a number of factors. The insights gained in this study will help us to understand the combinatorial control of gene regulation and how it evolves. They will pave the way for theoretical models that are cognizant of the important determinants of regulatory sequence evolution and will be critical in genome-wide identification of non-coding sequences under purifying or positive selection.
| The spatial–temporal expression pattern of a gene, which is crucial to its function, is controlled by cis-regulatory DNA sequences. Forming the basic units of regulatory sequences are transcription factor binding sites, often organized into larger modules that determine gene expression in response to combinatorial environmental signals. Understanding the conservation and change of regulatory sequences is critical to our knowledge of the unity as well as diversity of animal development and phenotypes. In this paper, we study the evolution of sequences involved in the regulation of body patterning in the Drosophila embryo. We find that mutations of nucleotides within a binding site are constrained by evolutionary forces to preserve the site's binding affinity to the cognate transcription factor. Functional binding sites are frequently destroyed during evolution and the rate of loss across evolutionary spans is roughly constant. We also find that the evolutionary fate of a site strongly depends on its context; a pair of interacting sites are more likely to survive mutational forces than isolated sites. Together, these findings provide new insights and pose new challenges to our understanding of cis-regulatory sequences and their evolution.
| Gene regulation is well recognized as a major determinant of how an organism functions [1], and is also gaining recognition as an important evolutionary substrate [2],[3]. Transcription control is one of the most common forms of gene regulation, and is known to be implemented through regulatory sequences often in the neighborhood of genes. Binding of transcription factors (TFs) to certain positions within regulatory sequences enhances or inhibits transcription and these bound sequences are called transcription factor binding sites (TFBSs). In the case that a gene has to be combinatorially regulated by multiple transcription factors, the cognate TFBSs of those regulating factors tend to be clustered together in ∼1 Kbp-length sequences called “cis-regulatory modules” (CRMs), or simply “modules” [4].
Despite significant recent efforts [5]–[8], we lack a good understanding of the organizational principles of CRMs, e.g., the requirements on strengths and arrangements of binding sites within a particular CRM. Inter-species comparison of modules provides a major opportunity to improve our understanding of such principles: (i) Evolution of CRM sequences is constrained by functional requirements, so the study of CRM evolution should allow us to infer which underlying features are more important, and to what extent. (ii) One may hope to find certain evolutionary signatures of CRM sequences through careful inter-species analysis [9], greatly facilitating the identification of yet unknown CRMs. (iii) The study of CRM evolution will also enable us to better understand the path “from DNA to diversity” [10].
Transcription factor binding sites are commonly predicted based on the assumption of their evolutionary conservation [11]. However, the exact nature of their conservation presents a complex picture. The study by Moses et al. [12] in yeast revealed that the rates of change of nucleotides of a TFBS depend on the binding profile of that TF–the positions of more specific protein-DNA binding permit lower rate of change. It should therefore be possible to leverage the position-specific substitution pattern to better predict TFBSs, as was done in [13]. This pattern has also been reported in bacteria [14] and vertebrates [15], but not in Drosophila. Given that this evolutionary pattern has already been assumed in practical analysis [16], it seems worthwhile to verify it in Drosophila. Moses et al. [13] further assumed that evolution of nucleotides at different positions are independent, and existing models of binding site evolution [17],[18] rely on this assumption; however, its validity is not obvious, given that a binding site typically functions as a unit. Empirical evidence either for or against this assumption has been lacking, except for a study in bacterial evolution [19] (where the evidence was against it). There is thus a clear need to test existing and new models of binding site evolution on the multi-species data from different phyla.
Even the most fundamental assumption of regulatory comparative genomics, that binding sites are evolutionarily conserved, has been challenged–Emberly et al. [20] found that binding sites are not substantially more conserved than their adjacent sequences in Drosophila; also, TFBSs are often found to have an unexpected amount of flux (gain or loss) in known CRM sequences [21]–[23] and in TF-bound regions in in vivo binding assays [24],[25]. It has been suggested that this flux is in part due to expression changes in the genes controlled by these sequences [24], and in part due to weak selection on individual sites even if the expression pattern of the target gene is conserved [26]. However, quantitative estimation of the strength of selection on binding sites has rarely been made, and requires extensive data on sets of orthologous binding sites. Moreover, the question of what leads to the observed levels of TFBS loss and gain is far from being resolved. For example, are the sites with higher binding affinities more likely to be conserved in evolution? How does the local context, i.e., the presence of other sites in the neighborhood, affect the probability of loss of a site? Does the loss probability correlate with overall selective pressure (substitution rate) of the CRM?
Cameron et al. [27] showed that insertions or deletions (“indels”) may be a powerful predictor of CRM sequences in sea urchin, as long indels were suppressed inside CRMs relative to their neighboring sequences. Lunter et al. [28] speculated that such a selection pattern may be particularly relevant to CRMs, as the “fitness” of these sequences may be sensitive to the length of the sequences between adjacent TFBSs, but not their exact nucleotide composition. In several earlier studies involving a number of well-studied CRMs in Drosophila, such a pattern has not been fully observed [23],[29]. So the following question remains: is indel-purifying selection in regulatory sequences a general evolutionary force, common to different organisms? The answer will affect our understanding of CRM organization; e.g., how tolerant a CRM sequence is to the change of spacing between TFBSs.
Earlier attempts to characterize the evolutionary patterns of regulatory sequences used a few well-studied CRM sequences. These studies were limited in their scope [21],[23],[29]. The availability of 12 Drosophila species [30] and a large collection of experimentally verified Drosophila CRM sequences [31] enable a large-scale and more systematic study of the evolutionary patterns of CRM sequences. Such studies also crucially depend on accurate computational tools for sequence comparison. Commonly used multiple alignment tools [32]–[34] that treat regulatory sequences as no different from other types of DNA (or for that matter amino acid) sequences are known to be a source of errors in evolutionary analysis [35],[36]. Even if the alignments are accurate, the step of annotating gaps as insertions or deletions (usually done by ad hoc parsimony criteria) may lead to inaccurate inferences [37]. We have previously developed new methods for inter-species sequence analysis, that are specially designed with the properties of regulatory sequences in mind. These include (i) Morph [38], which optimizes pair-wise sequence alignment by using the known binding profiles of relevant transcription factors, and (ii) Indelign [39], which uses a realistic probabilistic model of insertions and deletions to annotate “indel” events in a given multiple alignment. In this work, we take advantage of and extend these new methods to study the CRMs involved in Drosophila early development. This data set is ideally suited for such research because (i) the biological system is very well studied [8] and the relevant transcription factors are known, thereby limiting the false positives in binding site annotation, and (ii) much of the previous work on metazoan cis-regulatory evolution has been in this system [7],[23],[26]. Our study significantly extends the earlier work done on this dataset [40] and provides answers to many of the burning questions alluded to above.
We begin with our findings on the evolutionary behavior of transcription factor binding sites. We collected 68 D. melanogaster CRMs and seven TF motifs involved in the control of anterior-posterior segmentation in the blastoderm stage embryo. These CRMs (source: REDfly [31]) have been experimentally determined, without using evolutionary conservation for discovery, and are hence suitable for evolutionary studies without introducing ascertainment bias. Orthologous sequences of these CRMs were extracted from 11 other Drosophila species and were aligned by a special multiple alignment program, called “ProbconsMorph”. This is a new computational tool that we have developed, and is geared towards multiple alignments of regulatory modules in a TFBS-conscious manner (see Methods). It avoids propagating pair-wise alignment errors to the entire multiple alignment by combining the “consistency transformation” (see Methods) of Probcons [41] with posterior alignment probabilities obtained from Morph [38]. We also repeated most of our tests using the alignment tool “Pecan” [42] that does not use TF motifs, and we point out differences, if any, between results from the two types of alignment.
We annotated binding sites for each transcription factor, in the subset of D. melanogaster CRMs that overlap with ChIP-bound regions from Li et al. [43], if such data was available. Site prediction was based on the p-value of match to the respective PWM (“position weight matrix”) motif. We contrasted the density of these binding site predictions (in “bound” CRMs) with those in “unbound” intronic sequences, and typically found 2–3 fold enrichment in the former. (See Text S1, “False positive proportion estimation”.) We also predicted sites in each of the 11 other species separately, using the same method. Considering a binding site to be conserved if it is present in all other species in the D. melanogaster subgroup, we found that conserved sites were 2–3 fold enriched in CRMs than in intronic sequences. (See Text S1, “False positive proportion estimation”.) Our findings are consistent with earlier results in Li et al. [43], suggesting that the majority of predicted sites are likely to be functional.
Binding sites from different species, that overlap each other in the multiple alignment, are collectively referred to as an “orthologous TFBS set”. Sites in such an orthologous set were re-aligned locally in order to correct for any errors in their precise alignment. Graphic visualizations (Figure S1) of these 12-species CRM alignments, with binding site annotation, are available at our site http://europa.cs.uiuc.edu/TFBSevolution/.
Different positions in binding sites have different contributions to the binding affinity of the TF. Positions that form the core regions for TF-DNA binding are more specific (less variation allowed) in the motif, and should be under stronger selective constraints. We thus expect different positions of TFBSs to have different degrees of evolutionary conservation. The specificity of a position can be expressed by the information content (IC) of the corresponding column in the PWM (position weight matrix), and the evolutionary rate by the number of substitutions in that position in orthologous binding sites (see Methods). We observed highly significant negative correlations between specificity and evolutionary rate in five of seven TFs (i.e., all except Cad and Tll) (Table 1; Figure 1; Figure S2). Thus, our results confirm earlier similar findings in bacteria, yeast and vertebrates [12],[14],[15]. To avoid a bias introduced by the use of PWM-guided alignments, we used Pecan alignments (see Methods) of five closely related species for this particular analysis. The results were reproduced when using ProbconsMorph alignments (Table S1; Figure S3).
While substitution rates in a TFBS are position-specific, this does not imply that different positions evolve independently, although such an assumption is often made in existing evolutionary models [17],[18],[44]. It is easy to see that the exact same substitution can have drastically different effects on the functionality of a site, depending on how strong the site was to begin with. A site that is close to optimal will probably remain a site even if a crucial nucleotide is changed, thus this substitution is likely to be fixed. On the other hand, the same nucleotide change inside a weak site may have a larger functional consequence (the site loses its binding functionality), thus will be less likely to be fixed. It therefore seems plausible that the substitution rate of a position should depend on the entire site.
To study evolution at the level of binding sites, as opposed to nucleotides, we developed a simple mathematical model of binding site evolution, called “Site-level Selection” or “SS” model, that treats binding sites as single evolutionary units. Under this population genetics-based model, the fitness of a site can take two values, 1 if the binding affinity of this site is below some threshold, and if the affinity is above this threshold, for . (We use the same threshold as that used for defining a binding site.) The rate of substitution from site to , , is determined by the fitness difference between and according to this equation from population genetics theory [19],[45]:(1)where is effective population size, is the mutation rate of to , and is the fitness function defined above. When , we have ; when , i.e., there is a site gain, we apply the approximation that :(2)
When , i.e., there is a site loss, similarly we have:(3)
Note that and are inseparable in the above equations, so we will use the single quantity 4Ns as measuring the intensity of selection.
We tested how well this model fits the data on binding site evolution, and compared it to another model, called the “Halpern-Bruno” or “HB” model [17], which assumes positional independence and purifying selection at each position of the TFBS. The HB model has been used previously in cis-regulatory analyses (e.g., Moses et al. [13]). We considered predicted binding sites in D. melanogaster and their respective aligned sequences (whether designated binding site or not) in a closely-related species (D. yakuba), arbitrarily calling the former sites “ancestral” and the latter sites “descendant”. Assigning an “energy score” to each binding site based on its similarity to the PWM [46], we calculated the difference in energy score between the ancestral and descendant sites, and used this as the statistic to represent binding site evolution. We computed, for each TF, the histogram of this “energy difference” statistic, and asked how well this histogram fits theoretical predictions from simulations using either the SS or the HB model (Table 2). For every motif, the SS model showed a significantly better fit to the data than the HB model. (Table 2; Figure 2A; Figure S4). (See Methods for details of how statistical significance was estimated, while accounting for the additional free parameter in the SS model. The results were reproduced when using Pecan alignments; see Table S2 and Figure S5.) Our estimated level of selection (4Ns in the range 8–19) is consistent with an early estimate from bacterial regulatory sequences [19] and our results argue in favor of models treating entire binding sites as evolutionary units.
However, in absolute terms, neither model explains the data very well (Figure 2A; Figure S4), and there is a greater amount of conservation (energy differences close to zero) in the observed data than predicted even with strong selection. A similar analysis was performed with the evolutionary statistic being the number of substitutions between ancestral and descendant sites, and we found that there is an excessive number of fully conserved sites (no substitutions) than expected under either the HB or the SS model (Binomial test, p-value<10−12) (Figure 2B; Figure S6; Figure S7 with Pecan alignments). This seems to indicate that for many sites, the allowed binding affinities fall in some narrower range, instead of being determined by a single threshold (lower bound). It has been suggested that in order to produce the correct expression pattern, a binding site may prefer some specific affinity level, and both stronger and weaker binding tend to be less functionally optimal [47]. Our results provide support for this hypothesis.
Even though TFBS loss and gain (henceforth called “turnover”) have been commonly observed, it is not clear whether these changes are adaptive [48] or not [26]. If adaptive selection is the main force behind binding site turnover, it is likely that the process will show a lineage-specific pattern; on the other hand, a molecular clock has been known to be suggestive of the absence of adaptive selection, as per the neutral theory of evolution [49]. We considered the fraction of binding sites in D. melanogaster that have an ortholog (above threshold) in a second species, and plotted this fraction as a function of evolutionary divergence from the second species (Figure 3; Figure S8). For all transcription factors, the fraction of shared binding sites decreases linearly (R2>0.90, Table 3) as the divergence time increases, a clear sign of a molecular clock. One problem that may confound the analysis is the presence of false positive binding sites predictions, which are expected to follow a molecular clock. To examine this effect, we calculated a correction term in the fraction of conserved sites, and regressed this with divergence time, using the false positive proportion as a free parameter. High values of the adjusted R2 were obtained (Table 3), confirming the presence of the molecular clock. We repeated the exercise with sites for randomly created PWMs, and found a similar linear relationship. The rate of loss (negative slope of the line) for these random sites is higher than the rates for binding sites, for six of the seven transcription factors (Table 4), the difference being significant for Bcd and Kr. We note that the sites predicted by random PWMs do not represent neutral sequences, but reflect the average constraint in CRM sequences. This has been shown previously in [43]. The results were reproduced when using Pecan alignments (Table S3 and S4; Figure S9).
Having characterized some general patterns of TFBS evolution, in this section we study what specific factors may influence the conservation and turnover of binding sites.
Finally, we analyzed insertions and deletions in known regulatory sequences, to study the extent of indel-purifying selection. Among 370 non-overlapping D. melanogaster CRMs from the REDfly database [31], we chose 128 CRMs that have clear orthologous sequences in D. simulans, D. yakuba, and D. erecta. This choice of species was dictated by simulation-based assessment of the limits of our indel annotation capability (see Methods). Because insertions and deletions (indels) may have different functional consequences on CRMs, we treat them differently. We estimated the number of short insertions and deletions in CRMs using Pecan [42] for alignment and Indelign [39] to annotate the indels. For each CRM, the insertion or deletion count was defined as the average of the respective counts in the four species, weighted by the branch length. We compared indel frequencies in CRMs to those in “background sequences”, chosen to be the regions flanking the CRMs. We found that (i) the number of short deletions (less than 20 bp in length) in CRMs is significantly smaller than that in background regions (paired Wilcoxon signed-rank test, p-value 0.0074; 1970 in CRMs and 2183 in length-matched background regions) and (ii) there was no statistically significant difference (p-value 0.5464) in the number of insertions (1932 in CRMs and 1870 in background). The number of long indel events (20 bp or longer) in our data set was relatively small (CRM: 107 insertions and 175 deletions, background: 115 insertions and 178 deletions) and no significant difference was observed in this regard between CRMs and background regions.
Another related question is the indel pattern in the “spacer” region between CRMs and transcription start site (TSS) of the target genes. Transcriptional regulation depends on the communication between CRMs and promoter sequences [54], which may pose some requirements on the length of the spacer sequences. We thus repeated the above analysis on these spacer regions. (We only consider 63 upstream CRMs in this experiment.) No significant differences in frequencies of insertions or deletions were observed between these regions and background sequences (data not shown).
Our results show that indel-purifying selection exists on CRM sequences, but such selection acts most strongly on deletions. We did not find clear suppression of long-indels, as has been observed before [27].
The study of cis-regulatory evolutionary patterns has provided important insights on regulatory sequence function [23],[55], and proves valuable for prediction of these sequences in genomes [9],[56]. Yet, our understanding of cis-regulatory evolution is limited at best. While we have theories as well as a large volume of empirical data on protein evolution, we essentially have no theory and have made limited observations on the evolution of regulatory sequences. Our goal here is to begin to bridge the gulf between the vast amount of genomic sequence data and our poor understanding of regulatory sequences and their evolution. We have conducted a detailed evolutionary analysis of a large collection of experimentally verified CRM sequences, taking advantage of the recently sequenced 12 Drosophila genomes. Our analysis has revealed several interesting patterns, some along expected lines (but not confirmed previously), and some contrary to our expectations. We believe that our work will furnish evidence orthogonal to experimental characterization for understanding the organizational principles of CRMs, and will be important for developing a theory of regulatory evolution in the future.
There are several technical issues that were important to address in our analysis. Evolutionary comparison depends on the alignment of orthologous sequences, but in general, alignments cannot be perfectly determined and may be a source of biased conclusion [36]. This may be a particularly serious problem for the analysis using 12 Drosophila species because of the relatively large divergence. We addressed this concern by developing a new multiple alignment program tailor-made for regulatory sequences. It combines the power of a pair-wise regulatory sequence alignment tool, Morph [38], and a probabilistic multiple alignment framework Probcons [41]. We have made this new software (ProbconsMorph) available freely for public use, to facilitate future studies of this genre. Nevertheless, the use of motifs to construct alignment may artificially boost the conservation level of TFBSs. We carefully addressed this potential bias whenever it may affect our conclusion. For example, when testing the positional variation of substitution rates, we use Pecan-based alignments without using motifs and limited ourselves to five closely related species. Similarly, when testing the correlation of binding site strength to turnover rates, we use randomized PWMs (as “negative controls”) to validate our finding. We also repeated all our analyses with Pecan-based alignments. The various trends seen in Results were almost always reproduced. One notable difference was that the correlation between nearest homotypic site distance and evolutionary rate (Table 6) for Cad was statistically significant (p-value 0.02) in ProbconsMorph alignments, but insignificant (p-value 0.15) in Pecan alignments. We suspect that this may be due to the tendency of standard alignment tools (such as Pecan) to misalign one or two nucleotides at the boundary of binding sites, especially if the motif contains short repeats such as TTTT [38], as is the case for Cad.
Another critical component of our analysis is the prediction of TFBSs. By using the same PWMs for all the genomes, we have made the assumption that the PWM of any TF is fully conserved across 12 Drosophila genomes. This is questionable, as researchers have found in yeast that the change of TF binding specificities can be an important part of the evolutionary change of regulatory networks [57]. For the seven motifs we analyzed, however, there is prior computational evidence that the binding specificities have not changed between D. melanogaster and D. pseudoobscura [58]. Another issue related to TFBS prediction is that predicted binding sites tend to have a high proportion of false positives [59]. We believe this problem is mitigated by our focus on the segmentation network, the fact that we restrict ourselves to transcription factors and CRMs experimentally known to be involved in regulating the segmentation genes, and our use of ChIP-based binding information wherever possible. We also believe that within a CRM, any computationally predicted binding site for a relevant transcription factor can “attract” transcription factor molecules, and contribute to the expression pattern, and should thus be considered “functional” in a broad sense. The results from Janssens et al. [60] seem to support this point. In practice, we may still have a small number of false predictions because of inaccuracies of the PWMs and we have attempted to estimate the false positive proportion by various methods (see Text S1). Also note that while false site predictions may obscure the evolutionary pattern of functional binding sites, they will not, in general, introduce spurious patterns (since, by definition, these sites are not under selection). In cases where the false sites may affect our interpretation of results, for example, in the test of molecular clock for binding site turnover, we have tried to make appropriate corrections. In addition, in estimating TFBS turnover rates, we have emphasized on losses rather than gains of sites, because a predicted TFBS loss event has stronger supporting evidence than a gain event (the “gained” site is more likely to be a false positive prediction).
Our model of binding site evolution, the “Site-level Selection” (SS) model, is a special case of the population genetic model proposed by Mustonen and Lassig [19]. Under their model, the fitness of a site is determined by its binding energy. The difference of the energy distribution of known sites and of the neutral sites allows one to estimate the fitness of any energy value. A binding site evolves in the space of all possible sequences, with the transition rate between any two sequences determined by the fitness values of the two sequences, given by Equation (1). For most known TFs, however, the number of known sites is too small to reliably estimate a fitness function and the simplification introduced in our model is probably necessary. Our SS model is also similar to the model in Raijman et al [61]. Under this model, a site always tends to preserve its current functional status, that is, the substitution in a binding site that makes is nonfunctional will have a lower rate, and similarly, a substitution that creates a functional site in an originally neutral site will also have a lower rate. However, their model is not formulated in population genetic terms and the transition from a non-site to site is always selected against (this will be favored under the Mustonen-Lassig model and ours). We found that the SS model better explains the evolutionary pattern of binding sites than the HB model, which assumes the independence of substitutions at different positions of a site. A recent study [62] also reported this dependence of binding site positions, though without directly comparing two kinds of models. Admittedly, the presence of false sites may complicate our analysis. It is difficult to directly address this issue, say, through a mixture model approach as done in [22] because of the difficulty of computing probabilities under the SS model. However, we note that if we were to remove false sites from the observed data, we would see a greater proportion of conserved sites, implying that the SS model will continue to be closer to the observation than the HB model (see Figure 2). Next, we observe an overrepresentation of fully conserved sites (no mutations) compared to what is expected from both SS and HB models (Figure 2B). This argues for the conservation of precise affinities, a hypothesis consistent with our current knowledge about the dependence of expression pattern on precise binding affinities [63],[64], though this phenomenon has not been statistically observed previously. Finally, we note that the findings of position-specific substitution rates and site-level selection are not contradictory; as pointed out in [19], each position of the site contributes separately to the fitness of the site, which depends on the sum-total of these contributions.
Our findings of a molecular clock extend earlier results on a small number of well characterized CRMs [65] across three Drosophila species, suggesting that this is a property common to developmental CRMs across a large evolutionary range. Even though we cannot exclude the presence of adaptive selection in individual cases, our results seem to suggest that negative selection to maintain the existing binding sites is the dominant mode of evolution, coupled with the occasional loss of sites due to random drift. The rate of site loss likely reflects the strength of purifying selection.
Our tests point out that stronger binding sites are conserved more often than weaker sites. This is consistent with an earlier study [66], which found that stronger Dorsal binding sites were more likely to reside in conserved blocks. A simple explanation for this is that stronger sites are more likely to be important to CRM function, thus under stronger constraint. An alternative explanation is that there is a “quality” threshold that defines functionality and once a site drops below that threshold, it is impervious to selective forces. Assuming this is true, we note that a weaker site is closer to the threshold than a stronger site, and may thus be lost more easily. A recent paper [61] seems to support the latter hypothesis. It is likely that the forces of natural selection as well as those of mutation/random drift together determine the evolutionary fate of a binding site, as suggested by Mustonen and Lassig [19].
An unexpected result of our analyses is that the degree of homotypic clustering does not affect turnover rate. This is contrary to the notion that more binding sites of the same type will lead to greater redundancy, easing the selective pressure on the individual sites. Instead, the number of binding sites seems to be important to CRM function. This observation is similar to one of the implications of our findings of site-level selection: that exact affinities of binding sites are functionally important. Both observations are consistent with the so called “gradient threshold model” [47], which suggests that different genes may respond to different concentration levels of the same TF by harnessing different numbers of binding sites with varying affinities. The exact binding affinities and number of sites are important under this model. In a more detailed analysis of homotypic clustering, now considering the binding site arrangement, we observed that for some factors, if a site is adjacent to another site of the same factor, this site will be less likely to be lost during evolution. This may be indicative of cooperative activity of proximal homotypic binding sites, leading to stronger selective pressure. For instance, the significant result (p-value 0.0184, Table 6) for Cad is consistent with anecdotal evidence of Cad sites being located as proximal pairs [67]–[69], although we are not aware of any biochemical evidence for such cooperativity. There is also some evidence in the literature for DNA binding by homodimers of Tll [70] and Hb [71]. Our observation also suggests that sites that have a proximal “partner” are perhaps less likely to be spurious sites, which will provide a useful additional guideline to binding site prediction [72]. Surprisingly, we did not observe significant result for Bcd, even though it is known to bind cooperatively [52]. This negative result is a reminder that the sensitivity of our statistical tests may be reduced due to a variety of factors, e.g., alignment errors, false sites, etc. These factors are unlikely, however, to produce spurious statistical signals.
We found that the presence of a binding site for a different factor, either overlapping or proximal to a binding site, can strongly affect the latter's evolution. Different mechanisms of local interactions between sites are known in developmental CRMs, e.g., cooperative binding between two factors [73],[74], short-range quenching [75],[76], competitive binding to overlapping sites [74], etc. In all these cases, the loss of a single binding site may disrupt the interaction and create a larger change of expression than if the binding sites act in an additive fashion. As a consequence, these locally interacting site pairs may be under stronger selection. Our results support the importance of context in determining evolutionary fate of binding sites. A recent paper reports similar results for four CRMs of the even-skipped gene [53]. By working on a much larger set of CRMs, we confirm this context-dependence as a general evolutionary pattern. We also found some interesting specific cases, for example, the Kr sites that overlap with another TF site, appear more conserved, consistent with the known role of Kr as a repressor with the ability of competitive binding. In addition, the difference of the evolutionary patterns of the seven TFs suggests that they may depend on different mechanisms for their function. For example, both Kr and Tll are repressors, but Tll is more conserved if it is adjacent to some other site, while Kr is more conserved if it overlaps with another site. This seems to suggest that the relative importance of competitive binding and short-range quenching may be different in Kr and Tll.
We did not find strong evidence of suppression of large indels within CRMs relative to their flanking sequences. Our results are different from an earlier study of indel patterns of CRMs in sea urchins, which reports that large indels (>20 bp in length) are virtually absent inside CRM sequences [27]. There is an alternative explanation for this discrepancy: it has been known that Drosophila has a very compact genome as the neutral deletion rate is very high [77] and a large fraction (40–50% from different estimates) of intergenic non-coding sequences is under evolutionary constraint [48],[78]. Consequently, the flanking sequences of CRMs may not be entirely neutral, and the distinction between CRM and flanking sequences may not be as pronounced as in other species. (Our options were limited with respect to the “background” sequence to contrast with, since long repeats often used as neutral sequence in mammalian genomes [79] are rare in Drosophila.) The fact that short deletions are more constrained than short insertions is likely due to different effects of insertions and deletions on CRM sequences: any deletions that extend to an existing binding site will annul its functionality, while insertions, unless occurring exactly inside TFBSs, will only change the distance between sites, but not destroy them. In Text S1, we outline an illustrative calculation, suggesting that under simple but reasonable assumptions, short deletions are maybe twice as more likely to interfere with a binding site than are short insertions. These results combined with the lack of strong constraint on spacer sequences suggest that CRM structure is overall flexible, permits relatively quick evolutionary change, and functions without being very sensitive to the precise distances between binding sites. In terms of its implications for bioinformatics, our results seem to indicate that the indel signature can be a useful CRM predictor but not strong enough to work alone, somewhat contrary to prior expectations [27],[28].
12 Drosophila genome sequences from D. ananassae (Feb. 2006 assembly), D. erecta (Feb. 2006 assembly), D. grimshawi (Feb. 2006 assembly), D. melanogaster (Apr. 2006 assembly, release 5), D. mojavensis (Feb. 2006 assembly), D. persimilis (Oct. 2005 assembly), D. pseudoobscura (Feb. 2006 assembly), D. sechellia (Oct. 2005 assembly), D. simulans (Apr. 2005 assembly), D. virilis (Feb. 2006 assembly), D. willistoni (Feb. 2006 assembly), and D. simulans (Nov. 2005 assembly) were compiled from UCSC Genome Browser database [80]. To predict the positions of putative TFBSs, position weight matrices (PWMs) for seven TFs, Bcd (Bicoid), Cad (Caudal), Dstat, Hb (Hunchback), Kni (Knirps), Kr (Kruppel), and Tll (Tailless) were compiled from FlyReg [81] and the literature. We used the phylogenetic tree and branch lengths for the 12 species in [82] and for the four species (D. melanogaster, D. simulans, D. yakuba, and D. erecta) in [25]. Orthologous sequences of each D. melanogaster CRM were obtained by the liftOver program from the UCSC Genome Browser database. The background region corresponding to a CRM was defined as the region upstream of the farthest known CRM of its target gene, equal in length to its corresponding CRM.
For the analysis of TFBS evolution, we developed a new multiple alignment program, “ProbconsMorph”, by integrating Probcons [41], a consistency based multiple sequence alignment program, and Morph [38], a pair-wise sequence alignment program that is specially designed to align regulatory modules. Morph uses a pair-HMM as a generative model for alignment of two orthologous CRMs, and is parameterized by the given motifs, as well as various evolutionary rate parameters that it fits to the data. It uses maximum likelihood inference to simultaneously perform TFBS annotation and alignment. It reports for every pair of positions in the two sequences, the posterior probability that they are aligned. Morph was run to produce such a probabilistic alignment of every pair of species. Probcons takes such pair-wise alignment probabilities and builds a multiple sequence alignment progressively, while using the “consistency transformation”: the probability of alignment of two nucleotides and is updated based on the alignment probabilities of and and of and , where is a nucleotide from a third species. We have shown previously that Morph provides practical benefits for inference of evolutionary events and rates by computing a better alignment; ProbconsMorph is an effective and efficient extension of this program to more than two species. We made two simple modifications to Probcons to integrate it with Morph: firstly, Probcons was made to work on DNA sequences (the current implementation handles protein sequences only), and secondly, it was made to accept a phylogenetic tree as input, rather than estimate the tree at run-time. The ProbconsMorph software is publicly available at our site http://europa.cs.uiuc.edu/TFBSevolution/.
Pecan [42] was used for the alignment of four species in the analysis of indels in CRMs and spacers. We have performed extensive studies on simulated data to determine the limits of indel annotation, and estimated that accurate labeling of insertions and deletions is only possible for the four closely related species D. melanogaster, D. simulans, D. yakuba, and D. erecta. (Kim and Sinha, in preparation.) Pecan alignments of these four species, and D. sechellia, were also used for the study of position-specific substitution rates in binding sites (Table 1).
Insertion and deletion annotations were done using our previously published Indelign program [39] that is based on a probabilistic model of indels and annotates indels as being insertions or deletions based on maximum likelihood. We used a mixture of two geometric distributions as a model of the length distribution of indels. As shown in Figure S10, this mixture model is a much better fit to the indel length distributions empirically observed in D. melanogaster CRMs used in this study and their orthologous sequences in D. simulans, D. yakuba, and D. erecta. The new version of the Indelign program is available at our site http://europa.cs.uiuc.edu/TFBSevolution/.
|
10.1371/journal.pntd.0000271 | Landscape Diversity Related to Buruli Ulcer Disease in Côte d'Ivoire | Buruli ulcer disease (BU), due to the bacteria Mycobacterium ulcerans, represents an important and emerging public health problem, especially in many African countries. Few elements are known nowadays about the routes of transmission of this environmental bacterium to the human population.
In this study, we have investigated the relationships between the incidence of BU in Côte d'Ivoire, western Africa, and a group of environmental variables. These environmental variables concern vegetation, crops (rice and banana), dams, and lakes. Using a geographical information system and multivariate analyses, we show a link between cases of BU and different environmental factors for the first time on a country-wide scale. As a result, irrigated rice field cultures areas, and, to a lesser extent, banana fields as well as areas in the vicinity of dams used for irrigation and aquaculture purposes, represent high-risk zones for the human population to contract BU in Côte d'Ivoire. This is much more relevant in the central part of the country.
As already suspected by several case-control studies in different African countries, we strengthen in this work the identification of high-risk areas of BU on a national spatial scale. This first study should now be followed by many others in other countries and at a multi-year temporal scale. This goal implies a strong improvement in data collection and sharing in order to achieve to a global picture of the environmental conditions that drive BU emergence and persistence in human populations.
| Buruli ulcer (BU) is one of the most neglected but treatable tropical diseases. The causative organism, Mycobacterium ulcerans, is from the family of bacteria that causes tuberculosis and leprosy. This severe skin disease leads to long-term functional disability if not treated. BU has been reported in over 30 countries mainly with tropical and subtropical climates, but Côte d'Ivoire is one of the most affected countries. M. ulcerans is an environmental bacterium and its mode of transmission to humans is still unclear, such that the disease is often referred to as the “mysterious disease” or the “new leprosy”. Here, we explored the relationship between environmental and socioeconomic factors and BU cases on a nationwide scale. We found that irrigated rice field cultures areas, and, to a lesser extent, banana fields as well as areas in the vicinity of dams used for irrigation and aquaculture purposes, represent high risk zones for the human population to contract BU in Côte d'Ivoire. This work identifies high-risk areas for BU in Côte d'Ivoire and deserves to be extended to different countries. We need now to obtain a global vision and understanding of the route of transmission of M. ulcerans to humans in order to better implement control strategies.
| Buruli ulcer is a severe human skin disease caused by Mycobacterium ulcerans. It represents now the third mycobacterial infection in the world behind tuberculosis due to M. tuberculosis and leprosy, caused by M. leprae. The first clinical description of the disease agent was done in Australia in 1958 [1], even though disease cases have been recorded since the end of the XIXth century in Uganda, in the Buruli area [2]. During the last decades, a dramatic extension of the spatial distribution of Buruli ulcer disease as well as increase of the number of infected people has been reported in many parts of the world. Highest incidences are now observed in Western Africa with 20,000, 6,000 and 4,000 cases observed in 2005 in Côte d'Ivoire, Ghana and Benin, respectively [1],[3],[4].
Mycobacterium ulcerans is an environmental bacterium and its mode of transmission to humans is still unclear, this is why the disease is often referred to as the “mysterious disease” or the “new leprosy”. Recent findings on the life cycle of the Buruli ulcer's agent have enhanced current evidence on several points. First, it has been shown that M. ulcerans can develop as biofilms on the surface of aquatic plants [5]. More specifically, some freshwater aquatic plants might be involved in the mycobacterium life-cycle as potential intermediate hosts or trophic chain concentrators [5]–[7]. Secondly, contrasted animal species were found infected by the bacterium in natural conditions, e.g. fishes, frogs [8]–[9] or koalas (see [10] for review). Recently, an impressive field study also generated additional environmental data regarding M. ulcerans in nature [11]. Third, aquatic insects are also suspected to act as vector and to transmit the disease by biting. It has been demonstrated that M. ulcerans is present in salivary glands of African water bugs of the family Naucoridae, and that infected water bugs could transfer the pathogen to mice [6],[12],[13]. Infected insects were also found in endemic areas. Finally, it is generally admitted by medical and scientific communities that specific environmental niches, which still need to be precisely defined, favour the occurrence of the disease [11], [14]–[18]. Based on several case-control studies performed in different African countries, freshwater ecosystems like rivers, man-made ponds and lakes, or marshy zones and irrigated perimeters represent risk factors to BU [19]–[21]. At nation-wide or regional scale, other studies also showed relations between BU infections and different environmental factors [22]–[24], as for instance landscape cover attributes [25] or arsenic in water [26] in Ghana. Despite these specific studies, Buruli ulcer is still a mysterious disease and all new findings will contribute to help national and international public health authorities to fight this highly deleterious pathogen and neglected disease. For this reason, all information on the relations between the environment and the disease occurrence are highly relevant for a better understanding of the disease as a whole.
Here we propose to perform a first nation-wide scale study in Côte d'Ivoire of the link between environmental but also socio-economic factors and Buruli ulcer cases, based on spatial mapping and multivariate statistical analyses. First detection of BU in Côte d'Ivoire occurred in 1981 but the number of cases clearly increased in 1987 and became then a national public health problem [20].
The Buruli ulcer notifications database was generated by the Institute Raoul Follereau in Côte d'Ivoire. All notified cases were diagnosed by medical doctors from the institution based mainly on clinical signs and sometimes confirmed also by tissue biopsy or ulcer swab examination. Data correspond to annual cases reports from primary healthcare centers and central hospitals distributed all over the national territory in 1997. The central hospitals serve towns and cities with more than 5,000 inhabitants. In Côte d'Ivoire, up to 54 percent of the human population lives in a town with a primary health care center or hospital; and 91 percent lives less than 15 km from such a town [27]. Permission to use these data for the present study has been granted by the National Program for Buruli ulcer Control from the Ministry of Health and Public Hygiene of Côte d'Ivoire (Pr H. Asse).
Environmental data correspond to vegetation types and surfaces (forest or cultivated areas) and dams (location and superficies). Data concerning vegetation were extracted from two sources. To provide an overview of the global vegetation zones in the country, we first generated a map (Figure 1A) based on vegetation map by Guillaumet [28] which presents the principal zones of vegetation based on 1979 aerial photographs (50 centimeter resolution). Secondly, in order to obtain the most realistic information about forest patches corresponding to the BU reported period, we used the map of forest reservation in Côte d'Ivoire from the National Bureau of Technical and Development Studies [29] representing the recent state of the forest cover, based on 1993 and updated with 2000 SPOT images (20 meters resolution) to build a second map representing the current state of forest distribution in the country (Figure 1B). This recent map also enabled us to compute detailed geometrical characteristics (distances and areas) used in statistical analyses.
Côte d'Ivoire is characterized by some well-known contrasting ecosystems (Figure 1A): areas of wet dense forest and bush extending roughly South of the 8°N latitude line, and dry forest and Sudanian savanna ecosystems in the Northern region of the country. The origin of this division is primarily climatic, since the transition from wet dense forest to savanna-like ecosystems is associated with the change from four climatic equatorial seasons in the South to two tropical seasons in the North [28].
Data concerning dams were retrieved at the Côte d'Ivoire National Centre for mapping and remote-sensing, at the University of Abidjan-Cocody, by one of the authors (TB). They correspond to the dams' geographical position and total surface area for the year 1997. There are over 500 dams in Côte d'Ivoire and the majority of them were built between 1970 and 1990 in order to allow the cultivation of water-demanding crops, especially in the Central region and more specifically in the Northern region, where rainfall is not sufficient to sustain this kind of production. These water bodies have facilitated the establishment of irrigated rice cultures, and they are also a source of freshwater for cattle and fish farms.
Regarding socio-economic parameters, agricultural data were provided by the Côte d'Ivoire Ministry of Agriculture. They concerned i) the type of cultivated areas, and ii) the crop production (in tons) for the year 1997 for each of the 54 prefectures. For the present study, we only considered rice and banana fields, which require a constant high humidity and intensive water supply. Thus we did not consider crops which do not require irrigation, such as cassava ground nut and corn.
Epidemiological data correspond to cases detected at hospital or primary healthcare centers situated in each prefecture (Figure 2). Raw disease case data (Figure 3A) were transformed to numbers of cases per 100,000 inhabitants, considering the whole population of the prefecture of each 202 hospitals or health care centers. Moreover, it is known that each inhabitant lives in an average 10 km distance from a hospital in Cote d'Ivoire [27]. For this reason, we interpolated our epidemiological data on a regular 10 km×10 km grid-square by the inverse distance weighting method, using the Surfer software [30],[31] (Figure 3B). In order to illustrate environmental data, we generated maps representing both epidemiological and environment data by interpolating environment point data (i.e. dams and rice crops data) on the same regular grid (Figures 1 and 4). However, we did not use these interpolated data but the raw data for the following statistical analysis.
Surface data concerning forest area and vegetation zones were not transformed. The surface areas for the different types of ecosystems studied, e.g., evergreen forest patches and rice-field cultures were transformed according to the percentage of surface coverage on the total prefecture surface area size in km-square.
We used Geographical Information System (GIS) to explore the relationships between environmental and epidemiological data [32]. In this analysis, we used the original data instead of the interpolated data. We first arbitrarily defined 4 influence-by-distance classes (i.e., <5 km, 5–10 km, 10–15 km, 15–20 km) for Buruli ulcer disease occurrence, which corresponded to different distances between a given case and the nearest hydro-agricultural dam or the nearest forest patch respectively. Based on this distribution, we analysed the variation of the rate of disease incidence according to the different distances for the two environmental parameters under scrutiny, i.e., dams or forest patches. We thus considered here the location of dams and forest patches but not their surface. Concerning dams, we also made a regional study distinguishing three major hydrological regions: the Kossou region in the Centre of country, the Buyo region in the West, and the Southern region. Linear trends were tested by using simple linear model.
Finally, we used logistic regression [33] to analyse the statistical relationships between hydro-agricultural environmental conditions and incidence of Buruli ulcer disease by administrative division. The explicative variables considered were: forest surface area (in hectare), irrigated rice production (in thousands of tons), banana production (in thousands of tons), and dam surface area (in hectare). To make the analysis more complete, i.e., to better capture the multivariate dimension of the environmental conditions on disease emergence, we integrated the two-factors interaction terms.
Figure 3 illustrates Buruli ulcer cases and the aggregated distribution of incidences on a 10 km grid-square, in Côte d'Ivoire for the year 1997. Southern Côte d'Ivoire, rich in evergreen forest and forest/savanna-like landscapes (Figure 1A), had a very high Buruli ulcer incidence (more than 5,000 infected people for 100,000 inhabitants), whereas we observed a very low incidence in the North (less than 10 infected people for 100,000 inhabitants). In particular, the highest incidence (more than 10,000 cases for 100,000 inhabitants) was near the contact zone between forest and savanna, notably in proximity of the “V-Baoulé” region in the centre of Côte d'Ivoire. Figure 1B also illustrates the spatial relationship between Buruli ulcer cases and evergreen forest patches.
Figure 4A shows that the rice field areas exhibited the highest rate of disease notifications, notably near the large lakes in the Centre of the country, an area which provides nearly 40 per cent of the national rice production [34]. Together with the regions of high rice production, highest incidence spots of Buruli ulcer disease notification were mainly located in areas with a high density of dams (Figure 4B). It was particularly true for the area of the great lakes, in the Centre of the country, which hosts more than a half of the country's hydro-agricultural equipments, with an average reservoir surface area of 2 sq-km. When considering the geographical distance between Buruli ulcer disease cases and distance to dams or forest patches, we found significant negative relationships (Figure 5). In the Kossou region, located in the central part of Côte d'Ivoire, Buruli ulcer disease incidence was up to 100 notifications per 100,000 inhabitants within a 5 km-distance radius from the closest reservoir, but it dropped to 40 cases for a 5–10 km-distance, and to less than 20 cases for a 20 km-distance (r2 = 0.86, P<0.01; Figure 5A). Concerning the Buyo region in the West, and the whole Southern region, the relationship between Buruli ulcer incidence and distance to dams was also significant (r2 = 0.84, P<0.01; and r2 = 0.71, P<0.01, respectively; Figure 5). We observed the same negative relationship between disease incidence and increasing distance from a primary forest patch (r2 = 0.69, P<0.01; Figure 5B).
Table 1 illustrates the main results from the logistic regression analysis performed to explain Buruli ulcer disease incidence. This analysis showed that both the irrigated rice crops (P<0.001) and, to a lesser extent, the forest patch surface (P = 0.001) could increase the BU risk within communities by 77.5% and 3%, respectively. In contrast, the dam surface is clearly associated with a lower risk of BU infection (OR equals 0.74) even if we previously showed a significant positive correlation between BU and distance to dams (Figure 5). This apparent discrepancy can easily be explained by the different effects of the presence and the use by populations (contact with population) of dams, as discussed below.
However, since interactions were included in the model, these results are to be interpreted with care. Main effects are to be interpreted as odds-ratios for one covariate when all other covariates are fixed at their base value. For instance, the value 0.75 for dam area means that surface of dams is associated with a reduced Buruli ulcer incidence, only in places where there is no rice, no banana, etc. When rice is present, the odds-ratio for an increase of one unit in dam area is equal to 0.75×1.25 = 0.94 (with 1.25 corresponding to the OR value for the interaction between both parameters), not significant. The odds-ratio for a 1,000 ton increase in rice production and an increase of dam area with respect to no-rice is 1.77×0.75×1.25 = 1.66.
In reality, the different habitats are superimposed in space (see Figures 1 and 4), an observation which was in accordance with the statistical results of the influence of interactions, and, thus, could not be considered as specific variables. Two-way interaction terms between the forest patch coverage-banana crops and the dam surface-banana crops variables also appeared marginally significant to explain the increase of BU in Côte d'Ivoire, with odds-ratio 1.0067 (P<0.001; near 0.6%) and 1.0588 (P<0.001; near an order of 6%), respectively. Finally, the irrigated rice crops versus the banana crops interaction term was a significant protective factor (OR = 0.8938, P<10−6) in the present study.
Infectious agents indirectly transmitted to or between humans because of human-modified environments account for many emerging and re-emerging non-zoonotic or zoonotic diseases today [35]. The list of emerging diseases associated with human behaviours and environmental perturbations is still increasing [36],[37]. Concerning Buruli ulcer disease, although humans are dead-end hosts for the causative agent, Mycobacterium ulcerans, it is well known that the risk of infection is greatly increased by marked exposure to aquatic environments [38]. Even though the disease agent life-cycle is poorly understood, there is a consensus within the public health and scientific community that human behavior, i.e. frequentation of freshwaters for professional or recreational activities, and environmental aquatic alterations are major disease risk factors [19]. However, an extensive quantitative analysis of disease risk is dramatically lacking for Buruli ulcer [39], and risk factor analyses, like the one described in this paper, in parallel to the understanding of the evolutionary properties of the pathogenic microorganism, are definitely necessary to circumvent and control the spread of the disease.
Based on a country-wide statistical investigation on Buruli ulcer cases in Côte d'Ivoire, we have observed the highest incidence of Buruli ulcer disease in the Centre of Côte d'Ivoire, and more particularly in the sector of “V Baoulé”, where there is also the highest concentration of hydro-agricultural dams (the majority of the villages are located less than 5 km from a dam), used for the development of agriculture, especially semi-intensive rice-field and banana production. Interestingly, high occurrence of Buruli ulcer disease was also notified in the South-Western region of the country, where there is a very low density of man-made dams, but where environmental conditions, i.e. pristine ombrophilous dense forest, favour the existence of natural swamps and marshy aquatic ecosystems that could act as reservoirs for the microorganism. Regular agricultural activities, e.g., rice-field or banana production, in the vicinity of the forest remnants cause each year the transformation of the wet dense forest into an open-vegetation landscape, constituted by a mosaic of degraded forest patches, cultivated fields, fallows and degraded landscapes [40],[41]. Community encroachments and settlements in the vicinity of these patchy, altered environments may, thus, constitute foci for Buruli ulcer disease emergence and spread within human communities. On the contrary, in the South/South-East region of Côte d'Ivoire, where a vast coastal lagoon network, e.g. Ebrié, Aby, and two large dams, i.e. Ayamé 1 and 2, are associated with high human density settlements, Buruli ulcer disease is rare because coastal lagoons are brackish water ecosystems, thus potentially hostile to the development of M. ulcerans and/or its hosts/vectors, and man-made lakes are not associated with extensive agriculture on their shorelines. Finally, in the North of the country, small dams and rice fields are not associated to a high incidence of Buruli ulcer disease, although this area has known the development of the most important network of hydro-agricultural equipments in the last 20 years. The low rate of incidence recorded in the Northern area of Côte d'Ivoire can be best explained by the existence of a latitudinal limit possibly related to the bioclimatic North-South gradient. Indeed, contrary to the Southern region, due to the longer dry season period (more than 4 months) few dams and rivers are permanent throughout the year. This creates environmental conditions which are not favorable to the disease agent or its hosts/vectors persistence, thus, decreasing the contact opportunities for disease transmission to human populations. Indeed, the country-wide scale ecosystem signature of Buruli ulcer occurrence in Côte d'Ivoire, i.e. the highest incidence in the Centre of the country, where wet conditions are present all year around, in comparison to the North and the South, might be best explained by the existence of concomitant environmental factors, like the occurrence of specific biological species diversity that could be involved in the disease transmission.
The statistical regression analysis indicates that the most important risk factors for BU in human communities of Côte d'Ivoire are the irrigated rice crops (increase of risk by 77.5%). Other environmental factors, introduced into the analysis, are more marginal for the explanation of the disease risk: the interaction between dam surface and banana crops in the vicinity (increase of risk by 5.9%), and forest patch area and banana crops (increase of risk by 0.7%). We also showed that dam surfaces gave significantly lower odds for disease incidence whereas the increasing distance to dams was related to an increase in BU in Figure 5A. It seems that the distance to small dams is important for increased risk of BU infection, whereas the size of the dams is negatively related to the disease. This apparent discrepancy can be possibly explained by 2 main ways. First, this could be interpreted as a local-scale factor where small dams alter the environment in a way that is different from very large dams (similar to large lakes that usually do not have cases nearby). Small dammed areas can be much different in terms of the environment compared to large dams (lakes, reservoirs), and this is probably important to disease transmission and can imply different risk levels for BU. Here, we speculate that the risk of contamination is likely higher near small dams because of population behaviors. They used small dams for the irrigation of rice-field and banana crops. These small dams and irrigated crops are part of the everyday life of the neighbouring communities, and the dams constitute in most of cases, particularly in the Centre and North of Côte d'Ivoire, the only source of water supply for agro-pastoralism. Moreover, the potential interactions between primary forest patches and the development of agriculture may also constitute favorable ecotone zones that provide new environmental niches for the persistence and spread of M. ulcerans, and should be addressed.
To sum-up, it seems that the most important risk factor of BU regarding dams is not only its presence but mainly the contact between populations and these areas. To go further in this way and to be able to decide between (or define the proportion of) the two hypotheses discussed above (i.e. different environmental conditions and/or population behavior), field work is now required to determine the presence of the bacterium in environment and its potential different distribution in different areas (rice fields, small dams, large dams).
Thus, ecosystem dynamics and its evolution, e.g. land-use changes, biodiversity alteration, as well as socio-economic factors should be systematically taken into account in BUI research, putting this kind of study on a very promising way in order to better understand the transmission route of the bacterium from the environment to human populations, and then a better control of the disease. Further studies at different places and also at a multi-annual scale are now required to acquire a “bigger picture” of this disease.
All these findings are also consistent with two case-control studies in Benin and Ghana [19],[42] as well as a recent nation-wide study in Ghana [25], but they remain innovative since they define new risk factors (e.g. type of crops) for a new country, i.e. Côte d'Ivoire, and add socio-economic factors in the analyses. By extending the analysis of BU risk to a country-wide scale and to socio-economic factors, we highlight here, the importance of a multifaceted approach for disease surveillance. The ultimate goal of our research is to develop a quantitative, spatially realistic model for the BU system that will constitute the framework for the development of a sensible control plan. The present work demonstrates the importance of applying an environmental approach to the study of Buruli ulcer epidemiological problems and more generally highlights the strong necessity for an inter-disciplinary approach.
|
10.1371/journal.pgen.1005195 | Temporal Expression Profiling Identifies Pathways Mediating Effect of Causal Variant on Phenotype | Even with identification of multiple causal genetic variants for common human diseases, understanding the molecular processes mediating the causal variants’ effect on the disease remains a challenge. This understanding is crucial for the development of therapeutic strategies to prevent and treat disease. While static profiling of gene expression is primarily used to get insights into the biological bases of diseases, it makes differentiating the causative from the correlative effects difficult, as the dynamics of the underlying biological processes are not monitored. Using yeast as a model, we studied genome-wide gene expression dynamics in the presence of a causal variant as the sole genetic determinant, and performed allele-specific functional validation to delineate the causal effects of the genetic variant on the phenotype. Here, we characterized the precise genetic effects of a functional MKT1 allelic variant in sporulation efficiency variation. A mathematical model describing meiotic landmark events and conditional activation of MKT1 expression during sporulation specified an early meiotic role of this variant. By analyzing the early meiotic genome-wide transcriptional response, we demonstrate an MKT1-dependent role of novel modulators, namely, RTG1/3, regulators of mitochondrial retrograde signaling, and DAL82, regulator of nitrogen starvation, in additively effecting sporulation efficiency. In the presence of functional MKT1 allele, better respiration during early sporulation was observed, which was dependent on the mitochondrial retrograde regulator, RTG3. Furthermore, our approach showed that MKT1 contributes to sporulation independent of Puf3, an RNA-binding protein that steady-state transcription profiling studies have suggested to mediate MKT1-pleiotropic effects during mitotic growth. These results uncover interesting regulatory links between meiosis and mitochondrial retrograde signaling. In this study, we highlight the advantage of analyzing allele-specific transcriptional dynamics of mediating genes. Applications in higher eukaryotes can be valuable for inferring causal molecular pathways underlying complex dynamic processes, such as development, physiology and disease progression.
| The causal path from a genetic variant to a complex phenotype such as disease progression is often not known. Studying gene expression variation is one approach to identify the mediating genes, however, it is difficult to distinguish causative from correlative genes. This becomes a challenge especially when studying developmental and physiological traits, since they involve dynamic processes contributing to the variation and only single static expression profiling is performed. As a proof of concept, we addressed this challenge here in yeast, by studying genome-wide gene expression in the presence of the causative polymorphism of MKT1 as the sole genetic variant, during the time phase when it contributes to sporulation efficiency variation. Our analysis during early sporulation identified mitochondrial retrograde signaling and nitrogen starvation as novel regulators, acting additively to regulate sporulation efficiency. Furthermore, we showed that PUF3, a known interactor of MKT1 had an independent role in sporulation. Our results highlight the role of differential mitochondrial signaling for efficient meiosis, providing insights into the factors regulating infertility. In addition, our study has implications for characterizing the molecular effects of causal genetic variants on dynamic biological processes during development and disease progression.
| Identifying the causative genetic variants associated with complex human diseases is only the first step [1]. The major challenge is to understand how these genetic variants cause the disease. The mediating molecular pathways connecting these variants to phenotypes have been more systematically understood in model organisms than in humans [2]. However, even in model organisms there are several examples where a causal genetic variant is not a component of the annotated pathways associated to a trait, making it difficult to fully understand its molecular basis [3]. Having this complete knowledge for complex diseases has a huge potential for development and evaluation of available therapeutic and preventive strategies to counter these diseases [4].
Studying gene expression variation is a standard approach for identification of the causal path from a genetic variant to disease [5,6]. Many of these causal genetic variants have been resolved to single nucleotide polymorphisms (SNPs). Several studies in multiple organisms have been performed to study the effects of these variants called as expression quantitative trait loci (eQTLs) [7,8]. However, for making predictions for the molecular mechanisms underlying a disease, trans-acting SNPs are more challenging than cis-acting. This is due to the difficulty in distinguishing causative effects of these SNPs from the correlative effects since a SNP can: i) either affect gene expression and the phenotype independently, or ii) modulate gene expression of downstream molecular players, which in turn causes phenotypic variation (causal mediators), or iii) modulate the phenotype which then affects the gene-expression [5]. A few pragmatic approaches have been recently tested in model organisms to identify the causal mediators by studying gene expression changes. One approach, for instance, involved utilizing expression information for the causal genetic variants from multiple environments, which was a better predictor to identify the causal molecular intermediates by the fact that they interact persistently with the variant [9]. For developmental and physiological processes, gene expression follows complex dynamic patterns [10] and so the effect of eQTLs on gene expression can be highly context-sensitive, depending on the developmental stage, physiological phase or tissue type [11–13]. Therefore, when the causative molecular effects of a genetic variant are being studied by measuring gene expression, knowledge of the particular temporal phase when the causal variant transduces its molecular effects is crucial.
Allele replacement strains have been used extensively for fine-mapping the effects of causal genetic variants associated with a trait [14]. Studying allele-specific gene expression could be yet another useful approach which could be exploited in model organisms such as yeast, to study the precise molecular effects of the causal variant on the trait. This can be done by performing genome-wide expression profiling in a pair of allele replacement strains having the same genetic background except for the allele. Using allele replacement strains, MKT1(89G) was identified as a causal genetic variant for an efficient completion of sporulation in yeast, called its sporulation efficiency [15]. MKT1 is a putative endonuclease and its molecular role is beginning to be, but not completely understood [9,16]. MKT1 has been mapped as a causative gene for several stress-related complex phenotypes, highlighting its extensive pleiotropy [9,17–22], but its functional role in sporulation remains unclear. The developmental process of sporulation in yeast encompasses two meiotic divisions followed by spore formation [23,24]. A study performed parallel phenotyping analysis for the yeast deletion collection and identified around 200 genes required for optimal sporulation efficiency [25]. These genes are both sporulation-specific (i.e., required only during meiotic processes) and majorly sporulation-associated (i.e., required for general cellular functions during sporulation such as nutrient metabolism and respiration). However, the study did not identify MKT1 as one of these genes. It is also not known if MKT1(89G) affects any of these 200 genes or any other gene to increase sporulation efficiency. The first association of MKT1 and sporulation process was reported in the linkage mapping study between segregants of SK1 and S288c strains [15]. Moreover, MKT1(89G) was mapped for sporulation efficiency, the end-point of sporulation process. We do not know at which temporal phase during the course of sporulation (early entry into meiosis, middle progression through meiotic phases, or late spore wall formation), MKT1 affects meiosis.
In this study, we hypothesized that the use of allele replacement strains for studying genome-wide gene-expression during the temporal phase when the causal variant contributes to the phenotype could provide useful insights for identifying the causal molecular mediators underlying complex trait variation. In a pair of allele replacement strains differing solely for MKT1 causal allele, we characterized the molecular role of MKT1(89G) in yeast sporulation efficiency variation. Using genetic assays and mathematical modeling for the meiotic events, we identified the role of MKT1(89G) in the early phases of sporulation. In the specific context of MKT1(89G), we studied the genome-wide transcriptional response particularly in the early phase of sporulation and then genetically tested the candidate mediators. Using such an approach, we identified and confirmed novel pathways mediating the effects of MKT1(89G) in sporulation efficiency variation. The molecular findings resulting from our study demonstrate the advantage of studying allele-specific temporal gene expression dynamics to identify the causal pathways linking genetic variant to complex traits.
Allele replacement of MKT1 in the S288c strain from the endogenous adenine (89A) to guanine (89G), of SK1 strain, resulted in increased sporulation efficiency [15]. Whole-genome re-sequencing of the MKT1 allele replacement strain followed by a series of backcrosses (Methods), was done to confirm that MKT1(A89G) was the only sequence difference between the S288c parent (MKT1(89A) indicated as “S strain”) and the allele replacement strain (MKT1(89G) indicated as “M strain”), the two strains used in this study. After 48h, the high sporulating SK1 strain and the M strain showed increased sporulation efficiency compared to the S strain, which was consistent with the previous report [15] (Fig 1, Table 1, Methods). Compared to the S strain, the SK1 and M strains showed a 17- and a 9-fold increase, respectively (P = 1.9x10-28, P = 1.0x10-25, respectively, pair test in Methods). Deletion of MKT1 in the S strain resulted in sporulation efficiency similar to the S strain, showing that MKT1(89A) is a loss-of-function allele for its function in sporulation (Fig 1, Table 1). However, it is possible that the MKT1(89A) gene product may have an activity for other phenotypes.
To define the temporal phase during sporulation when MKT1(89G) contributes to sporulation efficiency, firstly the proportion of yeast cells completing Meiosis I and II (MI and MII) in the S, M and SK1 strains were quantified (Fig 2A, Methods). M strain started entering MI/II within 10h in sporulation medium, while S strain did not enter MI/II even after 48h. Using these data, multi-stage modeling for the M strain and the parent strains S and SK1 was done to study the distribution of the cell population in different stages of meiosis (Methods, S1 File). As expected, the model predicted that the difference between the M and the S strains occurred during entry into meiosis (initial lag phase of sporulation, S1 Fig). Hence, our observations and the model suggested an early role of the causal variant of MKT1 in sporulation, which was in agreement with a recent study that showed the contribution of causal variants in critical decision-making steps in the early stages of a phenotypic process [26]. In order to confirm this early role of MKT1(89G) in sporulation efficiency variation, a tetracycline-repressible dual-system was used to conditionally express MKT1(89G) (Methods). MKT1(89G) expression was switched off until 3h after initiation of sporulation, which led to a reduction in the sporulation efficiency of the M strain (PTet-MKT1) equivalent to the S strain (Fig 2B, S2 Fig). This result showed that activity of MKT1(89G) allele was essential within the first 3h of sporulation.
Meiotic initiation is regulated by multiple nutrient signaling pathways [27]. The functional allele of MKT1 has a fitness advantage during growth in glucose-rich conditions [9]. Therefore, we tested if increased sporulation efficiency of the M strain is influenced by expression of MKT1(89G) during the rich growth medium stage preceding sporulation (Methods). We observed that switching off MKT1(89G) during growth in glucose had no effect on sporulation efficiency of the M strain (Fig 2B). Altogether, these results indicated that the role of MKT1(89G) during sporulation was independent of its role during growth in glucose and that the allele played a role in the early response to sporulation.
To identify the pathways through which the MKT1(89G) allele affects early sporulation, we studied the entire range of transcriptional response in the S and M strains during the first 10h of sporulation, with denser sampling in the early phase of sporulation (Methods). An extensive remodeling of gene expression was observed in both strains, which increased as time progressed through sporulation (S5 Fig). As expected, the genes involved in sporulation showed a higher expression in the M strain than in the S strain (P = 2.0 x 10–37, permutation P = 0.16, Methods, S6 Fig). Amongst all genes, we identified 862 gene transcripts showing a statistically significant (10% FDR, Methods) differential expression as a function of time between the M and S strains. No enrichment of any functional category within these differentially expressed genes was observed, indicating the pleiotropic role of MKT1(89G) and that it might be affecting various aspects of the sporulation process. Comparison of expression profiles of the few known meiotic regulators in the M and S strains showed that IME1, the master regulator of meiosis [28], was not differentially expressed. However, NDT80, the other crucial regulator of meiosis, involved in meiotic commitment [29], was differentially expressed (S7 Fig, S4 Table). These results suggested that MKT1(89G) could affect sporulation at the post-transcriptional level of IME1 or at the transcriptional level of NDT80, both of which could have early regulatory consequences during meiosis [30]. This observation also suggested that the role of MKT1(89G) during sporulation might be early and upstream to the regulators of meiosis, in agreement with our earlier results (see Fig 2A and 2B).
To capture the early role of MKT1(89G) during sporulation, genes upregulated early in the M strain and either downregulated or expressed later in the S strain, were considered. Thus, differentially expressed genes were clustered based on their expression profiles, separately for the M and S strains (Methods). Clustering gave six and seven clusters in the M and S strains, respectively, from which four major clusters were identified in each strain (Fig 3A, S5 Table). Clusters I and II consisted of genes mostly expressed in the early stages of meiosis with an enrichment for the target genes of IME1 and NDT80, respectively. In particular Cluster I contained some of the earliest expression changes in the M strain. Comparison of this early cluster between the M and the S strains showed that while 46% (71/143) of its genes overlapped (Fig 3B, S5 Table), the remaining 72 early expressing genes were uniquely differentially expressed in the M strain (S6 Table). We posited that transcription factor(s) whose target genes were significantly enriched within these unique 72 early expressing genes of the M strain might be involved in regulating entry into meiosis. Forty one such transcription factors (P ≤ 0.05, odds ratio ≥ 1.5) were identified, which consisted of the regulators of metabolic and mitochondrial signaling (Methods, S7 Table), including sporulation-specific genes, such as IME1, SIN3 and WTM2 (a UME1 paralog). To evaluate if the approach we used indeed identified the causal mediating genes contributing to sporulation efficiency variation in the context of MKT1(89G), we selected a few candidate genes from this list of regulators for further investigation. One of the major concerns while studying gene expression is that transcriptional changes can be buffered at the level of phenotype and so do not always manifest themselves in phenotypic variation [31]. Hence, to avoid this buffering while identifying causal regulators of sporulation downstream MKT1(89G), a comprehensive literature survey was done for the selected 41 transcription factors to identify the prime candidate regulators. We did not consider those genes, which have been previously shown to have a causal relationship with sporulation efficiency variation [25]. While prioritizing candidate genes, specifically those regulators were chosen whose functional annotations were related to the processes associated with early regulation of sporulation, such as mitochondrial function and nutrient starvation, but a causal role in sporulation was not known [24,27,32–35]. From this list, RTG1, a regulator of mitochondrial retrograde signaling [36] and DAL82, a regulator of nitrogen metabolism [37] (Fig 4, S8–10 Figs, S8 Table) were selected for further investigation.
To test the role of RTG1 and DAL82 in sporulation efficiency variation, their deletions in both M and S strains were phenotyped. Another regulator of retrograde signaling RTG3 [38], a physical interactor and target gene of RTG1, showing differential expression in our data, was also deleted in the two strains. Deleting RTG1, RTG3 or DAL82 reduced the mean sporulation efficiency in the M strain significantly, by about two-fold (P = 6.2x10-10, P = 2.8x10-10, P = 1.6x10-7 respectively, Fig 5A, Table 1, pair test in Methods). This effect was specific to the M strain, because deletion of these genes in the S strain did not affect their mean sporulation efficiency (Fig 5A, Table 1, pair test in Methods); and for RTG1 and RTG3, significant interaction terms were found between the backgrounds (S and M strains) and the deletion for these genes (P = 5.8x10-5, P = 4.7x10-3 respectively, interaction test in Methods). RTG1, RTG3 and DAL82 have not been previously identified as involved in sporulation efficiency as determined from a genome-wide deletion screen [25]. Since this deletion collection was made in the S288c background, carrying the non-functional allele MKT1(89A), this could be a possible reason for the lack of functional implication. A deletion study in the SK1 strain that contains the functional MKT1(89G) allele, did not investigate the association of these early sporulation regulators with the process [39]. However, interestingly, an up-regulation of RTG1 in the early phase of sporulation has been observed in SK1 [40]. These results, thus, support our approach of studying the early effects of the causative allele and implicate novel roles for RTG1, RTG3 and DAL82 in the early phase of sporulation efficiency downstream to MKT1(89G).
To further investigate if RTG1/3 and DAL82 belonged to the same pathway (epistatic effect) or were in separate pathways (additive effect), double deletions for RTG3 and DAL82 were phenotyped in the M strain. Deletion of RTG3 and DAL82 together reduced the mean sporulation efficiency of the M strain by approximately 3-fold (Fig 5A, Table 1). A non-significant interaction term was obtained between RTG3 and DAL82 (interaction test in Methods), indicating that they regulated sporulation efficiency additively, downstream to MKT1(89G). Furthermore, because deletion of RTG3 and DAL82 in the M background only partially reduced the sporulation efficiency to that of the S strain (P [M (rtg3∆ dal82∆) vs. S] = 2.5x10-7, Fig 5A, pair test in Methods), these results indicated that these genes explained a partial role of MKT1(89G), and additional complementary pathways were at play.
The mitochondrial retrograde signaling pathway gets upregulated in response to altered mitochondrial function and nutrient starvation. This pathway fine-tunes the metabolic and stress response pathways of the cell by affecting glutamate synthesis and mitochondrial DNA maintenance [33,41]. Since mitochondrial function with regard to respiration is implicated as a critical regulator of sporulation [42], we speculated if differential mitochondrial activity was involved in sporulation efficiency variation in the presence of MKT1(89G). We evaluated the mitochondrial function in the M and S strains by assaying oxygen consumption flux during early sporulation (Methods). The M strain showed a better mitochondrial function than the S strain (Fig 5B) at 1h in sporulation. Deletion of RTG3 in the M strain decreased this oxygen consumption flux, though dal82∆ had no effect on the flux (Fig 5B). These results suggested a role of differential mitochondrial function in sporulation efficiency variation. However, a better understanding of the role of mitochondrial retrograde pathway in sporulation efficiency would require further investigation.
Differential mitochondrial activity in the presence of MKT1(89G) suggests a role for the Mkt1 interactor, Puf3, a Pumilio-family protein, which has been suggested to explain the extensive MKT1(89G) pleiotropy during mitotic growth in rich media as well as in stress environments [16,22,43]. Puf3 is an mRNA binding protein that regulates the fate of nearly 200 nuclear-encoded mitochondrial transcripts [44]. Even though we found a few PUF3 target genes (13/214 genes) differentially expressed during sporulation, none were in the set of unique early expressed transcripts in the M strain (S10 Fig). To further evaluate if PUF3 had a role in sporulation efficiency variation in the presence of MKT1(89G), we deleted PUF3 in the S and M strains and M strain with single deletions of rtg3∆ and dal82∆. If PUF3 has an independent role in sporulation, reduction in sporulation efficiency by puf3∆ deletion would be independent of the background (MKT1, RTG3 or DAL82), and we would observe an additive effect on sporulation efficiency. Any observed significant deviation from this expectation would imply dependence. One extreme case of dependence would be epistasis. In that case, deleting PUF3 in these backgrounds would not lead to decreased sporulation efficiency. We observed that PUF3 deletions in all the four backgrounds: M, S, M (rtg3∆) and M (dal82∆) reduced their sporulation efficiency (regression line y = 0.65x showing around 35% less sporulation efficiency for all strains, Fig 6A and 6B, Table 1, pair test in Methods). Furthermore, interaction terms (Methods) were non-significant for deletion of PUF3 between the M and the S strains (P = 0.49), the M and M (rtg3∆) strains (P = 0.53), and only mildly significant between the M and M (dal82∆) strains (P = 0.02). These results indicated that the effect of PUF3 on sporulation efficiency was independent of MKT1(89G) and its downstream genes RTG1/3 and DAL82.
Over the past decade a detailed genotype-phenotype map for complex traits including diseases has been determined [45], however, a functional map defining how causal genetic variants (alleles) modulate the underlying pathways resulting in phenotypic variation, is missing. Filling this functional gap will help to identify molecular candidates for therapeutic intervention in human diseases and to make useful predictions regarding response to a particular therapy and survival of a patient [1]. The first step to characterize this functional genotype-phenotype map requires identification of the causal mediating genes in a biological network regulating the phenotype. Investigation of the intermediate phenotypes viz. transcripts, proteins and metabolites, is routinely used to identify these causal mediators [46]. In this study we demonstrate a couple of steps essential for accurate identification of these causal molecular mediators: i) studying allele-specific temporal dynamics of the biological processes underlying complex traits, and ii) allele-specific functional validation of the predicted mediators. We report the characterization of molecular pathways modulated by a causal genetic variant in a dynamic biological process using the above approach. In particular, we studied the molecular effects of the essential MKT1(89G) allele on the yeast transcriptome during sporulation. We not only identified novel pathways regulating the phenotype, but also confirmed the independent role of a known interactor (Puf3) of MKT1(89G) in the phenotype (Fig 7). MKT1(89A) is not a naturally occurring allele, observed only in the S288c strain [20]. However, such rare polymorphisms are receiving increasing attention for their contribution to common human diseases [47]. In this sense, our approach has a general applicability since it can be applied to study the molecular basis of both common and rare variants.
Using our approach of studying early gene expression dynamics in response to the MKT1(89G) allele, we identified that regulators of mitochondrial retrograde signaling and of nitrogen starvation act additively to regulate sporulation efficiency (Fig 5A). Mitochondria responds to a wide array of stresses by inducing various complex cellular responses and promoting cellular adaptation to reduce the impact of further stressors [48]. Mitochondrial retrograde signaling is one of the stress signaling responses of the cell during mitochondrial functional alteration and glutamate starvation [33]. It affects mitochondrial DNA maintenance [49] and hence the respiratory competency of a cell. During meiosis in yeast cells, energy production occurs through the Krebs cycle [32,35,42], and hence respiration is a critical regulator of meiosis in yeast [42] and in humans. In humans, low mitochondrial DNA has been associated with ovarian insufficiency [50]. We observed an improved mitochondrial activity during early sporulation in the M strain compared to the S strain (Fig 5B). A reduction in this high mitochondrial activity in the absence of mitochondrial retrograde signaling regulator RTG3 indicated that MKT1(89G) might confer a better stress response through RTG3, with increased sporulation efficiency being one of the consequences. This role of retrograde signaling in regulation of developmental processes responding to nutritional stresses has been shown for pseudohyphal growth in yeast [51]. Further investigating this association of differential mitochondrial signaling, particularly retrograde signaling with meiosis and development in general can help provide insights into the factors regulating infertility.
In this study, we characterized the essential role of MKT1(89G) allele in sporulation efficiency. This allele was particularly interesting to study as this coding polymorphism of MKT1 is present in all laboratory strains (except strains isogenic to S288c), as well as clinical and natural isolates of yeast including the SGRP strain collection [15,18,20,52]. Since the previous genetic screens [39,25,53] or genome-wide expression studies [40,54] for sporulation and sporulation efficiency, were done in the S288c background carrying the MKT1 allele which is non-functional in sporulation, this could be a possible reason for not identifying MKT1 to be involved in the process. The founder strain of S288c, EM93 carries the MKT1(89G) allele suggesting that during domestication of S288c this functional allele was lost [20,55]. During evolution of S288c in low-glucose conditions, the native MKT1(89A) mutated to MKT1(89G) within 500 generations [56], also indicating the crucial role of MKT1(89G) in stress-related conditions. Altogether, these observations demonstrate the limitations of studying genotype-phenotype relationships in a single genetic background, especially in laboratory strains, which might have degenerated their stress response machinery partially or completely, as a result of domestication [57].
Using our approach, we further showed an MKT1(89G)-independent role of PUF3 in meiosis (Fig 6A and 6B). This was surprising since eQTL mapping studies have suggested MKT1 as a global regulator of gene expression [22,58] and have identified its most upstream interactors, such as PUF3, during mitotic growth in multiple environments [9,16]. Puf3 regulates translation and degradation of nuclear-encoded mitochondrial mRNAs by localizing them near mitochondria or P-bodies, which are cytoplasmic sites for mRNA decay and stalling [16,44,59,60]. Since MKT1 has a post-transcriptional regulatory role both in yeast [61] and in trypanosomes [62], its interaction with PUF3 suggested a probable mechanism for understanding the role of MKT1. However, for sporulation efficiency, we observed that Puf3 showed an MKT1(89G)-independent role. We, therefore, speculate that Puf3 might be a mitotic growth-specific interactor of MKT1(89G). Its role in sporulation efficiency, though, could involve post-transcriptional regulation of mitochondrial mRNAs through P-bodies during sporulation. In Drosophila, C. elegans, mice and mammals [63,64], P-bodies related RNA granules are known to be involved in translational control of germ cell transcripts. However, in yeast, P-bodies have been observed only during glucose starvation and stress conditions such as ethanol tolerance [22,65]. Therefore, our results indicate an interesting interaction between Puf3 and sporulation efficiency variation and this could be a future line of investigation to determine if P-body formation has a regulatory role in yeast meiosis.
Through our analysis, we attempted to understand the molecular basis of a complex trait. Using an allele-specific approach, we determined and functionally validated the molecular consequences of a single causative variant in phenotypic variation. This approach helped to identify novel associations between mitochondrial and metabolic pathways with meiosis. Further analyses of these expression data can identify additional regulators and pathways involved in sporulation efficiency variation in the presence of MKT1(89G) (Fig 7, S7 Table). This approach demonstrated in yeast can be applied to higher eukaryotes to study transcriptional dynamics of developmental processes or progression of diseases. This will assist in understanding the precise genetic effects of a causal variant, improving the existing genotype-phenotype functional relationship map.
Whole-genome resequencing of the MKT1 allele replacement strain (S9 Table) was performed to confirm the presence of the causative SNP (details in S1 Table, S1 Text (Section 1)). Backcrossing the haploid allele replacement strain to the S288c parent strain three consecutive times (details in S1 Text (Section 2)) confirmed that homozygous MKT1(A89G) was the only sequence difference between the diploid S288c parent (S strain) and the allele replacement strain (M strain). All the S (MKT1(89A)) and M (MKT1(89G)) strains used in this study were derivatives of S288c strain except SK1 strain (S9 Table). The strains were grown at 30°C in YPD (1% yeast extract, 2% bacto peptone, 2% dextrose) and YPA (1% yeast extract, 2% bacto peptone, 1% potassium acetate). Deletions were performed in the haploids by replacing the specific ORF with one of the dominant drug-resistance cassettes (hphMX4, kanMX4 or natMX4) which were PCR-amplified from their respective plasmids as described previously [66]. The strains were transformed using the standard lithium acetate-based method [67] and homologous integration of the deletion cassette was confirmed by performing a colony PCR for both the ends. Three confirmed independent transformants were selected to minimize random mutations during the transformation step, diplodized using pHS2 plasmid (containing a functional HO) and phenotyped. All further experiments were performed using the diplodized parent strains and their diploid derivatives. The primers for deletions and their confirmations are listed in S10 Table. Sporulation conditions and the calculation of sporulation efficiency was done as previously described [68] in liquid sporulation medium (1% potassium acetate supplemented with 20mg/ml uracil, 20mg/ml histidine, 30mg/ml leucine, 20mg/ml methionine and 30mg/ml lysine). For each strain, minimum three biological replicates were used and approximately 1,000 cells were counted per replicate. Fold difference was calculated as the ratio of mean sporulation efficiencies of the two strains A and B when the sporulation efficiency of A is greater than of B.
Two statistical tests were used: the pair test and the interaction test. The pair test tests the null hypothesis that two given strains have the same sporulation efficiency. To this end, the number yi,k of sporulated cells (4-nuclei count) among the total number of cells ni,k of strain i in replicate experiment k was modeled with a quasi-binomial generalized linear model using the logit link function and subject to a common log-odd ratio βi between replicates, i.e.:
log(μi,kni,k−μi,k)=βifor allk,
where μi,k = E(yi,k). The pair test tests the null hypothesis of equality of log odd-ratios for two strains i and j, i.e. H0: βi = βj.
The interaction test tests the null hypothesis that the effect of mutation A is independent of the effect of mutation B, taking the M strain as reference background. This test thus compares four strains: mutation A only, mutation B only, both A and B and neither A nor B (M strain). Here, the strain S was considered as a M strain mutated for MKT1(89). For every interaction test, we considered the dataset of the four strains of interest and fitted a quasi-binomial generalized linear model using the logit link function and subject to:
log(μi,kni,k−μi,k)=β0+βAAi+βBBi+βA,BAiBifor allk,
where, Ai and Bi are indicator variables of the mutations A and B in strain i respectively. The interaction test tested the null hypothesis that the odd ratio of sporulation in the double mutant equals the product of the odd ratios of each mutation, i.e. H0: βA,B = 0.
Both the pair test and the interaction test were implemented in the statistical language R with the function glm() assuming a constant variance function fitted by maximizing the quasi-likelihood and using the t-test on tested parameters (see S2 File for raw data and R script).
Aliquots of sporulating cells of M strain culture were fixed with ethanol at regular intervals (as indicated in Fig 2A) from 0 to 48h in the sporulation medium. These time-points were chosen to capture the progression through meiotic stages in the strain. Samples were stained with DAPI (4’-6’ diamidino-2-phenylindole) using the standard methods [69] for calculating the proportion of cells with 1-nucleus (Non-sporulating/G1), 2-nuclei (MI) and 4-nuclei (MII) using Carl Zeiss Axiovert 200 fluorescence microscope. For each strain, proportion of cells were counted till saturation was reached for two consecutive time points. Grey scale images were captured using a CCD camera and pseudo-coloured using the image acquisition software (Axiovision) supplied with the microscope. To estimate the sporulation efficiency and DAPI staining, 1,000 cells from the three biological replicates for each strain were counted.
A multi-stage modeling was performed (details and raw data in S1 File). Cells in G1/S phase of cell cycle are said to be in 1-nucleus state. Cells that have completed MI or MII are said to be in 2-nuclei or 4-nuclei state, respectively. Cells that did not progress from one cell cycle state to another are mentioned as inactive cells. The existence of inactive states is supported by the fact that at steady state, some cells still have one nucleus or 2-nuclei indicating they are trapped at these stages, which could be possibly due to nuclear destruction mechanism resulting in dyads [70]. Hence, cells could be either in a 1-nucleus active, 1-nucleus inactive, 2-nuclei active, 2-nuclei inactive or 4-nuclei state. Moreover the cells were assumed to only progress in one direction (no back transitions) from the 1-nucleus active to either the 1-nucleus inactive or the 2-nuclei active stage, and from the 2-nuclei active to either the 2-nuclei inactive or to the 4-nuclei state. The samples contain a large number of cells and thus we used Ordinary Differential Equations to describe the dynamics of the system. The dynamics was modeled with an initial lag phase (measured as τ) followed by first order kinetics between the stages (measured as α, β, γ and δ, as shown below).
(X1→αX2→γX4X1→βY1X2→δY2)
where, X1 is proportion of cells in 1-nucleus active stage, X2 in 2-nuclei active stage, X4 in 4-nuclei active stage, Y1 is proportion of cells in 1-nucleus inactive stage, Y2 in 2-nucleus inactive stage. The model was fitted by minimizing least square errors to the measured proportions of the cells with 1, 2, and 4-nuclei, measured along the time. Confidence intervals were obtained by bootstrap of the data.
tetO7-based promoter substitution cassette containing kanMX4, amplified from the plasmid pCM225 [71], was inserted to replace the endogenous MKT1 promoter (-300 to -1bp upstream start site) in the M strain (PTet-MKT1). M strains with the endogenous promoter (Pwt-MKT1) and the tetO7 promoter (PTet-MKT1) were grown in a glucose-rich medium (YPD) and synchronized in pre-sporulation medium (YPA) prior to initiating sporulation. To determine the concentration of doxycycline at which the effect of MKT1(89G) on sporulation efficiency is similar to MKT1(89A) (implying MKT1(89G) is not functional or OFF), the PTet-MKT1 strain was grown and sporulated in 2, 3 and 5μg/ml of doxycycline and phenotyped by estimating the sporulation efficiency after 48h. At 5μg/ml doxycycline, the sporulation efficiency of the PTet-MKT1 strain was similar to the S strain (S2 Fig) and this concentration was used for further experiments. To switch off MKT1(89G) expression only during the growth in glucose, the PTet-MKT1 strain was grown in YPD with doxycycline, washed and added to YPA and the sporulation medium in the absence of doxycycline. For switching off MKT1(89G) throughout the sporulation process, doxycycline was added to all the three media (YPD, YPA and sporulation). To switch off MKT1(89G) till 3h in sporulation medium, doxycycline was added in YPD, YPA and sporulation medium. Cells were washed after 3h in sporulation and resuspended in the sporulation medium without doxycycline till 48h, and were phenotyped. A complementary experiment where MKT1(89G) was switched ON till 3h in sporulation medium and switched OFF from 3h to 48h in sporulation was done by adding doxycycline in sporulation medium post 3h in sporulation medium (S2 Fig). For each strain in each condition, minimum three biological replicates were used and approximately 1,000 cells were counted per replicate per condition for estimation of sporulation efficiency. The means and variances were tested for significance using one-way ANOVA followed by Tukey’s multiple comparisons test (Prism, Graphpad Software Inc.). Statistical significance was determined at P < 0.05.
Temporal transcriptome profiling was performed for the sporulating yeast cells at 0h, 30m, 45m, 1h10m, 1h40m, 2h30m, 3h50m, 5h40m and 8h30m (logarithmic time-series) in the sporulation medium. For this, 100ml aliquots of the culture were pelleted and stored at -80°C. Transcriptome profiling was performed using the S. cerevisiae yeast tiling array (Affymetrix, Cat# 520055) as described previously [72]. Time-series arrays of M and S strains in sporulation were normalized by vsn (S1 Text (Section 3), S3 Fig) [73].
Using log2 transformed expression values, after normalization (S2 Table), the expression profiles of all transcripts of S and M strains were made continuous over time using locfit [74] with the bandwidth parameter ‘h’ optimized at 1.21 (S1 Text (Section 4), S4 Fig, S3 Table). A baseline transformation for each transcript, after smoothing, was done by subtracting each time point value from t = 0h (t0).
y'S(tn)=yS(tn)−yS(t0)y'M(tn)=yM(tn)−yM(t0)
where, y is the expression value of a transcript for a strain (S or M) at a specific time point and y’ is the transformed expression value.
To compare the sporulation genes (obtained from Deutschbauer et al. [25]) between the M and S strains, their expression in the two strains were tested using 1,000 permutations of Wilcoxon test on an equal number of randomly selected genes (S6 Fig). R scripts used for the analyses are given in the S3 File.
To identify differentially expressed genes (after removing tRNAs, snRNAs and transcripts from terminal repeats) between the two strains, the temporal expression profiles of each transcript was compared using the method implemented in the EDGE (Extraction of Differential Gene Expression) software [75]. One thousand permutations were done to calculate the null distribution with a random number seed. EDGE analysis identified transcripts of 862 significant differentially expressed genes across time (10% FDR, S4 Table). Within these 862 genes, a subset of differentially expressed transcription factors and differentially expressed targets of all the transcription factors (obtained from the YEASTRACT database, [76] were selected. This subset of 727 genes was used for further analysis.
The 727 differentially expressed genes were clustered according to their temporal expression patterns using time abstraction method implemented in the TimeClust software [77]. The smoothened and baseline transformed expression data of the 8 sporulation time-points was analysed with window span parameter set at 3. An absolute expression change of 0.1 was considered as a change. This clustering method was applied on the expression data separately for the two strains resulting in six and seven clusters in the M and S strains, respectively (S5 Table). The gene lists of the M and S strains for the Cluster I, consisting of early expressing genes, were compared. For the genes unique to the M strain in this cluster (S6 Table), the transcription factors regulating them were extracted using the YEASTRACT database (S7 Table) [76].
After 1h in sporulation, 5 x 106 cells from each of the three biological replicates were used for the assay. Oxygen consumption flux was determined, in total volume of 2.1ml sporulation medium at 30°C with 500 rpm, using OROBOROS O2k high-resolution respirometer (OROBOROS Instruments Corp., Innsbruck, Austria). Data acquisition and calculation of oxygen flux was done according to the manufacturer’s instruction in DatLab software. Unpaired Student’s t-test (Prism, Graphpad Software Inc.) was performed for comparing differences between the means of the two strains. Statistical significance was determined at P < 0.05.
The Supporting information is also available at: http://www.tifr.res.in/~dbs/faculty/hsinha/MKT1Spo
|
10.1371/journal.pntd.0004621 | Temporal Dynamics and Spatial Patterns of Aedes aegypti Breeding Sites, in the Context of a Dengue Control Program in Tartagal (Salta Province, Argentina) | Since 2009, Fundación Mundo Sano has implemented an Aedes aegypti Surveillance and Control Program in Tartagal city (Salta Province, Argentina). The purpose of this study was to analyze temporal dynamics of Ae. aegypti breeding sites spatial distribution, during five years of samplings, and the effect of control actions over vector population dynamics.
Seasonal entomological (larval) samplings were conducted in 17,815 fixed sites in Tartagal urban area between 2009 and 2014. Based on information of breeding sites abundance, from satellite remote sensing data (RS), and by the use of Geographic Information Systems (GIS), spatial analysis (hotspots and cluster analysis) and predictive model (MaxEnt) were performed. Spatial analysis showed a distribution pattern with the highest breeding densities registered in city outskirts. The model indicated that 75% of Ae. aegypti distribution is explained by 3 variables: bare soil coverage percentage (44.9%), urbanization coverage percentage(13.5%) and water distribution (11.6%).
This results have called attention to the way entomological field data and information from geospatial origin (RS/GIS) are used to infer scenarios which could then be applied in epidemiological surveillance programs and in the determination of dengue control strategies. Predictive maps development constructed with Ae. aegypti systematic spatiotemporal data, in Tartagal city, would allow public health workers to identify and target high-risk areas with appropriate and timely control measures. These tools could help decision-makers to improve health system responses and preventive measures related to vector control.
| As reported in Porcasi et al., in Argentina we are working on an integrated risk stratification system based in geospatial technologies that have moderately consolidated national scale, but need more understanding of its urban scale mechanisms. In this work, relevant results are shown on how Ae. aegypti breeding sites are distributed in dynamic spatial patterns in a small city on northern Argentina. 5 years of entomologic data were obtained by Mundo Sano Foundation, which is implementing an Aedes aegypti Surveillance and Control Program in Tartagal City (Salta Province, Argentina). The focus of this contribution is based on the difference that can found between one year data typical analysis and long term temporal evolution of spatial patterns. Although environmental sanitation activities and chemical control of breeding sites with larvicide were performed after each entomological surveillance all around Tartagal, sectors with higher densities of breeding sites remained present throughout study period. Nonetheless, the distribution of breeding sites showed a spatial dynamic with high density clusters in city outskirts.
| Mosquitoes of Aedes genus are the principal vectors of Dengue, Yellow Fever, Chinkungunya and Zika viruses in the Americas [1,2]. Aedes aegypti (Diptera: Culicidae) transmits Dengue virus in the tropical and subtropical South America regions, and its transmission is influenced by various factors, including vector mosquito density, circulating virus serotypes, and human populations susceptibility [3]. In Argentina, Ae. aegypti is the most relevant mosquito from epidemiologic point of view. This specie is characterized by its adaptation to the urban environment, its capacity and preference of breeding in artificial containers [4], the resistance of its eggs to desiccation and the feeding behavior of the female which bites in multiple occasions during each gonadotrophic cycle [5]. These characteristics, together with this vector wide distribution in Northern Argentina, constitute fundamental factors that influence circulation and transmission of Dengue and other related viruses in the region [6].
After a successful vector eradication campaign, at national level, in the 70´s [7], the first outbreak of dengue in Argentina was documented in 1998. Since then, intermittent outbreaks of the disease, with variable incidence rates, were registered in an almost continuous manner in the center and northern regions of the country [8]. A major dengue outbreak reached subtropical regions of Argentina in 2009, affecting more than 25,900 people including localities such as Buenos Aires and Córdoba [9,10]; although the largest percentage (over 90%) corresponded to case reports from Chaco, Catamarca and Salta provinces [8]. In this last province, in Tartagal city, around 665 dengue cases were confirmed including the first fatal case of this disease to be ever registered in Argentina [8]. From 2010 to 2014, a total of 338 suspected cases were registered in this city, from which 56 cases were confirmed (Hospital Provincial J.D. Perón, Tartagal, personal communication).
Taking into account 2009 epidemiological situation, in October of that year, Mundo Sano initiated an Ae. aegypti surveillance and control program with the objective of reducing the risk of dengue transmission in the city of Tartagal, Since then, a permanent surveillance system of breeding sites and key infestation determinant factors involved in was implemented to generate a systematic information record of high epidemiological value.
Considering Dengue native transmission, cases introduction from Paraguay, Bolivia and Brazil, and the absence of an effective vaccine [11], the north region of Argentine needs continuous vector control programs applications. Traditional Ae. aegypti control measures include elimination (breeding sources reduction) or larval habitats chemical treatment to prevent adults production, and space spray insecticides application to reduce adult population densities [12]. In this sense, current control methods require a clear identification of the areas and the periods of mayor entomologic risk, as well as the identification of the viral propagation flow in a community [13].
Multiple environmental factors, including biophysical and social ones, constitute a complex web that determines the spread of vector-borne diseases [14]. In this sense, Ostfeld and collaborators [15] indicated that despite the complexity, an analysis of the variables linked to vectors distribution and the identification of dengue cases can be a useful tool to generate future spatial and temporal scenarios for dengue. Spatial analysis of health events contribute to early detect situations involving diseases transmission [15], while the detection of disease clusters allows the identification of nonrandom events and the possibility of inferring their epidemiological determinants [16]. Surveillance tools, such as incidence maps, have been used to enhance public health preparedness for dengue outbreaks by providing a visual aid for decision-making [17,18].
On the other hand, the use of satellite images in epidemiological analyses allows the identification of key environmental factors (temperature, rainfall and humidity) that influence the dynamic of the vectors, as well as their interactions [19,20]. Since the beginning of remote sensing (RS) technology, studies on vector-borne diseases have focused on identifying and mapping vector habitats [21], assessing environmental factors related to vector biology [22–24] and studying disease epidemiology [25,26]. Recent studies investigated the application of RS and spatial analysis techniques to identify and map landscape elements, that collectively define vector and human population dynamics related to disease transmission risk [27,28]. In addition, the development of increasingly sophisticated Geographic Information Systems (GIS) and RS has provided a new set of tools for public health professionals to monitor and respond to health challenges [29,30].
In this frame, Louis and collaborators [12] have detected a great diversity of both predictors and modeling approaches employed to create dengue risk maps through a systematic review and determined that the field of predictive dengue risk mapping is young and still evolving. In addition, different studies propose measures of prevention and control of Ae. aegypti for the elaboration of maps based on the results obtained from a bounded availability (space-time) of recorded data from both, field data and satellite imagery [31–33]. In this sense, an increase in the quality (amount and accuracy) of the field data used for the development of predictive maps will allow public health workers to identify areas of high risk for adequate control of the disease [19,34,35]. Despite the knowledge of Ae. aegypti biology and the use of monitoring tools, the precise detection of high density spots of vector breeding sites, as places of occurrence of the disease, remains poorly understood.
Therefore, the purpose of this study was to analyze 5 years space-time dynamics of Ae. aegypti breeding sites and control actions effect on its populations, in Tartagal City (Prov. of Salta, Argentina). We discuss the predictive capacity of Ae. aegypti spatial distribution models, generated through environmental variables and minimal field data. This models constructed for dengue surveillance based on entomological risk maps, are considered a step in the generation for vector control strategies.
Tartagal city is located at the base of the Argentinean sub-Andean hills (22°32’ S, 63°49’ W; 450 m above sea level) in Salta Province (Fig 1). As the third largest urban center of the province, with 79,900 inhabitants, it includes several ethnic groups such as native Amerindians. The city is located 100 km northern of Capricorn Tropic and to 55 km Southern of Argentinean-Bolivian border (Fig 1). The city is surrounded by subtropical native forests and crops such as beans, cotton, soybean, maize, grapefruit and tomato.
The climate is subtropical, with an average annual temperature about 23°C; and an average maximum of 39°C (in summer) and average minimum of 9°C a(in winter) respectively. Annual cumulative precipitation is about 1,100 mm, with a dry season from June to October with a monthly average rainfall of 30 mm, that sharply contrasts with the wet season from November to May with a monthly average rainfall reaching 160 mm.
Tartagal is characterized by a cultural diversity based on the presence of several autochthonous ethnic groups and emigrant population and continued migration from the bordering country of Bolivia. This feature produces an important effect on the cultural, social and economic profile of this community.
The urban area of Tartagal city covers approximately 15 km2, and is composed by 1,027 blocks and 17,911 housing units. Each of the housing units was georeferenced by the use of GPS receiver. In Fig 1, sectors (a) and (b) refer to new neighborhoods that were incorporated in data collection and entomological control actions performed in 2011 and 2013, respectively.
Presence and abundance data of Ae. aegypti larval stages breeding sites were registered from 2009 to 2014 in Tartagal city, using a methodology called Focal Cycle (FC). This method consists in the entomological surveillance and chemical treatment of 100% of the housing units in the study area. A total of 10 FCs were performed; the first 8 were performed in a continuous manner between the years 2009 and 2012 (Table 1). The analysis of the data showed that during each year winter and spring, the presence of breeding sites and larval stages remained low. In order to optimize resources without losing any information, since 2012, during the winter-spring periods, random larval samplings were performed in 20% of the blocks of the city, alternating with FCs in the summer-autumn periods (Table 1). A total of 5 random larval samplings (denominated M1 to M5) were performed during the study period.
During winter-spring period, chemical treatment was substituted by physical management and/or removal of containers that could serve as breeding sites. For each period, entomological field records consisted in the complete inspection of the housing units, within each block, registering information on the type and number of containers, grouped by the following categories: tires, tanks, drums, barrels, vases, pots, building materials, auto parts, bottles, cans, plastic, wells, cisterns, natural receptacles, and others (washing machines, refrigerators, toilets, etc.). Total number of containers was counted, such as the number of containers with water, and with larval stages. Larval stages were collected in individual tubes by container, labeled and transported to Mundo Sano entomological laboratory, in Tartagal city, for taxonomic determination using a specific morphological key [36].
Housing units were considered positive when they presented at least one container, with one or more larvae or pupae of Ae. aegypti. Additionally, a series of places were inspected and identified as critical breeding sites, since they presented an elevated number of containers in comparison to those registered in the housing units. The cemetery, municipal garbage dump, tire repair shops, small garbage accumulation sites and other similar places were included in this category.
In each FC, the entomological indexes at the housing unit level were calculated using the House Index (HI) = number of positive homes/number of houses inspected) x 100 and Breteau Index (BI) = total number of breeding sites/number of houses inspected) x 100 [37,38]. These indexes are generally accepted for operational use [39,40].
After the entomological data collection, focal control actions were performed in each housing unit which entailed, for each FC round, mechanical treatment (modification, elimination or destruction) together with the application of the larvicide in a 1 mg/L dose, following the guidelines elaborated by TDR/WHO and the Argentinian Ministry of Health [41,42]. These actions were accompanied by a communication campaign through the use of printed pamphlets destined to inform the general population about the disease and its risks. Moreover, with the objective of reducing the environmental burden of active and potential breeding sites generally accumulated in the peridomicile, neighborhood rallies were organized in collaboration with the local municipality and the participation of local public and private entities, to get rid of containers that favor the accumulation of water during the weeks prior to the start of the rainy season and during the summer months.
In order to analyze the spatial and temporal distribution of the positive breeding sites, GIS vector layers were created including FCs data that were performed during summer and autumn of each year (Table 1). In this sense, and in order to comply with what was previously detailed, FC1 and 2 were combined since the interval of time between these FCs is equivalent to the time registered for the other FCs analyzed: FC5, FC8, FC9 and FC10 (Table 1). In order to avoid confusion, the unit of time of years will be used during the analysis to reference the FCs that correspond to the summer and autumn season of each year. Vector layers were generated using the free-access software Quantum GIS Desktop v2.6.1. Brighton (QGIS).
Density breeding site maps were elaborated using discreet information (sites of individual sampling) through QGIS tool “heatmap”, in order to analyze the manner in which Ae. aegypti breeding sites were distributed in the city. Annual density breeding site maps (summer-autumn) were generated using the Kern density algorithm that calculates the density of positive points (grouping of the points) for a determined area. Using this methodology, the heat map allows for a visual identification of the hot spots for a particular time and place.
The methodology of statistical spatial analysis exploration, developed by Kulldorf [43] was used to identify spatial clusters with Ae. aegypti larval stages presence, with greater density than those expected by a random distribution. The analysis would then indicate some areas with a greater presence of breeding sites than others. Sites with the presence of larval stages were indicated as positive cases (1) and those without any presence were indicated as negative controls (0).
The analysis consisted of a spatial scan through the superposition of exploratory circles, over sites with a record of larval presence. Each circle is a possible cluster and, taking into account the number of events inside and outside an expected number of events, each probability is calculated. The circle that presents the maximum probability, and an excess in the number of events observed versus expected, is defined as the most probable cluster [43]. In this case, the maximum size of the cluster was assigned as 30% of the total population under study.
In our analysis, for each place and window size (circle), the null hypothesis assumes that sites with the presence of larval stages are randomly distributed, while the alternative hypothesis indicates that there is a greater risk inside the window in comparison to the outside. A maximum of 999 Monte Carlo replications were performed in other to search for statistically significant composites. Only the composites that achieved statistical significance (p<0.05) under Bernoulli´s distribution were reported. The purely spatial exploration model was used for each year within 20102014 period in Tartagal. The statistical analysis was performed with SaTScan v9.3.1[44] software, while cartographic representations were done in QGIS software.
SPOT images (Satellite Pour l’Observation de la Terre) were used to characterize the types of environmental coverage in Tartagal. This is a commercial high-resolution optical imaging Earth observation satellite system, operating from space. In this case, we used the SPOT 5 J product, of 10 meter resolution in multispectral mode, with four bands on short wave infrared: green (0.50–0.59 μm)–red (0.61–0.68 μm)–nearest infrared (0.78–0.89 μm) and middle infrared (1.58–1.75 μm). SPOT image (16-11-2013) data was used to generate land cover classifications and macro-environmental products of the study area (Fig 2). All the images used were supplied by the Comisión Nacional de Actividades Espaciales (CONAE).
Unsupervised classification (k-means) classifiers have been used to classify the image of the study area as described by Rotela and collaborators [45]. Seven land cover classes were identified: bare soil, low vegetation (grass), high vegetation (trees), urban buildings, superficial water, shadows, and pasture and crops (Fig 2). A set of ground truth points (about 35/40 points for each class) were generated using Google Earth in order to validate classification accuracy. The confusion matrix, when using control points, showed an overall accuracy of 79.4% and a Kappa coefficient of 0.74. The classes of (low and high) vegetation, and bare soil and pasture reached values above 70% accuracy, and the urban class presented lower registers. QGIS and ENVI 5.1[46] software were used to create the vectors and assess the accuracy of the classification.
Based on the land cover classes previously created, two different types of macro-environmental variables were generated for each class, expressed as i) “distance maps or buffer image” (Fig 3) and ii) "percentage" of each land cover class (Fig 4), according to Rotela and collaborators [45]. In our study, the window size for the maps generated was 31x31 pixels, attributing to the central pixel the average value of the central window. A flight range of 150 m for Ae. aegypti [47,48] was used to generate the new land cover classes (distance and percentage), which could describe the environment that represents the average habitat of the species. All these analyses were performed using ENVI 5.1.
Tartagal information provided by the Instituto Nacional de Estadística y Censos (INDEC) was used to generate 2 layers that included demographic information related to the availability of drinking water (public network) (Fig 4h) and the distance to critical points (cemetery and garbage dump) (Fig 3g). First INDEC layer reflects the lack of this service, as an indicator of the use of containers for outdoor water storage, as potential generator of artificial Ae. aegypti breeding sites. INDEC information was transformed to a vector layer that included percentage values of the service per Census radio units by the use of QGIS software.
In order to assess the contribution of each of the selected variables to the prediction model, MaxEnt software [49] was used to predict suitable sites for the development of Ae. aegypti breeding sites in Tartagal, based on the environmental requirements of the species [19,50–52]. Ecological modelling calculates the probability of vector breeding sites presence using environmental and demographic variables, and the actual vector breeding sites presence as training sites. All the positive sites for Ae. aegypti larval stages from the sampling performed in Tartagal during 2014 were used. In order to analyze the possible relationship of a product that compiled all variables, values are generated. Thus, each pixel of the study area presents a landscape value (indicated by the set of variables, see Table 2) and may have an associated value indicating the probability. This analysis is based on two basic premises, i) the first one relates to the presence of sites where the species successfully grow, and the second ii) refers to the selected environmental variables that adequately represent the ecological requirements of the species. Each presence site is indicated by a pair of geographic coordinates (WGS84 Datum), and represents a place where Ae. aegypti breeding sites were found during the sampling period. The Maximum Entropy approach (MAXENT) was used to model and predict the ecological niche distribution of the vector. In general, this algorithm detects non-random relationships between two data sets: i) georeferenced records of the presence of the species, and ii) a set of land cover type "raster", digital data representing the environmental and demographic variables relevant to determine the distribution of the species in a particular scale of analysis [49]. The environmental data set consists of 19 variables in raster format, of 10 m pixel size (see Table 2 for data access to Tartagal). For the generation of vector presence probability maps, we applied the Maximum Entropy method based on the MaxEnt 3.3.3a software [49], available online at http://www.cs.princeton.edu/~schapire/maxent/, reserving 25% of Aedes aegypti presence points for validation, and with a 1000 repetitions run.
The temporal variation of Ae. aegypti spatial distribution of positive breeding sites presented a wide distribution all around Tartagal city(Figs 5 and 6), with the highest densities spatially concentrated in city outskirts, in comparison to the central areas (Fig 5). The temporal representation of positive sites in the city registered variations in distribution and number over the years, observing a similar spatial pattern as previously described (Fig 6).
Fig 7 shows that house indexes (HI and BI) values decreased between 2010 and 2014. Throughout this period, both indexes registered their highest levels during the summer and autumn seasons, which coincide with FC1, FC2, FC5, FC8, FC9 and FC10 (Table 1), while the lower values were associated with winter and spring (FC3, FC4, FC6, FC7, M1, M2, M3, M4 and M5) (Table 1 and Fig 8).
Sectors with presence of Ae. aegypti breeding sites were distributed throughout the entire study area, especially at the beginning of the program during 2010 and 2011 (Fig 8). Through the analysis of hotspots, three important aspects were detected: 1) the gradual reduction in the density of breeding sites detected each year, 2) the presence of sectors with a density of breeding sites that persist throughout the study period analyzed, located in the northeast, north, east and southeast regions of Tartagal, and 3) the highest density of breeding sites were associated with peripheral sectors while the lowest ones were registered in the central areas of the city (Fig 8).
In 2010, the northeast and southeast Tartagal sectors reached the highest density of breeding sites, with up to 20 positive breeding sites per housing unit. These two sectors remained positive throughout the study period although with variations in the density values (Fig 8). During 2011 and 2012, the north and east sectors of the city were identified as areas with high density of breeding sites (Fig 8). The year 2013 showed a spatial configuration that was similar to the previous years but with a marked reduction in breeding site density (Fig 8), while in 2014, the areas that registered the highest density of breeding sites were sectors located in the east and southeast (Fig 8).
Statistically significant (p < 0.05) differences were observed in spatial clusters between 2010 and 2014 (Fig 9). In general, the clusters with the largest dimensions were located in the northeast, north, east and southeast sectors of the city, with clusters that varied in size throughout the study period (Fig 9). In 2011 and 2014, the east sector registered the highest clusters with radius greater than 1.5 km (Fig 9). On the other hand, in 2010, 2012 and 2014 the southeast sector presents clusters with radius greater than 0.5 km. In the northeast sector, clusters with radius greater than 0.5 km were registered only during 2010 and 2013, while in the north sector these size clusters were only registered in 2013 (Fig 9). Another aspect observed using cluster analysis was the concentration of clusters in one or two sectors of the city for the year 2010 and 2014. For the rest of the years, numerous clusters of radius size between 100 and 300 m were registered throughout different sectors (Fig 9).
The predictive map obtained by the model (Fig 10) was assessed with measurement accuracy (MaxEnt software), therefore the area under the curve (AUC) in receiver operating characteristic (ROC) analysis was scored at 0.918. Its predictive ability for the 2014 data set is classified as acceptable according to Parolo and collaborators [53]. The model predicted that the environmental variables that best explain 75% of the distribution of Ae. aegypti breeding sites were: the percentage of bare soil (44,9%), percentage of urbanization (13,5%), and water distribution (11,6%) (Table 3).
In this study we have presented relevant results on the spatial pattern dynamics of Ae. aegypti breeding sites, in a small city in the north of Argentina (Tartagal). Although environmental sanitation activities and breeding sites chemical control, with larvicide, were performed after each entomological surveillance in the city, the sectors with higher densities of breeding sites remained present throughout the study period. Spatial analysis (Figs 8 and 9) and predictive risk map (Fig 10) results, allowed us to indicate, with some confidence, the difficulty of implementing control measures and dengue surveillance in urban areas such as Tartagal. The wide spatial data set used in the present study, to generate the predictive habitat quality map, enabled us to register that, although the predictors used to adjust the model explained 75% of the distribution of positive breeding larvae sites, they failed to indicate other areas containing breeding sites with high density values (Fig 9).
The spatial pattern of a dengue epidemics would be determined by multiple synergic factors that occur concurrently in the environment, including entomological, demographic and epidemiological factors, among others [6]. Consequently, one of the key aspects to reduce the abundance of the vector is associated with the identification of the breeding sites where Ae. aegypti lays its eggs which later develop into larvae [11]. The density (hotspots) and distribution maps of the breeding sites in Tartagal indicated a greater proportion of breeding sites in the outskirts of the city (Figs 5, 6 and 8). These areas are characterized by a deficit in the potable water supply, especially during the summer months, which consequently promotes the accumulation of a diverse array of containers for water storage in the peridomicile. Generally, most of these containers are maintained uncovered, constituting excellent breeding sites for Ae. aegypti. A previous study conducted in the city of Clorinda, in the neighboring Province of Formosa, where this phenomenon was also observed, concluded that the practices related to water accumulation are due to cultural patterns adopted to face the lack of access and availability of this critical element, constituting a complex set of factors that influence the abundance of breeding sites and the population dynamics of Ae. aegypti, jeopardizing the efficacy of vector control programs [54]. This availability of breeding sites, with higher and lower densities of larval stages, suggest a subsequent uneven abundance and distribution of adults, as was already seen in studies conducted in Colombia [31].
Analyses performed allowed the identification of breeding sites hotspots in specific sectors of the city (Figs 8, 9 and 10). This scenario allows us to assume a spatial pattern of vector presence, with areas that the favor the presence and abundance of Ae. aegypti. During the Dengue epidemic on 2004 in Tartagal, a study detected spatial groupings of confirmed cases of dengue [42], which showed that dengue cases were concentrated in city outskirts, in agreement with Ae. aegypti breeding sites distribution observed in the current study (Figs 8, 9 and 10). Therefore, since the presence of dengue cases and positive breeding sites coincide in space, surveillance and control activities based on vector data would allow the identification of low and high risk areas for dengue transmission, allowing the planning of activities according to risk probability. Other studies have already suggested that an increase in vector density could lead to an increase in the vector-human contact which would translate into a higher viral transmission rate among the population [55,56]. Therefore, the detection of changes in vector density presents itself as an important factor in the epidemiology of the disease [31]. Taking into consideration Ae. aegypti urban characteristic, directly associated with the presence of larvae and the number and type of containers suitable for larval development [56], a greater proportion of breeding sites supposes a greater abundance of mosquitoes which leads to an increased probability of bites. In this scenario, the inferred spatiotemporal relationship between dengue cases [19] and breeding sites detected in our study could offer information on the presence of hotspots of Ae. aegypti infestation in Tartagal, which are regularly maintained in space and time.
During this study, after each round of entomological surveillance, activities of physical and chemical (application of larvicide) control of breeding sites were performed, diminishing the availability of suitable breeding sites and therefore impacting on the level of viral transmission [12]. From 2009 to 2014, a reduction of Stegomyia indexes (both HI and BI) was observed, which would be associated with a lower density of positive breeding sites (Figs 7 and 8). This tendency would suppose a positive effect of the control activities performed in Tartagal during this period which resulted in a decreased number of registered breeding sites (Fig 8). Nonetheless, the distribution of the breeding sites showed a spatial dynamic with high densities in the outskirts of the city, at the same time presenting sectors were high values were always maintained and others that varied throughout the years (Figs 8 and 9). This observation could be associated with the different conducts with respect to the use and management of water, since the lack of piped water in the housing units forces inhabitants to store water in different types of containers. At the same time, there are different behaviors with respect to containers care that, together with eco-environmental characteristics (temperature, land cover, etc.), constitute favorable scenarios for the development of the mosquito.
The relationship between environmental and climatic conditions, and Ae. aegypti dynamic, is well known and has been shown to affect the abundance and distribution of the mosquito [32,57–61]. Regardless the complexity of variables that affect the distribution of anthropophilic vectors, such as Ae. aegypti, tools for spatial analysis and GIS applied in the current study allowed us to find spatial relations between the positive breeding sites and the variables derived from remote sensing satellite (SPOT 5), which confirmed the urban characteristic of the vector. The aptitude model generated for Tartagal (Fig 10) showed that the presence predictive power of Ae. aegypti breeding sites for test data was acceptable, in accordance with Parolo and collaborators [53]. The distribution map showed that the higher probabilities of vector presence, with values that could exceed 70%, would be associated with three sectors in the city: one located in the south, one in the northwest and one in the east. The three sectors identified as the most likely to have vector presence (Ae. aegypti breeding sites) are isolated and separated by sectors that show a low probability (0–30%), and best explained by variables that represent typical urban patterns (bare soil, urbanization and distance to water) (Table 3). In this regard, it is important to highlight the usefulness of the HRG Spot 5 images for the urban characteristics associated to the presence of positive vector breeding sites.
Although the model was acceptable, it had a limitation in the north sector of Tartagal, which is not indicated a high infestation area, even though it presented a high suitability for the occurrence of positive sites, according to density (Fig 7) and hotspots (Fig 8) analysis. This inconsistency allows us to question the suitability of our model. As seen in Fig 8, the aptitude model showed a distribution similar to the one registered in the hotspot map of positive sites for the year 2014, therefore, the model is reflecting a particular situation of the distribution of positive breeding sites. This means that the use of a single period of entomological data for the elaboration of the model raises certain concerns in the moment of thinking of a comprehensive tool for the prevention of the disease, highlighting the importance of implementing sustainable programs for the collection of data and the elaboration of risk maps, in order to be able to tailor vector control actions adapted to the local reality.
Cities supply most of the habitat characteristics required by Ae. aegypti [62]. Recent studies [45], in accordance with our own field data, indicate a direct association between the presence larvae and the number of containers suitable for larval development, which would derive in an increased risk of mosquito larval infestation as well as the presence of adults in a particular area [12]. In Tartagal, and in accordance with other studies [15,63,64], the information from remote sensing has been used with the objective to provide information about the type of land cover that would allow us to indirectly estimate conditions favorable for the presence of breeding sites and survival of the mosquito. Although the aptitude levels obtained in our model do not include sites which we could consider as high risk, it suggests a lack of precision in the model predictability. This is in agreement with what has previously been indicated [18], which is that it is not easy to find precise spatial information on dengue useful for modeling. This calls attention to the manner in which the information provided is used. The great quantity of field data collected and analyzed in Tartagal allows us to evaluate the predictive capacity of the model. Therefore, we agree with Lois and collaborators [18] in that the predictive models must use a large time series of local data (entomological and environmental), which would allow to model different scenarios to assess the risk of the temporary effect on the predictions to generate more efficient control actions.
The results presented in this study show how entomological information and geospatial data (RS / GIS) could be used to infer scenarios which could then be applied in epidemiological surveillance programs and strategies for dengue control. But on the other hand, we can obtain unreliable scenarios when using insufficient (in quality / spatial and temporal coverage) input data, as shown in this work (fig 10), using only one year sampling data. The considerations presented in our study could be used by those who develop predictive maps and by public health workers to identify and target high-risk areas for dengue transmission. The approaches generated in this study could contribute information to decision-makers that could improve health system responses and prevention measures related to vector control.
Population dynamics of Ae. aegypti observed in our study, during 5 years of continuous work, allowed us to evaluate and fine tune our control strategy in the local context. Therefore, in the future, we would have to consider certain variables not currently contemplated: 1) determine the periodicity of control actions in accordance to the operational capacity of the work groups, 2) provide a solution for closed or uncooperative housing units which escape control activities and constitute sources of re-infestation, and finally, 3) determine the volume of data necessary for the elaboration of a highly predictive model for dengue transmission.
|
10.1371/journal.pgen.1004915 | Transposable Elements Contribute to Activation of Maize Genes in Response to Abiotic Stress | Transposable elements (TEs) account for a large portion of the genome in many eukaryotic species. Despite their reputation as “junk” DNA or genomic parasites deleterious for the host, TEs have complex interactions with host genes and the potential to contribute to regulatory variation in gene expression. It has been hypothesized that TEs and genes they insert near may be transcriptionally activated in response to stress conditions. The maize genome, with many different types of TEs interspersed with genes, provides an ideal system to study the genome-wide influence of TEs on gene regulation. To analyze the magnitude of the TE effect on gene expression response to environmental changes, we profiled gene and TE transcript levels in maize seedlings exposed to a number of abiotic stresses. Many genes exhibit up- or down-regulation in response to these stress conditions. The analysis of TE families inserted within upstream regions of up-regulated genes revealed that between four and nine different TE families are associated with up-regulated gene expression in each of these stress conditions, affecting up to 20% of the genes up-regulated in response to abiotic stress, and as many as 33% of genes that are only expressed in response to stress. Expression of many of these same TE families also responds to the same stress conditions. The analysis of the stress-induced transcripts and proximity of the transposon to the gene suggests that these TEs may provide local enhancer activities that stimulate stress-responsive gene expression. Our data on allelic variation for insertions of several of these TEs show strong correlation between the presence of TE insertions and stress-responsive up-regulation of gene expression. Our findings suggest that TEs provide an important source of allelic regulatory variation in gene response to abiotic stress in maize.
| Transposable elements are mobile DNA elements that are a prevalent component of many eukaryotic genomes. While transposable elements can often have deleterious effects through insertions into protein-coding genes they may also contribute to regulatory variation of gene expression. There are a handful of examples in which specific transposon insertions contribute to regulatory variation of nearby genes, particularly in response to environmental stress. We sought to understand the genome-wide influence of transposable elements on gene expression responses to abiotic stress in maize, a plant with many families of transposable elements located in between genes. Our analysis suggests that a small number of maize transposable element families may contribute to the response of nearby genes to abiotic stress by providing stress-responsive enhancer-like functions. The specific insertions of transposable elements are often polymorphic within a species. Our data demonstrate that allelic variation for insertions of the transposable elements associated with stress-responsive expression can contribute to variation in the regulation of nearby genes. Thus novel insertions of transposable elements provide a potential mechanism for genes to acquire cis-regulatory influences that could contribute to heritable variation for stress response.
| Transposable elements (TEs), first described as “controlling elements” by Barbara McClintock [1], are now known to make up the majority of angiosperm DNA [2]–[4]. TE insertions within genes may result in mutant alleles by changing the reading frame or splice pattern, frequently negatively affecting gene function. However, TEs also have the potential to contribute to regulation of gene expression, potentially playing an important role in responses to environmental stress [2], [5]; McClintock initially referred to TEs as “controlling elements” based on their ability to influence the expression of nearby genes [1], [6]. Several specific examples of TE influence on the expression of nearby genes have now been documented (reviewed by [7]–[11]). TE insertions near genes may influence gene expression through several potential mechanisms, including inserting within cis-regulatory regions, contributing an outward reading promoter from the TE into the gene [12]–[15], or providing novel cis-regulatory sequences that can act as enhancers/repressors by facilitating transcription factor binding [16], or influencing the chromatin state of gene promoter regions [17]–[19].
Some TEs exhibit stress-responsive transcription or movement [20]–[25]. For example, expression of the tobacco Tnt1 element can be induced by biotic and abiotic stress [22]–[23]. The rice DNA transposon mPing can be activated in response to cold and salt stress [26]–[27]. The Arabidopsis retrotransposon ONSEN is transcriptionally activated by heat stress [16], [28]–[29]. Tissue culture is a complex stress that can result in the activation of DNA transposons in maize and retrotransposons in rice [30]–[31]. There is also evidence that some of these TE responses to environmental conditions can affect the expression of nearby genes. Novel mPing MITE insertions in the rice genome in some cases resulted in up-regulation of nearby genes in response to cold or salt stress with no change in expression in control conditions [26]–[27]. The ONSEN retrotransposon insertions near Arabidopsis genes exhibit similar properties: alleles containing ONSEN insertions often show heat-responsive regulation while alleles lacking ONSEN are not up-regulated by heat stress [16]. These studies suggest that TEs can provide novel regulatory mechanisms and influence the response to environmental stress.
Maize provides a good system for studying the potential influence of TEs on regulation of nearby genes. While TEs only account for ∼10% of the Arabidopsis genome [32] or ∼32% of the rice genome [33], they contribute ∼85% to the maize genome [34]–[35]. Many TEs are located in pericentromeric regions and heterochromatic maize knobs [34], [36], but there are also many TE insertions interspersed between maize genes [37]–[39]. The majority of maize genes (66%) are located within 1 kb of an annotated transposon [35]. In addition, allelic variation for the presence of TE insertions near genes is high in maize [39]–[41], creating the potential for allelic regulatory differences at nearby genes. For example, polymorphic TE insertions in different haplotypes of the tb1, Vgt1 and ZmCCT loci likely contribute to regulatory differences for these genes [42]–[44].
While there are good examples to suggest that specific TEs can influence the response of nearby genes to abiotic stress [16], [26] it remains unclear how widespread this phenomenon is, how many genes are activated in such a TE-dependent manner, and whether multiple TE families are capable of controlling stress response. We identified a subset of TE families over-represented in the promoters of maize genes that exhibit stress-responsive up-regulation or activation of gene expression. Based on our data, as many as 20% of genes that showed increased expression in response to stress are located near a TE from one of these families. We find that stress-responsive TEs appear to provide enhancer-like activity for nearby promoters and allelic variation for TE insertions is strongly associated with variation in expression response to stress for individual genes.
We extracted and sequenced RNA from 14 day old seedlings of inbred lines B73, Mo17 and Oh43 grown using standard conditions as well as seedlings that had been subjected to cold (5°C for 16 hours), heat (50°C for 4 hours), high salt (watered with 300 mM NaCl 20 hours prior to collection) or UV stress (2 hours) (see Materials and Methods for details). For each stress the plants were sampled immediately following the stress treatment and there were no apparent morphological changes in these plants relative to control plants. However, when the stressed plants were allowed to recover for 24 hours under standard conditions phenotypic consequences became apparent for several of the stress treatments (Fig. 1A–B). RNAseq data was generated for three biological replicates for cold and heat stress and one sample for the high salt and UV stress (SRA accessions and read number for each sample are provided in S1 Table). Differentially expressed genes (RPKM>1 in control or stressed samples, padj<0.1 in DESeq [45] analysis, and minimum of 2-fold change in stress compared to control) were identified in control relative to cold or heat treated plants for each genotype using both the filtered gene set (FGS) and working gene set (WGS) genes (S2 Table). For each stress by genotype combination we found that 18%–30% of the expressed genes (7–10% of all genes) exhibit significant changes in expression level with similar frequencies of up-and down-regulated expression changes (S2 Table). For the salt and UV stress we identified genes that exhibit at least 2-fold change in expression and RPKM>1 in at least one of the conditions. The analysis of data for heat/cold stress revealed that the genes identified as differentially expressed based on a single replicate of this data had>90% overlap with the genes identified as significant in the analysis of multiple replicates. The clustering of gene expression responses to abiotic stress suggests that each stress has a substantial influence on the transcriptome (Fig. 1C). The majority of genes that are differentially expressed exhibit low to moderate expression in the control condition (S1 Fig.). While all three inbred lines showed similar transcriptional responses to the stress conditions there is also evidence for genotype-specific responses (Fig. 1C). The expression for TE families was also compared in stress and control conditions by determining the reads per million (RPM) that mapped to 353 TE families that had insertions located near maize genes. A subset (3%–20%) of the TE families are 2-fold up- or down-regulated in response to specific abiotic stress conditions (S2 Fig., S3 Table).
To test the hypothesis that genes responding to abiotic stress may be influenced by nearby TE insertions we focused our initial analyses on expression responses in the inbred B73, for which a reference genome is available [35]. The TEs located within 1 kb of the transcription start site (TSS) of each gene were identified in the B73 reference genome. For each of 576 annotated TE families we determined whether genes located near the transposon were significantly enriched (p<0.001,>2 fold-enrichment and at least 10 expressed genes associated with the TE family) for responsiveness to each of the stress conditions (separate analyses for enrichment in up- or down-regulated genes for each stress) relative to non-differentially expressed genes (S3 Table). While the majority of transposon families are not associated with stress-responsive expression changes for nearby genes (Fig. 2A–B; S3 Table), 20 TE families are significantly enriched for being located near genes with stress-responsive up-regulation and 3 TE families are associated with genes down-regulated in response to stress (Fig. 2C; Table 1).
Examples of the expression changes for genes in different abiotic stresses are shown for two transposon families, ipiki and etug (Fig. 2D). Genes located near ipiki are enriched for up-regulation following salt and UV stress while genes located near etug elements are enriched for heat-responsive up-regulation. Another striking example is the joemon TE family for which 59 of 68 expressed genes containing an insertion within 1 kb are activated following cold stress (Table 1). Although similar numbers of genes exhibit increased and decreased gene expression genome-wide following abiotic stress conditions, the majority of enriched TE family – stress combinations (28/31) are associated with up-regulated gene expression. For each of the stress conditions there were 4–9 TE families that are associated with up-regulation of gene expression. Some TE families are associated with altered expression in multiple stress treatments (Table 1, S4 Table; Fig. 2C) and two of the TE families associated with down-regulation of gene expression under high salt stress were also associated with increased gene expression under UV stress.
The TE families enriched for genes activated in response to stress include all major super-families of TEs: TIR DNA transposons, LTR gypsy-like (RLG), copia-like (RLC), or unknown (RLX) retrotransposons, and LINE elements (Table 1,). These TE families vary substantially for the number of genes that they are located near: from 30 to 3052 genes (Table 1; S4 Table) and are spread uniformly across the maize genome. The presence of these TEs near genes is not fully sufficient for stress-responsive expression. For each of the TE families identified, 26–87% of genes located near a TE insertion show stress responsive expression depending on the stress and the TE family. The expression levels for the TEs themselves was assessed for each of the treatments and in the majority of TE family – stress combinations (14 of 21 with expression data) the TEs showed at least 2-fold increase in transcript levels in the stress treatment compared to control conditions (Table 1, S4 Table). There are several examples of TE families that exhibit increased levels of expression in a particular stress but the nearby genes are not enriched for stress-responsive expression (S3 Table), suggesting that not all TEs that are influenced by a particular stress influence nearby genes.
To understand what proportion of the transcriptome response to a specific abiotic stress may be explained by influences of specific TEs inserted near genes, up-regulated genes were classified according to whether they were located near a member of one of the stress-associated TE families (1 kb 5′ from TSS) and whether they are up-regulated (expressed under control and stress conditions) or activated in response to stress (only expressed following stress treatment). We found that a substantial portion of the transcriptome response to the abiotic stress could be associated with genes located near the set of 4–9 TE families that were identified as enriched for up-regulated genes (Fig. 2E). In total, 5–20% of the genome-wide transcriptome response to the abiotic stress and as many as 33% of activated genes could be attributed to the genes located near one of these TE families (Fig. 2E; S5 Table-6).
One possible mechanism by which these families of TEs could contribute to stress-responsive expression for nearby genes is that the TE may provide an outward-reading promoter that is stress-responsive. This model predicts that the orientation of the TE relative to the gene is important and that novel transcripts containing TE sequences fused to gene sequences would be present for up-regulated genes under stress conditions. In order to assess the importance of the orientation of the TE insertion relative to the gene, we compared the proportion of genes located on the same strand as a TE for genes up-regulated in response to stress and genes non-differentially expressed in response to stress for all TE families enriched for up-regulated genes (S7 Table). While most families showed no significant difference in the proportion of genes on the same strand as the TE between the up-regulated and non-differentially expressed genes, a minority of families (4/20) showed significant enrichment. For example, 97% of the stress-responsive genes located near etug elements are on the same strand as the TE (S7 Table).
The potential for TEs to provide novel promoters in stress conditions was also assessed by creating de novo transcript assemblies for each of the treatment conditions (S8 Table). These transcript assemblies were mapped to the reference genome to determine the transcriptional start sites in control and stress treatments. In particular, we focused on the 630 genes that had TE insertions at least 100 bp 5′ of the annotated transcription start site that had de novo transcript assemblies in both control and stress conditions. The location of the start site for the transcript assembly in control and stress conditions was compared to the location of the annotated start site and the location of the TE. There were a number of instances in which the transcript start site was located 5′ of the annotated site in both control and stress conditions and these likely reflect examples of incomplete annotation. There are 16 genes (out of 630 with data) that have a novel start site in the stress-treatment and not in the control that was located within or near the TE. There was not a significant enrichment for specific TE families among these 16 examples and these examples may simply reflect examples of inaccurate start site annotation without enough read depth in the control condition to identify the specific start site. These examples show that we could detect novel start sites but they suggest that it is rare for TEs to provide novel promoters in stress conditions.
Alternative models include the possibility that the TE may contain cis-regulatory sequences that can act as binding sites for stress-induced transcription factors, or that the TE could influence the local chromatin environment in such a way that the region is more accessible under stress conditions. The analysis of TE distance from transcription start sites of stress-responsive genes suggests that in many cases the effect of TE on stress-responsive gene activation quickly diminishes as the distance increases beyond 500 bp – 1 kb (S3A Fig.). The DREB/CBF transcription factors are often involved in transcriptional responses to abiotic stress in plants [46]. The consensus sequence for DREB/CBF binding (A/GCCGACNT [47]) was found in most of the TEs that were associated with stress-responsive expression for nearby genes, with the exception of elements that only exhibit UV stress response (S3B Fig.). While we did not have evidence to distinguish between the possibilities that TEs provide either a sequence-specific binding site that might act as a stress-specific enhancer or influence the chromatin state in a non-sequence specific manner, our data are consistent with the TE insertions acting predominantly as local enhancers of expression rather than as novel promoters.
Because individual TE copies are subject to frequent rearrangements and internal deletions, we investigated whether the presence of specific regions in each TE family were over-represented in insertions that confer stress-responsive expression. For six of the 20 TE families, this comparison revealed specific portions of the TE sequences enriched among insertions that convey stress-responsive expression. For example, naiba and etug insertions located near up-regulated genes are approximately four times as likely to contain a particular portion of the TE long terminal repeat (LTR; p-value <0.001; S4 Fig.), and this same sequence is found in a subset of insertions of the related family, gyma, that are associated with up-regulated genes. While we did not have evidence to rule out the possibility that TEs influence the chromatin state in a non-sequence specific manner, these data indicate that the presence of particular regions of TE elements likely provide enhancer functions associated with gene expression responses to stress and help explain the variable effect of different insertions of the same family on stress-responsive expression.
We assessed a number of properties of the TE-influenced stress-responsive genes in comparison with stress-responsive genes that are not associated with one of these TE families (Table 2). Stress-responsive genes located near the TE families tend to be substantially shorter in length with fewer introns. Analysis of developmental expression patterns for these genes using the B73 expression atlas [48] reveals that only 7% of the TE influenced genes are expressed in at least 5 tissues, compared to 41% of the non-TE influenced genes. The TE influenced genes are also less likely to be in the filtered gene set (FGS), and the proportion of the TE influenced genes with syntenic homologs in other grass species is much lower than the proportion of non-TE influenced genes (Table 2). Each of these features was assessed separately for each of the TE families (S7 Table) and there is some variation for these properties among different families. These observations are compatible with the notion that TE insertions may in some cases function as enhancers that can drive expression of cryptic promoters in non-coding regions of the genome. This will result in stress-responsive production of transcripts that may be annotated as genes but may not produce functional proteins. However, 37% of TE influenced genes are included in the FGS that has been curated to remove transposon-derived sequences and a substantial proportion of the TE influenced genes are syntenic with genes from other species, have GO annotations, and could contribute to functional responses to stress (Table 2, S7). These results suggest that many of TE influenced genes are not derived from TEs.
We were particularly intrigued by the question of whether polymorphic insertions of TEs from families associated with stress-responsive expression of nearby genes might contribute to allelic variation for stress-responsive gene expression. The consistency of stress-responsive expression of TE-associated genes across the three inbred lines surveyed varied widely across TE families (Fig. 3A; S5 Fig.). In order to assess whether insertions of TEs from the families associated with stress-responsive gene expression could contribute to allelic variation for gene expression regulation, we used whole-genome shotgun re-sequencing data from Mo17 and Oh43 [49] to find potential novel insertions of elements from the TE families identified in this study. We identified 23 novel (not present in B73) high-confidence insertions of TEs from these families located within 1 kb of the TSS of maize genes and validated them by PCR (S9 Table). Of the 10 genes with detectable expression in our RNAseq experiments, 7 showed stress-responsive up-regulation/activation associated with the TE-containing alleles (Fig. 3B). This analysis was expanded to additional genotypes by using PCR to detect the presence/absence of the TE insertion in a diverse set of 29 maize inbred lines that were selected to represent diverse North American germplasm from the stiff stalk, non-stiff stalk, iodent, tropical, sweet corn and popcorn population groups. The relative expression of the gene in stress compared to control treatment was also determined in each inbred using quantitative RT-PCR (S10 Table). For each of these genes we found that the alleles that lack the transposon insertion did not exhibit stress-responsive expression (Fig. 4), with the exception of one genotype for gene GRMZM2G108057. In contrast, the majority of the alleles that contain the TE (60–88%) exhibit stress-responsive up-regulation. Although for a single insertion we cannot rule out the possibility that differential expression is due to a different polymorphism on the same haplotype as the TE, the fact that we see TE-associated expression change in multiple genes for each of the TE families (Table. S10) argues strongly against such an explanation in general. These data thus provide evidence that insertion polymorphisms for the TE families identified here can generate novel expression responses for nearby genes.
Transposable elements are a major component of many eukaryotic genomes, and constitute the majority of plant nuclear DNA. TEs are usually considered as a deleterious or neutral component of these genomes. However, the interplay between TEs and genes may have important functional contributions to plant traits. There are clear examples of TE insertions that are linked to functionally relevant alleles in maize such as Tb1 [42] Vgt1 [43] and ZmCCT [44]. In these cases, a transposon insertion within a distant cis-regulatory sequence influences the regulation of adjacent genes. There are also examples of functionally relevant TE insertions in tomato, melons and citrus [50]–[52] that can influence gene expression, potentially through chromatin influences that generate obligate epialleles.
Previous research in several plant species has suggested that at least some families of transposable elements may become transcriptionally activated following environmental stress. Tissue culture has been shown to result in activation of transposons and retrotransposons in a number of plant species [30]–[31]. There are also examples of transcriptional activation of TEs in response to specific abiotic stresses in tobacco [22], rice [26]–[27] and Arabidopsis [16], [28]–[29]. It is expected that the stress responsive expression of these TEs involves local enhancers that result in up-regulation of the TE promoter in response to stress. These local enhancers could also act upon other nearby promoters. There are a handful of examples in which transposon insertions have been linked to stress-responsive expression of nearby genes including the mPING insertions associated with cold-responsive expression in rice [26]–[27] and ONSEN insertions associated with heat-stress responsive expression in Arabidopsis [16]. If this is a common occurrence then we might expect it to be even more prevalent in a genome such as maize where many genes are closely surrounded by TEs.
Our analysis suggests that a small number of TE families are associated with stress-responsive expression for nearby genes. While some TE families were associated with multiple stresses, we found a different subset of TE families for each abiotic stress that was evaluated. In most cases, these same TEs themselves were up-regulated in response to the stress treatment. However, we also noted that there were some TE families that themselves exhibit strong up-regulation but did not have apparent influences on a significant portion of nearby genes. Even though the majority of stress responsive regulation of gene expression is not associated with TEs, based on our data, up to 20% of genes up-regulated in response to stress and as many as 33% of genes activated in response to stress could be attributed to regulation by TEs. One of the alternative explanations would argue that only a small number of genes localized close to a TE are truly influenced by this TE insertion for their expression, while other up-regulated genes are secondary targets and are regulated by the TE influenced genes. Although some of the TE influenced genes we identified could be secondary targets, secondary target genes would not preferentially co-localize with TEs from specific families.
The analysis of the nearby genes that were influenced by TEs suggests that many of them may not actually be protein coding genes. In one sense, this is an expected result. If an enhancer sequence is mobilized within the genome it will have the potential to influence expression from both gene promoter as well as cryptic promoters that may not be associated with coding sequences. The gene annotation efforts in maize have relied upon EST and RNA-seq expression data from a variety of conditions. In many cases the genes that were found to exhibit stress-responsive expression associated with TEs were only annotated as genes based upon evidence of their expression. We would expect that insertions of the TEs that provide stress-responsive enhancer activity would influence cryptic promoters not associated with genes in many cases, but would also affect the expression of nearby protein coding genes. The frequency of each appeared to vary among TE families, with some, like nihep, showing little difference between TE-influenced and non-TE-influenced up-regulated genes (S7 Table). Overall, while TE influenced stress-responsive genes are enriched for short sequences with limited homology to sequences in other species, a significant proportion are longer, have several exons, are conserved in other species, and have GO annotations.
A particularly interesting aspect of these results is the potential mechanism for creating novel cis-regulatory variation. Our understanding of how particular genes might acquire novel regulatory mechanisms is limited. In many cases SNPs within promoters or regulatory sequences have limited functional significance. Therefore, it is difficult to envision how a novel response to a particular environmental or developmental cue would arise. Variation in TE insertions has the potential to create novel regulatory alleles by providing binding sites for transcription factors or influencing chromatin. We provide evidence that allelic variation for stress-responsive expression can be created by the insertion of certain TEs. Variation in TE insertions would generate allelic diversity that could influence an organism's response to environmental conditions and would provide phenotypic variation that could be acted upon by selection. As with other types of variation, most examples of novel stress-responsive expression are likely to be neutral or deleterious and would not be expected to rise in allele frequency. However, a subset of novel stress-responsive expression patterns could be beneficial and become targets of natural or artificial selection contributing to gene regulation networks of environmental stress response.
B73, Mo17, and Oh43 maize seedlings were grown at 24°C in 1∶1 mix of autoclaved field soil and MetroMix under natural light conditions in July 2013. For cold stress, seedlings were incubated at 5°C for 16 hours. For heat stress, seedlings were incubated at 50°C for 4 hours. For high salt stress, plants were watered with 300 mM NaCl 20 hours prior to tissue collection. UV stress was applied in the growth chamber conditions using UV-B lamps for 2 hours prior to tissue collection. UV stress causes accumulation of DNA mutations but most of such mutations would either have no immediate effect on gene expression or would lead to decrease or abortion of expression of specific genes. Light conditions were the same for all stress and control conditions. Whole above ground tissue was collected for 14 day old seedlings at 9am and six seedlings were pooled together for each sample. Three replicates for heat and cold-treated B73 and Mo17 seedlings were grown 3 days apart.
Three biological replicates of cold and heat stress and control conditions for B73 and Mo17 were prepared with eight plants pooled for each of the replicates. One biological replicate of high salt and UV stress conditions for B73 and Mo17 as well as all four stress and control conditions for Oh43 were prepared similarly. RNA was isolated using Trizol (Life Technologies, NY, USA) and purified with LiCl. All RNA samples were prepared by the University of Minnesota BioMedical Genomics Center in accordance with the TruSeq library creation protocol (Illumina, San Diego, CA). Samples were sequenced on the HiSeq 2000 developing 10–20 million reads per sample. Transcript abundance was calculated by mapping reads to the combined transcript models of the maize reference genome (AGPv2) using TopHat [53]. Reads were filtered to allow for only uniquely mapped reads. A high degree of correlation between replicates was observed (r>0.98). RPKM values were developed using ‘BAM to Counts' across the exon space of the maize genome reference working gene set (ZmB73_5a) within the iPlant Discovery Environment (www.iplantcollaborative.org). Genes were considered to be expressed if RPKM>1 and differentially expressed if log2(stress/control)> 1 or log2(stress/control) <-1. Statistical significance of expression differences was determined using DeSeq package for all fully replicated samples [45].
For each gene, transposons located within 1 kb of the transcription start site (TSS) were identified using the B73 reference genome annotation [35] and maize TE elements database [34]. TE distance from transcription start sites was determined using the closestBed tool from the BEDTools suite [54] where TEs upstream were given a positive distance value and TEs downstream were given a negative distance value. The transcriptional start site was defined as the 100-bp window intersecting the first base pair of a gene model from the maize genome gene set (ZmB73_5b). The proportion of up-regulated, down-regulated, and non-differentially expressed genes that have an insertion of a TE element from a particular family was calculated for 576 TE families for four stress conditions. Fold-enrichment of up-regulated genes relative to all expressed genes (the sum of up-regulated, down-regulated and non-differentially expressed genes) and relative to all genes was calculated for all TE family/stress combinations. Given the total number of expressed genes associated with each TE family and the proportion of up- and down-regulated genes, the expected numbers of up- and down-regulated genes and non-differentially expressed genes were calculated and a multinomial fit test was conducted. TE families that had over 10 expressed genes associated with them, fold enrichment of up- or down-regulated genes over 2, and p value <0.001 were considered “enriched” for up- or down-regulated, respectively. Similar analysis was conducted for working gene set and filtered gene set genes. The same set of “enriched” TE families was found for both groups of genes as well as when fold enrichment was calculated relative to all expressed genes or to all genes associated with TEs from a particular family.
To assess expression changes in response to stress for TE families, the overlap tool from BEDTools suite [54] was used to obtain read counts per each TE accession. The output file from alignment (BAM) was mapped to TE positions listed in the TE GFF file downloaded from maizesequence.org. Each read was required to have 100% overlap with a given TE region. The reads mapping to more than 5 locations in the genome were omitted. The reads were then summed across the entire TE region and combined for each of the TE families.
Tissue specific expression data is from the maize gene expression atlas [47]. Genes with RPKM of <1 were considered non-expressed. Orthologous and paralogous gene pairs were inferred from [55].
De novo assemblies for the control and each stress were performed for the B73 inbred line. Prior to assembly reads were cleaned using cutadapt version 1.4.1 [56] requiring a minimum read length of 30. Reads were further cleaned with the FASTX toolkit version 0.0.14 (http://hannonlab.cshl.edu/fastx_toolkit/) using the fastx_artifacts_filter and the fastq_quality_trimmer requiring a minimum read length of 30 and a minimum quality score of 20. Read pairs for which one read was discarded during the read cleaning pipeline were discarded from further analyses. Within each treatment all reads across biological replicates were combined for treatment specific assemblies. The transcriptome assembly was conducted using Trinity version r20140413 [57] using default parameters and requiring a minimum transcript length of 200. Each assembly was assessed based on the percentage of transcripts that could map back to the reference genome sequence and the percentage of input reads that could map to the final assembly. Transcripts were mapped to the maize v2 reference genome sequence (http://ftp.maizesequence.org) using GMAP version 2012-06-02 [58] with default parameters. Input reads were mapped back to the assembly using Bowtie version 0.12.9 [59] and TopHat version 1.4.1 [53] allowing a minimum and maximum intron size of 5 and 10,000 and the —no-novel-indels function. Assemblies were linked to stress differentially expressed genes based on the GMAP alignments. The start position for control and stress assembled transcripts were compared to identify transposable elements that act as either promoters or enhancers under stress conditions. Instances where the control assembled transcript starts within the gene model and the stress assembled transcript starts near or within the TE would provide evidence that the TE is acting as a promoter.
Nonreference TE insertions were detected for Oh43 and Mo17 using relocaTE [60], whole genome sequence from the NCBI SRA (Oh43: SRR447831-SRR447847; Mo17: SRR447948-SRR447950), and consensus TE sequences from the maize TE database [34]. Reads containing TEs were identified by mapping to consensus TE sequences, trimming portions of reads mapping to a TE, and mapping the remaining sequence to the reference genome. Nonreference TEs were identified when at least one uniquely mapped read supported both flanking sequences of the nonreference TE, overlapping for a characteristic distance that reflects the target site duplication generated upon integration (five nucleotides for all LTR retrotransposons, nine nucleotides for DNA TIR mutator). Primers for six TE polymorphic genes up-regulated under stress conditions in Oh43 or Mo17 but not in B73 were designed using Primer 3.0 software [61] and PCR reactions were performed using Hot Start Taq Polymerase (Qiagen, Ca, USA). Primer sequences are shown in S11 Table.
cDNA synthesis and qPCR analysis were performed as described in [62]. Primers for 10 differentially expressed genes and two control genes (GAPC and mez1) were designed using Primer 3.0 software [57]. Primer sequences are shown in S10 Table.
|
10.1371/journal.pbio.1000585 | Speeding Up Microevolution: The Effects of Increasing Temperature on Selection and Genetic Variance in a Wild Bird Population | The amount of genetic variance underlying a phenotypic trait and the strength of selection acting on that trait are two key parameters that determine any evolutionary response to selection. Despite substantial evidence that, in natural populations, both parameters may vary across environmental conditions, very little is known about the extent to which they may covary in response to environmental heterogeneity. Here we show that, in a wild population of great tits (Parus major), the strength of the directional selection gradients on timing of breeding increased with increasing spring temperatures, and that genotype-by-environment interactions also predicted an increase in additive genetic variance, and heritability, of timing of breeding with increasing spring temperature. Consequently, we therefore tested for an association between the annual selection gradients and levels of additive genetic variance expressed each year; this association was positive, but non-significant. However, there was a significant positive association between the annual selection differentials and the corresponding heritability. Such associations could potentially speed up the rate of micro-evolution and offer a largely ignored mechanism by which natural populations may adapt to environmental changes.
| The speed of evolutionary change in a phenotypic trait is determined by two key components: the amount of genetic variance underlying the trait and the strength of selection acting on it. Many studies have shown that both selection and expression of genetic variance may depend on the environmental conditions the population experiences. However, the possibility that the strength of selection and the expression of genetic variance become positively or negatively associated as a result of this environmental covariance, so as to speed up or hamper an evolutionary response, has been largely ignored. Here we show that, in a wild bird population, the annual strength of selection on and the expression of genetic variance in timing of breeding (a key life history trait) are positively associated due to changing environmental conditions (warmer temperatures). Such a positive association should potentially speed up any microevolutionary response to selection (such as that imposed by climate warming). Our results illustrate the existence of substantial temporal variation in response to environmental heterogeneity, and thus highlight a so far neglected mechanism that may be important in determining the evolutionary dynamics in natural populations.
| Predicting an evolutionary response to selection in a phenotypic trait requires knowledge of the strength of selection acting on the trait and its genetic basis. Although it has long been recognized that the strength, and direction, of selection may vary with environmental conditions (e.g., [1]), widespread recognition of the fact that additive genetic variance (and thus heritability) may also change with environmental conditions has been more recent [2],[3]. Taken together, these observations generate an expectation of an environmentally driven association between the two parameters that, in theory, has the potential to either enhance (positive association) or constrain (negative association) any response to selection. Surprisingly, however, to our knowledge only one study to date has quantified the association between annual estimates of selection and expression of genetic variance (measured as heritability) in a heterogeneous environment [4]. In this article, we present data from a long-term study of a great tit (Parus major) population known to be experiencing substantial shifts in climatic conditions, and test for the effects of the novel environmental conditions on the expression of additive genetic variance, and the selection on, a key life history trait, breeding time.
Many studies have found that selection is often strongest when environmental conditions are adverse (e.g., [4]–[8]), and there is a clear indication that “perturbed or stressed” populations have larger standardized selection differentials than “undisturbed” populations ([9] p. 208). For example, Garant and co-workers [5] examined selection on fledgling body mass in a population of great tits and found that selection differentials were greater in years when average body mass was low and when the proportion of individuals surviving to recruitment was low, both indicative of poor/adverse environmental conditions. In general, therefore, selection is often stronger when environmental conditions are adverse.
Unlike the general tendency for selection to be stronger in adverse environments, conclusions regarding the effects of good versus adverse environments on the expression of additive genetic variance are more mixed. Laboratory studies investigating the effect of environmental conditions have generally found a weak tendency for heritability to increase in stressful environments with this being caused by changes in both the expression of genetic variance as well as the environmental variance (reviewed in [10]). This pattern, however, is in contrast to most studies from natural populations that find, at least for morphological traits, that additive genetic variance and heritability is often relatively lower in unfavorable conditions [3],[10],[11].
It is important to realize that heritability (h2) may change under different environmental conditions either because of changes in additive genetic variance (VA) or other variance components (e.g., permanent environmental variance (VPE) or residual variance (VR)). However, changes in VA are of particular interest because they indicate a change in the “evolvability” [12], or the potential to respond to selection, of a trait. Furthermore, changes in VA can only be due to a change in the genetic architecture of a trait through mechanisms such as genotype-environment interactions, changes in mutation and recombination rates, and removal of alleles with low fitness by selection (reviewed in [10]). Depending on the direction and scale of these changes, both additive genetic variance and heritability may increase or decrease depending on the relative impact of each of the above factors [10].
The possibility that both the expression of additive genetic variance of a trait and the strength of selection acting on it may vary with environmental conditions is significant, as such environmentally induced variation may be important in determining the evolutionary dynamics of natural populations. In particular, the observation of a general increase in genetic variance of morphological traits [3],[10],[11] and a reduction in selection [4],[6] during favorable conditions in natural populations leads to the expectation of a negative relationship between genetic variance and the strength of selection, such that selection should be strongest in years in which the expression of additive genetic variance is least. This association could severely constrain a response to selection and provide one explanation for the frequently observed scenario of apparent stasis in natural populations [4]. However, in contrast to morphometric traits, life history traits do not appear to show a clear indication of increased heritability in stressful environments [3]. This makes it more difficult to predict how, or if, additive genetic variance and selection on life history traits may covary in a heterogeneous environment.
Surprisingly, despite the potential importance of environmentally induced associations between the strength of selection and expression of genetic variance, we are aware of only one previous study that has tested for such an association. Wilson et al. [4] found that the strength of selection on body weight in a free-living population of Soay sheep (Ovis aries) in a given year was negatively correlated with the expression of total genetic variance (assessed via the heritability) of body weight, suggesting a possible constraint on the potential for evolution of body weight in this species. However, so far no study has, to our knowledge, examined the association between strength of selection and VA (or h2) in a life history trait. Hence, we do not know if such relationships are common in nature, and whether they are generally negative, which may constrain an evolutionary response, or whether there are examples of positive associations between strength of selection and VA (or h2), which would speed up an evolutionary response.
Here we use data from an exceptionally long-term study population of great tits (Parus major) in the Hoge Veluwe, the Netherlands, to investigate how selection and expression of additive genetic variance of a key life-history trait (timing of breeding, or “laying date”) vary in relation to rapid changes in environmental conditions (spring temperature). The evolutionary response in a trait between generations can be predicted as R = VA * β [13],[14], where β is the selection gradient, defined as the covariance between relative fitness and trait value divided by the phenotypic variance in the trait (i.e., β = cov(ω,trait)/VPtrait) [15]; we therefore test the association between VA and the selection gradients β under different environmental conditions. We also consider the alternative format for predicted response, R = h2 * S [16], where S is the selection differential, defined as the covariance between relative fitness and trait value (i.e., S = cov(ω,trait)) and test for an association between heritability and selection differentials.
This system is particularly well suited to an exploration of the association between selection and VA in a variable environment because phenotypic data, pedigree data, and a thorough understanding of how environmental conditions influence laying date are available [17],. Previous studies in this population have reported a significant increase in spring temperature over the past four decades [18] and have also shown that warm spring temperatures lead to earlier laying dates [17]. Furthermore, warmer temperatures lead to reproduction being mistimed relative to the food peak [17], resulting in a decrease in both the number and size of fledglings [19], and in the proportion of females producing a second clutch [20]. Spring temperatures are thus not only directly related to observed variation in laying dates but can also be used as a measure of environmental quality in the population. In addition, spring temperatures are now significantly above those which the population has previously experienced [18], providing an ideal opportunity to study how novel environmental conditions may influence evolutionary dynamics. We therefore tested the temperature dependence of the selection gradients and differentials, how expression of additive genetic variance and heritability changed with temperature, and finally, how the measures of selection were associated with the amount of genetic variance present in the population.
We found, firstly, strong selection on laying date, with early breeding birds having higher fitness than late breeding individuals (Table 1). Indeed, 29 out of the 35 estimates of annual selection gradients and differentials were negative (Figure 1, Table S2), reflecting general selection for earlier breeding, as has previously been shown in this population [17],[21]. Secondly, the interaction between laying date and standardized spring temperature was significantly negative (Table 1), indicating that with increasing spring temperatures the relationship (slope) between fitness and laying date became more negative (i.e., slope steeper in warmer years). Consequently, selection for early breeding was significantly stronger (indicated by more negative values of β) in warm years than in cold years; i.e., the strength of selection on lay date varied with environmental conditions (Figure 1). This result was confirmed by regressing the annual selection gradients (β) against temperature: there was a significant increase in the (absolute) magnitude of the strength of selection with increasing temperatures (regression slope = −0.044, se = 0.019, t33 = −2.203, p = 0.035, Figure 1a). The results were the same for selection differentials (regression slope = −1.589, se = 0.450, t33 = −3.529, p = 0.001, Figure 1b).
Comparing a model in which the additive genetic and permanent environment components of variance (VA and VPE) in a given year were constant across different spring temperatures to one in which VA and VPE could vary with the temperature gave strong support for environmental dependence of VA and VPE (χ24 = 74.90, p<0.001). Consequently, we used the predictions from the model in which the two variance components varied with spring temperature to generate estimates of annual VA and h2 and to explore how these annual estimates corresponded to the observed changes in the strength of selection on laying date.
The estimated environment-specific G-matrix predicted a substantial increase in VA with increasing standardized spring temperatures (Figure 2a, each point represents an environment-specific VA estimate). Similarly, there was a corresponding increase in the year-specific heritability estimates with increasing temperature (Figure 2b, each point represents a environment-specific h2 estimate).
We then tested whether the effects of increasing temperature on selection and genetic variance generated an association between them.
The relationship between the selection gradients (β) and additive genetic variance (VA) for laying date was negative but non-significant (slope = −0.006, se = 0.005, t33 = 1.18, p = 0.25; Figure 3a, dotted line). However, as random regression models are known to give upwardly biased estimates at the endpoints of the polynomials [22], we also tested this relationship after removing the extreme VA outliers (VA >10, see Figure 3a). This resulted in a near-significant relationship between the two (slope = −0.014, se = 0.008, t31 = 1.84, p = 0.075; Figure 3a, solid line). Furthermore, there was a significant negative relationship between the selection differentials S and heritability (slope = −10.96, se = 4.43, t33 = 2.48, p = 0.019, Figure 3b), which was robust to excluding the two extreme heritability estimates (excluding h2 >0.25: slope = −13.82, se = 6.2, t31 = 2.23, p = 0.03). Finally, using standardized measures of selection, there was a negative although non-significant significant relationship between selection and additive genetic variance and a significantly negative relationship between strength of selection and heritability (see Text S1).
Note that because there is selection for early breeding, selection gradients and differentials are negative, but there is a positive association between the absolute strength of selection and levels of additive genetic variance (or heritability). As a result, in years in which selection on laying date was relatively strong, estimated VA (and h2) was higher than in years when selection was weak (Figure 3). This association resulted in a highly significant relationship between temperature and the magnitude of the predicted response to selection (Figure 4).
Our analysis of long-term records on an important life history trait in a wild bird population found evidence that in years when spring temperatures were highest, selection was strongest, and the magnitude of estimates of additive genetic variance VA (and hence heritability) was also highest. As a result, there was evidence of a positive association between the strength of selection and the expression of additive genetic variance, and heritability. A positive association such as this between the strength of selection and expression of genetic variance and heritability could make the magnitude of the response strongly environmentally dependent; in this case, warming temperatures would considerably enhance any expected response to selection.
As has generally been found in studies of selection on laying date in birds [17],[21],[23],[24],[25], selection gradients and differentials were generally negative, indicating that early-breeding individuals had higher fitness than late-breeding individuals. Furthermore, the strength of selection was strongest when temperatures were highest (Figure 1). It has previously been shown that reproductive success [26] has declined in this population over the study period, most likely because, with increasing spring temperatures, there is evidence of increased “mistiming” of reproduction relative to the peak in food abundance [17]. This decline in reproductive success suggests that high spring temperature is generally associated with adverse environmental conditions. Hence, our results confirm the expectation in natural populations of stronger selection in adverse environmental conditions [9]. It is important to point out, however, that high temperatures are not necessarily associated with adverse environmental conditions in other systems. For example, a population of great tits in the U.K. has also experienced increasing temperatures, but recruitment rates in this population have increased over time [27].
Previous studies on natural populations have found that heritability decreased when environmental conditions are stressful [3],[10], although we know less about how VA changes. Here, we found instead that both additive genetic variance and heritability of laying date increased rather than decreased (Figure 2). Although there was substantial evidence that VA and VPE changed with environmental conditions (see Results), the change in VA alone was not statistically significant [18], something that is reflected in the large standard errors in Figure 2a. However, the statistical power to detect significant changes in additive genetic variance in relation to varying environmental conditions using a random regression animal model approach may be limited [18],[28]. Most importantly, the increase in VA is very large and represents 81.4% of the total change in VP (Figure 2a). This increase in VA is, for example, much larger than the increase in maternal genetic variance (VM) for birth weight in Soay sheep [4]. Note also that in the Soay sheep analysis, maternal environmental effects were not fitted with the same order polynomials as the maternal genetic effects, so that some of the increase in maternal genetic effects variance estimates could potentially be driven by environmental rather than genetic effects (in the same way as permanent environment variance will inflate additive genetic variance if not fitted explicitly, [29]).
One possible explanation for why VA may increase with higher temperatures is that high temperatures constitute not only a stressful, but also a novel, environment. For example, 2005 and 2007 had the highest recorded spring temperatures since this population study began back in 1955. It has been suggested that VA could increase in novel environments because selection has not yet had the possibility to remove the most deleterious alleles, as it will have in the ancestral environment, thereby causing an increase in the standing genetic variation [30]; a suggestion that has been confirmed in some empirical studies [31],[32]. More generally, our finding adds support to the idea [3] that predicting the direction in which VA should change with environmental conditions is complicated when environmental changes also leads to novel conditions, as is often the case with human-induced changes [3].
The increase in VA, heritability and strength of selection with increasing spring temperature meant that there was a positive association between the strength of selection on laying date and the heritability as well as expression of additive genetic variance of laying date (Figure 3a and b, respectively). The relationship between selection and amount of genetic variance was in the same direction whether using β as the measure of selection and VA as the measure of the potential for the population to adapt, versus using S and h2, but it was stronger (and hence statistically significant) between S and h2 (Figure 3b) than between β and VA (Figure 3a). One possible explanation for this may be that in the S and h2 comparison, both parameters depend on VP whereas in the β and VA comparison only β depends on VP and thus a change in VP may more quickly lead to a disassociation between β and VA than between S and h2.
Nevertheless, we believe the fact that the relationships between β and VA and between S and h2 are in the same direction (as well as that between standardized selection and VA/h2; see Text S1) offers strong support for an environmental coupling between these two parameters. This conclusion is supported by a highly significant temperature dependence of the predicted response to selection (see below, Figure 4).
Following traditional methodology we predicted the expected response to selection (see Text S1) using the Lande equation: R = VA * β [13],[14] but correcting for overlapping generations and the sex-limited expression of laying date, with the year-specific VA and β estimates (see Table S2), which amounted to an advance of 1.81 days in total over the study period. Furthermore, using the average of the annual VA and β values gave a predicted response of 1.46 days advancement, which corresponds to only 81.1% of the predicted response using year-specific values. Thus, not incorporating environmental dependence of the expression of genetic variance and strength of selection may underestimate the predicted response by up to 20%, at least in this specific case. Failing to incorporate an environmentally dependent association between the strength of selection and genetic variance may further obscure our understanding of microevolution as the predicted response will be dependent on the environmental variable in question. For example, in our study the predicted response is strongly correlated with spring temperature, with a much larger predicted response in warmer temperatures compared to cold (Figure 4). We caution, however, that the breeder's equation (and equivalent Lande equation) has particularly poor success when applied to studies in natural populations [33], presumably because many of its underlying assumptions are not met (see Text S1 for further discussion on this topic).
Very few studies have simultaneously examined how environmental factors influence genetic expression and selection and the association between them. Indeed, we are only aware of this being examined in a Soay sheep population [4], where there was a negative association between the strength of selection and heritability of body size. Another example where there may be a negative association between the strength of selection and heritability is for juvenile growth rates in North American red squirrels (Tamiasciurus hudsonicus) [34]. Although this study did not explicitly consider the association between selection and genetic variance, they found that VA and maternal genetic variance increased in years with low cone abundance (poor environment) whereas viability selection was stronger in years of high cone abundance (due to competition for territories [34]). This should generate a negative association between selection and total genetic variance that may hamper a response to selection.
Our results thus demonstrate a relatively unexplored mechanism that could potentially increase the speed of adaptation to climate change in this population. As temperatures are expected to continue to increase [35], a positive association between strength of selection on laying date and its potential to evolve may prove an important factor allowing at least this specific population to adapt to the rapid environmental conditions experienced. As it is ultimately this rate of adaptation that is crucial if species are to cope with climate change [36], our findings suggest that models linking population viability to climate change should incorporate such dynamic processes.
The data were collected in the Hoge Veluwe National Park, the Netherlands (52°05′ N, 05°50′ E), during the period 1973 to 2007. Nest boxes were visited at least once every week during the breeding season (April–June). The laying date of the first egg of a female's clutch (laying date, LD) was calculated from the number of eggs found during the weekly checks, assuming that one egg was laid per day. Both parents were caught and individually marked on the nest using a spring trap when the young were 7–10 d old. Laying dates are presented as the number of days after March 31 (day 1 = April 1, day 31 = May 1). We only used information on the first clutch, defined as any clutch started within 30 d of the first laid egg in any given year. Replacement and second clutches (which currently compromise <5% of breeding attempts, 21]) were thus excluded from the analysis. In total, therefore, we had information about 3,852 breeding records from 2,394 females. More details about the study population can be found in van Balen [37].
Temperature data were obtained from the De Bilt weather station of the Royal Dutch Meteorological Institute (www.knmi.nl/klimatologie/daggegevens) and used to calculate the daily average temperature over the period March 13–April 20, which is the period that best predicts the onset of laying using a sliding window analysis (see [18] for more detail).
To test for a relationship between spring temperature and the strength of selection on laying date, we took two approaches. First, we used a generalized linear mixed effects model (GLMM) with a Poisson error link fitted in ASREML-R [38] to model the relationship between number of recruits a female produced for the given year (as the measure of fitness) and her laying date that year, and to test its dependence on spring temperature (as measured by the interaction term between laying date and spring temperature). Individual identity and year were included as random effects to account for repeated measures on the same individuals and on years. Second, we estimated the annual strength of selection using the number of recruits produced per year divided by the mean number of recruits produced in the given year as a measure of relative fitness (ω) for each individual. Selection was then measured as the selection gradients (β) defined as the covariance between relative fitness and observed laying date divided by the variance in observed laying date, i.e. β = cov(ω, LD)/VPLD. Using this measure of selection allows us to predict the response to selection using the Lande equation: R = VA*β [13]. Predicting the response to selection can also be done using the more familiar Breeder's equation, R = h2*S [16], which uses an alternative measure of selection, the selection differentials defined as the covariance between a female's relative fitness (ω) and her observed laying date (LD), i.e. S = cov(ω, LD) [15]. Because a previous study examining the association between strength of selection and expression of genetic variance used S and h2 as parameters [4], we also present our results using these parameters for comparison. We note, however, that using selection gradients may represent a better measure of selection when the phenotypic variance in a trait changes [14], which it does here. We then regressed the annual selection gradients (and differentials) against the environmental values using a least-squares regression (with 1/se2 as weights when considering the selection gradients) in R 2.8.0 [39].
Finally, to allow comparison with other studies [40], we repeated all selection analyses using variance-standardized laying dates (i.e. standardizing laying date values to have zero mean and unit variance within each year). This did not change our conclusions and we report the results from these analyses in the Supporting Information section (Text S1, Table S1).
Yearly spring temperature values, standardized spring temperature values, sample size, mean laying dates, selection gradients (β), selection differentials (S), standardized selection differentials, estimated additive genetic variance, and heritability estimates along with annual predicted responses to selection (VA*β) are all reported in Table S2.
Quantitative genetic analyses require knowledge about the relationships among individuals within a population. Here, a pedigree was constructed where all ringed females known to have bred were assigned a mother and father as determined from observational data. In cases where brood manipulation experiments had been carried out and chicks had been moved between nests, we assigned the genetic parent rather than the social parent. If only one parent was known, we “dummy coded” the missing parent to preserve sibship information (note that we did not assign a phenotype to this parent). The extra-pair paternity (EPP) rate is unknown in this population, but is generally found to be low (3%–9%) in other populations of great tits [41],[42] and as extra pair paternity rates of less than 20% have been shown to have a negligible impact on heritability estimates [43] using a social pedigree is unlikely to be problematic.
Phenotypic trait variances can be separated into genetic and environmental causes of variation using an “animal model” [44]–[46]. By maximizing the information available in an extensive multi-generational pedigree, the “animal model” minimizes upward inflation of estimates of additive genetic variance (VA) due to shared environmental effects between relatives; this approach has been shown in simulation studies to perform well in partitioning environmental and genetic components of variance [29]. There are several additional reasons to believe that the genetic and environmental components have been well separated here. First, a previous study found no indication that common environment effects in the form of maternal effects are important for laying date in this population (VM/VP = 0.0023 [47]). Second, although common environmental effects frequently play a major role in inflating covariances between relatives in nestling traits [48], this is rarely the case for traits that are only expressed as adults, like laying date which we study here. Third, we explicitly take common environmental effects into account by fitting a permanent environmental effect [45]. In summary, therefore, we believe that our estimates of VA and h2 are accurate and unbiased by inflation of common environment effects.
Rather than only estimating the amount of genetic and environmental variance in laying date, we are interested here in whether the variance components changed with environmental conditions, and we therefore used a “random regression animal model” [49]. Random regression models use covariance functions [50] to explicitly fit variance components as a function of the environment and hence allow a detailed examination of how environmental heterogeneity—in this case, spring temperature—influences genetic architecture. Thus our model was:(1)where LDi is the vector of the individual (i) laying dates and X, Z1, Z2, and Z3 are the design and incidence matrices relating to the fixed and random effects of the additive genetic (ai), permanent environment (pei), and year (yri) observations, respectively. T is the spring temperature each year standardized to a (−1, 1) interval (Table S2). Fixed effects (bi vector) included age as a two-level factor (first year breeder or older), to correct for the fact that laying date is generally later in young birds compared to older birds in great tits [51], and spring temperature to account for the population-level response in mean trait value. Year (yr vector) was included as a random effect in order to model variation between years not explained by spring temperature and a permanent environment effect (pei vector) was fitted because of the repeated sampling of the same individuals; this also reduces inflation of estimates of the additive genetic variance due to environmental factors [29]. The error term (e vector) was partitioned into three decade–specific (1973–1984, 1985–1996, 1997–2007) groups, thus allowing residual errors to vary between decades. φ(ai,n1,T) is the random regression function of order n1 of the additive genetic effect of individual i, which varies as a function of the temperature T in a given year, and similarly, φ(pei,n2,T), is the random regression function of order n2 of the permanent environment effect varying as a function of T.
Because we were only interested in whether the two variance components (and particularly VA) changed with the environment, we only fitted two models. The first model was a zero order function (n1 = n2 = 0) for both VA and VPE in which variance components are constant across the environment. In the second model, we fitted a first order polynomial (n1 = n2 = 1) for both VA and VPE, thus allowing both additive genetic effects and permanent environment effects, and hence their corresponding variance components, to vary across the environment T. These two models were then compared using a likelihood-ratio test by calculating twice the difference in log likelihood, which is chi-squared distributed with degrees of freedom equal to the difference in degrees of freedom between the two models [52], which is here equal to 4 (variance in slopes and covariance between elevation and slope for both VA and VPE). As the model where both variance components were allowed to vary was significantly better than a model in which they were assumed to be constant (see Results) we used the estimates from the first order polynomial model to generate predictions of annual values of VA (and VPE) across varying temperatures. The environment-specific additive genetic covariance matrix, G, was then obtained as G = zQzT, where z is the vector of orthogonal polynomials evaluated at standardized temperature values and Q is the additive genetic variance-covariance matrix of the random regression parameters. Approximate standard errors for the (co)variance components of G as a function of the temperature values were calculated according to Fischer et al. [22], with confidence intervals defined as twice the standard errors. Finally, environment-specific heritability estimates were calculated as the environment-specific VA estimate divided by the environment-specific VP estimate from the model in which both VA and VPE varied with the environment. Because it has been found that random regression models can be sensitive to “edge effects” [22],[53], we repeated our analyses where we look at the association between strength of selection and expression of genetic variance to be conservative. For more information about the use of random regression animal models in natural populations, see [18] and [54]. All animal models were fitted using REML methods implemented in ASReml v 2.0 [38].
In order to test for an association between the strength of selection operating on laying date and the expression of additive genetic variance in laying date, we used environment-specific (and thus annual) VA and h2 estimates generated from the random regression animal model and regressed the annual selection gradients on our annual estimates of VA; we then repeated the regression for annual selection differentials against h2. Regressions using selection gradients were weighted by the inverse of the square of the standard error.
Because some individuals bred in multiple environments (i.e. years), estimates of selection will not be entirely independent, potentially violating some of the assumptions of least squares regression analyses. Although this is an inherent problem to all longitudinal studies, we assessed the potential for it to bias our conclusions by repeating our selection analyses using only a single record per individual (its first breeding attempt). Because this did not change the direction or significance of our analyses (regression of β on VA using 1/se2 as weights: b = −0.009, se = 0.005, t33 = −1.70, p = 0.099; regression of S on h2: b = −14.54, se = 5.08, t33 = −2.86, p = 0.007), we conclude that the potential violation of the non-independence criteria caused by multiple breeding events from the same individuals is not a significant issue here.
Although annual estimates of VA and h2 are derived from the random regression model, note that in testing for a relationship between them and selection, we use them only as predictor variables in a regression, for which there need not be an assumption of independence of data points.
|
10.1371/journal.pgen.1005361 | A Genome Scan for Genes Underlying Microgeographic-Scale Local Adaptation in a Wild Arabidopsis Species | Adaptive divergence at the microgeographic scale has been generally disregarded because high gene flow is expected to disrupt local adaptation. Yet, growing number of studies reporting adaptive divergence at a small spatial scale highlight the importance of this process in evolutionary biology. To investigate the genetic basis of microgeographic local adaptation, we conducted a genome-wide scan among sets of continuously distributed populations of Arabidopsis halleri subsp. gemmifera that show altitudinal phenotypic divergence despite gene flow. Genomic comparisons were independently conducted in two distinct mountains where similar highland ecotypes are observed, presumably as a result of convergent evolution. Here, we established a de novo reference genome and employed an individual-based resequencing for a total of 56 individuals. Among 527,225 reliable SNP loci, we focused on those showing a unidirectional allele frequency shift across altitudes. Statistical tests on the screened genes showed that our microgeographic population genomic approach successfully retrieve genes with functional annotations that are in line with the known phenotypic and environmental differences between altitudes. Furthermore, comparison between the two distinct mountains enabled us to screen out those genes that are neutral or adaptive only in either mountain, and identify the genes involved in the convergent evolution. Our study demonstrates that the genomic comparison among a set of genetically connected populations, instead of the commonly-performed comparison between two isolated populations, can also offer an effective screening for the genetic basis of local adaptation.
| Where does a local adaptation take place? In general, an adaptive divergence is predicted to occur between isolated populations because gene flow will erode and prevent the divergence. Therefore, previous genome-wide studies that aim to find the adaptive genes have compared populations that are usually tens of hundreds of kilometers apart. However, because nearby populations are likely to be genetically connected or connected until recently, most of the genome should be undifferentiated, leaving the genetic footprints of natural selections more pronounced. Thus, if an adaptive divergence is to be found within a small spatial scale, such case may favor the screening for the adaptive genes. Here, we took advantage of a unique small-scale local adaptation in Arabidopsis halleri subsp. gemmifera, where similar phenotypic differentiation is found across an altitudinal cline on two distinct mountains. By scanning the genome with a focus on the presence of unidirectional allele frequency shift along the altitudes, we successfully obtained genes with functions that were in line with the known phenotypic and environmental difference between altitudes. Our approach is applicable to any species that show microgeographic divergence and should help understand the genetic basis of small-scale evolution.
| Recent advances in next-generation sequencing (NGS) technologies have enabled a genome-scale analysis to infer the phylogenetic history, demography, and selection of natural populations. One of the intriguing challenges in ecological genomics is to identify the genes underlying local adaptation [1]. Although ecological genomics has been applied to various study systems, screening methods to detect the selected loci can be represented by two approaches: those that focus on the adaptive differentiation, and those that focus on the genotype-environment correlations. The former differentiation-based approach assumes neutral genetic drift to affect the entire genome, so that unusual differentiation at a particular locus should indicate a presence of selection. FST-based outlier tests are among the earliest and most common method to detect the selected loci [2]. The latter correlation-based approach compares a set of subpopulations at heterogeneous environments to detect the loci with correlation between allele frequency and environmental variables [3]. Availability of the genome-scale datasets have facilitated improvements in these two approaches, along with the development of other methods that employ indicators such as reduced heterozygosity, skews in site frequency spectrum, and extended linkage disequilibrium (reviewed in [4]). Although ecological genomics have provided important insights into the genetic basis of local adaptation, each of the above mentioned approaches has drawbacks to its practical implementation, which includes false positive and false negative detection of the selected loci. For instance, FST-based outlier tests generally face problems in identifying the significant departure from neutral expectation. Without taking account the actual demographic history, outlier tests may suffer from false positives due to high variance in FST values among the neutral loci [5]. Within- and between-population structures can also increase the false positive rate of correlation-based approaches by creating spurious correlation between allele frequency and environmental variable [6]. In any case, complex demographic histories and entailing genetic structures are the major issues that challenge the genome-wide screening for adaptive genes, and a combination of different approaches is preferred to avoid false detections [6].
Because gene flow will erode and prevent a genetic divergence, adaptive differentiation is more likely to occur between populations that are reproductively isolated. Geographical distance can provide a strong reproductive barrier and also shape environmental differences (e.g., temperature along the latitudes), both of which may facilitate the adaptive divergence between populations. Indeed, most ecological genomic studies compare populations that are tens of hundreds of kilometers apart (e.g. representative study cases reviewed in [7]). The problem of comparing distantly isolated populations is that the periods since population divergence are usually long enough to allow the intervention of various demographic processes. As a consequence, complicated population structure seems as an intrinsic difficulty to conduct the genome-wide scan for adaptive genes. Recently, growing number of works reporting microgeographic-scale adaptation [8–12] have corroborated the theory that adaptive population divergence can take place even under high gene flow if selective pressure is sufficient [13]. Microgeographic-scale adaptation may in fact be a suitable system for ecological genomics because the evolutionary split between nearby populations should be relatively recent compared to that of distantly isolated populations. Furthermore, gene flow may benefit the screening procedure because most of the genome is expected to be undifferentiated between populations, leaving the genetic footprints of a natural selection more pronounced [14]. In plant species, NGS-based restriction-site associated DNA (RAD) sequencing has been used to study the distinct ecotypes that occur within few kilometers from each other in Senecio [15], and Helianthus [16]. Although these studies have provided insights into the phylogenetic history, population demography, and genomic structure dynamics during microgeographic-scale divergence, candidate genes that underlie the phenotypic differentiation were not identified.
An example of microgeographic-scale divergence has been reported from a self-incompatible perennial plant, Arabidopsis halleri subsp. gemmifera. In Mt. Ibuki, a mountain located in central Japan, populations of this plant are continuously distributed along the top to bottom of a hiking trail. Although the linear distance between the lowest and highest populations is smaller than 3 km, highland ecotypes characterized by dense trichomes on the leaves and stems [17] are found on the peaks (S1 Fig). A previous AFLP-based study on Mt. Ibuki demonstrated little genetic differentiation between normal and highland ecotypes collected from low and high altitudes [18]. Thus, it has been suggested that these two ecotypes share a similar genomic structure and the evolutionary split has occurred relatively recently. Interestingly, similar phenotypic divergence is also found along the altitudes of Mt. Fujiwara, which situate approximately 30 km from Mt. Ibuki. Highland ecotypes of the two mountains are regarded as a convergent evolution, however, no empirical evidences have yet been reported. In addition to denser trichomes, growth chamber measurements have confirmed other genetically based convergent characteristics of the highland ecotypes, such as shorter but thicker stems and leaves, increased resource investment to photosynthetic components, and increased accumulation of ultraviolet (UV) absorbing compounds [19]. Overall, these altitudinal differentiations are considered as a consequence of high altitude adaptation. Although trichomes in plants often serve in the defense against herbivores [20], a study in A. halleri subsp. gemmifera revealed no clear correlation with leaf beetle damage [21]. Interestingly, the hyperaccumulator plant A. halleri accumulates zinc and cadmium inside its trichome bases [22]. This finding suggests that denser trichomes in the highland ecotypes might indicate higher tolerance to heavy metals. Alternative trichome functions in other plant species, including the prevention of external ice formation [23], avoidance of excess transpiration under strong wind [24], and protection against UV radiation [25], are also considered to be related to the adaptive significance of dense trichomes at high altitudes. Other characteristics of the highland ecotypes are also associated with the common selective pressures in the two mountains, such as dwarf phenotype to resist strong wind, investment to photosynthetic component to compensate the reduced enzyme activity due to suboptimal conditions, and accumulation of UV absorbing compound to tolerate increased UV radiation [19]. However, mountain-specific altitudinal differentiations are also reported. For instance, freezing resistance [19] and rapid seed germination (Shin-Ichi Morinaga, personal communications) are found only in the highland ecotypes from Mt. Ibuki. Nevertheless, the two mountains share similar environmental characteristics in terms of altitudinal cline. Although both mountains are relatively low (1,377 and 1,144 m for Mt. Ibuki and Mt. Fujiwara, respectively), areas above approximately 1,000 m are host to open subalpine grasslands with calcareous scree and heavy snow cover in winter. In contrast, areas lower than approximately 400m occupy the understory of temperate forests. Annual temperature, snow depth, and canopy openness have been quantified to show gradient variation along the altitude in both mountains [19]. As in this case, mountain populations may be an excellent model for the analysis of microgeographic adaptation because steep environmental gradients can shape selective barriers on a small geographic scale.
Thanks to the genetic information accumulated in A. thaliana, ecological genomics has become a powerful approach to screen adaptive genes from wild Arabidopsis species [26–29]. However, while these studies have provided fruitful insights into the genetic basis of local adaptations, genomic comparisons have so far been conducted at the macrogeographic-scale, using distantly isolated populations. Here, we test the prediction that genomic comparison at the microgeographic-scale can also offer an effective screening for the genetic basis of local adaptation. If the screening procedure works as expected, we should be able to find some correlation between the candidate genes and the observable phenotypic or environmental differentiation. In addition, a replicated analysis in two independent but synchronizing environmental transects will have a good chance of finding the genes involved in a convergent evolution. Our study system take advantage of the above mentioned populations of A. halleri subsp. gemmifera on Mt. Ibuki and Mt. Fujiwara, where populations continuously distribute along a steep environmental cline and the populations at each extreme (the lowest and highest populations on each mountain) are locally adapted to their habitats. Within each mountain, the loci governing altitudinal adaptation should be highly differentiated between the lowest and highest populations. More importantly, theoretical models predict that, if a set of populations is distributed along an environmental continuum and neighboring populations are exchange their genes, clines of allele frequencies at the adaptive loci can be observed [30, 31]. Because neighboring populations of A. halleri subsp. gemmifera in both mountains are close enough to allow gene flow, we placed an emphasis on detecting correlations between allele frequencies and altitudinal clines. Thus, we employed both differentiation-based and correlation-based approaches to screen the selected loci from a genome-wide SNP dataset. Credibility of the screening procedure was evaluated by comparing the proportion of a certain Gene Ontology (GO) term between screened and unscreened set of genes. Here, we selected 30 GO terms that cover the representative phenotypic and environmental entries within the database. If we successfully retrieve the genes under natural selection, then we should be able to see coincidence between the enriched GO terms and the known phenotypic or environmental differentiation across the altitudes. Furthermore, the screened loci were narrowed based on the presence of genetic hitchhiking. The screening procedure was independently applied to each mountain, and we obtained two lists of candidate genes that are potentially involved in altitudinal adaptation. By comparing these gene lists, we distinguished between genes that are adaptive only in either mountain, and those involved in the convergent evolution.
To perform a genome-wide screen for loci associated with local altitudinal adaptation, we began by establishing a draft de novo reference genome for A. halleri subsp. gemmifera. The whole-genome shotgun method via next-generation sequencing (NGS) was applied to a single individual sampled from the base of Mt. Ibuki. Using 190× coverage sequence data (haploid genome size of A. halleri = 255 Mbp [32]), genome assembly resulted in 149,013 scaffolds, with an N50 of 4,825 bp and a total of 252 Mbp, which corresponds to 98.8% of the entire genome. The resulting reference genome was evaluated by mapping A. thaliana exon sequences from 33,602 genes deposited in the TAIR10 database (The Arabidopsis Information Resource; http://www.arabidopsis.org). For comparison, we mapped the same A. thaliana exon sequences to the high-quality reference genome of A. lyrata (695 scaffolds, with an N50 of 24.5 Mbp, totaling 207 Mbp [33]). As a result, 92.9% and 90.7% of the A. thaliana exons were mapped to the reference genomes of A. halleri subsp. gemmifera and A. lyrata, respectively. Although the number of scaffolds remains excessive compared with the actual chromosome number in A. halleri (2n = 16; [32]), our draft de novo reference genome sequence covers the entire genome well and will facilitate genomic studies in this species.
On both Mt. Ibuki and Mt. Fujiwara, four distinct populations associated with different altitudes were situated along hiking trails from the bottom to the top of the mountains. The four populations are found at the altitudes of 380, 600, 1,000, and 1,250 m on Mt. Ibuki and at 200, 400, 700, and 1,100 m on Mt. Fujiwara (Fig 1B and S1 Table). The linear distance between the lowest and highest populations is approximately 2.7 km on Mt. Ibuki and 1.9 km on Mt. Fujiwara. In addition to the main study sites, four reference populations were set apart from the mountains (Fig 1A and S1 Table). These populations were situated at low altitudes (220, 230, 370, and 520 m) with environments similar to the lowest populations from the main study sites. On the two mountains, five individuals from each altitude-specific population were collected for analysis, whereas four individuals were collected from the reference populations. Through genome-wide resequencing of each of these 56 individuals, we obtained a set of 527,225 reliable SNPs with a minimum read count of five per individual (S1 Table). The average inter-SNP spacing across the entire genome was 484 bp. The mapped A. thaliana exon information was used to examine the proximity of each SNP to a functional gene. Among the 527,225 SNPs, 327,980 overlapped with or were within 5 kbp of an exon for 22,395 genes. These SNPs and the associated functional gene information were used for the following analyses.
Genetic diversity (He) was significantly different (bonferroni-corrected p-value from pairwise Wilcoxon test < 0.01) among all paired populations within each mountain, except for IB0380 vs. IB0600 in Mt. Ibuki, and FJ0400 vs. FJ1100 in Mt. Fujiwara (Table 1). Although the statistical significance is somewhat overestimated, lower populations of Mt. Ibuki tended to have smaller genetic diversity compared to higher populations. To examine the population structure within and between the two mountains, we conducted a structure [34, 35] analysis of all 56 individuals (including the reference populations) using a set of 10,000 randomly selected SNPs. Based on 20 independent runs for each value of K (the number of subpopulations) from 1 to 12, both the log likelihood value and Evanno’s ΔK method [36] indicated the optimum K to be six (Fig 2B). Under K = 6, each cluster clearly corresponded to the two mountains and the four reference populations (Fig 2A). It is notable that the four altitude-specific populations on each mountain were not genetically subdivided. However, subdivision within each mountain were indicated with higher K values. Further structure analysis within each mountain supported the split in Mt. Ibuki, but not in Mt. Fujiwara (S2 Fig). Previous study has demonstrated that although snow depth and canopy openness increased with increasing altitude in both mountain, Mt. Ibuki showed steeper gradients for both environmental components [19]. Thus, the genetic split in Mt. Ibuki may indicate a restricted gene flow among the altitudes due to stronger environmental barrier. Nevertheless, interleaving populations of Mt. Ibuki (IB0600 and IB1000) seem to be comprised of some admixed individuals. These individuals indicate the presence of gene flow between the neighboring altitude-specific populations. In fact, although highland ecotypes from the top of the mountain are easily distinguished based on their appearance, plants with intermediate phenotypes are found at intervening altitudes. Because highland and normal ecotypes are highly cross-compatible (Shin-Ichi Morinaga, personal communications), these intermediate plants are likely to have originated from natural hybridization due to frequent gene flow between neighboring populations. In addition, pairwise G′ST values showed a pattern of genetic differentiation by distance in both mountains (Table 1). Thus, the population structure in each mountain can be regarded as a simple linear stepping-stone model proposed by Kimura and Weiss (1964 [37]).
We also examined the historical relationship among populations with TreeMix [38], a statistical model used to infer patterns of population splits and mixtures from genome-wide allele frequency data. The maximum likelihood tree based on 518,706 bi-allelic SNPs clearly demonstrated that the evolutionary split between the two mountains predated the differentiation of the altitude-specific populations (Fig 2C). In addition, the tree explained most (99.1%) of the variance in relatedness between the populations, which indicates that the tree captures the historical relationship without adopting migration events from distantly related populations. These results indicate that although the two mountains share a common ancestry, the differentiation of the altitude-specific populations took place independently on each mountain. Therefore, the morphologically similar highland ecotypes found on the two mountains may be considered to be a consequence of convergent evolution. Together with the results from structure analysis, these findings suggest that these populations are a suitable model for exploration of the genetic basis of microgeographic adaptation.
To identify the SNPs associated with altitudinal adaption, we conducted a screening based on the following assumptions: first, and most importantly, we anticipated a cline in the allele frequency as a result of natural selection across environmental gradients. Therefore, we focused on those loci that undergo a unidirectional change in allele frequency along the altitudinal cline. To further reduce the number of candidate loci, we adopted the following two selection criteria: 1) the SNP loci should be highly divergent between the lowest and highest populations; and 2) the frequency of the derived allele should be higher in the highest-altitude populations. We developed an index U to measure the unidirectional change in allele frequency, used G′ST proposed by Hedrick (2005 [39]) to measure the divergent between lowest and highest populations, and also developed an index ΔD′ to measure the frequency of the derived allele at the highest populations (see Materials and Methods section for details). Indices at each loci were averaged across a 4kbp window size and the upper 1.5 times the IQR (interquartile range) of a genome-wide frequency distribution (Fig 3) was determined as a screening threshold. Screening was conducted independently for the populations from each mountain, and only those SNP loci that fulfilled all three criteria were further considered. The number of SNPs that fulfilled the criteria was 5,523 for Mt. Ibuki and 5,407 for Mt. Fujiwara (Fig 4). The total number of identified SNPs in common between the two mountains were 358. Among the screened SNPs, 3,869 from Mt. Ibuki and 3,527 from Mt. Fujiwara were linked (overlapping or within 5 kbp of an exon) to a gene. The number of genes linked to the screened SNPs was 923 and 924 on Mt. Ibuki and Mt. Fujiwara, respectively.
To gain perspective into the biological process in which the screened SNPs are involved, we conducted a Gene Ontology (GO) enrichment analysis for each mountain. We tested for enrichment in 30 GO terms that cover the representative phenotypic and environmental entries within the database. To adjust for multiple comparisons, significant enrichment was accepted if the corresponding false discovery rate (FDR) q-value [40] was below 0.05. Here, we tested for enrichment using two approaches: one is an SNP-based method, where the ratio of SNPs that are associated and unassociated with a given GO term is compared between the lists of screened (SNP loci that fulfilled all three criteria mentioned above) and unscreened (all SNP loci) datasets. Another is a gene-based method, where the ratio of genes that are associated and unassociated with a given GO term is compared between the lists of screened and unscreened SNPs. Because the SNP-based method assumes that every screened SNP represents an independent observation, linkage between SNPs will cause bias, and the significance of enrichment will be overestimated [41]. However, the gene-based method ignores the joint effect of multiple SNPs within a gene, which may underestimate the significance of enrichment [41, 42]. As previously recommended for gene set enrichment analysis [43], we declare that our enrichment analysis is an exploratory procedure rather than a pure statistical solution. Not surprisingly, the SNP-based method detected more significant enrichment in GO terms compared with the gene-based method (Fig 5 and S2 Table). Here, we discuss the SNP-based enriched GO terms that were significant in both mountains. The four common GO terms were ‘response to red or far red light,’ ‘cellular response to DNA damage stimulus,’ ‘meristem development,’ and ‘trichome differentiation.’ It is noteworthy that the GO term related to trichomes, which constitute the most distinguishing characteristic of the highland ecotype [17], was detected in both mountains. In addition, enrichment for ‘trichome differentiation’ was also indicated by the gene-based method in both mountains. Detection of a major defining characteristic of the highland ecotype supports the validity of our screening procedure. Although the adaptive significance of the denser trichomes in the highland ecotypes remains unknown, our result strongly suggests that the trait has evolved under an common selective pressure between the two mountains. Another common GO term related to morphogenesis was ‘meristem development.’ This GO term can be related to the morphological differentiation where plants at the lower altitude are characterized by their tall and spindly appearance, and highland ecotypes by their dwarf-like appearance (S1 Fig).
Another common GO term ‘response to red or far red light’ is also interesting since previous observation has detected a positive correlation between canopy openness and altitude in both mountains [19]. Although we could not observe an enrichment in the term ‘photosynthesis,’ the increased investment to photosynthetic components in the higher altitudes in both mountains could be related to an adaptation against light environment variance. In this context, measurement based on cyclobutane pyrimidine dimer has demonstrated that opened canopy at higher altitudes induce increased UV induced DNA damage. At the same time, a correlation between altitude and UV tolerance via accumulation of UV absorbing compound was also detected in both mountains [19]. Although enrichment in the term ‘response to UV’ was not detected, we succeeded to find a significant enrichment in the term ‘cellular response to DNA damage stimulus’ in both mountains. These coincidence point out a possibility that light environment is an important selective pressure for the convergent evolution between the two mountains. On the other hand, although tolerance against freezing seems as an indispensable ability for high-altitude adaptation, previous observation detected an increased tolerance only from the highland ecotypes of Mt. Ibuki [19]. GO enrichment analysis were consistent with this result, where significant enrichment of the term ‘response to freezing’ was detected in Mt. Ibuki, but not in Mt. Fujiwara. Overall, consistency between the enriched GO terms and known features of the highland ecotypes suggests that our screening procedure provided a good estimate for the SNP loci associated with altitudinal adaptation.
Here, we also tested other popular approaches to find the loci under selection. We used BayeScan [44–46] to find the FST outliers between the lowest and highest populations, and LFMM (Latent Factor Mixed Models [47]) to find the loci that correlate with the altitude. As shown in S3 Fig, these typical outlier tests did not fit very well with our dataset, especially in terms of detecting statistically significant outliers. More specific, at the significance level of a FDR q-value = 0.01, none of the loci from both mountains were detected by the BayeScan analysis. In Mt. Ibuki,–log10(q-value) of even those with the most highly differentiated loci (loci that are fixed for one allele in the lowest, and fixed for another in the highest population) reached a ceiling around 1.0. The problem seems to be caused by our sampling design, where small number of individuals were collected from limited geographical points. According to the manual for BayeScan, statistical power to detect the outliers will be limited when small sample size is used. On the other hand, LFMM analysis detected 1,530 outliers (FDR q-value < 0.01) in Mt. Ibuki, however, none were detected in Mt. Fujiwara. In LFMM, the background population structure is modelled from a chosen number of latent factors (K), which corresponds to the number of neutral genetic structure of the data. Underestimated value of K leads to liberal tests with false positives, whereas overestimated K leads to conservative tests with false negatives. Here, we used K = 2 as a number of latent factor in both mountains. From the structure analysis, a genetic split was detected in Mt. Ibuki and K = 2 was statistically supported (S2A and S2C Fig). However, in Mt. Fujiwara, clear differentiation (K = 2) was not supported (S2B and S2D Fig). Thus, K = 2 for Mt. Fujiwara may have been an overestimate, leading to a conservative test with false negatives. Although we can run the LFMM with K = 1, such run will not account for background population structures and will produce a plethora of false positives because a large set of loci is correlated with the altitude. Overall, because typical outlier analyses expect a set of numerous individuals from variable locations (environment) as an input, our dataset would not be suitable for these tests. Another problem may be the linear stepping-stone population structure detected in our study sites (Table 1), where not only the adaptive loci but also a large set of neutral loci can be correlated with the altitude. Under this condition, it would be difficult to determine the cutoffs to correct for the underlying population structure.
Based on the screened SNPs linked to genes, we attempted to narrow down and sort the candidate genes according to the likelihood of having undergone natural selection. Here, we assumed that the presence of genetic hitchhiking represented a footprint of a selective sweep [48]. However, we acknowledge that variation in mutation rates, non-uniform recombination rates, and chromosomal rearrangements can also lead to differentiated genomic regions and clusters of genes that contribute to local adaptation are more likely to diverge together regardless of selective sweeps [49]. To detect local signatures of genetic hitchhiking, we scanned for continuous allele frequency clines (the primary criterion for screening the SNPs) around the screened gene-linked SNPs. Through an independent scanning procedure within each mountain, we identified 474 and 629 continuous hitchhiking regions, or ‘genomic islands,’ which included 573 and 721 genes in the populations from Mt. Ibuki and Mt. Fujiwara, respectively (see S3 Table for the genes within top 100 genomic islands). To reduce the false positive detection from a single SNP locus, genomic islands that contained only one screened SNPs were rejected and total of 350 and 203 genes from Mt. Ibuki and Mt. Fujiwara, respectively, were excluded. Based on the length of the continuous hitchhiking region (i.e., the length of linkage disequilibrium) and the steepness of the allele frequency clines (i.e., the difference in allele frequencies between lowest and highest populations), the genomic islands were ranked according to how likely they were to have undergone a selective sweep (see S4 Fig for workflow). Linkage disequilibrium can be disrupted by recurrent mutations and recombination events during the evolutionary time course; a higher ranking indicates that the genomic region experienced stronger and/or more recent natural selection.
Here, we considered the top 20 genomic islands as promising candidates that were recently subject to natural selection (Table 2). For example, we detected a steep allele frequency cline spanning approximately 10 kbp on Mt. Ibuki, with a peak near the 5’ UTR of EDA8 (AT4G00310; Fig 6A). EDA8 includes GO terms such as ‘regulation of flower development’, ‘response to freezing’, and ‘seed dormancy process’ [50]. Because freezing tolerance [19], flowering period, and seed dormancy (Shin-Ichi Morinaga, personal communications) differ between the lowest and highest populations from Mt. Ibuki, the functional annotations of EDA8 are in line with the known phenotypic and environmental differences between altitudes. However, an allele frequency cline was not detected in the same genomic region on Mt. Fujiwara (Fig 6B). Mountain-specific candidate genes, such as EDA8, may indicate the underlying differences in natural selection between the mountains or that each mountain utilizes distinct genes to overcome a common natural selective pressure. Other genes from Mt. Ibuki with notable GO terms included the following: FNR1 (AT5G66190), with ‘response to cold,’ and ‘photosynthesis’ [50]; LIS (AT2G41500), with ‘seed dormancy process,’ and ‘response to freezing’ [50]; EMB2788 (AT4G27010) with ‘regulation of flower development’ [50]; SAR1 (AT1G33410), with ‘regulation of flower development’ [50]; FTSH12 (AT1G79560) with ‘embryo development ending in seed dormancy’ [51]; and AT5G16280 with ‘vegetative to reproductive phase transition of meristem’ [50]. Specific genes from Mt. Fujiwara included the following: AT2G40270 with ‘response to bacterium,’ and ‘response to insect’ [50]; BAM7 (AT2G45880) with ‘vernalization response’ [50]; STO (AT1G06040), with ‘response to temperature stimulus,’ and ‘response to light stimulus’ [50, 52]; AVP1 (AT1G15690), with ‘response to water deprivation,’ and ‘response to salt stress’ [53]; and FWA (AT4G25530), with ‘photoperiodism, flowering,’ and ‘trichome morphogenesis’ [50] (see Table 2). Detailed analysis of the adaptive roles of these mountain-specific genes in A. halleri subsp. gemmifera would highlight unique characteristics of natural selection in the superficially similar habitats between the two mountains. We also found that some genes within the list shared a common function. For instance, four genes from Mt. Ibuki (EDA8, PBA1, FNR1, and LIS) and three genes from Mt. Fujiwara (PBA1, BAM7, and STO) had GO terms under ‘response to temperature stimulus.’ Among the 22,395 SNP-tagged genes, only 863 were associated with this GO term, and an empirical p-value for the observed result was 0.007. Although increased freezing tolerance was detected only in highland ecotypes of Mt. Ibuki [19], our results suggest that temperature variation can be an important selective pressure for altitudinal adaptation in both mountains. Inferring environments and ecological traits from genomic information, the so-called ‘reverse ecology’ approach [54], may give rise to a new era in ecological genomics on wild plant species.
The most novel findings of this study are candidate genes that are shared between the two mountains. In total, two genes were ranked within the top 20 genomic islands on both mountains. An empirical p-value to find two common genes between two independent gene lists from a set of 22,395 SNP-tagged genes was 0.001, which supports the presence of convergent evolution involving the same genes. Interestingly, both genes had functional annotations relevant to altitudinal adaptation. One of these ‘shared’ genes is GSL8 (AT2G36850), which is annotated with the GO terms ‘meristem initiation,’ ‘trichome morphogenesis,’ and ‘telomere maintenance in response to DNA damage’ [50]. On both mountains, the genomic region around GSL8 underwent a continuous unidirectional allele frequency shift that spanned at least approximately 15 kbp and most likely involved a longer region (Fig 6C and 6D). The long linkage distance observed in this case may be evidence of recent selection acting on this genomic region. In addition, anatomical observation of transposon-induced gsl8 A. thaliana mutant lines detected dwarfed growth, revealing the wild-type gene function in normal morphological development [55]. These results indicate that GSL8 is an ideal candidate gene for explaining the morphological convergence found between the highland ecotypes on the two mountains. Another candidate is PBA1 (AT4G31300), which presents the GO terms ‘response to temperature stimulus,’ ‘response to salt stress,’ and ‘response to cadmium ion’ [50, 56]. PBA1 shows an altered expression level in response to various stresses, such as NaCl [56], zinc [57], genotoxic agents [58], oxidants [59], and viral infection [60]. Furthermore, an RNAi knockdown lineage showed defects in plant immunity against bacterial pathogens [61]. Considering the variety of functions related to abiotic and biotic stresses, PBA1 appears to be a promising candidate for playing a role in altitudinal adaptation. Overall, these ‘shared’ genes may be a result of common natural selection acting on genetic variation that preceded the divergence of the two mountain populations, and they highlight the genetic basis of convergent evolution. Needless to say, other highly ranked genes without notable GO terms are also worth examining because they might retain unknown adaptive functions. To validate our result, the screened candidate genes must be further investigated by functional analyses of the genes, detecting loci that alter fitness, and field measurements including transplantation experiments.
Another ecological genomic study in A. halleri has been conducted at the Swiss Alps, where genome-wide SNP analyses were performed to search for the imprints from natural selection related to environmental variation [29]. By focusing on the highly differentiated genomic regions associated with environmental factors such as precipitation, slope, radiation, site water balance, and temperature, a list of 175 genes were obtained. Although the study case in the Swiss Alps was conducted in a wider geographical scale compared to the present study, the populations were situated at various altitudes ranging from 790 m to 2,308 m. Thus, we may have a chance to find common genes related to altitudinal adaptation between the mountains in central Japan (Mt. Ibuki and Mt. Fujiwara) and Swiss Alps. Unfortunately, none of the genes within the top 20 genomic islands from our study were found in the 175 genes from the Swiss Alps. However, three genes within each of the top 100 genomic islands from Mt. Ibuki and Mt. Fujiwara were also listed in the Alps (S3 Table). Although the coincidence is not surprising considering the large number of genes within each list (empirical p-value for the observed result was 0.09 for Mt. Ibuki and 0.06 for Mt. Fujiwara), we noticed that a single gene, CMT1 (AT1G80740), was detected in all three locations (empirical p-value = 0.006). This gene was ranked as the 51st and 40th in the gene list from Mt. Ibuki and Mt. Fujiwara, respectively, and was associated with site water balance in the Swiss Alps. Although we must further compare the selected loci and haplotypes between central Japan and Swiss Alps, the gene may be an evidence of convergent evolution to altitude in different continents.
Although theories for local adaptation have supported the development of population genomics, several central predictions remain untested, especially for predictions involving gene flow (reviewed in [1]). Under gene flow, adaptive differentiation requires an allele with high fitness in one environment to show lower fitness in the other environment [62]. Thus, fitness trade-offs of the adaptive traits are expected to be associated with trade-off at the loci level. Otherwise the allele with the highest fitness will invade the other population thereby causing the locus to become monomorphic [63]. In addition, the loci involved in local adaptation are expected to cluster together on the chromosomes [14, 49, 64]. Further investigations on our candidate genes should provide an opportunity to empirically evaluate the untested predictions, and help understand the evolutionary dynamics of adaptive genes during local adaptation. In this context, an improved reference genome with longer scaffolds would not only enhance accuracy of detecting the selected genes, but also would assist in clarifying the positional relationship among the adaptive loci. The Joint Genome Institute (JGI) has recently assembled another reference genome for A. halleri which is available at: http://phytozome.jgi.doe.gov/pz/portal.html#!info?alias=Org_Ahalleri_er. Although dataset usage is restricted prior to publication, the reference genome from JGI has shorter total genome size (145.5 Mbp versus 252.2 Mbp) but longer N50 value (24.4 Kbp versus 4.8 Kbp), compared to our present reference genome. However, we are also developing an improved version of the A. halleri subsp. gemmifera reference genome, which should be comparable to the A. halleri genome from JGI.
Our study demonstrates that typical outlier-based approaches (BayeScan [44–46] and LFMM [47]) have limitation in screening for the selected loci at a microgeographic-scale. Due to recent colonization event, not only the selected loci, but also a large set of neutral loci can show patterns of variation where allele frequencies change along the environmental gradient. In such cases, the selected loci may not differ from the genomic mean sufficiently to be considered as an outlier. We therefore suggest that a genomic region-based approach (genomic islands in the present study) which aims to detect the genetic hitchhiking regions may be more successful, rather than approaches that treat each locus as independent. Another promising approach would be a comparison between parallel environmental gradients. A study in sessile oak investigated whether SNP variation of candidate genes reflect the clinal pattern of bud burst along altitudinal and latitudinal gradients [65]. By comparing the results in the two parallel gradients, a set of genes showing imprints of selection in both gradients were obtained, which can be considered as evidence for convergent evolution. In the present study, we also utilized two independent but parallel altitudinal clines, where phenotypic observations indicate the presence of a convergent evolution. Because the probability of occasionally detecting the same gene from parallel environmental gradients is very low, the common genes appear intuitively promising. We anticipate that the number of ecological genomic studies on convergent evolutions will grow, as it provides an excellent opportunity to efficiently screen the candidate genes responding to natural selection.
Arabidopsis halleri subsp. gemmifera is a perennial, self-incompatible, clonal herb distributed in the Russian Far East, northeastern China, Korea, Taiwan, and Japan [66]. The highland ecotype, characterized by denser trichomes, was formerly treated as the variant Arabis gemmifera var. alpicola [17] and is found only in the higher altitudes of Mt. Ibuki and Mt. Fujiwara in central Japan. On both mountains, continuous variation in morphological characters is found along altitudes (Shin-Ichi Morinaga, personal communications). Our main study populations were located on Mt. Ibuki (IB0380, IB0600, IB1000, and IB1250) and Mt. Fujiwara (FJ0200, FJ0400, FJ0700, and FJ1100). The low-altitude reference populations were situated at Minoo (MN0220), Inotani (IN0230), Itamuro (IT0520), and Okunikkawa (OK0370). See Fig 1A and 1B and S1 Table for the location and coordinates. Leaf samples were collected from each of the 12 populations and silica-dried for subsequent DNA extraction. To avoid sampling of clones, the sampled individuals were at least 4 m apart from each other.
Genomic DNA was extracted from the dried leaf of a single individual using the DNeasy Plant Kit (QIAGEN). This individual was collected from population IB0380 and was not included in the resequencing analysis. DNA libraries were prepared using the Illumina TruSeq DNA Sample Preparation Kit for paired-end reads, the Roche GS Titanium Rapid Library Preparation Kit for 454 single reads, and the SOLiD Mate-Paired Library Construction Kit for mate-pair reads. Instead of SOLiD adapters, Illumina adapters were used in the final step of mate-pair library construction. Reads were generated using the Illumina GAIIx, HiSeq2000 (300 bp paired-end reads, 3 kbp and 5 kbp mate-pair reads), and Roche 454 GS FLX Plus Titanium (single reads) systems. Subsequent data processing was performed with CLC Genomics Workbench version 6 (CLC bio). Raw reads were trimmed based on quality scores of 0.05 and a maximum allowance of two ambiguous nucleotides. Reads shorter than 60 bp for the Illumina platform and 100 bp for the Roche 454 platform were discarded. De novo assembly was carried out using the “De Novo Assembly” function with the following parameters: Mismatch cost 3, Insertion cost 3, Deletion cost 3, Length fraction 1, Similarity 1, Minimum contig length 200. Single reads from the Roche 454 platform were used as guidance-only reads. The number of reads used to construct the reference genome was as follows: 74,102,134 (7,034,411,911 nt) Illumina 300 bp paired-end reads, 150,099,682 (13,756,599,514 nt) Illumina 3 kbp mate-pair reads, 127,910,808 (11,644,031,026 nt) Illumina 5 kbp mate-pair reads, 66,195,930 (6,338,573,278 nt) single reads from the broken pairs of Illumina 3 kbp mate-pair reads, 73,840,719 (7,058,674,210 nt) single reads from the broken pairs of Illumina 5 kbp mate-pair reads, and 3,534,305 (2,579,555,709 nt) Roche 454 single reads. The established de novo A. halleri subsp. gemmifera reference genome sequences is uploaded online and freely available. The quality of the assembled reference genome was validated by mapping the exon sequence of A. thaliana at the TAIR10 database (The Arabidopsis Information Resource; http://www.arabidopsis.org). A total of 217,183 A. thaliana exon sequences were mapped using the “Map Reads to Reference” function with the following parameters: Mismatch cost 2, Insertion cost 2, Deletion cost 2, Length fraction 0.3, Similarity 0.9. Using the same parameter settings, the A. thaliana exon sequences were mapped to the reference genome of A. lyrata [33] downloaded from the JGI’s PHYTOZOME portal (US Department of Energy Joint Genome Institute; http://www.phytozome.net/alyrata).
Genomic DNA from each of the 56 individuals was isolated with the DNeasy Plant Kit (QIAGEN). DNA libraries were constructed according to the Low-Throughput Protocol of the TruSeq DNA Sample Preparation Kit (Illumina). After quantification, 76 and 93 bp paired-end reads were obtained from the Illumina GAIIx platform and 101 bp paired-end reads from the HiSeq2000 platform. Raw short read sequences have been deposited at DDBJ and are freely available. Subsequent mapping and SNP calling procedures were performed using CLC Genomics Workbench version 6 (CLC bio). Prior to mapping, all sequences were trimmed based on a quality score of 0.05 and a maximum allowance of two ambiguous nucleotides. Broken pairs and reads shorter than 65 bp were discarded. For each individual, the reads were mapped to the reference genome with the following parameters: Mismatch cost 3, Insertion cost 3, Deletion cost 3, Length Fraction 0.97, and Similarity fraction 0.97. The reads from each individual were mapped to satisfy 9- to 15-fold coverage of the reference genome (S1 Table). We used 101 bp reads for mapping, but shorter reads were employed when the input was insufficient to meet the coverage demands. The short reads used for each individual are now undergoing the registration process and will be made freely available. SNPs were accepted if the locus had at least five reads per individual and the frequency of the antagonistic allele exceeded 20%. A total of 2 million provisional SNP loci were detected from the 56 individuals. Among these loci, those with a total read count over 10,000 were excluded because excessive read coverage may indicate nucleotide mismatches from paralogous copies of duplicated sequences. In addition, only those loci that had at least five reads in each individual were retained. Accordingly, a set of 527,225 SNP loci with an average read coverage per individual of 20 was obtained. Among these reliable SNP loci, 518,706 were bi-allelic, while 8,442 were tri-allelic, and 77 were tetra-allelic.
A Bayesian clustering analysis of population structure was performed with structure version 2.3.4 [34, 35]. All 56 individuals from the 12 populations were subjected to analysis, and 10,000 SNP loci were randomly selected for the input dataset. Twenty independent runs for each value of K (the number of subpopulations) ranging from 1 to 12 were performed. For the optional setting for each run, we chose 400,000 iterations, with the first 200,000 iterations discarded as burn-in, and we applied the admixture model with correlated allele frequencies. To decide the best number of genetic clusters for the 56 individuals, we plotted the values of LnP(D) (log likelihood of the observed genotype distribution) and estimated Evanno’s ΔK [36] for each K. Based on the largest value of LnP(D) and a clear peak of ΔK, we selected 6 as the best K (Fig 2B). As we found further subdivisions within the mountains in runs with K above 6, we conducted additional analysis within each mountain. Using the same SNP loci and settings mentioned above, 20 individuals from each mountain were subjected to a set of analysis with K from 1 to 4. Although LnP(D) and ΔK supported K = 2 for Mt. Ibuki, genetic subdivision was not supported in Mt. Fujiwara (S2 Fig). Graphical representations of the results were generated using the program Distruct [67].
A maximum likelihood tree of the 12 populations was constructed with TreeMix version 1.12 [38]. This program uses a set of genome-wide allele frequency data from populations to construct the maximum-likelihood tree. Population splits are represented as nodes, and branch lengths are proportional to the amount of genetic drift experienced by the population. Migration events are inferred for populations that fit the tree poorly. Input allele frequency data for the 12 populations were generated based on 518,706 bi-allelic SNP loci. We first inferred the maximum likelihood tree without adopting a migration event, using OK0370 as an outgroup. To judge the confidence of the topology, 100 bootstrap replicates were performed. We then calculated the fraction of the variance in relatedness between populations that was explained by the tree (f of Equation 30 in [38]).
Screening of the 527,225 SNP loci was carried out according to the following three distinct criteria.
For the first criterion, we defined an index (U) to evaluate the level of unidirectional change in allele frequencies across altitudes. For each locus, the following index, ranging from −1 to 1, was calculated for each mountain:
U=|FL−FH|+|FL−FH|−|FL−FM1|−|FM1−FM2|−|FM2−FH|2
where F indicates the allele frequency of a specific nucleotide in the lowest (L), lower-middle (M1), higher-middle (M2), and highest (H) altitude-specific populations. The nucleotide showing the largest allele frequency difference between the lowest and highest populations was used to calculate each F. The index yields greater values if the difference in the allele frequency between the lowest and highest populations is larger and if the allele frequency of the intervening population falls between that in the lower and higher populations. In other words, for a given allele frequency difference between the lowest and highest population, U value is highest when the frequency increases or decreases monotonically along the altitude. For each SNP locus, we calculated
U¯
, which is an average of the U values 2 kbp down- and upstream (4 kbp window size) from its genomic position to minimize the spurious noise from single SNP locus.
The second criterion was used to evaluate the genetic difference of a given SNP locus between the lowest and highest populations within each mountain. For each SNP locus, Hedrick’s G′ST [39] were calculated and averaged across 2 kbp down- and upstream from its genomic position to obtain
G′ST¯
.
Because the preceding two criteria basically filter those genes that are highly differentiated between lowest and highest populations, genes adaptive in the lower altitude can also be detected. While those genes are also interesting, our study system focus on high-altitude convergent evolution in two distinct mountains, and thus needed a third criteria to spot the genes that are related to high-altitude adaptation. Thus the third criterion was adopted to select those loci that show increased derived allele frequency (DAF) in the highest population compared with the low-altitude reference populations. Allele frequency data from the four reference populations (16 individuals in total) were combined and the allele with minor frequency was regarded as the derived allele. DAF of the reference populations ranges from 0 to 0.5, whereas DAF of the highest population ranges from 0 to 1.0. For tri- and tetra-allelic locus, we subtracted the major allele frequency from one and used it to calculate the DAF. An index to measure the increment of DAF in the highest populations (ΔD′) was calculated by:
ΔD′=(|DH−DR|)(1−DR)
where DH is the DAF in the highest population, and DR the DAF in the reference populations. As we are not sure whether the allele is really ‘derived,’ especially for locus with high minor allele frequency in the reference populations, absolute value for the DAF difference is used. In addition, a probability of the allele being derivative (1 − DR) was used to correct the absolute DAF difference between the highest and reference populations. As well as other indices, ΔD′ values were also averaged 2 kbp down- and upstream (4 kbp window size) from its genomic position to obtain
ΔD′¯
.
For all three indices (
U¯
,
G′ST¯
, and
ΔD′¯
), we analyzed the genome-wide frequency distribution and the upper 1.5 times the IQR of a genome-wide frequency distribution (Fig 3) was determined as a screening threshold. Screening was conducted independently for each mountain, and only those SNP loci that fulfilled all three criteria were considered further. Note that the three criteria are not completely independent. For instance, a steep monotonic allele shift along the altitude is likely to be found among loci that are highly differentiated between the lowest and highest populations. See Fig 4 for the overlaps between the sets of loci screened by different criteria.
To test for enrichment of a specific gene function among the screened SNPs, we conducted a Gene Ontology (GO) enrichment analysis with 30 GO terms that cover the representative phenotypic and environmental entries within the database. (See S2 Table for the complete list of the selected GO terms). Here, only those SNP loci that were linked (overlapping or within 5 kbp of an exon) to a mapped gene in the A. halleri subsp. gemmifera reference genome were used. The ratio between ‘the number of SNP loci (or genes) associated with a given GO term within the screened dataset’ and ‘the number of SNP loci (or genes) unassociated with a given GO term within the screened dataset’ was compared with the same ratio obtained from the unscreened dataset. Significant enrichment for each GO term was computed with a one-tailed Fisher’s exact test for a 2 × 2 table [68, 69], and p-values from multiple comparisons were adjusted using a 0.05 threshold of the FDR q-value [40].
We also applied our datasets to two popular outlier detection methods that take account of the underlying population structures. Both analysis were independently conducted in each mountain. BayeScan uses a hierarchal Bayesian approach to detect outliers from the locus-specific FST distribution [44–46]. The program is based on a multinomial Dirichlet model that covers a wide range of realistic demographic scenarios. In addition, the program can be used with small number of samples with the risk of a low power, but with no particular risk of bias. We run our dataset with BayeScan 2.1 using the default parameter settings (20 pilot runs for 5,000 length, 50,000 burn in followed by additional 50,000 iteration with a thinning interval of 10). Posterior probabilities for each locus were calculated and corrected by the FDR method implemented in the program. Outliers were identified at the 1% significant levels of the FDR q-value.
Another method LFMM (Latent Factor Mixed Models) uses a hierarchal Bayesian mixed model to detect outliers from correlations between environmental and genetic variation [47]. At the same time, the program infers the background levels of population structure based on principal component analysis. Population structure is modelled from a chosen number of latent factors (K), which corresponds to the number of principal components to describe the neutral structure of the data. Underestimated value of K leads to liberal tests with false positives, whereas overestimated K leads to conservative tests with false negatives. Here, based on the results from structure analysis, we used K = 2 as a number of latent factor in each mountain. Population altitudes shown in S1 Table were used as the environmental data for each individual. Using the program lfmm in the LEA package version 1.0 (LEA: an R package for Landscape and Ecological Association studies; http://membres-timc.imag.fr/Olivier.Francois/LEA/index.htm), we conducted 20 runs with a burn in number of 5,000 and a total of 10,000 iterations. FDR q-value [40] was calculated for each locus based on the outputted p-values.
The sorting process of the candidate genes was based on the level of unidirectional change in allele frequencies across altitudes (
U¯
described above) and the effect of genetic hitchhiking. First, the
U¯
values of all SNP loci were plotted and connected with a line across genomic regions. Continuous regions with positive
U¯
values, starting and ending at the x-intercept, were considered to be hitchhiking regions (genomic island). In addition to the x-intercept, the genomic islands were terminated if the neighboring SNP loci were more than 4 kbp apart. We then defined the x-axis as the base and computed the area inside each genomic island. Genomic islands that contained at least two screened SNP loci were sorted from those with the largest area. The top 20 genomic islands contained 38 and 32 genes in the Mt. Ibuki and Mt. Fujiwara populations, respectively (see S3 Table). Finally, to visualize the unidirectional change in allele frequency, the difference in allele frequencies between the lowest and higher populations was plotted using a sliding window approach with window size of 4 kbp and a step size of 1 kbp (see S4 Fig for workflow).
We also carried out a simulation-based analysis to confirm the statistical significance of our results. To calculate the empirical p-value for obtaining two common genes from the two independent gene lists, we performed one million trials of randomly selecting 38 (number of candidate genes within the list for Mt. Ibuki) and 32 (number of candidate genes within the list for Mt. Fujiwara) genes from a set of 23,395 genes (total number of analyzed SNP-tagged genes). For each trial, we examined the number of shared genes between the two lists. Similarly, we calculated the empirical p-value for detecting three and four genes with the GO term ‘response to temperature stimulus’ in two gene lists. Again, we performed one million trials of randomly selecting 38 and 32 genes from a set of 23,395 genes. This time, however, 863 of the 23,395 genes were tagged with the GO term ‘response to temperature stimulus’ and we counted the number of genes with the GO term in the two derived gene lists.
|
10.1371/journal.pcbi.1004071 | Accurate Computation of Survival Statistics in Genome-Wide Studies | A key challenge in genomics is to identify genetic variants that distinguish patients with different survival time following diagnosis or treatment. While the log-rank test is widely used for this purpose, nearly all implementations of the log-rank test rely on an asymptotic approximation that is not appropriate in many genomics applications. This is because: the two populations determined by a genetic variant may have very different sizes; and the evaluation of many possible variants demands highly accurate computation of very small p-values. We demonstrate this problem for cancer genomics data where the standard log-rank test leads to many false positive associations between somatic mutations and survival time. We develop and analyze a novel algorithm, Exact Log-rank Test (ExaLT), that accurately computes the p-value of the log-rank statistic under an exact distribution that is appropriate for any size populations. We demonstrate the advantages of ExaLT on data from published cancer genomics studies, finding significant differences from the reported p-values. We analyze somatic mutations in six cancer types from The Cancer Genome Atlas (TCGA), finding mutations with known association to survival as well as several novel associations. In contrast, standard implementations of the log-rank test report dozens-hundreds of likely false positive associations as more significant than these known associations.
| The identification of genetic variants associated with survival time is crucial in genomic studies. To this end, a number of methods have been proposed to computing a p-value that summarized the difference in survival time of two or more population. The most widely used method among these is the log-rank test. Widely used implementations of the log-rank test present a systematic error that emerges in most genome-wide applications, where the two populations have very different sizes, and the accurate computation of very small p-values is required due to the evaluation of a number of candidate variants. Considering cancer genomic applications, we show that the systematic error leads to many false positive associations of somatic variants and survival time. We present and analyze a new algorithm, ExaLT that accurately computes the p-value for the log-rank test under a distribution that is appropriate for the parameters found in genomics. Unlike previous approaches, ExaLT allows to control the accuracy of the computation. We use ExaLT to analyze cancer genomics data from The Cancer Genome Atlas (TCGA), identifying several novel associations in addition to well known associations. In contrast, the standard implementations of the log-rank test report a huge number of presumably false positive associations.
| Next-generation DNA sequencing technologies are now enabling the measurement of exomes, genomes, and mRNA expression in many samples. The next challenge is to interpret these large quantities of DNA and RNA sequence data. In many human and cancer genomics studies, a major goal is to find associations between an observed phenotype and a particular variable (e.g., a single nucleotide polymorphism (SNP), somatic mutation, or gene expression) from genome-wide measurements of many such variables. For example, many cancer sequencing studies aim to find somatic mutations that distinguish patients with fast-growing tumors that require aggressive treatment from patients with better prognosis. Similarly, many human disease studies aim to find genetic alleles that distinguish patients who respond to particular treatments, i.e. live longer. In both of these examples one tests the association between a DNA sequence variant and the survival time, or length of time that patients live following diagnosis or treatment.
The most widely approach to determine the statistical significance of an observed difference in survival time between two groups is the log-rank test [1, 2]. An important feature of this test, and related tests in survival analysis [3], is their handling of censored data: in clinical studies, patients may leave the study prematurely or the study may end before the deaths of all patients. Thus, a lower bound on the survival time of these patients is known. Importantly, many studies are designed to test survival differences between two pre-selected populations that differ by one characteristic; e.g. a clinical trial of the effectiveness of a drug. These populations are selected to be approximately equal in size with a suitable number of patients to achieve appropriate statistical power (Fig 1A). In this setting, the null distribution of the (normalized) log-rank statistic is asymptotically (standard) normal; i.e. follows the (standard) normal distribution in the limit of infinite sample size. Thus, nearly every available implementation (e.g., the LIFETEST procedure in SAS, and the survdiff R function, and coin and exactRankTests packages in R and SPlus) of the log-rank test computes p-values from the normal distribution, an approximation that is accurate asymptotically (see S1 Text).
The design of a genomics study is typically very different from the traditional clinical trials setting. In a genomics study, high-throughput measurement of many genomics features (e.g. whole-genome sequence or gene expression) in a cohort of patients is performed, and the goal is to discover those features that distinguish survival time. Thus, the measured individuals are repeatedly partitioned into two populations determined by a genomic variable (e.g. a SNP) and the log-rank test, or related survival test, is performed (Fig 1B). Depending on the variable the sizes of the two populations may be very different: e.g. most somatic mutations identified in cancer sequencing studies, including those in driver genes, are present in < 20% of patients [4–9]. Unfortunately, in the setting of unbalanced populations, the normal approximation of the log-rank statistic might give poor results.
While this fact has been noted in the statistics literature [10–12], it is not widely known, and indeed the normal approximation to the log-rank test is routinely used to test the association of somatic mutations and survival time (e.g. [13, 14] and numerous other publications). A second issue in genomics setting is that the repeated application of the log-rank test demands the accurate calculation of very small p-values, as the computed p-value for a single test must be corrected for the large number of tests; e.g. through a Bonferroni or other multiple-hypothesis correction. An inaccurate approximation of p-values will result in an unacceptable number of false positives/negative associations of genomic features with survival. These defining characteristics of genomics applications, unbalanced populations and necessity of highly-accurate p-values for multiple-hypothesis correction, indicate that standard implementations of the log-rank test are inadequate.
We propose to compute the p-value for the log-rank test using an exact distribution determined by the observed number of individuals in each population. Perhaps the most famous use of an exact distribution is Fisher’s exact test for testing the independence of two categorical variables arranged in a 2×2 contingency table. When the counts in the cells of the table are small, the exact test is preferred to the asymptotic approximation given by the χ2 test [15]. Exact tests for comparing two survival distributions have received scant attention in the literature. There are three major difficulties in developing such a test. First, there are multiple observed features that determine the exact distribution including the number of patients in each population and the observed censoring times. With so many combinations of parameters it is infeasible to pre-compute distribution tables for the test. Thus, we need an efficient algorithm that computes the p-value for any given combination of observed parameters. Second, we cannot apply a standard Monte-Carlo permutation test to this problem since we are interested in very small p-values that are expensive to accurately estimate with such an approach. (By the accurate estimation of a p-value using Monte-Carlo permutation test, we mean the calculation of a p-value and a confidence interval of the same order of magnitude of the p-value.) Third, there are two possible null distributions for the log-rank test, the conditional and the permutational [2, 16–18]. While both of these distributions are asymptotically normal, the permutational distribution is more appropriate for genomics settings [2, 19], as we detail below. Yet no efficient algorithm is known to compute the p-value of the log-rank test under the exact permutational distribution.
We introduce an efficient and mathematically sound algorithm, called ExaLT (for Exact Log-rank Test), for computing the p-value under the exact permutational distribution. (ExaLT computes an estimate of the p-value under the exact permutational distribution; for this reason we denote the p-value obtained from ExaLT as an exact p-value, in contrast to the approximate p-value obtained from asymptotic distributions.) The run-time of ExaLT is not function of the p-value, enabling the accurate calculation of small p-values. For example, obtaining an accurate estimate of p ≈ 10−9 is required if one wants to test the association of 1% of the human genome (e.g., the exome) with survival, and using a standard MC approach requires (with the Clopper-Pearson confidence interval estimate) the evaluation of ≥ 1011 samples, that for a population of 200 patients requires > 8 days; in contrast ExaLT is capable of estimating p ≈ 10−13 on 200 patients in < 2 hours. In contrast to heuristic approaches (see Materials and Methods) ExaLT provides rigorous guarantees on the relation between the estimated p-value and the correct p-value; moreover, it returns a conservative estimate of the p-value, thus guaranteeing rigorous control on the number of false discoveries. We test ExaLT on data from two published cancer studies [20, 21], finding substantial differences between the p-values obtained by our exact test and the approximate p-values obtained by standard tools in survival analysis. In addition, we run ExaLT on somatic mutation and survival data from The Cancer Genome Atlas (TCGA) and find a number of mutations with significant association with survival time. Some of these such as IDH1 mutations in glioblastoma are widely known; for others such as BRCA2 and NCOA3 mutations in ovarian cancer there is some evidence in the literature; while the remaining are genuinely novel. Most of these are identified only using the exact permutational test of ExaLT. In contrast, the genes reported as highly significant using standard implementations of the log-rank test are not supported by biological evidence; moreover, these methods report dozens-hundreds of such likely false positive associations as more significant than known genes associated with survival. These results show that our algorithm is practical, efficient, and avoids a number of false positives, while allowing the identification of genes known to be associated with survival and the discovery of novel, potentially prognostic biomarkers.
We first assessed the accuracy of the asymptotic approximation for the log-rank test on simulated data from a cohort of 500 patients with a gene g mutated in 5% of these patients, a frequency that is not unusual for cancer genes in large-scale sequencing studies [4–7]. We compared the survival times of the population 𝓟(g) of patients with a mutation in g to the survival of the population 𝓟 ‾ ( g ) of patients with no mutation in g. We computed p-values using R survdiff on multiple random instances (in order to obtain a distribution for the p-value of g) in which 𝓟(g) and 𝓟 ‾ ( g ) have the same survival distribution. S1 Fig shows that the p-values computed by the asymptotic approximation are much smaller than expected under the null hypothesis, with the smallest p-values showing the largest deviation from the expected uniform distribution.
The inaccuracy of the asymptotic log-rank test results in a large number of false discoveries: for example, considering a randomized version of a cancer mutation dataset (S1 Table) in which no mutation is associated with survival (i.e. no true positives), the asymptotic log-rank test reports 110 false discoveries (Bonferroni correction) or 291 false discoveries (False Discovery Rate (FDR) correction), with significance level α = 0.05 (Fig 2). We found that for the number of patients of interest to current genomic studies, the inaccuracy of the asymptotic log-rank test results mostly from the imbalance in the sizes of the two populations, rather than the total number of patients or the number of patients in the smaller population (see S1a Fig, S1b Fig, S1c Fig, S1d Fig, and S1 Text).
As noted above, there are two exact distributions for the log-rank test in the literature: the permutational distribution [2] and the conditional distribution [16]. We developed an algorithm, Exact Log-rank Test (ExaLT), to compute the p-value of the log-rank statistic under the exact permutational distribution. On simulations of cancer data, we found that the p-values from the permutational exact test are significantly closer to the empirical p-values than the p-values obtained from the conditional exact test (S2 Fig and S1 Text). Thus, we derived a fully polynomial time approximation scheme (FPTAS) for computing the p-value under the permutational distribution. In contrast to heuristic methods that do not provide any rigorous guarantee on the quality of the approximation of the p-value, our algorithm provides an approximation that is guaranteed to be within a user defined distance from the p-value, for any given sample size, in polynomial time. Furthermore, the output of our scheme is always a conservative or valid p-value estimate. The C++ implementation of ExaLT (that can be called from R) is available at https://github.com/fvandin/ExaLT.
To demonstrate the applicability of ExaLT we compared the p-values from the exact distribution to p-values from the asymptotic approximation reported in two recently published cancer genomics studies [20, 21]. Huang et al. [20] divides patients into groups defined by the number of risk alleles of five single nucleotide polymorphisms (SNPs), and compares the survival distribution of the resulting populations. In one comparison, the survival distribution of 2 patients (13% of total) with at most 2 risk alleles was compared with the survival distribution of 14 patients with more than 2 risk alleles, and a p-value of 0.012 is reported. Thus, this association is significant at the traditional significance level of α = 0.05. However, ExaLT computes an exact p-value of 0.17, raising doubts about this association. In another comparison patients at a different disease stage were considered, and the division of the patients into groups as above resulted in comparing the survival distribution of 8 patients (17% of total) with the survival distribution of 40 patients, and a p-value of 6×10−6 is reported. In contrast, ExaLT computes an exact p-value of 2×10−3, a reduction of three orders of magnitude in the significance level. Additional comparisons are shown in the S1 Text. Therefore in these cases the asymptotic approximation underestimates the exact permutational p-values resulting in associations deemed more significant than what is supported by the data.
In [21], the survival distribution of 14 glioblastoma patients (11% of total) with somatic IDH1 or IDH2 mutations was compared to the survival distribution of 115 patients with wild-type IDH1 and IDH2. The reported p-value from the asymptotic approximation is 2×10−3, while the exact permutational p-value is 5×10−4, indicating a stronger association between somatic mutations in IDH1 or IDH2 and (longer) survival than reported. Notably, this same association has been reported in three other glioblastoma studies [22–24].
We analyzed somatic mutation and survival data from studies of six different cancer types (S1 Table) from The Cancer Genome Atlas (TCGA). For the range of parameters of these datasets our simulations show that the asymptotic approximation is not accurate for genes with mutation frequency ≤ 10%; we then did not considered genes with mutation frequency > 10%. We also discarded genes mutated with frequency < 1%. For each mutated gene, we first obtained an estimate p ˜ of the p-value using an MC approach, and if p ˜ ≤ 0 . 01 we used ExaLT to compute a controlled approximation of the p-value. We compared the p-value obtained in this way with the one obtained by using the asymptotic approximation as computed by the R package survdiff (S2 Table, S3 Table, S4 Table, S5 Table). Fig 3 shows the exact p-values and the R survdiff p-values for the glioblastoma multiforme (GBM) dataset and ovarian serous adenocarcinoma (OV) dataset. The p-values for the other datasets are shown in S3 Fig.
For most datasets the asymptotic p-values obtained from R survdiff are very different from the ones obtained with the exact p-values obtained by ExaLT, and the ranking of the genes by p-value is very different as well (see S1 Text). For example, in GBM none of the top 26 genes reported by R are in the list of the top 26 genes reported by the exact permutational test. Since genomics studies are typically focused on the discovery of novel hypotheses that will be further validated, this striking difference in the ranking of genes by the two algorithms is important: a poor ranking of genes by their association with survival will lead to many false discoveries undergoing additional experimental validation. While some of genes ranked in the top 10 by ExaLT are known to have mutations associated with survival (e.g., IDH1 in GBM and BRCA2 in OV), none of the top 10 genes reported by R survdiff (S5 Table) have mutations known to be associated with survival. R survdiff ranks dozens-hundreds of presumably false positives associations as more significant than these known genes. Moreover, R survdiff reports extremely strong association with survival for many of these higher ranked, but likely false positive, genes; e.g., in uterine corpus endometrial carcinoma (UCEC), 13 genes have p < 10−8 and an additional 19 genes have p < 10−5, but none of these have a known association with survival.
The top 10 genes reported by ExaLT contain several novel associations that are supported by the literature and are not reported using R survdiff. In GBM, ExaLT identifies IDH1 (p ≤ 3×10−4), VARS2 (p ≤ 5×10−3) and GALR1 (p ≤ 6×10−3), among others. As noted above, the association between mutations in IDH1 and survival has been previously reported in GBM [21–24]. A germline variant in VARS2 has been reported to be a prognostic marker, associated with survival, in early breast cancer patients [25]. The expression of GALR1 has been reported to be associated with survival in colorectal cancer [26], and its inactivation by methylation has been associated with survival in head and neck cancer [27, 28]. In OV, ExaLT identifies BRCA2 (p ≤ 4×10−3) and NCOA3 (p ≤ 10−3), and others. Germline and somatic mutations in BRCA2 (and BRCA1) have been associated with survival in two ovarian cancer studies [4, 29]. A polymorphism in NCOA3 has been associated with breast cancer [30], and its amplification has been associated with survival in ER-positive tumors [31].
Thus, the exact test implemented by ExaLT appears to have higher sensitivity and specificity in detecting mutations associated with survival on the sizes of cohorts analyzed in TCGA. Finally, we note that the exact conditional test obtains results similar to R survdiff, confirming that the the exact permutational test implemented by ExaLT is a more appropriate exact test for genomics studies. (See S1 Text.) We have also assessed the difference between the result obtained using ExaLT and the results obtained using the asymptotic permutational test (S2 Table, S3 Table, S4 Table, S5 Table). In this case the difference in the ranking of genes is reduced but still present; in particular, in COADREAD 4 of the top 10 genes identified by ExaLT are not among the top 10 genes found using the asymptotic permutational test. Moreover, there are some genes for which there is a large difference in the p-value computed by ExaLT and the p-value from the asymptotic permutational test; for example, in UCEC data CTGF is ranked first by both ExaLT and the asymptotic permutational test, but it has p = 9.6×10−5 by ExaLT and p = 8.3×10−10 by the asymptotic permutational test.
In this work we focus on the problem of performing survival analysis in a genomics setting, where the populations being compared are not defined in advance, but rather are determined by a genomic measurement. The two distinguishing features of such studies are that the populations are typically unbalanced and that many survival tests are performed for different measurements, requiring highly accurate p-values for multiple hypothesis testing corrections. We show empirically that the asymptotic approximations used in available implementations of the log-rank test produce anti-conservative estimates of the true p-values when applied to unbalanced populations, resulting in a large number of false discoveries. This is not purely a phenomenon of small population size: the approximation remains inaccurate even for a large number of samples (e.g., 100) in the small population. This inaccuracy makes asymptotic approximations unsuitable for cancer genomic studies, where the vast majority of the genes are mutated in a small proportion of all samples [4–6] and also for genome-wide association studies (GWAS) where rare variants may be responsible for a difference in drug response or other phenotype, even if it is possible that for extremely large genomic studies (e.g., with 100000 patients) the asymptotic approximations would provide results accurate enough even for imbalanced populations (e.g., when the small population is 1%).
The problem with the log-rank test for unbalanced populations has previously been reported [10, 11], but the implications for genomics studies have not received attention. Note that the issue of unbalanced populations is further exacerbated by any further subdivision of the data: e.g. by considering mutations in specific locations or protein domains; by considering the impact of mutations on a specific therapeutic regimen; by testing the association of mutations with survival in a particular subtype of cancer; by grouping into more than two populations; or by correcting for additional covariates such as age, stage, grade, etc. All of these situations occur in genomics studies.
We considered the two versions of the log-rank test, the conditional [16] and the permutational [2], and we found that the exact permutational distribution is more accurate in genomics settings. We introduce ExaLT, the first efficient algorithm to compute highly accurate p-values for the exact permutational distribution. We implemented and tested our algorithm on data from two published cancer studies, showing that the exact permutational p-values are significantly different from the p-values obtained using the asymptotic approximations. We also ran ExaLT on somatic mutation and survival data from six cancer types from The Cancer Genome Atlas (TCGA), showing that our algorithm is practical, efficient, and allows the identification of genes known to be associated with survival in these cancer types as well as novel associations. We note that ExaLT can be employed as part of permutation tests that require the computation of p-values for a large number of genomic features measured on the same set of patients [12]. While the current implementation of ExaLT handles ties in survival times by breaking them arbitrarily, its extension to different tie breaking strategies and their impact is an important future direction.
The method we propose can be generalized to assess the difference in survival between more than two groups, by considering the exact permutational distribution for the appropriate test statistic. For this reason, our method can be adapted to test the difference in survival between groups of patients that have homozygous or heterozygous mutations, or to test whether the presence of a group of genomic features has a different effect on survival compared to the presence of the single genomic features. For the same reason, our method can incorporate categorical covariates, while it is unclear how methods based on the log-rank test, as ours, can incorporate continuous covariates or how they can be used to assess specific (e.g., additive) models of interactions between genomic features and survival.
While our focus here was the log-rank test, our results are relevant to more general survival statistics. First, in some survival analysis applications, samples are given different weights; our algorithm can be easily adapted to a number of these different weighting schemes. Second, an alternative approach in survival analysis is to use the Cox Proportional-Hazards model [3]. While in the Cox regression model one can easily adjust for categorical and continuous covariates, it is not clear how to incorporate continuos potential confounders in the log-rank test that we consider. While this constitutes a limitation of our method, the Cox regression models is often used to compare two populations even when no adjustment for confounders is performed. In this case, the significance of the resulting coefficients in the regression is typically done using a test that is equivalent to the log-rank test, and thus our results are relevant for this approach as well. See S1 Text and S4 Fig.
The challenges of extending multivariate regression models to the multiple-hypothesis setting of genome-wide measurements is not straightforward. Direct application of such a multivariate Cox regression will often not give reasonable results as: there are a limited number of samples and a large number of genomic variants; and many variants are rare and not associated with survival. Witten and Tibshirani (2010) [32] recently noted these difficulties for gene expression data stating that: “While there are a great number of methods in the literature for identification of significant genes in a microarray experiment with a two-class outcome … the topic of identification of significant genes with a survival outcome is still relatively unexplored.” We propose that exact tests such as the one provided here will be useful building blocks for more advanced models of survival analysis in the genomics setting.
We focus here on the two-sample log-rank test of comparing the survival distribution of two groups, P0 and P1. Let t1 < t2 < … < tk be the times of observed, uncensored events; in case of ties, we assume that they are broken arbitrarily. Let Rj be the number of patients at risk at time tj, i.e. the number of patients that survived (and were not censored) up to this time, and let Rj,1 be the number of P1 patients at risk at that time. Let Oj be the number of observed uncensored events in the interval (tj−1, tj], and let Oj,1 be the number of these events in group P1. If the survival distributions of P0 and P1 are the same, then the expected value E [ O j , 1 ] = O j R j , 1 R j. The log-rank statistic [1, 2] measures the sum of the deviations of Oj,1 from the expectation, V = ∑ j = 1 k ( O j , 1 − O j R j , 1 R j ).
(In some clinical applications one is more interested in either earlier or later events. In that case the statistic is a weighted sum of the deviations. Our results easily translate to the weighted version of the test.) Under the null hypothesis of no difference in the survival distributions of the two groups, E[V] = 0, and Pr(∣V∣ ≥ ∣v∣) is the p-value of an observed value v. Two possible null distributions are considered in the literature, the permutational distribution and the conditional distribution (see S5 Fig).
In the permutational log-rank test [2], the null distribution is obtained by assigning each patient to population P0 or P1 independently of the survival time. Let n be the total number of patients, and n1 the number of patients in group P1. We consider the sample space of all ( n n 1 ) possible selections of survival times and censoring information from the observed data for the n1 patients of group P1. Each such selection is assigned equal probability ( n n 1 ) − 1.
In the conditional log-rank test [16], the null distribution is defined by conditioning on Oj, Rj, and Rj,1 for j = 1,…, k. If at time tj there are a total of Rj patients at risk, including Rj,1 patients in P1, then under the assumption of no difference in the survival of P0 and P1 the Oj events at time tj are split between P0 and P1 according to a hypergeometric distribution with parameters Rj, Rj,1, and Oj.
We considered the two versions of the log-rank test, the conditional [16] and the permutational [2], that differ in the null distribution they consider. The conditional log-rank test is preferred in clinical trials because it does not assume equal distribution of censoring in the two populations. This is important in clinical trials when patients in the two groups are subject to different treatments that may affect their probability of leaving the trial. However, unequal censoring is not a concern in genomic studies, since we do not expect a DNA sequence variant to have an impact of the censoring in the population. Moreover, in the genomic studies of interest to this work, the patients are not assigned to the two populations at the beginning of the study, and the measured individuals are instead repeatedly partitioned into two populations determined by a genomic variable. Therefore under the null hypothesis of no association between a genomic variable and survival time the two populations can be assumed to have the same distribution of potential follow-up times, and the correction for unequal follow-up, that is a concern in clinical trials [33], is not required in our scenario. In both versions of the test, under the null distribution the prefix sums of the log-rank statistic define a martingale, and by the martingale central limit theorem [3], the normalized log-rank statistic has an asymptotic 𝓝(0,1) distribution. (Sometimes the log-rank test is described using an asymptotic χ2 distribution; the two version of the tests are related, and our results hold for the version of the log-rank test based χ2 distribution as well (S1 Fig).) The normalizing variance is different in the two null distributions and this may be reflected in differences between the p-values obtained from the two null distributions, but asymptotically the two variances are the same [17], leading to the same p-values in the two versions of the test for large balanced populations. Therefore, the distinction between the two versions of the test is largely ignored in practice, where most papers that use the log-rank test or software packages that implement the test do not document the specific version test they consider. This can be explained, in part, by the widespread use of the log-rank in other scenarios, like clinical trials, where the issues specific to genomic settings (e.g., the imbalance between populations) do not arise. The differences between the tests are also rarely discussed in the literature, although there is some discussion [17, 19] on which variance is the appropriate one to use to compute p-values from the asymptotic approximation.
In the case of small and unbalanced populations, the two null distributions yield different p-values, and the normal approximation gives poor estimates of both (S1 Fig). On simulations of cancer data, we found that the p-values from the permutational exact test are closer to the empirical p-values than the p-values obtained from the conditional exact test (S2 Fig and S1 Text). Moreover, we prefer the permutational null distribution because it better models the null hypothesis for mutation data.
While the exact computation of p-values in the conditional null distribution can be computed in polynomial time using a dynamic programming algorithm [33], no polynomial time algorithm is known for the exact computation of the p-value in the permutational null distribution: current implementations are based on a complete enumeration algorithm, making its use impractical for large number of patients (e.g., the StatXact manual recommends using the enumeration algorithm only when the number of samples is at most 20). Several heuristics have been developed for related computations including: saddlepoint methods to approximate the mid-p-values [34], methods based on the Fast Fourier Transform (FFT) [35–37], and branch and bound methods [38]. Such heuristics are shown to be asymptotically correct, converging to the correct p-value as the number of samples and computation time grows to infinity. However, no explicit bounds are known for the accuracy of the computed p-value when these heuristics are applied to a fixed sample size and under a bounded computation time. Therefore, when run on a specific input, these heuristics do not provide guarantees on the relation between the p-value and the approximation they report. Given the systematic error we report below for the standard asymptotic implementation of the log-rank test, we argue that such guarantees are essential in this and many similar settings.
We developed an algorithm, Exact Log-rank Test (ExaLT), to compute the p-value of the log-rank statistic under the exact permutational distribution. In particular, we designed a fully polynomial time approximation scheme (FPTAS) for computing the p-value under the permutational distribution. Our algorithm gives an explicit bound on the error in approximating the true p-value, for any given sample size, in polynomial time. Furthermore, the output of our scheme is always a conservative or valid p-value estimate. Conceptually, our algorithm is similar to the one presented in [39].
Since the log-rank statistic depends only on the order of the events and not on their actual times, we can without loss of generality treat the survival data (including the censored times) as an ordered sequence of events, with no two patients having identical survival times. Let nj = ∣Pj∣, for j = 0,1, be the number of patients in each population and let n = n0+n1 be the total number of patients. We represent the data by two binary vectors x ∈ {0,1}n and c ∈ {0,1}n, where xi = 1 if the ith event was in P1 and xi = 0 otherwise; ci = 0 if the ith event was censored and ci = 1 otherwise. Note that n 1 = ∑ i = 1 n x i. In this notation the log-rank statistic is
V = V ( x , c ) = ∑ j = 1 n c j ( x j - n 1 - ∑ i = 0 j - 1 x i n - j + 1 ) ⋅ (1)
Let V t ( x ) = ∑ j = 1 t c j ( x j − n 1 − ∑ i = 0 j − 1 x i n − j + 1 ) be the test statistic V(x) at time t. Note that since n, n1, and c are fixed, the statistic depends only on the value of x. Assume the observed log-rank statistic has value v. The p-value of the observation v is the probability Pr(∣V(x)∣ ≥ ∣v∣) computed in the probability space in which the n1 events of P1 are uniformly distributed among the n events.
For any 0 ≤ t ≤ n and 0 ≤ r ≤ n1, let P ( t , r , v ) = P r ( V t ( x ) ≤ v and ∑ i = 1 t x i = r ) denote the joint probability of Vt(x) ≤ v and exactly r events from P1 occur in the first t events. Let Q ( t , r , v ) = P r ( V t ( x ) ≥ v and ∑ i = 1 t x i = r ) ⋅ denote the joint probability of Vt(x) ≥ v and exactly r events from P1 occur in the first t events.
At time 0: P(0,r,v) = 1 if r = 0 and v ≥ 0, otherwise P(0,r,v) = 0. Similarly Q(0,r,v) = 1 if r = 0 and v ≤ 0, otherwise Q(0,r,v) = 0.
Given the values of P(t,r,v) for all v and r, we can compute the values of P(t+1,r,v) using the following relations: If ct+1 = 1 then
P ( t + 1 , r , v ) = ( 1 - n 1 - r n - t ) P ( t , r , v + n 1 - r n - t ) + n 1 - ( r - 1 ) n - t P ( t , r - 1 , v - ( 1 - n 1 - ( r - 1 ) n - t ) ) ⋅
If ct+1 = 0 then
P ( t + 1 , r , v ) = ( 1 - n 1 - r n - t ) P ( t , r , v ) + n 1 - ( r - 1 ) n - t P ( t , r - 1 , v ) ⋅
Analogous equations hold for Q(t,r,v).
The process defined by these equations guarantees that the n events always include n1 events of P1. Thus, the p-value of the observation v is given by Pr(∣V(x)∣ ≥ ∣v∣) = P(n,n1,−∣v∣)+Q(n,n1,∣v∣).
For fixed t and r, P(t+1,r,v) and Q(t+1,r,v) are step functions. For example, if ct+1 = 1, then as we vary v, P(t+1,r,v) changes only at the points in which P ( t , r − 1 , v − ( 1 − n 1 − ( r − 1 ) n − t ) ) or P ( t , r , v + n 1 − r n − t ) change values. Thus, we only need to compute the function P(t+1,r,v) at these points. At t = 0 the function P(0,r,v) assumes up to 2 values. If P(t,r,v) assumes m(t,r) values and P(t,r−1,v) assumes m(t,r−1) values, then P(t+1,r,v) assumes up to m(t,r)+m(t,r−1) values.
Similar relations hold for P(t+1,r,v) when ct+1 = 0, and for computing Q(t,r,v) in the two cases. Thus, in n iterations the process computes the exact probabilities P(n,r,v) and Q(n,r,v), but it may have to compute probabilities for an exponential number of different values of v in some iterations.
We construct a polynomial time algorithm by modifying the above procedure to compute the probabilities of only a polynomial number of values in each iteration. We first observe that since the probability space consists of ( n n 1 ) equal probability events, all non-zero probabilities in our analysis are ≥ n − n 1. For ɛ > 0, fix ɛ1 such that (1−ɛ1)−n = 1+ɛ. Note that ε1 = O(ε/n). We discretize the interval of possible non-zero probabilities [ n − n 1 , 1 ], using the values (1−ɛ1)k, for k = 0,…,ℓ, where ℓ = − n 1 log n log ( 1 − ɛ 1 ) = O ( ɛ − 1 n n 1 log n ). The approximation algorithm estimates P(t,r,v) with a step function P ˜ ( t , r , v ) defined by a sequence of ℓ points v k , r t, k = 0,…,ℓ, such that v k , r t is an estimate for the largest v such that P(t,r,v) ≤ (1−ɛ1)k. We prove that if iteration t computes a function P ˜ ( t , r , v ) ( 1 − ε 1 ) t ≤ P ( t , r , v ) ≤ P ˜ ( t , r , v ), then starting from P ˜ ( t , r , v ) the t+1 iteration computes an estimate P ˜ ( t + 1 , r , v ) ( 1 − ε 1 ) t + 1 ≤ P ( t + 1 , r , v ) ≤ P ˜ ( t + 1 , r , v ). (S6 Fig provides the intuition for how the approximation at time t+1 is computed from the approximation at time t.) Thus, after n iterations we have an ε-approximations for P(n,n1,v). Similar computations obtain an ε-approximation for Q(n,n1,v). The details of the algorithm and analysis are given in the S1 Text.
We implemented the FPTAS in our software ExaLT and evaluated its performance as n and ɛ varies, and by comparing its running time with the running time of the exhaustive enumeration algorithm for the permutational test (S7 Fig). Our implementation of the FPTAS is very efficient, with significant speed-up compared to the exhaustive algorithm.
We note that, given parameters n,n1,ɛ, and given the censoring vector c, ExaLT computes the entire distribution of the test statistic under the null hypothesis. Once this distribution is computed, it can be used for any vector x ∈ {0,1}n with n 1 = ∑ i = 1 n x i and the same censoring vector c. We therefore implemented a variant of ExaLT that given multiple vectors x’s for the same set of patients (e.g., mutation data from multiple genes on the same set of patients) as input stores the distributions already computed and uses these distributions for quick lookup whenever possible. (This feature cannot be used when the vectors x’s are defined for different sets of patients, for example when there are missing gene measurements or genotypes.)
We used synthetic data to assess the accuracy of the asymptotic approximations. We generated data as follow: when no censoring was included, we generated the survival times for the patients from an exponential distribution, and the group labeling (mutated or not) were assigned to patients independently of their survival time; when censoring in c% of the patients was included, the survival times come from the exponential distribution with expectation equal to 30, and censoring variable from an exponential distribution resulting in c% of censoring (in expectation).
We used synthetic data to compare the empirical p-value and the p-values from the exact tests as well. In this case we generated synthetic data using two related but different procedures. In the first procedure, we mutate a gene g in exactly a fraction f of all patients. In the second procedure, we mutated a gene g in each patient independently with probability f. The second procedure models the fact that mutations in a gene g are found in each patient independently with a certain probability. In both cases the survival information is generated from the same distribution for all patients, as described above.
We analyzed somatic mutation and clinical data, including survival information, from the public TCGA data portal (https://tcga-data.nci.nih.gov/tcga/). In particular we considered single nucleotide variants and small indels for colorectal carcinoma (COADREAD), glioblastoma multiforme (GBM), kidney renal clear cell carcinoma (KIRC), lung squamous cell carcinoma (LUSC), ovarian serous adenocarcinoma (OV), and uterine corpus endometrial carcinoma (UCEC). We restricted our analysis to patients for which somatic mutation and survival data were both available. We only considered genes mutated in > 1% of patients. We also removed genes with mutation frequency > 10%. Since genes mutated in the same set of patients would have the same association to survival, they are all equivalent for an automated analysis of association between mutations and survival; we then collapsed them into metagenes, recording the genes that appear in a metagene. For the remaining genes we first obtained an estimate p ˜ of the p-value using a MC approach, and we then used ExaLT (with ɛ = 1.5) to compute the exact permutational p-value whenever p ˜ was ≤ 0.01. For any given TCGA dataset we used the variant of ExaLT that stores previously computed distribution and uses them for quick lookup whenever possible. The runtime was reasonable for the dataset we analysed (e.g., 310 minutes for the COADREAD dataset).
The results published here are in whole or part based upon data generated by The Cancer Genome Atlas pilot project established by the NCI and NHGRI. Information about TCGA and the investigators and institutions who constitute the TCGA research network can be found at http://cancergenome.nih.gov/.
|
10.1371/journal.pcbi.1006103 | Allostery in the dengue virus NS3 helicase: Insights into the NTPase cycle from molecular simulations | The C-terminus domain of non-structural 3 (NS3) protein of the Flaviviridae viruses (e.g. HCV, dengue, West Nile, Zika) is a nucleotide triphosphatase (NTPase) -dependent superfamily 2 (SF2) helicase that unwinds double-stranded RNA while translocating along the nucleic polymer. Due to these functions, NS3 is an important target for antiviral development yet the biophysics of this enzyme are poorly understood. Microsecond-long molecular dynamic simulations of the dengue NS3 helicase domain are reported from which allosteric effects of RNA and NTPase substrates are observed. The presence of a bound single-stranded RNA catalytically enhances the phosphate hydrolysis reaction by affecting the dynamics and positioning of waters within the hydrolysis active site. Coupled with results from the simulations, electronic structure calculations of the reaction are used to quantify this enhancement to be a 150-fold increase, in qualitative agreement with the experimental enhancement factor of 10–100. Additionally, protein-RNA interactions exhibit NTPase substrate-induced allostery, where the presence of a nucleotide (e.g. ATP or ADP) structurally perturbs residues in direct contact with the phosphodiester backbone of the RNA. Residue-residue network analyses highlight pathways of short ranged interactions that connect the two active sites. These analyses identify motif V as a highly connected region of protein structure through which energy released from either active site is hypothesized to move, thereby inducing the observed allosteric effects. These results lay the foundation for the design of novel allosteric inhibitors of NS3.
| Non-structural protein 3 (NS3) is a Flaviviridae (e.g. Hepatitis C, dengue, and Zika viruses) helicase that unwinds double stranded RNA while translocating along the nucleic polymer during viral genome replication. As a member of superfamily 2 (SF2) helicases, NS3 utilizes the free energy of nucleotide triphosphate (NTP) binding, hydrolysis, and product unbinding to perform its functions. While much is known about SF2 helicases, the pathways and mechanisms through which free energy is transduced between the NTP hydrolysis active site and RNA binding cleft remains elusive. Here we present a multiscale computational study to characterize the allosteric effects induced by the RNA and NTPase substrates (ATP, ADP, and Pi) as well as the pathways of short-range, residue-residue interactions that connect the two active sites. Results from this body of molecular dynamics simulations and electronic structure calculations are highlighted in context to the NTPase enzymatic cycle, allowing for development of testable hypotheses for validation of these simulations. Our insights, therefore, provide novel details about the biophysics of NS3 and guide the next generation of experimental studies.
| Flaviviruses (family Flaviviridae) are small (∼11 kilobases) positive-sense, single-stranded RNA (ssRNA) viruses that include members such as dengue (serotypes 1-4), Zika, West Nile, yellow fever, and Japanese Encephalitis viruses. The dengue virus (DENV) is a public health threat that causes serious morbidity and mortality globally [1, 2]. Infection with DENV can result in “break-bone” fever, an extraordinarily painful disease with symptoms ranging from a mild fever to a fatal hemorrhagic syndrome [3]. There are approximately 50 million serious infections and 20,000 deaths each year, and dengue infections are a leading cause of mortality in children in a number of Latin and Asian countries [1]. Dengue viruses have re-emerged in the United States, and a growing number of locally acquired infections in Florida, Texas, and Hawaii have been reported over the last decade. Despite a reinvigorated effort due to the recent Zika epidemic [4], there are currently no approved small molecule antivirals to treat Flavivirus-induced diseases.
One of the primary antiviral targets in Flaviviridae is the nonstructural protein 3 (NS3), which plays a critical role in the viral replication cycle [5–15]. NS3 is a multifunctional protein found in all Flaviviridae, possessing an N-terminal serine protease domain responsible for proteolytically cleaving the viral polyprotein during translation [16] and a C-terminal helicase/nucleotide triphosphatase (NTPase)/RNA triphosphatase domain [17–22]. In a nucleotide triphosphate (NTP) hydrolysis-dependent mechanism, the NS3 helicase domain (NS3h) unwinds double-stranded RNA (dsRNA) while translocating along the nucleic polymer. These functions are required to resolve the dsRNA replication intermediate into fully-mature positive strand RNAs (see Ref. [23] for a recent review). Mutations in the NS3 helicase and NTPase active sites are seen to abrogate NS3 function as well as decrease viral survival [24–26], demonstrating the importance of these enzymatic functions to the flavivirus life cycle. Drugs identified to inhibit DENV NS3h suffer from specificity issues because they are either NTPase inhibitors [27] or RNA/DNA mimics such as ivermectin [13], suramin [14] or aurintricarboxylic acid [15]. Therefore, it is of interest to further elucidate the mechanism of DENV NS3h with molecular resolution to help identify new and specific target regions for antiviral therapeutics.
The Flaviviridae NS3h have been classified as a superfamily 2 (SF2) helicase (NS3/NPH-II subfamily; a DEx/H helicase) where the NTPase cycle (Fig 1) provides the free energy needed to unwind dsRNA and translocate along the nucleic substrate in a 3′ to 5′ direction [28]. Structurally, NS3h are monomeric helicases composed of three subdomains; subdomains 1 and 2 (red and orange in the inset of Fig 1) are RecA-like folds that are structurally conserved across all SF1 and SF2 helicases, whereas subdomain 3 (green) is unique to the NS3/NPH-II subfamily and contains some of the least conserved portions of the protein. In Fig 1, an adenosine triphosphate (ATP; purple) molecule is bound within the NTPase active site between subdomains 1 and 2. Also, an RNA substrate (blue) is bound within the RNA-binding cleft, separating subdomains 1 and 2 from subdomain 3. The 5′ terminus of the RNA is positioned at the top of the protein in Fig 1 and the ds/ss RNA junction is hypothesized to be just above this region of the protein.
The NS3/NPH-II subfamily of SF2 helicases exhibit both RNA-stimulated NTPase activity and NTPase-dependent helicase activity [17–22]. These experimentally observed phenomena suggest that (1) the presence of RNA affects the NTPase active site, thereby activating the NTPase cycle and (2) this cycle is the source of free energy needed to perform work on the RNA (translocation and unwinding). In Fig 1, the enzymatic cycle for the NTPase function is depicted by four dynamic events: RNA is bound within the RNA-binding cleft and activates the NTPase cycle, NTP binds, NTP is hydrolyzed, and finally products (nucleotide diphosphate—NDP—and inorganic phosphate—H2PO4-, Pi) are released. To date, it is unclear which stage(s) of the cycle are responsible for the translocation and unwinding functions of NS3h. Furthermore, the biophysical couplings between NTPase and helicase active sites are still poorly understood [28].
One of the better studied Flaviviridae NS3h is that of the Hepatitis C virus (HCV; family: Flaviviridae hepacivirus) [29–40]. Utilizing both ensemble [29–35] and single molecule [36–38, 41, 42] techniques, studies have provided insights into the kinetic steps of the HCV NS3h translocation function. These studies, alongside crystallography studies of various Flaviviridae NS3h, suggest that the NS3 enzyme tracks along the phosphodiester backbone of the nucleic oligomer, unwinding one base-pair per hydrolysis event [35–37]. To explain these experimental results, various models describing the translocation mechanism have been reported, depicting NS3h as a Brownian [33–35] or backbone stepping motor [36, 39–41] protein. These models envision the coupling between NTPase and helicase functions through different biophysical mechanisms, yet the models are not mutually exclusive and are limited in temporal and spatial resolution [43, 44].
Luo et al. reported a set of crystal structures of the DENV NS3h in important protein-substrate complexes of the NTPase cycle (bolded text in Fig 1) [45]. From these structures, major allosteric influences of RNA-binding were seen in the NTPase active site. For example, Luo and coworkers noted that the presence of an RNA substrate shifts the carboxylate group of Glu285 (motif II) into a more catalytically relevant structure for the hydrolysis reaction. Mutation of the Glu285 residue abrogates NTPase and helicase activities [25]. These static structures have provided novel insights into RNA-induced protein structural changes yet provide limited insight into the NTPase cycle or translocation and unwinding functions of NS3h.
Previous theoretical studies of helicases have focused on a broad range of enzymes such as PcrA (SF1) [46–49], transcription terminator Rho (SF5) [50], SV40 (SF3) [51], and various NS3h enzymes [52–56]. Of the theoretical studies on NS3h, Perez-Villa et al. reported microsecond-long molecular dynamics (MD) simulations of the HCV NS3h-ssRNA systems in the presence and absence of ATP and ADP. The reported simulations were used to interrogate the thermodynamics of these substrate states with various conformations of the NTPase active site [52]. While the reported results are of interest for NS3h, the authors provide limited insight into the molecular mechanisms at play during the NTPase cycle. Other theoretical studies of the NS3h enzyme are limited in timescales (tens to hundreds of ns of simulation), substrate states modeled, or spatial resolution (e.g. coarse grained elastic network model) [53–56]. Therefore, theoretical modeling of the NS3h enzyme has yet to elucidate further details about the structural and dynamic couplings within NS3h in light of the NTPase cycle.
We report here a multiscale theoretical study of the DENV NS3h enzyme at each substrate state along the NTPase cycle. RNA-induced allostery on the NTPase active site is reported wherein the presence of an RNA substrate alters the positioning and dynamics of waters within the hydrolysis active site. Inspired by this observation, minimum energy electronic structure calculations are performed to investigate the energy landscape of the hydrolysis reaction. Additionally, investigations into NTPase substrate-induced allostery on the RNA-binding cleft suggest that NS3h interacts with RNA in a NTPase substrate-dependent manner. Umbrella sampling (US) simulations are performed to enhance the sampling of a proposed elementary step of the translocation mechanism observed during the unbiased simulations. Finally, analyses of the correlated motions between residues are used to identify allosteric pathways that connect the two active sites. It is through these pathways that we hypothesize that free energy released during the NTPase cycle is transduced to the RNA-binding cleft and utilized to perform work on the RNA. This study of the substrate states of DENV NS3h lays the foundation for further study of the NTPase cycle and marks the most complete picture of the molecular mechanism of the NS3 NTPase/helicase to date.
A subset of the crystal structures reported by Luo et al. [45] of the Dengue NS3h (serotype 4) are used as the initial structures for all-atom, explicit solvent MD simulations. Specifically, the binary complex of NS3h with a seven-residue ssRNA substrate (PDB ID: 2JLU) is used to model the ssRNA substrate state, while the ternary structures of ssRNA+ATP (2JLV), ssRNA+ADP+Pi (2JLY), and ssRNA+ADP (2JLZ) model the pre-hydrolysis, post-hydrolysis, and product release states of the NTPase cycle, respectively. The Apo (2JLQ) and ATP (2JLR) substrate states are also simulated and used as experimental controls for our investigation into allostery.
The RNA-bound structures of DENV NS3h were crystalized as dimers of the protein [45]. For these systems, chain A of the structure is used as the starting conformation. Furthermore, the A conformers are chosen for residues with multiple side chain conformations. In all crystal structures with ATP substrates, the crystalized Mn2+ divalent cation is converted into a Mg2+. For the ATP crystal structure (2JLR), residues of the protease linker region were poorly resolved and so are transferred from the Apo (2JLQ) structure after aligning the neighboring amino acid backbones in both systems.
All-atom, explicit solvent MD simulations are performed for the six substrate states of DENV NS3 and presented in Fig 1 (denoted Apo, ATP, ssRNA, ssRNA+ATP, ssRNA+ADP+Pi, and ssRNA+ADP). The simulations are performed using the GPU-enabled AMBER14 software [57], ff14SB [58] parameters for proteins, and ff99bsc0χ OL3 [59, 60] parameters for RNA. Parameters for ATP [61], ADP [61], Pi (provided in Supplementary Information (SI); S2 File), and Mg2+ [62] are also used. For each system, the crystal structures are solvated in TIP3P water boxes with at least a 12 Å buffer between the protein and periodic images. Crystallographic waters are maintained. Sodium and chloride ions are added to neutralize charge and maintain a 0.10 M ionic concentration. The Langevin dynamics thermostat and Monte Carlo barostat are used to maintain the systems at 300 K and 1 bar. Direct nonbonding interactions are calculated up to a 12 Å distance cutoff. The SHAKE algorithm is used to constrain covalent bonds that include hydrogen [63]. The particle-mesh Ewald method [64] is used to account for long-ranged electrostatic interactions. A 2 fs integration time step is used, with energies and positions written every 2 ps. The minimum amount of simulation performed for each system is one trajectory of 1.5 μs, with the first 200 ns of simulation sacrificed to equilibration of the starting structures. Simulation of the ssRNA system is performed to 2 μs. For both the ATP and ssRNA+ATP systems, two 1.5 μs simulations are performed. The total amount of unbiased simulation reported here on the described structures is 12.5 μs.
US simulations are performed to enhance sampling of a hypothesized elementary translocation event wherein the biased collective variable is the distance between the central carbon of the guanidinium group of Arg387 to the phosphorous atom of phosphate 4 in the RNA. These simulations are run for the ssRNA, ssRNA+ATP, ssRNA+ADP+Pi, and ssRNA+ADP systems, using the same protocol as the unbiased simulations with the addition of a bias. For each substrate state, a minimum of 22 sampling windows are simulated for 50 ns each with harmonic wells positioned every 0.5 Å and ranging from 3.50 to 14.00 Å. Harmonic force constants are 20 kcal mol-1 Å-2. Further simulation and additional windows are run in regions of collective variable space with poor sampling. The weighted histogram analysis method (WHAM) [65] is used to analyze the results of these simulations, with bin sizes of 0.1 Å. Bootstrapping is used to approximate error bars for the probability density and free energy plots shown. The total amount of biased simulations reported here is 5.12 μs.
Electronic structure calculations are performed at the ωB97X-D/6-31+G* level of theory [66] using the Guassian 09 version B.01 program [67]. The ωB97X-D functional is chosen due to its broad applicability [68, 69] and a recent study demonstrating its energetic accuracy for a variety of phosphate hydrolysis reactions [70]. The QM system is composed of a truncated ATP molecule (truncated to methyl triphosphate, MTP), functional groups of nine surrounding protein residues (Pro195, Gly196, Lys199, Glu285, Ala316, Gly414, Gln456, Arg460, and Arg463), a Mg2+ ion, and seven water molecules. The amino acids are truncated at various positions (more detail in S1 File) using hydrogen atoms. For each residue, the position of the terminal heavy atom is frozen to maintain the active site geometry. This yielded a total of 138 atoms in the QM calculations.
These calculations are performed on active site conformations pulled from the unbiased MD simulations of the ssRNA+ATP and ATP substrate states, thereby investigating the influence of observed RNA structural allostery on the hydrolysis reaction mechanism and energy landscape. Frames used for the initial reactant state structures were selected by visualizing MD frames in which a lytic water is present. Through visual and RMSD analyses of such frames, a single frame was chosen to represent the population of catalytically relevant structures. The hydrolysis reaction is then monitored by optimizing the reactants (MTP+lytic water), products (MDP+HPO42−), and a single transition state (TS) in between. The initial TS and product state structures were created from the previous optimized structure. The minima are confirmed using a Hessian calculation. The TS is confirmed by examining the direction of the single imaginary frequency. Following geometry optimization, frequency calculations are performed to obtain gas-phase, zero-point energy corrected free energies for each active site conformation.
Unless stated otherwise, analyses of MD trajectories are performed using Python 2.7 and the MDAnalysis module (version 0.15.0) [71]. Matplotlib is used for plotting data [72]. VMD is used for visualization of trajectories and production of structural figures [73–75]. For each substrate state, a single frame from the trajectories is used when presenting structural details of the respective substrate state. Further information on choosing these “exemplar” structures is given in the S1 File. Additionally, details of all analyses performed can be found in the S1 File. All scripts for the analyses are available on Github (https://github.com/mccullaghlab/DENV-NS3h).
For clarity, we present and discuss our results in three sections. The first and second sections independently report observed RNA-induced and NTPase substrate-induced structural allosteries, respectively. The focus of the RNA-induced allostery section is on the structural changes seen in the NTPase active site due to bound RNA. Similarly, the NTPase substrate-induced allostery section highlights changes seen in the structure and dynamics of the RNA-binding cleft due to the presence of different nucleotide substrates. In the final section, correlated motions between residues are used to highlight pathways through which these structural allosteric effects are induced.
To date, no biophysical explanation has been proposed for the 10 to 100-fold increase in NTPase turnover rate observed for DENV NS3h in the presence of RNA [22]. Crystallographic studies of the DENV NS3h structure have identified static structural allostery due to RNA binding [45], yet a dynamic picture and interpretation of these influences are still missing. In this section, comparisons of the simulations of the Apo, ATP, ssRNA, and ssRNA+ATP substrate states are used to depict structural rearrangements induced by RNA. These RNA-induced allosteries are observed to affect the positioning and dynamics of waters within the NTPase active site. These novel insights gained from the comparisons of the MD simulations inspire the reported electronic structure calculations of the reactant, transition, and product states of the hydrolysis reaction. In combination, these results demonstrate that the observed enhancement of NTPase activity originates from the RNA-induced destabilization of the lytic water.
Experimental studies have shown that the NS3h helicase functions (translocation and unwinding) are NTPase dependent, yet it is unclear which equilibrium states and/or dynamic events of the NTPase cycle are the source of the necessary free energy for these functions [20, 21]. All previously developed models describing these functions have deduced that the NTPase cycle drives conformational changes in the RNA-binding cleft, thereby cycling the protein-RNA interactions leading to unidirectional translocation and melting of the duplex/single stranded nucleic junction [33–36, 39–41]. Yet, limited structural allostery attributed to the NTPase substrates (e.g. ATP, ADP, and Pi) is observed in the crystal structures of DENV NS3h [45]. Therefore, a subset of the MD simulations reported here (ssRNA, ssRNA+ATP, ssRNA+ADP+Pi, and ssRNA+ADP) is used to interrogate protein-RNA interactions as well as identify protein structural changes that have NTPase substrate-dependent behaviors.
The current view of allosteric regulation focuses on signal transduction through complex, 3-dimensional networks, brought about by intrinsic structural and/or dynamic changes along pathways connecting two distal, non-overlapping active sites [96–98]. These allosteric pathways are described by coupled short-range, residue-residue interactions that lead to long-range correlations. In the previous two sections, RNA-induced and NTPase substrate-induced structural rearrangements have been presented. In this section, these allosteric structural changes are absorbed into a unified description of the allosteric pathways connecting the RNA-binding cleft with the NTPase active site.
Dynamic network analyses, such as residue-residue correlations, have been used to identify allosteric pathways within proteins from simulation [96–100]. A growing body of literature has highlighted the functional importance of such pathways as well as the fundamental residue-residue interactions leading to their emergence [96–102]. We report here residue-residue distance correlation analyses that are used to identify the allosteric pathways present within the DENV NS3h protein. Focus is given to the motifs discussed in the previous sections (α2, motifs II and IVa) due to the observed structural rearrangements. Additionally, the correlation heat maps are used to identify segments of the protein that experience strong correlations with numerous other regions of the protein, such as motif V. While motif V does not experience substrate-induced structural rearrangements, the strong correlations between motif V and motifs in both the NTPase active site and RNA binding cleft are hypothesized to have functional importance in the signal transduction mechanism of allosteric regulation. Unlike the previous two sections, comparisons between substrate states (RNA-bound and NTPase substrate-bound) are not considered here. Instead, focus is given to the discussion of the residue-residue distance correlation analysis of the ssRNA+ATP substrate state.
Through analyses of the reported simulations, molecular observables of RNA- and NTPase substrate-induced allostery were identified. Specifically, an RNA bound within the RNA-binding cleft affects the dynamics and positioning of water molecules within the NTPase active site. This allosteric influence is conferred from the RNA-binding cleft to the hydrolysis active site through structural rearrangements of Lβ3β4, α2, and motif II. These RNA-induced structural changes lead to an entropic destabilization of the NTPase active site as well as a direct destabilization of the lytic water. Inspired from these results, electronic structure calculations were used to investigate the energetics of NTP hydrolysis reaction. The energetic landscapes obtained from the DFT calculations demonstrate that RNA decreases the activation barrier as well as affects the mechanism of the hydrolysis reaction. Combining these results into a kinetic model allowed for the calculation of a theoretical RNA-stimulated NTPase activity enhancement factor of 150, which qualitatively matches the experimentally observed enhancement factor. Therefore, results from MD and DFT calculations provide novel, multiscale insight into the RNA-induced allosteric effects that stimulate the catalysis of the NTP hydrolysis reaction in DENV NS3h.
Unlike RNA, the NTPase substrates are smaller perturbations to the NS3h structure and dynamics. Protein-RNA interaction energies were used to investigate the NTPase substrate-dependence of protein-RNA contacts in the unbiased MD simulations. From these analyses, the protein-RNA phosphodiester backbone interactions were observed to be NTPase substrate-dependent. The presence of the γ-phosphate (or Pi) of the NTPase substrate was observed to strengthen the protein-RNA contacts. Furthermore, the localized nonbonding interaction energies demonstrate a large shift in protein-RNA contacts, originating in part from the side chain conformational states of Arg387. Results from US simulations demonstrate that the Arg387 side chain conformational states exemplify NTPase substrate-dependent protein-RNA interactions. With the purview of the NTPase cycle, transitions between conformational states leads to 3′ to 5′ translocation. Therefore, we hypothesize that the transition between Arg387 side chain conformations is an elementary step in the unidirectional translocation mechanism of NS3h along the phosphodiester backbone of RNA.
Finally, consideration of these allosteric effects independent of one another provides an incomplete picture of the biophysics of the NS3h protein. Residue-residue correlation analyses were used to identify structural regions of the protein that experienced correlated motions with other regions. These analyses were used to describe the allosteric pathways that connect α2 with motifs II and IVa. The short-range, residue-residue interactions were presented that connect the RNA-binding cleft to the NTPase active site. Furthermore, the correlation heat maps allow for identification of regions of the protein that experience strong correlated motions with numerous other regions. Motif V is one such example, where the segment of 13 residues has strong coupled motions with seven other motifs in subdomains 1 and 2. This highly correlated nature suggests that motif V functions as a centralized communication hub that connects distal portions of the protein structure.
Complete modeling of a revolution of the NTPase cycle in the DENV NS3h presents a significant challenge for current computational methodologies. Rather, we have divided the cycle into equilibrium substrate states and dynamic events where the protein transitions from one substrate state to another. The simulations reported have modeled the important NTPase cycle substrate states, leading to novel insights into the function and underlying biophysics of the DENV NS3h enzyme with focus given to the allosteric connections between the RNA-binding cleft and NTPase active site. We hypothesize that the observed allosteric effects and pathways have important roles in the transduction of energy from one active site to the other during the dynamic events of the NTPase cycle. Therefore, the results reported have laid an initial foundation for theoretical investigations into the dynamic events of the NTPase cycle.
Beyond further theoretical modeling of NS3h, transverse relaxation-optimized spectroscopy (TROSY) NMR, mutational biochemical studies, and targeted small molecule binding experiments can be envisioned to test the hypothesis and results presented. Nuclear magnetic resonance has been previously used to study dynamics within isolated HCV NS3 helicase subdomains [103–105], but the size of the full dengue NS3h domain is too large for traditional NMR approaches due to line width increasing with increased molecular mass. However, TROSY NMR has been developed that may allow for experimental monitoring of fast NS3h dynamics [106]. We anticipate that perturbation of the wild-type structure or dynamics of the allosteric pathways in NS3h will lead to abrogation of NTPase and/or helicase functions, and are currently developing assays to test this hypothesis with dengue NS3h. Residues active in the allosteric pathways are viable targets for mutational studies where varying a specific amino acid residue is expected to alter the short-range, residue-residue interactions and lead to a destabilization of the pathway connecting the two active sites. This is hypothesized to result in reductions in enzymatic activity and would be observable in our biochemical assays. Additionally, these pathways are viable targets for theoretical and experimental small molecule drug docking experiments with focus given to molecules that disrupt the residue-residue interactions along the pathway. Co-crystallization of specific conformation-binding molecules may lock NS3h into transition-state conformations that can help verify our computational studies. Molecular candidates also have the potential for inhibiting NS3h function during replication and being specific to the NS3/NPH-II subfamily of SF2 helicases.
|
10.1371/journal.ppat.1005459 | When Parasites Are Good for Health: Cestode Parasitism Increases Resistance to Arsenic in Brine Shrimps | Parasites and pollutants can both affect any living organism, and their interactions can be very important. To date, repeated studies have found that parasites and heavy metals or metalloids both have important negative effects on the health of animals, often in a synergistic manner. Here, we show for the first time that parasites can increase host resistance to metalloid arsenic, focusing on a clonal population of brine shrimp from the contaminated Odiel and Tinto estuary in SW Spain. We studied the effect of cestodes on the response of Artemia to arsenic (acute toxicity tests, 24h LC50) and found that infection consistently reduced mortality across a range of arsenic concentrations. An increase from 25°C to 29°C, simulating the change in mean temperature expected under climate change, increased arsenic toxicity, but the benefits of infection persisted. Infected individuals showed higher levels of catalase and glutathione reductase activity, antioxidant enzymes with a very important role in the protection against oxidative stress. Levels of TBARS were unaffected by parasites, suggesting that infection is not associated with oxidative damage. Moreover, infected Artemia had a higher number of carotenoid-rich lipid droplets which may also protect the host through the “survival of the fattest” principle and the antioxidant potential of carotenoids. This study illustrates the need to consider the multi-stress context (contaminants and temperature increase) in which host-parasite interactions occur.
| Virtually all free-living organisms are infected by parasites. Moreover, both parasites and hosts may be exposed to increasing levels of pollution and might be affected by climate change. However, few studies have considered the environmental context in which parasites and hosts interact, and the relationships between these factors remains poorly understood. It is assumed that infection with parasites increases mortality under a cause of stress such as pollution. We studied the combined effect of arsenic (As) pollution, temperature increase and infection by tapeworms on the health of the economically and ecologically important brine shrimp Artemia. We found that tapeworms make Artemia more resistant to As, a major pollutant in aquatic environments, even under increased temperature conditions. These parasites increase the capacity of antioxidant enzymatic defenses, allowing infected individuals to cope better with As. Moreover, tapeworms increase fat reserves in their hosts, which may be advantageous due to the ability of lipids to sequester pollutants (“survival of the fattest” principle). Although our results may be unusual, we find a clear explanation for them. This makes them of broad significance and general interest.
| Many aspects of host-parasite interactions have been studied in detail, from molecular mechanisms to adaptive strategies and their ecological and evolutionary consequences (reviewed in Schmid-Hempel 2011 [1]). In contrast, few studies have considered the context of multiple environmental stressors in which host-parasite interactions occur in natural conditions. Consequently there are important limitations to our understanding of the ecology and evolution of host-parasite interactions, and to our ability to reach reliable conclusions. For example, increasing numbers of taxa (both, free living and parasites) are exposed to pollution and impacted by climate change [2]. However, the complex relationships between these factors (i.e. parasites, pollution and climate change) are not well understood.
Parasites and hosts can react differently to pollutants, influencing their mutual interactions. For instance, if the parasite is more susceptible than the host to the pollutant, the exposure to pollution provides the indirect benefit (for the host) of protecting against the parasite. Conversely, the epidemiology of the parasite may be altered negatively if the pollutant impacts the life history of the host (e.g. reducing survival), thus compromising parasite transmission. Moreover, parasites and pollution can interact to affect the health of the host, the central topic of the emerging field of “Environmental Parasitology” [3]. Most studies evaluating the joint effect of parasites and pollution on the health of free-living organisms find that there are additive or synergistic effects between these stressors [4,5,6]. For example, coinfection of amphipods by acanthocephalans and microsporidians led to a reduction in antitoxic defenses when exposed to cadmium [7].
However, parasites can influence multiple facets of host phenotype [8], including physiology, behavior and biochemistry, so they may change the host response to a pollutant in diverse and complex ways. The physiology of the parasite itself may have a direct influence, e.g. through the capacity observed in several parasite groups to bioaccumulate contaminants [9,10]. Therefore, more studies are required to understand the diversity of ways parasites and pollution can interact to affect the health of organisms. Of particular concern are the mechanistic (proximate) bases of host-parasite interactions; for example the potential of oxygen-free radicals and other reactive oxygen species (ROS) to induce oxidative damage in tissues and cellular components, leading to adverse health effects and diseases. The evaluation of the levels of antioxidant enzymes in infected and uninfected organisms can provide valuable information on antioxidant status and their capacity to confront multiple stress conditions.
Furthermore, all these effects are likely to depend on physiological variation which is influenced by temperature and hence by climate change. The projected temperature increase this century [11] is likely to increase the toxicity of pollutants [12] as well as the transmission, distribution and abundance of many parasites [13,14,15]. To date, studies of interactions between parasites and pollutants on hosts have only been carried out at a single temperature, with the exception of one investigation [16] dealing with the effect of trematode infection and seasonal temperature on bioaccumulation of xenobiotics in freshwater clams.
Brine shrimps Artemia spp. (Branchiopoda, Anostraca) are economically and ecologically important. They are used as model organisms in aquatic toxicology [17, 18] and are the most important live food used in aquaculture worldwide [19]. Artemia are the dominant macroinvertebrate in hypersaline ecosystems around the world, and many waterbirds depend on them as food [20,21,22]. They control phytoplankton populations [23] and are also intermediate hosts for a rich community of avian cestodes [24]. These parasites cause strong physiological and behavioural changes in Artemia [25,26,27] which may be expected to influence their response to pollutants.
We studied the Artemia parthenogenetica population from the Odiel and Tinto estuary, SW Spain, one of the most polluted estuarine systems in Western Europe [28]. A. parthenogenetica from Odiel have high levels of cestode infection [29], especially of Flamingolepis liguloides which uses flamingos as final hosts, and Confluaria podicipina which infects grebes [30]. Arsenic is a highly toxic and bioaccumulable metalloid that originates from both anthropogenic and natural sources and causes detrimental effects in humans and wildlife [31]. Arsenic in the Odiel and Tinto estuary originates from historic and current mining activity [32,33,34]. An estimated 12 t yr-1 and 23 t yr-1 of As is transported by the Tinto and Odiel rivers (respectively) into the Atlantic Ocean [35]. Previous studies have found high levels of inorganic arsenic in sediments (85–610 ppm) in the study area [36], exceeding the ERM (Effects Range Median) for marine and estuarine sediments (70 ppm [37]), and also the Canadian sediment quality guideline (CSQG) value for the protection of aquatic life (7.2 ppm [38]).
This study was designed to test the individual and combined effects of infection by cestode parasites and a 4°C temperature change on the sensitivity of Artemia to acute As exposure. We conducted toxicity tests to compare As median lethal concentrations at 24 h (24-h LC50) for infected and uninfected individuals on two separate dates one month apart, in order to test the effect of different parasite assemblages on arsenic sensitivity. Since cestode parasites in A. parthenogenetica have a strong seasonal pattern [29], and different parasite species induce different physiological effects in their host, we chose to consider the effects of different parasite compositions (i.e dates). We also evaluated the effect of parasites on the antioxidant defense mechanisms of Artemia, in order to measure the capacity of infected animals for detoxification of reactive oxygen species caused by factors such as pollution or climate change. We included the activity of four important enzymes: superoxide dismutase (SOD) responsible for the detoxification of the highly reactive oxygen species superoxide anions; catalase (CAT) which catalyses the decomposition of hydrogen peroxide in 02 and H20; glutathione peroxidase (GPX) which acts as a scavenger for high concentrations of hydrogen peroxide; and glutathione reductase (GR) implicated in the reduction of glutathione disulphide to the sulphydryl form glutathione, which is a critical molecule in combating oxidative stress. In order to evaluate the efficacy of the above enzymatic mechanisms for control of reactive oxygen species, we also measured the levels of thiobarbituric acid reactive substances (TBARS), which indicates the extent of lipid peroxidation as a consequence of oxidative damage. Finally, we quantified the carotenoid-rich lipid droplets of infected and uninfected individuals. These droplets serve as intracellular lipid storage but also play a protective role in protecting cells against oxidative stress.
It was impossible to analyse oxidative stress in individuals used in acute toxicity tests because of the need to crush the specimens in order to check for parasites, so we conducted a separate sampling in May 2015. Parasite prevalence has a seasonal pattern [29], so in order to have a similar parasite assemblage, samples were collected in the same calendar month as those used in toxicity tests and from the same ponds of medium-high salinity (140–220 g/l). One subsample was used for oxidative stress analysis and another to characterize parasite assemblage and to quantify lipid droplets. Infected and uninfected individuals were recognised on the basis of their colour [392]. Table 3 shows the infection index of infected individuals. As expected, infection was dominated by F. liguloides, and its prevalence did not differ from the sampling of May 2014 (Mann-Whitney U test, U = 13158, P = 0.585). Prevalence and abundance of minority cestode species varied among dates.
The analysis of oxidative stress in infected and uninfected Artemia revealed the association of parasites with strong changes in the antioxidant capacity of the host. Infected individuals showed significantly higher levels of CAT (Fig 5, t = -2.892, P = 0.02) and GR activities (Fig 5, t = -2.881, P = 0.002), whereas SOD was higher in uninfected individuals (t = 2.739, P = 0.03). Conversely, parasites had a negligible effect on GPX (t = -0.814, P = 0.447) and TBARS (t = -1.781, P = 0.125).
We also quantified lipid droplets, which are readily visible throughout the body and appendices of adult Artemia (Fig 2B). Lipid volume estimates are strongly correlated with biochemical measurements of lipids [41], and are highly related with the ability of organisms to protect themselves against pollutants [42]. Infection with cestodes was associated with increased number of lipid droplets (Mann-Whitney U test, U = 45.5, P < 0.001, n = 20; Fig 6).
This study tested the effect of parasites on Artemia sensitivity to As, and explored the physiological crosstalk between the parasite and the host, measuring oxidative stress and lipid content in infected and uninfected Artemia. Our study provides the first empirical evidence that parasites can increase resistance to metal or metalloid pollution, rather than decrease it. It is also the first study to consider the influence of temperature change on parasite-pollutant interactions. In three separate acute toxicity experiments, Artemia infected with cestodes consistently showed lower sensitivity to As than uninfected individuals. The higher sensitivity of infected Artemia in April suggests that multiple infections may reduce the benefits of cestode infection to host resistance.
Our results contradict the pre-existing view that pollution and parasites are stressors that both have negative effects on the health of free living organisms. This view was based on previous field and laboratory investigations (including chronic and acute exposure to a wide variety of toxicants, in vertebrates and invertebrates, intermediate and definitive hosts, and in several groups of parasites [43,44,45].
The results of oxidative stress analysis provide a mechanistic explanation for our findings. Infected individuals exhibited much higher levels of CAT and GR, reflecting a superior ability to combat the effects of exposure to pollutants with oxidative potential, such as As. The particularly high levels of CAT in infected individuals (nearly double that of uninfected Artemia) is related to the increased levels of hemoglobins in F. liguloides-infected Artemia compared with uninfected individuals (Fig 2C and 2D). CAT is a haematin protein complex with four porphyrin heme groups that allow the enzyme to react with hydrogen peroxide. There is a close linear relationship between CAT activity and hemoglobin concentration in human blood [46]. CAT also has one of the highest rates of all enzymes; one CAT molecule is able to catalyse the conversion of 5 million molecules of hydrogen peroxide per second to water and oxygen. Thus, even if the level of SOD was lower in infected individuals, the control mechanisms via CAT and GR seem to be sufficient to avoid the establishment of an oxidative stress condition, as indicated by the lack of changes in TBARS between infected and uninfected individuals. TBARS is a by-product of lipid peroxidation, so this result indicates that parasites are not inducing damage by oxidative stress. Our results conflict with what most studies have shown up to now (but see for example Marcogliese et al 2005 [6]). Infection by parasites and pathogens of a wide range of taxa are generally associated with the inhibition or weakening of the host antioxidant system, and the concomitant increase in TBARS [47,48,49].
This enhancement of the anti-oxidative defense mechanisms is probably connected with the trophic transmission mode of cestodes that infect Artemia, which means they require “healthy” hosts in order to increase transmission success (through predation by the definitive host). Decrease in antioxidant status enhances short-term survival prospects [50], so potentiating it may be part of the transmission strategy of the parasite to increase longevity and probability of transmission (as in the “parasite manipulation hypothesis” [51]).
We also found that there were more lipid droplets in infected individuals, which is consistent with previous work indicating that F. liguloides increases total lipid levels [52] and that C. podicipina increases triglyceride levels in Artemia [26] (Fig 2B). Lipids have a high heavy metal binding capacity, and lipid content has a significant effect on the accumulation of As in other organisms [53]. Neutral lipids such as those in lipid droplets can protect organisms against pollutants, sequestering them away from sensitive target sites [54,55]—a principle known as survival of the fattest [41]. Although many studies support this principle, none have addressed parasite-mediated effects. The only previous study to suggest that parasites can increase host survival under polluted conditions through a lipid-related effect was on the freshwater clam Pisidium amnicum [56]. Clams infected by digenean trematode larvae are less sensitive to pentachlorophenol, perhaps because the high lipid content of the parasite changes the internal distribution of the toxicant. Pentachlorophenol is moderately lipophilic so is expected to accumulate in adipose tissue, whereas the target sites for toxicity are the mitochondria [56]. In parasitized Artemia, additional lipids accumulate in the host, not in the parasite as in the case of Pisidium.
Lipid droplets in infected Artemia are associated with carotenoids (see Fig 2B) and this, together with hemoglobins induced by parasites, largely explains the colour change that allowed us to separate infected individuals with the naked eye (Fig 1). Both F. liguloides and C. podicipina increase the concentration of carotenoids in infected Artemia [52]. In contrast, carotenoid concentrations in other animals are often negatively correlated with parasite load [57] and with pollutants [58]. Carotenoids are potent lipid-soluble antioxidants [59] and are able to inhibit lipid peroxidation in liposomes [60]. The accumulation of carotenoids in infected Artemia is also considered a parasite strategy to increase the probability of being predated by birds (the final host) by increasing visibility [61] and enhancing nutritive value [26]. Carotenoids provide protection against oxidative stress in many free living organisms [62,63,64] so they may also increase longevity in infected Artemia. Given that oxidative stress is a common marker of toxicity, not only for As in plants, invertebrates and vertebrates [65,66,67] but also for heavy metals (e.g. lead, cadmium and mercury [68]), cestode parasites may protect Artemia against a broader range of pollutants.
Unlike carotenoids, a positive effect of parasites on host lipid content is common in nature, e.g. in acanthocephalans infecting gammarids [69] or trematodes infecting bivalves [70]. Therefore, our finding regarding increased resistance to As in the presence of parasites may not be an isolated case, and more studies are needed to evaluate how frequently this occurs in nature.
The differences in sensitivity to As observed in infected Artemia collected on different dates have several possible explanations, including a negative effect of the generally higher infection levels, or a higher pathogenicity of C. podicipina which was absent in May. Alternatively, it could be related to seasonal changes in the ages of the parasites or their hosts, or changes in the lipid or carotenoid levels in the host. Increased heavy metal accumulation with age in cysticercoids has been shown in other cestodes [71,72]. Previous studies of the interactions between parasites and pollutants on toxicity have focused on individual parasite species (but see Gismondi et al. [7]). In nature, co-infections of different parasites are extremely common, and our study illustrates the need to consider the effects of co-infections in environmental parasitology.
The toxicity of As is highly dependent on its bioavailability, which in turn is dependent on its chemical form and the capacity to be released from environmental matrices [73]. In marine environments, As occurs predominantly as the inorganic forms arsenate (As(V)), and arsenite (As(III)), and is significantly more bioavailable from seawater and porewater than from sediments [74]. Moreover, temperature can strongly affect the chemical behaviour of pollutants and their bioavailability [75,76,77], but also the physiology of aquatic organisms. There are many studies of the effect of temperature on heavy metal toxicity, but we are not aware of any that integrate the influence of parasites. A temperature rise of 4°C caused a significant decrease in the LC50 for both infected and uninfected Artemia. The decrease in dissolved oxygen in hypersaline waters with increasing temperature coincides with higher respiratory demands in Artemia. Water permeability and drinking in Artemia increase markedly with temperature [78], hence uptake of pollutants will also increase. Indeed, copper uptake in Artemia increases with temperature owing to increased activity and diffusion rate [79]. Differences in As toxicity in fish have also been associated with higher uptake at higher temperatures [80].
Parasites and pathogens are conventionally considered as detrimental for a host, but they can also have positive impacts with consequences for non-host species and even the whole ecosystem [81]. We provide evidence that parasites can also benefit their hosts by increasing resistance to pollutants in contaminated environments. Infection by parasites was associated with an improved antioxidant defense system (CAT and GR) without oxidative damage, as confirmed by unaffected values of TBARS. Parasites also increased the number of lipid droplets in infected individuals, which is a common phenomenon in intermediate hosts manipulated by parasites, so more studies in other host-parasite systems are needed to evaluate the wider relevance of our findings.
Our results provide an important advance in our understanding of host-parasite interactions and underline the importance of considering interactions between parasites, pollutants and temperature in combination, particularly given expected climate change and the likelihood that toxicity will increase with temperature.
Naturally infected and uninfected adults of A. parthenogenetica were collected with a plankton net (0.5 mm) within the Odiel saltpan complex (see Sánchez et al. [21] for details of the study area).
Sampling was carried out on two dates (14th of April and 15th of May 2014) at which the relative abundance of different cestode parasites changed considerably. On the 14th of April, Artemia were collected from three ponds with salinities of 140, 150 and 190 g/l (measured with a refractometer). On the 15th of May, samples were collected from another pond with a salinity of 200 g/l. These four ponds were all within the same stage of the solar evaporation process, were hydrologically interconnected, and had similar sediment type, depth, and invertebrate and bird communities [21]. These ponds were selected on the basis of the abundance of live Artemia. Artemia were transported to the laboratory in 25 litre containers. Infected and uninfected Artemia individuals of a similar size were then selected on the basis of their colour (Fig 1); infected individuals are visually recognisable by the bright red colouration induced by the cestodes, whereas uninfected individuals are practically transparent [39]. Similar size (as a proxy of age) was selected since this may influence As concentrations or sensitivity [82, 83]. There is no difference in growth rates between infected and uninfected shrimp [52]. Of the individuals visually allocated to the infected group prior to experiments, 98% were truly infected when inspected afterwards (n = 989). Among individuals allocated to the uninfected group, 98% were truly uninfected (n = 988).
Toxicity experiments were conducted after 24 h of acclimation of the Artemia at 100 g/l salinity (artificial salt mixture of Instant Ocean dissolved in distilled water). Two experiments were carried out. The first (with Artemia collected on 14th of April) was conducted at 25°C (the mean annual temperature in the Odiel salt pans [21]). The second (Artemia collected on 15th of May) was conducted at both 25 and 29°C. During summer months, Artemia are often exposed to temperatures ≥ 29°C [21, 22] and the frequency of these events is expected to increase in future.
Median lethal concentration (LC50) was used to quantify As toxicity in infected and uninfected adult Artemia. Arsenic, as reagent-grade sodium arsenate, NaAsO (CAS No. 10048-95-0) was used to prepare a concentrated stock solution. The study design concentrations were prepared by mixing different proportions of the stock solution and saltwater (Instant Ocean prepared with MilliQ). The saltwater used for the dilutions was prepared within 24 h of the start of the experiment and As added one hour before conducting the experiment (to prevent oxygen depletion). Ten concentrations of As between 5 and 140 mg/l were used for the experiments at 25°C (0, 5, 20, 35, 50, 65, 85, 95, 110, 125, 140 mg/l), and ten between 4 and 67 mg/l were used for the experiment at 29°C (0, 4, 11, 18, 25, 32, 39, 46, 53, 60, 67 mg/l) in order to estimate the LC50. Experimental concentrations were adjusted to the observed mortality (higher mortality at 29°C implied lower tested concentrations). These concentrations were selected after preliminary tests. Three replicates were used per concentration, with each replicate made up of a group of 10 individuals. Infected (red) and uninfected (transparent) individuals were transferred to 25 ml beakers (10 individuals per beaker) and placed in climatic chambers at the chosen temperature, with a 12:12 photoperiod and without food. After a 24 h exposure period, dead individuals (considered to be those with no movement of the appendages observed within 10 seconds) were counted and all (alive or dead) individuals were mounted on slides to confirm parasitic status under the microscope. After observations of the cysticercoids (cestode larval stage in the intermediate host) in situ, each infected specimen was gently pressed under the coverslip. If the identification of the cysticercoids recorded was not possible at this stage, whole infected brine shrimps or isolated cysticercoids were prepared as permanent mounts in Berlese’s medium in order to facilitate observations on the morphology of rostellar hooks. Cysticercoids were identified following Georgiev et al. [30] and Vasileva et al. [84]. Prevalence (P) and mean abundance (MA) were calculated separately for the “infected group” on both dates.
Artemia were sampled in May 2015 from Odiel salt ponds of intermediate-high salinity. Brine shrimps were transported to the laboratory and placed in artificial sea water (Instant Ocean, 100 g/l) before the experiment. A subsample was used to characterize the exact parasite composition (n = 60 infected individuals) and quantify the number of lipid droplets (n = 20 infected and 20 uninfected Artemia). The numbers of cysticercoids, prevalence, mean abundance and mean intensity (see Bush et al. 1997 [85] for definitions) were calculated for each cestode species. The number of lipid droplets was estimated according to Wurtsbaugh & Gliwicz 2001 [86]. We quantified lipid levels by inspecting individuals at 30x magnification and counting the number of lipid droplets along the right side of the 5th and 6th segments of the body.
The rest of the specimens were acclimated during 24h to the experimental salinity with continuous aeration and fed ad libitum with lyophilized Tetraselmis chuii algae. The toxic concentrations of 4.69 mg/l As was selected on the basis of preliminary LC50 tests (the lowest concentration at which mortality was detected). Infected and uninfected A. parthenogenetica of the same size range were allocated to 1L glass vials (100 individuals per vial) with 600 ml of experimental solution (either control (no As) or 4.69 mg/L As) during 24h at 25°C (12:12 photoperiod) without food.
After 24h exposure, individuals were gently washed in distilled water and, after removing excess water, stored at ‒80°C until biochemical analysis. All operations were performed at 4°C to prevent enzyme or tissue degradation. We performed the biochemical analysis on pools of 20 individuals per treatment. Number of replicates varied from 1 to 12 depending on Artemia availability. The different biomarkers were determined in the whole soft tissues after homogenization and centrifugation. Tissues were homogenized with an electrical homogenator (Miccra D-1 Art Moderne Labor Technik) in cool homogenization buffer (Tris-HCl 100 mM, EDTA 0.1 mM, Triton X-100 0.1%, pH7.8) using 200 μl of buffer per sample (20 individuals). The sample was centrifuged at 14,000 rpm at 4°C for 30 minutes and supernatant stored at −80°C until enzymatic determination.
We quantified five parameters as a proxy for oxidative status of Artemia: activity of four enzymes (catalase CAT, superoxide dismutase SOD, glutathione peroxidase GPx and glutathione reductase GR) and lipid peroxidation levels (thiobarbituric acid reactive substances TBARS). Total protein content in the supernatant fluid was determined following a standard Bradford’s procedure [87]. Enzyme activity was determined colorimetrically. Concentration of each indicator was estimated following the specific procedures below.
The median acute lethal concentration (LC50) and 95% confidence limits were estimated and compared between infected and uninfected individuals, different temperatures and dates using Trimmed Spearman-Karber (TSK) analysis for lethal tests [94]. The criterion of non-overlapping 95% confidence limits was used to determine a significant difference (p < 0.05) between LC50 values [40].
To test the effect of date, temperature, parasitic status and As concentration on mortality we performed GLZ with a Poisson error distribution, log link function and correction for overdispersion. Date, temperature and parasitic status were included as categorical factors, and As concentration as a continuous variable. Stepwise backwards removal was used to obtain a final model containing only significant factors.
Differences in prevalence of different cestode species between the two sampling dates were evaluated with Z-tests. Comparisons of cestode abundance were performed with Mann-Whitney U tests. We also used Mann-Whitney tests to compare enzymatic activity and lipid peroxidation, as well as lipid droplets between infected and uninfected individuals. Statistica 12 software for Windows was used for all statistical analyses [95].
|
10.1371/journal.pntd.0007030 | The effects of trans-chalcone and chalcone 4 hydrate on the growth of Babesia and Theileria | Chemotherapy is a principle tool for the control and prevention of piroplasmosis. The search for a new chemotherapy against Babesia and Theileria parasites has become increasingly urgent due to the toxic side effects of and developed resistance to the current drugs. Chalcones have attracted much attention due to their diverse biological activities. With the aim to discover new drugs and drug targets, in vitro and in vivo antibabesial activity of trans-chalcone (TC) and chalcone 4 hydrate (CH) alone and combined with diminazene aceturate (DA), clofazimine (CF) and atovaquone (AQ) were investigated.
The fluorescence-based assay was used for evaluating the inhibitory effect of TC and CH on four Babesia species, including B. bovis, B. bigemina, B. divergens, B. caballi, and T. equi, the combination with DA, CF, and AQ on in vitro cultures, and on the multiplication of a B. microti–infected mouse model. The cytotoxicity of compounds was tested on Madin–Darby bovine kidney (MDBK), mouse embryonic fibroblast (NIH/3T3), and human foreskin fibroblast (HFF) cell lines. The half maximal inhibitory concentration (IC50) values of TC and CH against B. bovis, B. bigemina, B. divergens, B. caballi, and T. equi were 69.6 ± 2.3, 33.3 ± 1.2, 64.8 ± 2.5, 18.9 ± 1.7, and 14.3 ± 1.6 μM and 138.4 ± 4.4, 60.9 ± 1.1, 82.3 ± 2.3, 27.9 ± 1.2, and 19.2 ± 1.5 μM, respectively. In toxicity assays, TC and CH affected the viability of MDBK, NIH/3T3, and HFF cell lines the with half maximum effective concentration (EC50) values of 293.9 ± 2.9, 434.4 ± 2.7, and 498 ± 3.1 μM and 252.7 ± 1.7, 406.3 ± 9.7, and 466 ± 5.7 μM, respectively. In the mouse experiment, TC reduced the peak parasitemia of B. microti by 71.8% when administered intraperitoneally at 25 mg/kg. Combination therapies of TC–DA and TC–CF were more potent against B. microti infection in mice than their monotherapies.
In conclusion, both TC and CH inhibited the growth of Babesia and Theileria in vitro, and TC inhibited the growth of B. microti in vivo. Therefore, TC and CH could be candidates for the treatment of piroplasmosis after further studies.
| Protozoa of the genus Babesia are the second most common blood-borne parasites of mammals after the trypanosomes. Babesia and Theileria are the etiological agents of piroplasmosis, a tick-transmitted disease causing substantial losses of livestock and companion animals worldwide and has recently gained attention as one of the emerging zoonosis in humans. Diminazene aceturate (DA) and imidocarb dipropionate are still the first choices for the treatment of animals. However, these drugs cause many adverse effects. Furthermore, they are not approved for human medicine. Therefore, the development of alternative treatment remedies against babesiosis is urgently required. In the present study we evaluated the effects chalcone 4 hydrate (CH) and trans-chalcone (TC), against the growth of four species of Babesia and T. equi. Furthermore, we studied the chemotherapeutic potential of TC on B. microti in mice. The effects of the combined treatment of TC with DA, CF and AQ revealed that TC was found to diminish the adverse effects of these drugs.
| Babesiosis is one of the most severe infections of animals worldwide and has recently gained attention as one of the emerging zoonosis in humans [1, 2]. Babesia bovis, Babesia bigemina, and Babesia divergens infect cattle and cause bovine babesiosis. Of these, B. bovis is much more virulent than B. bigemina and B. divergens due to its ability to sequestrate in the capillaries, causing hypotensive shock syndrome and neurological damage [3]. In horses, Babesia caballi and Theileria equi (formerly Babesia equi) infect horses, causing equine piroplasmosis. T. equi parasitizes leucocytes and erythrocytes for the completion of its life cycle, causing anemia, weight loss, lethargy, and fever [4], whereas B. caballi directly infects horse erythrocytes in a manner similar to B. bovis and B. bigemina in cattle. Human babesiosis is caused by Babesia microti in North America, while in Europe, it is caused by Babesia divergens. Human babesiosis manifests as an apparently silent infection to a fulminant, malaria-like disease, resulting occasionally in the death of the infected individual [5].
Prevention of babesiosis relies on vector control, vaccination, and chemotherapy. Thus far, chemotherapy has been the most successful method due to the availability of efficacious compounds such as diminazene aceturate and imidocarb dipropionate for animals and atovaquone, azithromycin clindamycin, and quinine for humans [5]. Unfortunately, atovaquone-resistant Babesia gibsoni has been reported [6, 7], and Mosqueda et al. reported the emergence of parasites resistant to diminazene aceturate (DA) [8]. Therefore, research to discover new drugs and drug targets is the fundamental approach toward addressing current limitations.
Chalcones (trans-1, 3-diaryl-2-propen-1-ones) are natural products belonging to the flavonoid family that are widespread in plants and are considered as intermediate in the flavonoid biosynthesis [9, 10]. They are recognized for their broad-spectrum biological activities, including antimalarial [11], anticancer, antileishmanial, antitrypanosomal [12, 13, 14, 15, 16], anti-inflammatory, antibacterial, antioxidant, antifilarial, antifungal, antimicrobial, larvicidal, and anticonvulsant ones [17, 18]. Based on the wide range of pharmacological effects, it is implied that chalcones have several mode of actions in different parasites. For instance, Go et al. showed that chalcones modulate the permeability pathways of the Plasmodium-infected erythrocyte membrane, affecting its growth and multiplication [9]. Frölich et al. showed that chalcones inhibit glutathione (GSH)-dependent haemin degradation and binding to the active site of the cysteine protease (falcipain) enzyme involved in hemoglobin degradation in the Plasmodium parasite [11, 19, 20]. Chalcones inhibit the components of mitochondrial respiratory chain bc1 complex (ubiquinol–cytochrome c reductase) (UQCR) [21]. Additionally, chalcones inhibit the cyclin-dependent protein kinases (CDKs) (Pfmrk and PfPK) and plasmepsin II [22]. These various pathways demonstrate the importance of chalcones as chemotherapeutic candidates against malaria. However, the effect of chalcones had never been evaluated against Babesia and Theileria parasites. Therefore, this study evaluated the effects of chalcones, namely, chalcone 4 hydrate (CH) and trans-chalcone (TC), against the growth of B. bovis, B. bigemina, B. divergens, B. caballi, and T. equi in vitro. Furthermore, we studied the chemotherapeutic potential of TC on B. microti in mice.
The growth inhibitory assay was conducted on five species: B. bovis, B. bigemina, B. divergens, B. caballi, and T. equi. Trans-chalcone (TC) and chalcone 4 hydrate (CH) inhibited the multiplication and growth of all species tested in a dose-dependent manner (Figs 1 and 2).
The IC50 values of TC and CH on B. bovis, B. bigemina, B. divergens, B. caballi, and T. equi were 69.6, 33.3, 64.8, 18.9, and 14.3 μM and 138.4, 60.9, 82.3, 27.9, and 19.2 μM, respectively (Table 1).
In this study, DA showed IC50 values at 0.35, 0.68, 0.43, 0.022, and 0.71 μM against B. bovis, B. bigemina, B. divergens, B. caballi, and T. equi, respectively. AQ showed IC50 values at 0.039, 0.701, 0.038, 0.102, and 0.095 μM against B. bovis, B. bigemina, B. divergens, B. caballi, and T. equi, respectively. CF showed IC50 values at 8.24, 5.73, 13.85, 7.95, and 2.88 μM against B. bovis, B. bigemina, B. divergens, B. caballi, and T. equi, respectively (S1 Table). The effectiveness of chalcones was not influenced by the diluent, since there was no significant difference in the inhibition between wells containing the DMSO and untreated wells. The precultivation of RBCs with TC and CH was conducted to determine their direct effect on host RBCs. Bovine and equine RBCs were incubated with TC or CH at 10, 100, and 200 μM for 3 h prior to the subculture of B. bovis and T. equi. The multiplication of B. bovis (S1A Fig) and T. equi (S1B Fig) did not significantly differ between TC- or CH-treated RBCs and normal RBCs for either species.
A viability assay was performed to determine whether the concentrations of TC and CH could completely clear parasites after 4 days of successive treatment, followed by withdrawal of the drug pressure. B. bovis, B. bigemina, and B. caballi treated with TC could not regrow at a concentration of 2×IC50, while B. divergens could not regrow at 4×IC50. B. bovis, B. bigemina, B. divergens, and B. caballi treated with CH could not regrow at a concentration of 4×IC50. T. equi treated with TC and CH could regrow at a concentration of 4×IC50 (Table 2).
Micrographs of TC- and CH-treated B. bovis (Fig 3), B. bigemina, B. divergens, B. caballi (Fig 4), and T. equi consistently showed degeneration of the parasites by loss of the typical shapes at 24 h, whereas further observations at 72 h showed deeply stained dot-shaped remnants of the parasites lodged within the erythrocytes.
A drug-combination analysis was performed to determine whether the combined treatments are synergistic (have a greater effect), additive (have a similar effect), or antagonistic (have a reduced effect or block the effect). Five dilutions of CH or TC (S2 Table), as recommended in the Chou–Talalay method [25], were combined at a constant ratio with DA, AQ, or CF. The inhibition percentage for the single drug and each combination was analyzed using CompuSyn software to generate the combination index (CI) values (S3 Table). The combination treatments of TC–DA and CH–DA showed an additive effect against B. bovis and B. bigemina and a synergistic effect against B. divergens, B. caballi, and T. equi. The combination treatments of TC–AQ showed a synergistic effect against B. bigemina, B. caballi, and T. equi and an additive effect against B. bovis and B. divergens. The combination treatments of CH–AQ showed a synergistic effect against B. bovis, B. divergens, B. caballi, and T. equi and an additive effect against B. bigemina. The combination treatments of TC–CF showed a synergistic effect against B. bigemina, B. divergens, B. caballi, and T. equi and an additive effect against B. bovis. The combination treatments of CH–CF showed a synergistic effect against B. bigemina, B. caballi, and T. equi and an additive effect against B. bovis and B. divergens. None of the combinations showed an antagonistic effect (Table 3).
Trans-chalcone and chalcone 4 hydrate showed an inhibitory effect on the in vitro culture of Babesia and Theileria parasites. Therefore, the effect of TC and CH on the host cells was evaluated using MDBK, NIH/3T3, and HFF cell lines to see the cytotoxicity of the two compounds (Table 1). The EC50 values of TC on MDBK, NIH/3T3, and HFF cell lines were 293.9 ± 2.9, 434.4 ± 2.7, and 498 ± 3.1 μM, respectively. The EC50 values of CH on MDBK, NIH/3T3, and HFF cell lines were 252.7 ± 1.7, 406.3 ± 9.7, and 466 ± 5.7 μM, respectively (Table 1). In a separate assay, DA and AQ at 100 μM did not show any inhibition of MDBK, NIH/3T3, or HFF cell viability, while CF showed inhibition only of MDBK with an EC50 value of 34 ± 3.4 μM (S1 Table). The selectivity index, defined as the ratio of EC50 of the drugs tested on the cell line to IC50 of the tested drugs on in vitro culture of parasites. For TC, the highest selectivity index was achieved on T. equi, for the MDBK cell line the selectivity index was found to be 20.6 times higher than its IC50 on T. equi, while in the case of the NIH/3T3 cell line was found to be 30.4 times higher than the IC50 and in the case of the HFF cell line showed selectivity index 34.8 times higher than its IC50 on T. equi. For CH, the highest selectivity index was achieved on T. equi as in the case of the MDBK, NIH/3T3 and HFF cell lines was found to be 13.2 times, 21.2 times and 24.3 times higher than IC50, respectively (Table 1).
For further evaluation of TC efficacy in comparison with other drugs, the chemotherapeutic effect of TC was examined in mice infected with B. microti (Fig 5).
In the DDW control group, the multiplication of B. microti increased significantly and reached the highest parasitemia at 57.7% on day 8 post infection (p.i). In all treated groups, the level of parasitemia was cleared at a significantly lower percent of parasitemia than the control group (p < 0.05) from days 6–12 p.i. In the monochemotherapy-treated mice, the peak parasitemia level reached 16.3% on day 9, 4.4% on day 8, 5.5% on day 8, and 4% on day 6 with 25 mg/kg TC, 25 mg/kg DA, 20 mg/kg AQ, and 20 mg/kg CF, respectively (Fig 5). The parasitemia was undetectable via microscopy starting on day 13, 15, and 13 p.i. in mice treated with 25 mg/kg DA, 20 mg/kg AQ, and 20 mg/kg CF, respectively. In the combination-chemotherapy-treated groups, the peak parasitemia level reached 2.6%, 3.2%, and 10.4% with 12.5 mg/kg TC–12.5 mg/kg DA, 12.5 mg/kg TC–10 mg/kg CF, and 12.5 mg/kg TC–10 mg/kg AQ, respectively, on day 9 (Fig 6).
The parasitemia was undetectable in microscopic examination on day 21 p.i. in mice treated with 25 mg/kg TC. The parasitemia was undetectable in mice by microscopic examination on days 13, 17, and 21 p.i. with 12.5 mg/kg TC–12.5 mg/kg DA, 12.5 mg/kg TC–10 mg/kg CF, and 12.5 mg/kg TC–10 mg/kg AQ, respectively. The parasite DNA was not detected on day 45 with 25 mg/kg DA IP, 12.5 mg/kg TC–10 mg/kg CF, or 12.5 mg/kg TC–12.5 mg/kg DA. In all other groups (20 mg/kg AQ oral, 20 mg/kg CF oral, 25 mg/kg TC IP, and 12.5 mg/kg TC–10 mg/kg AQ), the parasite DNA was detected until day 45 (Fig 7).
Furthermore, infection with B. microti reduces the RBC count (Fig 8A), hemoglobin concentration (Fig 8B), and hematocrit percentage (Fig 8C) in mouse blood, as observed in the DDW control group on days 8 and 12 p.i. Significant differences (p < 0. 05) in RBC count were observed between the DDW control group and all drug-treated groups on days 8 and 12.
The treatment of bovine and equine piroplasmosis is limited to diminazene aceturate (DA) and imidocarb propionate, while clindamycin–quinine and atovaquone–azithromycin combinations have been utilized to manage human babesiosis [8, 26]. Unfortunately, toxic effects and resistance of the piroplasms against the current drug molecules have been documented [2]. To overcome this challenge, research is urgently needed to discover new drug candidates and drug targets against piroplasms [8]. Therefore, the current study assessed the chemotherapeutic potential of chalcone 4 hydrate (CH) and trans-chalcone (TC) against Babesia and Theileria parasites in vitro and B. microti in vivo. Furthermore, the effects of combining CH and TC with the currently available drugs, namely, AQ, CF, and DA, against Babesia and Theileria parasites were assessed in vitro.
In the current study, both CH and TC were effective against the Babesia and Theileria parasites in vitro (Figs 1 and 2). It is noteworthy that CH and TC were most effective against T. equi, followed by B. caballi, B. bigemina, and B. divergens, whereas they were least effective against B. bovis. Interestingly, previous reports revealed that many drugs showed varying effectiveness against different species of Babesia and Theileria parasites and this may be due to different nutrition requirements for parasite growth, the basic biochemistry, physiology, and molecular biology of parasites and of their interactions with their hosts, thus, the inhibitory effectiveness of the drugs on each species becomes different [27]. For instance, AbouLaila et al. reported that epoxomicin exhibited IC50 values of 21.4 ± 0.2, 4 ± 0.1, 39.5 ± 0.1, 9.7 ± 0.3, and 21.1 ± 0.1 nM for B. bovis, B. bigemina, B. ovata, B. caballi, and T equi [28]. Therefore, it is not surprising that the IC50s of CH or TC against some parasites were higher or lower comparing to other species. The IC50 values shown by CH and TC against Babesia and Theileria parasites were comparable to 11.5 and 21.7 μM shown by CH and TC against P. falciparum, respectively, Trypanosoma, and Leishmania [9, 10, 13, 14, 22, 29]. This emphasizes that CH and TC are effective against many protozoan parasites. However, the mode of action has yet to be understood comprehensively in comparison with the existing data. In a previous study, Mi-Ichi et al. documented that chalcones are mitochondrial electron transport inhibitors that block ubiquinone (UQ) from binding to cytochrome b (bc1) in Plasmodium parasites and exhibit potent antimalarial activity[21]. Torres-Santos et al. and Chen et al. reported that chalcones inhibit the growth of Leishmania and Trypanosomes by inhibiting the activity of fumarate reductase (FRD), one of the enzymes of the parasite respiratory chain that it is very important in the energy metabolism of the parasites [12, 15]. Since this enzyme is absent from mammalian cells, it could be an important target for drugs against protozoan parasites. Based on the previous findings, it is possible that chalcones also inhibit the mitochondrial respiratory chain enzymes in Babesia and Theileria parasites, which could be elucidated in future studies.
The viability assay showed that TC and CH were more effective against Babesia parasites than against Theileria parasites. B. bovis, B. bigemina, and B. caballi could not relapse at 2×IC50 treatments of TC, while B. divergens could not relapse at 4×IC50 treatments of TC. B. bovis, B. bigemina, B. caballi, and B. divergens could not relapse at 4×IC50 treatments of CH. In contrast, T. equi recovered even at 4×IC50 treatments of CH and TC. This finding was similar to deductions by Reece et al., who revealed that different species of malaria parasites utilized different coping mechanisms by varying in their investment in gametocytes during infections depending on aspects of their environment [30]. However, these patterns appear confusing, they can explain how parasites can respond to changes experienced during infection. Therefore, T. equi might act by different mechanism of action than that for Babesia species for coping and endure the stress caused by TC and CH treatments. That may be the reason that T. equi can be revived again after the withdrawal of the drug pressure on the fourth day. However, further studies are required to detect the mechanism preventing T. equi from being completely killed by CH and TC. In an attempt to visualize the morphological changes of CH- and TC-treated Babesia and Theileria parasites, micrographs were taken at various incubation times. The observations showed deformed and dividing parasites and irregular parasite shapes at 24 h and pyknotic remnants within the iRBCs at 72 h (Figs 3 and 4). This showed that chalcones have a time-dependent effect on the Babesia and Theileria parasites. Although the exact mode of action is yet to be elucidated, the parasites progressively lost their shape and became smaller. This could be attributed to the ability of chalcones to interfere with the metabolic pathway, as documented in P. falciparum and Leishmania parasites [21, 31].
Combination chemotherapy has been recommended against drug-resistant protozoan and bacterial pathogens. Additionally, combination chemotherapy reduces drug dosages, thereby reducing their toxic side effects. Hence, the current study explored the combination of TC and CH with drugs such as DA, AQ, and CF in vitro. The findings of this study show that the effects of TC and CH combined with DA, AQ, or CF were either synergistic or additive against Babesia and Theileria parasites. The ability of TC and CH to combine with the current effective drugs is a property that can be explored in the development of chemotherapies against Babesia and Theileria [32]. That study showed that the CF–DA combination has additive effects on the in vitro growth of B. bovis, B. bigemina, and B. caballi and synergistic effects on that of T. equi, and the combination chemotherapy with low-dose regimens of CF and DA has a more potent inhibitory effect on B. microti in mice than did their monochemotherapies.
The experiments to understand toxicity showed that CH and TC affected the viability of MDBK, NIH/3T3, and HFF cell lines with a dose-dependent inhibitory effect and a modest selectivity index. This finding is consistent with the results reported by Echeverria et al. [33]. They examined the cytotoxic activities of synthetic 2’-hydroxychalcones against hepatocellular carcinoma cells, demonstrating that synthetic 2’-hydroxychalcones show apoptosis induction and dose-dependent inhibition of cell proliferation without cytotoxic activities on normal cell lines [33]. Mi-Ichi et al. explained that the low cytotoxic activity of TC and CH against mammalian cell lines is attributed to the fact that the ubiquinol–cytochrome c reductase (UQCR) and succinate ubiquinone reductase (SQR) of P. falciparum mitochondria are different from those of the mammalian host cells [21]. With reference to the above, chalcones might be safe for use in animals and humans following further in vivo clinical studies.
The promising efficacy of TC in vitro prompted us to evaluate TC performance in vivo. TC administered intraperitoneally at a dose of 25 mg/kg resulted in a 71.8% inhibition of the parasitemia on day 9 p.i. However, the inhibition rate was lower than those in the presence of 25 mg/kg DA, 20 mg/kg AQ, and 20 mg/kg CF, which were 92.5%, 90.8%, and 93.1%, respectively (Fig 5). Certainly, the additive and synergistic effects in these combinations were indicated by the high degree of association observed in vitro, which prompted studies in vivo. Therefore, the TC–DA and TC–CF combinations were evaluated in mice to determine whether combination treatment would enable the reduction of DA, AQ, and CF dosages without altering the therapeutic efficacy against B. microti infection. Interestingly, the combination treatment of TC and DA at a dose of 12.5 + 12.5 mg/kg improved the efficacy to 95.6%, while the combination of TC and AQ at a dose of 12.5 + 10 mg/kg resulted in 81.9% efficacy. The combination of TC and CF at a dose of 12.5 + 10 mg/kg resulted in a 94.4% inhibition in the parasitemia level at day 8 p.i. (Fig 6). The potentiation of TC that was achieved in in vivo combination therapy confirms the result that was observed in the in vitro combination experiment, emphasizing that chalcones are good combinatorial drugs. With regard to the chemotherapeutic effects of TC against Leishmania in mice, Piñero et al. showed that a single dose of 4 mg/kg TC by subcutaneous administration could completely inhibit the pathogenicity of the Leishmania parasite in vivo [29]. In addition to being efficacious, chalcones enhanced the production of nitric oxide, which kills the intra-erythrocytic parasites and stimulates the host immune system [34].
In order to confirm the ability of TC to eliminate B. microti, a PCR assay was performed on samples collected on day 45 p.i. to be analyzed for the presence of DNA. Interestingly, this study confirmed the absence of B. microti DNA in groups treated with a combination chemotherapy of TC+DA or TC+CF (Fig 7) as compared to monotreatment. These results underscore the importance of combination chemotherapy in the effective control of piroplasmosis. This finding further emphasizes the need for combination therapy to achieve the most optimum efficacy and prevent the relapse of infection or development of a carrier state [32].
Furthermore, TC did not show toxic side effects to mice (Fig 8A–8C), consistent with a previous study [14]. Taken together, these findings advocated the anti-piroplasmic potential of two compounds of chalcones against the growth of several Babesia species and T. equi in vitro and B. microti in vivo that can open the way to search for other synthetic chalcone analogs with great potency and lesser toxicity. However, it should be noted that there are some limitations to the present study. Therefore, further studies are required to confirm the mechanisms of TC and CH against Babesia and Theileria so as to better understand the effect of interactions with other drugs such as DA, AQ, and CF.
CH and TC showed growth inhibitory effect against Babesia and Theileria in vitro. Furthermore, TC showed chemotherapeutic efficacies against B. microti in vivo. TC effectiveness in vivo was comparable to that shown by DA, and it showed no toxicity to mice. The TC–DA and TC–CF combinations showed higher efficiency against piroplasms than did TC, DA, or CF monotherapies. This implies that TC could be used as a chemotherapeutic drug against piroplasmosis. Moreover, the results suggest that the TC–DA and TC–CF combination chemotherapies will be better choices for the treatment of piroplasmosis than TC, DA, or CF monotherapies.
|
10.1371/journal.ppat.1004650 | Cytosolic Access of Mycobacterium tuberculosis: Critical Impact of Phagosomal Acidification Control and Demonstration of Occurrence In Vivo | Mycobacterium tuberculosis (Mtb) uses efficient strategies to evade the eradication by professional phagocytes, involving—as recently confirmed—escape from phagosomal confinement. While Mtb determinants, such as the ESX-1 type VII secretion system, that contribute to this phenomenon are known, the host cell factors governing this important biological process are yet unexplored. Using a newly developed flow-cytometric approach for Mtb, we show that macrophages expressing the phagosomal bivalent cation transporter Nramp-1, are much less susceptible to phagosomal rupture. Together with results from the use of the phagosome acidification inhibitor bafilomycin, we demonstrate that restriction of phagosomal acidification is a prerequisite for mycobacterial phagosomal rupture and cytosolic contact. Using different in vivo approaches including an enrichment and screen for tracking rare infected phagocytes carrying the CD45.1 hematopoietic allelic marker, we here provide first and unique evidence of M. tuberculosis-mediated phagosomal rupture in mouse spleen and lungs and in numerous phagocyte types. Our results, linking the ability of restriction of phagosome acidification to cytosolic access, provide an important conceptual advance for our knowledge on host processes targeted by Mtb evasion strategies.
| The intracellular fate of the agent of the human tuberculosis agent in phagocytes is a question of great biological relevance. Among the mycobacterial survival strategies, the escape of Mycobacterium tuberculosis from phagosomes has been subject of scientific debate for a long time. However, technically improved methods recently reinforced the occurrence of this phenomenon. Here, we focused on the host factors involved in phagosomal rupture and provide first and singular evidence of M. tuberculosis-mediated phagosomal rupture in vivo in mouse lungs and inside the granuloma. We show that partial blockage of phagosomal acidification, induced by mycobacteria, is a prerequisite for efficient vacuolar breakage by M. tuberculosis and link maturation arrest, cytosolic contact and the corresponding immune responses. From our results we conclude that vacuolar breakage induced by M. tuberculosis is not an ex vivo artifact of cell cultures, but an important process that occurs inside infected phagocytes within organs during several days that strongly determines the outcome of infection with this key pathogen.
| The pathogenic potential of Mycobacterium tuberculosis (Mtb), the etiologic agent of human tuberculosis (TB), depends largely on the type VII secretion system ESX-1 [1,2], which is responsible for the secretion of the 6-kDa Early Secreted Antigenic Target (ESAT-6), its protein partner, the 10-kDa Culture Filtrate Protein (CFP-10), and several ESX-1 associated proteins (Esps) [3,4]. ESX-1 secretion is evolutionary conserved in most members of the M. tuberculosis complex [5], and the more distantly related tubercle bacilli of the Mycobacterium canettii clade [6,7], as well as in some non-tuberculous mycobacteria such as Mycobacterium marinum [8]. This secretion system governs numerous aspects of interaction between pathogenic mycobacteria and the host cell [1,2], including membrane-damaging activity [9–11], thought to be implicated in phagosomal escape at later stages of infection [12–16]. Although this phenomenon is a matter of debate [2,17–20], by use of a single-cell Fluorescence Resonance Energy Transfer (FRET)-based technology [21], we recently demonstrated that ESX-1-proficient Mtb and recombinant Mycobacterium bovis BCG::ESX-1 were able to induce phagosome rupture in human THP-1 macrophage (MΦ)-like cells [15]. This assay uses the ability of the surface-exposed BlaC β-lactamase of Mtb [22,23] to cleave the FRET substrate CCF-4, which consists of a cephalosporin core linking 7-hydroxycoumarin to fluorescein that has also been used for exploring effector injection and intracellular localization of Gram-negative bacteria [21,24,25]. The ESX-1-induced rupture of the phagosomal membrane, which results in the exit of mycobacterial products from the endosomal pathway and in extra-phagosomal localization of bacilli [13–16] is of relevance for the outcome of the immune control and bacterial dissemination [26–29]. Phagosomes are reported to be specialized platforms for pathogen recognition [30] and there is also growing evidence of a link between the functionality of the ESX-1 secretion system and the presence of mycobacteria-associated molecular patterns in the host cytosol. Peptidoglycans [31,32] and extracellular mycobacterial DNA [33] were reported to be sensed by the cytosolic receptors of the innate system with multiple biological consequences. Indeed, the Mtb-mediated induction of Nucleotide binding Oligomerization Domain (NOD)-Like Receptor pathways, i.e., NOD2 / Receptor-interacting protein 2 kinase (Rip2) / TANK-Binding Kinase 1 (TBK1) / Interferon regulatory factor (Irf) 5, is responsible for a significant part of type I interferon (IFN) production [31,32]. On the other hand, the signaling through the Stimulator of IFN Genes (STING) / TBK1 / Irf3 pathway [33] leads to a type I IFN signature on which depends the expression of CCL5, CXCL10 and Nitric Oxide Synthase 2 [34,35]. The formation of Nucleotide-binding domain and Leucin-rich Repeat pyrin–containing Protein-3 (NLRP-3) / ASC (Apoptosis-associated Speck-like protein containing a carboxy-terminal CARD) / caspase-1 inflammasome complex, is required in humans for the processing of the pro-IL-1β into biologically active pleïotropic immune mediatorIL-1β following Mtb infection [36,37]. Moreover, the ubiquitination of Mtb prior to its delivery to the autophagic machinery also necessitates the ESX-1-dependent translocation of extracellular Mtb DNA to the cytosol [16,33,38,39]. Thus, the events arising from mycobacterial cytoplasmic access may substantially influence both the immune control of Mtb and the inflammation-induced tissue damage.
The impact of selected components of the ESX-1 system on phagosomal rupture has recently been assessed [13,15,16], however, other potential intervening factors, including those from the host cell remain largely unexplored. Here, we have investigated the host parameters modulating the Mtb-mediated vacuolar breakage, by developing a CCF-4 FRET-based approach that can be used for the study of Mtb-infected cells by flow cytometry. This approach, which permits to combine the detection of phagosomal rupture with the analysis of numerous host cell phenotypic and functional parameters, allowed us to explore multiple phagocyte types, including those isolated from mouse airways. Our results provide first and unique evidence that Mtb-induced phagosomal rupture does occur in vivo inside the lungs and spleens of infected experimental animals and lasts over several days. Moreover, we here explore the impact of vacuolar acidification that constitutes a fundamental cellular defense mechanism [40] and demonstrate that the characteristic partial prevention of phagosomal acidification by Mtb is a prerequisite for phagosomal escape of the pathogen. Our study thus reveals novel details and presents a refined model of cellular events during infection with Mtb.
To evaluate mycobacteria-mediated phagosomal rupture in different phagocyte types and different physiological contexts, we adapted the previously used microscopy-based CCF-4 FRET technique [15] for flow cytometry. The latter approach not only allows monitoring of bacteria-induced phagosomal rupture or tracking of endosome-to-cytosol antigen translocation [25,41], but also permits the simultaneous inspection of surface markers and analysis of hundreds of thousands of host cells. At first, we infected differentiated THP-1 cells at a multiplicity of infection (MOI) of 1 either with Mtb H37Rv WT or the isogenic ΔESX-1 derivative, Mtb H37Rv-ΔRD1 [10], which both display similar β-lactamase activity [15]. These THP-1 cells were then incubated with CCF-4-AM, an esterified, lipophilic form of the CCF-4 substrate that can readily enter into cells, where it is converted by endogenous cytoplasmic esterases into negatively charged CCF-4, which is retained in the cytosol and emits green fluorescence (500–550 nm) upon stimulation at 320–380 nm, due to FRET from the coumarin moiety to the fluoroscein part (S1 Fig.). In the case of Mtb-induced phagosomal rupture, cleavage of CCF-4 by the intrinsic Mtb BlaC β-lactamase leads to loss of FRET and a change of the CCF-4 emission spectrum from green to blue coumarin fluorescence (410–470 nm). As depicted in Fig. 1A, the CCF-4 emission signals of CD11b+ gated THP-1 cells, infected with wild-type (WT) Mtb H37Rv, showed a marked shift of the CCF-4 emission towards blue at 4 days post infection (dpi). In contrast, a much weaker shift of the CCF-4 spectrum was observed for Mtb ΔESX-1-infected cells, validating our experimental setup and confirming the fundamental virulence differences between the used ESX-1-proficient and ESX-1–deficient Mtb strains [10,15]. The residual blue shift in Mtb ΔESX-1-infected cells relative to non-infected cells is likely a consequence of paraformaldehyde (PFA) fixation prior to signal acquisition (S2A–B Fig.). These results were further corroborated by ratios of Mean Fluorescence Intensities (MFI) of blue vs. green signals (Fig. 1B), and blue MFI447 nm (Fig. 1C). Moreover, we also used fluorescent Mtb (DsRed-Mtb H37Rv) to infect THP-1 cells, at a weaker initial dose (MOI = 0.3), and thereby observed that the CCF-4 blue emission shift selectively occurred in cells that had engulfed the bacteria (Fig. 1D). This approach thus allowed a quantitative study of phagosomal rupture in host cells that have engulfed Mtb, and whose subtype can be identified/determined by staining of the specific surface markers. Hence, our experimental setup was adapted to be used for various cell types and physiological situations, including the detection of vacuolar rupture in rare (infected) cells that were dispersed in a large and heterogeneous cell background population.
Dendritic cells (DC) and MΦ do not play the same roles during the infection. DCs that have engulfed Mtb, are more prone to process and present pathogen-derived antigens and to prime T cells than Mtb-laden MΦ which are thought to initiate the inflammatory program and are considered as long-term Mtb reservoirs. We thus comparatively evaluated the potential of Mtb to induce phagosome rupture in bone-marrow-derived (BM)-DC and -MΦ At first, by using fluorescent DsRed WT and ΔESX-1 Mtb variants, we showed similar uptake and infectivity of both strains at the beginning of the infection (Fig. 2A). Infection of BM-DC and BM-MΦ with WT Mtb then resulted in a strong blue shift at 3 dpi and thereafter, whereas for cells infected with the ΔESX-1 Mtb strain only a minor blue shift was detected (Fig. 2B). The relatively stable CCF-4 green signal and its progressively increasing blue shift for WT Mtb resulted in a blue/green ratio of 15 in BM-DC and 10 in BM-MΦ respectively, at 6 dpi (Fig. 2C). Similar as observed for THP-1 cells (Fig. 1D), infection of BM-DC with DsRed expressing Mtb showed that cells, which had engulfed DsRed Mtb, progressively increased their CCF4 blue shift over the observation period of 3 to 5 dpi (S3 Fig.). Together, these results suggest that ESX-1-dependent, Mtb-induced phagosomal rupture does occur in DC and MΦ.
To ascertain that the absence of FRET inhibition in cells infected with the ΔESX-1 Mtb mutant was not due to other molecular reasons than the absence of the ESX-1 secretion system, we complemented the Mtb ΔESX-1 strain with the integrative cosmid p2F9, containing 32 kb of the ESX-1 encoding genomic region from Mtb H37Rv [42]. This complementation reconstituted the ability of the resulting strain to induce phagosomal rupture, and thereby validated the ΔESX-1 mutants used throughout this study (S2C Fig.).
When uncontrolled inside the host cell, Mtb infection may lead to necrosis [27,43], which could theoretically allow exchanges between phagosome and cytosol and thereby establish a contact between mycobacterial β-lactamase located within the phagosome and CCF-4 located inside the cytosol. To investigate this key question, we determined whether the cytosolic access of Mtb was a consequence of host cell necrosis. In a dose-response experiment, changes in the FRET signal for the Mtb WT strain were seen as a function of the MOI (Fig. 2D). Except for an MOI below 1, the proportions of BM-DC displaying FRET inhibition were higher than the percentages of necrotic cells (Fig. 2E). In contrast, BM-DC infected with Mtb ΔESX-1 at the same MOIs displayed much weaker CCF-4 blue shifts. These data suggest that ESX-1-mediated phagosomal rupture progressively occurs in phagocytes in an MOI-dependent manner and that the resultant presence of mycobacterial β-lactamase activity in the host cell cytosol does not arise from host cell necrosis but rather precedes cell death.
So far, Mtb-induced phagosomal rupture has only been observed at later stages of infection, i.e, 3–5 dpi, a kinetic situation, which cannot explain the very early, ESX-1-dependent release of type I IFNs or IL-1β, that requires recognition of mycobacterial components by the host cytosolic sensors [44]. However, our highly sensitive approach allowed now detection of minor levels of FRET inhibition indicated by enhanced MFI447 nm (blue), as early as 3 hours post infection (hpi) with WT Mtb (Fig. 3A-B). The blue shift then progressively increased at 24 and 48 hpi, although it remained still low compared to values obtained for later time points (Fig. 2B-C). Comparison of these results with those from infection experiments using the Mtb ΔESX-1 deletion mutant, which overall showed much lower MFI447nm (blue) values (Fig. 3B), suggests that Mtb-mediated phagosomal rupture begins already at such early time-points, likely caused by initial ESX-1-induced pore forming activity, and progresses into stronger phagosomal disassembly over time. These findings suggest that the time during which the Mtb-infected host cell displays phagosomal rupture and Mtb cytosolic access, prior to host cell death, is longer than previously estimated [15].
Considering the long Mtb replication time of ≈ 20h, such early initiation of Mtb-mediated phagosomal rupture suggests that this phenomenon does not depend on bacterial replication, but on the functions of the implicated bacterial virulence factors. The levels of phagosome disruption were entirely proportional to the amounts of secreted IFN-β (Fig. 3C). A partially ESX-1-dependent increase in the IFN-α secretion was also detected, which might be linked to the induction of Irf7 subsequent to IFN-β induction [45]. Therefore, minute levels of early phagosomal rupture are in direct correlation with the kinetics of the induction of type I IFN production. In contrast, no differences were found between ESX-1-proficient and ΔESX-1 Mtb strains when IL-1β secretion was studied (S4 Fig.), which is consistent with the inflammasome/caspase-1-independent IL-1β secretion in mice during Mtb infection [46] and which is different to the situation in humans [36].
We next evaluated whether the characteristic Mtb-mediated partial inhibition of phagosome acidification was connected to the phenomenon of phagosome rupture. Given the previously established role of Natural resistance-associated macrophage protein (Nramp)-1, a phagosomal bivalent cation transporter, in phagosomal acidification and pH regulation [47–49], we evaluated its possible impact on mycobacteria-mediated phagosomal rupture. We thus used Mtb WT or ΔESX-1 strains to infect cells from the murine MΦ cell line Raw264.7, deficient in functional Nramp-1, which had been transfected with a non-functional nramp-1S (Sensitive) or a functional nramp-1R (Resistance) allele [50]. At 3 dpi, intense CCF-4 blue shifts were observed in WT Mtb-infected parental Raw264.7 cells and Raw264.7::Nramp-1S cells, whereas much less FRET inhibition was detected in Raw264.7::Nramp-1R cells (Fig. 4A-C). As assessed for various MOI, the intracellular mycobacterial load inside parental, Nramp-1S- or Nramp-1R-transfected Raw264.7 cells was comparable at 3 dpi, when the phagosomal rupture was monitored (Fig. 4D). Thus, the functional Nramp-1R seems to provide protection against Mtb-induced phagosomal rupture for the benefit of the host cell. The Nramp-1-mediated rescue of the host cells occurred at any MOI and independently of the host cell proliferation rate, which as we noticed, both influence the control of the infection (S5 Fig.). We obtained further confirmation of our results by using an nramp-1 gene silencing strategy in Raw264.7::Nramp-1R cells (Fig. 4E), which reversed the phenotype and promoted Mtb-mediated phagosomal rupture (Fig. 4F-G).
We further treated Raw264.7::Nramp-1R cells or, as primary phagocytes, BM-DC from Sv129 (nramp-1R) mice with bafilomycin, a specific inhibitor of vacuolar proton ATPases, prior to infection with WT Mtb H37Rv. As shown in Fig. 5A-B, the bafilomycin-mediated reduction of phagosomal acidification resulted in enhanced phagosomal rupture in both cell types. This observation provides additional evidence for a link between restriction of phagosome acidification and the strength of observed phagosomal rupture. In this FRET-based method, the β-lactamase operates on CCF-4 located in the host cytosol, where the pH remains neutral [25,41]. However, to further ascertain that the micro-environmental acidity did not affect the functionality of mycobacterial BlaC, we tested the β-lactamase enzymatic activity of Mtb at different pH levels by the use of nitrocefin, a chromogenic β-lactamase substrate. These experiments confirmed that Mtb, grown at different pH, ranging from 5 to 7, preserves entirely its β-lactamase enzymatic activity (Fig. 5C).
Thus, acidification of the phagosomal lumen seems to be a critical host cell parameters, which exerts an antagonistic effect on Mtb-mediated phagosomal rupture in phagocytes. The finding that both phenomena are linked provides a new basis for elucidating the molecular key players that govern the host-pathogen interaction during Mtb infection.
Previous studies on vacuolar rupture and phagosomal escape of M. marinum [12,51] and Mtb [13,15,16] used infected MΦ or DC under in vitro conditions. To extend our investigations towards cells from the lung, we examined the Mtb-mediated phagosomal rupture in different phagocyte types of mouse airways. To this end, low-density cells isolated from mouse lung parenchyma were infected ex vivo at an MOI of 1 with ΔESX-1 or WT Mtb strains. CCF-4 signals obtained from monocytes/MΦ (CD11bhi CD11c-) and DC (CD11bint CD11c+) were analyzed at 4 dpi, when changes in the FRET signal were detected in lung monocytes/MΦ and DC (Fig. 6A), showing the occurrence of Mtb-mediated phagosomal rupture in the primary lung phagocytes.
To assess the relevance of mycobacteria-mediated phagosomal rupture in phagocytes in vivo, in a first attempt we used T-/B-cell deficient recombination activation gene (rag) 2 knock-out mice in which infection with Mtb is more persistent and the innate cell compartments more developed than in their immunocompetent counterparts. However, flow cytometric analysis of lung- or spleen-derived MΦ/monocytes, DC and neutrophils obtained from infected (1 x 106 CFU i.v. /mouse of WT or ΔESX-1 Mtb) or uninfected rag2°/° mice displayed indistinguishable CCF-4 blue profiles (S6 Fig.). The apparent failure in the detection of phagosomal rupture in this experimental setting seems to be related to the very low frequencies of mycobacteria-infected cells within each innate cell subset and/or a possible furtive feature of the phenomenon in vivo due to possible efferocytosis [52] of the primary phagocytes, in which phagosomal rupture and certain damage signals would have been initiated.
To distinguish infected and non-infected cells, we then used fluorescent DsRed-WT Mtb (1 x 106 CFU/mouse) for intravenous (i.v.) infection of C57BL/6 mice, which allowed us to focus on the relatively few Mtb-infected phagocytes present during the initial phase of chronic infection. At 3 weeks p.i. mice were sacrificed, the spleens homogenized and resulting cells enriched and subjected to flow cytometric analysis. We have focused on the phagocytes of the spleen because this organ is particularly targeted by the i.v. route of infection. When the CCF-4 blue signal of the innate immune cells that contained DsRed Mtb was compared to the other cells inside each cell subset in the spleen (Fig. 6B), a slight increase in CCF-4 blue signal was notably detected in Mtb-containing cells in the subsets of neutrophils (CD11bhiCD11c-Ly6G+) and MΦ/monocytes (CD11bintCD11c-Ly6G-) (Fig. 6C), which suggests the occurrence of weak, albeit reproducible, levels of phagosomal rupture in these infected cells. Interestingly, no DsRed+ cells were detected inside the CD11bloCD11c+Ly6G- DC subset, which might be due to possible rapid turnover of infected DC or to their CD11b up-regulation. In this chronic infection model, it was however not possible to compare WT and ΔESX-1 Mtb strains, because of the non-persistence of the latter. To overcome this limitation we developed an alternative in vivo model whereby mice were instilled intra-nasally with cells that were infected with Mtb in vitro prior to transfer, and whose infection status in vivo could be specifically monitored. To this end, BM-DC from mice with CD45.1 hematopoietic allelic marker were infected in vitro with WT or ΔESX-1 Mtb, in conditions that allowed up to 70% of the cells to be infected (Fig. 2A), whereas control cells were left uninfected. At 16 hpi, the cells were instilled into the airways of congenic CD45.2 recipients. At different time points post-transfer, the lung low-density cells were isolated and the CCF-4 blue shift in the CD11b+ CD45.1 cell subset of the different experimental groups assessed (Figs. 7A and S7). Strikingly, at day 4 and day 6 post-transfer, in the CD11b+ CD45.1 population infected with WT Mtb, a blue shift was detected in comparison to the non-infected or ΔESX-1-infected transferred cells (Fig. 7B-C).
Moreover, independent flow cytometric examination of cells extracted directly from surface lung granuloma tissue of Mtb-infected C57BL/6 mice revealed a small, distinct cell population that displayed a clear-cut blue signal and a CD11b+ CD11c+ phenotype (Fig. 7D), which points to the presence of innate cells in these lungs wherein Mtb-mediated phagosomal rupture had occurred.
Altogether, our data suggest that the Mtb-induced phagosomal rupture does indeed happen in vivo, in Mtb-infected cells in the organs of small laboratory animals. The detected phagocytes containing intracellular bacteria seem to have a life-time of several days, which however does not exclude the possibility that a portion of the total number of infected phagocytes might get eliminated by efferocytosis [52], as suggested by the relatively modest differences in blue shift observed in the in vivo settings.
The pathogenic potential of Mtb is intimately linked to the interplay between the host defense and the persistence of the mycobacteria. The intracellular localization and cytosolic access of the bacterium has substantial consequences on the recognition of mycobacteria-associated patterns by the cytosolic receptors of the innate immunity that determine innate and adaptive immune responses and ultimately the fate of the host cell and the bacterium [27]. Subsequent to phagocytosis, in order to avoid the acidified environment generated by the phagosome-lysosome fusion, some specialized intracellular bacteria, such as S. flexneri, Listeria monocytogenes or Francisella tularensis, evolved to rapidly escape from phagosomes into the cytosol [21,53,54]. In contrast, Mtb has been described as a bacterium that resists degradation in the phagosome by inhibiting the fusion with lysosomes, a characteristic feature that seems to protect the bacilli from bactericidal mechanisms of the phagocytes and allows intracellular survival and multiplication [10,18,55–57]. However, recent reports based on in vitro infection of phagocytes also suggest that at later stages of infection ESX-1-dependent vacuolar breakage might be an important requirement for the pathogenic potential of Mtb, given that ESX-1-deficient bacilli that are unable to perforate and lyse the phagosomal membrane are—in general—attenuated [13,15,16,18,56–59].
In previous studies, Mtb-mediated phagosomal escape has only been reported at late time points like 2–5 dpi, a kinetic feature that was not reconcilable with the intracellular host immune events, like type I IFN induction, which require the early recognition of mycobacterial components by cytosolic sensors. Here, the use of highly sensitive FRET-based cytometry enabled us to highlight minor levels of cytosolic contact of Mtb and its products initiated as soon as 3 hpi, which is kinetically concordant and proportional with the amounts of IFN-β released by DC. While we cannot exclude the possibility that some of this effect may have been caused by bacterial products translocating through permeable phagosomal membranes [30], the reproducible differences observed between the WT and the ΔESX-1 Mtb strains argue for a specific, ESX-1-mediated impact. We also noted that distinct cell types might display different susceptibility to phagosomal rupture, with THP-1 cells as the most susceptible ones, followed by BM-DC/BM-MΦ, and the Raw264.7 MΦ as the least affected cell types, tested.
Our results show that the phagosomal bivalent cation transporter Nramp-1 interferes with Mtb-induced phagosomal rupture as observed at 3 dpi, i.e., a time point at which mycobacterial loads were still comparable in Mtb-infected MΦ harboring Nramp-1S (non functional)- or Nramp-1R (functional) allelic forms. In line with that, the effect of bafilomycin, reported to inhibit phagosomal acidification [60], reconstituted in Nramp-1R-proficient phagocytes the capacity of Mtb to enhance phagosomal rupture to the level of Nramp-1S phagocytes. Thus, the partial inhibition of phagosome acidification emerges as a prerequisite to mycobacterial phagosomal rupture. Plausibly, only when phagosome acidification is partially inhibited, mycobacteria may survive, use their virulence factors and induce phagosomal membrane disruption.
Although cellular models may provide important new insights into cell biological mechanisms, evaluation of the accuracy of the findings in an in vivo model, i.e. in tissues or organs is of crucial importance to emphasize their relevance. Previous electron microscopy analyses of lung innate cells isolated from TB patients or mycobacteria-infected mice have led to discrepancies with regards to intracellular location [18]. In alveolar MΦ of TB patients and in granuloma or lung homogenates of infected mice, Mtb has been detected as single bacterium or pairs of bacilli inside phagosomes [61,62], whereas Mtb has also been observed in membrane-disrupted compartments or free in the cytosol in the mouse granulomas [63,64]. Moreover, heavily infected human alveolar MΦ [62] and damaged mouse MΦ of inflammatory sites [65] contain multiple mycobacteria per phagosome. In this context, our results from carefully designed in vivo infection experiments add new elements to the discussion. Although the strength of the FRET-inhibition was found weaker under in vivo conditions (Figs. 6 and 7) than observed for the cell culture-based infection assays (Figs. 1 and 2), the reproducibility and complementarity of the results from the three distinct in vivo settings analyzed, point to biological relevance of mycobacteria-induced phagosomal rupture in the organs of Mtb-infected laboratory animals. It should be noted that in our experiment with BM-DC from mice with the CD45.1 hematopoietic allelic markers (Fig. 7), we cannot exclude that in the infected DC some minor cytosolic contact might develop already in vitro, prior to their instillation to the CD45.2 recipient mice. However, the finding that FRET inhibition remains detectable for several days after the transfer into the lungs of the CD45.2 recipients suggests that the phagocytes in which cytosolic access of Mtb progressively builds up, can survive in the host environment for some days. Together with ex vivo results from MΦ/monocytes and DC isolated from the lung parenchyma, the in vivo demonstration of cytosolic access of Mtb provides important new insights into the cellular events during infection inside the organs. Our data suggest that after infection, the concerned phagocytes may persist in the organs long enough to have a potential impact on host defense mechanisms that likely also include key cellular processes, such as autophagy, which requires Mtb ubiquitination in an ESX-1-dependent manner [16,33,38,39].
The intracellular localization of mycobacteria and mycobacteria-mediated phagosomal rupture have been subject of numerous controversies, which may be explained by the differences between the level of virulence of mycobacterial strains used, the MOI and the conditions of the mycobacterial cultures in vitro [18]. For the virulent strains, here we used WT and DsRed Mtb previously passaged in immunocompetent mice to maintain a normal degree of virulence and to remain as close to natural infection as possible. We only used mycobacterial cultures in mid-log10 growth phase to minimize bacterial mortality, and we cultured the bacteria in the presence of Tween 80 to avoid clumping, as phagocytosis of non-viable or clumped mycobacteria may lead to rapid phagosome-lysosome fusion and prevent visualization of phagosomal rupture [18]. In addition, we systematically compared the ESX-1-proficient and ESX-1-deficient mycobacterial strains and detected a relevant phagosomal rupture only with ESX-1-proficient strains.
Previous observations with numerous virulent and attenuated Mtb strains suggest that the capacity of a strain to induce phagosomal rupture in vitro is often correlated with its virulence [15,16]. Hence, the ESX-1-dependent, mycobacteria-induced phagosomal rupture emerges as a major characteristic feature of Mtb infection, which likely initiates the first damages caused by this intracellular pathogen to the host cell. Consequently, modulation of the parameters, which orchestrate this phenomenon, may constitute a promising base for vaccinal or therapeutic interventions against TB. For example, we have previously noticed that recombinant BCG and M. microti strains with a reconstituted ESX-1 secretion system showed enhanced protective efficacy [66,67]. More recently, a dedicated study identified small molecule inhibitors belonging to the benzyloxybenzylidene-hydrazine and the benzothiophene chemical classes, which interfered with ESAT-6 secretion and thereby protected host cells from Mtb-induced lysis [68]. Molecules belonging to closely related chemical scaffolds were also identified in a high content phenotypic screen as agents that interfered with the intracellular growth and the virulence of Mtb [69]. Hence, it is conceivable that future phenotypic library screening might identify novel pharmacological compounds that inhibit Mtb-mediated phagosomal rupture in the host cell. Such molecules would represent interesting anti-virulence compounds to be tested as addition to conventional treatment regimens against TB.
In conclusion, our study suggests that Mtb is not the passive pathogen that induces pathology only by the over-boarding reaction of the host immune system. We show that ESX-1-mediated phagosomal rupture contributes in a significant way to establish mycobacterial cytosolic contact, which is however only possible if the maturation / acidification of the phagosome is limited in a first process. In this direction, our study also opens new perspectives for future studies on the mycobacterial components involved in the modulation of phagosomal acidification such as the phthiocerol dimycocerosates and other mycobacterial factors, reported to intervene in this process [70,71].
The ESX-1 system might thus represent one of the final members in a chain of virulence factors that determine the pathogenicity of Mtb through the induction of phagosomal rupture, and its function might therefore have been evolutionary preserved [5,7]. As such, our work has the potential to reconcile the outcome of previous studies on mycobacterial virulence factors that interfere with vacuolar acidification [71–74] and studies on cellular localization of Mtb [13–16] and establishes Mtb-mediated phagosomal rupture as a basic biological mechanism involved in TB pathogenesis.
C57BL/6 mice, rag2°/° or CD45.1 were obtained from Animal Facilities of Institut Pasteur. C57BL/6 mice were purchased from Janvier Le Genest-Saint-Isle France). CD45.2 mice were anesthetized by i.p. injection of 100 mg/kg Ketamine (Lyon, France) and 10 mg/kg Xylazine (KCP Kiel, Germany) before cell transfer by i.n. route. Mouse infection with Mtb via aerosol route was performed as previously described [75]. Granuloma were recovered from the surface lung parenchyma of infected C57BL/6 mice at 6 weeks p.i.
Mouse studies were approved by the Institut Pasteur Safety Committee, in accordance with French and European guidelines and regulations (Directive 86/609/CEE and Decree 87–848 of 19 October 1987) and the Animal Experimentation Ethics Committee Ile-de-France-1 (reference number 2012–0005).
THP-1 cells (our laboratory stock collection, initially originating from ATCC provided cells) were maintained in RPMI, complemented with 10% heat-inactivated FBS and were treated with 20 ng/ml of Phorbol 12-Myristate 13-Acetate for 72h to induce their differentiation into MΦ. Raw264.7 cells transfected with nramp-1S or -1R allele (kind gift of Pr J. Blackwell) [50] were treated with 8 μg/ml of the selective antibiotic puromycin.
BM-MΦ or -DC were generated from femur hematopoietic precursors, respectively by use of M-CSF or GM-CSF. Rat anti-mouse IFN-α mAb (RMMA-1), biotinylated polyclonal rabbit anti-mouse IFN-α (R&D), rat anti-mouse IFN-β (8.S.415) (LifeSpan BioSciences) and biotinylated polyclonal rabbit anti-mouse IFN-β (R&D) were used to quantify the cytokines produced in the culture supernatants by ELISA.
Mtb H37Rv, WT, ΔESX-1 (kind gift of Pr. W. Jacobs) [10] or ΔESX::ESX-1 [42] were maintained in 7H9 medium supplemented with ADC (Difco). Seven-to-10 days before cell infection, bacteria were transferred into Dubos medium, which contains Tween 80, to avoid mycobacterial clumping. DsRed-WT or -ΔESX-1 strains were obtained by complementation with the pMRF plasmid containing a DsRed cassette, under the hsp60 promoter (kind gift of Dr. S. Cho) and were cultured in the continuous presence of 20 μg/ml of the selective antibiotic kanamycin. In in vivo experiments, we used an Mtb H37Rv strain with a plasmid containing the DsRed and hygromycin resistance genes (kind gift of Dr. O. Neyrolles). Only mycobacteria grown to mid-log10 phase were used to minimize the frequency of death bacteria.
Raw264.7 cells were infected at various MOI with Mtb in complete antibiotic-free RPMI. At 3 dpi, equal numbers of cells were lysed by addition of 0.1% Triton X-100 in PBS and the intracellular CFU was determined by plating serial dilutions of cell lysates on 7H9 Agar medium and incubation at 37°C for 3 weeks.
The principle of the β-lactamase CCF-4 FRET assay is summarized in S1 Fig.. To measure the Mtb phagosomal rupture, cells were stained during 1h at RT, with 8 μM CCF-4 (Invitrogen) in EM buffer (120 mM NaCl, 7 mM KCl, 1.8 mM CaCl2, 0.8 mM MgCl2, 5 mM glucose and 25 mM Hepes, pH 7.3) complemented with 2.5 μM probenecid. Cells were then stained with anti-CD11c-PE-Cy7, anti-CD11b-PerCp-Cy5.5 (eBiosciences) or anti-CD11b-APC (BD) mAbs andfixed with 4% PFA overnight at 4°C. Cell mortality in the same cultures of infected cells was determined by use of Pacific Blue Dead/Live reagent (Invitrogen), which reacts with free amines both inside and outside of the plasma membrane, yielding log10 1 more intense fluorescent staining of dead cells. Anti-CD45.1-PE-Cy7 and anti-CD45.2-PerCpCy5.5 were from eBiosciences. To avoid fluorochromes with emission signals overlapping with those of CCF-4 (λem 500–550 nm and λem 410–470 nm), APC (λem 660 nm)-, PerCp-Cy5.5 (λem 696 nm)- or PE-Cy7 (λem 778 nm)-conjugated mAbs were chosen for concomitant cell surface staining. Cells were analyzed in a CyAn cytometer using Summit software (Beckman Coulter, France). At least 100,000 events per sample were acquired for in vitro assays. For in vivo detection of CCF-4 signal in CD45 congenic mouse model, 1,000,000 events per sample have been acquired. Data were analyzed with FlowJo software (Treestar, OR).
siRNA transfection to cells was performed by using reverse transfection method. A pool of four Nramp-1-specific siRNAs, GGUCAAGUCUAGAGAAGUA, GAUCCUAGGCUGUCUCUUU, GGGCGACUGUGCUAGGUUU and GAAGUCAUCGGGACGGCUA, at final concentration of 50 nM, was mixed with 6 μl of lipofectamine (Invitrogen) in 500 μl of PBS in 6-well plates. After 30 min incubation at RT, 3 x 105 cells contained in 2 ml of complete RPMI were added to the mixture and incubated for 3 days at 37°C. The efficiency of gene silencing was determined by qRT-PCR before the infection. One mg of total RNA was transcribed into cDNA. Then, 4 μL of cDNA was tested by qRT-PCR with LightCycler 480 SYBR Green using GCCACTGTGCTAGGTTTGCT and AATGGTGATCAGTACACCGC primers. All experiments were run in triplicate and the Livak method [76] was applied for relative quantification with β-actin.
The β-lactamase activity of Mtb, grown in Dubos broth with various pH, was measured by use of the chromogenic β-lactamase substrate, nitrocefin. Briefly 1 x 106 bacteria, re-suspended in 100 μl of Dubos broth at indicated pH, were incubated in 96-well plates with 50 μl of nitrocefin, reconstituted at 0.5 mg/ml in PBS which contained 5% DMSO. Absorbance by nitrocefin at 486 nm was measured after 3 hours of incubation at 37°C.
Lungs or spleen were removed aseptically and were digested by treatment with 400 U/ml type IV collagenase and DNase I (Roche). Following a 45 min incubation at 37°C, single-cell suspensions were prepared by use of a Gentle Macs (Miltenyi) and by passage through 100-μm nylon filters (Cell Strainer, BD Falcon). When indicated, cell suspensions were enriched in low-density cells on iodixanol gradient medium (OptiPrep, Axis-Shield), according to the manufacturer’s protocol. Notably this gradient only selects alive cells, as confirmed by blue Trypan exclusion assay. These cells were either used directly in flow cytometry analyses or were plated in 12 well culture plates in complete RPMI to be infected ex vivo with mycobacteria.
|
10.1371/journal.pgen.1000579 | A Missense Mutation in the SERPINH1 Gene in Dachshunds with Osteogenesis Imperfecta | Osteogenesis imperfecta (OI) is a hereditary disease occurring in humans and dogs. It is characterized by extremely fragile bones and teeth. Most human and some canine OI cases are caused by mutations in the COL1A1 and COL1A2 genes encoding the subunits of collagen I. Recently, mutations in the CRTAP and LEPRE1 genes were found to cause some rare forms of human OI. Many OI cases exist where the causative mutation has not yet been found. We investigated Dachshunds with an autosomal recessive form of OI. Genotyping only five affected dogs on the 50 k canine SNP chip allowed us to localize the causative mutation to a 5.82 Mb interval on chromosome 21 by homozygosity mapping. Haplotype analysis of five additional carriers narrowed the interval further down to 4.74 Mb. The SERPINH1 gene is located within this interval and encodes an essential chaperone involved in the correct folding of the collagen triple helix. Therefore, we considered SERPINH1 a positional and functional candidate gene and performed mutation analysis in affected and control Dachshunds. A missense mutation (c.977C>T, p.L326P) located in an evolutionary conserved domain was perfectly associated with the OI phenotype. We thus have identified a candidate causative mutation for OI in Dachshunds and identified a fifth OI gene.
| Osteogenesis imperfecta (OI) is a genetic condition of humans and dogs characterized by extremely fragile bones and teeth. Most human OI cases are caused by defects in one of two collagen genes. Mutations in two other genes related to collagen maturation can also lead to OI in some patients. We studied Dachshunds with OI and initially investigated the two known collagen genes that are normally mutated in OI but did not find a mutation. Subsequently, we performed a search for shared segments across the entire genome in five affected Dachshunds. This experiment revealed that the causative mutation for OI in Dachshunds is located on dog chromosome 21. The SERPINH1 gene known to be involved in collagen maturation is located in this shared genome region. We sequenced the SERPINH1 gene in healthy and affected Dachshunds and found a single mutation exclusively shared by all affected dogs but not by healthy controls. Thus we have identified SERPINH1 as a fifth OI gene and a mutation within this gene as the most likely cause of OI in Dachshunds. The knowledge of this mutation enables genetic testing and will allow breeders to eradicate the deleterious allele from the Dachshund breeding population. SERPINH1 mutations might also be responsible for some human OI forms, where the causative mutation has not yet been identified.
| Collagen I is the most abundant protein in the human body and its highly ordered fibril structure is responsible for its special mechanical properties. Together with inorganic hydroxylapatite it is the main component of bones and gives them elasticity while the hydroxylapatite alone would be very brittle. Defects in the structure of the highly ordered collagen I triple helix lead to osteogenesis imperfecta (OI), a disease characterized by extremely fragile bones and teeth. OI is sometimes also accompanied by blue sclera, hearing loss, dwarfism, dentinogenesis imperfecta, and other complications. Seven subtypes of human OI are distinguished based on the underlying genetic defects and phenotypic severity [1]. OI affects an estimated 6 to 7 per 100,000 people worldwide [http://ghr.nlm.nih.gov/condition=osteogenesisimperfecta/]. Approximately 85–90% of the human OI cases are caused by mutations in the COL1A1 or COL1A2 genes encoding the two different subunits of collagen I. More than 800 distinct mutations in these two genes have been described and most of them lead to autosomal dominant forms of OI [2]. The maturation and correct folding of collagens is a complicated process, which involves a large number of accessory proteins and chaperones. Recently mutations in two of these accessory proteins were found in patients with autosomal recessive forms of OI [3]–[7]. Both of these proteins are involved in the 3-hydroxylation of a specific proline residue in collagen I. One represents the enzymatically active prolyl-3-hydroxylase 1 itself and is encoded by the LEPRE1 gene [3]. The other is called cartilage-associated protein (CRTAP) and forms a complex with the prolyl-3-hydroxylase [4]. For some human OI cases the underlying mutation has not yet been found.
OI also occurs in dogs and the dog may represent a better model for human OI than genetically engineered mice because of its larger body size and the resulting similarity of mechanical forces that act on the skeleton. OI in dogs has been described in Golden Retrievers, Beagles, Collies, Poodles, Norwegian Elkhounds, and Bedlington Terriers [8]–[12]. In Golden Retrievers a COL1A1 mutation and in Beagles a COL1A2 mutation has been reported to cause OI [8],[9]. For other canine OI cases the underlying genetic defect has not been elucidated. We have observed a severe form of OI in rough-coated Dachshunds that is inherited as a monogenic autosomal recessive trait [13]. In our initial analysis of the OI Dachshunds we did not find any mutations in the COL1A1 or COL1A2 genes. Therefore, we hypothesized that a mutation in a novel OI gene may be responsible for the observed bone defects in Dachshunds. Consequently, we started a positional cloning approach to identify this mutation.
We collected samples from six Dachshund families segregating for congenital OI (Figure 1; Video S1). The parents of all available cases were healthy (Figure S1). The pedigrees were consistent with a monogenic autosomal recessive inheritance although the ratio of affected Dachshunds from the presumed carrier x carrier matings was slightly higher than expected with 14 out of 36 total pups affected instead of the expected 9/36. The available pedigree records indicate that the affected dogs from the German Dachshund breeding population share common ancestors and most likely trace back to a single common founder. As the OI phenotype in Dachshunds shows striking clinical similarities to human OI forms, we initially hypothesized that mutations in COL1A1 or COL1A2 might cause the canine disease. In order to validate whether a mutation in one of these genes might be responsible for OI, we genotyped three gene associated microsatellite markers derived from the surrounding genome sequence of COL1A1 (located on chromosome 9 (CFA 9) at 29.5 Mb) and COL1A2 (located on CFA 14 at 22.8 Mb), respectively. Two-point linkage analysis in the six available families clearly excluded the COL1A2 gene but indicated a suggestive linkage of OI to the region of COL1A1 with a positive LOD score of 1.5 (Table S1). However, the re-sequencing of COL1A1 using DNA of four affected and four healthy Dachshunds did not reveal any disease associated sequence polymorphism within the 51 coding exons and flanking intron regions of the COL1A1 gene. Furthermore, haplotype analysis revealed five different CFA 9 microsatellite marker haplotypes in affected dogs, which was not compatible with our assumption of a single founder mutation in all OI affected dogs.
Based on the pedigrees of our samples we hypothesized that the affected Dachshunds most likely were inbred to one single founder animal. Under this scenario the affected Dachshunds were expected to be identical by descent (IBD) for the causative mutation and flanking chromosomal segments. Therefore, we decided to apply a homozygosity mapping approach to determine the position of the mutation in the canine genome. We genotyped approximately 50,000 evenly spaced SNPs from five affected dogs and five obligate carriers. We analyzed the cases for extended regions of homozygosity with simultaneous allele sharing. Only one genome region fulfilled our search criteria (Table S2). On CFA 21 all five affected genotyped dogs were homozygous and shared identical alleles over 102 SNP markers corresponding to a 5.82 Mb interval from 23.58–29.40 Mb (Figure 2). We then examined the five obligate carriers for the mutation and reconstructed one copy of the disease-associated haplotype in each dog. One of the carriers showed a recombination event, which allowed us to narrow down the critical interval harboring the causative mutation to 4.74 Mb from 24.66–29.40 Mb (Figure 2).
As the quality of dog genome annotation is still far from perfect, we inferred the gene annotation of the mapped interval from the corresponding human interval. The dog OI interval corresponds to two human segments from 3.59–3.84 Mb and from 71.30–76.48 Mb on HSA 11. The two human intervals contain 98 annotated genes including 8 annotated pseudogenes (NCBI MapViewer, build 36.3). A careful inspection of these genes and database searches of their presumed function revealed SERPINH1 as a functional candidate gene within the critical interval at 26.0 Mb on CFA 21. SERPINH1 encodes a serine protease inhibitor, also called heat shock protein 47 (HSP47) or collagen binding protein 1. Serpinh1 deficient mice die at around day 11 of development due to defective collagen synthesis [14] and it was shown that Serpinh1−/− fibroblasts produce abnormally thin and branched collagen type I fibres [15]. In order to further validate SERPINH1 as positional candidate gene for OI, we genotyped two gene associated microsatellite markers derived from the surrounding genome sequence of CFA 21 (Table S1). The obtained LOD score of 4.1 conclusively confirmed the linkage of OI to the candidate gene region in the Dachshund families. All OI affected dogs showed homozygosity at both tested microsatellites and all genotyped parents had one copy of the disease associated haplotype. We therefore investigated whether mutations in the canine SERPINH1 gene might be responsible for the OI phenotype. We designed PCR primers for the amplification of the four coding exons and determined the genomic sequence of two affected and two control dogs. This analysis revealed twelve polymorphisms including three non-synonymous substitutions (Table 1). Of these polymorphisms only a single SNP located in SERPINH1 exon 5 (c.977T>C; Figure 3) showed perfect association to the OI phenotype (Table 2, Figure S1). All 11 affected dogs were homozygous C/C and all 13 known carriers were heterozygous C/T. One grandmother and 16 out of 22 healthy full- and half-sibs of OI affected dogs were also heterozygous C/T. None of 66 unrelated healthy Dachshunds had the homozygous C/C genotype, but twelve of them were also presumed carriers with the C/T genotype. Thus the allele frequency of the deleterious C-allele within the unrelated Dachshunds was 18%. The mutation was encountered in wire-haired and short-haired Dachshunds. The mutant C-allele was absent from 79 control dogs from 75 diverse dog breeds (Table S2). RT-PCR on bone cDNA confirmed that the SERPINH1 RNA is normally spliced in an affected dog. We sequenced the cDNA from an affected dog and it did contain the mutant C-nucleotide at position +977 confirming that the mutant mRNA is expressed at normal levels. The c.977T>C substitution is predicted to result in an exchange of a highly conserved leucine to a proline in the SERPINH1 protein sequence (p.L326P, Figure 4). We modeled the wildtype and mutant SERPINH1 protein structures based on experimentally determined structures of serpins and found that the p.L326P mutation indeed affects the three-dimensional structure of SERPINH1 (Figure 5).
We have applied an efficient SNP-based homozygosity mapping strategy to map the causative gene for OI in Dachshunds using only five affected dogs and five obligate carriers. The special population structure of purebred dog breeds with a limited amount of inbreeding on the one hand increases the occurrence of recessive phenotypes and on the other hand provides ideal prerequisites to map the underlying genes for these traits [16]. In this study we did not have enough high-quality DNA samples of unrelated Dachshunds for a genome-wide association study. However, the homozygosity mapping approach for this recessive trait basically required only samples from the five affected dogs to map the causative gene to one unique chromosome segment of 5.82 Mb. Adding the five obligate carriers further reduced this interval to 4.74 Mb. Thus the use of genome-wide canine SNP genotyping data enables very efficient positional cloning projects of Mendelian traits even if only very few samples are available.
The mapped OI interval contains a very good functional candidate gene, SERPINH1. We found a non-synonymous mutation in this gene, which is perfectly associated with the OI phenotype in Dachshunds, and confirmed the presence of this mutation on the genomic DNA and mRNA level. Although we cannot provide functional proof of the causality of the mutation at this time, the wealth of functional data, which are available for the SERPINH1 gene, strongly supports the hypothesis that p.L326P is indeed the causative mutation. SERPINH1 or HSP47 is a molecular chaperone of the serpin family. It promotes the correct folding of the collagen I triple helix [17]. This triple-helical structure would normally not be stable at temperatures above 35 °C. SERPINH1 is present in high concentration in the endoplasmic reticulum and specifically binds to and stabilizes the triple helices of nascent collagens [18]–[20]. Apparently the complete absence of SERPINH1 leads to embryonic lethality due to deficiencies in several types of collagen [14]. It is an evolutionary conserved protein with 97% identity between human and dog and 64% identity between human and zebrafish. The p.L326P mutation lies within the conserved serpin domain and the wildtype leucine is conserved across all SERPINH1 sequences while in other, more distantly related serpins like antitrypsin or ovalbumin, it is conservatively replaced by isoleucine, valine or methionine. It is located at the interface of helices hB, hC and hI (Figure 5A) [21]. Leu326 has backbone dihedral angles Φ/Ψ of about −90/+88 degrees. Proline has a Φ-angle restricted to about −60 degrees due to its five-membered ring and most frequently Ψ-angles of −45 or +135 degrees. Therefore, it is likely that the mutation L326P results in an increased strain. Furthermore, in the wildtype Leu326 donates a main-chain H-bond to Leu321 that is not possible with the imino acid proline (Figure 5B). Thus it is conceivable that this mutation affects the proper folding and stability of the native conformation, possibly reducing the protein level significantly. Additionally, this non-conservative amino acid substitution could affect the ability of SERPINH1 to bind and stabilize collagen triple helices. We speculate that the p.L326P mutation in OI affected dogs probably does not represent a complete null allele but has some residual activity, which results in live-born dogs with a severe form of OI instead of the embryonic lethality seen in Serpinh1 knock-out mice. The phenotype of OI affected dogs primarily indicates a deficiency in collagen I, the most abundant collagen, whereas basal membranes, which contain collagen IV, do not seem to be severely altered [13].
Our finding of a SERPINH1 p.L326P mutation in dogs with OI provides a valuable model for human medicine and identifies SERPINH1 as a fifth OI gene in addition to COL1A1, COL1A2, CRTAP, and LEPRE1. It has already been shown that a functional SNP in the promoter of the human SERPINH1 gene is associated in African American women with an increased risk for preterm premature rupture of membranes [22]. Our study indicates that coding mutations of the SERPINH1 gene might be responsible for recessive forms of human OI, where no mutation in the four known OI genes has been found. It has been proposed to develop SERPINH1 binding molecules as drugs against fibrosis [23]. The findings of our study emphasize that such a therapeutic strategy will have to be very carefully adjusted in order not to have adverse effects on the physiological production of collagen.
In conclusion, we have identified the p.L326P mutation in the canine SERPINH1 gene as the candidate causative mutation for OI in Dachshunds. This result allows genetic testing and eradication of a lethal disease from the Dachshund breeding population. Our study also provides a defined animal model and a novel genetic mechanism for a lethal or severely debilitating human hereditary disease.
We collected samples from OI affected rough-coated Dachshunds (n = 11), their healthy littermates (n = 22), sires (n = 4), dams (n = 7), and one grandmother. We performed parentage verification to confirm the pedigree documentation from the breeders (Figure S1). In addition, we collected two Dachshunds recorded as sires of OI affected puppies. Parents of affected offspring were classified as obligate carriers (n = 13). We also collected 66 unrelated healthy Dachshunds resulting in a total of 113 samples from the Dachshund breed. Furthermore, we sampled 79 control dogs from 75 different breeds for the re-sequencing of SERPINH1 exon 5 (Table S3).
Genomic DNA was isolated from blood or tissue using the Nucleon Bacc2 kit (GE Healthcare). Total RNA was isolated from bone or skin using Trizol reagent according to the manufacturer's instructions (Invitrogen).
Microsatellite markers were amplified using the Multiplex PCR Kit (Qiagen) and fragment size analyses were determined on an ABI 3730 capillary sequencer (Applied Biosystems) and analyzed with the GeneMapper 4.0 software (Applied Biosystems). Twopoint parametric linkage analysis under the assumption of OI segregating as a biallelic autosomal recessive trait with complete penetrance was performed with Merlin software version 1.1.2 [24]. The frequency of the mutant allele in the considered population was unknown and there were no data available that would have made it possible to estimate the frequency in a reliable manner. For the calculations a frequency of 0.001 for the mutant allele was assumed. The LOD score test statistic was used to estimate the proportion of linked families and the corresponding maximum heterogeneity LOD score. Within the available families, a maximum LOD score of 5.573 would have been possible. To reconstruct the most likely haplotypes, we applied the ‘best’ option of the Merlin software.
Genomic DNA from five affected Dachshunds and five carriers was genotyped on the canine Affymetrix version 2 SNP genotyping microarray (49,663 SNPs). The results were analyzed with PLINK [http://pngu.mgh.harvard.edu/~purcell/plink/]. To identify extended homozygous regions with allele sharing across all five affected animals the options –homozyg-group and –homozyg-match were applied. All given positions correspond to the build 2.1 dog genome assembly [http://www.ncbi.nlm.nih.gov/projects/mapview/map_search.cgi?taxid=9615].
Primers for the amplification of each of the four SERPINH1 coding exons with flanking regions were designed with the software Primer3 [http://frodo.wi.mit.edu/cgi-bin/primer3/primer3_www.cgi] after masking repetitive sequences with RepeatMasker (Smit, A.F.A and Green, P. [http://repeatmasker.genome.washington.edu/]). The sequences of the primers are listed in Table S4. For the mutation analysis PCR products were amplified of two affected and two unrelated healthy dogs using TopTaq polymerase (Qiagen). The subsequent re-sequencing of the PCR products was performed after rAPid alkaline phosphatase (Roche) and exonuclease I (New England Biolabs) treatment using both PCR primers with the ABI BigDye Terminator Sequencing Kit 3.1 (Applied Biosystems) on an ABI 3730. Sequence data were analyzed with Sequencher 4.8 (GeneCodes).
Aliquots of 1 µg total RNA were reverse transcribed into cDNA using 20 pmol (T)24V primer and Omniscript reverse transcriptase (Qiagen). Two microliters of the cDNA were used as a template in PCR. PCR reactions were performed as described above and primer sequences are given in Table S5. The canine SERPINH1 cDNA sequence was deposited under accession FN395288 in the EMBL nucleotide database.
Models of wildtype and mutant SERPINH1were produced employing 3D-JIGSAW [25],[26] with template PDB entries1OO8 and 1OPH [27]. Figures were prepared using PyMOL [http://www.pymol.org].
|
10.1371/journal.ppat.1005045 | Dimerization-Induced Allosteric Changes of the Oxyanion-Hole Loop Activate the Pseudorabies Virus Assemblin pUL26N, a Herpesvirus Serine Protease | Herpesviruses encode a characteristic serine protease with a unique fold and an active site that comprises the unusual triad Ser-His-His. The protease is essential for viral replication and as such constitutes a promising drug target. In solution, a dynamic equilibrium exists between an inactive monomeric and an active dimeric form of the enzyme, which is believed to play a key regulatory role in the orchestration of proteolysis and capsid assembly. Currently available crystal structures of herpesvirus proteases correspond either to the dimeric state or to complexes with peptide mimetics that alter the dimerization interface. In contrast, the structure of the native monomeric state has remained elusive. Here, we present the three-dimensional structures of native monomeric, active dimeric, and diisopropyl fluorophosphate-inhibited dimeric protease derived from pseudorabies virus, an alphaherpesvirus of swine. These structures, solved by X-ray crystallography to respective resolutions of 2.05, 2.10 and 2.03 Å, allow a direct comparison of the main conformational states of the protease. In the dimeric form, a functional oxyanion hole is formed by a loop of 10 amino-acid residues encompassing two consecutive arginine residues (Arg136 and Arg137); both are strictly conserved throughout the herpesviruses. In the monomeric form, the top of the loop is shifted by approximately 11 Å, resulting in a complete disruption of the oxyanion hole and loss of activity. The dimerization-induced allosteric changes described here form the physical basis for the concentration-dependent activation of the protease, which is essential for proper virus replication. Small-angle X-ray scattering experiments confirmed a concentration-dependent equilibrium of monomeric and dimeric protease in solution.
| Herpesviruses encode a unique serine protease, which is essential for herpesvirus capsid maturation and is therefore an interesting target for drug development. In solution, this protease exists in an equilibrium of an inactive monomeric and an active dimeric form. All currently available crystal structures of herpesvirus proteases represent complexes, particularly dimers. Here we show the first three-dimensional structure of the native monomeric form in addition to the native and the chemically inactivated dimeric form of the protease derived from the porcine herpesvirus pseudorabies virus. Comparison of the monomeric and dimeric form allows predictions on the structural changes that occur during dimerization and shed light onto the process of protease activation. These new crystal structures provide a rational base to develop drugs preventing dimerization and therefore impeding herpesvirus capsid maturation. Furthermore, it is likely that this mechanism is conserved throughout the herpesviruses.
| The family Herpesviridae is divided into the three subfamilies alpha-, beta- and gammaherpesviruses. These infectious agents cause a variety of diseases in many different hosts including humans. Pseudorabies virus (PrV) is a neurotropic porcine alphaherpesvirus [1] and the causative agent of Aujeszky's disease. The pig is the only susceptible species that can survive a PrV infection depending on the age of the animal and virulence of the virus, while most other mammals die within a few days. Only higher primates including humans and equids are resistant to infection. Due to its broad host range PrV has become an important model system to study herpesvirus biology in cell culture and in the natural host. PrV genome organization and protein content exhibit significant homology to that of the human herpes simplex virus type 1 (HSV-1) [2,3], which is among the best-studied herpesviruses.
Capsid assembly of HSV-1 has been intensively analyzed. However, since herpesvirus capsid proteins are well conserved, it is very likely that the process leading to mature, DNA-filled nucleocapsids is also similar. The proteolytic activity of the serine protease is essential for this process [4]. In HSV-1 and PrV, this protease is encoded by the UL26 gene [5], which is the longest open reading frame in a family of in-frame overlapping genes [5–8]. UL26 overlaps in frame with UL26.5 [3]. UL26 and UL26.5 possess identical 3'-termini, which encode a scaffold protein while the unique 5'-terminus of UL26 contains the protease domain. There are at least two target sites for the protease in the full-length UL26 protein (pUL26) [8,9]. Autoproteolytic activity at the release site (R-site) results in release of the N-terminal protease domain (pUL26N, also called VP24 or generic: assemblin) and the C-terminal part containing the scaffold protein (pUL26C, also called VP21 or generic: assembly protein) [10]. Due to the presence of a linker region pUL26C is 21 amino-acid residues longer than pUL26.5 (Fig 1). Near the C-terminus of the scaffold protein is the maturational site (M-site) where pUL26.5 and pUL26C are cleaved [10]. The scaffold protein binds to the major capsid protein pUL19 (VP5) and directs it to the nucleus [11–17]. During capsid assembly, the scaffold protein forms a scaffold core with the major capsid protein bound to the C-termini of pUL26/pUL26C/pUL26.5 [17]. When the capsid is fully assembled the scaffold is cleaved at its M-site releasing the ring-like scaffold structure from the capsid, which is then expelled during DNA packaging. In contrast, the protease remains in the nucleocapsid [18]. Without protease activity, the scaffold remains in the capsid resulting in capsids without viral DNA designated as B-capsids. Upon activation of the protease, the B-capsids mature and subsequent steps of viral replication occur as shown with a temperature-sensitive HSV-1 pUL26 mutant [19].
The function of pUL26.5 can be taken over by pUL26C but with reduced efficiency as well as loss of an apparent core structure in the resulting capsids [22]. Proteases of several herpesviruses, such as human cytomegalovirus (HCMV) and Kaposi's sarcoma-associated herpesvirus (KSHV), have one or more internal cleavage sites to regulate activity or promote destabilization of the protease [23–25]. For proteases of PrV, HSV-1 and herpes simplex virus type 2 (HSV-2) no internal cleavage sites have been reported.
Herpesvirus maturational proteases exist in a monomer-dimer equilibrium. Dimers are weakly associated with dissociation constants (KD) in the micromolar range [26] and are active, while monomers are almost inactive [27]. It was shown that dimerization of the HCMV assemblin is dependent on protein concentration. The fraction of dimeric protease at 0.2 μM was demonstrated to be ~0.3 increasing to ~0.7 at 4.5 μM [27]. Additionally, dimerization is favored by high concentrations of kosmotropic compounds like glycerol [26,27].
Activity of the herpesvirus protease has to be strictly regulated. The enzyme is expressed as full-length pUL26 in the cytosol of infected cells where its concentration is low and thus the inactive, monomeric form is predominant [28]. The major capsid protein is bound by pUL26.5 and translocated to the nucleus via the nuclear localization sequence within pUL26.5 [29]. Since pUL26 encompasses pUL26.5, it also contains this nuclear localization sequence. Therefore, autoproteolytic activity of the protease in the cytosol would prevent its localization to the nucleus. When capsid assembly is completed, the protease has to become active to release the scaffold protein from the capsid and to allow DNA packaging. In the capsids, the concentration of protease is much higher than in the cytosol, thus promoting dimerization [28]. Additionally, it was proposed that the capsid environment itself might enhance proteolytic activity of the protease [30].
Several structures of homologous assemblins of other herpesviruses have been published, revealing the overall fold, active site and biological assembly [31–38]. The sequence identities of these structures to the PrV assemblin range from 60% (alphaherpesviruses) to 30% (beta- and gammaherpesviruses) [39,40]. All assemblins consist of 6–9 α-helices surrounding a β-barrel formed by two β-sheets. The catalytic triad is unique among the serine proteases and consists of one serine and two histidine residues. The active site is solvent accessible and distal to the dimer interface. Nevertheless, dimerization drastically influences the activity [27]. There is evidence, that upon dimerization the oxyanion hole is formed by structural changes of a loop containing two strictly conserved arginine residues [41,42].
Currently, crystal structures are available for native dimeric and covalently inhibited dimeric assemblins. Recently, three structures of the truncated KSHV assemblin in complex with helical-peptide mimetics were published [43,44]. These compounds bind to the dimerization area and disrupt the dimerization interface of full-length KSHV assemblin [45]. Here, we report crystal structure analyses of the active dimer, the diisopropyl phosphate-inhibited dimer as well as the non-inhibited monomer of pUL26N from PrV (224 amino-acid residues). The latter is the first-ever structure of an assemblin in its native monomeric state. Comparison of the monomeric and dimeric forms provides insight into the regulation of protease activity by dimerization and a structural basis for rational design of therapeutic substances that trap the protein in the inactive monomeric state.
Additional details and information about herpesvirus proteases and its involvement in capsid maturation are reviewed elsewhere [21,46–52].
The active site serine is solvent accessible and the catalytic residues are part of β-strands β5 (Ser109) and β6 (His128) and the β2-β3 loop (His43), like in earlier structures [31–37]. In our inhibitor complex, covalent binding of diisopropyl fluorophosphate has formed a phosphate ester with the active-site serine in a manner also observed in crystal structures of homologous assemblins [34]. In the native structure, a chloride ion occupies the oxyanion hole, which is formed by the β6-β7 loop (residues 133–142, further referred to as oxyanion-hole loop, OHL, Fig 3). Two consecutive arginine residues in this loop (Arg136 and Arg137) are strictly conserved throughout all herpesvirus maturational proteases (S3 Fig). The backbone N-H of Arg136 provides the first hydrogen-bond donor of the oxyanion hole. A water molecule as second hydrogen-bond donor supports anion stabilization, as does the positive local environment established by these two arginine residues. For HCMV assemblin it was shown that this water molecule plays a role in catalysis [56]. It is kept in place by a second water molecule that is positioned by hydrogen bonds of the peptide backbone oxygen of Leu10 (loop β1-α1) and the peptide N-H of Leu110 (β5). Both water molecules are present in the dimeric structures of assemblins from PrV (this report), HSV-2 (pdb entry 1at3), KSHV (pdb entries 1fl1, and 2pbk) and HCMV (pdb entries 1cmv, 1wpo, 1id4, 1iec, 1ied, 1ief, and 1ieg). The positions of these water molecules seem to be conserved in all active assemblins. Our findings for the composition of the oxyanion hole are consistent with those reported for HSV-2 [34].
An extensive network of hydrogen bonds stabilizes the OHL (S4 Fig). This network includes conserved water molecules and parts of the α1 region, α8 and β5. There are five strictly conserved residues in the OHL and each of these is involved in the network of hydrogen bonds that positions Arg136 and stabilizes the oxyanion hole.
The side chain of Arg137 forms hydrogen bonds to the peptide backbone of Leu20 (α1) and Leu110 (β5), which are also conserved in assemblins (S3 Fig). The OH-group of the conserved Thr140 forms a hydrogen bond to the backbone oxygen of Arg137 whereas the backbone oxygen of Thr140 accepts a hydrogen bond from the backbone N-H of Ala108 (β5). Inspection of other assemblin structures showed that these hydrogen bonds are present in all dimeric structures. In the PrV assemblin, the peptide oxygen of Val138 is connected to the peptide N-H of Leu12 via hydrogen bonds mediated by a conserved water molecule (found in HSV-2 and KSHV assemblins). Most likely this water molecule and hydrogen bonding pattern are present in all herpesvirus maturational proteases.
Additionally, the OHL of PrV pUL26N is maintained by hydrogen bonds of the side chains of Asp16 (α1), Glu214 and Arg211 (both from α8) with the peptide backbone of Val138, Gly139 and Gly135, respectively. These hydrogen bonds are also present in HSV-2 assemblin with Glu214 replaced by Gln. Identical residues are present in the VZV assemblin, but the side chains of Glu and Asp point in different directions and do not form the corresponding hydrogen bonds. This ambiguity may result from the limited resolution of that structural model (pdb entry 1vzv with a resolution of 3 Å, no structure factors were deposited). Most likely, the loop is stabilized comparable to HSV-2 and PrV assemblins.
Diffraction datasets of dimeric pUL26N were collected from needle-shaped crystals and phasing by molecular replacement was successful either by using a monomer or the complete dimer as search models. For datasets derived from morphologically different, plate-shaped crystals, molecular replacement was only successful when using a monomer as a search model. In the resulting structure, two chains are present in the asymmetric unit. These chains and their adjacent symmetry mates are not in a proper position to form the known dimer. PDBePISA [53] suggests one assembly. This putative dimer has no local dyad, which is unusual for biologically relevant dimers [57,58]. The interface area (978 Å2) is much smaller than average for homodimers of this molecular weight (~1,500 Å2) [57]. Furthermore, the helices forming the interface have B-factors (60–100 Å2) above average (52 Å2). Thus, the suggested dimeric interaction is a mere crystal contact and the crystal structure actually corresponds to monomeric pUL26N.
The region corresponding to helix α1 in dimeric PrV pUL26N is disordered in the monomeric form. For ease of comparison, in this report numbering of the helices in the monomeric protease is adapted to that of the dimeric protease. The β-barrel and distal side of the dimerization area of chain A align very well with chain B of the asymmetric unit. The OHL and helices α3, α7 and α8 of the dimerization area, on the other hand, differ in chain A and B (Fig 4). The OHL has two alternative conformations in chain A with equal occupancies. One conformation correlates to the OHL of chain B and the other one is shifted towards α8 due to crystal contacts. Compared to chain B the helices α3 and α7 of chain A are bent and helix α8 is slightly tilted. Accordingly, the overall r.m.s. deviation of chain A and B is 1.03 Å (207 aligned Cα atoms). Furthermore, there is no electron density observed for residues 16–19 of chain A (β1-α2 loop) and very weak electron density for residues 14–18 and 194–196 of chain B (β1-α2 loop and the short α7-α8 loop, respectively).
The core of the monomer is rigid as indicated by low B-factors of Cα atoms in the β-barrel, helix α4 and most of helix α8 of chain A (Fig 4). In contrast, the periphery of both monomers is flexible as evidenced by increased B-factors of Cα atoms at the distal side of the dimerization area, the OHL, and the dimerization helices α3 and α7 as well as in helix α8 of chain B. B-factors around 100 Å2 are also observed at the N-termini of helices α8 in both chains.
Taken together, these observations show that the dimerization area and the parts necessary for formation of the oxyanion hole in dimeric pUL26N are flexible and not strictly ordered in monomeric pUL26N.
The tertiary structures of monomeric and dimeric PrV pUL26N are partially similar. The r.m.s. deviations of inhibited dimeric pUL26N chain B with chain A and chain B of monomeric protease are 0.93 Å (192 aligned Cα atoms) and 1.01 Å (190 aligned Cα atoms), respectively. The β-barrel, the distal side of the dimer interface, and helix α4 of the dimer interface are almost identical in monomeric and dimeric pUL26N (Fig 5). The r.m.s. deviations of these parts of inhibited dimeric pUL26N chain B with chain A and chain B of monomeric protease are 0.54 Å (151 aligned Cα atoms) and 0.63 Å (149 aligned Cα atoms), respectively. In contrast, significant differences are observed for the OHL, N- and C-termini of helices α3, α7 and α8 of the dimer interface, and the α7-α8 loop (Fig 5).
In the monomer, the N-terminus of helix α8 is one and a half turn longer, but the C-terminus of helix α7 is one turn shorter. Thus, the α7-α8 loop conformation is shifted towards α8 upon dimer formation. In dimeric pUL26N helices α3 and α7 form a hydrophobic cleft that is occupied by helix α7 of the second monomer within the dimer. This cleft is closed in the monomer by bending the C-termini of these helices towards each other (Fig 5).
In the monomeric pUL26N, the OHL is positioned near the N-terminus of α8. Thus, the activation of pUL26N relies on dimerization-induced allosteric changes, which shift the OHL towards the active site (Fig 6A). The top of the OHL (Arg136Cα) moves approximately 11 Å. In the dimer, the N-terminus of α8 is unwound, so some residues which point towards the protein core in the monomeric pUL26N are buried by the adjacent monomer of the dimer (Fig 6B). However, the side chain positions of the catalytic triad are unchanged as suspected by Batra et al. [41].
Comparison of the active and inactive conformations of the OHL reveals a key location in the structure that is occupied by alternative hydrophobic residues. In dimeric PrV pUL26N, Ile134 is present at this position, whereas in the monomeric form Val138 takes its place (Fig 6A). The hydrophobic character of the Ile134 position is highly conserved in all sequences and three-dimensional structures of assemblins (S3 Fig). The hydrophobic character of Val138 is type-conserved in alphaherpesvirus sequences. Both residues are kept in place by hydrophobic interactions with two side chains of helix α8 (Leu207, Val208) or the corresponding helices in related structures. The hydrophobic character of these residues is conserved in herpesvirus proteases, although Thr and long, partially aliphatic side chains appear to be tolerated (e.g. Thr in Epstein-Barr virus (EBV) assemblin, or Lys in KSHV assemblin, S3 Fig).
Another consequence of the different conformation of the OHL in the monomeric form is that the Asp16-containing region and α1 are disordered. This loop, the OHL and major parts of the dimer interface area are flexible in monomeric pUL26N as indicated by partially disordered segments and high B-factors. Thus, the crystallographically observed increase of order upon dimerization of PrV assemblin is in line with the disorder-to-order mechanism of dimerization previously proposed for KSHV assemblin [28,43].
In our crystallization assays we used a construct of PrV pUL26N that was shortened by one C-terminal amino-acid residue (Ala225) compared to the in vivo form. The resulting overall fold, conformation of the OHL, and position and orientation of the residues of the catalytic triad are identical to those in previously determined dimeric full-length assemblin structures of related herpesviruses, strongly suggesting that the C-terminal deletion does not affect the activity or structure of the protease. Moreover, native and inhibited dimers of PrV pUL26N crystallized isomorphously and the diisopropyl phosphate-ester of the active-site serine is quantitatively observed in the inhibited structure.
The inhibitor diisopropyl fluorophosphate reacts specifically with the active-site serine, but hydrolyzes in aqueous solutions with a half-life of one hour at pH 7.5 and even faster at higher pH [59]. After incubation of PrV pUL26N with a sevenfold excess of diisopropyl fluorophosphate (5 mM) in a buffer at pH 7.5 at room temperature for one hour, the protein was crystallized at pH 8. Based on these constraints quantitative inhibition of PrV pUL26N (minimum 90% in the crystal) prior to inhibitor hydrolysis can be calculated to occur at a second order reaction rate of at least 0.09 s−1 M−1. This is consistent with the known reaction rate of full-length HSV-1 assemblin with diisopropyl fluorophosphate of 1 s−1 M−1 at higher pH (pH 8.0) and higher temperature (30°C) [60]. Thus, the catalytic site of PrV pUL26N remains fully reactive with respect to this substrate-like inhibitor. Furthermore, earlier work has shown that proteolytic activity with respect to substrates containing the M-site is not significantly reduced by 3- or 8-residue C-terminal truncations [61]. C-terminally extended assemblins on the other hand do show a considerable decrease in activity [37,61,62], presumably because such an extension sterically interferes with the proper positioning of helix α8, which is required to establish the correct conformation of the OHL. The absence of Ala225, however, does not interfere in any way with helix α8, the conformation of the active site, and the OHL. Indeed, in the crystal structure of VZV assemblin the C-termini are disordered [35], further confirming that the C-terminal region does not have a substantial impact on the overall protein fold.
Structural properties and self-association behavior of PrV pUL26N in solution were characterized by small-angle X-ray scattering (SAXS). Data were recorded at protein concentrations between 0.5 mg/ml and 10 mg/ml. Normalized SAXS intensities vary with protein concentration, suggesting the presence of more than one oligomerization state. The experimental SAXS curves can be accurately described assuming a monomer-dimer equilibrium and using form factors for the individual states derived from the monomeric and dimeric crystallographic models (a representative example is shown in S5 Fig). Monomer volume fractions resulting from curve fitting with the program OLIGOMER [63] are shown in S5 Fig and S1 Table. At the highest protein concentration investigated (10 mg/ml, which approximately corresponds to the starting concentration in our crystallization experiments), a considerable volume fraction (30%) of pUL26N is present in the monomeric form. This observation is consistent with the fact that crystals of the monomer could be obtained under these conditions. In contrast, inclusion of 0.2 M MgCl2 in the buffer used for SAXS measurements markedly shifts the monomer-dimer equilibrium, resulting in nearly complete dimerization at a protein concentration of 10 mg/ml. This result may explain why crystallization of the dimeric form of pUL26N required the presence of MgCl2. Particle shape under conditions that result in virtually complete dimerization according to the analysis with OLIGOMER (i.e. 10 mg/ml protein in the presence of MgCl2) was also determined independently by means of an ab initio approach. The size and oblong shape of the final bead model obtained with DAMMIN [64] are in good agreement with the crystallographic structure of the pUL26N dimer (S5 Fig), further confirming the nature of the concentration-dependent self-association that is observed here. The dissociation constants with and without MgCl2 can be estimated to 200 μM and 50 μM, respectively (S5 Fig). Higher oligomers than dimers are not observed.
Three structures of KSHV assemblin (KA) with helical-peptide mimetics (HPMs) have been published to date with pdb entries 3njq, 4p2t, and 4p3h [43,44]. These HPMs disrupt dimer formation in full-length KA as shown by size-exclusion chromatography [45]. In solution, both C-terminal helices of the monomeric KA are unfolded according to NMR- and CD-spectroscopic data [28,43]. Therefore, 34 C-terminal residues of KA were truncated for crystallization of these HPM complexes. One of these truncated helices is the major dimer-interface helix, so this truncated KA is obligate monomeric. Consequently, the models of these HPM complexes were stated as monomeric [43,44]. Inspection of these models led us to the conclusion, that the HPMs function as an artificial dimer interface for this truncated KA. PDBePISA suggests dimers or higher association states for HPM complexes (buried surface area of ~1,200 Å2 per monomer). The monomer is termed as unstable in solution, because the hydrophobic HPMs would be heavily solvent exposed. Compared to the native dimeric structure of full-length KA, one monomer is rotated approximately 80° around an axis roughly perpendicular to the dimer interface in the artificial, inactive HPM complexes of truncated KA (S6 Fig).
Although artificial, the HPM-interface underlines the importance of hydrophobic interfaces as suitable drug targets. The used HPMs were reported to bind to assemblin dimer interfaces of all herpesvirus classes. The IC50-values for these HPMs against the representative alphaherpesvirus assemblin (HSV-2 assemblin) were significantly higher than those against HCMV, EBV and KSHV assemblins, indicating much weaker binding to HSV-2 assemblin [44]. The C-terminal helices of alphaherpesvirus assemblins are likely ordered due to the hydrophobic key position being occupied by alternative conserved hydrophobic residues of the OHL. Thus, the HPMs have to compete against the C-terminal helices for binding which explains the observed higher IC50-values. Additionally, the most important residues for dimerization in KA, the so-called “hot spot” residues [65], are Trp109 (α4) [43] and maybe Phe76 (α3). These correspond to Tyr (α4) and Leu (α3), respectively, in PrV, VZV, HSV-1 and HSV-2 assemblins (S3 Fig). Neither Leu nor Tyr are considered as “hot spot” residues [65] and thus, binding of these HPMs is likely weakened. Screening for suitable mimetics is necessary to achieve specific and efficient inactivation of alphaherpesvirus proteases.
As mentioned above the HPMs force an artificial assembly of truncated KA in contrast to our native monomeric PrV assemblin. Therefore, a detailed comparison of monomeric PrV assemblin and HPM complexes of truncated KA is discussed in the supplement (S1 Text and S7–S10 Figs).
A putative cation was found in both dimeric structures from PrV. A distorted octahedral arrangement of the coordinated water molecules with mean distances of 2.1 Å strongly suggests the presence of a cation rather than water. Since the crystallization buffers contain MgCl2, it is highly probable that these cations are Mg2+ ions. For verification, divalent cations with higher electron density were tested in the crystallization procedure. Mn2+ was able to substitute for Mg2+ but the resulting crystals diffracted considerably less well. The best dataset that we managed to collect at a wavelength near the Mn absorption edge had a resolution of 3 Å, but no significant anomalous signal was detected and electron density at the putative Mg2+/Mn2+ position was too weak to unambiguously confirm the presence of a Mn2+ ion. Presumably, the presence of putative Mg2+ ions in the structures is a direct result of the crystallization conditions (containing 400 mM MgCl2) and does not reflect functional Mg2+ binding by the native protein since the metal ions compensate negative charges from two adjacent dimers, rather than within one dimer (S11 Fig). Thus, Mg2+ ions or divalent cations with related properties are presumably necessary for the observed crystal packing of dimeric pUL26N from PrV.
For HSV-1 protease a temperature-sensitive (ts) phenotype was described by mutating Y30F and A48V [66]. Mutating the corresponding residues in PrV protease (Y13F/A30V), however, did not result in the desired phenotype. Hence, these mutations may not sufficiently destabilize PrV pUL26N. In HSV-1 protease, it is highly probable that the result of these mutations is a destabilization of the region around helix α1. This helix and surrounding loops stabilize the OHL in the active conformation. In the PrV assemblin Arg24 is located at the N-terminus of α2. This residue stabilizes the α1 region, because its side chain forms a hydrogen bond to the peptide oxygen of Asp59 (Fig 7). This hydrogen bond is missing in HSV-1 assemblin, since the corresponding residue is Pro42. Consequently, we propose a Y13F/R24P/A30V mutant for achieving the desired ts phenotype in PrV assemblin. If the additional R24P exchange completely inactivates the protease, an R24K variant could also be taken into account. The slightly shorter side chain of Lys may cause a varied hydrogen-bond network and a weaker hydrogen bond to the peptide oxygen of Asp59.
During capsid assembly, pUL26 accumulates in the nascent capsids because of its C-terminal scaffold-protein part. Thus, the local high concentration of the protease is promoting dimerization and autoproteolysis occurs to release the scaffold from capsids for DNA packaging. In comparison with its dimeric structure, the monomer of pUL26N reveals changes at the dimerization area, in line with allosteric changes of a loop forming the oxyanion hole. As previously anticipated, the core of the protease including the positions and orientation of the active-site residues remains unchanged, but the oxyanion hole is disrupted in the monomeric form [41,42]. The monomeric structure presented here is not truncated and does not contain any inhibitors. Thus, it constitutes the first reliable model for native monomeric structures of other assemblins.
Dimerization induces the following allosteric events: helix α7 of a monomer interleaves between helices α3 and α7 of a second monomer moving these helices farther apart. At the same time, helix α7 is elongated by one turn at its C-terminus to the cost of one N-terminal turn of helix α8. This allosteric process forces the OHL to shift to the vicinity of the active-site serine and builds a far-reaching network of hydrogen bonds with side chains of helix α8 and the polypeptide of strand β5. Furthermore, residues 13–20 of the β1-α2 loop become ordered and form helix α1 in the dimer by getting involved in that network of hydrogen bonds. In this position, the OHL forms the oxyanion hole and activity of the protease is established. The extent of disorder at the dimerization area will vary in assemblins of different herpesviruses, but the general disorder-to-order mechanism of dimerization [28,43] will very likely hold for all assemblins.
The molecular structure of dimeric pUL26N will help to engineer a temperature-sensitive phenotype of the PrV protease. A temperature-sensitive variant will be a powerful tool to observe subsequent steps of viral replication in a synchronous wave [19]. This will provide valuable data on kinetics for cleavage, packaging of the DNA, nuclear egress and intracellular trafficking of the virions.
The structure of monomeric PrV assemblin is the paradigm for monomeric states of other assemblins, primarily from alphaherpesviruses. Detailed knowledge of this structure, conformational changes and sequence specific contacts upon dimerization are a rational basis for the development of drugs binding to the dimerization area and, thus, trapping the inactive monomeric state [45]. Inhibition of dimerization suppresses protease activity and therefore prevents the assembly of fully functional virus capsids.
Small-angle X-ray scattering (SAXS) of PrV assemblin in solution revealed that dimerization increases with protein concentration and in the presence of MgCl2. Dissociation constants in the micromolar range are comparable to those observed for other assemblins [26,27]. Divalent cations like Mg2+ or Mn2+ are required for crystallization of the dimeric PrV assemblin, because these cations support crystal packing by compensating negative charges of neighboring dimers. Since MgCl2 shifts the condition of equilibrium towards dimeric pUL26N, the concentration of the monomeric form is likely below the critical nucleation concentration resulting in crystals of the dimeric form only. Accordingly, crystallization of the monomeric assemblin requires absence of divalent cations. The monomeric fraction of ~0.3 is sufficient for nucleation and the monomer-dimer equilibrium provides a steady supply of monomeric pUL26N. The monomeric form is increasingly favored since crystallization of monomeric pUL26N decreases the concentration of pUL26N in solution. Similar cases with a minor monomeric fraction crystallizing from a monomer-dimer equilibrium were reported earlier [67].
Full-length pUL26 cleaves itself at two positions, and therefore expression of the full-length protein leads to an inhomogeneous product that is unlikely to crystallize. Accordingly, cleavage was prevented by cloning a stop codon behind Gln224. The resulting coding sequence contains the protease fraction of pUL26 (pUL26N) only.
N-terminally (His)6-tagged pUL26N was expressed using a pET28a+ vector in E.coli strain BL-21 (DE3). Cells were grown in LB medium to an OD600 of 0.5–0.8 at 37°C and then induced by addition of isopropyl β-D-1-thiogalactopyranoside to a final concentration of 1 mM. Cells were lysed by sonication. The protein was purified by performing immobilized metal-ion affinity chromatography using a Poros MC 20 column loaded with Ni2+ ions (0.5 M NaCl, 50 mM Tris/HCl pH 7.5, 5% glycerol, eluted with a gradient of 0–250 mM imidazole). The protein was checked for its purity by SDS-PAGE and then concentrated to ~20 mg/ml. Aliquots were stored at −80°C. The purified protein was not tested for enzymatic activity.
Crystals were obtained using the hanging-drop vapor diffusion method at 22°C. First crystals grew in drops containing 1 μl pUL26N concentrate and 1 μl precipitant solution (0.1 M Hepes pH 7.5, 20% PEG 8,000) within several days. The quality and size of the crystals could be increased by optimization of the composition of the precipitant solution and the ratio of protein solution to precipitant solution. The morphology of the crystals changed from plate-shaped to needle-shaped when MgCl2 was used as an additive in the crystallization procedure. Plate-shaped crystals turned out to be monomeric pUL26N, whereas the dimer formed needle-shaped crystals. For crystallization with inhibitor, the concentrated protein solution was incubated with a final concentration of 5 mM diisopropyl fluorophosphate for 1 hour prior to crystallization.
Best crystals of native dimeric, inhibited dimeric and monomeric pUL26N grew in drops containing 2 μl pUL26N concentrate and 1 μl precipitant solution. Precipitant solution for monomeric pUL26N consisted of 0.1 M Tris/HCl pH 8, 8% PEG 8,000 and crystals grew within one week. Precipitant solutions for native dimeric and inhibited pUL26N consisted of 0.1 M Tris/HCl pH 8, 14% PEG 8,000, 0.4 M MgCl2 and 0.1 M Tris/HCl pH 8, 20% PEG 8,000, 0.2 M MgCl2, respectively.
All crystals were cryo-protected by soaking for 15 s in drops of precipitant solution with increasing amounts of PEG 400. Final concentrations of PEG 400 were 25% (monomer), 18% (native dimer) and 17% (inhibited dimer). Cryo-protected crystals were flash-frozen in liquid nitrogen.
Datasets were measured at 100 K with a PILATUS-6M detector at beamline BL14.1, operated by the Helmholtz-Zentrum Berlin (HZB) at the BESSY II electron storage ring (Berlin-Adlershof, Germany) [68]. 1,800 images were collected at X-ray wavelength 0.91841 Å with an exposure time of 0.5 s and an oscillation range of 0.1°.
All datasets were processed using XDS and Aimless [69–72]. Further details on data collection and processing are listed in Table 1. For determination of the Matthews coefficient and solvent content of the unit cell the CCP4 Program suite (version 6.4.0) was used [73–76].
The structures of monomeric and native dimeric pUL26N were solved via molecular replacement using Phaser [77]. The starting model for native dimeric pUL26N was the A-chain of pdb entry 1at3 (the homologous protein from HSV-2) edited with Chainsaw [78,79]. The A-chain of native dimeric pUL26N (pdb entry 4v07) was used as the starting model for monomeric pUL26N. Since both dimeric forms were isomorphous, the inhibited structure could be solved by refining the native structure against the dataset of the inhibited form. Cycles of model building and refinement were carried out using Coot (version 0.7.1) and Refmac5 (version 5.8.0073), respectively [80–88]. Further details on refinement are listed in Table 1. The oligomerization states were confirmed by PDBePISA [53]. All figures representing structural models were prepared using PyMOL version 1.7.1.3 [89].
The diffraction data and refined models of monomeric, native dimeric and inhibited dimeric pUL26N were deposited with the Protein Data Bank under entry codes 4v0t, 4v07, and 4v08, respectively. A preliminary dataset of the monomeric pUL26N at 2.5 Å resolution was deposited with entry code 4cx8.
SAXS data were recorded at beamline P12 of the EMBL outstation at PETRA III, DESY, Hamburg [90], using a PILATUS 2M pixel detector, a sample-to-detector distance of 3.1 m and a wavelength of 1.24 Å. Solutions contained 0.5 M NaCl, 50 mM Tris pH 7.5, 0.25 M imidazole, 5% glycerol, 50 mM urea and pUL26N as indicated. In all experiments the sample temperature was 283 K. Measurements covered the momentum transfer range 0.008 < s < 0.47 Å-1 (s = 4π sin(θ) / λ, where 2θ is the scattering angle and λ is the X-ray wavelength). To monitor radiation damage, 20 successive 50 ms exposures of protein solutions were compared, revealing no significant change. The data were normalized to the intensity of the transmitted beam and radially averaged. Scattering of the buffer was subtracted and the difference curves were scaled to unity protein concentration (1 mg/ml). For further data analysis, version 2.6.0 of the ATSAS package was used [91]. Form factors were generated from the monomeric and dimeric crystallographic models by means of the program FFMAKER. For subsequent curve fitting, the program OLIGOMER [63] was used. An ab initio model corresponding to the highest protein concentration in the presence of MgCl2 was generated using the programs DAMMIF [92], DAMAVER [93] and DAMMIN [64] via the PRIMUS interface [63], in "slow" mode and without imposing particle symmetry.
The scattering data, structural models and curve fittings of dimeric PrV pUL26N were deposited with the small-angle scattering biological data bank (SASBDB) with entry code SASDA58 [94].
PrV UL26: Gene ID 2952508
HSV-1 UL26: Gene ID 2703453
PrV UL26.5: Gene ID 2952525
HSV-1 UL26.5: Gene ID 2703454
PrV pUL26: Q83417 in UniProtKB, S21.001 in MEROPS [95]
PrV pUL26N: Amino-acid residues 1–224 of Q83417 in UniProtKB, S21.001 in MEROPS [95]
PrV pUL26C: Amino-acid residues 226–524 of Q83417 in UniProtKB
PrV pUL26.5: Q83418 in UniProtKB
KA: Q2HRB6 in UniProtKB, S21.006 in MEROPS [95]
|
10.1371/journal.pcbi.1003437 | Timing over Tuning: Overcoming the Shortcomings of a Line Attractor during a Working Memory Task | How the brain stores information about a sensory stimulus in working memory is not completely known. Clues about the mechanisms responsible for working memory can be gleaned by recording from neurons during the performance of a delayed response task. I focus on the data recorded during such an experiment, a classic tactile discrimination task. I describe how the observed variability in the firing rate during a trial suggests that the type of attractor that is responsible for holding the stimulus information is not a fixed-point type attractor. I propose an alternate mechanism to a line attractor that allows the network to hold the value of an analog stimulus variable for the duration of the delay period, but rather than maintain a constant level of activity, the cells' firing rate varies throughout the delay period. I describe how my proposed mechanism offers a substantial advantage over a line attractor: The tuning requirements of cell to cell connections are greatly eased from that of a line attractor. To accommodate a change in the length of the delay period, I show that the network can be altered by changing a single parameter - the timing of an executive signal that originates outside of the network. To demonstrate the mechanism, as well as the tuning benefits, I use a well known model of propagation in neuronal networks.
| The ability to retain stimulus information after the stimulus has ceased is important for survival. The term “working memory” refers to storage of stimulus information for a short period of time, so that this information can be recalled for a decision making process. A common way to probe for the cellular basis of working memory is recording of neurons during a delayed response task. This study focuses on one of these studies - the now classic experiments of Romo et al.. This experiment demonstrates that the frequency of a tactile vibration is held in memory using a type of encoding where the cellular output depends monotonically on the stimulus variable (frequency). In this paper, I develop a model that accounts for a number of features found in the data. Using the model, I am able to account for a diversity of cellular responses, as well as variability during a trial. This paper builds on previous modeling studies of this data set. The advance is an executive input that controls the behavior of the network, and reduces the burden of tuning compared to previous models.
| In order to survive, animals must be able to receive sensory stimuli and hold this information in memory after the stimulus has ceased. The ability to recall sensory information allows the animal to process information and make decisions, such as fight or flight. Certain areas of the brain are known to play a role in the ability to hold sensory information, but precisely how the information is held is not completely known. This type of memory, where the information from a transient stimulus is stored for a short period of time, for use in a task or recall in a decision making process, is referred to as working memory.
In order to probe for the neuronal basis of working memory, recordings of cellular activity are made during delayed response tasks. In these tasks, an initial stimulus or cue is given to an animal, and then removed. The relevant cue information is held in memory for the duration of a delay period. At the conclusion of the delay period the animal is asked to demonstrate memory of the stimulus. This is generally done with a motor response (button push, bar grab/release, eye saccade, etc.). In most delay period studies, the cellular responses during the delay period vary widely, both from cell to cell, and even within one cell across trials. Much attention has been given to this variability [1]–[6].
The work in this paper is motivated by the experimental work done in one such study [1], [2]. In these classic experiments, an animal was presented with a tactile stimulus, a vibration briefly applied to a finger. After a delay (3 seconds), the animal is presented with a second stimulus. The animals' task is to correctly signal which of the two frequencies was higher. Thus, for successful completion of the task, the animal is required to hold the frequency of the first vibration (the analog stimulus variable) in memory for the duration of the delay period. Consequently, recordings made during the delay period provide clues about the mechanisms responsible for storing this stimulus variable.
The neuronal correlate of working memory is presumed to be persistent cellular activity [7]–[9]: meaningful neuronal activity that continues after the causal stimulus is removed. Often, the level of persistent activity - the firing rate of the relevant cells - depends on the stimulus itself, and therefore can encode information about the stimulus identity. One example, and a focus of this paper, is the case of monotonic encoding. This type of encoding refers to a scenario in which the level of cellular activity depends monotonically on an analog stimulus variable, such as the frequency of a tactile vibration.
The recordings [1], [2] show that there are cells in the frontal cortex that have a monotonic relationship with the stimulus frequency. A structure commonly used to model this type of relationship with a continuous variable is the line attractor [10]. However, perfect line attractors are unlikely to exist in nature as they require exact tuning. Moreover, even if a perfectly tuned line attractor was possible, they cannot stably hold information since they are only neutrally stable along the axis of the attractor, allowing for corruption by noise [11].
A number of features found in the data further suggest that a true line attractor is not the correct type of attractor. There is a lot of variability in the data, both from cell to cell, and at the single cell level. The first type of variability is a large diversity of behaviors among cells. The authors [1] divide cells into three classes - early, persistent, and late. The classification refers to when, during the delay period, the cell is monotonically tuned to the stimulus variable. Early cells are those cells that encode the stimulus during the first part of the delay period, but then lose tuning with the stimulus. Late cells do not begin the delay period tuned to the stimulus, but they are activated and are monotonically tuned to the stimulus at the end of the delay period. The persistent cells are monotonically tuned to the stimulus variable for the entire delay period.
This division in behavior results in a second feature of the data - a systematic change in the number of cells that encode the stimulus at any given time. At the start of the delay period, both early cells and persistent cells are tuned to the stimulus. As the early cells fall out of tune the number of encoding cells decreases, until only persistent cells represent the stimulus variable. As the late cells become tuned to the stimulus, and the persistent cells remain tuned, the total number of encoding cells grows. These changes generate the U-shaped curve describing the number of cells encoding the stimulus as a function of time [1].
The third important feature of the data is the variability of the firing rates during the delay period. It is evident from the experiments that the persistence is not a fixed-point type of persistence, where the cell assumes an invariant firing rate. In these persistent cells, the level of persistent activity is not generally constant for the duration of the delay period. Rather, persistent cells demonstrate changes in their activity level during the course of the delay period. Still, these cells maintain a monotonic relationship with the stimulus variable.
I will show how these observations suggest two things: First, the changes in the number of encoding cells, and the division into early, persistent and late, suggest that the neural representation of the stimulus variable is held as a wave. Second, the variations during the delay period are the result of a poorly tuned line attractor combined with a time-aware correction mechanism to account for the imperfections.
I aim to describe how a network of cells that are not tuned well enough to act as a line attractor can still hold a signal for the duration of a delay period. The key ingredient is a time aware task input [12] that allows the network to amplify its activity, correcting for the deviation of the tuning from that of a line attractor. I also show how this network of cells can be tuned to delay periods of different length by changing a single parameter - the timing of the task input - rather than by manipulating the cell-to-cell connectivity. This task input is assumed to be an executive input originating from outside of the network.
My goal is to describe a mechanism that can account for the cellular activity recorded during the tactile discrimination task. The key aspects of the mechanism are traveling wavefronts. In choosing an illustrative model, I only require that the model admit traveling wave solutions. I use a biologically motivated model of propagation between cells [9], [13]. This model admits traveling wave solutions, but is otherwise generic:(1)where is the activity variable, is the noise for the th node (modeled as a Wiener process), and is the matrix of connection strengths (The element is the connection strength from node to node ). The function is a response function, converting presynaptic activity to postsynaptic input. This function is very important for the mechanism that I propose, and is discussed in detail below (section and the task manipulations).
I assume feed forward, nearest neighbor connectivity, ie.(2)This type of connectivity is chosen for its simplicity, as well as the existence of traveling wave fronts. The results that we obtain here generalize to more complicated types of connectivity. Later, I will demonstrate the mechanism for a model where the connectivity matrix, , is completely symmetrical.
The function is central to the mechanism. Manipulation of this function is how external events (stimulus information, for example) influence the network. In this section, the different roles this function fills are outlined.
There are four different configurations of . Three of these parallel the “loading”, “maintenance”, and “comparison” task components described by Machens et al. [7]. The fourth role of the function is to prevent a cell from responding to input (“quiescent”). Figure 1 shows the four different configurations of : a stable equilibrium point (Row B, left), a line attractor (Row B, center), an unstable equilibrium point (Row B, right), and a quiescent mode (Row D, right). Each of these configurations corresponds to a subtask of the working memory task. The first three are implemented as Machens does [7]: The stable fixed point is for loading the stimulus variable into the network. The maintenance configuration is a line attractor. The comparison configuration is an unstable fixed point.
Chronologically, the first component of the task is “loading” the stimulus variable into the network. During this component of the task, the network is exposed to a stimulus that is described by an analog scalar variable. Here, this variable is the frequency of the tactile vibration. Since the cells tune monotonically to the stimulus variable, the cellular response to the stimulus will be a monotonic function of the stimulus variable. In terms of , this is done by creating a stable fixed point at the desired activity level (left panel of item C in figure 1). The activity level at this fixed point is determined by the slope of . Thus, the slope of is a monotonic function of input frequency. So, for a cell that is positively monotonically encoding the stimulus, the slope of is a monotonically increasing function of frequency. The stable fixed point draws the activity toward this frequency specific level, and the stimulus variable is loaded into the network.
The second component of the task is “maintenance”. This begins once the stimulus is removed. This is the actual memory component of the task, where the network contains information about a stimulus that is no longer present. The information is to be retained for the duration of the delay period. In terms of , this configuration is shown in the center panel of item C in figure 1. In this configuration, the network behaves as a line attractor. The line attractor holds the stimulus dependent values throughout the delay period. The important result of this study is how the brain might overcome the drift that occurs when this configuration is not perfect - when the lines do not perfectly overlap.
The final component of the task is “comparison”. This component occurs at the end of the delay period. In the experiments, a second stimulus of frequency arrives and the objective for the animal is to compare this stimulus to the original stimulus (frequency ). In terms of , this configuration is shown in the rightmost panel of item C in figure 1. There is an -dependent unstable fixed point, where the slope of is determined by the stimulus frequency. This unstable fixed point acts as a separatix. If the -dependent activity levels are above this separatix, they will increase. Conversely, if the -dependent levels are below this separatix, they will quickly decrease. So, whether the activity level increases or decreases upon the arrival of the second stimulus determines whether the network assessed the first stimulus frequency () to be higher or lower than the second (), thus providing a comparison.
The right panel of item D in figure 1 shows the “quiescent” configuration. With this , the cells do not respond meaningfully to input. All activity will quickly decrease to zero. The usefulness of this configuration is described in the next section.
Earlier, I described three characteristics of the data recorded during a delay response task [1], [2]. These features suggest a mechanism that the brain can employ to store an analog stimulus variable for the duration of a delay period. These features are:
In this section, I show how each of these features shape the proposed memory mechanism. I divide the section into three parts, as itemized above. Though done sequentially, it will become apparent that, in my interpretation, these features are tightly intertwined. Once the model is built, I will show how this model is tuned, and why this provides a substantial advantage in feasibility, with regard to tuning, over a regular line attractor.
The keystone of the mechanism is an input that originates externally to the population of cells that we focus on. This input is an executive one, and provides an interpretation of time to the network. The assumption that the timing originates externally is supported by the data. Machens et al. [14] use principal component analysis to show that the cellular responses to the stimulus can be divided into two groups - those components whose variance is due to the stimulus frequency and those components whose variance is due to time. The authors show that the variance due to time is external to the network. The timing of this task input determines the behavior of the network and can be used to tune the network without changing any of the intrinsic properties of the network (eg. connection strength between cells).
I begin by demonstrating how a traveling wave can account for the division of the population into early, persistent, and late cells. Figure 2 shows a solution to equations (1) (2). The array plot shows the solution as a function of time and position in the chain. Also shown are temporal profiles of three cells - one that loses activation quickly (early), one that is active throughout the delay period (persistent), and a cell that does not tune to the stimulus variable until the end of the delay period (late).
Figure 2 demonstrates how a traveling pulse can generate a pattern that would allow cells to be classified as early, persistent, or late. There are actually a pair of wavefronts - a leading wave front and a trailing wavefront. The leading wavefront tunes cells to the stimulus variable. Cells lose their relationship with the stimulus as the trailing wavefront passes. By definition, the arrival of the stimulus tunes the early and persistent cells to the stimulus variable. What separates early cells from persistent cells is the position in the chain. Early cells are first in the chain, and the trailing wavefront passes through these cells early in the delay period. Persistent cells are further along in the chain, and the trailing front does not reach these cells during the delay period. Late cells, in this illustrative scenario, are later in the chain than the persistent cells. They are not tuned to the stimulus initially, rather they only become tuned as the leading wavefront reaches them.
Of importance is that the leading and trailing wavefronts have the same slope in the array plot of figure 2. Thus, these wavefronts have the same speed causing the number of cells encoding the stimulus to be constant throughout the delay period. This is not what the experiments show. Rather, the number of cells encoding the stimulus variable decreases at first and then, at some point near the middle of the delay period, begins to increase. The simple pulse described so far is not capable of this. In the next section, I discuss a modulation of the leading wavefront that can account for the initial decrease in the number of encoding cells.
The second feature of the data that was identified was a systematic decrease in the number of encoding cells during the first half of the delay period, followed by an increase. In the previous section, I showed how a pair of traveling wavefronts can account for the existence of early, persistent, and late cells. The problem remaining at the end of the section was that the trailing and advancing wavefronts have the same speed, and so the number of encoding cells is constant. In this section I describe how the wavefronts can be modulated during the delay period to account for this characteristic.
By definition, the initial decrease in the number of encoding cells is due to the early cells losing their monotonic relationship with the stimulus. Similarly, the subsequent increase can only be due to the late cells assuming a stimulus dependent activity level. Thus, the transition from decreasing to increasing number of encoding cells is equivalent to the transition from early cell decay to late cell activation. An important result from the experiments [1], [2] is that the transition from decreasing to increasing occurs roughly halfway through the delay period. This is regardless of the length of the delay period. To illustrate, the authors show what happens when the length of the delay period is changed from 3 seconds to 6 seconds. They show that the late cell response, which began roughly halfway through the delay period for the 3 second delay period, is stretched to roughly halfway through the 6 second delay period after a couple of trials. So, a change in the length of the delay period modulates the time that the transition from decreasing number of cells to an increasing number of encoding cells occurs.
I suggest that the external executive input is responsible for the transition in the number of encoding cells. Suppose that, prior to the arrival of this input, the late cells are not allowed to tune the stimulus variable. The effect of the executive input, then, is to allow the late cells to participate in the task. Prior to the arrival of the input, the leading wavefront is frozen. The early cells are simultaneously falling out of tune, and so the net result is a decreasing number of encoding cells, prior to the arrival of the input.
To incorporate these changes into the model (1), I need to specify for the late cells, separately from the early and persistent cells. I do this by designating the late cells as “quiescent”, (Figure 1), prior to the arrival of the executive input. When the input arrives, its action is to shift the configuration from “quiescent” to the “maintenance” configuration (Figure 1). This switch allows the leading wavefront to advance into the late cells, tuning them to the stimulus variable.
Figure 3 shows the network with the late cells modulated as above. Notice the frozen leading wavefront for the first half of the trial. Prior to the inclusion of the late cells, The number of encoding cells decreases. When the activity is allowed to propagate into the late cells, they become tuned to the stimulus. A difficulty also arises: the persistent cells are losing their monotonic relationship with the stimulus and so there is no net gain in the number of encoding cells. The experiments clearly show an increase in the number of encoding cells. To account for this, either the advancing front (the front going into the late cells) must be faster that the trailing front, or the trailing front (the front that causes the early cells to fall out of tune) must slow down.
Simply increasing the speed of propagation in the late population, thus speeding up the advancing front, would accomplish the growth in the number of encoding cells, but the data suggests that this is not the case. Persistent cells are also impacted by the arrival of the executive input. Many persistent cells show a dramatic change in behavior simultaneously with the incorporation of the late cells. It is then natural to suspect that the executive input is involved with this change of behavior. I posit that the late cells project back onto the persistent cells, freezing the trailing wavefront, and allowing the number of cells that encode the stimulus to increase. The change in the behavior of persistent cells is the topic of the next section.
Persistence, as a mechanism, is a staple of working memory. It is how cells can hold information about the past - a stimulus that is no longer present. Persistence is often modeled as a fixed point, or for the case studied here - monotonic encoding - a line attractor. Each of these attractors holds a cellular variable (eg. firing rate) constant for the duration of the delay period. However, the persistent cells recorded in many working memory studies do not behave as a fixed point. Rather, the firing rates of persistent cells vary widely during the delay period. For the experiments that are the focus of this paper, the large variation of the persistent cells is divided into four categories - cells that initially decrease and then increase, cells that decrease for the duration of the delay period, and the opposite behaviors. Here, I only consider those persistent cells that initially decrease, and after the executive input, are amplified (a typical example is shown in Figure 2 of [1]).
A line attractor is often used to store an analog variable [7], [10], but this mechanism requires very precise tuning. However, a network that admits a near-line attractor (a line attractor where the tuning is not perfect) is still capable of maintaining a monotonic relationship with the stimulus variable. That is, for two stimuli , and firing rates , , it may be possible to tune the network well enough so that for all in the delay period, even if is not constant. This is possible if the tuning is close to that of a line attractor, but not perfect. The imperfections will result in a slow drift from the original value, as shown in [10]. If the tuning is good enough to make this drift sufficiently slow, then a monotonic relationship between the stimulus variable and the cellular output over the course of the delay period can be achieved.
In the model, the cell-to-cell connections are determined by the entries of the connectivity matrix, (equation (1)). We are assuming nearest neighbor, feed forward connectivity, so all non-zero entries reside on the first sub-diagonal (equation (2)). For a perfectly tuned line attractor, each of these entries are . The slow drift is modeled by allowing these entries to be less than (the slow drift will be a decreasing one).
The change in connection strengths will cause the amplitude of the wave to decrease slowly toward zero. If the entries of are close enough to , the exponential decay will be slow and monotonicity will be preserved. (Note: If the initial stimuli are very close in scalar value, and there is sufficient noise, this monotonicity can be broken. For example, if in one trial the stimulus is at Hz and in another trial the stimulus is Hz, for a small value of , one would expect the noise to destroy the monotonicity. Accordingly, there is a minimum separation between the frequency of the first and second stimuli in the experiments).
In order to reflect the stimulus information, the signal must be amplified to recover from the initial decay. The switch from decreasing activity to increasing activity takes place at nearly the same time the late cells begin to encode the stimulus, so it is natural to view the late cells as implicit in this transition. I propose that the late cells project back onto the early and persistent cells. This feedback accomplishes two things: 1). It amplifies the decayed signal so that the firing rate at the end of the delay period is indicative of that at the beginning - the stimulus induced value - and 2). it stops the trailing wave front from advancing, allowing the total number of cells encoding the stimulus to increase. This addresses the issue that I ended the previous section with - the leading and trailing wavefronts no longer have the same speed.
In order for the feedback from the late cells to amplify the signal, the strength of this feedback must be sufficiently strong. As in [9], the solution of the th cell in the chain, after the stimulus has been removed, is given by(3)Following Goldman [9], the late cells become part of the wave roughly one time unit for each connection in the chain after the initial blockade (an implicit assumption is that the time constant of a cell is much shorter than the length of the delay period). Once the propagation of activity is allowed to advance into the late cells, they assume this solution as well. So, in general for , the onset time of the task input. After the inclusion of the late cells, we can approximate the evolution of a persistent cell with(4)where is the strength of input from other persistent cells (the value of ) and is the strength of the feedback from the late cells. If , there will be amplification. I will derive the specific restraints on these parameters in the next subsection, and their relationship to the tolerance in the timing of the executive input.
It is my claim that the decay and amplification mechanisms can greatly ease the cell-to-cell connectivity restrictions of a line attractor. There is another important advantage to the proposed mechanism - it can be used to tune the network to delay periods of different lengths without changing any of the individual connections between cells. Figure 4 shows a demonstration of this. In the top panel of figure 4, the network is tuned so that it successfully stores the stimulus value for a delay period of length . If, on the next trial, the delay period is increased without warning or preparation, say doubled (), the task input will not move, resulting in an unreliable cellular response. However, after a couple of trials, the timing input is shifted to a time that results in correct trials (bottom panel of figure 4). This is consistent with the data; as reported in [1] there is a slight increase in the error rate directly after the switch from 3 to 6 seconds. After a few trials, the animal's performance improves. Moreover, raster plots show that the onset of late-cell activity gradually adapts to the longer delay period length, after a few trials [1].
This mechanism is further supported by the data. Brody et al. [1] utilize a descriptive model to determine to what extent the activity during the 6 second delay period is a stretched version of the activity during a 3 second delay period. They show that the early cells behave in much the same way for each delay period length. That is, the time course for an early cell does not stretch or contract. The model agrees with this, since an early cell for the 6 second delay period will evolve the same way as it would for a 3 second delay period. In either case, the trailing wavefront has passed. The authors show that the timing of the late cells is stretched by a factor of 2. This also agrees with the model, since the timing of the task input - which determines when the late cells become active - is roughly halfway through the delay period. Doubling the delay period will then roughly double the time at which the late cells begin to encode, yielding an approximate stretch factor of 2.
At the end of the previous section, I concluded that the external signal also causes the freezing of the trailing wavefront, so that persistent cells can remain active. How the trailing wavefront is frozen is interesting, and may not be immediately obvious. The late cells feed back onto all of the early and persistent cells. The early cells are those cells that, by definition, have lost their monotonic relationship with the stimulus. In other words, they have decayed below the level where signal can be differentiated from noise.
Still, with the addition of late cell input, the noise in the early cells' activity is amplified and propagates through the medium. Though it is summed noise and does not have a relationship to the stimulus, it does generate enough activity to “prop up” the trailing wavefront. This allows the persistent cells to maintain their monotonicity with the stimulus variable, freezing the trailing wavefront. Figure 5 shows early cells at different locations in the chain, and how they contribute to the maintenance of a persistent cell.
Figure 6 shows a simulation of the full network. I implement the feedback from the late cells to the early and persistent cells as a connection from a single late cell, though more general patterns would work as well. As the late cells are incorporated into the network, the persistent cells begin to increase their firing rates. The time courses for an early, persistent, and late cell are also shown in figure 6. The decreasing and increasing of the firing rate in a persistent cell is clearly demonstrated, as is the match between the initial stimulus dependent activity level and the level after amplification. Figure 7 shows how this this tuning works for a range of stimuli.
In this section, I show that the proposed mechanism offers a substantial advantage over a true line attractor in terms of the tuning requirements for the connections between cells; i.e. the closeness to the line attractor configuration in figure 1.
The first step is to determine, as a function of the decay and amplification rates, an interval during which the executive input must arrive to correctly amplify the cellular activity. The goal is to determine the length of this interval as a function of the decay and amplification rates ( and in equation (4)). I show that, for decay (and amplification) rates well outside those acceptable for a true line attractor, this interval of times is within the abilities of networks in the brain.
To quantify the tuning requirements in the context of our model, I first establish how accurate the network needs to be. I define as the resolution of the network. That is, the network has to be able to differentiate between stimulus variable values that are separated by more than . Frequencies closer than this are assumed to be too close to differentiate. Therefore, it is necessary to determine when the task input must arrive so thatwhere is the length of the delay period. Assume a decay rate of , and an amplification rate of , where is the coupling strength and is the strength of the feedback connections, as in equation (4).
First, we derive the requirements on the coupling strengths that are necessary so that a line attractor can hold the value of the stimulus variable for the duration of the delay period. Following Goldman [9], the -th cell in a chain of cells satisfying equation (1) has the solutionFor large , this can be approximated by(5)For the regular line attractor, we require thatSolving for yields(6)
For the proposed mechanism, the tuning requirement is on the timing of the external task input. To determine this requirement, we determine the allowable values of the task timing () given the rate of decay, and the amplification rate due to feedback from late cells. With decay followed by amplification, equation (5) can be extended to yield(7)Where specifies the timing of the executive input. We require . Solving for gives(8)The length of this interval iswhere . From this inequality, one can see that the length of the interval scales with the sum . In other words, doubling the sum will decrease the length of the interval by a factor of 2.
Now, to show that this is advantageous over tuning a line attractor, we consider how the bound on varies as the decay rate increases past the limit allowed by a line attractor. Inserting the bound (6) into the expression for the length of the interval, and letting givesThis means that for any choice of and , the length of the timing interval will be, approximately,(9)Thus, this mechanism is feasible for values of and well outside of values that will yield an effective line attractor. As an example, if and , then the interval where the external signal can arrive has length . As another example, if and then the interval will have length . For a 3 second delay period, these examples give interval lengths of ms and ms, respectively. Based on measured dynamics of neural operations in the brain, these intervals are within the limits of feasibility.
In addition to the length of the interval, equation (8) gives where, during the delay period this interval resides. The interval is centered atThus, if the amplified network acts as a line attractor, and so the timing input should arrive right away. Accordingly, the interval hugs the left endpoint of the delay period. On the other hand, if , then the network is a line attractor during decay, and the executive input never needs to arrive. Thus, if either the decay or the amplification meet the requirements of a line attractor, then the network is, by default, accurate enough.
In this section, I demonstrate the performance of the model, in terms of accuracy (correct or not) and how this varies with the noise level and the decay and amplification parameters. I also show how the decay-amplify model integrates noise, and compare to a line attractor.
First, I determine how often the mechanism results in a correct response at the end of the delay period for different executive input times and different levels of noise. In all of these numerical experiments, the stimulus variable is , the delay period begins at time and ends at time . My criterion for success is that the activity at the end of the delay period () lies between and (so that in equations (6) (8) is equal to ).
In figure 8, I show the accuracy (as a percentage) of the model for three decay-amplify sets - . For a delay period of length , inequality (6) gives the bound on acceptable values for a line attractor:This corresponds to a coupling strength of (). Expression (9) predicts an approximate interval width of for , for , and for . These are, respectively, , and of the delay period lengths. Figure 8 clearly shows that the correctness rate lies at values that are consistent with this calculation. At the correctness rate, the mean (which is the trajectory discussed in the previous section) is right on the boundary. The noise causes a distribution of values for the activity level at the end point. If the network is tuned properly, the width of this distribution (variance) determines how reliable the recall will be. It is necessary to show that the decay-amplify model does not suffer from a noise integration disadvantage over a line attractor, otherwise the tuning advantageous would be nullified.
To determine the relationship between the cellular noise strength (the variance of the Wiener process) and the variance of the output, I ran the simulation 200 times, recording the activity at the end of the delay period for a single cell - the last persistent cell in the chain (cell #100).I repeat this for three instances of the model: A line attractor () and two decay-amplify models with and . I simulate each of these with four different noise strengths (). The results are shown in figure 9. The figure shows that the decay-amplify mechanism integrates noise no worse (or better) than a line attractor. This is not surprising. During the decay phase, the effect of noise is reduced. Conversely, during amplification the effect of noise is also amplified. The net effect of having a decay followed by an amplification of noise results in roughly the same distribution as a line attractor, where the effect of the noise uniform throughout the delay period.
I have demonstrated the decay and amplify mechanism using a simple feed forward model of neuronal activity. The choice of coupling was made to simplify the calculations and make the behavior of the model as transparent as possible. In this section, I describe a more general connectivity, where there is no bias, to demonstrate the decay-amplify mechanism is not dependent on a specific type of network architecture.
In place of the feed forward chain, I model the cells as(10)where is the activity, is location, is the coupling strength, and is a connectivity kernel. We assume that (symmetric), , and (on an unbounded domain, I use to scale for a bounded domain). For the simulations, I useso that . If then if the networks admits a line attractor. Because we are considering a bounded domain, the valuecorresponds to the maintenance configuration in figure 1, where cells on the interior behave as a line attractor until the arrival of the wavefront. For simulations, the integral is discretized and 300 early/persistent cells are used.
There are countless ways to implement the late cells. I choose to implement the late cells as a another line of cells, coupled to the activity described by (10) according to(11)and implement the feedback as one to one by rewriting equation (10) as(12)where is the strength of the feedback from the late cells. Here, I have implemented the feedback from the late cells as a one to one relationship, and the connections from the early and persistent cells are divergent. Figure 10 shows the network schematic of this configuration.
Simulations of the network for a range of stimulus variable values are shown in figure 11. Shown are example time series for an early cell, a persistent cell, and a late cell, for a range of values of the stimulus variable. This figure demonstrates the decay-amplify model for a network that is not a feed-forward chain.
I demonstrate a mechanism that allows a network of cells to store an analog stimulus variable for a delay period, greatly easing the tuning requirements that would be necessary to accomplish the same feat with a line attractor. Using a simple mathematical representation of cellular activity, I demonstrate how wave fronts can account for the different types of activity observed in the experiments [1], [2] (early, persistent, and late), the systematic change in the number of cells encoding the stimulus, as well as the in-trial variability of persistent cells. The keystone of the mechanism is an external signal that is executive in nature and provides timing to the network.
I show that the proposed mechanism eases the requirements on cell-to-cell connections by initially allowing the cellular activity to decrease. A subsequent amplification, initiated by the executive input, corrects for the decay. I show that the restriction on the timing of the executive input depends on the decay and amplification rates in a way that is feasible for networks in the brain, even when the decay and amplification rates are well outside those allowable for a line attractor. The memory mechanism that I describe has an additional advantage: It allows the network to quickly adapt to delay periods of different lengths.
The tuning strategy involves changing the time when a wave front is allowed to propagate into a previously response-less group of cells, late cells, which do not encode the stimulus until the latter half of the delay period. This strategy agrees with the data shown in [1], where upon lengthening the delay period from 3 seconds to 6 seconds, the activation of late cells is pushed back. Importantly, this transition is not accomplished in one step, but rather over a series of steps. Consistent with this gradual transition is an increase in the error rate for a few trials after the change of delay period length takes place [1]. The mechanism that I describe suggests that there should be an increase in the error rate upon changing the length from 6 to 3 seconds - a testable hypothesis that is somewhat counter-intuitive.
The analysis that I provide is for a feed forward network. This choice for the connectivity matrix was made so that the behavior of the model is as transparent as possible. I make the claim that this choice of connectivity is not crucial for the mechanism to work, and I demonstrate the mechanism using a recurrent network. In general, a linear filter that slowly decays the signal (the eigenvalues of the network are negative, with a few near 0 for slowness -ie. near a line attractor) will work for the mechanism, since all that is needed is that monotonicity be preserved. The evolution of early and persistent cells depends on the relative decay rates. If the network of linear filters has a range of decay rates which are spatially localized, then the activity level of some cells will decay quickly, while the activity level of others will decay slowly. Those that decay quickly are candidates to be early cells. Those that decay slowly maintain a monotonic relationship with the stimulus and are more likely to be persistent cells. Any such network that is not normal can be viewed as a feed forward network under an appropriate change of basis [9]. If the network is normal, then the behavior can be viewed as independent modes (the eigenvectors) and will not be, strictly speaking, feed forward. Such a network still is capable of various decay rates, and so can admit early and persistent cells just as a feed forward network can.
I treat late cells as a distinct group, defined by the inability to respond to input until an executive input is received. There are many ways to implement the late cells. In the feed forward model, I attach them to the end of the chain. In the symmetric model, I treat them as a separate line of cells that are reciprocally connected with the early and persistent cells. In either case, once allowed to participate in the task, they assume the activity of the persistent cells, and amplify the network. Any late cell configuration that does this will work, there are no other requirements on the late cells.
A major claim of this paper is that the late cells are governed by a timing input, executive in nature. Prior to the arrival of the executive signal, these cells do not respond to input. There are many plausible mechanisms for this. The most obvious to me is a shunt, as described by Torre and Poggio [15]. If ion channels that have a reversal potential near the membrane resting potential and are held open, the impact of other channel openings (eg. sodium) will be greatly reduced. So, in this scenario, the action of the executive input would be to remove this shunt by allowing the responsible channels to close. Another possibility is inhibition. Inhibitory control is known to be an important element of prefrontal function [12]. A constant inhibitory drive onto the late cells would hold them quiescent. Removal of this inhibition would serve the purpose of the executive input.
There have been numerous modeling studies of the delayed discrimination experiments. When comparing and contrasting the proposed decay-amplify mechanism with these models, I focus on the three most important features: 1). The decay-amplify model is an extension of a line attractor. 2). The proposed model accounts for the division into early, persistent and late cells using a wave front.3). An external executive input is used as a timing mechanism for the model. This external input is independent of the stimulus variable.
A line attractor is a natural choice to store the value of an analog variable. Other authors have used a line attractor to model the Romo data [7], [16]. In [7], Machens et al. use the interplay between cells that respond to the stimulus in different ways (monotonically increasing relationship versus a decreasing relationship with the stimulus variable) and inhibition to form the attractor. Singh and Eliasmith [17] use a “neural integrator”, which is similar to a line attractor in that the output of the system does not change without a change in the input, and that the connection strengths between cells needs to be precise. The decay-amplify model is a novel extension of these models. Rather than require the very tight tuning necessary for a line attractor, I allow the cellular activity level to drift slowly. The value of the stimulus variable is lost, but the activity of the decaying system maintains the monotonic relationship with the stimulus variable.
It is important to note that the line attractor has not been the only proposed means of modeling the data. Barak et al. [16] explore two types of models, in addition to a line attractor. They show that a network with random connectivity can perform the task by using a linear sum over the constituent neurons. They also demonstrate a learning model that begins like the random network (random connections, linear readout) but then adjusts the connection strengths between neurons based on past performance. Each of these models are capable of the performing the task, though none of them account for the late cells. Additionally, a change in the length of the delay period would require a complete recalculation of the weights applied in the linear readout scheme, rather than changing a single parameter.
There have been studies that use completely different strategies to store the stimulus variable. Miller et al. [18] tune a model so that it approximates a line attractor near a degeneracy. The attractor holds the memory by holding the activity level constant for the duration of the delay period. Miller includes inhibition and cells that encode the stimulus variable both positively and negatively. The attractor is formed through the interplay of these different cells. Another strategy that has been described is to store the stimulus as a level of facilitation in the cells [19], [20]. The initial stimulus facilitates the synaptic connections between cells, and these facilitated cells later respond to a recall signal. The facilitation decays slowly, so that the memory is stored at the synaptic level. Neither of these models attempt to describe the diversity of the cellular responses, in particular the division into early, persistent and late cells.
Barak et al. [20] show that the stimulus information is held in a dynamic way. They quantify a population state for the recorded cells by trends in how the cells are tuned to the stimulus. They look at two representations of the population state, sensory and memory. The sensory representation of the stimulus is the population state at the beginning of the delay period. The memory representation is the population state at the end of the delay period. The authors show that applying the sensory representation to cells at the end of the delay period, or applying the memory representation to the cells at the beginning of the delay period, provides no stimulus information. Thus, they demonstrate that the stimulus information is held in a dynamic way, by different cells at different times. Those neurons that are classified as sensory correspond to the early cells in the model I propose. At the beginning of the delay period, they are tightly tuned with the stimulus. At the end of the delay period, they are devoid of information. Similarly, the late cells begin with no stimulus information, but gain it later in the delay period. The authors note that the classification of sensory or memory is only weakly correlated with the classification of early or late. The model that I propose is too simple to account for all of the data, but the wave front hypothesis for diversity of responses neatly accounts for the transition from a sensory representation to a memory one, as described.
Singh and Eliasmith [17] offer an alternative mechanism to account for the diversity. They build a network of cells, each having a preferred orientation to a state space variable. As the state space variable evolves, it passes through the tuning curves of the cells. The distribution of preferred orientations yields a diverse array of responses including early, persistent (ramping type, as described in [4]), and late.
The centerpiece of the decay-amplify mechanism is an executive input that adds timing to the network. This is a novel addition to the modeling literature, though the separation of stimulus and time has been explored before. Use of an external executive input for timing purposes is in agreement with other studies that separate the stimulus component of the activity from a time component. Machens [14] shows that there are two separate causes of variance in the data: stimulus and time. They show that the variance attributed to time is likely external in origin. This strongly supports the use of an external executive input to time the network.
Singh and Eliasmith [17] also separate stimulus and time. The state space variable that evolves through the tuning curves has two components, the stimulus and a variable that is akin to elapsed time. They do not implement their timing component as an external signal. Moreover, they model it as a passive process that is ongoing throughout the delay period. In contrast, the executive input that I propose is external, and upon arrival the behavior of the network drastically changes. These drastic network changes can be seen in the data, where there is an obvious change in behavior near the mid point of the delay period [1]. The number of tuned cells begins to increase, and the behavior of individual cells changes. The tuning curves in the Singh and Eliasmith model are monotonic along both axis (stimulus and time), and so the activity pattern that is the focus of this paper, a decay of activity followed by an amplification, is not possible in the Singh-Eliasmith model without an external intervention.
The decay-amplify model is the only model that directly addresses the two-mode behavior that Brody et al. [1] describe - behavior that is different during the first half of the delay period than during the second half. I focus on activity that decreases during the beginning of the delay period and then is amplified to recover the stimulus information. Another type of persistent activity that occurs is ramping, where a cell either increases or decrease for the duration of the delay period while maintaining a monotonic relationship with the stimulus variable [3], [4]. The simple linear filter that I use is capable of generating all of these types of behavior (figure 12), but they cannot coexist for the simple chain of neurons that I describe.
In conclusion, this study suggests another potential means of storing a stimulus variable as a firing rate for the duration of a delay period. This mechanism stands apart from previous models that do not take the variability during the delay period into account. Moreover, this variability is revealed as part of the solution to the memory problem, rather than a confound. The major claim that I make is that there is an external timing signal that causes the network to switch modes. There are many features of the data that I do not take into account (eg. inhibition and different monotonic encodings). These features are almost certain to play a role in the prefrontal calculations, and figuring out how everything works together is an ongoing process.
|
10.1371/journal.pgen.1007694 | Glucocerebrosidase deficiency promotes protein aggregation through dysregulation of extracellular vesicles | Mutations in the glucosylceramidase beta (GBA) gene are strongly associated with neurodegenerative diseases marked by protein aggregation. GBA encodes the lysosomal enzyme glucocerebrosidase, which breaks down glucosylceramide. A common explanation for the link between GBA mutations and protein aggregation is that lysosomal accumulation of glucosylceramide causes impaired autophagy. We tested this hypothesis directly by measuring protein turnover and abundance in Drosophila mutants with deletions in the GBA ortholog Gba1b. Proteomic analyses revealed that known autophagy substrates, which had severely impaired turnover in autophagy-deficient Atg7 mutants, showed little to no overall slowing of turnover or increase in abundance in Gba1b mutants. Likewise, Gba1b mutants did not have the marked impairment of mitochondrial protein turnover seen in mitophagy-deficient parkin mutants. Proteasome activity, microautophagy, and endocytic degradation also appeared unaffected in Gba1b mutants. However, we found striking changes in the turnover and abundance of proteins associated with extracellular vesicles (EVs), which have been proposed as vehicles for the spread of protein aggregates in neurodegenerative disease. These changes were specific to Gba1b mutants and did not represent an acceleration of normal aging. Western blotting of isolated EVs confirmed the increased abundance of EV proteins in Gba1b mutants, and nanoparticle tracking analysis revealed that Gba1b mutants had six times as many EVs as controls. Genetic perturbations of EV production in Gba1b mutants suppressed protein aggregation, demonstrating that the increase in EV abundance contributed to the accumulation of protein aggregates. Together, our findings indicate that glucocerebrosidase deficiency causes pathogenic changes in EV metabolism and may promote the spread of protein aggregates through extracellular vesicles.
| Mutations in the GBA gene, which encodes the enzyme glucocerebrosidase, are common and increase the risk of Parkinson disease. A widely accepted explanation for the increased risk is that the fatty substance normally broken down by glucocerebrosidase builds up in the lysosome, which is the cell’s recycling center, until the cell can no longer get rid of damaged parts. At that point, proteins that should be destroyed in the lysosome form large clumps (aggregates) throughout the cell. We used mutant fruit flies without glucocerebrosidase to test this theory, and we were surprised to see no evidence that the lysosome was failing. The destruction of proteins usually recycled by the lysosome was not slowed down in the mutant flies. Instead, we saw evidence that the mutants’ cells might be producing too many extracellular vesicles, tiny spheres that transport cargo and messages from cell to cell. Some researchers have also suggested that extracellular vesicles carry the protein aggregates that spread between cells as Parkinson disease get worse. Our study supports this idea. It suggests that increased spread of aggregates through extracellular vesicles, rather than failure of the lysosome, might explain why GBA mutations increase the risk of neurodegenerative disease.
| Mutations in the gene encoding the lysosomal enzyme glucocerebrosidase, glucosylceramidase beta (GBA), are associated with neurodegeneration and brain protein aggregation [1, 2]. Homozygous mutations in GBA cause the lysosomal storage disorder Gaucher disease, which in some cases includes devastating neurological symptoms [3], while heterozygous GBA mutations are the strongest risk factor for both Parkinson disease (PD) and the related disorder dementia with Lewy bodies [1, 2, 4]. Up to 10% of individuals with nonfamilial PD carry a GBA mutation [5]. In addition, PD patients with a GBA mutation have faster progression of both motor and cognitive symptoms [6]. To study the mechanisms underlying the association between GBA mutations and neurodegeneration, we created a Drosophila model of glucocerebrosidase (GCase) deficiency. Drosophila has two GBA homologs, designated Gba1a and Gba1b. The Gba1a gene is expressed exclusively in the midgut [7], and deletion of this gene does not appear to confer deleterious phenotypes [8]. By contrast, the Gba1b gene is ubiquitously expressed [7], and Gba1b deletion causes marked abnormalities. We previously reported that Gba1b null mutants exhibit phenotypes including shortened lifespan, locomotor and memory deficits, neurodegeneration, accumulation of the autophagy adaptor Ref(2)P (p62/SQSTM1), and accumulation of ubiquitinated protein aggregates [9]. Similar phenotypes were subsequently seen in an independently generated Gba1b null mutant [8].
The protein aggregation and elevated Ref(2)P levels in Gba1b mutants suggested that they had impaired autophagy, as did morphological changes in the autolysosomal system noted by Kinghorn et al. [8, 9]. These findings are consistent with previous reports of autolysosomal impairment upon loss of GCase activity [1, 10–15]. Based on such findings, we and others hypothesized that lysosomal accumulation of glucosylceramide, the normal substrate of GCase, leads to impairment of autophagy [12, 16–18]. However, none of the work implicating autophagy in the pathogenic effects of GCase deficiency has yet established that GCase loss of function causes global impairment of autophagic degradation.
To investigate the autophagy failure model of GBA pathogenesis, we used proteomics-based techniques to measure protein turnover and abundance in Gba1b mutants and controls, as well as in flies with mutations in key autophagy (Atg7) or mitophagy (parkin) genes [19, 20]. While Atg7 mutants showed marked and widespread slowing of autophagy substrate turnover, Gba1b mutants did not. The effects of Gba1b mutation on the turnover and abundance of autophagy substrates also failed to correlate with those of Atg7 or parkin mutations. Moreover, we detected no deficits in turnover mediated by the proteasome, microautophagy, or endocytosis. However, we found high incidences of faster turnover and increased abundance among proteins associated with extracellular vesicles (EVs), which have been previously suggested as a mechanism for the spread of protein aggregates in neurodegenerative disease. Biochemical studies confirmed increased abundance of EV marker proteins in isolated EVs from Gba1b mutants, and nanoparticle tracking analysis showed that the mutants had markedly increased numbers of EVs. Genetic manipulations to reduce EV production decreased the accumulation of ubiquitinated protein aggregates and Ref(2)P in Gba1b mutants, supporting the model that excessive EV abundance promotes the accumulation of protein aggregates. Together, our findings suggest that the most important pathological consequence of Gba1b loss of function is not failure of autophagic protein degradation but excessive production of extracellular vesicles.
To test the hypothesis that GCase deficiency causes impaired autophagic turnover, we compared protein degradation rates in heads from Gba1b mutants and controls using stable isotope labeling. In brief, our method involves feeding flies a stable heavy isotope of leucine and then using mass spectrometry to monitor the rate at which unlabeled proteins are degraded and replaced with labeled proteins [21]. We measured the influence of Gba1b loss of function on all proteins with data that met quality standards in both Gba1b mutants and controls (1297 proteins for turnover analysis, 4221 for abundance; S1 Data). We analyzed turnover data with Topograph [22], software specifically designed for measurement of protein turnover via stable isotope labeling. We also compared protein abundance in Gba1b and control flies using Skyline [23] and MSstats [24]. Fold change in turnover and fold change in abundance were then calculated for every protein. Fold change for a protein was calculated as the value in Gba1b mutants divided by the value in controls.
We predicted that autophagy substrates would show slower turnover (longer half-lives) in Gba1b mutants, and that they might show increased abundance if synthesis did not decrease to match the slower degradation rate (Fig 1A). We defined autophagy substrates as proteins from mitochondria, cytosolic ribosomes, endoplasmic reticulum (ER), and peroxisomes, all previously identified as targets of autophagy [25–30]. We validated our prediction using turnover and abundance data from autophagy-deficient Atg7 mutants, which we characterized in previous work [21] (S1 Data). We first plotted Atg7 fold change in turnover against fold change in abundance for autophagy substrates to observe the overall pattern of proteostasis changes (Fig 1B). Atg7 mutants showed changes consistent with our prediction: the vast majority of autophagy substrates (72%) had slower turnover (fold change in half-life >1) and increased or unchanged abundance. We therefore used Atg7 mutant data as a reference for the effects of autophagy impairment. When we plotted fold change in turnover against fold change in abundance for Gba1b mutants, the pattern of changes was markedly different; only 15% of autophagy substrate proteins had slower turnover and increased or unchanged abundance (Fig 1C). Proteostasis of autophagy substrates in Gba1b mutants thus did not overall resemble the pattern seen in Atg7 mutants.
To compare in more detail the effects of Gba1b and Atg7 mutations on protein turnover and abundance, we performed several additional analyses, beginning by calculating Gba1b and Atg7 mean fold change in turnover (half-life) for the autophagy substrate proteins mentioned above. Each mutant was compared to its own control. Turnover of proteins from all three classes of autophagy substrates was significantly slowed in Atg7 mutants (p < 0.001 by nested ANOVA), but in Gba1b mutants there was no overall change in the half-lives of ribosomal or ER/peroxisomal proteins (Fig 1D; p = NS by nested ANOVA) and only a very mild slowing of mean mitochondrial protein turnover (mean fold change 1.15 ± 0.32; p = 0.02 by nested ANOVA; Fig 1D).
To test further for evidence of impaired autophagy in Gba1b mutants, we compared the effects of Atg7 and Gba1b mutations on individual proteins. We began with turnover, plotting the fold change in half-life for Gba1b mutants (Gba1b mutant half-life/Gba1b control half-life) against the fold change for Atg7 mutants (Atg7 mutant/Atg7 control). We compared Gba1b and Atg7 effects on individual proteins from each of the three autophagy substrate categories. There was no statistically significant relationship between the effects of Gba1b and those of autophagy ablation for mitochondrial, ribosomal, or ER/peroxisomal proteins (Fig 2A–2C). We also tested for a relationship between Gba1b and Atg7 effects on protein abundance (Fig 2D–2F) and found no significant correlation for any of the three autophagy substrate groups. The effects of Gba1b loss of function on protein turnover and abundance thus do not resemble the effects of autophagy ablation, and we find no evidence that Gba1b mutation causes global impairment of autophagic protein degradation.
One reported consequence of GBA loss of function is accumulation of dysfunctional mitochondria due to defective mitophagy [31, 32]; the slight but statistically significant slowdown of mitochondrial protein turnover in Gba1b mutants therefore raised the possibility of a mild mitophagy deficit. We had previously found a mitochondrial protein turnover deficit in flies with mutations in the mitophagy factor parkin [21], and we now compared the effects of Gba1b mutation on mitochondrial proteostasis with those of parkin. In parkin mutants, turnover was slowed for the vast majority of mitochondrial proteins (Fig 3A). In Gba1b mutants, changes in mitochondrial protein turnover were both milder and less consistent (Fig 3A, S1 Data). We considered the possibility that Gba1b mutants had a mitophagy defect that was obscured by compensatory upregulation of other mitochondrial protein turnover mechanisms, as we previously found in PINK1B9 mutants, which lack a mitophagy factor upstream of Parkin [21]. In PINK1B9 mutants, while the mean fold change in mitochondrial protein half-life was not significantly altered, the effects of PINK1B9 mutation on individual proteins correlated strongly with those of parkin mutation. We therefore tested whether the effect of Gba1b on mitochondrial proteostasis would also correlate with the effect of parkin. However, we detected no significant correlation between Gba1b and parkin effects on mitochondrial protein turnover (Fig 3B) or abundance (Fig 3C). Our findings therefore do not support either globally impaired autophagy or selectively impaired mitophagy in Gba1b mutants.
The lack of evidence for autophagy failure led us to consider alternative explanations for the accumulation of ubiquitin-positive protein aggregates in Gba1b mutants. We first considered the possibility that these aggregates could arise because of reduced proteasome function, which is known to lead to the formation of large ubiquitin-positive protein aggregates called aggresomes [33]. We tested proteasome activity using fluorescent substrates, and found that all three enzyme activities were normal (Fig 4A). We also considered the possibility that delivery of substrates to the proteasome might be impaired [34], and used our proteomic data to determine whether actual proteasome substrates were degraded normally in Gba1b mutants. We identified cytosolic proteasome substrates based on data from Wagner et al. [35] (S2 Data) and compared turnover and abundance changes in this group of proteins to the changes in all other cytosolic proteins. If proteasomal degradation were impaired, we would expect substrates of the process to have slowed turnover and possibly increased abundance, as we predicted for autophagy impairment. In fact, however, the percentage of proteins with slowed turnover in Gba1b mutants was significantly lower for proteasome substrates than for other cytosolic proteins, and proteasome substrates did not show a greater incidence of increased abundance in Gba1b mutants (Fig 4B). Together, these results indicate that proteasome dysfunction does not underlie the accumulation of ubiquitin-positive aggregates in Gba1b mutants.
We next examined whether the protein aggregation in Gba1b mutants could be the result of altered endosomal functioning. As Gba1b mutants have markedly increased levels of glucosylceramide (S1 Fig, [8]) and moderately decreased levels of ceramide (S1 Fig), abnormal membrane composition could compromise functioning of the endosomal system [36, 37]. We therefore tested for impairment of endosomal microautophagy, an Hsc70-4–dependent process that degrades cytosolic proteins with specific targeting sequences (“KFERQ-like motifs”) [38]. To test whether microautophagy is impaired in Gba1b mutants, we searched the Drosophila proteome for proteins with KFERQ-like motifs, and compared the effects of Gba1b on cytosolic proteins with and without such motifs. Compared to proteins without KFERQ-like motifs, proteins with one or more KFERQ-like motifs did not have an increased incidence of proteins with slower turnover or increased abundance (Fig 4C, S2 Data). We also tested whether overexpression of Hsc70-4, which has been shown to increase microautophagy in Drosophila [39], would influence the accumulation of insoluble ubiquitinated protein. However, this manipulation had no effect on the abundance of ubiquitinated protein aggregates (Fig 4D). We thus found no evidence that impaired endosomal microautophagy is responsible for the accumulation of ubiquitinated protein aggregates in Gba1b mutants.
We then investigated whether Gba1b mutations impaired the functioning of another endosomal degradation pathway, endocytic turnover. Using FlyBase [40] and other annotation resources, we identified typical substrates of this pathway, primarily integral cell membrane proteins (n = 90 in turnover data, 437 in abundance data; S2 Data). We also identified a separate group of “endosomal machinery” proteins, which reside in endosomes or are required for endocytosis (n = 32 in turnover data, 102 in abundance data; S2 Data). We found no evidence that degradation of endocytic turnover substrates was compromised; compared to all other proteins, endocytic turnover substrates did not have a higher frequency of significantly slowed turnover or increased abundance (Fig 5A). When we examined endosomal machinery, however, we found a higher prevalence of proteins with increased abundance (p < 0.0001 vs. all other proteins by Fisher exact test; Fig 5B). Thus, Gba1b mutants had no evidence of compromised endocytic turnover, but proteostasis of the endosomal machinery was clearly altered.
Many endosomal machinery proteins also play roles in the creation and release of extracellular vesicles (EVs), a heterogeneous population of membrane-delimited structures originating from the multivesicular endosome and plasma membrane [41, 42]. EVs transport varied cargoes of protein and nucleic acids from cell to cell and play roles in signaling, waste disposal, and intercellular resource transfer [43–45]. EVs have also been implicated in the spread of protein aggregates in neurodegenerative disease [41]. Given that Gba1b mutants have altered turnover and abundance of endosomal machinery proteins but not endocytic turnover substrates, we considered the alternative possibility that GCase deficiency influences EV biology. To explore this hypothesis, we first tested whether proteins known to be associated with EVs showed significant alterations in turnover or abundance in Gba1b mutants. We compiled a list of proteins detected in EVs from Drosophila cultured cells [46–49]; the resulting list contained 544 nonredundant proteins (S3 Data), 329 of which were found in the Gba1b turnover data and 499 in the abundance data. Compared to all other proteins in the dataset, a smaller percentage of EV-associated proteins had slowed turnover, and a higher percentage had faster-than-normal turnover (p < 0.0001 by Fisher exact test; Fig 5C). In addition, a greater proportion of EV proteins had increased abundance in Gba1b mutants (p < 0.0001 by Fisher exact test; Fig 5C). To confirm that EV-associated proteins had faster turnover and increased abundance in Gba1b mutants, we repeated our analysis using an independent list of EV proteins. We obtained the ExoCarta [47] “top 100” list of proteins most frequently identified in mammalian EVs and identified their Drosophila orthologs using DIOPT v6.0 [50] (n = 97; S3 Data). Once again, compared to the rest of the dataset, EV-associated proteins had higher frequencies of faster turnover and increased abundance in Gba1b mutants (Fig 5D), suggesting that GCase deficiency may cause dysregulation of EV biology.
To test whether faster turnover and increased abundance of EV-associated proteins are specifically associated with Gba1b loss of function, we investigated whether these proteins were also disproportionately affected by other conditions that alter protein turnover. We evaluated the pattern of changes, as we had done for autophagy substrates, by plotting fold change in turnover against fold change in abundance for all EV-associated proteins. In Gba1b mutants, 59% of the datapoints representing EV proteins appeared in the quadrant representing faster turnover and increased abundance (Fig 6A); in Atg7 mutants, only 3% of EV-associated proteins showed the same pattern (Fig 6B). We also looked at the pattern of EV proteostasis in other mutants described in our previous work [21]: the mitophagy mutants parkin and PINK1, and the oxidative stress mutant Sod2. Because abundance data for these mutants lacked enough significant changes for analysis, we analyzed turnover only. None of these mutants showed faster turnover of EV proteins (S2 Fig).
We also investigated whether the EV proteostasis alterations in Gba1b mutants represented a distinctive pathological process or simply an acceleration of normal aging, given that ubiquitinated protein aggregates accumulate with age even in wild-type flies [51–53]. To do this, we measured protein turnover and abundance in old flies (55 to 60 days at the start of labeling) and young flies (5 days). Old flies had dramatically slower turnover of most proteins (mean fold change in half-life for all proteins 2.46 ± 4.31) and milder changes in protein abundance (both increases and decreases; S4 Data). In old flies, only 4% of EV-associated proteins were represented by datapoints in the faster turnover/increased abundance quadrant (Fig 6C), indicating that the altered EV proteostasis observed in Gba1b mutants does not represent an acceleration of normal aging. Together, our findings indicate that altered proteostasis of EV-associated proteins is a specific and novel feature of Gba1b mutants.
As mentioned above, all of our proteomic analyses were performed using protein extracts from fly heads. To test whether the observed alterations in EV protein abundance were also evident in EVs themselves, we performed western blotting for known EV markers on EV fractions from hemolymph, the Drosophila equivalent of blood. To do this, we collected cell-free hemolymph extracts containing the full range of circulating EVs, which we designated total EVs (tEVs). We also prepared extracts containing only EVs under 220 nm in size, which we designated small EVs (sEVs). We then performed western blot analysis on tEVs or sEVs compared to whole-fly homogenate to measure the abundance of two EV marker proteins: Rab11 and an HA-tagged form of ALiX (PDCD6IP) [48, 54]. We also used western blotting to verify EV isolation by the absence of microsomal markers Calnexin (Cnx99A) and Golgin (Golgin84; Fig 7A and 7E) according to International Society for Extracellular Vesicles standards [55].
Rab11 and ALiX-HA were significantly increased in abundance in Gba1b mutants vs. controls in both tEVs and sEVs, but not in whole-fly homogenate (Fig 7A–7E). Although the Rab11 detected in sEVs was 3–5 kDa smaller than in the whole-fly homogenate, this finding is consistent with previous work demonstrating altered molecular weights for several proteins when detected in EVs [56]. A GFP-tagged form of Rab11 also showed increased abundance in sEVs from Gba1b mutants (S3 Fig). The findings using tagged forms of EV proteins are particularly informative because these exogenous proteins were expressed at equivalent overall levels in controls and Gba1b mutants (Fig 7G, S3 Fig). The increased abundance of these markers in EVs from Gba1b mutants indicates that either more of each marker protein is loaded into each EV, or that Gba1b mutants produce more EVs.
One of the most striking abnormalities in Gba1b mutants is their accumulation of Ref(2)P [9], the Drosophila p62 ortholog, which was markedly elevated by proteomic measurement (S1 Data). This is especially noteworthy given that accumulation of Ref(2)P is usually interpreted as an indication of impaired autophagic flux [57–59], and yet we find no evidence of impaired autophagic degradation in Gba1b mutants. Ref(2)P/p62 has multiple functions, however, and mammalian p62 has been detected in EVs [47, 60]. We therefore performed western blotting for Ref(2)P on sEVs from Gba1b mutants and controls to test whether Ref(2)P accumulates in EVs. The sEVs contained very little monomeric Ref(2)P, but did reveal a marked increase in higher molecular weight Ref(2)P oligomers (Fig 8A–8C), which were approximately three times as abundant in Gba1b mutants as in controls (Fig 8C). We confirmed that these high molecular weight bands represented Ref(2)P by performing RNAi knockdown of Ref(2)P in Gba1b mutants (S4 Fig). The increased Ref(2)P abundance in Gba1b mutant EVs suggests that changes in EVs may contribute to the markedly increased Ref(2)P seen in Gba1b mutant heads.
As mentioned above, the increased abundance of multiple EV-associated proteins in Gba1b mutants suggests either that more of each protein is loaded into each EV, or that more EVs are produced. To distinguish these possibilities, we performed nanoparticle tracking analysis on EVs from the hemolymph of Gba1b mutants and controls. For these experiments, we chose to use a 0.65 μm rather than a 0.22 μm filter to retain EVs of as many sizes as possible while still ensuring removal of all cell debris. While the mean size of EVs was comparable in Gba1b mutants and controls (Fig 9A), the concentration of EVs was approximately six times higher in the mutants (Fig 9B). The mean concentrations were 4.55 x 1011 particles/mL (± 1.87 x 1011) for Gba1b mutants and 7.28 x 1010 particles/mL (±3.36 x 1010) for controls (p = 0.013 by Student t test). Thus, the increased abundance of EV proteins in Gba1b mutants is best explained by the increased production of EVs. Together, our findings give clear evidence of altered EV biology in Gba1b mutants.
As previously noted, EVs have been repeatedly described as possible vehicles for the spread of brain protein aggregation in neurodegenerative disease [41]. Our finding that Gba1b mutants had more EVs led us to hypothesize that increased EV release promotes protein aggregation by increasing cell-to-cell transmission of aggregation-prone proteins. As a first step toward testing this model, we determined whether the accumulation of protein aggregates in Gba1b mutants could be suppressed by knocking down components of the ESCRT (endosomal sorting complexes required for transport) pathway, which are required for production of many types of EVs [41, 54]. Using a pan-neuronal driver, we expressed RNAi against proteins from three of the four ESCRT complexes: Mvb12 (Multivesicular body subunit 12; ESCRT-I), lsn (larsen/Vps22; ESCRT-II), and CHMP2B (Charged multivesicular body protein 2b; ESCRT-III). We found that knockdown of each of the three ESCRT proteins significantly reduced accumulation of Ref(2)P in Gba1b mutants, and that knockdown of Mvb12 and lsn also reduced the accumulation of insoluble ubiquitinated protein (Fig 10A–10F). These findings support the model that excessive production of EVs is responsible for the accumulation of protein aggregates caused by GCase deficiency.
Impairment of autolysosomal degradation is widely thought to explain the increased risk of neurodegeneration associated with mutations in GBA, which encodes the lysosomal enzyme glucocerebrosidase (GCase) [1, 16], and multiple studies have found hallmarks of impaired autophagy associated with GCase loss of function. These hallmarks have included accumulation of ubiquitinated protein aggregates, increased abundance of autophagic flux markers such as p62/SQSTM1 and LC3-II, impairment of autophagosome-lysosome fusion, and changes in the size and number of autophagosomes and lysosomes [12, 13, 61–66]. These indications that GCase deficiency leads to autophagy impairment have been found in diverse experimental systems, including multiple animal models, cultured cells, iPSC-derived human neuronal models, and postmortem patient samples [8, 11–13, 31, 66–70]. Our own initial characterization of Drosophila Gba1b mutants, which revealed extensive ubiquitinated protein aggregates and markedly elevated levels of the p62 ortholog Ref(2)P, also appeared to support the model that GCase deficiency impairs autophagic degradation [9]. In our current work, however, proteomic measurement of protein turnover and abundance showed no evidence that degradation of autophagy substrates was globally impaired in Gba1b mutants. The mutants also showed no evidence of failure in other protein degradation pathways. Instead, we found faster turnover and increased abundance of proteins associated with extracellular vesicles (EVs). Followup experiments on isolated EVs confirmed increased abundance of EV marker proteins and revealed a strikingly increased number of EVs. Furthermore, genetic manipulations that reduced EV formation suppressed both the increased protein aggregation and the increased Ref(2)P abundance observed in Gba1b mutants. Our findings suggest that dysregulation of extracellular vesicles, rather than failure of autophagic degradation, may be the primary mechanism by which GCase deficiency leads to protein aggregation and neurodegeneration.
Although the many previous reports of autophagy impairment in GCase-deficient organisms appear incompatible with our current protein turnover findings, we do not believe that our findings contradict previous work. When we measure common markers of autolysosomal function such as Ref(2)P/p62 and insoluble ubiquitinated protein, Drosophila Gba1b mutants show results comparable to those seen in vertebrate models of GCase deficiency [10, 15, 68, 69, 71]. Our proteomic measurements of protein abundance are also consistent with previous reports of increased lysosomal mass in GCase deficiency [1, 8, 66]. The abundance of the lysosomal marker Lamp1 was nearly tripled in Gba1b mutants, and 41% of lysosomal proteins were significantly increased in abundance (S1 Data). Nevertheless, our protein turnover measurements reveal that the overall rates of degradation through lysosomal processes are not grossly altered. Thus, one possible explanation of our findings is that the efficiency of autolysosomal degradation is decreased, with lower throughput per unit of autolysosomal mass, but that the organism has compensated by increasing the amount of autolysosomal machinery available. Because this compensation is sufficient to maintain degradation rates, we would describe Gba1b mutants as being under autolysosomal stress rather than in autolysosomal failure. Over time, the degree of stress may exceed the capacity to compensate, and aged Gba1b mutants may show overt failure of lysosomal degradation. Even if this is the case, late failure of autolysosomal degradation cannot explain the behavioral and biochemical abnormalities that begin in early adulthood [8, 9].
Another explanation for the apparent discrepancy between our findings of normal autophagic substrate turnover and previous reports of impaired autophagy is that commonly used autophagy markers are not solely representative of autophagic flux [57, 72]. This is especially true of Ref(2)P, or p62, which has multiple nonautophagic functions and is transcriptionally upregulated by stress [57, 73]. In addition, p62 and LC3 have recently been detected in mammalian EVs [47, 74, 75], and we found increased levels of oligomeric Ref(2)P in EVs from Gba1b mutants (Fig 8). It is therefore possible that the increased Ref(2)P levels detected in Gba1b mutants result from a combination of stress response and EV dysregulation.
Our work leaves unanswered the question of how GCase deficiency results in increased EV abundance, but does suggest two possible explanations. Increased production of EVs could be caused either by lysosomal stress or by changes in membrane lipid composition. Lysosomal stress has been shown in cultured cells to promote the release of exosomes, a major type of EV [75, 76]. Exosomes are generated when a multivesicular endosome (MVE) fuses with the plasma membrane rather than the lysosome, releasing its intraluminal vesicles into extracellular space [41, 54]. Lysosomal blockade increases the probability that an MVE will fuse with the plasma membrane [75, 76]. If lysosomal stress rather than outright failure is sufficient to trigger increased exosome release, it could account for the overabundance of EVs in Gba1b mutants.
A second explanation for increased EVs in GCase-deficient animals is that abnormal membrane lipid composition may directly alter EV biogenesis. Lipid composition determines membrane fluidity and curvature, and thus controls the size, shape, and fusion kinetics of EVs [77–79]. In fact, lipid rafts, particularly those enriched in ceramide, are required for formation at least one type of EV [78]. Membrane changes such as those caused by GCase deficiency, including accumulation of glucosylceramide and altered ceramide levels [80, 81], could alter EV functioning at any stage from formation to internalization by a recipient cell. Either increased or decreased probability of ceramide-dependent EV formation could lead to increased overall EV production, as suppression of one type of EV has been shown to cause overproduction of another type [82].
While understanding the mechanism by which GCase deficiency causes increased EV release is an important goal of future work, an equally important question is how increased EV abundance in Gba1b mutants promotes the accumulation of protein aggregates. EVs have been increasingly implicated in the pathogenesis of neurodegenerative disease. Many disease-associated proteins, including prion protein, α-synuclein, β-amyloid, and tau, are detected in EVs [41, 83, 84], which have been proposed as vehicles for the well-documented progressive spread of protein aggregates from one brain region to another [83, 85, 86]. In support of this model, toxic forms of these disease-associated proteins are more abundant in EVs from humans with neurodegenerative diseases such as Alzheimer disease, dementia with Lewy bodies, and Parkinson disease (PD) [84, 87, 88], and EVs from these patients can induce protein aggregation in recipient cells under experimental conditions [89, 90]. However, progression of these diseases has not yet been conclusively demonstrated to be mediated by EVs. Perhaps the strongest evidence that EVs promote the spread of protein aggregates has been found for prion protein. Stimulating the release of EVs increased the cell-to-cell spread of misfolded prion protein, and decreasing EV release reduced the spread [91]. Our findings appear to follow the same pattern: genetic interference with EV production suppressed protein aggregation in Gba1b mutants. If the same holds true for other aggregation-prone proteins, conditions that increase EV release could promote the spread of protein aggregates and thus be risk factors for neurodegenerative disease.
Fig 11 illustrates this model. When GCase activity is normal (Fig 11A), EVs travel between cells, carrying both factors that promote protein aggregation (e.g., disease-associated proteins such as α-synuclein) [88, 92] and factors that oppose it (e.g., chaperones) [93]. Some cells likely generate more aggregates than others, and may therefore release more aggregate-promoting factors, including small aggregate “seeds.” Quality control mechanisms in recipient cells successfully combat protein aggregation, and aggregates accumulate only slowly with age. If GCase activity is absent or reduced, however (Fig 11B), more EVs are generated; this results in greater cell-to-cell transfer of aggregate-prone proteins, perhaps simply because these proteins are normally part of EV cargo. In particular, they may be normal cargo of ESCRT-dependent EVs, given our finding that knockdown of ESCRTs in Gba1b mutants ameliorated the mutants’ protein aggregation phenotype. Alternatively, GCase deficiency may alter cargo selection so that more aggregation-prone proteins are loaded into EVs. The net effect of the EV changes is transfer of aggregation-producing factors in quantities that overwhelm quality control mechanisms, leading to excessive accumulation of ubiquitin-protein aggregates in recipient cells.
GBA mutations are the strongest single risk factor for PD and dementia with Lewy bodies, affecting up to 10% of PD patients worldwide [2, 5]. Our finding that GCase deficiency causes increased EV release offers new insight into these prevalent disorders. For example, increased transmission of protein aggregates via EVs could explain the earlier onset and faster disease progression in PD patients with GBA mutations [6, 94–97]. Future investigations should determine how glucocerebrosidase deficiency increases EV abundance, and how manipulations of EV production might prevent or delay the progression of neurodegenerative disease.
Fly stocks were maintained on standard cornmeal-molasses food at 25°C. The Gba1b null (Gba1bΔTT), Gba1b control (Gba1brv), Atg7d4, Atg7d77, Sod2n283, Sod2wk, park25, PINK1B9, and PINK1rv alleles, as well as the UAS-PINK1#2 strain, have been previously described [9, 19, 20, 98, 99]. The UAS-ALiX-HA strain was obtained from the former Bangalore Fly center (National Centre for Biological Sciences, Bangalore, India). The UAS-Ref(2)P-RNAi strain (v108193) was obtained from the Vienna Drosophila Resource Center. Other strains and alleles were obtained from the Bloomington Stock Center: elav-GAL4 (458), Act5C-GAL4 (3953), UAS-Hsc70-4 (5846), w1118 (3605), UAS-Rab11-GFP (8506), Gba1bMB03039 (23602) [100], UAS-Mvb12-RNAi (43152), UAS-larsen-RNAi (38289) [101], and UAS-CHMP2B-RNAi (38375) [102]. Atg7 null mutants were Atg7d4/Atg7d77 transheterozygotes. Sod2 mutants were null/hypomorph compound heterozygotes (Sod2n283/Sod2wk). The full genotype of parkin mutants was If/CyO; park25/park25. The WT controls for Atg7 and parkin mutants were a composite dataset derived from four groups of healthy flies with intentionally diverse genetic backgrounds (see protein turnover rate calculations section). The control for PINK1B9 was its revertant (precise excision) strain, PINK1rv, and the control for Sod2 was CyO/+. The control strain for Gba1b was the revertant Gba1brv. In Fig 7 we used the following genotypes for the experiments involving the ALiX-HA transgene: control = Gba1brv/Gba1bMB03039; Gba1b = Gba1bΔTT/Gba1bMB03039. This combination of Gba1b mutant alleles, which we used for ease of recombination with the ALiX-HA transgene, produced the same biochemical abnormalities found in Gba1bΔTT homozygotes (S5 Fig).
Lipidomic analysis was performed at the Northwest Metabolomics Research Center at the University of Washington, Heads were isolated from 10-day-old control and Gba1b flies flash-frozen in liquid nitrogen, and lipids were then extracted from the frozen head tissue. Levels of glucosylceramide and ceramide were measured by a high-performance liquid chromatography/mass spectrometry (LC-MS/MS) method, using a sphingolipids mix as internal standard (Avanti Sphingolipids Mix II LM-6005). Results were expressed as lipid levels per mass of starting tissue. For each lipid species, three independent samples were analyzed.
[5,5,5 – 2H3] leucine (D3-leucine; 99 atom % deuterium) was obtained from Isotec/Sigma-Aldrich. Synthetic complete medium without leucine (C-Leu) was supplemented with glucose and 60 mg/L D3-leucine. A strain of Saccharomyces cerevisiae auxotrophic for leucine (BB14-3A, Brewer Lab, University of Washington [103]) was grown to saturation at 30°C, then spun down, flash-frozen in liquid nitrogen, lyophilized, and stored at −80°C.
Because brewing in-house produced limited quantities of labeled yeast, we made labeled fly food in batches of ~40 mL using a microwave. We did this by substituting cornstarch for cornmeal in the lab’s standard recipe (2.35% yeast w/v) and dispensing the cooked food in small amounts into vials lined with wet Whatman paper to maintain moisture. Unlabeled transition food for the first 24 hours after eclosion was made and dispensed in the same way, substituting Red Star yeast.
Atg7, parkin, PINK1, and Sod2 mutant samples were processed as previously described [21]. GBA1b and old/young samples were processed as follows: Fused silica microcapillary columns of 75 μm inner diameter (Polymicro Technologies, Phoenix, AZ) were packed in-house by pressure loading 30 cm of Jupiter 90 Å C12 material (Phenomenex). Kasil (PQ Corporation) frit microcapillary column traps of 100 μm inner diameter with a 2-mm Kasil frit were packed with 4 cm of Jupiter 90 Å C12. A retention time calibration mixture (Pierce) was used to assess quality of the column before and during analysis. Three of these quality control runs were analyzed prior to any sample analysis, and another quality control run was performed after every six sample runs. One microgram of each sample digest and 150 femtomoles of the quality control sample were loaded onto the trap and column by the NanoACQUITY UPLC system (Waters Corporation). Buffer solutions used were water, 0.1% formic acid (buffer A), and acetonitrile, 0.1% formic acid (buffer B). The 60-minute gradient of the quality control consisted of 30 minutes of 98% buffer A and 2% buffer B, 5 minutes of 65% buffer A and 35% buffer B, 10 minutes of 40% buffer A and 60% buffer B, 5 minutes of 95% buffer A and 5% buffer B, and 18 minutes of 98% buffer A and 2% buffer B at a flow rate of 0.3 μL/min. The 240-minute gradient for the sample digest consisted of 120 minutes of 98% buffer A and 2% buffer B, 80 minutes of 65% buffer A and 35% buffer B, 20 minutes of 20% buffer A and 80% buffer B, and 20 minutes of 98% buffer A and 2% buffer B at a flow rate of 0.25 μL/min. Peptides were eluted from the column and electrosprayed directly into an Q-Exactive HF mass spectrometer (Thermo Fisher) with the application of a distal 3 kV spray voltage. For the quality control analysis, a cycle of one 60,000 resolution full-scan mass spectrum (400–1600 m/z) was followed by 17 data-independent MS/MS spectra using an inclusion list at 15,000 resolution, 27% normalized collision energy with a 2 m/z isolation window. For the sample digests, a cycle of one 120,000 resolution full-scan mass spectrum (400–1600 m/z) followed by 20 data-dependent MS/MS spectra on the top 20 most intense precursor ions at 15,000 resolution, 27% normalized collision energy with a 1.5 m/z isolation window. Application of the mass spectrometer and UPLC solvent gradients was controlled by the Thermo Fisher XCalibur data system.
The quality control sample data were analyzed using Skyline [23]. High-resolution MS data were processed by BullsEye to optimize precursor mass information [22]. The MS/MS output was searched using COMET [104] with differential modification search of 3.0188325 Da for leucine and 15.994915 methionine and a static modification of 57.021461 Da for cysteine, against a FASTA database containing all the protein sequences from FlyBase plus contaminant proteins. Peptide-spectrum match false discovery rates were determined using Percolator [105] at a threshold of 0.01, and peptides were assembled into protein identifications using an in-house implementation of IDPicker [106].
Turnover rates were calculated using Topograph software [22]. For a full description of Topograph settings, see Vincow et al. [21]. A protein’s turnover rate was computed based on data from all peptides detected, and values from all biological replicates were pooled for turnover calculations. A protein’s turnover rate was calculated based on at least 6 measurements per genotype of percent turnover for GBA1b mutants and old/young flies, and at least 15 measurements per genotype for Atg7, parkin, PINK1, or Sod2 mutants. Peptides that could be the product of more than one gene were excluded from analysis. For a small percentage of genes (2%-5%), Topograph clustered peptides corresponding to a single gene into 2-3 nonoverlapping “isoform groups.” For example, isoform group 1 might include peptides mapping only to the COX6B-PA isoform, while isoform group 2 peptides could have come from COX6B-PA, -PB, or -PC. While in most cases the isoform groups for a single protein had essentially identical turnover rates, occasionally they displayed significant differences in turnover behavior. Each isoform group was therefore analyzed as a separate protein.
We excluded proteins with excessive inter-replicate variability of turnover rates, defined as coefficient of variation ≥ 0.25. We calculated the turnover rate separately for each biological replicate and determined the coefficient of variation across replicates. Proteins were analyzed only if they met inclusion criteria in both mutants and controls.
In previous work, we had compared Atg7 and parkin null mutants to their respective heterozygotes [21]. However, we later found that both Atg7 and parkin heterozygotes had mild but significant slowing of mitochondrial protein turnover compared to WT flies, and we selected the WT dataset as a more appropriate control. For turnover analyses, Atg7 and parkin nulls were both compared to a composite WT dataset derived from four separate groups of healthy flies (w1118, PINK1rv, CyOActGFP/+, and a mixture of`CyO/Hsp70-GAL4 and CyO/UAS-PINK1#2). Turnover rates are the mean values for all genotypes in which the protein was detected; the rates are highly consistent across genotypes, as previously reported [21]. Each mean value for a genotype was treated as one replicate for statistical purposes. Statistical significance of fold change in turnover was calculated for groups of proteins using nested ANOVA [107], and significance of change for individual proteins was calculated using t tests. The following subgroups of proteins had enough replicates for t tests: 148 mitochondrial, 36 ribosomal, 15 ER/peroxisomal, and 275 nonorganellar proteins.
We measured protein abundance from the same raw mass spectrometry data used in the turnover study, using Skyline [23] and MSstats [24]. Prior to MSstats analysis, we obtained total abundance (labeled plus unlabeled) for each peptide using a custom R script. The statistical significance of intergroup differences was calculated using a linear mixed model, then adjusted for multiple comparisons by the Benjamini-Hochberg procedure with a false discovery rate of 0.05. All abundance comparisons were made at the second time point, when differences between genotypes were most marked. In abundance analyses, parkin and Atg7 mutants were compared to their original heterozygote controls rather than WT flies (see calculations above). While the composite control group approach was appropriate for measurement of turnover, which is more consistent and less noisy than abundance [21], measurement of relative protein abundance required mutant and control samples that had been run at the same time.
General: Drosophila protein localization was determined from a variety of resources including gene and protein information databases (FlyBase [108], MitoDrome [109], NCBI [110], UniProt [111]), protein localization prediction algorithms (WoLF PSORT [112], MitoProt [113], Predotar [114], SignalP [115], NucPred [116], and PTS1 Predictor [117, 118]), BLAST [119], and primary literature.
Proteasome substrates: We identified proteins as proteasome substrates (Fig 4) if their mammalian orthologs had one or more regulated ubiquitinated sites according to Wagner et al. [35]. These sites showed altered abundance of ubiquitinated peptides after proteasome inhibitor treatment. We identified Drosophila orthologs of proteins from the Wagner et al. data with the DRSC Integrative Ortholog Prediction Tool (DIOPT) v6 [50], minimum score 5.
Microautophagy substrates: We identified microautophagy substrates by searching for targeting sequences (Fig 4), also called KFERQ-like motifs. These motifs were defined as sequences of five amino acids (AAs) that fit criteria established by Dice [120, 121]:
We wrote an algorithm using Python 2.7 to search protein sequences for these motifs and applied it to the fly proteome (FASTA sequences downloaded from FlyBase). We then identified cytosolic proteins by annotation as described above, and compared the effects of GCase deficiency on cytosolic proteins with and without KFERQ-like sequences.
Endocytic turnover substrates: Proteins designated endocytic turnover substrates in Fig 5 were identified using FlyBase annotation and search terms such as receptor, transmembrane, extracellular matrix, integral component of plasma membrane, and channel. Endosomal machinery proteins (see below) were excluded.
Endosomal machinery: Proteins designated “endosomal machinery” in Fig 5 were identified by a FlyBase search for the string “endosom*” in at least one of the following fields: GO Molecular Function, GO Biological Process, GO Cellular Component, Gene Snapshot, or UniProt Function.
Extracellular vesicle proteins: To identify extracellular vesicle proteins, we compiled a list of proteins detected in EVs in mass spectrometry studies of Drosophila cultured cells [46–49]. The list contained 544 unique proteins, 329 of which were found in Gba1b mutant protein turnover data and 499 in abundance data. In addition, we obtained from ExoCarta [47] the list of “top 100 [mammalian] proteins that are often identified in exosomes,” and identified 97 Drosophila orthologs of these proteins using DIOPT v6 as previously described [50]. Fifty-nine proteins from this list were found in Gba1b mutant turnover data and 86 in abundance data.
The significance of intergroup differences was evaluated using the Fisher exact test except when the total number of proteins was too large, in which case we performed a χ2 test of homogeneity.
Proteasome activity was measured in heads from male and female flies 10 to 11 days old (50 per sample) according to the method of Tsakiri et al. [122], with the following modifications: We used 26S lysis buffer only. We obtained substrate buffer and fluorescent substrates from the UBPBio Proteasome Activity Fluorometric Assay Kit II (J4120), and we used epoxomicin 20 μM for proteasome inhibition. Specifically, we divided the lysate in half and added DMSO to one half and epoxomicin to the other. We measured the protein concentration of lysates using the Pierce BCA Protein Assay Kit (23227), and measured sample fluorescence with a Synergy H1 BioTek plate reader (excitation 350 nm, emission 450 nm). We subtracted the activity measured in the epoxomicin-treated homogenate from the activity in the DMSO-treated homogenate. The experiment was repeated three times.
For total EVs (tEVs), hemolymph was obtained from 20 flies (10 males and 10 females, 10 to 11 days old) per sample. In order to obtain whole-fly homogenate from the same animals used for collection of hemolymph, hemolymph was extracted manually from the first four flies and their bodies were reserved for later use. These flies were decapitated, following which their hemolymph was collected by pressing on the thorax with the head of a butterfly pin. The hemolymph from these four flies was collected by capillary action into 1 μL PBS, and the total sample was transferred to a 1.7-mL microfuge tube containing 9 μL PBS. The heads and bodies of the four flies were then homogenized in RIPA buffer for whole-fly protein homogenates. Two holes were made with a 25-g needle in the bottom of a 0.5-mL tube, and 16 more flies were decapitated and placed in this tube. The 0.5-mL tube was then seated in the PBS-containing 1.7-mL tube for centrifugation. The tubes were centrifuged at 5000 x g for 5 min at 4°C, after which the extracted hemolymph was centrifuged for 30 min at 10,000 x g at 4°C to remove cell debris and the cell-free supernatant was collected. An equal volume of 2x Laemmli buffer (4% SDS, 20% glycerol, 120 mM Tris-Cl pH 6.8, 0.02% bromophenol blue, 2% β-mercaptoethanol) was added to the cell-free supernatant and also to the whole-fly protein homogenates, and all samples were boiled for 10 min and then stored at −80°C. The experiment was repeated at least three times.
Small EVs (sEVs) were prepared as for tEVs, with the following changes: 50–60 adult flies were used per sample. The hemolymph was collected into a volume of PBS scaled to the number of flies used (1 μL/fly) to minimize sample loss during filtration. After the 10,000 x g spin, Total Exosome Isolation Reagent for Cell Culture (Thermo Fisher/Invitrogen, 4478359) was used as in Tassetto et al. [123] except that we used Ultrafree 0.22 μm spin filters (Fisher, UFC30GV0S). The resulting filtrate was boiled and stored as for the tEVs.
Heads from 10-day-old flies (6 females and 6 males per sample) were homogenized in Triton lysis buffer (50 mM Tris-HCl pH 7.4, 1% Triton X-100, 150 mM NaCl, 1 mM EDTA), and then spun at 15,000 x g for 20 min. The detergent-soluble supernatant was collected and mixed with an equal volume of 2x Laemmli buffer, and the same buffer was used to resuspend the Triton-insoluble pellet. All samples were boiled for 10 minutes. The Triton-insoluble protein extracts were then cleared of debris by centrifugation at 15,000 x g for 10 minutes, followed by collection of the supernatant. At least three independent experiments were performed.
Proteins were separated by SDS-PAGE on 4%-20% MOPS-acrylamide gels (GenScript Express Plus, M42012) and electrophoretically transferred onto Immobilon PVDF membranes (Fisher, IPVH00010). Immunodetection was performed using the iBind Flex Western Device (Thermo Fisher, SLF2000). Antibodies were used at the following concentrations: 1:25,000 mouse anti-Actin (Chemicon/Bioscience Research Reagents, MAB1501), 1:250 mouse anti-Rab11 (BD Transduction Laboratories, 610657), 1:200 rabbit anti-Ref(2)P (Abcam, ab178440), 1:500 mouse anti-ubiquitin (Santa Cruz, sc-8017), 1:800 mouse anti-Cnx99A (DHSB, Cnx99A 6-2-1), 1:100 mouse anti-Golgin-84 (DHSB, Golgin84 12–1), and 1:500 rat anti-HA (Sigma-Aldrich, 11867423001). HRP secondary antibodies were used as follows: 1:500 to 1:1000 anti-mouse (BioRad, 170–6516), 1:100 anti-rat (Sigma-Aldrich, A9037), and 1:500 to 1:1000 anti-rabbit (BioRad, 172–1019). Signal was detected using Pierce ECL Western Blotting Substrate (Fisher, 32106). Densitometry measurements of the western blot images were measured blind to genotype and condition using Fiji software [49]. For homogenates, signal was normalized either to Actin or to Ponceau-S [124, 125]. For EVs, signal was normalized to loading volume. Normalized western blot data were log-transformed when necessary to stabilize variance before means were compared using Student t test. Each experiment was performed at least three times.
EVs were prepared for nanoparticle tracking analysis (NTA) as described for western blotting through the 10,000 x g step, after which they were passed through a 0.65 μm Ultrafree-MC filter (Fisher, UFC30DV0S) to ensure removal of any remaining cellular debris and stored at −80°C. Hemolymph was obtained from 60 flies per sample, and four biological replicates per genotype were collected.
EV size and concentration were measured using NTA by Alpha Nano Tech LLC (Research Triangle Park, NC). NTA was performed using a ZetaView instrument equipped with an sCMOS camera and 532 nm laser. Instrument parameters were as follows: temperature setting 23°C, Max Area 500, Min Area 20, Min Brightness 20. Two cycles of analysis at 11 positions were performed for each sample. Data were analyzed using ZetaView software version 8.04.02.
Standard laboratory protection equipment was used during all steps of sample preparation and analysis to prevent sample contamination with dust particles. The 1x PBS solution (Amresco) used to dilute samples was filtered on the day of analysis through a 0.22 μm Millex-GV syringe filter (Millipore), and its purity was confirmed by NTA analysis prior to the study. Instrument qualification was performed by analyzing a polystyrene bead standard (100 nm, Particle Metrix) in 1x PBS prior to each study. Instrument accuracy and precision were confirmed to ± 5% of the target value.
|
10.1371/journal.pcbi.1007033 | Reduced level of docosahexaenoic acid shifts GPCR neuroreceptors to less ordered membrane regions | G protein-coupled receptors (GPCRs) control cellular signaling and responses. Many of these GPCRs are modulated by cholesterol and polyunsaturated fatty acids (PUFAs) which have been shown to co-exist with saturated lipids in ordered membrane domains. However, the lipid compositions of such domains extracted from the brain cortex tissue of individuals suffering from GPCR-associated neurological disorders show drastically lowered levels of PUFAs. Here, using free energy techniques and multiscale simulations of numerous membrane proteins, we show that the presence of the PUFA DHA helps helical multi-pass proteins such as GPCRs partition into ordered membrane domains. The mechanism is based on hybrid lipids, whose PUFA chains coat the rough protein surface, while the saturated chains face the raft environment, thus minimizing perturbations therein. Our findings suggest that the reduction of GPCR partitioning to their native ordered environments due to PUFA depletion might affect the function of these receptors in numerous neurodegenerative diseases, where the membrane PUFA levels in the brain are decreased. We hope that this work inspires experimental studies on the connection between membrane PUFA levels and GPCR signaling.
| Our current picture of cellular membranes depicts them as laterally heterogeneous sheets of lipids crowded with membrane proteins. These proteins often require a specific lipid environment to efficiently perform their functions. Certain neuroreceptor proteins are regulated by membrane cholesterol that is considered to be enriched in ordered membrane domains. In the brain, these very same domains also contain a fair amount of polyunsaturated fatty acids (PUFAs) that have also been discovered to interact favorably with many receptor proteins. However, certain neurological diseases—associated with the inadequate functioning of the neuroreceptors—seem to result in the decrease of brain PUFA levels. We hypothesized that this decrease in PUFA levels somehow inhibits receptor partitioning to cholesterol-rich domains, which could further compromise their function. We verified our hypothesis by an extensive set of computer simulations. They demonstrated that the PUFA–receptor interaction indeed leads to favorable partitioning of the receptors in the cholesterol-rich ordered domains. Moreover, the underlying mechanism based on the shielding of the rough protein surface by the PUFAs seems to be exclusive for multi-helical protein structures, of which neuroreceptors are a prime example.
| Cellular membranes host functional membrane domains (“lipid rafts’’) rich in proteins and cholesterol (CHOL) [1]. Many G protein-coupled receptors (GPCRs) and cognate G proteins are found in these domains [2], and numerous reports have suggested that CHOL is involved in GPCR function [3–7]. Moreover, impaired CHOL homoeostasis and raft disruption have been linked to different neurodegenerative diseases [2, 8], where GPCRs play a pivotal role. However, the mechanism driving the partitioning of GPCRs to their native functional CHOL-rich environments is still not well understood.
Polyunsaturated fatty acids (PUFAs) such as docosahexaenoic acid (DHA, 22:6(n-3)) are likewise key membrane components of brain cells [9]. PUFAs esterify to phospholipids together with a saturated chain to form a hybrid lipid. Intriguingly, despite their disordered nature, hybrid lipids are found in raft extracts [10–12], and they also partition surprisingly well to cholesterol-rich ordered membrane regions [13]. However, raft PUFA levels are reduced in various neuropsychiatric and mental disorders [14] including Alzheimer’s [10] and Parkinson’s diseases [11]. This lack of PUFAs could thus affect GPCR function. In fact, experiments have shown that DHA-containing lipids enhance the function of the prototypical GPCR rhodopsin [15–17], which simulation studies have explained to take place as a result of the high conformational flexibility of DHA chains. This provides hybrid lipids with high affinity for the rough surface of GPCRs, [18–21] further promoting protein–protein interactions [22].
We recently reported the high affinity of DHA for the adenosine A2A receptor (A2AR) [23], a GPCR with an important role in the central nervous systems, where different antagonists of A2AR have shown promising neuroprotective effects [24, 25]. Membrane CHOL is also known to closely interact with A2AR [7, 26–28], modulating its function [29] and ligand binding properties [7]. The partitioning of A2AR into ordered membrane domains [30] is therefore quite expected, though the mechanism rendering it possible has been suggested to be complex [31]. Moreover, given the numerous factors affecting protein partitioning [32] and the limited ability of model systems to capture in vivo behavior [33], it is not surprising that the role of PUFAs in A2AR partitioning remains to be investigated. Given the central role of GPCRs in cell signaling, unlocking how DHA interacts with GPCRs is the key to understanding why GPCR function is impaired in severe brain diseases associated with a lowered membrane DHA level.
Here, we studied the role of PUFAs in the partitioning of GPCRs into CHOL-rich (raft-like) liquid-ordered (Lo) and CHOL-depleted liquid-disordered (Ld) phases. Combining all-atom and coarse-grained molecular dynamics (MD) simulations with free energy calculations, we demonstrate for A2AR that in the absence of DHA, corresponding to brain tissue of diseased individuals, partitioning to the Ld phase is energetically favored. However, in membranes including DHA-containing hybrid lipids, corresponding to brain tissue of healthy individuals, DHA drives A2AR to partition to the Lo phase, as a favorable structural arrangement of DHA around A2AR minimizes the structural perturbations therein. Furthermore, based on our studies on a number of distinct membrane proteins, we demonstrate that the observed effect of DHA could be limited to rough helical multi-pass membrane proteins, which include GPCRs.
We calculated the free energy of transfer of A2AR between Lo and Ld phases in the coarse-grained (CG) scheme using the non-polarizable Martini 2.2 model [34–36]. First, we embedded the protein in an Lo phase membrane containing distearoylphosphatidylcholine (DSPC, Fig 1B), 20 mol% CHOL (Fig 1E), and different concentrations of stearoyldocosahexaenoylphosphatidylethanolamine (SDPE, Fig 1D) with a polyunsaturated DHA chain, see Fig 1A. In line with lipidomics experiments, DHA was paired with the PE head group. [37] Next, we mutated Lo-forming DSPC to Ld-forming dioleoylphosphatidylcholine (DOPC, Fig 1C) in a set of simulations and extracted the free energy change ΔGLoProt→LdProt using the free energy perturbation approach. Here, a coupling parameter λ has a value of 0 for DSPC and 1 for DOPC. Then, we obtained ΔGLo→Ld by repeating this calculation in the absence of the protein. As discussed in Section B.7 in the S1 File, it is possible that the experimentally observed microscopic phase separation in this DOPC/DSPC/CHOL mixture [38] is associated by a fairly large line tension and hence only takes place in membranes larger than those currently within the reach of MD simulations. This limits us from studying protein partitioning in DOPC/DSPC/CHOL mixtures with coexisting domains. Nevertheless, the lipid chain order parameters, shown in Fig 2A as a function of λ, demonstrate a smooth transition between distinct Lo and Ld phases in both sets of the simulations. We therefore believe that our approach is able to capture the physical properties of the coexisting phases in isolation. Further analyses shown in Section B.1 in the S1 File also support this view. Following the thermodynamic cycle depicted in Fig 1F, we carried on to extract the free energies of transfer as (ΔGLoProt→LdProt−ΔGLo→Ld). Additionally, we also used a more realistic composition—based on the tie lines measured for the DOPC/DPPC/CHOL mixture—where the DSPC/DOPC ratio was 2.3 in the Lo phase, and then reversed to 1/2.3 in the Ld phase. For further details on our computational approach, the system compositions, and the simulation parameters, see Methods and the S1 File.
The free energy of transfer of A2AR as a function of SDPE concentration is shown as a solid line in Fig 3A. Strikingly, the free energy of transfer changes sign at the SDPE concentration of ∼8 mol%. This highlights that for dilute concentrations of SDPE, A2AR partitions to the Ld phase. However, at higher SDPE concentrations the picture changes completely and the protein favors partitioning to the Lo phase.
Concluding, the data provide compelling evidence that the presence of SDPE, and therefore DHA, makes A2AR compatible with the Lo phase.
Fig 4A shows the 2D radial distribution functions (RDFs) of all lipid chain types around A2AR in the Lo phase with 4 mol% of SDPE. The data are extracted from well-equilibrated membranes in the CG scheme. Fig 4A demonstrates that A2AR is fully coated by SDPE with polyunsaturated DHA forming the first solvation shell, followed by the saturated acyl chain of SDPE and CHOL in the second shell. The formation of these shells is illustrated in the movie at DOI:10.6084/m9.figshare.5903881. With increasing SDPE concentration, the right tail of the RDF peaks of all lipids extends further away from the protein (see Fig. E in the S1 File), indicating that the A2AR surface becomes saturated with DHA.
Interestingly, Fig 4A shows that CHOL penetrates the shell formed by the saturated chains of SDPE, and occupies annular binding sites, in line with experimental and computational studies on CHOL–A2AR interaction [26–28]. Indeed, cholesterol finds the suggested binding sites in the absence (Fig 4B) but also in the presence of (Fig 4C) an SDPE shell.
These lipid shells around A2AR are dynamic. Lipids exchange in the time scale of ∼100 ns, as evidenced by the decay time constants found through double exponential fits to the contact data, shown in Table B in the S1 File (see also Section B.2 in the S1 File). For the Lo phase, the rates of SDPE and CHOL exchange increase as SDPE concentration increases. In the Ld phase, the SDPE corona dissolves (see Fig. G in the S1 File). This lack of a tightly-bound SDPE shell leads to higher SDPE and CHOL exchange rates. Similarly, CHOL exchange rates are also higher in the absence of SPDE. These findings demonstrate that the formation of an SDPE shell also affects the dynamics of CHOL association by stabilizing the neighborhood of A2AR.
Concluding, the strong affinity of DHA to interact with A2AR leads to coating of the protein by SDPE lipids. DHA is in contact with the protein, whereas the saturated chains favor interactions with CHOL.
Partitioning of a membrane protein to either the Lo or the Ld phase is driven by the mutual structural compatibility between the protein and the lipids forming the membrane phase. Possible parameters describing this compatibility include hydrophobic mismatch, the conformational entropy of the protein, and perturbation of lipid chain order. We evaluated the contribution of all these factors in the CG scheme.
Membrane thickness is shown in Fig. K in the S1 File as a function of distance from protein surface. The presence of SDPE has a clear effect on the thickness. Based on the mattress model [39] and using the hydrophobic mismatch parameter from Ref. [32] and the hydrophobic thickness of A2AR from the OPM database [40], we estimate that hydrophobic mismatch contributes to the free energy of transfer by approximately 1.8 kJ/mol, favoring the Ld phase. However, the presence of 8 mol% of SDPE has an insignificant effect on this value, indicating that negating hydrophobic mismatch is not the mechanism through which SDPE shifts partitioning of A2AR towards the Lo phase. Notably, this conclusion is insensitive to the value of the hydrophobic mismatch parameter, which might be different between experiment and our simulation model.
Next, we evaluated whether the SDPE corona promotes protein flexibility, hence resulting in a favorable entropic contribution for partitioning to the Lo phase in the presence of SDPE. We plot the residue-wise root mean squared fluctuations (RMSF) of the protein structure in both the Lo and Ld phases in Fig. L in the S1 File. Curiously, in the absence of SDPE, the average RMSF value is slightly higher in the Lo phase. However, at 8 mol% of SDPE the average RMSF becomes larger in the Ld phase than in the Lo phase (see inset in Fig. L in the S1 File). This suggests that the entropic contribution due to the presence of SDPE actually promotes A2AR partitioning to the Ld phase and hence acts against the observed effect of SDPE on the free energy of transfer. Moreover, we note that the omitted lipid entropies also likely play a role on partitioning.
How about protein-induced changes in lipid acyl chain order? The outer layer of the SDPE corona around A2AR is formed by the saturated stearic acid chains of SDPE (see Fig 4A). This layer is likely more compatible with the Lo phase than the rough surface of A2AR. This idea is indeed backed up by results from CG systems, which show that the effects of SDPE on membrane properties are reduced in the presence of A2AR and vice versa (see Fig 2A).
We note here that the CG approach is not well-suited to fully characterize acyl chain order. Therefore, we also studied the effects of SDPE and A2AR on membrane order in all-atom detail. To this end, we fine-grained selected coarse-grained Lo phase systems and carried out all-atom simulations using the CHARMM36 force field [41, 42] as described in Methods. The averaged stearic acid chain order parameters from both all-atom and coarse-grained simulations of the Lo phase membranes are shown in Fig 2B.
It is evident that both A2AR and SDPE lower the average order of the membrane. However, at an SDPE concentration of 8 mol%, the presence of A2AR actually increases membrane order, and this observation holds for both all-atom and coarse-grained schemes. The explanation to this behavior is that when both SDPE and A2AR are present, the DHA–A2AR interactions shield the order-lowering effects of both SDPE and A2AR. Importantly, the values from coarse-grained and atomistic simulations are in the same ballpark. The spatial variation of membrane order due to the presence of A2AR is studied in detail in Section B.3 in S1 File.
To conclude, in the Lo phase, the association of the flexible DHA chains and the rough surface of A2AR weakens their perturbations on membrane (acyl chain) order.
Previous simulations and experiments have demonstrated the favorable interactions of DHA and GPCRs, including A2AR and dopamine D2 receptor (D2R) [18, 20–23, 43]. Here, we systematically studied four distinct membrane protein types—one β-barrel and three α-helical structures with 1, 2, or 7 transmembrane passes, including A2AR as a representative GPCR. The proteins are 1) the transmembrane domain of the human receptor tyrosine kinase (ErbB1, PDB id: 2M0B), a single helix; 2) a dimer formed by two Glycophorin A peptides [44] (GpA dimer, PDB id: 1AFO); 3) A2AR (PDB id: 3EML) [45], a heptahelical bundle employed in the CG free energy calculation; and 4) the voltage-dependent anion channel (VDAC, PDB id: 3EMN) [46], a β-barrel. These proteins are depicted in the middle column of Fig 5. Notably, the lengths of the hydrophobic spans of the helical proteins were all equal to 3.2 nm [40], so this factor cannot lead to differences in lipid–protein interactions. However, the β-barrel is substantially thinner at 2.3 nm.
We simulated these proteins in membranes comprised of lipids, whose chains’ unsaturation level was varied (chains with 0, 1, 2, or 6 double bonds per chain). We evaluated how the lipids solvated the proteins in these membranes using unbiased all-atom simulations together with the CHARMM36 force field [41, 42]. We paired all lipid chains with a PC head group in order to study only the effect of lipid chains. The final structures of the simulated systems are shown in the rightmost column of Fig 5. The details are given in Methods and in Section A.4 in the S1 File. The RDFs of the fatty acid chains around the proteins were determined after full lipid mixing had taken place.
It is evident from these RDFs (see leftmost column of Fig 5) that the non-GPCR proteins (here ErbB1, GpA dimer, and VDAC) do not show any clear preference for DHA. Meanwhile, A2AR, as a representative example of GPCRs, interacts mostly with the DHA chain of SDPE, and the saturated chain of SDPE again forms an outer layer of the lipid corona that is in contact with the protein. This observation, in agreement with the results of CG simulations (Fig 4A and our earlier study [23]), suggests that DHA adapts to the rough surface of A2AR. Protein roughness (i.e. the degree of irregularity of a protein surface) [47] is known to correlate with its propensity to interact with small molecules [48]. Therefore, it has been used to predict binding sites at the protein surface [49]. Importantly, surface roughness is a general feature of GPCRs [50] and explains the preferential interaction of the flexible and kinked DHA chain with A2AR [19]. The fact that a smoother β-barrel (VDAC) surface is not solvated by DHA is in favor of this view. Since this phenomenon is also not observed for proteins with a smaller number of helices (ErbB1 and GpA dimer), its origin likely lies in the preference of DHA for the creviced tertiary structure instead of the helical secondary structure.
To further quantify the DHA adaptation onto the A2AR surface, we calculated the mean number of the residues in the helical TM region of A2AR that were in the vicinity (<0.3 nm) of a lipid chain in the fine-grained simulations. We found a systematic increase: +10% for the membrane with 4 mol% of SDPE and +15% for the membrane with 8 mol% of SDPE as compared to the SDPE-free case. This effect was not dependent on the chosen cutoff, as values of +7% and +17% were calculated for a cutoff of 0.4 nm. While this calculation clearly shows that DHA chains adapt better to the A2AR surface, a further and more systematic study on the effects of the surface topology and the amino acid content therein on DHA–protein interactions is required in the future to verify our findings. The favorable interaction between flexible DHA chains and GPCR surfaces is highlighted in Fig 6, which shows representative configurations sampled in the fine-grained all-atom simulations, where a DHA chain has adapted its conformation to the rough protein surface and entered a crevice on the A2AR surface (see Fig 6A), or penetrated into the helical bundle of A2AR (see Fig 6B).
Concluding, hybrid lipids with a DHA (or likely other PUFA) chain and a saturated chain seem to be favored by GPCRs, and this is likely due to the rough surface of the transmembrane region in these multi-helical proteins.
Based on the observation that the DHA–protein interaction is characteristic for proteins with multi-pass helical bundles, we extended our free energy of transfer calculation in the CG scheme to two additional proteins of this kind. We also note that the effects for other membrane protein types might be similar in the Martini scheme as many proteins seem to interact favorably with PUFAs [51], likely due to unbalanced entropic and enthalpic contributions to this interaction. However, based on our all-atom simulations, we abstain from studying the free energies of transfer for protein types without multiple TM helices. D2R is linked to many neurological and psychiatric disorders [52] associated with lowered PUFA levels [10, 11, 14]. The DHA–D2R interaction was recently demonstrated by us [23]. We also considered the brain-associated glucose transporter GLUT1, whose function is also dependent on PUFAs [53, 54]. While GLUT1 is not a GPCR, it also has a multi-pass structure consisting of 12 helices.
We estimated the free energies of transfer for all three proteins—D2R, GLUT1, and A2AR—in the absence and in the presence (16 mol%) of SDPE and hence DHA. We also note that while the phase-separation of the commonly used lipid mixtures in the Martini model is complete and the phase boundaries are sharp [55], experiments report less distinct compositions between the Lo and Ld phases [56]. We therefore considered both the situation mimicking complete separation (such as above), as well as a more realistic situation in which the Lo phase had a realistic DSPC/DOPC ratio of 2.3, which is reversed during the mutation into an Ld phase (see Methods and Section A.2 in S1 File).
The free energies of transfer for all three proteins are shown in Fig 3B. The effect of SDPE is clearly demonstrated for all proteins. Moreover, while the absolute values are smaller in the membranes with more realistic compositions, the change of sign, i.e. the change in the favored phase changes consistently upon the addition of SDPE. This behavior is in line with the two phases now being less distinct, as demonstrated by the order parameters shown in Fig. D in the S1 File. Moreover, the strong association of D2R and GLUT1 with DHA (see Fig. F in the S1 File) is again responsible for the effect—similar to what was observed for A2AR (see Fig 4A). It is also worth pointing out that while we paired DHA with a PE head group (to form SDPE), the calculations performed with SDPC instead of SDPE show an almost equal effect on protein partitioning (see Section B.8 in S1 File).
Concluding, the SDPE-induced partitioning to the Lo phase is reproduced across three multi-helical brain-associated proteins—two of which are GPCR neuroreceptors—whose function is compromised by changes in membrane DHA levels. This effect is also consistently observed with less distinct and more realistic phase compositions.
Using multi-scale simulations and free energy calculations, we showed that a small amount of SDPE, a DHA-containing hybrid lipid, enhances A2AR partitioning to the Lo phase. Without DHA, the protein favors partitioning to the Ld phase instead. The change in this behavior stems from the rough surface of A2AR that favors interacting with DHA and, presumably, also with other PUFAs over saturated chains. This interaction leads to a well-organized SDPE corona where the DHA chains face the receptor, while the saturated chain of SDPE in the outer layer of the corona interacts with CHOL and saturated phospholipid chains in the Lo phase. Through this mechanism, the perturbations of the flexible DHA chains and the rough receptor surface on the Lo phase are largely eliminated. The striking finding made in this work is that the lipid corona could play a decisive role in the partitioning of membrane proteins. We showed that this holds true for A2AR used in this work as a prototypical GPCR. The additional results strongly suggest that the same conclusion holds for helical multi-pass proteins such as GPCRs with rough surfaces, yet not for other protein topologies with smoother surfaces.
We acknowledge that while coarse-grained models are designed to capture the correct trends, the absolute free energy values should be taken with caution. Still, our values are in line with [32] if not smaller than (compare the data for WALP23 in Methods with Ref. [55]) the values obtained with the Martini model exploiting different free energy techniques. We discuss other possible methodological limitations in detail in Section B.7 in the S1 File.
Our results suggest that small concentrations of lipids not included in model membranes might have drastic effects on the partitioning behavior of membrane proteins studied in vivo, and can explain why raft-associated proteins partition to the Ld phase in phase-separated giant unilamellar vesicles [33]. Further, the present simulation results are in line with experiments suggesting that other structural features such as post-translational modifications, protein surface roughness, and hydrophobic mismatch modulate the affinity of membrane proteins for lipid rafts [32]. Given that the solvation of a GPCR by a DHA-containing hybrid lipid is based on a layer where DHA stands next to the protein surface and saturated chains occupy the outermost shell of the protein, this arrangement can increase the raft affinity of the GPCR protein in three ways: it provides the protein with non-covalently bound saturated lipid anchors, it complements the surface roughness of the protein, and with an appropriate choice of the saturated chain in the hybrid lipids, hydrophobic mismatch can be reduced.
The concentration of DHA in raft membrane domains in the brain of healthy subjects is ∼7 mol-% [11]. Assuming a protein area coverage similar to that in red blood cells [57] and an average protein and lipid area of 10 nm2 and 0.7 nm2, respectively, the protein-to-lipid ratio would be approximately 1 to 50 per leaflet. With an SDPE content of ∼14 mol-%, and considering that the membrane has two leaflets, the estimated protein-to-SDPE ratio is 1 to 13. Strikingly, the saturation of the A2AR surface in our simulations takes place around this protein-to-SDPE ratio (see Figs. G and H in the S1 File). Hence, this consideration suggests that in the brain tissue of healthy subjects the DHA concentration is sufficiently large to favor the partitioning of A2AR to ordered regions with structural similarity to the Lo phase. However, one has to keep in mind that our simplified model membranes do not capture either the heterogeneity or leaflet asymmetry of membranes in the brain, which can fine-tune the partitioning behavior of proteins. Moreover, the membranes considered in this study are planar, yet GPCRs with high intrinsic curvature are also sorted by curvature [58], and the DHA corona might have an effect therein. Studies of asymmetry or curvature are beyond this work, yet might need to be taken into account when experimental validation for our findings is sought.
Then what happens if the DHA level is decreased? It is known that the level of DHA in the brain of people suffering from neurodegenerative diseases is substantially reduced [10, 11, 59]. It is tempting to speculate that the reduced DHA content would alter the partitioning of A2A or D2 receptors, displacing them from CHOL-rich domains to disordered regions, compromising GPCR signaling. It has been shown that cholesterol binds to GPCRs such as beta-2-adrenergic receptor in an allosteric manner [6], affecting its conformational distribution, thus the concern of compromised GPCR signalling due to a lowered DHA level is justified. In brief, the effect observed in the present study on partitioning has implications on health. While DHA is promising in the treatment of neurodegenerative diseases [60], the mechanism behind this protective effect, despite rendering membranes more fluid, remains elusive. In our earlier study [23], we showed that the formation of A2AR homo- and hetero-oligomers with the dopamine D2 receptor is decreased when the DHA levels are reduced. In the current work, we postulate that DHA-containing lipids have a dual role in preventing neurodegenerative diseases by lipid–protein interactions: 1) they can influence raft partitioning, therefore indirectly 2) modulating key aspects of the GPCR biology, such as protein oligomerization. The proper function of these oligomeric and mutually regulatory receptor units in a suitable lipid environment is essential for the properly functioning healthy brain. Our findings could explain some of the beneficial effects of DHA-based therapies previously shown for certain brain disorders [61].
All simulations are listed in Table A in the S1 File. The approach for extracting free energies of transfer is described in Section A.1 in the S1 File, and details of simulation models and methods are given in Sections A.2–A.4 of the S1 File.
We embedded A2AR (PDB id: 3EML [45]) to an Lo membrane consisting of DSPC and 20 mol% CHOL. Next, varying amounts of DSPC was replaced by the hybrid lipid SDPE with a saturated (C18:0) and a polynsaturated (DHA) chain. The protein and the lipids were modeled in the coarse-grained (CG) scheme using the non-polarizable Martini 2.2 model [34–36] together with the elastic network for A2AR [62].
Next, DSPC was transformed into DOPC, resulting in the change of membrane phase from Lo to Ld. This process was performed as an alchemical transformation using the dual topology paradigm with 27 windows. We verified the change in phase thoroughly (see Section A.1 in the S1 File), and validated our approach using the 27-residue WALP peptide that favored the Ld phase (free energy of transfer of 17.2±1.0 kJ/mol), in line with eperiments and simulations [55]. The associated free energy changes were estimated by the Bennett acceptance ratio (BAR) method [63] implemented in the gmx bar tool of GROMACS, and the free energy of transfer was obtained as ΔGLoProt→LdProt−ΔGLo→Ld where the two terms correspond to this phase change in the presence and absence of the protein.
To study the generality of the effect of SDPE on the partitioning of helical multi-pass membrane proteins, we considered two additional brain-associated cases, with relation to DHA—dopamine D2 receptor (D2R) and glucose transporter GLUT1 (PDB id: 4PYP [64]), whose free energies of transfer were calculated in the absence of SDPE and in the presence of 16 mol%SDPE. The systems were set up identically to the ones containing A2AR, and the same equilibration and simulation protocols were followed. In the simulations, performed using GROMACS v5.0.x [65], the recently suggested “New-RF” simulation parameters [66] were employed. See Section A.2 in the S1 File for further details.
Finally, the free energies of transfer were also calculated for A2AR, D2R, and GLUT1 in the absence of and in the presence of 16 mol% SDPE in membranes whose compositions mimicked those of coexisting phases in model membranes (see Table A in the S1 File).
To study how DHA affects the adaptation of the protein into the membrane, we fine-grained the well-equilibrated CG systems containing 0, 4, and 8 mol% SDPE into all-atom resolution using the backward tool [67]. Additionally, we simulated membranes with identical lipid ratios yet in the absence of the protein as a control. All all-atom simulations, performed using GROMACS v5.0.x [65], employed the CHARMM36 force field [41, 42]. The last 150 ns of 200 ns simulations was used in the analyses. The default input parameters provided by CHARMM-GUI were used [68]. See Section A.3 in the S1 File for further details.
We studied whether certain protein types are more prone to be solvated by DHA in all-atom detail. To this end, we simulated four structurally different transmembrane proteins: 1) the transmembrane domain of the human receptor tyrosine kinase (ErbB1, PDB id: 2M0B), a single helix; 2) a dimer formed by two Glycophorin A peptides [44] (GpA dimer, PDB id: 1AFO); 3) A2AR (PDB id: 3EML) [45], a heptahelical bundle employed in the CG free energy calculation; and 4) the voltage-dependent anion channel (VDAC, PDB id: 3EMN) [46], a β-barrel. These proteins were embedded in a lipid bilayer consisting of equimolar concentrations of CHOL, dipalmitoyl-phosphatidylcholine (DPPC, two saturated chains; di-16:0), DOPC (two monounsaturated chains; di-18:1), dilinoleoyl-phosphatidylcholine (DLiPC, two diunsaturated chains; di-18:2), and stearoyl-docosahexaenoyl-phosphatidylcholine (SDPC, one saturated 18:0 chain and one polyunsaturated 22:6 (DHA) chain). The input structures for GROMACS were generated using the CHARMM-GUI Membrane Builder [68], and the systems were simulated for 4 μs using the input parameters provided by CHARMM-GUI [68]. The last 500 ns were used in the analyses. See Section A.4 in the S1 File for further details.
|
10.1371/journal.pmed.1002516 | Progression of diabetes, heart disease, and stroke multimorbidity in middle-aged women: A 20-year cohort study | The prevalence of diabetes, heart disease, and stroke multimorbidity (co-occurrence of two or three of these conditions) has increased rapidly. Little is known about how the three conditions progress from one to another sequentially through the life course. We aimed to delineate this progression in middle-aged women and to determine the roles of common risk factors in the accumulation of diabetes, heart disease, and stroke multimorbidity.
We used data from 13,714 women aged 45–50 years without a history of any of the three conditions. They were participants in the Australian Longitudinal Study on Women's Health (ALSWH), enrolled in 1996, and surveyed approximately every 3 years to 2016. We characterized the longitudinal progression of the three conditions and multimorbidity. We estimated the accumulation of multimorbidity over 20 years of follow-up and investigated their association with both baseline and time-varying predictors (sociodemographic factors, lifestyle factors, and other chronic conditions).
Over 20 years, 2,511 (18.3%) of the women progressed to at least one condition, of whom 1,420 (56.6%) had diabetes, 1,277 (50.9%) had heart disease, and 308 (12.3%) had stroke; 423 (16.8%) had two or three of these conditions. Over a 3-year period, the age-adjusted odds of two or more conditions was approximately twice that of developing one new condition compared to women who did not develop any new conditions. For example, the odds for developing one new condition between Surveys 7 and 8 were 2.29 (95% confidence interval [CI], 1.93–2.72), whereas the odds for developing two or more conditions was 6.51 (95% CI, 3.95–10.75). The onset of stroke was more strongly associated with the progression to the other conditions (i.e., 23.4% [95% CI, 16.3%–32.2%] of women after first onset of stroke progressed to other conditions, whereas the percentages for diabetes and heart disease were 9.9% [95% CI, 7.9%–12.4%] and 11.4% [95% CI, 9.1%–14.4%], respectively). Being separated, divorced, or widowed; being born outside Australia; having difficulty managing on their available income; being overweight or obese; having hypertension; being physically inactive; being a current smoker; and having prior chronic conditions (i.e., mental disorders, asthma, cancer, osteoporosis, and arthritis) were significantly associated with increased odds of accumulation of diabetes, heart disease, and stroke multimorbidity. The main limitations of this study were the use of self-reported data and the low number of events.
Stroke was associated with increased risk of progression to diabetes or heart disease. Social inequality, obesity, hypertension, physical inactivity, smoking, or having other chronic conditions were also significantly associated with increased odds of accumulating multimorbidity. Our findings highlight the importance of awareness of the role of diabetes, heart disease, and stroke multimorbidity among middle-aged women for clinicians and health-promotion agencies.
| In an aging population, it is common for women to experience two or more of diabetes, heart disease, and stroke.
Few published studies have investigated how women progress from a “healthy” state to having one of diabetes, heart disease, and stroke and then to multimorbidity.
No prospective evidence is available on the roles of time-varying common risk factors (i.e., high blood pressure or obesity) in the accumulation of multimorbidity from diabetes, heart disease, and stroke—information that may be important for health promotion and risk modification.
In this national prospective cohort study, 13,714 middle-aged Australian women (45–50 years old) were recruited in 1996 and have been followed for 2 decades. We collected data on their health conditions, including diabetes, heart disease, and stroke, as well as potential risk factors every 3 years until 2016.
From early to late middle age, many more women developed a single condition than multimorbidity. However, the odds ratio for accumulation of multimorbidity (i.e., from none or one to two or three, or from two to three of diabetes, heart disease, or stroke) was much higher than the odds of developing only one new condition, compared to women who did not develop any new condition over 3 years.
Nearly one-quarter of women who were initially diagnosed with stroke subsequently progressed to other conditions, a much higher percentage than those who were initially diagnosed with diabetes (9.9%) or heart disease (11.4%).
Women who were obese, had hypertension, were physically inactive, were smokers, or had other chronic conditions had higher odds of accumulating diabetes, heart disease, and stroke multimorbidity over the following 3-year period than women without these characteristics.
To our knowledge, this is the first study to delineate the accumulation of diabetes, heart disease, and stroke multimorbidity and investigate its association with both baseline and time-varying predictors in a prospective cohort study.
These findings suggest that health promotion, interventions for modifying lifestyle factors (obesity, high blood pressure, physical inactivity, and smoking), and treatment of other chronic conditions would be potentially beneficial for preventing the accumulation of diabetes, heart disease, and stroke multimorbidity.
For women who have had a stroke, health promotion, intervention, and treatment would be particularly important, as they appear most likely to progress to other conditions.
| The prevalence of diabetes, heart disease, and stroke multimorbidity (co-occurrence of two or three from these conditions) has increased rapidly over the past few decades [1–3], potentially translating to excess morbidity and mortality [4,5]. As the three conditions may interact with each other and be driven by common risk factors through the life course [6,7], the relationships among them are complex.
Diabetes is a well-established risk factor for heart disease and stroke, with a stronger effect in women than in men [8,9]. Although limited, there is evidence that patients with heart disease or stroke may progress to diabetes [10]. However, much less is known about how the three conditions progress from one to another sequentially through the life course [10]. The relationship between common risk factors (e.g., high blood pressure, obesity) and individual conditions has been researched intensively [6,7], but far less is known about the roles of these risk factors in the accumulation of diabetes, heart disease, and stroke multimorbidity from a longitudinal perspective [11].
Understanding the course of diabetes, heart disease, and stroke multimorbidity and related risk factors is essential for developing medical strategies to interrupt the progression from one condition to additional conditions. Considerable evidence has already shown that early intervention with diabetes or common risk factors might have the potential to ameliorate the course or even prevent the onset of cardiovascular disease (CVD) [12–14].
The aims of this article were (a) to investigate the progression of diabetes, heart disease, and stroke multimorbidity in middle-aged women over 20 years of follow-up and (b) to determine the roles of common risk factors in the accumulation of these conditions.
The Australian Longitudinal Study on Women’s Health (ALSWH) is an ongoing population-based cohort study that aims to investigate factors associated with health and well-being over time. The women were randomly selected from the national database of the Health Insurance Commission, the universal health insurance scheme that includes all citizens and permanent residents of Australia. Details of the study design, recruitment methods, and response rates have been described elsewhere [15,16]. The data analyses for the present study were performed following a prespecified analysis plan (S1 Text). Changes in the analysis plan were also described in S1 Text. This study is reported as per the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) guidelines (S1 Checklist).
The study was approved by the Human Research Ethics Committees of the Universities of Queensland and Newcastle. All participants signed informed consent, and all data used in the analyses were de-identified.
This study includes data from women born between 1946 and 1951, also known as the 1946–1951 cohort in ALSWH. A total of 13,714 women aged 45–50 years responded to the first survey in 1996. Response rates to the first mailed survey (baseline) cannot be exactly specified, as some women selected for the sample may not have received the invitation (e.g., if they had died or had changed their address without notifying the Health Insurance Commission). It is estimated that 53%–56% responded to the initial invitation to participate [16]. Self-administered questionnaires were sent to the women every 3 years (apart from a 2-year interval between the first and second surveys) until 2016.
Women who participated in at least two consecutive surveys with relevant information on exposures and outcomes of interest were included in the analysis (see Fig 1). Attrition rates and reasons for dropout at each survey are shown in S1 Table.
The main outcomes were the cumulative incidence of diabetes, heart disease, and stroke and accumulation of multimorbidity from these conditions. At each survey, women were asked “Have you ever been told by a doctor that you have diabetes (high blood sugar), heart disease (including heart attack, angina), and stroke over the past 3 years?” The three self-reported conditions were validated with hospital discharge data in a subset of the cohort (women living in New South Wales, Australia). The following International Statistical Classification of Diseases and Related Health Problems-Tenth Revision-Australian Modification (ICD-10-AM) diagnosis codes were used for the validation: diabetes mellitus (E10, E11, E13, and E14), ischemic heart diseases (I20–I25), and stroke (I60–I64). The prevalence and bias-adjusted kappa statistics for the three conditions were 0.93, 0.91, and 0.98, respectively [17].
The incidence of each of the conditions was based on the first report of that condition. Accumulation of multimorbidity was based on the first report of two or three of these conditions and the progression from two to three conditions.
Additionally, we decomposed multimorbidity into 4 unique outcomes that had developed over the 20-year follow-up: diabetes only, CVD only, CVD followed by diabetes, and diabetes followed by CVD [11]. CVD followed by diabetes refers to women with heart disease or stroke who subsequently developed comorbid diabetes. Diabetes followed by CVD refers to women with diabetes who subsequently developed comorbid heart disease or stroke.
There were four groups of covariates used in the analysis, and these covariates were collected at each survey unless indicated otherwise. Covariates included are study design variables: age in single years at 1996 (baseline), time period (Surveys 1–8); sociodemographic factors: country of birth, marital status, area of residence, education (at baseline), and ability to manage on income; lifestyle factors: body mass index (BMI), hypertension, physical activity, and smoking; and other chronic conditions: depression/anxiety, cancer, asthma, arthritis, osteoporosis, and chronic obstructive pulmonary disease (COPD). All covariates were time-varying except for country of birth and level of education.
BMI was calculated as weight in kilograms divided by height in meters squared and categorized as underweight (<18.5 kg/m2), normal weight (18.5–24.9 kg/m2), overweight (25–29.9 kg/m2), or obese (≥30 kg/m2), according to the World Health Organization classification [18]. Physical activity was categorized as sedentary (0–39 metabolic equivalent [MET] min/week), low (40–599 MET min/week), moderate (600–1,199 MET min/week), and high (≥1,200 MET min/week) [19].
Baseline characteristics were described by the number of conditions from diabetes, heart disease, and stroke developed during the 20-year follow-up. Differences among the groups were examined using analysis of variance (ANOVA) or chi-squared tests. A proportional Venn diagram was drawn to display the number of women with single or overlapping conditions. A Sankey diagram was constructed to characterize the dynamic changes of different combinations of these conditions over time. We estimated the association among the three conditions using repeated measures logistic regression of existing conditions on the incidence of each of the other two conditions, adjusted for age at baseline and time period.
To investigate the longitudinal odds of developing these conditions, we used a repeated measures logistic regression model fitted using generalized estimating equations. We calculated odds ratios (ORs) and 95% confidence intervals (CIs) for incidence one of the three conditions or two or three over each 3-year period, compared to a reference group of women who did not develop any new conditions. We also calculated cumulative incidence of the other two conditions after the first onset of each single condition.
To estimate the associations between predictors (sociodemographic factors, lifestyle factors, and other prior chronic conditions) in the progression of diabetes, heart disease, and stroke multimorbidity, we used a generalized linear mixed model for the multinomial outcome of cumulative incidence of 0, 1, or ≥2 conditions (≥2, including the transition from 2 to 3). We calculated ORs with 95% CIs for the association between the three outcomes at each survey and the risk factors (including time-varying covariates except for education and country of birth) at the previous survey; women with 0 conditions were the reference group. For example, the cumulative incidence of 1 or ≥2 conditions from Surveys 6–7 were modelled using predictors from Survey 6. We also used multinomial logistic regression to investigate the association between the three outcomes and the predictors at baseline (rather than time-varying).
We also used multinomial logistic regression to investigate the association between the predictors at baseline and four unique outcomes that had developed over the 20-year follow-up: diabetes only, CVD only, CVD followed by diabetes, and diabetes followed by CVD [11].
We conducted sensitivity analyses to check the robustness of our findings. All analyses were rerun using complete case data (i.e., only data from those women who responded to all 8 surveys).
All analyses were performed using SAS (version 9.4, SAS Institute Inc.). All statistical tests were two-sided, and P < 0.05 was considered to be statistically significant.
Of 13,714 women aged 45–50 years at baseline, 3.3% (n = 451) reported a history of any of the three conditions and were excluded. The final analytic sample comprised 11,941 women who provided information at two or more consecutive surveys from 1996 to 2016 (Fig 1).
Descriptive statistics for women at baseline, categorized according to the number of conditions developed during follow-up, are shown in Table 1. Over a follow-up period of 20 years, women who developed multimorbidity were more likely to be separated, divorced, or widowed; to have difficulty in managing on their available income; to be overweight or obese; to have hypertension; to have low levels of physical activity; to be smokers; and to have other chronic conditions (depression/anxiety, asthma, cancer, and arthritis) at baseline.
During the 20 years of follow-up, 2,511 (18.3%) of the women progressed to at least one condition, of whom 1,420 (56.6%) had diabetes, 1,277 (50.9%) had heart disease, and 308 (12.3%) had stroke; 423 (16.8%) had 2 or 3 of these conditions (Fig 2). As expected, there was an increase in the proportion of women with one or more conditions over time, culminating in 19.2% with 1 condition and 3.9% with multimorbidity at Survey 8. Proportions of women with each combination of these conditions at each survey are presented in S2 Table.
The cumulative incidence of the three individual conditions is presented in Fig 3, with 12.9%, 12.0%, and 2.7% of women progressing to diabetes, heart disease, and stroke by Survey 8, respectively. The three conditions were associated with each other (Fig 4). For example, compared with women without stroke, the OR for progressing to diabetes was 2.29 (95% CI, 1.49–3.52) and to heart disease was 3.84 (95% CI, 2.58–5.72) in women with stroke.
The odds for progressing to one condition and accumulation of multimorbidity increased over time, with the accumulation of multimorbidity accelerating between Surveys 4 and 5 (Fig 5). Over a 3-year period, the age-adjusted odds of two or more conditions were approximately twice that of developing one new condition, compared to women who did not develop any new conditions. For example, the odds for developing one new condition between Surveys 7 and 8 were 2.29 (95% CI, 1.93–2.72), whereas the odds for developing two or more conditions were 6.51 (95% CI, 3.95–10.75).
The Sankey diagram in Fig 6 illustrates the longitudinal progression and transitions among different combinations of diabetes, heart disease, and stroke in women who developed at least one of these conditions. This diagram shows the transitions from “healthy” state (no condition) to any one condition and then to multimorbidity. Over the 20 years of follow-up, the highest rate of progression from one condition to another was observed in women with stroke (Fig 7). The overall incidences of multimorbidity after the first onset of diabetes, heart disease, and stroke were 9.9% (95% CI, 7.9%–12.4%), 11.4% (95% CI, 9.1%–14.4%), and 23.4% (95% CI, 16.3%–32.2%), respectively.
After controlling for other covariates, being separated, divorced, or widowed; having lower education level; being born in another country; and having difficulties in managing on income were significantly associated with increased odds of accumulating multimorbidity (Table 2). A U-shaped relationship between BMI and the incidence of one condition and accumulation of multimorbidity was seen; obesity was associated with 2.6- and 3.0-fold increased odds of one condition and accumulation of multimorbidity (OR = 2.58, 95% CI, 2.28–2.92 and 3.01, 95% CI, 2.21–4.08, respectively). Compared with women without prior hypertension, the OR for progressing to one condition was 1.55 (95% CI, 1.40–1.71) and for accumulation of multimorbidity was 2.19 (95% CI, 1.74–2.75). Being sedentary and current smoking were also associated with increased odds of one condition and accumulation of multimorbidity, as were a number of other chronic conditions (Table 2).
Similar findings were observed based on multinomial logistic regression of the predictors at baseline (S3 Table). However, being sedentary, having depression/anxiety, asthma, cancer, or osteoporosis at baseline were not associated with the incidence of multimorbidity during follow-up using this simpler model.
Similar findings were also observed with the multinomial regression models of the different combinations of these conditions (S4 Table). Although obesity was consistently associated with multimorbidity, the odds varied depending on the condition sequence. For example, the OR for diabetes followed by CVD was 1.85 (95% CI, 1.25–2.73), whereas for CVD followed by diabetes it was 2.43 (95% CI, 1.28–4.29).
The trends of cumulative incidence of the three individual conditions and the accumulation of multimorbidity were similar when analyses were performed on the complete case data (N = 6,718 women), with 12.1%, 11.8%, and 2.9% of women progressing to diabetes, heart disease, and stroke at Survey 8, respectively (S1 Fig). The ORs for the incidence of one condition and accumulation of multimorbidity among sociodemographic and lifestyle factors and other chronic conditions were also consistent in complete cases, compared with the main results using all cases (S5 Table).
We found that the age-adjusted odds of two or more conditions were approximately twice that of developing one new condition compared to women who did not develop any new conditions, and the onset of stroke is associated with increased risk of progression to heart disease and diabetes. Furthermore, being separated, divorced, or widowed; being born outside Australia; having difficulties with management on income; being overweight or obese; being physically inactive; having hypertension; currently smoking; and having other chronic conditions (i.e, mental disorders, asthma, cancer, osteoporosis, and arthritis) were also associated with this progression.
There has been an extensive discussion in the literature on the associations of diabetes with heart disease and stroke [5–9,12,13,20,21]. However, the longitudinal progression of these conditions remains unclear. To our knowledge, this is the first study to delineate the progression of diabetes, heart disease, and stroke multimorbidity.
The Emerging Risk Factor Collaboration of 91 cohort studies showed that cardiometabolic multimorbidity (defined as a history of two or three from diabetes, stroke, myocardial infarction) is associated with increased mortality risk (hazard ratios for mortality of only one condition, a combination of any two, and a combination of all three were about 2, 4, and 8, respectively), suggesting that development of cardiometabolic conditions accelerates progression to death [4]. In this study, we focused on the early stages of the disease course to investigate progression to individual conditions and multimorbidity. Our results showed that progression to cardiometabolic conditions accelerates as women age, in particular after their mid-50s (Fig 5). The findings of these two studies collectively suggest that interventions that could slow progression to cardiometabolic conditions may have the potential to lead to significant increases in life expectancy.
The relationships among the three conditions are complex. Evidence from systematic reviews suggests that compared with men with diabetes, women with diabetes have a greater risk of coronary heart disease and stroke [8,9]. One population-based cohort study suggested that one-third of patients with heart disease developed new-onset diabetes or impaired fasting glucose during 3.5 years of follow-up [10]. Our study is consistent with these findings. Furthermore, we found that all three conditions were associated with each other, and the onset of stroke may be associated with the progression to other conditions. With self-reported data, a potential explanation for the onset of diabetes following heart disease or stroke is that these women had undiagnosed diabetes (or “prediabetes”) before their heart disease or stroke event. However, statins are recommended for secondary prevention following heart disease or stroke in Australia [22]; hence, heart disease or stroke could have led to statin use, which may have increased the risk of diabetes. Our previous study using data from the 1921–1926 cohort of ALSWH found a 33% increased risk of new-onset diabetes associated with statin exposure [23].
A pooled analysis of individual-level data from 16 cohort studies suggested that overweight and obesity (at baseline) were strongly associated with the incidence of cardiometabolic multimorbidity (at least two from type 2 diabetes, coronary heart disease, and stroke) during a mean follow-up of 10.7 years [11]. Our study is consistent with this study, using multiple follow-up intervals of 3 years and treating BMI as a time-varying variable. However, we found smaller ORs for obesity at baseline on the different combinations of conditions during follow-up. In our study, the ORs associated with obesity varied between 1.10 and 2.43 (95% CI varied between 0.93 and 4.29) depending on the combination of conditions, compared to the previously reported ORs that ranged from 1.5 to 29.8 (95% CI varied between 1.2 and 40.8). Possible explanations for this difference are that the participants in our study were women aged 45–50 at baseline, and they were followed up for 20 years, whereas the previous study included both men and women with a wider age range of 35–105 and a shorter period of follow-up (10.7 years).
The relationship between hypertension and individual diabetes, heart disease, and stroke is well established [24,25], but the results are controversial when considering the sequences of these conditions [24]. We found that women with prior hypertension are more like to progress to multimorbidity. This finding suggests that hypertension might play a major role in the longitudinal progression of the three conditions over time [25]. There have been few studies that have investigated the associations of other chronic conditions with diabetes, heart disease, and stroke multimorbidity [21,26]. A cross-sectional study using data from the UK Biobank found rheumatoid arthritis was associated with the increasing numbers of the three conditions [21], a result consistent with our findings.
A strength of this study is the large, nationally representative sample and the longitudinal nature of the data, which allowed us to document disease progression over time [27]. Also, as the onset and experience of chronic conditions often occur during midlife [28], the use of a cohort of middle-aged women who were initially healthy reflects the “real-world experience” of disease progression, particularly in the early stages of the course of disease [29]. Furthermore, we used time-varying lifestyle factors and other chronic conditions to capture the dynamic changes of these predictors over time [30]. This may be important, as illustrated by the differing results for the effect of physical activity on the incidence of multimorbidity depending on whether this variable was treated as time-varying or fixed at baseline.
Several methodological limitations of the study should also be noted. First, diabetes, heart disease, stroke, and other chronic conditions are based on self-report, which may have introduced bias in outcome ascertainment. However, the cumulative incidence of the three conditions at the age of 65–70 (Survey 8) in our study population was almost the same as that for women of this age reported in the Australian National Health Survey (2014–2015) [31]. Furthermore, previous studies have validated self-report of various chronic diseases in the ALSWH. For example, the prevalence and bias-adjusted kappa for diabetes, heart disease, stroke, and hypertension from hospital data were 0.93, 0.91, 0.98, and 0.53, respectively [17]. Cancer information included in the study has been validated against Cancer Registry data with 89% sensitivity and 97% specificity [32]. Several other studies have demonstrated the validity of self-reported conditions, BMI, physical activity, and other risk factors [19,33–38].
Second, the small number of cases with more than one condition, potential undiagnosed cases of diabetes, and unknown information on heart disease and stroke type [39,40] might all have led to increased uncertainty and potential underestimation of the odds obtained from the regression models. Third, although we conducted a sensitivity analysis to check the robustness of our findings with complete case data, we did not account for the competing risk of women dying before they could potentially progress to one or more of diabetes, heart disease, and stroke [41]. However, we compared the mortality of all participants and of just the women who developed the three conditions, and the proportions were quite similar and low (4.63% [553/11,941] versus 4.70% [118/2,511], respectively). Fourth, although the attrition rates were low during the 20-year follow-up (S1 Table), there is potential for bias in the estimates presented due to attrition or restricting the analysis to complete cases. Fifth, the study sample was women aged 45–50 at baseline, which limits the generalizability of the findings to other groups. However, the study sample is broadly representative of all women born in 1945–1950 in Australia [15]. Finally, the inclusion of information on treatments in the analysis may have influenced the results.
Further studies with a large number of validated cases and for women and men in different age groups are needed to establish the generalizability of these findings and to explore interventions that might slow cardiometabolic progression [27,42].
Our findings indicate that the odds for accumulation of diabetes, heart disease, and stroke multimorbidity was higher than the odds of progressing to one condition in middle-aged women. Stroke was associated with increased risk of progression to diabetes and heart disease. Social inequality, overweight and obesity, hypertension, physical inactivity, smoking, and having prior chronic conditions (i.e., mental disorders, asthma, cancer, osteoporosis, and arthritis) are all associated with increased risk of accumulation of multimorbidity.
These findings could have significant clinical and public health implications for the treatment and prevention of diabetes, heart disease, and stroke multimorbidity. Delineation of the disease progression may assist in the evaluation of risk for additional conditions during clinical practice. The identified risk factors from this study are appropriate targets for reducing the risk of multimorbidity.
|
10.1371/journal.ppat.1007530 | Recruitment of Vps34 PI3K and enrichment of PI3P phosphoinositide in the viral replication compartment is crucial for replication of a positive-strand RNA virus | Tombusviruses depend on subversions of multiple host factors and retarget cellular pathways to support viral replication. In this work, we demonstrate that tomato bushy stunt virus (TBSV) and the closely-related carnation Italian ringspot virus (CIRV) recruit the cellular Vps34 phosphatidylinositol 3-kinase (PI3K) into the large viral replication compartment. The kinase function of Vps34 is critical for TBSV replication, suggesting that PI(3)P phosphoinositide is utilized by TBSV for building of the replication compartment. We also observed increased expression of Vps34 and the higher abundance of PI(3)P in the presence of the tombusviral replication proteins, which likely leads to more efficient tombusvirus replication. Accordingly, overexpression of PI(3)P phosphatase in yeast or plants inhibited TBSV replication on the peroxisomal membranes and CIRV replication on the mitochondrial membranes. Moreover, the purified PI(3)P phosphatase reduced TBSV replicase assembly in a cell-free system. Detection of PI(3)P with antibody or a bioprobe revealed the enrichment of PI(3)P in the replication compartment. Vps34 is directly recruited into the replication compartment through interaction with p33 replication protein. Gene deletion analysis in surrogate yeast host unraveled that TBSV replication requires the vesicle transport function of Vps34. In the absence of Vps34, TBSV cannot efficiently recruit the Rab5-positive early endosomes, which provide PE-rich membranes for membrane biogenesis of the TBSV replication compartment. We found that Vps34 and PI(3)P needed for the stability of the p33 replication protein, which is degraded by the 26S proteasome when PI(3)P abundance was decreased by an inhibitor of Vps34. In summary, Vps34 and PI(3)P are needed for providing the optimal microenvironment for the replication of the peroxisomal TBSV and the mitochondrial CIRV.
| Replication of RNA viruses infecting various eukaryotic organisms is the central step in the infection process that leads to generation of progeny viruses. The replication process requires the assembly of numerous viral replicase complexes within the large replication compartment, whose formation is not well understood. Using tombusviruses and the model host yeast, the authors discovered that a highly conserved cellular lipid kinase, Vps34 phosphatidylinositol 3-kinase (PI3K), is critical for the formation of the viral replication compartment. Expression of PI3K mutants and the PI(3)P phosphatase revealed that the PI(3)P phosphoinositide produced by Vps34 is crucial for tombusvirus replication. Tombusviruses co-opt Vps34 through interaction with the viral replication protein into the replication compartment. In vitro reconstitution of the tombusvirus replicase revealed the need for Vps34 and PI(3)P for the full-activity of the viral replicase. Chemical inhibition of Vps34 in yeast or plants showed that PI(3)P is important for the replication of several plant viruses within the Tombusviridae family and the insect-infecting Nodamuravirus. These results open up the possibility that the cellular Vps34 PI3K could be a target for new antiviral strategies.
| Positive-strand RNA viruses replicate inside the infected plant or animal cells by utilizing subcellular membranes and co-opting multiple host proteins. These viruses generate membranous viral replication compartments, often harboring numerous vesicle-like membrane invaginations with narrow openings towards the cytosol [1–5]. The viral replication compartment help sequestering viral proteins, viral RNAs and co-opted host factors in confined areas, which facilitate efficient viral replicase complex (VRC) assembly and robust viral RNA replication. The replication compartment also protects the viral RNA from cellular defense mechanisms [5,6]. In spite of intensive research on the formation of viral replication compartments, it is still incompletely understood how the VRC assembly process is guided by viral and host factors.
The plant-infecting tombusviruses, such as Tomato bushy stunt virus (TBSV), induce complex rearrangements of cellular membranes, alter metabolic processes with the help of a number of co-opted host proteins [7–9]. The pro-viral host proteins include protein chaperones, translation elongation factors, DEAD-box helicases, glycolytic enzymes, the actin network, and cellular membrane remodeling proteins, such as the endosomal sorting complex required for transport (ESCRT) machinery [3,8,10–12]. TBSV also exploits sterols and phospholipids to induce a membranous replication compartment harboring numerous spherules, which are vesicle-like invaginations in the peroxisomal membranes [7,13–15].
Tombusviruses have a wide host range and they are among the best-characterized viruses [16–19]. They have one component (+)RNA genome of ~4.8 kb [20]. They belong to Flavivirus-like supergroup that includes important human, animal and plant pathogens. Tombusviruses code for five proteins including two essential replication proteins, p33 and p92pol, which is the RdRp protein and it is translated from the genomic RNA via readthrough of the translational stop codon in p33 ORF. The second replication protein, p33, is an RNA chaperone involved in recruitment of the viral (+)RNA for replication [20–22]. The TBSV replicon (rep)RNA, which is based on DI-72 RNA, contains four non-contiguous segments from the gRNA, can replicate efficiently in yeast and plant cells expressing p33 and p92pol [20,23]. The replication of repRNA, which produces a double-stranded RNA replication intermediate, occurs in vesicle-like structures, called spherules in cells [7,8].
Intriguingly, tombusviruses take advantage of various cellular compartments for VRC assembly [24]. TBSV and the closely related cucumber necrosis virus (CNV) use peroxisomal membranes, whereas carnation Italian ringspot virus (CIRV) utilizes the outer membranes in mitochondria. The ER could support TBSV replication efficiently in the absence of peroxisomes in yeast [25,26]. Moreover, the formation of membrane contact sites (MCSs) between the ER and peroxisomes promote sterol-enrichment at replication sites [13,27]. Also, TBSV hijacks the Rab5-positive endosomes to build large replication compartments in yeast and plant cells [28].
Unlike mammals, yeast and plants have only one phosphatidylinositol (PI) 3-kinase (PI3K), namely Vps34, which produces phosphatidylinositol-3-phosphate, PI(3)P, a critical signaling and structural lipid molecule [29,30]. The multiple functions of Vps34 are to facilitate the formation of early endosomes involved in protein secretion and recycling; and autophagic structures involved in protein/lipid recycling during starvation [31–33]. Vps34 PI3K has attracted immense interest due to its role in human diseases, such as numerous forms of cancer, heart problems and neurodegeneration in humans [34,35]. The best-known role of Vps34 PI3K is in hepatitis C virus (HCV) replication [36–38]. Also, Rab5 GTPase and Vps34 are involved in HCV NS4B-induced autophagy [37]. Many DNA and RNA viruses are also influenced by the PI3K signaling pathway [39,40]. Since (+)RNA viruses remodel subcellular membranes to facilitate virus replication and avoid antiviral responses [3,9,41–44], it is possible that subversion of Vps34 PI3K and PI(3)P could be wide-spread among viruses.
Although PI(3)P is a minor lipid in the cell, it is a key player in endosomal vesicle trafficking by conferring identity to endosomes [29,30]. Also, PI(3)P plays a crucial role in regulating vesicle fusion and autophagosome formation. The accumulation of PI(3)P helps the recruitment of its numerous protein effectors. Many intracellular microbes and parasites exploit the cellular PI(3)P to establish infections. For example, the SopB effector of the Salmonella bacterium recruits Vps34 to the bacteria-containing vacuole, leading to enrichment of PI(3)P and maturation of the bacteria-containing vacuole [45]. Mycobacterium tuberculosis, Phytophtora and Plasmodium parasites also exploit PI(3)P to regulate endosomal functions [45,46]. In addition, elevated PI(3)P level induced artemisinin resistance in malaria parasites [47]. All these examples highlight the central roles of Vps34 and PI(3)P in microbe-host intracellular interactions.
In our previous paper, we showed the recruitment of the Rab5-positive endosomes to the large replication compartment formed in tombusvirus-infected cells [28]. Through specific interaction of the p33 replication protein with Rab5 small GTPase, tombusviruses enrich endosomal lipids, most importantly phosphatidylethanolamine (PE), but also PI(3)P, in peroxisomes or mitochondria for different tombusviruses. This raised the question if PI(3)P plays a role in tombusvirus replication. Accordingly, in this paper we show that TBSV hijacks Vps34p PI3K that leads to enrichment of PI(3)P in the large viral replication compartment in model yeast and plant hosts. Altogether, the direct interaction between p33 replication protein and Vps34p is needed for the biogenesis of the replication compartments.
The recruitment of the components of the early endosome to the tombusvirus replication compartment leads to enrichment of PI(3)P within the replication sites [28], suggesting that PI(3)P phosphoinositide might be involved in the viral replication process. To study the putative role of PI(3)P in tombusvirus replication, first we analyzed if Vps34 PI3K, which is the only PI(3)P kinase in yeast [31,48,49], affects TBSV replication in yeast. We launched TBSV replication in a yeast strain lacking VPS34 gene (vps34Δ) by expressing the p33 and p92pol replication proteins and the repRNA from plasmids, followed by measuring TBSV repRNA level by Northern blotting. These experiments demonstrated that TBSV replicated only at a ~6% level in the absence of Vps34 protein when compared to the replication level in the WT yeast (Fig 1A, lanes 13–15 versus 1–3). Interestingly, the p33 replication protein accumulated poorly in vps34Δ yeast (Fig 1A). Complementation of tombusvirus replication by expression of wt Vps34p from a plasmid in vps34Δ yeast supported 2-fold higher level of viral repRNA accumulation than in WT yeast (Fig 1A, lanes 16–18), demonstrating pro-viral function for TBSV replication. Interestingly, expression of two inactive forms of Vps34 (N736K and D749E) [49,50], which are defective in producing PI(3)P, in vps34Δ yeast could not complement the pro-viral function (Fig 1A, lanes 19–24), suggesting that PI(3)P is required for TBSV replication. Moreover, the expression of the two inactive mutant forms of Vps34 inhibited TBSV repRNA accumulation in WT yeast (Fig 1A, lanes 7–12 versus 1–3), likely due to the competition between the WT Vps34p expressed from the chromosome and the Vps34p mutants expressed from plasmids to participate in protein complexes (see below).
To test if the closely-related CIRV, which replicates on the outer mitochondrial membranes, also requires Vps34p, we measured repRNA replication when the CIRV replication proteins were expressed in vps34Δ yeast. The accumulation of repRNA in yeast decreased by ~30-fold in the absence of Vps34p (Fig 1B, lanes 13–15), confirming that CIRV also requires the pro-viral functions of Vps34p. Expression of the plasmid-borne Vps34p in vps34Δ yeast increased repRNA accumulation by ~3.5-fold in comparison with wt yeast (Fig 1B, lanes 16–18), whereas N736K and D749E mutants of Vps34p behaved as dominant negatives in viral replication when expressed in WT yeast (Fig 1B, lanes 7–12). Both p36 and p95pol replication proteins showed reduced accumulation in vps34Δ yeast (Fig 1B). Overall, we conclude that Vps34 provides critical pro-viral functions for both the peroxisomal TBSV and the mitochondrial CIRV replication in yeast cells.
To obtain further evidence on the pro-viral role of Vps34p, we used a specific inhibitor (AS604850) [51] to block Vps34 activity in yeast cells. We found ~8-fold reduction in TBSV repRNA accumulation when the highest concentration of the inhibitor was used (Fig 2A). Similar to the deletion of VPS34, chemical inhibition of Vps34p activity also resulted in reduced accumulation of p33 replication protein in WT yeast (Fig 2A). Similar studies using AS604850 inhibitor of Vps34p showed that replication of CIRV, the closely related cucumber necrosis virus (CNV) and the unrelated Nodamura virus (NoV, an insect-infecting alfanodavirus) is also dependent on Vps34p functions in yeast cells (S1 Fig). The NoV RNA1 replication was more sensitive to AS604850 inhibitor than the transcription of subgenomic RNA3, which was decreased only when the largest amount of inhibitor was applied (S1C Fig)
To test the fate of p33 replication protein when Vps34 activity is blocked, we used MG132 inhibitor to block the function of the 26S proteasome in yeast [52]. We found a 40% increased level of p33 replication protein in MG132-treated versus DMSO-treated wt yeast spheroplasts (Fig 2B). Moreover, MG132-treatment also reversed the negative effect of AS604850 inhibitor on the p33 level by resulting in ~5-fold higher level p33 in yeast spheroplasts (Fig 2B), suggesting that p33 becomes degraded by the proteasome when Vps34 function is inhibited and PI(3)P is not produced in the viral replication compartment.
Testing the activity of the in vitro assembled tombusvirus replicase based on purified recombinant replication proteins in yeast cell-free extracts (CFEs) revealed ~5-fold reduced activities for CFEs obtained from vps34Δ yeast in comparison with wt yeast CFEs (Fig 2C). The accumulation of both the dsRNA replication intermediate and the newly-made (+)RNA progeny decreased in CFEs prepared from vps34Δ yeast, indicating that Vps34 activity is likely required during the replicase complex assembly step in vitro. This was further supported by another CFE-based replicase assembly experiment when WT yeast was grown in the presence of Wortmannin, an inhibitor of Vps34p (Fig 2D). These results also suggest that Vps34p PI3K activity is required not only for maintaining p33 replication protein level in yeast, but during the actual replication process as well. This is because we provided the same amounts of purified recombinant replication proteins for the above CFE-based assays (Fig 2C and 2D) [53]. Thus, these in vitro results with Vps34p are different from the Rab5-based studies [28], suggesting that the pro-viral role of Vps34p is more complex than that of co-opted Rab5 and the endosomes in tombusvirus replication.
To explore if tombusviruses depend on Vps34p functions in plants, first we knocked down Vps34 level based on gene-silencing in N. benthamiana plants. Similar to yeast, plants also have only a single PI3K kinase, namely Vps34 [54]. Knockdown of Vps34 in N. benthamiana led to ~5-fold reduction of TBSV genomic (g)RNA and ~3-fold reduction in CIRV RNA accumulation, respectively (Fig 3A–3C, lanes 4–6). These findings confirmed the pro-viral function of Vps34 in tombusvirus replication. Knockdown of Vps34 delayed the symptom formation and necrosis in young leaves infected with TBSV or CIRV (Fig 3B–3D). Second, we measured the amount of Vps34p, which is increased by 60% from its native promoter and original chromosomal location upon expression of p33 and p92 replication proteins in WT yeast (Fig 3E). In addition, we analyzed Vps34 mRNA levels in TBSV-infected versus mock-treated Nicotiana benthamiana leaves. RT-PCR results showed up-regulation of Vps34 mRNA level in TBSV-infected leaves (Fig 3F), suggesting that TBSV replication induces increased Vps34 expression in plants and yeast.
To learn if Vps34p lipid kinase is co-opted by TBSV for supporting its replication, first, we co-expressed RFP-tagged Vps34p with GFP-tagged TBSV p33 replication protein in wt yeast cells, followed by confocal imaging. These experiments revealed partial co-localization of TBSV p33 replication protein and Vps34 (71±26% co-localization from the p33 point of view, Fig 4A). Partial co-localization of FLAG-tagged Vps34 and p33 (using anti-p33 antibody) in yeast was also observed using super-resolution microscopy (Fig 4B). Co-expression of Vps34-RFP with Pex13-GFP peroxisomal marker protein in the presence of replicating TBSV replicon (rep)RNA in yeast also showed partial co-localization pattern (Fig 4C), whereas Vps34-RFP and Pex13-GFP did not co-localize in the absence of viral components (Fig 4D). The co-localization of Pex13-GFP and p33 replication protein is complete within the viral replication compartment consisting of aggregated peroxisomes [55]. In addition, similar partial co-localization pattern was observed when the GFP-p36 replication protein of the closely related CIRV, which localizes to the outer mitochondrial membranes [56], was co-expressed with Vps34-RFP in yeast cells (Fig 4E). Based on these experiments, we conclude that Vps34p lipid kinase is partially retargeted by tombusvirus replication proteins to the tombusvirus replication compartment in yeast.
To confirm that comparable outcomes take place in the native plant cells, we co-expressed TBSV p33-BFP with the Arabidopsis AtVps34 and RFP-SKL (peroxisomal luminar marker protein) in N. benthamiana leaves infected with TBSV. Confocal microscopy imaging revealed the partial co-localization of AtVps34 with p33-BFP and RFP-SKL (Fig 5A and 5B), suggesting that Vps34 is partially recruited into the large TBSV replication compartment in plant cells. On the contrary, the peroxisomes did not show significant co-localization with AtVps34 in the absence of TBSV infection (S2 Fig).
To confirm that the plant Vps34 is recruited into the viral replication compartment through the interaction with the TBSV p33 or CIRV p36 replication proteins, we have performed bimolecular fluorescence complementation (BiFC) experiments in N. benthamiana leaves. The BiFC assay using confocal microscopic imaging revealed the interaction between Vps34 and p33/p36 replication proteins within the replication compartment, consiting of either aggregated peroxisomes or mitochondria (detected with the help of peroxisomal RFP-SKL and the mitochondrial CoxIV-RFP) (Fig 5C and 5D).
To further confirm the interaction between the plant Vps34 and the TBSV replication protein, we agroinfiltrated N. benthamiana leaves to express the Flag-tagged AtVps34, followed by inoculation of the agroinfiltrated leaves with sap containing TBSV. Flag-affinity purification of AtVps34 from the detergent-solubilized membrane fraction resulted in detection of the co-purified p33 replication protein via anti-p33 antibody using Western blotting (Fig 6A, lane 2).
Since the above experiments indicated that Vps34 is recruited to the viral replication compartment in yeast, we also tested interaction between the replication proteins and the yeast Vps34p using co-purification experiments with Flag-tagged Vps34 and His6-taged TBSV replication proteins from yeast. Western-blot analysis of the Flag-affinity purified Vps34 samples showed the efficient co-purification of both p33 and p92 replication proteins from the membrane fraction of yeast (Fig 6B). To exclude that Vps34p could only be part of the tombusvirus replicase complex when over-expressed, we performed the reciprocal approach by Flag-affinity purification of Flag-p33 from the membrane fraction of yeast expressing Vps34-3xHA from its natural promoter and the original chromosomal location in wt yeast. These experiments confirmed that Vps34-3xHA could be co-purified with the TBSV replicase (Fig 6C, lane 2). Thus, the co-purification experiments supported interaction between the tombusvirus replication proteins and Vps34p PI3K.
We then performed pull-down experiments with purified recombinant p33 replication protein, which also supported direct interaction with Vps34p expressed in yeast (Fig 6D). The kinase inactive mutants of Vps34p bound to p33 in the pull-down assay, suggesting that the kinase activity is not needed for this interaction (Fig 6D). To test if the interaction between p33 replication protein and Vps34p depends on Rab5, which is also present in the early endosomes and interacts with p33 [28], we performed pull down experiments with purified recombinant MBP-p33 and GST-Vps34p expressed in E. coli, which does not have Rab5. These experiments also showed direct interaction between p33 and Vps34p and the kinase inactive mutant of Vps34p (Fig 6E). Vps21p, a Rab5 protein in yeast was co-purified with FLAG-p33 from the membrane fraction of yeast lacking VPS34 gene, suggesting that p33 replication protein could interact separately with Vps21p and Vps34p (S3 Fig). Altogether, all these data support direct interaction between the tombusvirus replication proteins and Vps34p that results in partial recruitment of Vps34p into the viral replication compartment in both yeast and plant cells.
To examine if Vps34p is a permanent component of the tombusvirus replicase, we shut down the formation of new tombusvirus replicase complexes by stopping ribosomal translation through adding cycloheximide to the yeast growth media [8]. Flag-affinity-purification of the tombusvirus replicase from the membrane fraction of yeast at various time-points showed the decreasing amounts of the co-purified Vps34p in the purified replicase preparations (Fig 6F, lanes 3–4 versus 2). Therefore, Vps34p seems to be released from the replicase, indicating that Vps34p is a temporarily co-opted host factor in the tombusvirus replicase complex.
Based on the recruitment of Vps34p PI3K into the viral replication compartment, we assumed that the replication compartment is enriched in PI(3)P. Accordingly, detection of PI(3)P with anti-PI(3)P antibody in yeast cells replicating TBSV repRNA using confocal microscopy revealed the enrichment of PI(3)P at replication sites, which were also decorated with the peroxisomal marker (Fig 7A and S4A Fig). In the absence of TBSV components, PI(3)P was not co-localized with the peroxisomal marker (Fig 7A, bottom panel, and S4A Fig). These observations were confirmed by using a biosensor (RFP-2xFYVE), which specifically binds to PI(3)P in cells [30,57]. Expression of the RFP-tagged 2xFYVE protein domain showed partial co-localization of PI(3)P and GFP-p33 replication protein in the large replication compartment in yeast cells, whereas PI(3)P was co-localized with the Rab5-decorated (Vps21 in yeast) endosome in the absence of TBSV components (Fig 7B).
Because Vps34 level is increased in yeast and plants (Fig 3E and 3F), we also tested if PI(3)P amount is changed due to tombusvirus replication in yeast. Interestingly, we observed 2.5x fold increase in PI(3)P level in yeast replicating TBSV repRNA (Fig 7C). Thus, the increased level of Vps34p results in higher abundance of PI(3)P, which likely promotes more efficient tombusvirus replication.
Studying the PI(3)P distribution via either anti-PI(3)P antibody or RFP-tagged FYVE protein domain in N. benthamiana plant cells infected with TBSV also showed PI(3)P enrichment within the large replication compartment (Fig 7D and 7E, and S4B Fig). On the contrary, PI(3)P was not co-localized with the peroxisomal marker and it was co-localized with Rab5B (early endosome) in plant cells not infected with TBSV. Based on these results, we conclude that PI(3)P is enriched in the TBSV replication compartment in yeast and plant cells during TBSV replication.
To test if PI(3)P has a proviral role during TBSV replication, we over-expressed the yeast Ymr1p PI(3)P phosphatase, which converts PI(3)P to PI [58]. Over-expression of Ymr1p in wt yeast reduced TBSV repRNA replication by more than 50%, whereas CIRV replication was inhibited by ~3-fold (Fig 8A and 8B), suggesting that PI(3)P is needed for both TBSV and CIRV replication. On the contrary, over-expression of Ymr1p in vps34Δ yeast, which cannot synthesize PI(3)P, does not affect TBSV or CIRV replication when compared with the accumulation level of repRNA in vps34Δ yeast (Fig 8A and 8B). We found that over-expression of Ymr1p reduced PI(3)P level by ~40% in yeast, based on immunofluorescense assay with antibody against PI(3)P (Fig 8C). Over-expression of Ymr1p in either pex3Δ or pex19Δ yeasts, which are deficient in peroxisome biogenesis and support TBSV replication via the ER membrane [26,59], led to 50–60% reduction in TBSV repRNA replication (S5 Fig). Thus, PI(3)P is required for tombusvirus replication in different subcellular locations.
We also expressed PI(3)P-binding proteins (FYVE and PX) [30], that by binding to PI(3)P, might sequester PI(3)P, thus this lipid would not be readily available for supporting TBSV replication. Indeed, these proteins inhibited TBSV replication by 40-to-60% when expressed in yeast, whereas the mutated version of PX [46,60] with reduced PI(3)P-binding did not inhibit TBSV replication in yeast (Fig 8D). We cannot exclude that the steric hindrance effect of the FYVE and PX proteins recruited to the PI(3)P-rich replication compartment also contributes to the inhibitory effect of these proteins on tombusvirus replication. Moreover, expression of the Arabidopsis ortholog of the yeast Ymr1, called Mtm1, also inhibited TBSV and CIRV replication in N. benthamiana plants by ~70–80% (Fig 8E and 8F).
To obtain further evidence that PI(3)P has pro-viral function during TBSV replicase reconstitution in vitro, we affinity-purified Flag-Ymr1p from yeast, followed by pre-incubation with CFE obtained from WT yeast. Then, we added (+)repRNA and the purified recombinant TBSV replication proteins to the CFE-based replicase reconstitution assay and measured repRNA replication. We observed a ~40% reduction in repRNA replication when the CFEs were pre-incubated with Flag-Ymr1 in comparison with pre-incubation of CFE with the buffer (Fig 8G). The accumulation of both the dsRNA replication intermediate and the newly-made (+)RNA progeny decreased in CFEs pre-incubated with Flag-Ymr1, indicating that PI(3)P function is likely required during the replicase complex assembly step in vitro.
Deletion of YMR1 in yeast did not affect TBSV or CIRV replication (S6A and S6B Fig). Similarly, CFE prepared from ymr1Δ yeast supported TBSV repRNA replication close to wt level (S6C Fig), indicating that Ymr1p has no antiviral activities in yeast. All these data point at PI(3)P as a proviral host component during tombusvirus replication.
Vps34p forms at least four different complexes in yeast with different functions [48]. While Vps15p, the activator of Vps34p, is present in all these complexes, Atg14p is a unique component in the signaling complex regulating autophagy, whereas Vps38p is needed for the endosome trafficking function of Vps34p complex [61]. One of the two yeast Vps34p complexes regulating endosome trafficking also contains Beclin1-related VPS30 (also called ATG6).
To gain insights into the pro-viral functions of Vps34p and PI(3)P, we have analyzed TBSV repRNA accumulation in yeast missing proteins that form complexes with Vps34p. Deletion of VPS15 reduced TBSV or CIRV repRNA accumulation by ~4-fold (Fig 9A and 9B, lanes 9–10). Deletion of VPS38 component of the Vps34/Vps15p complex reduced TBSV or CIRV repRNA accumulation as much as deletion of VPS34 (Fig 9A and 9B, lanes 13–14). The effect of deletion of Beclin1-related VPS30 was dramatic in TBSV replication, but lesser in CIRV repRNA accumulation (Fig 9). Deletion of VPS15, VPS30 (Beclin1) and VPS38 also resulted in reduced accumulation of the p33 and p36 replication proteins of TBSV and CIRV (Fig 9A and 9B), suggesting that the pro-viral functions of Vps34p are performed by the four-component Vps34/Vps15/Vps30/Vps38 complex. Based on these data, we suggest that the vesicle trafficking function of Vps34 PI3K is important for tombusvirus replication.
Vps34p is also involved in providing PI(3)P for the initiation of autophagy, a recycling mechanism for the cells [31,33,34]. However, we found that deletion of the critical autophagy genes, such as ATG1, ATG8 and ATG14, the latter which participate in autophagic signaling complex formation with Vps34p, did not affect TBSV or CIRV RNA accumulation in yeast under the given growth conditions (Fig 9). Similarly, deletion of additional autophagy genes ATG5, ATG7 and ATG12 did not have a major effect on TBSV replication in yeast (S6 Fig). Further experiments will be needed to analyze the roles of the autophagy genes under different conditions in tombusvirus replication. Altogether, these experiments indicated that the Vps34/Vps15/Vps30/Vps38 complex, involved in endomembrane trafficking is essential for tombusvirus replication in yeast.
Rab5 GTPase is important for the maturation of the early endosome and it interacts with Vps34p to perform this function [62,63]. Moreover, the p33 replication protein interacts with Rab5 that leads to the recruitment of the early endosome to the replication compartment [28]. Therefore, we tested if Rab5 facilitates PI(3)P production within the replication compartment. We found that deletion of all three Rab5 genes (vps21Δypt52Δypt53Δ) in yeast interfered with the production of PI(3)P within the replication compartment based on the lack of co-localization of RFP-2xFYVE and GFP-p33 (Fig 10A).
Previously, we found that a major pro-viral function of Rab5, whose deletion results in ~20% TBSV replication [28], is to facilitate the enrichment of PE in the viral replication compartment [28]. Vps34p and PI(3)P might be needed for this process due to their documented functions with Rab5. Accordingly, deletion of VPS34 in yeast prevented the enrichment of PE within the replication compartment (Fig 10B and S7 Fig). Similarly, knockdown of Vps34 level in N. benthamiana resulted in poor enrichment of PE within the replication compartments, which were visibly smaller than those in the control plants (Fig 10C). Based on these data, we suggest that Vps34p is involved in facilitating the tombusvirus replication protein-driven PE enrichment within the replication compartment.
To test if Vps34p PI3K and PI(3)P play a role in the replication of other plant viruses, we treated N. benthamiana protoplasts (cell wall-free plant cells) with PI3K inhibitors, Wortmannin and AS604850, respectively [51]. We found that the replication of all five plant viruses within the Tombusviridae family, namely, TBSV, CIRV, cucumber leaf spot virus (CLSV), turnip crinkle virus (TCV), and red clover necrotic mosaic virus (RCNMV), was greatly inhibited (by up to 80–100%) by these two PI3K inhibitors (Fig 11). Thus, targeting of the cellular PI(3)P and the PI3K could lead to broad spectrum resistance against several plant viruses.
One of the emerging themes in RNA virus replication is that (+)RNA viruses build large replication compartments inside the cytosol of the infected cells during the infection process to sequester viral and host-components into membranous replication compartments for robust viral replication [5,9,64–66]. The replication compartment also provides protected subcellular environment for the viral dsRNA replication intermediates [5,6]. Formation of the replication compartments is driven by viral replication proteins with the help of numerous co-opted host proteins, cellular membranes and lipids [1,67–72].
In this paper, we exploited tombusviruses and yeast as simple model systems to identify new host factors involved in building the viral replication compartments. We have found a major role for the cellular Vps34p PI3K and PI(3)P phosphoinositide in the formation of tombusvirus replication compartment. The observed recruitment of Vps34p PI3K allows the production of PI(3)P within the replication compartment, which is required for robust tombusvirus replication. Accordingly, we measured little tombusvirus replication in the absence of Vps34 in yeast or when we knocked down Vps34 level in plant cells. We also observed increased expression of Vps34 and higher abundance of PI(3)P in the presence of the tombusviral replication proteins, which likely leads to more efficient tombusvirus replication.
An intriguing feature of tombusviruses is that they can utilize different subcellular organellar membranes for their replication. Accordingly, TBSV and CNV recruit Vps34p to the peroxisome/ER-based replication compartment, whereas CIRV co-opts Vps34p to the outer membranes of mitochondria. The retargeting of Vps34p into the viral replication compartment by tombusviruses seems to be based on direct interaction between the TBSV p33 or the CIRV p36 replication proteins and Vps34 PI3K. Vps34p interaction with the viral replication proteins seems to be temporal based on the decreasing amount of Vps34p in the purified replicase preparations when the assembly of new VRCs is stopped by cycloheximide. Vps34p is likely needed at the early replicase assembly/activation step since the synthesis of TBSV dsRNA replication intermediate was inhibited in CFEs prepared from vps34Δ yeast.
Vps34p participates in different complexes that perform separate functions, such as endomembrane/vesicular trafficking, autophagy, pheromone signaling and cytokinesis [31,33]. A gene deletion approach, which removed critical components of the various Vps34p complexes, indicated that all four of the known proteins in the vesicular trafficking pathway, such as Vps34p, Vps15p, Vps30p (human Beclin-1 homolog) and Vps38p, affected TBSV or CIRV replication in yeast. On the contrary, deletion of ATG14, which binds to Vps30 and is essential component of the autophagy and pheromone signaling pathways, had no effect on TBSV replication under the conditions used (Fig 9). Moreover, Vps34p affected PE recruitment to the site of replication (Fig 10) and PE is enriched in the endosomal membranes [28]. This further supports that the endomembrane/vesicular trafficking function of Vps34p is crucial for tombusvirus replication. Also, the findings further strengthen our previous model based on the pro-viral function of Rab5 small GTPase [28], which is required for TBSV to enrich PI(3)P within the replication compartment. Thus, tombusviruses exploit the early endosomes to usurp proteins, Vps34p and Rab5, and for enrichment of lipids, such as PE and PI(3)P, within the viral replication compartment.
The lack of early endosomal proteins Vps34p and Rab5 small GTPase had similar as well as unique effects on TBSV replication. For example, in the absence of Vps34p and Rab5, the accumulation levels of TBSV p33 and CIRV p36 replication proteins are greatly decreased, possibly due to protein degradation. Moreover, the enrichment of PE within the replication compartment depends on both co-opted host proteins. These overlapping, but not redundant, functions of Vps34p and Rab5 in tombusvirus replication might be due to the requirement of these proteins to maintain the proper endosomal network, including high PE level in the endosomal membranes, which then can be hijacked by tombusviruses. The unique function of Vps34p in comparison with Rab5 has been observed using the vps34Δ CFE-based replicase reconstitution assay, which showed reduced TBSV repRNA accumulation in vitro, whereas the comparable CFE assay from vps21Δypt52Δypt53Δ yeast supported as efficient TBSV repRNA replication as the CFE from WT yeast [28]. This difference indicates that the product of Vps34p PI3K, namely PI(3)P, likely plays a very important role in the VRC assembly. This model is supported by additional CFE-based assays, when addition of purified yeast Ymr1p PI(3)P phosphatase to the CFE from WT yeast reduced TBSV repRNA replication in vitro.
Vps34p seems to be recruited to the replication compartment to produce PI(3)P phosphoinositide, leading to highly enriched PI(3)P level in the membranes utilized for TBSV replication. The key component for tombusvirus replication is PI(3)P phosphoinositide, since Vps34p kinase mutants could not complement TBSV replication deficiency in vps34Δ yeast. And these mutants acted as dominant negative mutants by inhibiting TBSV replication in WT yeast. Moreover, expression of the PI(3)P phosphatase (yeast Ymr1p or Arabidopsis Mtm1) reduced TBSV and CIRV RNA accumulation. Also, sequestering PI(3)P with PI(3)P-binding proteins inhibited TBSV replication. PI(3)P is an important signaling lipid in the eukaryotic cells through interacting with many effector proteins [46,73,74], which might function in tombusvirus replication. PI(3)P level seems to affect p33 replication protein stability since p33 accumulation was reduced in vps34Δ yeast or when PI(3)P phosphatase was over-expressed.
The presented data with the proteasome inhibitor (Fig 2B) point out that the overall level of TBSV and CIRV replication is likely affected by both the reduced amount of replication proteins due to increased replication protein degradation in the absence of Vps34 and the direct contribution PI(3)P and/or Vps34 PI3K to formation of viral replication complex. The direct contribution of these host factors is also supported by in vitro CFE-based experiments, which demonstrated the less efficient TBSV replicase assembly in the absence of Vps34. Although the direct function of PI(3)P in tombusvirus replication is currently unknown, PI(3)P might be involved in recruitment of additional host factors, which bind to PI(3)P.
Characterization of the proviral effect of cellular PI(3)P and PI3K also opens up new approaches to target viral infections with antivirals, such as the PI3K inhibitors Wortmannin and AS604850 (Fig 11 and S1 Fig). Accordingly, we demonstrate strong inhibition of replication of tombusviruses and other related plant viruses and the unrelated NoV insect alfanodavirus. Therefore, it seems that chemical inhibition of PI3K might result in broad range protection against viruses. The NS4B replication protein of hepatitis C virus has been shown to interact with Vps34 and Rab5 [37], leading to induction of autophagy, which is beneficial for HCV replication [36].
Many (+)RNA viruses, such as picornaviruses and HCV, require PI(4)P instead of PI(3)P [70]. PI(4)P is the signature lipid of the secretory pathway, whereas PI(3)P is critical for the endosomal pathway. This difference among viruses might reflect the different membrane origins of their replication organelles. For example, poliovirus (PV) and coxsackievirus B3 hijack the Golgi and TGN and recruit PI4K to the replication organelle to produce PI(4)P in situ [67,75,76]. The 3C replication protein binds directly to PI(4)P and PI(4)P is required for viral RNA synthesis [75,77]. PI(4)P is also required for sterol enrichment within the replication organelles of Rhinovirus and PV [78,79].
In summary, tombusviruses take advantage of PI3K and enrich PI(3)P phosphoinositide within the viral replication compartment. This allows tombusviruses to provide optimal microenvironment for efficient VRC assembly and robust virus replication. Additional work will define if more endosomal components and PI(3)P effectors are exploited by tombusviruses for replication. Our observations of virus-mediated re-targeting of major cellular components should also be useful to understand the multiple and complex functions of cellular components and their roles in disease states.
Parental yeast strain BY4741 (MATa his3Δ1 leu2Δ0 met15Δ0 ura3Δ0), and deletion strains vps34Δ, ymr1Δ, atg1Δ, atg8Δ, vps15Δ, vps30Δ, vps38Δ, atg14Δ, atg5Δ, atg7Δ and atg12Δ were purchased from Open Biosystems. SC1 (MATa his3Δ1 leu2Δ trp1Δ289 uraΔ52) yeast strain was purchased from Invitrogen.
The list of plasmids and primers are described in S1 Table.
To measure the effects of deletion of specific yeast genes on replication of TBSV or CIRV, yeast strains BY4741, vps34Δ, atg1Δ, atg8Δ, vps15Δ, vps30Δ, vps38Δ, atg14Δ, atg5Δ, atg7Δ, atg12Δ and ymr1Δ were separately transformed with plasmids pESC-His-p33/DI72, pYES-His-p92 and pRS315-cflag for TBSV replication or pESC-strep-p36/DI72, pYES-strep-p95 and pRS315-cflag for CIRV replication. The transformed yeast cells were pre-grown in synthetic complete medium lacking uracil, leucine and histidine (ULH-) supplemented with 2% glucose at 29°C for overnight. Then, tombusviral repRNA replication was induced by changing the media to synthetic complete medium (ULH-) supplemented with 2% galactose at 23°C for 24 h for TBSV or 30 h for CIRV. Yeast total RNA and total protein were extracted and analyzed by Northern blot and Western blot, respectively [23].
PI3K inhibitors AS604850 (Selleck Chemicals, Cat#S2681) or Wortmannin (Alfa Aesar, Cat#AAJ63983) were added to the yeast culture to test the role of Vps34 in TBSV, CIRV, NoV and CNV replication in yeast. PI3K inhibitors dissolved in DMSO were used in different concentrations (presented in the figure legends). PI3K inhibitor AS604850 was added to the cultures when virus replication was induced. Yeast strain BY4741 was transformed with plasmids pEsc-His-p33/DI72 and pYes-His-p92 for TBSV, pEsc-His-p36/DI72 and pYes-His-p95 for CIRV, pEsc-His-CNV p33/DI72 and pYes-His-CNV p92 for CNV. Tombusviral repRNA replication was induced by changing the media to synthetic complete medium lacking uracil and histidine supplemented with 2% galactose at 23°C for 24 h for TBSV and CNV, and for 30 h in case of CIRV. For measuring NoV RNA1 and RNA3 accumulation, yeast strain BY4741 was transformed with plasmid pEsc-His/Cupm/NoV/RNA1/TRSVrz. Replication was induced by adding 50 μM CuSO4 to the synthetic complete medium lacking histidine containing 2% glucose and then yeast was cultured for 48 h at 29°C. After the induction of viral replication and drug treatments, total RNA and total protein were isolated to evaluate virus replication level. Each experiment was repeated.
To measure the effects of sequestering cellular PI(3)P on TBSV replication, yeast strain BY4741 was transformed with plasmids pEsc-His-p33/DI72 and pYes-His-p92 and with either pRS425-Cup1-RFP or pRS425-Cup1-RFP-Flag-2xFYVE or pRS425-Cup1-RFP-Flag-PX or pRS425-Cup1-RFP-Flag-PXm. The transformed yeast cells were pre-grown in synthetic complete medium lacking uracil, leucine and histidine supplemented with 2% glucose and 50 μM CuSO4, followed by culturing at 29°C for 16 h. Tombusviral repRNA replication was induced by changing the media to synthetic complete medium lacking uracil, leucine and histidine supplemented with 2% galactose and 50 μM CuSO4 for 24 h at 23°C. Yeast total RNA and total protein were extracted and analyzed by Northern blot and Western blot, respectively.
NbVps34 gene expression was silenced using tobacco rattle virus (TRV) -based virus-induced gene silencing (VIGS) system in N. benthamiana [80,81]. After 12 d post agroinfiltration, the upper NbVps34-silenced leaves were inoculated with TBSV or CIRV sap. Plant leaf discs from inoculated leaves were collected for total RNA extraction at 2 d post inoculation (dpi) for TBSV and 2.5 dpi in case of CIRV. The control plants were treated the same way, except using TRV-cGFP. AtMtm1 was expressed in N. benthamiana by agroinfiltration. The agroinfiltrated leaves were inoculated with TBSV or CIRV sap at 44 h or 30 h post agroinfiltration. Then, the leaf discs from the inoculated leaves were collected at 2 dpi for RNA and protein detection [80].
Protoplasts were prepared from N. benthamiana callus as described previously [82]. About 5 x 105 protoplasts were electroporated with 2 μg in vitro transcribed full-length TBSV, CIRV, CLSV, TCV genomic RNA or RCNMV RNA1. Different amounts of AS604850 or Wortmannin PI3K inhibitors (see figure legends) and DMSO (as the negative control) were added to the electroporated protoplasts, and incubated in 35 x 10 mm petri dishes in dark at room temperature for 16 hours. Total RNAs were then isolated from these protoplasts and subjected to Northern blot analysis for detection of viral RNA accumulation. Radioactive-labeled probes for Northern blotting were prepared by in vitro RNA synthesis using T7 RNA polymerase in the presence of [α-32P]UTP using PCR-amplified templates.
Yeast strain BY4741 was transformed with pESC-His-p33/DI72, pYES-His-p92 and pRS315-Vps34-Flag plasmids. The transformed yeast cells were pre-grown in synthetic complete medium lacking uracil, leucine and histidine supplemented with 2% glucose medium at 29°C overnight. Tombusvirus repRNA replication was induced by changing the media to synthetic complete medium lacking uracil, leucine and histidine supplemented with 2% galactose for 21 h at 23°C. Immunofluorescence analysis was conducted as described previously [83]. Briefly, yeast cells were digested with Zymolase 20T to remove yeast cell wall, followed by simultaneous incubation with anti-p33 monoclonal mouse antibody and anti-Flag rabbit antibody (Sigma-Aldrich, Cat#F7435). Subsequently, the spheroplasts were incubated with anti-mouse secondary antibody conjugated to Alexa Flour 647 and anti-rabbit secondary antibody conjugated to ATTO488 for 1 h. Cells on the glass-bottom dishes were subjected to super-resolution microscopic observation (N-STORM Super Resolution Microscopy, Nikon).
To examine the subcellular localization of Vps34 in plants, N. benthamiana leaves were co-infiltrated with Agrobacterium carrying plasmids pGD-AtVps34-GFP or pGD-RFP-SKL together with pGD-p33-BFP (0.3 OD600, each). The agroinfiltrated leaves were inoculated with TBSV sap at 14 h post agroinfiltration. After 2dpi, the agroinfiltrated leaves were subjected to confocal microscopy (FV3000 confocal laser scanning microscope, Olympus) using 405 nm laser for BFP, 488 nm laser for GFP and 559 nm for RFP. Images were captured successively and merged using the FLUOVIEW software [15].
To analyze the subcellular localization of Vps34 in yeast cells, pYes-His-p92, pEsc-GFP-p33/DI72 and pRS315-Vps34-RFP were transformed into BY4741 yeast. The transformed yeast cells were pre-grown in synthetic complete medium lacking uracil, leucine and histidine supplemented with 2% glucose at 29°C overnight. Tombusviral repRNA replication was induced by changing the media to synthetic complete medium lacking uracil, leucine and histidine supplemented with 2% galactose for 21 h at 23°C. Yeast cells were subjected to confocal microscopy analysis using 488 nm laser for GFP and 559 nm for RFP in an Olympus FV1200 confocal laser scanning microscope. The co-localization between p33 and Vps34 in yeast was quantitated from the p33 point of view using the FLUOVIEW software.
To test whether Vps34 localizes to the peroxisome upon virus replication in yeast cells, pYes-His-p92, pEsc-His-p33/DI72, pRS314-Pex13-GFP and pRS315-Vps34-RFP were transformed into SC1 yeast. The transformed yeast cells were pre-grown in synthetic complete medium lacking uracil, leucine, tryptophan and histidine supplemented with 2% glucose medium at 29°C overnight. Tombusviral repRNA replication was induced by changing the media to synthetic complete medium lacking uracil, leucine, tryptophan and histidine supplemented with 2% galactose for 21 h at 23°C. Yeast cells were subjected to confocal microscopy analysis using 488 nm laser for GFP and 559 nm for RFP in an Olympus FV1200 confocal laser scanning microscope.
To observe PI(3)P localization upon virus replication, plant protoplasts or yeast spheroplasts were isolated and treated (as described above), then subjected to immunofluorescence analysis. The permeabilized cells were incubated with purified anti-PI3P mouse antibody (Echelon Biosciences Inc. Cat#Z-P003), and then incubated with anti-mouse secondary antibody conjugated with Alexa Fluor 568 (Thermo Fisher Scientific, Cat#A11031). The cells were imaged with Olympus FV1200 confocal laser scanning microscope. The intensity profiles of the images were processed and exported using the FLUOVIEW software [15].
PI(3)P biosensor was used to determine PI(3)P localization upon virus replication. Yeast strains BY4741 and vps21Δypt52Δypt53Δ were transformed with pEsc-GFP-p33/DI72, pYes-His-p92 and pRS315-RFP-2xFYVE plasmids. The transformed yeast cells were pre-grown in synthetic complete medium lacking uracil, leucine and histidine supplemented with 2% glucose, followed by culturing at 29°C overnight. Tombusviral repRNA replication was induced by changing the media to synthetic complete medium lacking uracil, leucine and histidine supplemented with 2% galactose for 21 h at 23°C. Yeast cells were subjected to confocal microscopic analysis using 488-nm laser for GFP and 559-nm for RFP in an Olympus FV1200 confocal laser scanning microscope.
To identify interaction between Vps34 and TBSV p33 or CIRV p36 replication proteins in vivo, bimolecular fluorescence complementation (BiFC) assay was conducted [13]. The plasmids pGD-p33-cYFP or pGD-p36-cYFP, pGD-nYFP-AtVps34, pGD-nYFP-MBP, pGD-RFP-SKL and pGD-CoxIV-RFP were transformed to Agrobacterium strain C58C1. The obtained Agrobacterium transformants were co-agroinfiltrated (0.3 OD600, each) into the leaves of four weeks-old N. benthamiana plants. Agroinfiltrated leaves were inoculated with TBSV or CIRV 14 h after agroinfiltration. Agroinfiltrated leaves were subjected to confocal laser microscopy at 48 h post infiltration.
Additional methods used are described in S1 Text.
|
10.1371/journal.ppat.1005000 | HIV Reactivation from Latency after Treatment Interruption Occurs on Average Every 5-8 Days—Implications for HIV Remission | HIV infection can be effectively controlled by anti-retroviral therapy (ART) in most patients. However therapy must be continued for life, because interruption of ART leads to rapid recrudescence of infection from long-lived latently infected cells. A number of approaches are currently being developed to ‘purge’ the reservoir of latently infected cells in order to either eliminate infection completely, or significantly delay the time to viral recrudescence after therapy interruption. A fundamental question in HIV research is how frequently the virus reactivates from latency, and thus how much the reservoir might need to be reduced to produce a prolonged antiretroviral-free HIV remission. Here we provide the first direct estimates of the frequency of viral recrudescence after ART interruption, combining data from four independent cohorts of patients undergoing treatment interruption, comprising 100 patients in total. We estimate that viral replication is initiated on average once every ≈6 days (range 5.1- 7.6 days). This rate is around 24 times lower than previous thought, and is very similar across the cohorts. In addition, we analyse data on the ratios of different ‘reactivation founder’ viruses in a separate cohort of patients undergoing ART-interruption, and estimate the frequency of successful reactivation to be once every 3.6 days. This suggests that a reduction in the reservoir size of around 50-70-fold would be required to increase the average time-to-recrudescence to about one year, and thus achieve at least a short period of anti-retroviral free HIV remission. Our analyses suggests that time-to-recrudescence studies will need to be large in order to detect modest changes in the reservoir, and that macaque models of SIV latency may have much higher frequencies of viral recrudescence after ART interruption than seen in human HIV infection. Understanding the mean frequency of recrudescence from latency is an important first step in approaches to prolong antiretroviral-free viral remission in HIV.
| During treatment of HIV infection the virus persists in infected cells in a quiescent or ‘latent’ state. If treatment is stopped, then virus rebounds to detectable levels usually within 2–3 weeks. This is thought to occur due to release of infectious virus from a reservoir of long-lived latently infected cells. Reducing the number of latently infected cells should allow a prolonged period of HIV remission without antiviral treatment. A fundamental question is ‘how frequently does infectious virus emerge from the pool of latently infected cells?’, and thus how much would we need to reduce the number of latently infected cells to produce remission? Here we directly estimate the frequency of successful viral reactivation in four independent cohorts of patients undergoing treatment interruption. We find that active infection is initiated on average once every 5–8 days, considerably more slowly than previously thought. This has important implications for how much we need to reduce the number of latent cells in order to produce remission. Whereas previous analyses suggested that we would need to reduce the latent cell number 2000 fold to produce an average one-year remission, we show that reducing the latent cell number by 50–70 fold could achieve this aim.
| The development of highly potent antiretroviral therapy (ART) for HIV means that the virus can be effectively controlled in most treated patients. However, ART must be taken continuously, as interruption of ART is followed by the rapid recrudescence of virus from a quiescent ‘latent reservoir’ of infected cells. A major thrust of HIV research is to reduce the latent reservoir so that prolonged antiretroviral-free HIV remission can be achieved. A number of ‘latency reversing agents’ (LRA) are currently being developed to reduce the latent reservoir by reactivating latently infected cells [1–4]. Clinical studies of LRA in HIV-infected patients on ART have shown the ability to significantly increase cell-associated unspliced HIV RNA and in some studies increase plasma HIV RNA. However, these studies have not resulted in decreases in HIV DNA—a crude surrogate marker of latently infected cells—or measurable reductions in various measurements of the latent reservoir or antiretroviral free HIV remission [5–10]. A fundamental question in achieving HIV remission is what level of reduction of latently infected cells is required? It is currently estimated that the reservoir of latently infected cells may be between one and 60 million cells [11–13]. Complete elimination of this would thus require reducing the size of the reservoir by at least one million fold. However, reducing the reservoir by smaller amounts may still produce significant delays between ART-interruption and viral recrudescence, allowing potentially for prolonged interruptions of therapy before viral recrudescence. Understanding factors that predict the duration of viral remission will be critical for the future design of eradication studies [14].
The dynamics of HIV reactivation from latency have not been examined in detail experimentally. However, it is clear from a number of studies that after ART-interruption there is generally a delay of about a week before viral rebound can be detected, and about half of the patients often experience rebound within the first two weeks or so [8,15–17]. However, a proportion of patients usually remains virus-free even after a month, suggesting variable dynamics of reactivation. From this observation, we can attempt to predict the underlying dynamics from our understanding of infection. Firstly, we expect that the latently infected cells might ‘attempt’ reactivation even during successful ART, because ART itself is not expected to affect the rate of initial latent cell reactivation. These reactivation attempts by latent cells may occur at some average frequency, and some fraction of these events will produce replication competent virus and thus be capable of initiating successful viral rebound. These reactivation events do not result in successful viral growth while therapeutic levels of ART are present. Thus, after ART there will be a period of ‘drug-washout’ before the virus is able to grow (which will vary depending on the pharmacokinetics of the ART regime). Once ART levels have declined sufficiently that viral growth is possible, there may be some delay until the first replication-competent viral reactivation event occurs (assuming reactivation is occurring randomly). Following the first successful reactivation, we expect virus levels will start at some low level, and then take some time to grow to the level of viral detection. Together, the time for drug-washout and viral growth create a ‘fixed delay’ before reactivation can be detected, and likely explain the fact that little rebound is usually detected in the first week after ART-interruption. After this fixed delay, if latent cells are reactivating randomly at some average frequency, this will lead to an exponential distribution in the time-to-reactivation observed, and may explain why some patients remain virus-free for longer periods. By fitting of the time-to-reactivation curve, we can estimate the frequency of successful reactivation from latency.
In this study, we directly estimate the frequency of HIV recrudescence from latency following ART-interruption by analysing the time to detection of viral rebound from 4 independent patient cohorts undergoing ART-interruption. We find that the average frequency of successful reactivation from latency is approximately once every 6 days, around 24 times lower than previously estimated [18,19]. This low rate of successful reactivation has important implications for designing future eradication studies.
After interruption of successful ART, HIV rebounds to detectable levels within a few weeks in the majority of patients. This requires reactivation of latent cells bearing replication competent virus. The frequency with which this reactivation occurs is likely a function of the size of the latent reservoir (which may vary substantially between individuals [20–22]), and the per-latent-cell probability of successive reactivation. If the initiation of viral growth after ART-interruption is a random event then the distribution of time-to-initiation will be exponential, and we could estimate the average frequency of initiation directly from the ‘survival curve’ of time-to-initiation of viral growth. However, since we are usually unable to detect the initiation of viral growth after ART-interruption, we instead measure ‘time-to-detection of virus’ at some threshold viral level. The actual time when we first detect virus is delayed both because of drug washout preventing viral growth immediately after interruption, and the time taken for the virus to grow from its level at initial reactivation to our threshold for detection of plasma virus (Fig 1A). The duration of the delays due to drug washout and viral growth to the level of detection only affect the “shoulder” of the curve of time-to-detection by delaying the time until we could first detect virus (Fig 1B). The average delay to viral detection is thus the sum of the average time between initiation of successful reactivation events, and the delay from reactivation to detection. There may be a distribution in both time-to-initiation as well as from reactivation to detection, both of which may affect the shape of the subsequent time-to-detection curve.
To test this approach we first analysed the kinetics of time-to-detection of HIV in a published cohort of nine patients treated with the LRA panobinostat, and undergoing therapy interruption and biweekly monitoring of viral loads [8]. The threshold of detection of HIV viremia was 20 copies ml-1, and virus was first detected between day 10 and day 45 across the patient group (Fig 2A). To see if the observed time-to-detection was consistent with an exponential process, we plotted the ‘survival curve’ of time-to-detection in the cohort (Fig 2B). This plot demonstrates an initial shoulder (as expected due to drug washout and the time taken for viral growth), followed by a survival curve that conformed well to an exponential process. The exponential rate can be estimated from the survival curve, and equates to a frequency of viral reactivation of once every 7.6 days (95% confidence intervals (CI) = 6.5, 9.1).
To test whether the exponential model was suitable, we performed a Chi-squared goodness-of-fit analysis, which indicated a good fit to the data (p = 0.988). Although the analysis above is consistent with an exponential process, this does not prove that this is the only source of delay. It has also been proposed that early stochastic events, differences in the initial level of replicating virus, or differences in viral growth rate may contribute to the delay until viral detection [18,23]. However, a comparison between a survival curve based on an exponential distribution with one based on a gamma distribution showed that the gamma distribution (which incorporates multiple delays) did not provide a significantly better fit (p = 0.72 F-test). We can also use a modelling approach to understand the effects of these different factors on time-to-detection. For example, Pearson et al. [24] have estimated the distribution of delays arising from early stochastic events following primary HIV infection under different assumptions. Under most scenarios, the expected distribution of delays from early stochastic events is of the order of 1–3 days. Moreover, the importance of stochastic delays only becomes relevant in the presence of a low frequency of reactivation. In the presence of frequent reactivation, the stochastic delay in any individual reactivation event is overcome by the rapid arrival of the next reactivation (see S1 File). Another potential cause of differences in time-to-detection is differences in initial levels of virus. That is, if the first latent cell to reactivate were to ‘seed’ the infection with a lower initial level of virus, then it will take longer for the virus to grow to the level of detection. However, given the growth rates of virus observed in these patients, the initial level of virus would have to vary by many orders of magnitude to produce the delays observed (see S2 File).
It is also possible that the distribution in time-to-detection arose because of slower viral growth in some patients. To investigate whether the observed differences in detection times could be due to slower viral growth, we estimated the viral growth rate from the serial viral load measurements after virus became detectable, and investigated whether later detection was associated with slower viral growth. We found no correlation between viral growth rate and when virus was first detected (Fig 2C), indicating that slow viral growth did not explain the distribution of time-to-detection.
Taken together, these results are consistent with the observed time-to-detection in this cohort being determined by a low rate of viral recrudescence from latency, with an average frequency of initiating viral replication of once every 7.6 days. A potentially confounding factor with this analysis is that patients were part of a trial of the LRA panobinostat, a histone deacetylase inhibitor, to assess its effect on the HIV reservoir under ART. We note that although panobinostat increased HIV in plasma and cell-associated unspliced HIV RNA, there were no changes in HIV DNA. Thus, it seems unlikely that panobinostat treatment significantly reduced the HIV reservoir. Alternatively, it is possible that panobinostat-induced activation might increase the frequency of viral recrudescence, although this seems unlikely given that the last dose of panobinostat was administered >36 weeks before ART interruption.
Given the small number of patients in cohort 1 and their prior treatment with an LRA, it is important to confirm the estimated frequency of reactivation in other patient cohorts that have not received LRA. Therefore we obtained data on time to recrudescence for another three cohorts of patients undergoing ART-interruption, comprising an additional 91 subjects (summarised in Table 1). In the second cohort, 59 patients treated in primary infection underwent treatment interruption and weekly monitoring [15] (Fig 3A). Estimating the frequency of initiation from the time-to-detection of virus (at a threshold of 50 copies ml-1) we found an average frequency of once every 6.3 days (CI = 5.7, 7.1) (Fig 3B), similar to our estimate from the panobinostat cohort. Estimation of viral growth rate was less accurate in this cohort, as patients were only sampled weekly. We compared viral growth rate in this cohort with the time-to-detection to once again check whether a difference in viral growth rate could explain the different time-to-detection of virus. We estimated viral growth using a ‘two-point’ growth estimate to compare growth rates of virus in patients where virus was first detected in different weeks. Using this approach, there was no difference in growth rate estimates for patients with virus detected in weeks two and three, but a slightly higher growth rate in week one (Fig 3C). However, this estimate of growth rate is biased by the fact that viral loads at detection were lower in week one, and therefore we were estimating viral growth rate earlier in the growth phase, before it slows towards peak (Fig 3D). Overall, differences in viral growth did not appear to play a major role in time-to-detection of infection in this cohort.
We also analysed two other cohorts using data extracted from earlier publications on ART-interruption. The third cohort included 18 patients undergoing ART-interruption, where time-to-detection at a threshold of detection of 50 copies ml-1 was measured and viral growth rates were estimated (Table 2 of reference [16]). The fourth cohort included 14 patients monitored on days 4, 8 and 14 following ART-interruption (using data from Fig 1 of reference [17]). Because of the small number of patients and timepoints in the fourth cohort, we included data from five sequential interruption cycles. Using the same method for estimating the frequency of initiation from the time-to-detection curves, we found very similar frequencies of viral recrudescence in these two cohorts (every 5.1, days CI (4.2, 6.5) and 7.2 days CI (6.0, 8.7) respectively, see Fig 4A–4B). In the third cohort viral growth rate was also estimated independently in the original study (reference [16], Table 2), and again, viral growth rate was not significantly correlated with time-to-detection of infection (Fig 4C), confirming that differences in viral growth played little role in the time-to-recrudescence in this study. Comparing all cohorts together, we found a trend for slightly higher frequencies of reactivation in cohort 3, who initiated ART in chronic infection, and slightly lower frequencies of reactivation in patients treated in primary infection (cohort 2) or with the LRA panobinostat (cohort 1). However, the frequency of recrudescence was not significantly different between the cohorts (p- value = 0.059, F-test). In addition, we used a Chi-squared test to assess whether the exponential model of reactivation frequency was suitable across the four datasets, and found that the data conformed well to this model (p = 0.996) and that a survival curve based on a gamma distribution did not provide a better fit (p = 0.5, F-test).
Overall, despite the different sampling regimens and study designs, the estimated frequencies of reactivation were similar across the four cohorts studied (once every 5.1, 6.3, 7.2, and 7.6 days), with an average frequency of once every 6.0 days (CI 5.5,6.6).
The analysis of time-to-detection of HIV following ART-interruption suggests a relatively low frequency of recrudescence from latency, and thus a significant delay between successive reactivation events. If each reactivation event is ‘founded’ by virus produced by a single latently infected cell, this predicts that early after ART-interruption, the viral population would often be the progeny of a single latent cell (in much the same way as virus observed early after sexual transmission is thought to arise from a single founder virion). Joos et al [25] have compared the diversity of the HIV plasma viral population present soon after ART interruption with the diversity present prior to commencing ART. They found a major narrowing of diversity after ART-interruption, suggesting monoclonal or oligoclonal origins of the plasma virus. Although the viral population after ART-interruption was not entirely homogeneous, they observed one or more ‘families’ of closely relate viruses, differing by only a few nucleotides, similar to the founder viruses observed after sexual transmission. They concluded from this that the viral population after ART-interruption represented random reactivation of latently infected cells, rather than continual seeding of virus.
We accessed the viral sequence data from the Joos study (Genbank accession numbers listed in the original publication [25]) and reanalyzed this data in order to investigate the ratios of different ‘reactivation founder’ viruses in these patients. We observed six patients in whom it was possible to identify and count the frequency of founder viruses early after ART-interruption, and investigated the ratio of the number of copies of the most frequently observed founder to the next most frequently observed founder (see S3 File). This ratio of founder copies is determined by both the delay until the next founder starts growing, and the overall growth rate of the virus. We then considered the distribution of these founder ratios, and used this to estimate the distribution of reactivation events and thus the average frequency of reactivation.
We used maximum likelihood estimation to fit the ratios of founder copies observed in the Joos study to the theoretical distribution of ratios we would expect if founders reactivated λ times per day (described in detail in methods). We found the average frequency of reactivation events (1/λ) to be once every 3.6 days (CI 1.98–6.62 days). The real delay between reactivation events is likely more than this, because in some cases (marked with an asterisk in Fig 5A) we can only estimate the minimum ratio (for example, if all 16 sequences in a patient are from the same founder we can only say the frequency of the next founder is likely <1/16, whereas it could be much lower). On the other hand, it is also possible that two latent cells bearing founder viruses that were identical in the sequenced region reactivated sequentially, and thus were classified as a single founder. We aimed to minimise the likelihood of this occurring by only selecting patients for this analysis with sufficient diversity of virus pre-treatment (sequenced in the same region of the virus). In addition, it is likely that the latent reservoir was indeed more diverse than the circulating virus immediately before treatment, as it contains an archive of different viral strains. Thus, it seems unlikely that we are aggregating multiple identical founder viruses. Future studies using larger regions of the virus, and / or more in-depth sequencing approaches should provide more accurate estimates of the ratio of reactivation founders and the frequency of reactivation. However, our analysis of the ratios of reactivation founder viruses leads to very similar estimates of reactivation frequency to those obtained studying time-to-detection.
Recent studies in macaques have suggested that very early treatment after SIV infection may also lead to delayed time-to-recrudescence after ART-interruption [26]. In this study they found that time-to-recrudescence was very short and significantly correlated with area-under-the-curve of viral load since infection, although significantly longer delays were seen only animals treated within 3 days of infection [26]. Using the same approach to estimate frequency of reactivation from time-to-detection of virus in the animals treated with ART at days 7, 10 and 14 (ie: excluding animals treated at day 3), we found that the average frequency of initiation of viral replication in macaques was once every 1.7 days, compared to every 6 days in HIV (Fig 4D). One explanation for this might be higher levels of HIV DNA in the macaques. However, the total number of HIV copies per million PBMC measured in the macaques just prior to ART-interruption seems similar to that reported in patients during ART [8,27]. Another reason for the higher reactivation rate in macaques may be the generally shorter periods of treatment (24 weeks of ART in SIV versus >12 months in the HIV studies (Table 1)), which may have allowed less time for activated cells to decay and a steady-state of latently infected cells to be attained [28]. Alternatively, differences in immune activation or cytokine levels may also play a role. Regardless of the mechanism, this work suggests that short-term treated macaques may experience much higher rates of reactivation from latency compared to HIV patients even if treated early after infection.
The primary goal in tackling HIV latency is to allow prolonged HIV remission in the absence of ART. Thus, a major question is how much we would need to decrease the latent reservoir in order to produce a durable delay in time-to-recrudescence and subsequent recommencement of ART? A previous study estimated that a reduction in reservoir size of >2000 fold would be required to provide a one year average delay until reactivation [18]. However, that study assumed that viral reactivation was over 24 times faster than our estimates (reactivation every 0.25 days), based on indirect modelling approaches published previously [18,19]. Our analysis demonstrates a much lower rate of viral reactivation, and thus much smaller reductions in the size of the latent pool would be needed for a one-year delay to reactivation. The required reduction in latent reservoir can be calculated as:
R=Td
(1)
Where R is the required reduction in size of the latent reservoir, T is the length of delay until viral recrudescence and d is the average time between viral reactivation events (ie: the baseline frequency of viral reactivation). For a baseline frequency of reactivation of once every 6 days (the average over the four cohorts), our analysis predicts that a 61-fold reduction in the reservoir would provide an average one-year delay until recrudescence. Thus, for example, 12 rounds of therapy using an LRA that reduced the reservoir (and reactivation rate) by 30% would achieve an ≈72-fold reduction in the reservoir and hence an average one year ART-free control of viremia.
Several recent case reports have suggested that very prolonged remission is possible if the reservoir can be reduced by early treatment or other interventions such as bone marrow transplantation. In the case of the ‘Mississippi baby’, viral recrudescence was not observed until 27 months after ART-interruption [29]. Similarly, in two cases of haematopoietic stem cell transplantation in adults, viral recrudescence was not observed until 84 and 225 days after ART-interruption [30]. Our analysis indicates that the average frequency of reactivation is once every 6 days across the four cohorts we analysed. Therefore these three cases are respectively 135, 14 and 37 fold longer than expected on average in these cohorts. One might speculate from these delays that the reactivation rate and reservoir size were respectively 135, 14 and 37 fold smaller than average (using Eq 1). However, since reactivation is a random process, recrudescence is not always observed at the average time expected. Using the data we can estimate bounds for the likely frequency of initiation of viral replication (and the extent of reservoir reduction) based on observed time-to-detection of virus (see Fig 5B). For example, given an observed time-to-detection of 84 days or more, it is highly unlikely that the reservoir was of the average size determined by analysis of our four cohorts (probability for this is 8.4x10-7). For a time-to-detection of 84 days to lie within the range expected for 95% of subjects, then the average frequency of recrudescence would have to be bounded below by 23 days and above by 3318 days. This suggests that in this case the latent reservoir was most likely between 3.8 fold and 553 fold smaller than the average size estimated from our four cohorts. Using the same approach in the case of the Mississippi baby, the maximum predicted reduction in viral reservoir (top border of 95% CI) is 5,300 fold. A corollary to these observations is that the rate of reactivation from latency and level of viral reservoir in the transplant patients is not decreased as much as might be predicted from the degree of chimerism seen in peripheral blood (<0.001% of PBMC were of donor origin [30]). However, as noted by the authors of that study, the degree of chimerism in the patients’ tissues are likely significantly higher than that seen in PBMC [30] particularly as the patients received a reduced intensity conditioning regime. In addition, a number of recent studies have suggested that lymphatic sites may be a significant source of virus under therapy [31–33]. Thus, we speculate that reactivation from chimeric tissue sites might contribute to the observed reactivation rate. Overall, the wide error bars on estimates of potential reservoir size based on time-to-detection of individual patients (Fig 5B) suggest significant limitations in the use of time-to-detection to estimate reservoir size. Therefore, we also investigated the usefulness of time-to-detection assays in detecting the effects of LRA.
A major question in clinical trials of LRA is how to measure changes in the latent reservoir. Approaches using detection of plasma HIV RNA, cell-associated HIV RNA, cell associated HIV DNA, as well as ex vivo quantitative outgrowth assays have been studied [7–10,34,35]. However, it is not clear whether these measures will reflect time to viral recrudescence after ART-interruption in vivo [8,36]. Since HIV remission involves essentially a prolonged time-to-detection of virus, direct measurement of time-to-detection following ART-interruption will ultimately be the most clinically meaningful endpoint.
Treatment interruption studies to measure time-to-detection pose a number of ethical questions. Firstly, frequent treatment interruptions may increase morbidity or mortality compared to continuous treatment [37], although it is less clear that occasional interruptions would have the same effect. Secondly, interruption may act to ‘replenish’ the viral reservoir, although this does not appear to occur quickly [22]. Thirdly, such studies would ideally require a control group, in order to compare time-to-detection in treated versus untreated patients. However, in addition to these factors, there are also a number of issues with the statistical power to detect delays in time-to-detection. Firstly, as indicated by the extremely wide error bars in our estimates of relative reservoir size in the Boston patients and Mississippi baby (Fig 5B), time-to reactivation is not a useful measure of reservoir size in an individual, because of the random nature of reactivation. Time-to-detection is only useful in cohorts of patients. Once we understand time-to-detection as an exponential process, we can apply a power analysis to estimate how many patients would be required to identify differences in time-to-detection. Such an analysis suggests that to detect a 30% decrease in the reservoir (and a 30% increase in the frequency of initiation of viral replication) assuming a 100 day follow-up one would require >120 patients in each arm to have an 80% chance of detecting a difference (Fig 5C). Thus, such studies are only likely to be useful in detecting rather large changes in the reservoir and rate of reactivation.
Our work provides the first direct estimates of the frequency of viral recrudescence from latency based on analysis of time-to-detection of plasma viremia from ART-interruption cohort studies. Previous studies have modeled the recrudescence of virus following ART-interruption, using a variety of approaches. This work has often focused on the dynamics of virus within the individual, and either did not estimate the frequency of viral reactivation [16,38], or estimated a constant rate of production of virus, rather than the frequency of events [23]. Rong et al used a similar modeling approach to understand viral ‘blips’ during ART, and estimated that these were infrequent [39]. Pennings et al estimated the frequency of successful latent cell reactivation under ART as five times per day, based on the rate of development of drug resistance under ART and viral mutation rate [19], and more recently Hill et al used this frequency to model the affects of LRA [18]. Our estimate of the frequency of successful reactivation from latency (once every 5–8 days) is based upon analysis of the distribution in time-to-detection in the patient population, and is substantially slower than these previous estimates. The relatively slow frequency of recrudescence has important implications for understanding how to prolong anti-retroviral-free viral remission. In addition, careful consideration of the dynamics of viral recrudescence is critical to designing successful future eradication studies. Our work suggests that rather than using indirect approaches to estimate reductions in the reservoir and predict delays in time-to-recrudescence, we should measure this directly.
Previous studies of the number of latently infected cells under therapy have estimated that there are between 1 million and 60 million latently infected cells in the body, and the half-life of the latent reservoir is around 44 months [11–13]. One question that arises from this is whether reactivation from latency plays a significant role in the observed rate of decay of the latent reservoir, and whether periodic reactivation may lead to the reservoir ‘running dry’ [18,39,40]. If the current estimates of the number of latently infected cells and their decay rate are correct, this means the reservoir ‘loses’ on average at least 500 cells per day. Since we observe a successful reactivation from latency and reseeding of the viral reservoir only every 6 days, this suggests that reactivation would play a minimal role in the decay of the latent reservoir (unless the reservoir is much smaller than previously estimated). Previous modelling has assumed that there may be many ‘abortive’ reactivation events for every successful reactivation leading to recrudescence [18]. This might occur, for example, if very early events in viral reactivation are controlled by the host immune response. However even considering this possibility, it seems unlikely that reactivation from latency is a major factor contributing to the observed half-life of the latent reservoir. Finally, if we assume that previous estimates of the reservoir size are correct, we can also estimate the average probability of an individual latently infected cell successfully initiating viral recrudescence on a given day, and the average time until an individual cell is likely to achieve this. Assuming a conservative reservoir size of one million latently infected cells per patient and the fact that on average a patient has only one successful reactivation every 6 days, we can calculate that an individual latent cell has only a 1.7 x 10−7 probability of initiating viral recrudescence each day. Thus, the average time for an individual latent cell to initiate infection (assuming they all have the same probability of this) is around 16,500 years (ie: most latent cells will not successfully initiate an infection within the lifetime of the host). Although this per-cell probability of reactivation appears very low, it is perhaps worth considering how the reservoir is generally measured experimentally. The estimate of one million cells comes from the frequency of cells able to initiate viral growth in an in vitro viral outgrowth assay [12]. This assay involves stimulation of cells with PHA, and thus aims to estimate the number of ‘reactivatable’ latent cells with this strong and generalized stimulus. Reactivation in vivo may rely on antigenic stimulus, or random weak reactivation events, which activate a much smaller proportion of latent cells at any one time. Thus, it is not surprising that a much smaller number of cells is estimated to reactivate in vivo, than can be stimulated in vitro. Nonetheless, the in vitro quantitative viral outgrowth assay likely gives us a valuable measure of the size of the reservoir, as long as we recognize that only a fraction of these will actually reactivate in a given time. Given this potential for very prolonged quiescence of latent cells, it is not surprising that reactivation can be observed after prolonged periods of remission, as has been observed after transplantation and in the case of the Mississippi baby [29,30]. In our analysis we estimated a frequency of initiation of viral rebound for the different cohorts as if all patients in a cohort had the same frequency. However, recent studies have shown that reservoir size may vary substantially between patients and appears correlated with time to recrudescence after ART-interruption [20–22], and it is highly likely that our patients also differed in reservoir size and rebound rate. To investigate this, we looked at whether time-to-detection was correlated for individual patients undergoing successive ART-interruptions in the SSITT trial (cohort 4). We found that time-to-detection was indeed correlated over multiple interruptions in individual patients (Kendal’s concordance W = 0.47, p = 0.013). Thus, it seems likely that the frequency of reactivation we estimate for the cohorts represents the average frequency in the cohort, and there will be a distribution amongst individuals. In addition, we model reactivation as if it were the only mechanism affecting time to detection, and disregard the effects of viral growth because it is not correlated with time-to-detection. It is clear that differences in growth rates will inevitably affect time-to-detection, as slower growing virus will be seen later. However, unless the distribution of delays due to growth is large compared to the delays due to time-to-initiation, we would not expect growth to correlate well with time-to-detection (as we have recently illustrated in the context of malaria infection [41]).
There are clear limitations of our analysis of time-to-detection after ART-interruption, including the use of diverse cohorts, capturing patients at different times of infection, or after different interventions (including the use of an LRA)(summarised in Table 1). Our analysis was limited to ART-interruption studies with regular sampling after interruption, as this is required to capture the time-to-detection. Despite these obvious differences in the cohorts, we find very similar estimates of the average frequency of reactivation. These estimates were confirmed by a completely different approach, analyzing the ratios of ‘reactivation founder’ virus after ART-interruption. Our analysis suggests that much larger and more frequently sampled cohorts may be required to demonstrate differences in time-to-recrudescence amongst patients treated at different stages of infection or for different times (consistent with the predictions of the power analysis). One apparent paradox of any estimate of frequency of reactivation is the observation of persistent low viral loads in patients on ART [42,43]. If infectious virus were continuously present, then there is no real concept of delay-to-reactivation (and this argument applies equally to estimates of five reactivation events per day, or one every 6 days). However, the presence of reactivation founder virus suggests that viral growth is initiated by discreet reactivation events, rather than a constant ‘dribbling’ of virus from the latent reservoir [25]. Therefore it seems likely that the low levels of circulating virus detected under ART do not provide an immediate source of virus for reactivation.
The development of therapies to purge the latent reservoir of HIV and produce prolonged antiretroviral-free HIV remission is a major priority. Understanding the frequency of recrudescence from latency is a crucial parameter in predicting the impact of interventions. Establishing the ‘normal’ rate of recrudescence from latency in HIV also allows us to assess the appropriateness of animal models and interventions, which can be judged on their ability to alter this parameter. A variety of approaches have been proposed to assess the effectiveness of latency reversing drugs. However, ultimately the test of LRA efficacy is the length of remission after ART-interruption. Future studies should determine the best predictors of time-to-recrudescence, so that these measures may be used as proxies to assess the efficacy of HIV eradication interventions.
This manuscript involves the analysis of previously published data from original human and animal studies published elsewhere (summarized in Table 1). Details of the ethical approval for the original studies may be found in the original publications.
To study the dynamics of viral recrudescence, we assumed that the initiation of viral replication after ART interruption is a random event, occurring at a given frequency. Thus, the time-to-initiation will be exponentially distributed, and the proportion of patients without reactivation (P) will follow the equation:
P=P0e−k(t−t0)
(2)
Where P0 is the initial number of patients, k is the frequency of reactivation (ie: reactivation occurs once every 1/k days), and t0 is the minimal time to detection (as a result of ART-washout and the time taken for virus to grow from the initial level of viral infection to the level of detection (summarized in Fig 1). The equation was fitted to the data using the least squares method.
In order to compare rates of reactivation between studies, we allowed the initial delay to detection to be an independent parameter for each study (since both the ART drugs and threshold of detection varied between groups), and estimated the optimal frequency of recrudescence (1/k) across all groups. To investigate whether the frequency of recrudescence (1/k) was significantly different between groups we used an F-test.
In order to estimate whether differences in viral growth could account for the observed delays to detection, we estimated viral growth rates from the viral load data, assuming exponential growth of the virus. We then investigated whether growth rate was correlated with time-to-detection, as would be expected if delayed detection occurred due to slower viral growth. In the first cohort of 9 patients sampled frequently, we used linear regression to estimate the slope of log-transformed viral load with time, using 2–5 sequential viral load measurements. In the second cohort of 59 patients sampled weekly, we estimated viral growth rate using a two-point estimate of the growth between the first and second positive viral load samples. Note that this may tend to underestimate viral growth if it growth slows as viral load increases. Moreover, there will be a tendency for patients detected with a lower viral load to have a faster growth rate (because growth is measured at an earlier (and thus faster) stage of infection). The observation of lower viral loads in patients detected in the first week (Fig 3E) is likely an artefact of the pharmacokinetic delays before drug was fully eliminated and viral growth was possible in the first days after interruption.
An additional assumption in our analysis is that the viral growth rate in plasma at the time of detection is reflective of (or at least proportional to) viral growth early after viral reactivation.
In order to detect statistically significant differences of hazard ratios (HR) by Log-Rank Test with the level of significance α and power 1-β we need to have sufficient number of patients in each arm of the experiment. For estimation of this number we first need to estimate the number of events (recrudescence of virus) (m) and for this purpose used a formula, which assumes the equal number of patients in each arm [44];
m=4(zα/2+zβ)2/θ
(3)
where θ = Ln(HR).
However, if the rate of detection is not high enough to observe all patients in given time window of follow up, then the number of events will be lower than the total number of patients. Thus we need to correct the value of m by the fraction of patients with detectable virus at the end of the study. Assuming the exponential time to detection with the rate of detection estimated in our study (k) we can write the formula that relates the reduction in reactivation rate and the number of patients in one arm of the study.
n=4(zα/2+zβ)2ln(1−p/100)2(1−e−kt)(1−e−p100kt)
(4)
where p is the percent reduction in reactivation rate, t is the time window of analysis.
Sequence data on viral quasispecies after ART-interruption from the Joos study were obtained from Genbank (Genbank accession numbers listed in the original publication [25]). The data were analysed using a ‘highlighter plot’ (www.hiv.lanl.gov) to identify the relationships between different viral species within a given patient (see S3 File). Six patients were identified in whom we could distinguish and count the frequency of founder viruses early after ART-interruption, and this data was used to find the ratio of the number of copies of each observed founder virus in a patient to the next largest founder.
To estimate the frequency of reactivation from the ratio of founder viruses, we assumed an exponential time-to-initiation of viral growth, and exponential growth of virus during the initial phase of infection. We can then write down a formula for the expected ratios (R) between the sizes of subsequent founders:
R=V0egt1V0egt2=egΔ
(5)
Where g is the growth rate of virus (= 0.4 day-1), Δ = t1 ‒t2 is the delay between successive initiation events at times t1 and t2, and V0 is the initial concentration of virus. The distribution of delays between the initiation of growth of different founders (and thus their ratios) will be determined by the frequency of initiation of viral growth after ART-interruption. We assume that that Δ has an exponential distribution with parameter λ and can then derive a formula for the probability density function (PDF) of the expected ratios (h(y)) using the formula for distribution function of a random variable.
Where fexp(λ, x) is the probability density function (PDF) of the exponential distribution. The cumulative distribution function (CDF) of the ratios, H(y), can be defined by:
H(y)= Fexp(λ, ln(y)/g)
(7)
Where Ftrexp(λ, x) is the CDF of the exponential distribution.
By using maximum likelihood estimation to fit the observed ratios between the number of copies of founders to h(y) we are able to estimate the rate of successful reactivation λ.
We note that this analysis implicitly assumes that different founders grow at the same rate. It is also possible that individual founder viruses grow at different rates. However, as long as the growth rate is independent of the reactivation time, this should not significantly affect the expected distribution of founder ratios.
|
10.1371/journal.pbio.1001966 | Melanoma Cells Break Down LPA to Establish Local Gradients That Drive Chemotactic Dispersal | The high mortality of melanoma is caused by rapid spread of cancer cells, which occurs unusually early in tumour evolution. Unlike most solid tumours, thickness rather than cytological markers or differentiation is the best guide to metastatic potential. Multiple stimuli that drive melanoma cell migration have been described, but it is not clear which are responsible for invasion, nor if chemotactic gradients exist in real tumours. In a chamber-based assay for melanoma dispersal, we find that cells migrate efficiently away from one another, even in initially homogeneous medium. This dispersal is driven by positive chemotaxis rather than chemorepulsion or contact inhibition. The principal chemoattractant, unexpectedly active across all tumour stages, is the lipid agonist lysophosphatidic acid (LPA) acting through the LPA receptor LPAR1. LPA induces chemotaxis of remarkable accuracy, and is both necessary and sufficient for chemotaxis and invasion in 2-D and 3-D assays. Growth factors, often described as tumour attractants, cause negligible chemotaxis themselves, but potentiate chemotaxis to LPA. Cells rapidly break down LPA present at substantial levels in culture medium and normal skin to generate outward-facing gradients. We measure LPA gradients across the margins of melanomas in vivo, confirming the physiological importance of our results. We conclude that LPA chemotaxis provides a strong drive for melanoma cells to invade outwards. Cells create their own gradients by acting as a sink, breaking down locally present LPA, and thus forming a gradient that is low in the tumour and high in the surrounding areas. The key step is not acquisition of sensitivity to the chemoattractant, but rather the tumour growing to break down enough LPA to form a gradient. Thus the stimulus that drives cell dispersal is not the presence of LPA itself, but the self-generated, outward-directed gradient.
| Melanoma is feared because it spreads very rapidly when tumours are relatively small. It is not known why this metastasis is so efficient and aggressive. In particular, it is not known what drives melanoma cells to start to migrate out from the tumour. Here, we have studied the chemical signals that guide the migration of melanoma cells. We find that a component of serum, lysophosphatidic acid (LPA), functions as a remarkably strong attractant for all of the melanoma cells that we examined. We also observe that melanoma cells rapidly break down LPA. We conclude that melanomas create their own gradients of LPA, with low LPA in the tumour and high LPA outside. Since melanoma cells are attracted by LPA, this LPA gradient around the melanomas serves as a signal that drives the tumour cells out into the surrounding skin and blood vessels. Finally, we show that such gradients exist in a mouse model of melanoma. Self-generated LPA gradients are therefore an intriguing new driver for melanoma dispersal.
| Melanoma is an unusually aggressive cancer, which often metastasizes early during tumour development [1]. Tumours that have not clinically metastasized are frequently curable, but patients are far less likely to survive if tumours have metastasized before they are surgically removed, and metastasis is the principal cause of cancer mortality [2]. The most influential prognostic factor in predicting metastasis and survival is the thickness of the tumour (the “Breslow depth”) [3]. There is a dramatic increase in the risk of metastasis with only millimeter increases in Breslow depth [3]. This characteristic is unlike most solid tumours, in which the cytological morphology of the tumour cells and the individual genes mutated in the cancer are more important than size alone. Metastasis is therefore an important, and undermedicated, potential target for cancer therapy [4],[5].
One principal reason behind the aggressiveness of melanoma derives from the developmental history of melanocytes, the pigment producing cells in the skin that mutate to form melanomas. During mammalian development melanoblasts, the melanocyte precursors, emerge from a restricted location at the neural crest, and migrate rapidly from there throughout the developing dermis, before maturing into melanocytes on the basement membrane of the epidermis [6]. Thus a substantial level of cell migration is required for even skin pigmentation. Even in adults—for example following treatment for vitiligo—melanocytes can spread significant distances from the hair follicles to repopulate the surrounding skin. The melanocyte lineage is thus inherently migratory.
However, several questions about melanoma progression remain unanswered. The first is what drives melanomas to change from the relatively benign radial growth phase (RGP) to the far more invasive vertical growth phase (VGP) (see schematic diagram in Figure 1A). In RGP melanomas, cells only spread horizontally along the basement membrane, compared to VGP melanoma cells, which are also capable of spreading both upwards into the epidermis (Pagetoid spread) and downwards, into and through the dermis (invasion). This spread raises the related question, of what drives cells to migrate away from the primary tumour. Simple, random migration is an extremely inefficient way of dispersing cells and also unlikely to drive cells to invade through matrix and basement membranes. Chemotaxis—cell migration directed by gradients of soluble signalling molecules—is implicated as an important driver of metastasis by a wide range of data [7],[8], and is considered necessary to drive efficient invasion. In breast cancer, for example, some tumour cells migrate towards epidermal growth factor (EGF) [9]. However, EGF gradients have only been inferred in vivo, never measured, and their sources are usually unclear. In the case of breast cancer, the EGF is thought to be secreted by macrophages recruited in a paracrine loop by the tumour [10], but for other attractants and cell types the sources of chemotactic signals are not known.
In the melanoma literature, most chemotaxis is attributed to growth factors such as platelet-derived growth factor (PDGF) and EGF [11] and the CXCR4 ligand SDF-1 [12], though a wide variety of potential attractants have been discussed [13]. Gradients of growth factor or SDF-1 have not been identified in vivo, they can only be inferred from the cells' behaviour or pattern of responses in vitro.
Chemotaxis assays are typically performed in transwell chambers, in which cells are grown on one side of a membrane filter and potential attractants are added to the other side. Chemotaxis is assayed by the number of cells observed on the far side of the filter after a fixed interval. These assays are subject to a wide range of artifacts. Cells' behaviour during chemotaxis cannot be studied, which makes it extremely difficult to distinguish chemotaxis from directionless changes in migratory behaviour (i.e., chemokinesis [14]). Potential attractants form extremely steep and rather short-lived concentration gradients, unlike the physiological conditions the assay aims to reproduce. More seriously still, conditions either side of the filter may be discretely different; cells may grow, survive, or adhere better on one side of the filter than the other, giving changes in the numbers of cells that can be artifactually interpreted as chemotaxis. Direct viewing chambers, such as Dunn, Zigmond, or Insall chambers, are more laborious to use but yield a far higher quality of data, with fewer artifacts [15]–[17]. In work described here, we use direct-viewing chambers to identify lysophosphatidic acid (LPA) as a far more potent chemoattractant for melanoma cells than other previously described attractants. We have developed and refined two direct-viewing assays to assess mechanisms of cell dispersal and chemotaxis, allowing us to distinguish chemotactic from chemokinetic and contact-driven responses under defined conditions that minimize artifacts. Furthermore, the use of direct-viewing chambers makes comparison of attractants' relative efficiencies practical.
The suggested role of chemoattractants in cancer dispersal—whether growth factors, chemokines, or LPA—raises the crucial question of how gradients are generated. Chemotaxis will only work with signals that are presented as gradients—homogeneous signals contain no directional information—and the steeper the gradient, the more efficient the chemotaxis. Chemical gradients are typically effective over distances of less than a millimetre—limits on the efficiency of diffusion make larger gradients impractical [18]. Thus for a gradient to be formed there must be a gradient source that is close to the tumour.
Alternatively, local gradients may be formed from signals that are widely produced, but are absorbed or broken down locally. This local depletion mechanism is potentially just as effective as local production, but less often invoked. In the cancer literature, only localised sources are typically invoked, for example individual macrophages within the vasculature attracting cancer cells within the tumour [10].
If cells that are responding to a stimulus are also responsible for breaking it down, the result is a self-generated gradient. Under these conditions the gradient is always oriented away from the current location of the cells. One such example has been shown during the development of the zebrafish lateral line primordium [19]–[21], in which a dummy receptor locally absorbs an SDF-1 stimulus to set up a gradient that is detected by a different receptor. In this work we find that melanoma cells self-generate chemotactic gradients from unlocalised, exogenous LPA. These gradients tend to direct cells to disperse outwards from tumours, thus directly promoting metastasis. Furthermore, we measure LPA gradients across real melanomas in vivo. Since melanomas of sufficient size both generate their own LPA gradients and respond to them, chemotaxis-steered spread of melanomas is almost inevitable.
To examine the signals that drive the spread of melanoma cells, we set up 2-D assays for tumour cell spread using a direct-viewing chemotaxis chamber that allows detailed analysis of cell migration [15]. The chamber contains two wells, connected by a bridge that allows diffusion of attractants but not flow. Both cells were homogeneously filled with complete medium, but cultured melanoma cells [22] were only seeded in one well, at a range of different densities.
Our initial results were surprising: Cells consistently spread outwards from the well in which they started, even in uniform medium without an externally applied gradient (Figure 1B; Movie S1). This effect was density-dependent; cells plated at 2×103 or 6×103 cells/well barely migrated, while 2×104 cells/well migrated up to 350 µm in 24 hours (Figure 1C and 1D). This behaviour strikingly resembles the behaviour of real melanomas, in which the chance of metastasis is more correlated with tumour thickness than any other parameter [3].
This type of density-dependent spreading requires individual cells (or small clusters of cells) to migrate away from the bulk population. This dispersal occurred in our assays; cells moved directly away from the well they resided in with unprecedented accuracy (Movie S1). This directed, non-random migration can only occur if the moving cells perceive a directional cue from the bulk population of the cells to spread. We therefore analyzed the nature of the signal that was directing cells away from the bulk population. The most probable signalling mechanisms are contact inhibition of migration [23] or chemotaxis. We therefore examined these potential mechanisms in turn.
Contact inhibition (of migration, as opposed to the more frequently described contact inhibition of growth) is an effective mechanism for short-range dispersal in which cell∶cell contact directs cells away from one another. It has been shown in other neural crest-derived cell types [24]. However we found no evidence to suggest it drives cell dispersal in our assays. Movie S2 shows one example in which cells spread both individually and while contacting one another. Some cells steer accurately outwards through multiple cycles of new pseudopods independently of cell∶cell contact. Others continue to migrate outwards when contacting the cell in front, where contact inhibition predicts these cells should reverse into the space behind them. Analysis of the paths of individual cells (Figure S1) shows that cell-cell contact is not steering cells; the paths of cells that are contacting others, have recently contacted others, and are not in contact are strikingly similar. The one apparent example of contact inhibition (Movie S2, cell 2) changed the cell's direction but did not improve its outward accuracy. Thus while these cells may experience contact inhibition, we considered chemotaxis as the most likely mechanism steering them away from the main population.
Cells could generate chemotactic gradients to drive dispersal by either of two mechanisms. They could secrete an autocrine chemorepellent and migrate away from it. We have previously shown this to be a key driver of Entamoeba pathogenesis [25], in which chemotaxis away from ethanol generated by the amoebas themselves causes cells to migrate from the lumen of the gut into the walls of the gut and eventually the liver of the patient. Alternatively, the melanoma cells could locally break down or consume a chemoattractant that is produced externally, but spatially homogeneously [26],[27], as seen in the zebrafish lateral line primordium [19],[21]. In either case, dense populations of cells create a gradient that consistently directs migration away from themselves. We considered that homogeneous attractants would most likely derive from the serum added to full medium. To find if dispersal used a repellent or a consumed attractant, we compared cell dispersal in serum-free and normal medium. Cells in serum-free medium are healthy and motile in control movies, but do not migrate away from one another (Figure 1E), demonstrating that the cells do not secrete chemorepellents. We also compared cells moving out of fresh medium into serum-free and full medium. Cells dispersed far more efficiently into the rich medium (Figure 1F), implying that they are driven by attractants in fresh medium rather than an inhibitor whose production depends on serum.
To test whether consumption of a component of serum produces a positive chemotaxis response, we compared migration in uniform serum to an assay in which cells are exposed to a gradient between serum-free medium and medium supplemented with 10% serum (Movie S3). We found that both assays produced similar directed migratory responses; cells migrated towards the opposite well with or without a preformed serum gradient (Figure 1G). This finding further supports the concept that the outward migration is driven by positive chemotaxis, most likely towards a chemoattractant globally present in the serum but depleted around the cells.
We tested this hypothesis using a more traditional chemotaxis assay, in which cells are spread homogeneously over the field at the start of the assay, giving the cells the opportunity to move in any direction [14]. We loaded cells into the chamber in complete medium that had been conditioned by melanoma cells for 48 hours, then replaced the medium in one well with fresh medium containing 10% serum. The cells migrated towards the well containing fresh medium very efficiently (Figure 2A and 2B), showing that an attractant in fresh medium is consumed by the melanoma cells.
We confirmed that chemoattractants are present in normal serum by exposing melanoma cells—again homogeneously seeded in the chemotaxis chamber—to exogenous gradients of serum. In homogeneous serum-free medium the cells were healthy, and migrated, but randomly (Figure 2C). When a gradient of serum was applied, the cells migrated towards the higher concentrations with unprecedented precision (Figure 2D); their paths are overwhelmingly oriented up-gradient, in a manner more usually associated with neutrophils and Dictyostelium [28] than cancer cells, which typically chemotax less accurately [29]. The high chemotactic index was maintained throughout a sustained period, with narrow and accurate confidence interval, and strongly significant Rayleigh test [30] for directional migration (Figure 2E). Thus serum contains a remarkably potent chemoattractant for melanoma cells.
We therefore conclude that melanoma dispersal across the chamber is driven by positive chemotaxis towards an attractant that is present in serum. The attractant is broken down by the cells themselves into a gradient that efficiently disperses cells.
One potential explanation for cancer cells becoming metastatic is that they evolve chemotactic competence as the tumours develop [13],[31],[32], and thus move from unsteered to steered migration. We therefore examined the ability of a panel of cell lines isolated from different tumour stages and selected for physiologically appropriate behaviour (Figure 3A) [22]. Surprisingly, all the lines we examined responded chemotactically to serum gradients (Figure 3B). Cells from metastases were more motile than cells from earlier stages (Figure 3C); highly invasive (VGP) cells were slightly more accurate, but not significantly faster than the biologically earlier, RGP cells. Cells from more advanced tumours responded more robustly, but the progression from nonmetastatic to metastatic was not marked by the cells newly acquiring responsiveness—all lines examined were chemotactic enough to spread away from the tumour efficiently in the presence of an appropriate gradient. Several lines of data suggest that genetic and epigenetic changes during progression from RGP to VGP increase cells' ability to survive [33]; our data imply that it is cell survival, rather than chemotactic sensitivity, that defines the difference. The increase in migratory ability could modulate cells' ability to escape from a primary tumour, but our principal conclusion is that melanoma cells from all stages are chemotactic.
There are multiple reports of chemotaxis driving metastasis of melanoma and other tumour cells, in particular breast cancer. Published accounts of chemotactic invasion most often describe growth factors as the attractants—for example EGF for solid tumours [34], and EGF, hepatocyte growth factor (HGF), and stem cell factor (SCF)/KitL for melanoma [13]. However these attractants were often identified in transwell chambers, which as earlier discussed are subject to a range of artifacts, in particular false positive. For example, the positive well might promote survival, growth, or adhesion of cells that move randomly across the membrane. Our direct-viewing chambers provide a far more rigorous analysis. We therefore tested a broad range of attractants in our assays. To our surprise, no growth factor acted as an attractant to any measurable degree (Figure 4A); steep or shallow gradients gave no obvious movement upgradient, and no significant chemotactic index towards any growth factor tested (Figure 4B). We therefore conclude that the chemotaxis towards serum we observed was unlikely to be towards growth factors. This does not, of course, demonstrate that melanoma cells are never chemotactic towards growth factors; but it clearly shows the surprising and efficient chemotaxis towards serum observed earlier is mediated by another molecule.
EGF and PDGF did increase cells' speed (Figure 4C), but they did not provide directional specificity. They therefore acted as chemokines, regulating overall cell behaviour, rather than as chemoattractants that could steer the cells.
The striking accuracy of chemotaxis demonstrated by melanoma cells towards serum was more reminiscent of neutrophil chemotaxis towards formyl peptides, or Dictyostelium towards cAMP, which signal through G-protein coupled receptors (GPCRs) rather than growth factor receptors like EGFR and PDGFR. We therefore investigated SDF-1, the ligand for the GPCR CXCR4, which has been associated with poor prognosis and malignancy of melanoma [35]; but again, it was not measurably attractive to cells in our assays (Figure 4B, compare with strong response to serum).
However, LPA, another well-known component of serum that signals through GPCRs, was strikingly attractive to melanoma cells. A gradient from 0 to 1 µM LPA across the chamber (consistent with the approximate concentration of LPA in serum; see below) induced chemotaxis almost as effectively as 0%–10% serum (Figure 4D), yielding a comparable chemotactic index (Figure 4E). This was a surprise: LPA is more typically described as an inflammatory mitogen, acting on haematopoietic cells such as macrophages. It appears frequently in the cancer literature, but more often as a mitogen and chemokine for cancer cells, acting via autotaxin, which catalyzes the production of LPA from lysophosphatidylcholine [36]. However in our assays the chemotaxis of melanoma to LPA was again remarkably accurate compared with the weaker chemotaxis typically seen in cancer cells [37].
To examine whether LPA was the principal attractive component of serum, we assayed chemotaxis in the presence of the antagonist Ki16425, which specifically inhibits binding to LPA receptors 1 and 3 [38]. The effects were again remarkably clear. 10 µM Ki16425 blocked cell spread in our original, density-dependent assay (Movie S4) and chemotaxis towards 10% serum (Figure 5A; Movie S5), reducing the chemotactic index from more than +0.4 to zero (Figure 5B). Ki16425-treated cells were obviously healthy, and moved similarly to untreated cells, with similar track lengths, showing that the treatment was not making the cells nonspecifically sick or non-motile. Knockdown of LPAR1 by siRNA had a similar effect (Figure S2A), showing that LPAR1 is the key receptor for this process, and 10 µM Ki16425 also blocked chemotaxis towards pure LPA (Figure S2B). Again, LPA chemotaxis is not tumour stage-specific; Ki16425 blocked chemotaxis in all cell lines from all stages of cancer progression (Figure 5C). RGP and VGP cell lines were completely inhibited, and the highly motile metastatic lines were substantially inhibited. The residual chemotaxis in the presence of inhibitor could represent either incomplete inhibition by the antagonist, or a small amount of chemotaxis to another agent. From these data, we conclude that LPA is overwhelmingly the dominant chemoattractant in serum for all lines examined.
While chamber-based assays are optimized to allow accurate and detailed recording, they provide a 2-D view of a process that more often happens in 3-D tissues in vivo [39]. We therefore examined the role of LPA in a widely used organotypic tumour cell invasion model [40]. In this system melanoma cells are added to the top of a plug of collagen in which fibroblasts are growing, and over time they migrate vertically downwards into the 3-D matrix. During the course of the assay, the collagen plug is set so only its bottom face contacts the medium, at which point malignant melanoma cells invade downwards [41]. We hypothesized that the melanoma cells were driven by a self-generated LPA gradient as in Figure 1B, once fresh LPA could only be supplied from the bottom. This hypothesis is supported by assays in which the collagen plugs remain submerged, and no invasion is seen (Figure S3), further rejecting contact inhibition of migration as a mechanism of invasion. When the gels were treated with Ki16425, the melanoma cells did not invade downwards into the gel (despite comparable numbers of cells at the end, showing no change in growth or survival). Quantitative analysis confirms that Ki16425 strongly inhibited invasion in both cell lines that were invasive in this assay (Figure 5D and 5E). Thus LPA is a dominant steering system for 3-D organotypic assays, as well as for 2-D chamber assays.
Our earlier data (Figures 1B and 2A, in particular) showed that melanoma cells disperse by depleting a chemoattractant from serum. We therefore tested whether melanoma cells are able to deplete LPA from their surroundings. Full medium with and without serum was incubated with different densities of melanoma cells for different times, then LPA was extracted from the conditioned medium and analyzed by mass spectrometry [42]. This confirms that the melanoma cells effectively break down LPA; the conditioned medium was depleted in a density-dependent manner (Figure 6A) and in a timescale that correlates with the medium conditioning experiments in Figure 2A and 2B. One advantage of using mass spectrometry is the identification of molecular subspecies. The biological activity of LPA is known to vary with its structure [43],[44]. In particular, there is a strong correlation between biological activity and the degree of polyunsaturation, and also acyl chain length [45]. Melanoma cells broke down the biologically active species more rapidly than the others (Figure 6B), ensuring that the most active species also formed the steepest gradients.
The results we have obtained conflict with the established dogma that growth factors are primary melanoma chemoattractants [13]. To reconcile these accounts with our data, we examined the role of growth factors during chemotaxis towards LPA. As shown previously (Figure 4C), EGF and (particularly) PDGF increased the basal speed of cells. Gradients of EGF and PDGF, and mixtures of both, enhanced the accuracy of chemotaxis to LPA (Figure 7); LPA, EGF, and PDGF together in serum-free minimal medium were as effective as 10% serum. Most tellingly, however, when cells were presented with LPA and growth factor gradients oriented in opposite directions, they chemotaxed towards the LPA not the growth factors; if anything they migrated towards the LPA with enhanced efficiency (Figure 7B, bottom two lines). Thus when examined in the high levels of detail afforded by our chambers, the growth factors are potentially important accessory factors that increase cell speed and efficiency of chemotaxis, but they do not themselves act as chemoattractants.
These results are reminiscent of observations of development in vivo, in which the growth factor SCF promotes migration but not direction of melanoblast migration [46]. It is possible that the melanoma chemotaxis to growth factors observed in other work [13] is due to changes in speed alone, which as discussed earlier can cause a false positive in transwell assays. It has also been shown that growth factors can cause cancer cells to secrete LPA [47], which could also provide an element of indirect chemotaxis in many types of assay.
We have clearly shown that LPA is a potent chemoattractant for melanoma cells of all biological stages. To determine whether this chemotaxis was an important driver of melanoma chemotaxis in vivo, we investigated whether the tissue surrounding real melanomas contained LPA gradients that would direct cells out of tumours. Mice that are heterozygotes for the driver mutation BrafV600E (the most prevalent driver of human melanomas) and deletion of the tumour suppressor PTEN develop sporadic melanomas (Figure 8A) genetically and cytologically comparable to human tumours (Figure 8B). We took punch biopsies from the tissue in and across melanomas (Figure 8C) from several mice, extracted total lipids, and examined LPA levels using mass spectrometry. In all non-ulcerated melanomas we examined, LPA levels were low inside the tumour, higher at the edges, and higher still in the tissues immediately outside the tumour (Figure 8D). Cells at the edges of the tumour are therefore experiencing an outward-oriented LPA gradient tending to drive them out into surrounding tissues and vasculature.
We further examined the LPA species in the tissue. Forms that are strongly associated with signalling, in particular 18∶2-LPA and 20∶4-LPA [48], formed the steepest gradients (Figure 8E), while gradients of non-signalling forms such as 18∶0-LPA were flatter. This finding further supports the idea that the gradients of LPA are specifically produced as signals targeted at LPA receptors.
This study is, to our knowledge, the first time a chemotactic gradient has been directly measured around tumours in vivo. There are a number of situations where the presence of a gradient has been inferred from cellular behaviour, most prominently in the paracrine loops shown by Segall and others [10]. However, such gradients must by definition be local and tend to be transient. The gradients we observe in melanomas are clear, large-scale, and provide a convincing driver for cell dispersal, and one highly plausible explanation of why melanomas above a certain size, and hence Breslow thickness, always tend to be metastatic.
In this work, we have shown that LPA is a potent chemoattractant for melanoma cells in general, and that outward-oriented gradients of LPA are self-generated by melanoma cells. Because self-generated gradients are always oriented away from tumours, this combination provides a plausible mechanism for driving tumour cell dispersal. We do not exclude other mechanisms; it has for example been proposed that LPA regulates cadherin levels [49], which would not be visible in our assays. Growth factor chemotaxis may be visible under the appropriate conditions (though, as discussed previously, many data are from transwell assays, which are artifact-prone and unreliable). Likewise, we do not exclude other mechanisms than chemotaxis. Contact inhibition of migration occurs in many cell types derived from the neural crest and so is probably found in melanoma, and defects in cell growth and survival in inappropriate locations are of course important factors. But the mechanism we have found that overwhelmingly dominates the dispersal in our assays is robust and is apparently active in a high proportion of melanomas. It is therefore likely to be a particularly important mediator of tumour cell dispersal. We hypothesize that similar mechanisms will be common in cancer metastasis.
The source of LPA around melanomas is unknown. In many tumours, including melanoma, expression of autotaxin and thus autocrine production of LPA has been associated with tumour progression [50]. This LPA production appears to be a mechanism for promoting melanoma growth, rather than driving chemotaxis and invasion. LPA generated by the tumour itself would be found at a higher level in the tumour than outside it, which would oppose outward dispersal and thus metastasis. Rather, we find that the melanoma cells in culture and in tissues break down externally generated LPA, making outward-facing gradients. LPA is therefore more likely to be generated through inflammatory processes—haematopoietic cells, in particular, are a principal source of LPA in tissues [51] —or by inducing LPA production from stromal cells. In metastatic breast cancer xenografts, expression of LPA receptor promotes cell growth and metastasis, but the LPA is made locally by platelets, which are in turn recruited by many tumours [52]. Platelets are also a rich source of growth factors [53]. Our data therefore implicate inflammation in initiating melanoma spread. This finding has important implications for therapy. Interventions that promote inflammation without removing the entire tumour could be extremely dangerous—diagnostic punch biopsies, in particular, could promote a wave of metastasis in response to LPA released by inflammation. From a therapeutic perspective, data from epidemiological studies suggests the anti-inflammatory drug aspirin can protect against metastasis [54].
The increased speed of the metastatic cells may be important, but may also be an artifact of selection. It remains unclear whether the increased speed of migration is clinically important, or whether the fastest strains will metastasize earlier, and thus be the first to be identified. Our data suggest that even less invasive cells move rapidly and accurately enough to metastasize, but our assays may miss factors that retard cell migration.
We have shown that cultured melanoma cells from throughout tumour evolution are chemotactic towards LPA in transwell assays. A recent paper has reported the opposite, that LPA is a chemorepellent for B16 cells [37]. This seems a cell-line specific effect, as these highly derived and divergent cells do not express the LPAR1 and LPAR3 receptors, which are usually highly expressed and dominate LPA chemotaxis in our assays (Figure S2A).
We have found that melanomas generate their own chemotactic gradients from homogeneous LPA that is exogenously provided. LPA chemotaxis is an essential feature driving melanoma invasion in 3-D organotypic assays. We have also shown that real tumours create a chemotactic gradient of LPA in vivo. Taken together, these lines of evidence suggest a model of chemotaxis towards self-generated LPA gradients is a major driving force for melanoma dispersal (Figure 9). One unforeseen advantage of this model is that it also provides a simple unifying explanation for upward or pagetoid spread, which is a hallmark of the invasive VGP stage melanoma.
We have measured actual LPA gradients in animals with experimentally induced melanomas. We have also shown that all the melanoma cells we tested perform chemotaxis towards LPA gradients, in both 2-D and 3-D assays. It is thus reasonable to conclude that LPA gradients are sufficient signals to mediate melanoma cell dispersal.
To test whether LPA is necessary for melanoma metastasis in vivo will be very difficult. Our hypothesis is that LPA gradients drive intravasation from the tumour towards local blood vessels. Many widely used metastasis assays, for example tail-vein injection, completely miss this step. Slower assays, for example subcutaneously injected xenografts, metastasize impractically slowly, and to nonphysiological targets, in particular the lymph nodes. Pharmacological approaches, for example blockade of the LPA signalling system by LPA antagonists, are confounded by the importance of LPA to the vascular and haematopoietic systems. A mouse model of melanoma that metastasizes through a physiological route and can be crossed with inducible LPA receptor knockouts does not currently exist; when it is developed, such a model will be the ideal system for testing our model in vivo.
The most important message from this work is that it is the gradient of LPA—not the presence of LPA per se—that contains the information. LPA is a very prevalent molecule. It is present at high levels in serum, and may be generated within tumours by cancer cells or exogenously by, for example, platelet activation. Interestingly, cells ahead of the main group do not respond even when an external gradient is applied (in Movie S3, for example). Presumably these cells reach a region where LPA levels are homogeneously high, at which point there is little or no guidance information available to them. Likewise, if too few cells are used in the spread cell assay, no chemotaxis is observed, suggesting that LPA breakdown is important even in classical chamber assays. We suspect that LPA is not a chemoattractant for low densities of cells, because they cannot break it down rapidly enough to form an appropriate local gradient. In our invasion assays, LPA becomes an attractant when—counterintuitively—cells are present at high enough densities to break down most of it. This means that the LPA gradient is self-generated by the melanoma.
Self-generated gradients are currently highly topical. Recent papers showing the detailed roles of the CXCR4 and CXCR7 receptors (which respond to and deplete SDF-1, respectively) during the formation of the zebrafish lateral line have caused a spike in interest, but other methods whereby cells drive creation of attractant gradients then respond to them occur in multiple systems, especially during embryonic development [26],[27],[55],[56]. More generally self-generation provides a means whereby cells can maintain a directional cue over distances that are far too large for premade gradients. Furthermore, with externally formed gradients, the information that specifies the gradient must come from somewhere else. If an external gradient attracts cells during development, the secret to understanding the process lies with understanding where and by whom the attractant is being made. Self-generated gradients are different; there is no need for external information. The gradient is generated as an emergent property of the interaction between the cells and their environment. Thisconclusion is perhaps the most interesting feature of this work. In LPA chemotaxis during melanoma metastasis, there is no need for any other cell type to set up a local gradient. The melanoma cells first generate a gradient—once the tumour is thick enough—and then respond to it by migrating away. Thus the melanoma drives its own metastasis.
All mice used were control cohorts from other studies. Before they were humanely killed, all mice had reached the primary or secondary end-points of their designated study.
All melanoma cell lines used are listed by biological stage of derivation and were transferred from the Wellcome Trust Functional Genomics Cell Bank (Biomedical Sciences Research Centre, St. George's, University of London).
Cells were maintained in Roswell Park Memorial Institute (RPMI, Invitrogen) 1640 medium, supplemented with 10% fetal bovine serum (FBS) (PAA Labs), 2 mM L-Glutamine (Gibco, Invitrogen), and 1% penicillin and streptomycin (Gibco, Invitrogen).
siRNA constructs were obtained from QIAGEN and transfected as per instructions. WM239A cells were challenged twice with siRNA, 48 hours apart, then used in the assay 48 hours after the second transfection.
Insall chambers were manufactured and used as described [15]. The chambers were drilled in advance with a 1.3 mm drill bit using an overhead drill press. During drilling, the chamber was secured in a small machine vice sitting inside a V-block at 45° and a hole was drilled into each “rabbit ear” of the outer well to allow reverse filling. Cells were starved in PBS for 12 hours then seeded at a density of 5.5×104 cells/ml in CGM. Each cover slip was coated with 2 ml of the seeding suspension. After seeding cells, the six-well dish was shaken in the x and then y planes for 5 seconds each and placed in a CO2 incubator at 37°C on top of a shock absorbent base to prevent vibration induced patterns of cell accumulation.
VALAP sealant (vaseline, lanolin, and paraffin) was prepared by combining the three components together in a weight ratio 1∶1∶1 and melting at 100°C on a heat block. A fine artist's paint brush was used to apply the VALAP.
Cover slips were treated with human fibronectin (BD Biosciences) 1 mg/ml throughout, generating an adsorbed concentration of 4.17 µg/cm2 in the range of 1–5 µg/cm2 as suggested by the manufacturer. Following fibronectin coverslips were passivated with 0.5% (w/v) heat-treated BSA solution in PBS.
Chemoattractants were added to serum-free RPMI medium as required. Addition of 5 mM HEPES to the media in the sealed chamber is essential to buffer the pH of the media throughout the experiment. LPA (Sigma) was dissolved in a 1∶1 ratio of distilled water: absolute ethanol to generate a 1 mM stock solution and stored at −20°C. To use this as a chemoattractant, BSA was diluted to a final concentration of 0.05% (w/v) to SFM-H (SFM-HB) and then 1 µl LPA was added to 1 ml to generate a 1 µM LPA solution. EGF (Peprotech), PDGF, BB Homodimer (Calbiochem), HGF/Scatter Factor (Peprotech), and SDF-1α/CXCL12 (Peprotech) were dissolved in PBS to a stock concentration of 10–100 µg/ml, stored at −20°C and used as indicated.
Ki16425 (Cambridge Bio) was stored in absolute ethanol at a stock concentration of 10 mM as per the manufacturer's instructions. In Insall chamber assays, cells were pre-incubated for 5 minutes with a 10 µM solution before combining with reagents in the chamber at the same concentration.
We used a Nikon TE2000-E inverted time-lapse microscope equipped with a motorised stage (Prior) and Perfect Focus System (PFS) to prevent focal drift due to thermal fluctuations. The entire microscope was enclosed in a plexiglass box, humidified and maintained at 37°C with 5% CO2. The Insall chamber experiments did not require the addition of supplementary CO2. Our microscope system was driven by Metamorph software (Molecular Devices) and the x, y positions were manually selected and pre-loaded.
Images were processed using ImageJ (http://rsb.info.nih.gov/ij/), if necessary using the Image stabilizer plugin (http://www.kangli.org/code/Image_Stabilizer.html) to correct for drift. Cells were tracked using MtrackJ (http://www.imagescience.org/meijering/software/mtrackj/) to follow the path of the cell nucleus For consistency, we attempted to track a minimum of 40 cells in every chamber assay; in most cases this was sufficient to ensure statistical significance. The following criteria were used for deciding which cells to track: cells that moved more than 1 cell length in 24 hours; cells that tracked continuously until the end of the experiment or until the cell migrated off the bridge or rounded up in preparation for mitosis; cells were excluded that migrated onto the bridge during the experiment; avoided tracking post-mitotic cells.
We developed an Excel spreadsheet (written by DMV and AJM-M) to facilitate the processing, analysis, and quantification. This spreadsheet automatically produces spider plots, speed, and chemotaxis index data over time. A time window was selected (e.g., 6–12 hours for melanoma cells) and values zeroed within this window to produce end-point data. Chemotaxis index (cosθ) plots are presented as mean ± standard error of the mean (SEM). Cosθ is a function of the distance migrated in the direction of the gradient divided by the euclidian distance (the linear distance between the start and end position of the cell). These data were also processed in the Circstat toolbox for MATLAB by GK [30]. This process generated rose and polar plots with 95% confidence intervals and a Rayleigh test.
Conditioned media were generated as follows. A sub-confluent 10 cm petri dish of WM239A cells was washed 3× with PBS then cells were split in a 1∶5 ratio into five new 10 cm petri dishes and combined with fresh CGM to a final volume of 10 ml. Conditioned medium was then harvested from one dish per time-point, staggered between 0–48 hours (Marked T0, T6, etc.). All 10 ml was aliquoted into 10×1 ml eppendorf tubes. The samples were immediately frozen on dry ice before storing at −80°C. The cells in each dish were then counted. When needed aliquots of conditioned media were thawed at 37°C and centrifuged for 10 min using a lab top centrifuge, then filtering with a sterile 0.2 µm filter.
Collagen gels were prepared by combining 2 mg/ml rat tail collagen solution, 10× Minimum Essential Medium (MEM, Invitrogen), and 0.22 M NaOH in a ratio 8∶1∶1. The pH was finely adjusted to pH 7.2 with the 0.22M NaOH. One volume of FBS containing 7.5×105 primary human skin fibroblasts (passage 5–7) was immediately combined with 10 ml of the gel mixture on ice. After pipetting well, 2.5 ml of the gel and cell mixture was added to each 35 mm petri dish. The gels were then placed in a humidified incubator with 5% CO2 to set for 15–30 minutes. A further 1 ml MEM was added to each petri dish and the gels were carefully detached to enable gel contraction in the same incubator. The media was changed every 3 days. After 6–7 days the gels measured approximately 1.5 cm in diameter and were transferred to a 24-well dish ready for tumour cell seeding. 1–2×105 tumour cells were then counted and allowed to seed on the surface of each gel. The gel was carefully transferred with forceps to an elevated stainless steel grid (Sigma, screens for CD-1, size: 40 mesh) and placed in a 6 cm petri dish and this was denoted day 0. CGM was added to cover the grid and was then carefully aspirated to leave a meniscus around the base of each gel, thereby generating an air-liquid interface. Three gels were loaded onto each grid and the medium was changed three times weekly. In experiments using Ki16425, the gels with adherent cells were pre-incubated for 5 minutes with 10 µM Ki16425 in the CGM before raising the gels to the air-liquid interface. 10 µM Ki16425 was maintained in the CGM throughout the experiment with thrice weekly media changes as before. A typical experiment lasted 7–12 days.
At the end of the invasion assay, each gel was divided into two with a scalpel, fixed in 4% formaldehyde at 4°C and sectioned before being stained with haematoxylin and eosin.
We used the inducible Tyr::CreERT2 BRAFV600E/+ PTENlox/− melanoma model [57], in which the melanomas were all generated in mixed background mice from 6–12 weeks of age. Animals were treated with 2 mg tamoxifen topically to shaved back skin daily for 5 days. There was no discernable phenotype until naevi or primary melanomas started developing 6–8 weeks after induction predominantly on the treated area. Typical grooming behaviour spread the tamoxifen to other parts of the skin and/or was ingested leading to activation in other cutaneous regions. All mice used were control cohorts from other studies. Before they were killed, all mice had reached the primary or secondary end-points of their designated study.
Suitable mice were identified with at least one and up to four tumours, ideally located on the back. The smallest tumour size was 4×4 mm to enable at least two areas to be sampled. Skin containing the tumours was rapidly dissected off the back and pinned slightly taut to paper overlying a corkboard. Sterile Punch Biopsy tools (Stiefel) were used to punch circular samples from the tumour and surrounding skin. The size of punch biopsy depended on the tumour size and varied from 3–6 mm in diameter. Samples were taken at various locations across the tumour and were coded as follows: within the tumour (A), across the margin (B), 5 mm from the margin (C), and 10 mm from the margin (D). Samples were immediately snap frozen in liquid nitrogen and transferred to a −80°C freezer for storage.
Control samples of normal appearing skin in the same melanoma model activated with tamoxifen were used to calculate the basal level of LPA. Each section of mouse skin underwent a series of nine punch biopsies (A, B, and C in three replicate series).
Mice and human melanoma/skin samples (1–20 mg) were pulverised after thoroughly cooling with liquid nitrogen. The pulverised powder was suspended in 750 µl water then used for LPA extraction. For cell culture media samples, 750 µl of cell culture media was used for LPA extraction. Media or tissue samples were spiked with 50 ng of 17∶0-LPA as an internal standard before extraction. LPA was extracted with 1 ml n-butanol three times at room temperature.
The combined LPA extract was dried under vacuum with SpeedVac (Thermo) and re-dissolved in 60 µl chloroform/methanol/water 2∶5∶1. 14 µl was injected for liquid-chromatography with tandem mass spectrometry (LC-MS/MS) analysis. For LC-MS/MS analysis, we used a Thermo Orbitrap Elite system (Thermo Fisher) hyphenated with a five-channel online degasser, four-pump, column oven, and autosampler with cooler Shimadzu Prominence HPLC system (Shimadzu) for lipids analysis. High resolution/accurate mass and tandem MS were used for molecular species identification and quantification. The identity of the lipid subspecies was further confirmed by reference to appropriate lipids standards. All the solvents used for lipid extraction and LC-MS/MS analysis were LC-MS grade from Fisher Scientific.
The final amount of LPA (ng) is presented as a concentration per 750 ml of conditioned media analysed or per mg tissue. The data are represented graphically plotting mean ± SEM for the concentration of LPA versus conditioning time (for conditioned media samples); and distance from tumour margin (for tumour samples). Samples were normalised to position “A” for comparison between tissue samples.
|
10.1371/journal.pgen.1002386 | Genome Instability and Transcription Elongation Impairment in Human Cells Depleted of THO/TREX | THO/TREX connects transcription with genome integrity in yeast, but a role of mammalian THO in these processes is uncertain, which suggests a differential implication of mRNP biogenesis factors in genome integrity in yeast and humans. We show that human THO depletion impairs transcription elongation and mRNA export and increases instability associated with DNA breaks, leading to hyper-recombination and γH2AX and 53BP1 foci accumulation. This is accompanied by replication alteration as determined by DNA combing. Genome instability is R-loop–dependent, as deduced from the ability of the AID enzyme to increase DNA damage and of RNaseH to reduce it, or from the enhancement of R-loop–dependent class-switching caused by THOC1-depletion in CH12 murine cells. Therefore, mammalian THO prevents R-loop formation and has a role in genome dynamics and function consistent with an evolutionary conservation of the functional connection between these mRNP biogenesis factors and genome integrity that had not been anticipated.
| THO/TREX is an eukaryotic conserved complex, first identified in budding yeast, that acts at the interface between transcription and mRNP (ribonucleoprotein) export. In yeast, THO mutants show gene expression defects and a transcription-associated recombination phenotype. Despite the structural conservation of THO/TREX, it is unclear whether the functional relevance is the same in mammals, in which several reports have identified a role of THO/TREX separated from transcription. We have asked whether mammalian THO/TREX function is connected to transcription and whether this function is required to prevent R-loop formation and to maintain genome integrity. Our study reveals that depletion of human THO subunits, in particular THOC1/hHPR1, reduces transcription elongation, affects mRNA export, and increases genome instability associated with the accumulation of DNA breaks. This genome instability is R-loop–dependent and is accompanied by an alteration of global replication patterns and an increase in recombination. We conclude that human THO/TREX prevents the formation of R-loops that can compromise genome integrity. This work, therefore, provides experimental evidence for a role of mRNP biogenesis factors and R loops in genome integrity in humans.
| Transcription is a central cellular process occurring in the nucleus of eukaryotic cells in coordination with other nuclear processes. During transcription, the nascent pre-mRNA associates with mRNA-binding proteins and undergoes a series of processing steps, resulting in export-competent mRNA ribonucleoprotein complexes (mRNP) that are transported into the cytoplasm. The different steps of mRNP biogenesis are coupled to each other via an extensive network of physical and functional interactions [1], [2].
THO is a structural and functional unit identified first in budding yeast that is composed of four-protein (Hpr1, Tho2, Mft1, Thp2) and is associated with Tex1 and the mRNA export factors Sub2 and Yra1 forming a larger complex termed TREX [3], [4]. THO mutations lead to gene expression defects that are particularly evident for long and GC-rich DNA sequences [3], as well as for repeat-containing genes [5]. Such defects are the consequence of an impairment in transcription elongation as determined both in vivo and in vitro [3], [6], [7]. THO mutants show a hyper-recombination phenotype that is associated with transcription and is dependent on the nascent RNA molecule and on the co-transcriptional formation of RNA-DNA hybrids (R-loops) [8], [9]. In the current view, yeast THO would participate in the co-transcriptional formation of export-competent mRNP during transcription elongation by controlling the assembly of heterogeneous nuclear ribonucleoproteins (hnRNPs) onto the mRNA [10].
THO/TREX is conserved in all eukaryotes, and has been purified in Drosophila and human cells [4], [11], [12]. The human TREX (hTREX) complex is composed of the multimeric THO (hTHO) complex, containing hTHO2/THOC2, hHpr1/THOC1, fSAP79/THOC5, fSAP35/THOC6, fSAP24/THOC7 and hTex1/THOC3, the DEAD-box RNA helicase Sub2/UAP56 and the mRNA export adaptor protein Yra1/Aly/THOC4 [12]. Interestingly, it is associated with the spliceosome proteins and with spliced RNA, the latter interaction being independent of transcription, which raises the question of whether or not the involvement of THO/TREX in transcription is general from yeast to humans [12]. There is also evidence for transcription-dependent recruitment of THO to chromatin in both Drosophila and human cells [13], [14], but whether or not this is due to the known co-transcriptional function of the splicing machinery is still an open question. In this sense, hTREX has been shown to be recruited to the 5′ cap site of the mRNA via an interaction between ALY and the cap-binding complex CBC during splicing, ensuring mRNA export to the cytoplasm in a 5′ to 3′ direction [13]. ALY is a well-conserved RNA-binding protein that physically interacts with the conserved mRNA export Mex67/Tap/NXF1 allowing the mRNA-protein complex to be exported through the nuclear pore [15].
Despite the conservation of THO/TREX it is unclear whether the functional relevance is the same in all eukaryotes, which is important to know the degree of coupling between transcription and RNA export in higher eukaryotes. Thus, for example, in Drosophila the THO complex, is not essential for bulk poly(A)+ RNA export, whereas this is the case for UAP56 [16]–[19]. Whether human THO depletion impairs transcription elongation, mRNP biogenesis or RNA export or has genome-wide or transcript-specific effect is still an open question [11], [12], [19]–[22].
A distinctive phenotype of yeast THO mutants is their hyper-recombination phenotype associated with transcription, which is shared by other mRNP biogenesis/export factors from yeast to humans [23]–[26]. It has long been established that transcription enhances homologous recombination from bacteria to mammalian cells, a phenomenon termed TAR (transcription-associated recombination) [27]. However, whereas TAR in yeast THO mutants is dependent on the nascent mRNA molecule and is associated with R-loop formation, this has not been shown for human THO depletion.
In this work the effect of human THO depletion has been investigated on cell proliferation, transcription elongation and genome stability. Our study reveals that depletion of human THO subunits, in particular THOC1/hHPR1, reduces transcription elongation and RNA export, as determined by nuclear mRNA accumulation. hTHO depletion in different cell lines increases instability associated with the accumulation of DNA breaks, such instability being R-loop-dependent. Consistently, R-loop-dependent class-switching recombination is enhanced by THOC1 depletion in murine CH12 cells. Altogether, this work provides evidence for a functional role of THO in transcription and RNA-dependent genome instability, supporting a function of human THO/TREX in chromatin dynamics and function. These results indicate that the connection of transcription and mRNP biogenesis with genome instability is more functionally conserved from yeast to humans than previously anticipated.
To assay the implication of hTHO/TREX in gene expression, the effect of gene silencing of different THO/TREX components by RNA interference was investigated. HeLa cells were transfected with siRNAs against hHpr1/THOC1, THOC5, UAP56 and ALY and total RNA was analyzed by RT-qPCR. After transfection THOC1, THOC5, UAP56 and ALY mRNA levels were reduced to 33%, 26%, 51% and 19%, respectively, compared to the levels of the cells transfected with the siC control (Figure 1A).
Next we asked whether this depletion had a significant effect on gene expression of a reporter gene. For this purpose 72 h siRNA depleted HeLa cells were newly transfected with plasmid pmax containing a GFP cDNA. GFP gene expression was analyzed 24 h later by flow cytometry. Figure 1B shows the levels of GFP expression in the siRNA-transfected HeLa cells. The relative percentage of GFP positive cells was calculated for each siRNA transfected cell line. A reduction of approximately 40% of GFP expression was observed in cells transfected with siTHOC1 and siTHOC5, a similar percentage (about 50% of GFP) detected for cells treated with the transcription inhibitor α-amanitin, used as positive control in the experiments. The most drastic change, however, was observed in cells transfected with siUAP56 and siALY with a reduction close to 70% in the percentage of GFP positive cells. A reduction in the expression of a constitutive endogenous HPRT gene by RT-qPCR was also detected (Figure S1A). Altogether these results support that hTHO/hTREX depletion in human cells causes transcription defects.
To investigate whether the conserved human complex has a role in transcription elongation we used a tandem system (TAN1) to measure transcription elongation in cells depleted of THOC1 and other mRNP factors. This system consists of a single transcriptional unit covering two reporter genes, FLUC and hRLUC, under the control of the doxycycline inducible tet promoter [28] (Figure 1C). As the two reporter sequences are transcribed from a unique promoter, the ratio of expression of the downstream reporter versus the upstream reporter provides a measure of the relative rate of successful elongation through the intervening sequence. First, we evaluated the expression of the reporter FLUC in cells transfected with siTHOC1 versus the siC control and α-amanitin treated cells. A reduction of 50% of the reporter expression was observed in α-amanitin treated cells and a comparable but slightly lower reduction was detected with siTHOC1 (Figure 1C, upper panel). Interestingly, when the siTHOC1 transfected cells were treated with α-amanitin, a synergistic effect was observed (90% of reduction in FLUC expression). The ratio of hRLUC to FLUC activities in THOC1-depleted cells was reduced and in the presence of α-amanitin a synergistic effect was observed again (Figure 1C, lower panel), indicative of a role of THOC1 in transcription elongation. A strong reduction of FLUC activity was observed with siTHOC5, siUAP56 and siALY, consistent with a relevant role of these three subunits in transcription. However, due to these low FLUC values, it was not possible to obtain reliable hRLUC/FLUC ratios in these cases. To confirm that the effect observed in transcription elongation deduced from the hRLUC and FLUC activities, we determined the levels of transcripts containing each segment by qRT-PCR. The results clearly indicate that the hRLUC∶FLUC ratios of mRNA levels was significantly reduced in the cell lines depleted of the 4 subunits analyzed, THOC1, THOC5, UAP56 and ALY (Figure S1B), confirming a general role of THO/TREX in transcription elongation in human cells.
Since homologous recombination in mitotically dividing cells is the consequence of the repair of DNA breaks, the accumulation of DNA breaks was determined in THO/TREX-depleted cells by measuring the accumulation of γH2AX foci, one of the first components of the DNA damage response [29]. γH2AX in situ localization in HeLa cells transfected with siRNAs against hHpr1/THOC1, THOC5, UAP56 and ALY show clearly that transient depletion of these factors causes an accumulation of DNA damage, as deduced from the 2.6–10-fold increase in the number of cells containing γH2AX foci (Figure 2). Experiments were also performed in HeLa cell lines in which THO/TREX subunits were depleted by shRNAs. We first showed that shTHOC1, shUAP56 and shALY reduced the levels of THOC1, UAP56 and ALY mRNA to 30–40% of the levels of the cells transfected with the shTM blank control, as determined by RT-qPCR after 48 h of transfection (Figure S2A). In this case, in addition to γH2AX foci we analyzed the levels of the 53BP1 DNA-damage checkpoint protein [30]. The levels of γH2AX and 53BP1 foci were increased (Figure S2B and S2C), although to a lesser extent as in the siRNA-depleted cells, likely due to an earlier and more efficient protein depletion with siRNAs transfection.
Next, DNA damage was directly assessed by single-cell electrophoresis (Comet assay) by which, following DNA unwinding under alkaline conditions, broken DNA fragments (damaged DNA) migrate away from the nucleus (see Materials and Methods). First, we performed comet assay at different times in HeLa cells transfected with siTHOC1 and siTHOC5. As can be seen in Figure 3A, THO depleted cells show a significant increase in the tail moment. Similar results were obtained in siTHOC1 and siTHOC5-depleted MRC5 cells (Figure S3), a fibroblast cell line derived from normal lung tissue, indicating that THO depletion leads to an accumulation of DNA breaks in both normal and tumoral cell lines. Finally, to assay whether this accumulation occurred in cells depleted of other THO/TREX subunits, we performed the same experiments in HeLa cells transfected with siUAP56 and siALY siRNAs. As can be seen in Figure 3B, after 72 h of siRNA transfection, the cells showed a significant increase in DNA breaks as determined by the Comet assay.
Once demonstrated that THO/TREX depletion has an impact on gene expression, regardless of the subunit depleted, we decided to continue the analysis with the THOC1 conserved subunit as a representative THO subunit. For this reason, we first constructed stable HeLa cells lines for the depletion of THOC1. HeLa cells were stably transfected with an inducible shRNA for THOC1 (see Materials and Methods). Stable integration of the inducible THOC1 shRNA vector allowed the rapid production of siRNAs upon doxycycline induction. Among 4 stable clones obtained, HeTH-1 and HeTH-4 were chosen, showing about 50% reduction on THOC1 mRNA levels as determined by RT-qPCR (data not shown) and an efficient knock-down of the THOC1 protein as determined by Western analysis (Figure 4A). As expected from previous works [31], the growth rate of these stable cell lines was significantly reduced (Figure S4).
The impact of THOC1 depletion on transcription of endogenous genes was analyzed in HeTH-4 cells (+DOX) (+doxycycline) compared to control cells HeTH-4 cells (-DOX) (Figure 4B). These kinds of analyses have been previously used to study the role in transcription elongation of splicing factors [32]. RT-qPCR on DNase I-treated total RNA was performed using primer pairs covering different regions of three different-sized genes: PTBP1 (polypyrimidine tract binding protein 1), LIG3 (ligase III, DNA, ATP-dependent) and UTRN (utrophin). A reduction in the amount of mRNA at the 3′ proximal regions versus the 5′ ones were observed for the three endogenous genes analyzed upon doxycycline addition (Figure 4B). A similar reduction was observed when we added the transcription inhibitor α-amanitin (Figure S5). These results suggest that THOC1 depletion has a negative effect on transcription elongation in human cells.
We performed a global analysis of transcription to see whether the effect was general and to explore whether an effect on a specific transcription or mRNP biogenesis factor could indirectly explain the previous results. Comparison of the gene expression profiles between THOC1-depleted cells (HeTH-4 +DOX) cells with mock-treated controls (HeTH-4 –DOX) revealed that out of 28869 genes, 94 were down-regulated (32 well annotated genes and 62 non-coding RNAs (ncRNA)) and 140 up-regulated (36 well annotated genes and 104 ncRNAs), taking as a threshold set at 1.5-fold difference. (Table S2). Gene-GO term enrichment analysis does not show any relevant GO term associated with the list of gene deregulated. These data suggest that the effect of THOC1 depletion on transcription could be direct and not mediated by the altered expression of other genes.
Yeast THO mutants have a global poly(A)+ mRNA export defect [4], whereas in Drosophila THO is required for nuclear export of heat-shock mRNAs but it seems dispensable for nuclear export of total mRNA [11]. However, in human cells ambiguous data about the role of the THO complex in export of bulk poly(A)+RNA have been reported [20], [21]. To explore whether THOC1 is required for nuclear export of bulk poly(A)+ RNA in human cells in situ hybridization assays were performed with a fluorescently labeled oligo(dT) probe in HeTH-4 cells (+DOX) (Figure 4C). A series of optical sections through the entire cell was analyzed by confocal microscopy and the fluorescence signal in the cytoplasmic and nuclei was compared. The analysis revealed that whereas the non-induced shTHOC1 control cells showed uniform poly(A)+ distribution in the cell, similar to that of untransfected HeLa cells, poly(A)+RNA accumulated in a non-uniform manner in HeTH-4 cells (+DOX). The pattern was similar to that observed in HeLa cells transfected with a plasmid bearing a shRNA specific of ALY, used as positive control (Figure 4C). Accordingly, a significant reduction in the cytoplasmic-nuclear (C/N) ratio was observed with respect to that of HeLa control cells. Altogether these results suggest a general role of THOC1 in transcription and RNA export.
To further investigate how THO depletion induces DNA damage, we focused our efforts on THOC1-depletion using as a tool the stable HeTH-4 cell line expressing the inducible shTHOC1. First, we confirmed that the accumulation of 53BP1 foci after depletion of this factor also take place in the stable cell line. The strong reduction of THOC1 (+DOX) was accompanied by a 2-fold increase in 53BP1 foci with respect to the control (−DOX) (Figure 5A).
Next, DNA damage was assessed by single-cell electrophoresis (Comet assay) by which, following DNA unwinding under alkaline conditions, broken DNA fragments (damaged DNA) migrate away from the nucleus (see Materials and Methods). A two-fold increase in the tail moment in THOC1-depleted HeTH-4 cells (+DOX) demonstrates the accumulation of DNA breaks (Figure 5B).
Finally, to assay whether the increase in DNA breaks in THO/TREX depleted cells resulted in an increase in recombination, we designed and constructed a direct-repeat recombination construct, pIREC (Figure 6A). It consists of two GFP truncated repeats sharing 200-bp of homology and placed under the control of the doxycycline inducible tet promoter that were stably integrated into the genome (HeRG) (see Materials and Methods). Recombinants in this assay could be detected by FACS analyses as GFP positive cells. As can be seen in Figure 6 in both the HeRG stable cell line transfected with siRNA to deplete THOC1 (Figure 6B) and in HeTH4 cells (+DOX) transfected with the pIREC system (Figure 6C), the spontaneous recombination frequency increased with respect to their respective controls, consistent with an increase in DNA breaks and subsequent recombination events taking place upon THO depletion.
Next we wondered whether DNA breaks in THO-depleted human cells were a consequence of R-loop formation. THOC1-depleted HeTH-4 cells (+DOX) were transfected with vectors expressing RNaseH (RNH1 and/or RNH2), which degrades the RNA strand of DNA-RNA hybrids, and the 53BP1 foci formation was measured. Overexpression of RNH1, RNH2 or both reduced 53BP1 foci to values close to control levels, consistent with R-loop formation, although we can not rule out other sources of genome instability (Figure 7A). To further confirm this result, we tested whether human AID, a cytidine deaminase which works in single-stranded DNA as those formed in R-loops, increased DNA breaks in THOC1-depleted cells, an assay that has been used successfully in yeast [9]. As shown in Figure 7B AID expression in HeTH-4 cells (+DOX) increased the percentage of cells containing γH2AX foci 1.8-fold. Indeed, it is worth noting that after AID overexpression we detected PARP degradation in THOC1-depleted cells as determined by western (Figure 7C), suggesting that under these conditions a cell death program could be induced. According to these data an increase in the number of apoptotic cells, as measured by sub-G1 DNA content, was detected (Figure 7D).
Class switching is a natural phenomenon of recombination that is dependent on R-loop formation at the switch S regions of Immunoglobulin genes [33]. If THO-depletion facilitates R-loop formation in mammalian cells, we reasoned that in B cells AID would enhance its spectrum of action. To test this possibility class switching was assayed in murine CH12 cells derived from B cell lymphoma that were depleted of different subunits of murine THO by siRNA against different exons of THOC1. As can be seen in Figure 8B, there is a clear and consistent increase of class switching, in both unstimulated and stimulated cells, as determined by IgM to IgA conversion measured by FACS analyses (see Materials and Methods). The enhancement of the basal level of class switching detected in unstimulated cells can be explained by the direct action of AID on the S region, as it has been previously reported that AID induction augments class switching of unstimulated CH12 cells, which are known to express germline transcripts even without stimulation [34], [35]. In agreement with these data we detected AID and Iμ transcripts in unstimulated cells as determined by RT-PCR (Figure S6). The increase in class-switching in CH12 cells after THOC1 depletion support our hypothesis that THO-depletion could enhance the ability of mammalian cells to form recombinogenic R-loops.
One main function of recombination is to repair the DNA breaks that occur spontaneously as a consequence of DNA replication stalling or collapse. We asked whether the breaks and hyper-recombination of THOC1-depleted cells were accompanied by replication defects. Therefore, HeLa cells were transfected with siTHOC1 and siC and pulse-labeled with CldU (Chlorodeoxyuridine) to monitor replication by DNA combing. This analysis revealed that CldU tracks, which visualize newly replicated regions, are longer in siTHOC1 cells (54.5 kb) than in the siC control cells (34.0 kb) (Figure 9), suggesting that replication was 30% faster in THOC1-depleted cells. Similar frequency of replication initiation, as inferred from the distance between the centers of two CldU tracks, were observed in THOC1-depleted siTHOC1 (101.1 kb) and siC control cells (92.5 kb) (Figure 9). To measure replication elongation, cells were pulse-labeled with IdU and CIdU and the distance covered by individual forks during the pulse was determined. Results showed that replication forks travel at an apparent faster speed in THOC1-depleted siTHOC1 (2.3 kb/min) than in siC (1.6 kb/min) cells (Figure 9). Similar results were obtained with the stable cell line HeTH-4 (Figure S7). Instead, no clear differences were observed in the frequency of origin firing, as the inter origin distance in siTHOC1 was similar to that of control cells (Figure 9). It seems therefore clear that THOC1 depletion alters the progression of replication fork.
In this study we provide evidence that in human cells the THO/TREX complex has a role in mRNP biogenesis that connects transcription elongation, mRNA export and genetic instability. Reducing the expression of human THO/TREX components by RNA interference experiments results not only in a reduction of gene expression and mRNA export, but also in an impairment of transcription elongation. Moreover, we show that human THO depletion increases instability associated with DNA breaks, as determined by hyper-recombination and γH2AX and 53BP1 foci accumulation. Notably, such instability is dependent on R-loop formation, as determined by different in vivo approaches, and correlates with an alteration of global replication patterns as determined by DNA combing. Altogether these data suggest that human THO is a key player for mRNP formation and genome integrity that connects transcription elongation with genome dynamics and reveals that the connection of transcription and mRNP biogenesis with genome instability is more conserved than previously anticipated.
Our analyses of a tandem transcription reporter construct, and nascent mRNAs from different endogenous genes by RT-qPCR (Figure 1, Figure S1 and Figure 4) indicate that THO has a role in transcription elongation. The impact of THO depletion seems to be general and direct. The microarray analysis does not identify a significant reduction of expression of genes involved in mRNP biogenesis that could explain the results (Table S2). Different results have been reported for THO/TREX co-precipitation with the transcription apparatus [12], [31], [36], but our data suggests that THO/TREX have a functional role coupled to transcription elongation. Consistently, an early recruitment of THO to the 5′ end of mRNAs has been shown in a splicing- and cap-dependent manner [13]. This recruitment requires the cap-binding subunit CBP80, which interacts with the ALY/REF subunit of human TREX, and could explain that mRNA export takes place through the nuclear pore in a 5′ to 3′ direction.
RNA interference and biochemical studies in metazoans and genetic analyses in yeast indicate that the conserved THO/TREX complex functions in mRNA export [15], [20] (Figure 4C; [21]). However, in Drosophila the nuclear export of only a subset of mRNAs is affected by depletion of the THO subunits, which depends on the subunit depleted [11], [19]. In light of these results, the existence of various nuclear mRNA export pathways in multicellular eukaryotes has been suggested, which may be dictated by different adaptor RNA binding proteins. Consistently, it has been shown that THOC5, a subunit of the metazoan THO complex with no apparent orthologue in yeast, is not required for bulk mRNA export. However, it interacts with TAP-p15 and ALY, and functions in the export of specific mRNAs such as HSP70 [20].
The number of factors working at the interface transcription elongation and mRNA export reveals an increasing importance of the tight association between transcription and RNA biogenesis steps. Thus, Drosophila THO and ENY2/Sus1, a component of the histone-acetyltransferase complex SAGA/TFTC involved in transcription activation, interacts with the THSC/TREX-2 complex, required for mRNA export [14]. Also the human hnRNP CIP29 protein, the ortholog of yeast Tho1 hnRNP functionally related with THO, has been shown to be recruited to THO and to participate in mRNA export [37]. Spt6, a transcription elongation factor and histone H3 chaperone, binds to the Ser2P CTD of RNAPII and recruits Iws1 and the REF1/Aly mRNA export adaptor to facilitate mRNA export [38]. Iws1, which recruits the HYPB/Setd2 histone methyltransferase to the RNAPII elongation complex forms a megacomplex that affects mRNA export as well as the histone modification state of active genes in yeast [39]. Also noteworthy is the association of transcribed genes with the nuclear pore complex [40]–[43]. Our data, therefore, indicate that the human THO/TREX complex is another important factor in the coupling of transcription elongation with mRNA export.
One key feature of the yeast THO complex is its functional relevance in maintaining genome integrity, in particular by limiting the co-transcriptional formation of R loops. A similar R-loop-dependent co-transcriptional genome instability is observed in mammalian cells with loss of the splicing factor ASF/SF2 [25], [44], [45]. A recent genome-wide siRNA screening performed to identify genes involved in genome stability by monitoring phosphorylation of the histone variant H2AX suggests that a specific class of RNA processing factors may help prevent genome instability [26]. In a number of cases the accumulation of γH2AX foci are suppressed by RNase H overexpression, as would be expected if they were mediated by R-loops, whether completely or partially. The relevance of R-loops in the origin of chromosome instability has been studied at the S regions of the Immunoglobulin genes of B cells. In this case the R-loop provides the substrate for the action of the AID deaminase, which specifically acts on the ssDNA displaced by the DNA∶RNA hybrid [33]. Interestingly, an involvement of THO in genome instability had not been shown in humans. This is of key importance, as it would clarify whether human THO/TREX functions in vivo during transcription to prevent R-loop formation and whether its function would be related to the co-trancriptional formation of an mRNP. Our study clearly shows an increase in DNA damage, as determined as a larger percentage of cells with γH2AX and 53BP1 foci, in cells depleted of human THO/TREX (Figure 2 and Figure S2).
Accumulation of γH2AX foci of THO-depleted HeLa cells is suppressed by overexpresssion of RNAse H (Figure 7A) and enhanced by overexpression of the human cytidine deaminase AID (Figure 7B). These results are explained by the formation of DNA∶RNA hybrids, implying that the nascent mRNA could interact with the transcribed region behind the advancing RNAPII in the absence of human THO. This analysis demonstrates that indeed THO prevents R-loop formation in human cells (Figure 10). It is worth noting that a previously reported genome-wide analysis [26] failed to identify THO/TREX components among the affected RNAi-depleted cells leading to the accumulation of DNA breaks. THOC1 and THOC2-depleted cells appeared as showing a low proportion of cells (2–2.7%) with γH2AX foci. We believe that this could be due to the limitation of such a genome wide-analysis on cells depleted of essential factors that strongly affects their proliferation capacity, as is the case of THO-depleted human cells.
High levels of DNA breaks have been determined with the comet assay in cell lines depleted of different subunits of the THO/TREX complex (Figure 3A, 3B). The high accumulation of DNA breaks correlated with a hyper-recombination effect observed in the direct-repeat recombination construct pIREC after THOC1 depletion (Figure 6). Such a hyper-recombination phenotype was in any case lower than that of yeast THO mutants, which may be due to the fact that DSBs in mammals are more efficiently repaired via Non Homologous End Joining. Class switching, which is linked to transcription and R-loop formation at the S regions of the Ig genes [33], increased significantly in murine CH12 cells transfected with different siRNAs against THOC1 (Figure 8), consistent with a function of human THO preventing formation of R-loops and DSBs. Interestingly, THOC5 has recently been re-isolated in a screening of genes with a potential effect in CSR [46]. Altogether these data suggest that THO could play a role during normal B cell development, although further in vivo analysis would be needed to confirm this possibility. THO could contribute to the mRNP packaging of S regions. However, the structure and the G-richness of these S regions might make them difficult to assemble as an optimal ribonucleprotein even in the presence of THO. Consequently a basal level of R-loops could form at the S regions and promote the events necessary for normal development [9].
Finally, consistent with the idea that DNA breaks in THO-depleted human cells are linked to replication failures, we provide evidence in this study of an alteration in the pattern of replication in THOC1 depleted cells determined by DNA combing (Figure 9). This is consistent with the observation that transcription-associated recombination (TAR), as most forms of homologous recombination, is highly dependent on replication both in yeast and mammalian cells [47]–[49]. Highly transcribed genes are impediments for replication fork progression [50] and TAR may be linked to collisions between DNA replication and transcription machineries [51]. Thus, Topoisomerase I suppresses genome instability in mammalian cells by preventing conflicts between transcription and DNA replication [52]. Interestingly, however, our DNA combing analyses show that replicons seem longer. One possibility could be a putative incapacity of THOC1-depleted cells to trigger the S-phase or DNA damage response checkpoint and/or an incapacity to finish replication properly, leading to an apparent higher speed and longer replicons of THOC1-depleted cells. Indeed, yeast THO mutants activate the S-phase checkpoint and require an active S-phase checkpoint for viability under replicative stress [53]. Another plausible explanation of the longer replicons could be a reduction in the levels of transcribing RNAPII on the DNA, due to abortive transcription elongation. Faster replication forks have been detected for yeast sgs1Δ cells, which also show hyper-recombination [54]. Further molecular analyses of replication in THO-depleted human cells would be needed to understand the molecular basis for the DNA combing pattern observed.
In summary, our work shows that human THO controls transcription elongation at the interface with RNA processing and export, implying a physical connection with active chromatin. THO prevents the formation of R-loops that can compromise genome integrity by altering replication progression and leading to an accumulation of recombinogenic DNA breaks (Figure 9). This works, therefore, provides experimental evidence for a role of mRNP biogenesis factors in genome integrity in humans and reveals that the functional interconnection between mRNP biogenesis and the maintenance of genome integrity is more conserved than previously anticipated.
Commercial antibodies used were anti-ß actin, anti-THOC1 (Abcam), anti-γH2A (clone JBW301 Upstate), anti-53BP1 (NB100-304 Abyntec Biopharma), and mouse and rabbit polyclonal antibodies. For immunobloting, anti-mouse or anti-rabbit antibodies conjugated with horseradish peroxidase were used as secondary antibodies.
pSUPER-RETRO GFP was used to clone specific DNA sequences for shRNA with BglII and HindIII and as indicated by the manufacturer (OligoEngine VEC-PRT-0005/0006). The TAN 1 system and the method for the measurement of luciferase and renilla activities have been described [28]. pcDNA6/TR (Invitrogen) and pTER [55] were used to generate stable inducible shRNA cells. pcDNA3-RNaseH1 and pcDNA3-RNaseH2 were kindly provided by F. Baas [56]. For the plasmid pIREC, the mutated EGFP from the vector pI [57] was replaced for GFP repeats generated by PCR. pcDNA3 (Invitrogen) was used to clone the open reading frame of human AID.
All cell lines used in this study, except CH12, were maintained in DMEM (Gibco) supplemented with 10% heat-inactivated fetal bovine serum at 37°C (5% CO2). Transient transfection of plasmid (4 µg) or siRNA (100 nM) was performed using Lipofectamine 2000 (Invitrogen, Carlsbad, CA) according to the manufacturer's instructions.
HeLa stable cell lines with THOC1 shRNA were established by Lipofectamine 2000-mediated transfection of pTER-THOC1, a TetR-expressing construct, pCDNA6TR, followed by selection with 5 µg/ml blasticidin and 100 µg/ml of zeocin. The two positive clones selected were named HeTH-1 and HeTH-4. The shRNA target sequence is available upon request. CH12 cell line was maintained in RPMI 1640 supplemented with 10% FBS, 10 mM of 2-mercaptoetanol and 5% NCTC (Invitrogen).
HeRG stable cell lines were established by Lipofectamine 2000-mediated transfection of the pIREC plasmid in a HeLa stable cell line carrying pcDNA6TR (a TetR-expressing construct), followed by selection with 5 µg/ml blasticidin and 500 µg/ml of G418. The construction was verified treating the cells with different drugs as camptotecyn and neocarzinostatin.
Hypodiploid apoptotic cells were detected by flow cytometry according to published procedures [58]. Basically, cells were washed with phosphate-buffered saline (PBS), fixed in cold 70% ethanol, and then stained with propidium iodide while treating with RNase. Quantitative analyses of sub-G1 cells were carried out in a FACScan cytometer using the Cell Quest software (BD Biosciences).
cDNA was synthesized from cytoplasmic RNA (1 µg) by reverse transcription using Super-Script TM First strand synthesis for RT-PCR (Invitrogen) and random primers. RT-qPCR was performed with SYBR qPCR Mix (Applied Biosystems) and analyzed on an ABI Prism 7000 (Applied Biosystems, Carlsbad, CA). Primers sets for this analysis are described in Table S1.
Six independent microarray expression experiments were conducted. Affymetrix array experimental procedures were performed according to manufacturer's instructions at the CABIMER's Genomic Unit. Human Gene 1.0 ST array (Affymetrix, Santa Clara, USA) were used. The probe set signals were calculated using the Affymetrix GeneChip Operating Software 1.4.0.036. Linear fold-change cutoffs were analyzed at 95% confidence levels (p-values<0.05) in 1.5-fold down-regulated or up-regulated genes of THOC1-depleted cells. The microarray data were submitted to Gene expression Omnibus (GEO; accession number: GSE27091).
Cells cultured on glass coverslips were transfected with siRNA or plasmids (30% or 60–80% of confluency, respectively). After transfection, cells were cultured for 48 h, fixed in 2% formaldehyde in phosphate-buffered saline (PBS) and treated with Ethanol 70% for 5 min at −20°C, 5 min at 4°C, and washed twice in PBS. After blocking with 3% bovine serum albumin (BSA) in PBS, the coverslips were incubated with primary antibodies in 3% BSA in PBS followed by secondary antibodies conjugated withTexas Red goat anti-mouse or Alexa Fluor 568 goat anti-rabbit (Invitrogen). DNA was stained with DAPI.
RNA in situ hybridization was carried out as described [20]. Cells cultured on glass coverslips were either transfected with plasmid shRNA or treated with doxycycline (in the case of stable shRNA clones). Samples were processed after 48 h and 96 h respectively. Cells were fixed in 4% formaldehyde in phosphate-buffered saline (PBS). Quantitation of the nuclear-cytoplasmic distribution of poly(A)+ RNA was done using the Multiwavelength-MetaMorph v7.5.1.0. software. The cellular periphery was defined with phase contrast images and the nucleus with DAPI staining. The cytoplasmic: nuclear ratio of the mean fluorescence intensities was determined. All experimental analyses were performed with 7.5 104 cells for both HeLa or HeTH-4 cell lines
DNA DSBs were analyzed using a commercial comet assay (Trevigen, Inc.) following the manufacturer's protocol. For quantification, comet-positive cells were scored in random fields of cells. More than 100 cells from each sample were scored. The quantitative analysis was performed with the Comet-score software (version 1.5).
CH12 cells were transfected with pSUPERshRNA targeting THOC1 using Nucleofector (Amaxa). CH12 cells were stimulated for 12 h by adding 1 ng/ml of TGFB (R&D Systems), 5 ng/ml of IL4 (Bionova) and 0.5 µg/ml of anti-CD40 (Pharmigen). Surface IgA was stained with anti-mouse IgA-RPE antibody (AbDSerotec) and analyzed by flow cytometry 72 h after transfection. Double positive GFP-RPE cells were counted.
DNA combing was performed as described [59]. Briefly, DNA fibres were extracted from cells in agarose plugs immediately after CldU labeling and were stretched on silanized coverslips. DNA molecules were counterstained with an anti-ssDNA antibody (MAB3034, Chemicon; 1/500) and an anti-mouse IgG coupled to Alexa 546 (A11030, Molecular Probes, 1/50). CldU and IdU were detected with BU1/75 (AbCys, 1/20) and BD44 (Becton Dickinson, 1/20) anti-BrdU antibodies, respectively. DNA fibres were analysed on a Leica DM6000 microscope equipped with a DFC390 camera (Leica). Data acquisition was performed with LAS AF (Leica). Representative images of DNA fibers were assembled from different microscopic fields of view and were processed as described [60].
|
10.1371/journal.ppat.1005068 | C-di-GMP Regulates Motile to Sessile Transition by Modulating MshA Pili Biogenesis and Near-Surface Motility Behavior in Vibrio cholerae | In many bacteria, including Vibrio cholerae, cyclic dimeric guanosine monophosphate (c-di-GMP) controls the motile to biofilm life style switch. Yet, little is known about how this occurs. In this study, we report that changes in c-di-GMP concentration impact the biosynthesis of the MshA pili, resulting in altered motility and biofilm phenotypes in V. cholerae. Previously, we reported that cdgJ encodes a c-di-GMP phosphodiesterase and a ΔcdgJ mutant has reduced motility and enhanced biofilm formation. Here we show that loss of the genes required for the mannose-sensitive hemagglutinin (MshA) pilus biogenesis restores motility in the ΔcdgJ mutant. Mutations of the predicted ATPase proteins mshE or pilT, responsible for polymerizing and depolymerizing MshA pili, impair near surface motility behavior and initial surface attachment dynamics. A ΔcdgJ mutant has enhanced surface attachment, while the ΔcdgJmshA mutant phenocopies the high motility and low attachment phenotypes observed in a ΔmshA strain. Elevated concentrations of c-di-GMP enhance surface MshA pilus production. MshE, but not PilT binds c-di-GMP directly, establishing a mechanism for c-di-GMP signaling input in MshA pilus production. Collectively, our results suggest that the dynamic nature of the MshA pilus established by the assembly and disassembly of pilin subunits is essential for transition from the motile to sessile lifestyle and that c-di-GMP affects MshA pilus assembly and function through direct interactions with the MshE ATPase.
| The human pathogen Vibrio cholerae causes the debilitating disease cholera through ingestion of contaminated food and water. V. cholerae is a natural inhabitant of aquatic environments. Transmission of V. cholerae to the human host is dependent on survival of the pathogen in aquatic reservoirs where it is challenged with many stressors, including changes in the physiochemical parameters of environments and predation by protozoa and phages. One method utilized to endure these assaults is to form a multicellular community called a biofilm. The signaling molecule cyclic dimeric guanosine monophosphate (c-di-GMP) is utilized to induce biofilm formation in many bacteria, including V. cholerae. We demonstrate that c-di-GMP promotes the production of a cell surface structure called MshA pili by binding the molecular motor responsible for polymerizing pilus subunits. These pili are adhesive appendages that are essential for attachment to surfaces. This study identifies a novel mechanism for c-di-GMP regulation of pilus production through interactions with the molecular motor responsible for pilus assembly. Since many bacteria utilize pili for attachment to surfaces and c-di-GMP as a pro-biofilm signaling molecule, the mechanism for pilus regulation and biofilm formation described here may be widespread among many pathogens.
| Vibrio cholerae, the causative agent of the human intestinal disease cholera, is a natural inhabitant of aquatic ecosystems [1]. Cholera infection results from consumption of food and water contaminated with V. cholerae. Subsequently the bacteria turn on regulatory networks that facilitate bacterial growth and survival during the infection process [2]. They also activate production of virulence factors including the toxin-coregulated pilus (TCP), essential for intestinal colonization, and the cholera toxin (CT), responsible for production of massive watery diarrhea that results in dissemination of V. cholerae back to aquatic ecosystems [3]. V. cholerae’s ability to cause epidemics is tied to its dissemination and survival in aquatic habitats and its transmission to the human host. One critical factor for dissemination, environmental survival and transmission of the pathogen is its ability to form matrix enclosed surface-associated communities termed biofilms [4–6]. V. cholerae forms biofilms on the surfaces of phytoplankton and zooplankton [7], and exists in the surface waters of cholera endemic areas in matrix-enclosed aggregates thought to arise from biofilm-like populations of V. cholerae present in human stools [5]. Removal of particles >20 μm in diameter from water can reduce cholera incidence by 48% [8]. Additionally, growth in a biofilm induces a hyper-infectious phenotype [9]. Collectively, these studies highlight the importance of the biofilm growth mode in both the intestinal and aquatic phases of the V. cholerae life cycle.
Biofilm formation by V. cholerae begins with surface attachment, and subsequent development of microcolonies and mature biofilm structures [10–16]. The biofilm matrix is primarily composed of Vibrio exopolysaccharide (VPS) [17], extracellular DNA [18], and biofilm matrix proteins (RbmA, RbmC, and Bap1) [13,14,16,19], which are required for cell-cell and cell-surface interactions and development of mature V. cholerae biofilms [15].
Two cell-surface structures, a single polar flagellum and type IVa mannose-sensitive hemagglutanin pili (MshA), are critical for initial attachment and biofilm formation [11,20]. Type IVa pili have the ability to rapidly extend and retract, contributing to twitching and swarming motility in many bacteria [21]. Though V. cholerae produces the type IVa MshA pilus, twitching motility has not been reported. Genes required for biogenesis of MshA pilus are clustered into 16.7 kb region in V. cholerae chromosome-I and organized into two operons: the first operon harbors 9 genes from mshI-mshF predicted to encode proteins required for assembly and secretion and the second harboring 7 genes from mshB—mshQ encoding pilus structural components mshB-Q [22]. The pilus is comprised of repeats of the major pilin, MshA. Predicted function of the proteins encoded by MSHA gene locus are included in Table 1.
MshA pili and flagellum are also crucial for two distinct near-surface motility trajectories of V. cholerae: ‘roaming’ and ‘orbiting’, [10]. Low curvature ‘roaming’ trajectories meander over large distances and result from weak MshA-surface interactions. In contrast, orbiting trajectories repeatedly trace out tight circular tracks over the same region, and are the result of strong MshA-surface interactions. Cells that attach to the surface come from the orbiting subpopulation, while the roaming subpopulation pass over the surface without attaching. That orbiting motility is ablated when a mannose derivative is added to the medium to saturate MshA pilus binding further indicates that interactions between MshA pili with the surface are important [10].
The second messenger cyclic dimeric guanosine monophosphate (c-di-GMP) is an important promoter of the switch from motile planktonic growth mode to biofilm growth mode [23–25]. c-di-GMP synthesis is catalyzed by diguanylate cyclases (DGC) harboring a GGDEF domain while degradation is catalyzed by phosphodiesterases (PDE) harboring an EAL or HD-GYP domain. Subsequently, c-di-GMP is sensed by different classes of receptor proteins or RNAs and thereby converted to specific phenotypic outputs affecting motility, biofilm formation, and virulence. Elevated intracellular levels of c-di-GMP inhibit motility both by post-transcriptional and transcriptional mechanisms. In Salmonella enterica and Escherichia coli, the PilZ class of c-di-GMP receptor protein YcgR affects flagellar motor functions through interaction with FliG and FliM subunits of the flagellar rotor or the stator subunit MotA [26,27]. In Pseudomonas aeruginosa and V. cholerae, c-di-GMP inhibits motility by repressing transcription of flagellar genes through the AAA+ ATPase enhancer binding class of c-di-GMP receptor and transcriptional regulators FleQ [28] and FlrA [29], respectively. In addition to the regulation of flagellum production and activity by c-di-GMP, there are reports of c-di-GMP regulating the assembly and activity of Type IV pili. In Klebsiella pneumoniae, c-di-GMP is bound by PilZ class of c-di-GMP receptor protein MrkH, which upregulates the transcription of the fimbrial subunit mrkA [30–32]. In P. aeruginosa, the degenerate GGDEF-EAL domain class of c-di-GMP receptor FimX modulates Type IV pili production in an intracellular c-di-GMP concentration-dependent manner [33,34].
The V. cholerae genome encodes 31 GGDEF domain, 12 EAL domain and 10 dual GGDEF/EAL domain proteins [35]. Systematic analysis of in-frame deletion mutants of all V. cholerae genes encoding proteins with GGDEF and/or EAL domains for motility phenotypes revealed that four DGCs (CdgH, CdgK, CdgL, and CdgD) and two PDEs (CdgJ and RocS) affect motility in an LB soft agar motility assay [36]. Though deletion of the PDE cdgJ affected motility, no difference in intracellular c-di-GMP concentration was observed between the ΔcdgJ mutant and WT [36]. This measurement was conducted from a population, so there may be subcellular localized differences or population differences that affect motility. The molecular mechanism of c-di-GMP mediated motility repression and contribution of these DGCs and PDEs to switch from motile to surface-associated lifestyle remains elusive. In this study, we demonstrate that c-di-GMP inversely regulates motility and biofilm formation through direct regulation of the assembly and activity of the MshA pilus. Swimming motility is impaired in strains lacking the phosphodiesterase cdgJ, and disruption of the assembly or disassembly of the MshA pilus restores motility to WT levels by reducing the interactions with surfaces. Quantitative measurements indicate that c-di-GMP leads to increased production of MshA pili, which in turn bind surfaces and reduce motility. We demonstrate that the ATPase responsible for pilus polymerization, MshE, functions as a c-di-GMP receptor thereby providing an input for the c-di-GMP signal into the assembly of the MshA pilus. Collectively, this study elucidates how type IV pili and swimming motility are regulated by c-di-GMP in V. cholerae by presenting the first characterization of the complex involved in the assembly and disassembly of the MshA pilus and how c-di-GMP regulates the production and function of this complex.
CdgJ is a PDE and a cdgJ mutant displays a decrease in motility, enhanced VPS production, and increased biofilm formation compared to WT [36]. To begin investigating the mechanism by which CdgJ impacts motility, we performed transposon mutagenesis in a cdgJ deletion mutant (ΔcdgJ) and screened the resulting mutants for enhanced motility phenotype using LB soft agar motility assay. We screened 7054 transposon mutants and identified 42 extragenic suppressor mutants with increased motility and mapped the transposon insertions site to 22 different genes (Table 1, Fig 1). As previously reported, we found that mutations in the DGC encoding genes cdgH and cdgK in a ΔcdgJ strain enhance motility [36]. We also found that insertion in the gene encoding a pilus retraction motor, PilT, and insertions into different genes predicted to be required in mannose-sensitive hemagglutinin type IV pilus (MshA) biogenesis restored swimming motility.
To further investigate the suppression of a ΔcdgJ motility phenotype, we generated in-frame deletions of several of the genes identified in the transposon screen in wild-type and ΔcdgJ strains and analyzed the mutants for motility phenotype using LB soft agar motility assay. We focused on mshA (encoding major pilin subunit), mshE (encoding putative polymerizing ATPase), and pilT (encoding putative depolymerizing ATPase) as they are crucial for production and function of MshA pili. As previously reported, ΔcdgJ has a significant motility defect compared to the parental WT strain [36] (Fig 2A). Deletion of mshA, mshE, and pilT in a ΔcdgJ background restored motility similar to the WT strain, confirming that those mutations mediate suppression of the flagellar motility defect in the ΔcdgJ mutant. Deletion of pilU (predicted to encode a second copy of putative depolymerizing ATPase) in ΔcdgJ had no effect on the motility compared to the parental ΔcdgJ mutant. Mutants of mshA, mshE, and pilT in a WT background were assayed for motility to determine if their enhanced motility in the ΔcdgJ background is dependent on this mutation, or if this could be a case of bypass suppression. The ΔmshA, ΔmshE, and ΔpilT mutants exhibited enhanced motility compared to the WT strain (Fig 2A). The ΔpilU strain had a similar motility phenotype to the WT strain, suggesting that the functions or expression profiles of pilT and pilU are different. These data demonstrate that deletion of mshA, mshE, and pilT enhances motility regardless of the presence of a wild-type copy of cdgJ. However, we could not rule out the possibility that cdgJ could directly or indirectly control the production or function of the MshA pilus.
To determine if mshA-mediated suppression of the flagellar motility phenotype is specific to the cdgJ mutation or if it occurs in other PDE deletion backgrounds, we mutated mshA in a ΔrocS strain. RocS has both GGDEF and EAL domains and is predicted to function mainly as a PDE as ΔrocS mutants have reduced motility along with enhanced VPS production and biofilm formation [36–38]. We generated a ΔrocSmshA double mutant and determined that this mutant exhibited a wild-type motility phenotype (S1 Fig). These findings suggest that MshA negatively impacts V. cholerae flagellar motility and is involved in general c-di-GMP mediated repression of motility.
To evaluate further the ability of MshA to repress motility, we analyzed the effect of the expression of a WT copy of mshA provided in trans in an expression plasmid with an IPTG-inducible promoter. Motility assays confirmed that expression of mshA, upon induction with IPTG, significantly reduced motility in a ΔmshA strain (Fig 2B). Additionally, expression of mshA in the WT strain reduced motility, suggesting that overproduction of MshA impairs motility. IPTG had no effect on motility in strains harboring an empty vector control.
Since the MshA pilus is critical for initial stages of surface attachment and subsequent biofilm formation [10,11,20], we hypothesized that the ΔmshE and ΔpilT mutants would phenocopy the reduced biofilm phenotype of a ΔmshA mutant. Biofilms of these mutants were grown using a flow cell system, imaged using confocal microscopy, and analyzed using the COMSTAT image analysis software package to evaluate biofilm structural properties. As expected, the ΔmshA mutant attached poorly to the substrate and formed biofilms with low biomass (Fig 3A and 3B). The ΔmshE and ΔpilT mutants grew biofilms with significantly less biomass, thickness, and substrate coverage than WT. COMSTAT analysis revealed that while biofilm biomass of ΔpilT and ΔmshA was similar, the biomass of ΔmshE was significantly less than the ΔpilT strain (Fig 3B). Additionally, surface coverage of ΔpilT mutant was greater than that of ΔmshA and ΔmshE. These differences suggest that though the ΔmshA, ΔmshE, and ΔpilT mutants are deficient at forming biofilms, there are subtle differences in the phenotypes of the strains. The ΔpilU mutant produced biofilms that were indistinguishable from WT.
Since our initial interest in the MshA pilus was sparked by the discovery that mutations in pilus genes can suppress the motility defect in mutants of the PDE cdgJ, we determined whether the MshA pilus mutations were epistatic to the cdgJ mutation. As previously reported, the ΔcdgJ strain produced thicker biofilms than the WT strain (Fig 3C and 3D) [36]. The ΔcdgJmshA, ΔcdgJmshE, and ΔcdgJpilT strains formed biofilms with significantly less biomass, thickness, and surface coverage than the ΔcdgJ strain. These findings are consistent with the hypothesis that these proteins are involved in formation of the MshA pilus and that a functional MshA pilus is required for biofilm formation, surface attachment, and the inhibition of flagellar motility. Biofilms formed by the ΔcdgJmshE mutant had a significant reduction in biomass, thickness, and surface coverage compared to the ΔcdgJmshA or ΔcdgJpilT strains.
In addition to using bulk differences in biofilm formation as an index of the transition between motile and sessile behavior, we also directly monitored surface attachment of V. cholerae with single cell resolution using high-speed microscopy and cell tracking (at 5 ms frame rate) to elucidate how genes involved in MshA pilus production impact microscopic outcomes such as initial surface attachment. ΔmshA mutants do not attach to the surface in significant numbers, with no attached cells observed in the first 15min after inoculation. In contrast, well over 100 cells of the WT strain attach to the surface over the same time interval (Fig 4A). Using the same metrics, the ΔmshE mutant was also unable to attach to the surface, exhibiting binding profiles that were similar to the ΔmshA strain. The ΔpilT strain was able to attach to the surface more than the ΔmshA; however, it is unable to attach with the same efficiency as the WT strain. This suggests the following hierarchy of behavioral categories for comparison with non-WT backgrounds: the strong binding strain (WT), the intermediate binding strain (ΔpilT), and the weak binding strains (ΔmshA and ΔmshE). To dissect the origin of this aggregate statistical behavior of surface attachment, we needed more single-cell metrics for bacterial behavior near a surface. We examined the temporal aspects of single bacterium interactions with the surface in the form of single cell residence time (number of seconds that each stationary cell remained associated with the surface, Fig 4B). The WT strain formed prolonged associations with the surface, with a mean residence time of 2.6 seconds and a maximum of 78.9 seconds. The ΔmshA, ΔmshE, and ΔpilT mutant strains demonstrated more transient interactions with the surface, with reduced mean residence times compared to WT (0.69, 0.58, and 0.58 seconds, respectively). This is most evident in the inset of Fig 4B, where the entire adherent populations of these mutants have residence times of less than 6 sec, while the WT residence times extend out to nearly 80 seconds.
We used high-speed microscopy and near-surface cell tracking to record the trajectories of cells within one micrometer of the coverslip surface in a microscopy chamber (Fig 5). All strains with flagella are capable of swimming motility in 3D, and can exhibit trajectories that come in and out of focus. Consistent with previous reports, the WT strain exhibits orbiting and roaming behavior with regards to near surface motility [10]. All of the surface attached cells come from the orbiting subpopulation. The ΔmshA mutant does not show orbiting or roaming behavior, consistent with the model that MshA pili-surface interactions are responsible for these near-surface motility phenotypes. Moreover, also consistent with the model, the mutant shows greatly reduced surface attachment (Fig 4A) [10]. The ΔmshE strain tracks phenocopy ΔmshA, exhibiting predominantly a swimming phenotype with little observable attachment. The ΔpilT mutant has an intermediate phenotype; a few cells appear to exhibit behavior similar to WT orbiting, but has greatly reduced attachment compared to WT. These data are consistent with the biofilm and attachment data described in Figs 2 and 3.
To investigate the role of the PDE CdgJ on initial surface attachment, we observed mshA, mshE, pilT mutants in a ΔcdgJ background with high-speed microscopy using similar experiments. We determined that the ΔcdgJ mutant exhibits strong attachment to the surface, with a sharp increase in the number of attached cells as a function of time during the initial few minutes compared to WT. (Fig 4C, time = 1). This strain rapidly associated with the surface, with nearly all cells binding within the first two minutes of observation. As predicted from the biofilm and motility data (Figs 1 and 2), the ΔcdgJmshA, ΔcdgJmshE, and ΔcdgJpilT strains exhibited reduced surface attachment compared to the parental ΔcdgJ strain and WT (Fig 4C). As in the WT background, the ΔcdgJpilT strain exhibited an intermediate level of attachment and the ΔcdgJmshA and ΔcdgJmshE strains were poor at attachment. The ΔcdgJmshA, ΔcdgJmshE, and ΔcdgJpilT strains exhibited short residence times with mean values of 0.43, 0.53, and 0.57 s, respectively (Fig 4D). By contrast, the ΔcdgJ strain exhibits much longer mean surface residence times (3.59 s with and a maximum of 75.8 seconds), which surpass even WT (Fig 4D, red vs black bars). These longer residence times indicate a strong tendency of the ΔcdgJ strain to associate with surfaces, which is also evident in the rapid decrease in density of tracks of the ΔcdgJ over time (Fig 5): since only motile cells are displayed in each image, the density of tracks diminishes over time in the ΔcdgJ strain due to the increasing proportion of adherent, nonmotile cells. These results demonstrate the ΔcdgJ strain adheres to surfaces more rapidly than WT, and that mshA is epistatic to cdgJ.
The motility and biofilm phenotypes of the ΔmshE and ΔpilT mutants, combined with homology to known Type IV pilus motor proteins suggests that these genes are involved in the production of a functional MshA pilus. MshE shares 75% amino acid similarity with the Type IV extension ATPase PilB of P. aeruginosa and 77% similarity to the Type II extension ATPase EpsE of V. cholerae, suggesting that MshE is the extension ATPase of the MshA pilus (S2 and S3 Figs). We investigated the role of these genes in the production of MshA pili using a surface MshA pilin ELISA (Fig 6A), which detects only assembled pili. ΔmshA, ΔmshE, and ΔpilT mutants produced significantly less surface MshA pili than the WT strain. This supports the hypothesis that MshE and PilT are involved in the production of a functional MshA pilus. The ΔpilT strain produced significantly more surface pili than ΔmshA or ΔmshE, correlating with the intermediate phenotype of this strain observed in biofilm and near surface motility assays (Fig 3A and 3A). Surface MshA pilus production was determined in deletion mutants of other genes in the secretory operon with multiple transposon insertions (S4 Fig). Deletion of mshL, which encodes the putative outer membrane pore protein, resulted in no surface pili. While deletion of mshM (predicted to encode an ATPase) or mshN (predicted to encode a tetratricopeptide repeat domain) resulted in similar pilus production to the WT strain, pilus production was increased in a ΔmshI mutant, The specific mechanisms by which lack of these genes results in suppression of a motility defect in a ΔcdgJ strain is yet to be determined. The ΔcdgJ mutant produced significantly more surface MshA pili than WT. Surface MshA pili were reduced in ΔcdgJmshA, ΔcdgJmshE, and ΔcdgJpilT mutants to the level of the ΔmshA strain, indicating that none of these strains could produce a functional pilus. Whole cell western analysis indicated that all of the strains except ΔmshA and ΔcdgJmshA produced similar amounts of MshA protein, suggesting that the altered surface MshA is not a result of altered MshA production (Fig 6B and 6C). These data support the hypothesis that a functional MshA pilus inhibits motility and enhances surface attachment and biofilm formation. This effect is likely due to the interactions between the pili and surfaces, however, there may be additional mechanisms affecting motility and surface attachment, as discussed below. Additionally, the elevated production of MshA pili by the ΔcdgJ mutant could explain the motility and biofilm phenotypes observed in this mutant.
To confirm further that mshE was responsible for the lack of pili in a ΔmshE mutant, we generated chromosomal replacements at the native locus with either the WT mshE or a sequence encoding a mutation in the Walker A ATPase active site (K329A) (S5 Fig). Pilus production was restored to WT levels when the WT sequence was inserted, however the K329A produced no detectable surface pili. These data confirm that MshE, and specifically an intact ATPase domain, are required for MshA pilus production.
We also utilized transmission electron microscopy (TEM) to assess presence of MshA on the cell surface (Fig 7). We determined that WT produces several pili along the cell body (range 2–5) that were about one half to one cell body length. These pili were not present in the mshA mutant, suggesting that the pili observed are in fact MshA pili. Similarly, the ΔmshE strain produced no visible pili, which is consistent with the prediction that MshE is the motor protein responsible for extension of the MshA pilus. These images reveal that there were no differences observed between pili produced by WT and the ΔpilT, ΔpilU and ΔcdgJ mutants. Due to the fragile nature of MshA pili, quantitative measurements were not possible with TEM though these images observing the presence or absence of MshA pili support the quantitative measurements observed by ELISA.
The pilus motor proteins MshE and PilT contain Walker A ATPase domains, which are utilized to energize the assembly and disassembly of pili. Baraquet et al. demonstrated that the activity of the P. aeruginosa regulator FleQ is modulated by binding c-di-GMP at its Walker A site [39]. We hypothesized that one or both of the Msh pilus motor proteins could function as a c-di-GMP receptor. Isothermal calorimetry was utilized to investigate the interaction of these proteins with c-di-GMP. We determined that MshE binds c-di-GMP, while PilT and PilU were unable to bind c-di-GMP (Fig 8A). ATPase activity of purified MshE, PilT, and PilU confirmed that these preparations contain functional protein (S6 Fig). VpsT was purified and included as a positive control, as it has been demonstrated to bind c-di-GMP [40]. Fitting the data to a single binding site model indicates that MshE has a slightly lower affinity for c-di-GMP (K = 1.14x105 ± 3.24x104 M-1) than VpsT (K = 9.9x104 ± 2.06x104 M-1). These data suggest that c-di-GMP does affect MshA pilus production through interactions with MshE.
Since MshE, PilT, and PilU contain conserved Walker A ATPase domains, but only MshE bound c-di-GMP, we purified the N terminal domain of MshE (amino acids 1–180) for analysis of the interactions with c-di-GMP. This domain lacks homology with either PilT or PilU and therefore is a likely candidate for binding c-di-GMP. Fluorescence thermal shift assays were utilized to determine the interaction between MshE or MshE-N terminal domain and nucleotide. Purified proteins are stabilized by a bound ligand; therefore the temperature midpoint of unfolding (Tm) of the protein in the presence of ligand is higher than the Tm of the protein in buffer [41,42]. We observed that the Tm of full-length MshE in buffer was 38.43 ± 0.52°C (Fig 8B). When full-length MshE was incubated with ATP or c-di-GMP, the Tm increased to 40.15 ±1.02°C and 42.77±0.70°C, respectively. This indicates that full-length MshE binds both ATP and c-di-GMP. The N-terminal domain of MshE had a Tm of 66.36 ± 0.79°C in buffer. There was no increase in Tm in the presence of ATP (Tm = 62.72 ±0.46°C). In contrast, incubation of the MshE N-terminal domain with c-di-GMP increased the Tm to 70.18 ±0.55°C, indicating that this domain binds c-di-GMP (Fig 8C). Additionally, these interactions are specific, as neither the full-length nor the N-terminus of MshE bind to cAMP, c-di-AMP, GTP, or cGMP in thermal shift assays (Fig 8B and 8C).
To further evaluate the link between c-di-GMP production and pilus assembly, we performed a surface pilin ELISA to determine the production of MshA pili over a range of c-di-GMP concentrations (Fig 9). To generate a range of c-di-GMP concentrations, we introduced a construct harboring an IPTG-inducible copy of the DGC VCA0956 (Ptac0956) into the chromosome of the V. cholerae O1 El Tor strain A1552. This strain was grown in the presence of varying concentrations of IPTG, followed by detection of extracellular MshA with the surface pilin ELISA. We observed that upon induction of the expression of VCA0956 with IPTG, there was an increase in intracellular c-di-GMP (Fig 9, Grey bars). Additionally, the surface pilin ELISA detected increased amounts of extracellular MshA that coincided with the increased c-di-GMP (Fig 9, Black bars), though the total MshA production was not increased (S7B Fig) suggesting that the increased surface pilin is due to enhanced assembly. There was a significant correlation between elevated c-di-GMP and increased surface MshA (Pearson correlation, p = 0.0025, R2 = 0.9199). Increased c-di-GMP resulted in reduced motility and increased biofilm maximum thickness compared to WT (S7 Fig). Collectively, these data indicate that c-di-GMP promotes assembly of the MshA pilus. Induction of VCA0956 in a ΔmshE strain does not increase the production of surface MshA pili, further supporting that MshE is required for the c-di-GMP-dependent increase of MshA pili (S8 Fig).
High c-di-GMP levels inhibit motility of bacteria and studies have highlighted some of the mechanisms involved. c-di-GMP can repress transcription of flagellar genes, or can act post-transcriptionally to regulate flagellar reversals by interactions with particular flagellar motor proteins or by altering the chemotaxis signal transduction system [26–29,43–45]. Our work revealed a role for c-di-GMP in the regulation of MshA pilus production with effects on near surface motility, motile to sessile transition, and biofilm formation via a post-translational mechanism. c-di-GMP also promotes the assembly and activity of Type IV pili in P. aeruginosa [34,46]. In this example, ectopic expression of a DGC results in elevated amounts of c-di-GMP and increased surface pili. We demonstrate that production of the MshA pilus in V. cholerae is increased in response to high concentrations of c-di-GMP. These similar results in two organisms suggest that c-di-GMP may promote type IV pili assembly and activity as a more general mechanism of pilus regulation than previously identified. Many bacteria rely on type IV pili for motility and attachment, so integration of c-di-GMP in post-translational control of these structures could be a conserved mechanism across species.
The V. cholerae genome encodes 31 GGDEF domain proteins but we have found that only a subset of these impact motility, biofilm formation, or both. We previously reported that four DGCs CdgH, CdgK, CdgL, and CdgD and two PDEs CdgJ and RocS affect motility. Furthermore, a strain lacking all four DGC-encoding genes (ΔcdgDΔcdgHΔcdgKΔcdgL) has a markedly high motility phenotype, suggesting the effect of these proteins on motility is additive [36]. Screening for suppressor mutants of ΔcdgJ that restored swimming motility identified CdgH and CdgK, suggesting that c-di-GMP produced by these DGCs could be a substrate for CdgJ. To test if CdgJ and DGCs that control motility physically interact, we analyzed interactions of CdgJ with CdgH, CdgK, CdgL, and CdgD using the commercially available Bacterial Adenylate Cyclase Two-Hybrid System (BACTH) (Euromedex Strasbourg, France). We did not observe any interaction between DGCs and CdgJ or between CdgJ and MshE or PilT, suggesting that these proteins do not need to be in physical contact, or the interaction is too weak or brief to be detected with this method.
Pili are dynamic structures that are generated by the assembly and disassembly of pilin subunits. The motor proteins responsible for this process have been characterized in many other systems. The two putative motor proteins of interest in this report, MshE and PilT, were identified based on homology to the PilT motor protein in P. aeruginosa. These proteins belong to AAA+ ATPase family proteins and are responsible for energizing the addition and disassembly of pilin subunits. Previous studies revealed presence of a bacterial AAA+ ATPase enhancer binding class of c-di-GMP receptors [28]. Here, we demonstrate that MshE binds to c-di-GMP. This is an important finding, as it establishes that ATPases beyond enhancer binding proteins are also capable of binding c-di-GMP. We note that PilT and PilU, which both harbor AAA+ ATPase domain, are unable to bind to c-di-GMP under the conditions tested. We also showed that N-terminal domain of MshE, which is not present in PilT and PilU is capable of binding to c-di-GMP. Future studies will elucidate the effect of c-di-GMP binding by MshE, as well as the specific mechanisms of this interaction.
We have also characterized several proteins necessary for the production of a functional MshA pilus. Although the function of members of the MshA operons were predicted based on homology to proteins of known function in other bacteria, function of these genes in MshA biogenesis were not characterized [22]. We have demonstrated that MshE is responsible for the assembly of MshA pilin subunits into a functional pilus. These data indicate that the decreased attachment and biofilm phenotypes, as well as the enhanced motility of the mshA, mshE and pilT mutants relies on the dynamic nature of the MshA pilus. If the presence of a pilus could enhance attachment and biofilm formation, a ΔpilT strain would in principle produce biofilms that had similar, or even greater biomass than the WT strain. The observation that the ΔpilT strain phenocopies ΔmshA and ΔmshE, which lack extracellular pili, suggests that both extension and retraction of the pili are critical for normal substrate attachment. The ELISA indicates that the ΔpilT strain produces fewer pili than WT, though the TEM images confirm that there are pili on the surface. Future studies will further investigate the role of PilT in the retraction of the MshA pilus. It is important to note that PilT has already been characterized as the retraction pilus of the ChiRP pilus (chitin-induced competence), suggesting that PilT can function in more than one type IV pilus system [47]. This could be a mechanism for genomic conservation, where one promiscuous retraction ATPase is encoded, while several extension ATPases allow for specificity of the system in regulation. Future studies will address the mechanism of co-regulation of these systems to determine whether pilT is expressed constitutively while the specific extension ATPases are regulated, or if there is overlap in the regulation between the extension and retraction ATPases.
Several studies have investigated how bacteria sense and respond to surfaces, often times by rapidly upregulating production of adhesins and polysaccharides [48–50]. This regulation is typically mediated by c-di-GMP [23,28,40,50,51]. Many bacteria also regulate motility in response to surfaces. E. coli uses the resistance to flagellar rotation as a mechanosensor and adapt by adding force-generating motor subunits to the stator complex [43]. This allows the bacterium to adjust the force of flagellar rotation to match the viscosity of the environment. An additional example of “stator swapping” to modulate flagellar force was recently published for P. aeruginosa [44]. In this example, c-di-GMP represses motility by excluding the swarming proficient MotC/D proteins from the stator complex in favor of the swarming deficient MotA/B proteins. P. aeruginosa also utilizes the altered chemotaxis protein WspA as a surface sensor, which results in production of c-di-GMP by the DGC WspR upon interaction with a surface [52,53]. In B. subtilis, production of flagella, and therefore swarming motility, is regulated by Lon-dependent proteolysis of the master regulator of flagellar biosynthesis SwrA upon surface contact [54]; and flagellar function is modulated by EpsE which synergizes exopolysaccharide biosynthesis with flagellar motility by acting as a clutch through interaction with the flagellar protein FliG to limit rotation and therefore motility [55].
This study presents a possible mechanism for how c-di-GMP production affects motility and biofilm formation through modulating MshA pilus production. Both pili and flagella contribute to near surface motility and initial attachment [10]. Besides the generation of near-surface motility modes conducive to surface attachment, it is known that van der Waals forces depend crucially on the extent of surface contact [56]. That V. cholerae has a comma-like helicoid shape with smaller surface contact areas implies that adhesive forces between the surface and the cell body will be decreased relative to more cylindrically symmetrical species such as P. aeruginosa for most cell orientations, so adhesive contributions from appendages like MSHA to ‘anchor’ the cell on surface will be comparatively more important. In fact, recent measurements of TFP adhesive forces show that they can be quite strong, in the hundreds of pico-Newton (pN) range, and are surface chemistry dependent, in agreement with our results [57]. That V. cholera select for surfaces that interact with MSHA strongly (and thereby generate ‘orbiting’ motility) implies that the existence of more functional MSHA induced by c-di-GMP can better anchor a cell mechanically and mitigate against flagellum driven motion. This work further strengthens the notion that there is a mechanistic link between c-di-GMP and initial attachment through modulation of flagellar motility and pilus activity.
The bacterial strains and plasmids used in this study are listed in S1 Table. All V. cholerae and Escherichia coli strains were grown aerobically, at 30°C and 37°C, respectively, unless otherwise noted. All cultures were grown in Luria-Bertani (LB) broth (1% Tryptone, 0.5% Yeast Extract, 1% NaCl), pH 7.5. LB agar medium contains 1.5% (wt/vol) granulated agar (BD, Sparks, MD). Concentrations of antibiotics and inducers used, where appropriate, were as follows: ampicillin, 100 μg/ml; rifampicin, 100 μg/ml; gentamicin, 50 μg/ml, kanamycin, 50 μg/ml, and arabinose, 0.2% (w/v), 6.25–400μM IPTG. In-frame deletion and GFP-tagged strains were generated according to protocols previously published [13,14].
DNA manipulations were carried out by standard molecular techniques according to manufacturer’s instructions. Restriction and DNA modification enzymes were purchased from New England Biolabs (Ipswitch, MA). Polymerase chain reactions (PCR) were carried out using primers purchased from Bioneer Corporation (Alameda, CA) and the Phusion High-Fidelity PCR kit (New England Biolabs, Ipswitch, MA), unless otherwise noted. Sequences of the primers used in the present study are available upon request. Sequences of constructs were verified by DNA sequencing (UC Berkeley DNA Sequencing Facility, Berkeley, CA).
A region encompassing the PlacIq-lacI and the Ptac promoter elements was amplified from the pMAL-c5x plasmid (New England Biolabs, Ipswitch, MA) by PCR. The amplified product was joined by overlapping PCR to amplicons of ~500 bp that correspond to sequences upstream and downstream of the VCA0956 translational start site. The resulting amplicon was cloned into the suicide plasmid pGP704sacB28 and mobilized into Vibrio cholerae A1552 by biparental mating. The selection of double recombinants with the desired insertion of the PlacIq-lacI and Ptac promoter elements was performed as described in [14]. Sequences of constructs were verified by DNA sequencing (UC Berkeley DNA Sequencing Facility, Berkeley, CA).
To generate a library of transposon mutants, V. cholerae O1 El Tor strain A1552 ΔcdgJ was conjugated with the donor E. coli S-17-l λpir containing the Mariner transposon on the pSC189 backbone [58]. Transconjugants were selected on LB agar containing kanamycin 50μg/ml and rifampicin 100μg/ml. A total of 7054 mutants were isolated and screened for motility phenotypes on LB soft agar (0.3%) motility plates.
Motility plates consist of LB containing 0.3% agar supplemented with 100μM IPTG where appropriate. Plates were poured and allowed to dry at room temperature for 4 h prior to inoculation. Colonies from overnight LB agar plates grown at 30°C were transferred to motility plates and incubated for 16 h at 30°C. Motility diameter was measured and normalized to the average of WT on each plate. Experiments were performed with three biological replicates in triplicate and data were analyzed with a Oneway ANOVA followed by Dunnett’s multiple comparison test.
Inoculation of flow cells was done by diluting overnight-grown cultures to an OD600 of 0.04 and injecting into a μ-Slide VI0.4 (Ibidi, Martinsried, Germany). To inoculate the flow cell surface, bacteria were allowed to adhere at room temperature for 1 h. Flow of 2% v/v LB (0.02% tryptone, 0.01% yeast extract, 1% NaCl; pH 7.5) was initiated at a rate of 7.5 ml/h and continued for 24 h. Confocal images were obtained on a Zeiss LSM 5 PASCAL Laser Scanning Confocal microscope. Images were obtained with a 40X dry objective and were processed using Imaris (Bitplane, Zurich, Switzerland). Quantitative analyses were performed using the COMSTAT software package [59]. Statistical significance was determined using Oneway ANOVA with Dunnett’s Multiple Comparison test. Three biological replicates were performed in triplicate. Images presented are from one representative experiment.
Bacteria were cultured in full strength Luria–Bertani (LB) broth overnight under shaking at 30°C. Immediately prior to inoculation, cultures were diluted into 2% LB (containing 171 mM NaCl) to an OD600 0.01–0.03. The V. cholerae cells were then injected into a sterile flow-cell containing the same media and imaged immediately.
Imaging was performed with a Phantom V12.1 high-speed camera (Vision Research) collecting ~20,000 bright-field images at 5 ms resolution with a 100x oil objective on an IX71 Olympus microscope. All movies were recorded at the same frame rate, for the same duration, after the first, third and 15th minute post inoculation.
For cell-tracking algorithms and analysis protocol, every frame of a movie was preprocessed in Matlab (Mathworks) by subtracting the background, scaling, smoothing and thresholding. Image processing this way causes the bacteria appear as bright regions. Tracking is done by locating all bright objects that overlap objects in the next frame by combining the two frames into a three-dimensional (3D) matrix and then by locating 3D connected components.
Results are stored in a tree-like data structure with multiple roots; every newly detected bacterium that appears is recorded as a ‘root’ of the tree. When bacteria interact, they are recorded as a ‘node’ of the tree; when they depart, they are recorded as a ‘leaf’. Each root or node stores the sequence of pixel lists that comprise the bacterium in all frames until the next interaction or detachment event. We measure the instantaneous shape properties of the bacteria using the Matlab regionprops function [10].
Surface pili composed of MshA were quantified using an ELISA based on a previously published protocol [46]. Briefly, overnight culture was diluted 1:100 in fresh LB medium and grown to OD600 0.5 at 30°C. Cells (125μL) were added to a 96-well plate (Greiner Bio-One, Monroe, NC) and incubated at 30°C for one hour. Cells were fixed with 100μL of methanol for 10 minutes at room temperature, then washed twice with PBS. Samples were blocked in 5% nonfat dry milk and immunoblotted with polyclonal rabbit anti-MshA (1:1000 dilution, gift of J. Zhu) and horseradish peroxidase (HRP)-conjugated secondary antibody (Santa Cruz Biotechnology, Santa Cruz, CA). After three washes in PBS, 100μL of TMB (eBioscience, San Diego, CA) was added and incubated for 30 minutes at room temperature followed by the addition of 100μL of 2N H2SO4. Absorbance was recorded at 490nm and the samples were normalized to the change in WT. Three biological replicates were assayed in triplicate and statistical significance was determined with a Oneway ANOVA followed by a Dunnett’s Multiple Comparison test.
Samples were grown to mid-exponential phase (OD6000.5) in LB or LB with IPTG. Cells were collected via centrifugation and cell pellets were resuspended in 2% SDS and boiled for 5 minutes. Lysates were cleared via centrifugation and total protein was quantified via BCA assay (Pierce, Rockford, IL). Two hundred μg of protein was separated on a 12% SDS PAGE gel and transferred to a PVDF membrane using a semi-dry transfer apparatus (Bio-Rad, Hercules, CA). Blots were blocked in 5% nonfat dry milk and immunoblotted with polyclonal rabbit anti-MshA (1:2000 dilution, gift of J. Zhu) and horseradish peroxidase (HRP)-conjugated secondary antibody (Santa Cruz Biotechnology, Santa Cruz, CA). Chemiluminescence was detected with the SuperSignal West Pico reagents (Pierce, Rockford, IL) on the ChemiDoc MP Imager (Bio-Rad Hercules, CA). Densitometry was performed using the Image Lab software v4.0.1 (Bio-Rad, Hercules, CA) using the band in the WT lane as a reference. Blots were performed in triplicate for densitometry analysis and a representative image is shown.
Bacteria were prepared for electron microscopy by inoculating a single colony of each V. cholerae strain in LB broth grown overnight at 30°C with shaking at 200 rpm after which each culture was diluted 1:100 in LB broth and allowed to grow similarly to OD 0.4. An aliquot of each culture was diluted to yield an optical density of 0.1–0.2 and then applied to a 300 mesh carbon-coated Formvar grid (Electron Microscopy Sciences, Hatfield, PA). After 2 minutes, each grid was washed five times with deionized water, and negatively stained with 2% (w/v) aqueous uranyl acetate solution for 90 seconds. Imaging was performed with a JEOL JEM-1400 transmission microscope.
E. coli BL21 harboring plasmids for gene expression were grown to an OD600 of 0.4 at 30°C in LB containing 100μg/mL ampicillin. Cultures were shifted to 18°C and IPTG was added to a final concentration of 100μM. 16h post induction, cells were harvested by centrifugation at 10,000 x g for 15 minutes and stored at -80°C.
Cell pellets were resuspended in GST Lysis Buffer (50mM Tris (pH 8.0), 1M NaCl, 0.5% Tween-20 containing PI cocktail tablets (Roche Life Science, Indianapolis, IN). Cells were lysed by sonication and cell lysate was cleared via centrifugation. Cleared lysate was loaded onto GST FPLC column as follows.
GSTPrep FF16/10 column (GE Healthcare, Piscataway, NJ) was equilibrated in lysis buffer at 1mL/minute using a BioLogic DuoFlow FPLC system (Bio-Rad, Hurcules, CA). Sample was loaded and washed with 1 column volume of GST Lysis Buffer (20mL). Subsequent washes were performed with 20mL of Wash Buffer 2 (50mM Tris (pH 8.0), 0.25M NaCl, 0.5% Tween-20, 0.5mM DTT) and 3 (50mM Tris (pH 8.0), 0.25M NaCl, 0.5mM DTT). Bound protein was eluted with 80mL of GST Elution Buffer 4 (50mM Tris (pH 8.0), 0.25M NaCl, Glutathione 1.5g/L) and collected in 15mL fractions.
These fractions were pooled and concentrated to approximately 10mL using an Amicon 10KDa cutoff spin fliter (EMD Millipore, Darmstadt, Germany). Samples were dialyzed against ITC buffer (25mM Tris-HCl, 150mM NaCl, 250μM DTT, pH 7.5) overnight using 12 kDa cutoff dialysis tubing (Fisherbrand, Pittsburgh, PA). An aliquot of dialyzed protein was diluted in 6M guanidinium HCl and concentration determined via A280.
MshE (18.9μM), PilT (20.2μM), PilU (21.1μM), VpsT (19.5μM) and c-di-GMP (250μM) were prepared in 25mM TrisHCl pH 7.5, 150mM NaCl, and 200μM DTT and degassed prior to analysis. ITC was performed in with a VP-ITC (MicroCal, Northampton, MA) with the following parameters: 3 initial injections of 2μL followed by 40 injections of 10μL spaced at 180 seconds. The data were normalized to a run injecting c-di-GMP into buffer to account for the heat of dilution. Data were processed in Origin v7.0 software (OriginLab, Northampton, MA) and fit to a single site model.
Thermal shift assays were performed as previously described with modifications [41,42]. Briefly, purified protein was added to the reaction to a final concentration of 3μM in the presence or absence of 2mM concentration of the indicated nucleotide in buffer (25Mm TrisHCl pH 7.5, 100mM NaCl, 1:1000 dilution of SYPRO Orange Dye (Invitrogen), and 0.2mM MgCl2. A melt curve protocol was run on an Applied Biosystems ViiA7 qPCR instrument. The fluorescence was measured using the ROX reporter with a temperature gradient of 20–95°C in 0.5°C increments at 30 second intervals. Melt curve data were trimmed to three data points after maximum and the data were plotted with Boltzmann model to obtain the temperature midpoint of unfolding (Tm) of the protein in each condition using Prism 5.0 software (GraphPad). The fluorescence baseline of each sample was normalized to the buffer control for visualization purposes. Three biological replicates were assayed in triplicate and statistical significance was determined with a Oneway ANOVA followed by a Dunnett’s Multiple Comparison test.
ATPase activity of purified proteins was determined by measuring the production of inorganic phosphate from ATP using the Enzchek Phosphate Assay Kit (Invitrogen). The standard reaction mixture was prepared with the addition of 2mM MgCl2, 10mM KCl, and 1mM DTT. Purified protein in buffer (25mM TrisHCl pH 7.5, 100mM NaCl) was added to the standard reaction mixture to a final concentration of 5μM. After a 10 minute incubation at room temperature, ATP was added to a final concentration of 10mM and reactions were incubated at 22°C for 30 minutes. Production of inorganic phosphate was monitored by reading OD360 and compared to a standard curve of solutions of KH2PO4. The data are reported as specific activity (nmol Pi/min/mg of protein). BSA was included as a negative control. Three independent experiments were run in triplicate.
c-di-GMP extraction was performed as described previously [36]. Briefly, 40 ml of culture grown to OD600 ~0.4 was centrifuged at 3220 x g for 30 min. Cell pellets were allowed to dry briefly then re-suspended in 1 ml extraction solution (40% acetonitrile, 40% methanol, 0.1% formic acid, 19.9% water), and incubated on ice for 5 min. Samples were then centrifuged at 16,100 g for 5 min and 800 μl of supernatant was dried under vacuum and lyophilized. Samples were re-suspended in 50 μl of 184 mM NaCl and analyzed by liquid chromatography-tandem mass spectrometry (LC-MS/MS) on a Thermo-Electron Finnigan LTQ mass spectrometer coupled to a surveyor HPLC (Thermo, Waltham, MA). The Synergin Hydro 4u Fusion-RP 80A column (150 mm x 2.00 mm diameter; 4-μm particle size) (Phenomenex, Torrance, CA) was used for reverse-phase liquid chromatography. Solvent A was 0.1% acetic acid in 10 mM ammonium acetate, solvent B was 0.1% formic acid in methanol. The gradient used was as follows: time (t) = 0–4 min, 98% solvent A, 2% solvent B; t = 10–15 minutes, 5% solvent A, 95% solvent B. The injection volume was 20 μl and the flow rate for chromatography was 200 μl/minutes.
The amount of c-di-GMP in samples was calculated with a standard curve generated from pure c-di-GMP suspended in 184 mM NaCl (Biolog Life Science Institute, Bremen, Germany). Concentrations used for standard curve generation were 50 nM, 100 nM, 500 nM, 2 μM, 3.5 μM, 5 μM, 7.5 μM, and 10 μM. The assay is linear from 50 nM to 10 μM with an R2 of 0.999. c-di-GMP levels were normalized to total protein per ml of culture.
To determine protein concentration, 4 ml from each culture was pelleted, the supernatant was removed, and cells were lysed in 1 ml of 2% sodium dodecyl sulfate. Total protein in the samples was estimated with BCA assay (Pierce, Rockford, IL) using bovine serum albumin (BSA) as standards. Each c-di-GMP quantification experiment was performed with four biological replicates. Levels of c-di-GMP were compared to WT with Oneway ANOVA followed by a Dunnett’s Multiple Comparison test.
|
10.1371/journal.pgen.1006382 | Local Adaptation of Sun-Exposure-Dependent Gene Expression Regulation in Human Skin | Sun-exposure is a key environmental variable in the study of human evolution. Several skin-pigmentation genes serve as classical examples of positive selection, suggesting that sun-exposure has significantly shaped worldwide genomic variation. Here we investigate the interaction between genetic variation and sun-exposure, and how this impacts gene expression regulation. Using RNA-Seq data from 607 human skin samples, we identified thousands of transcripts that are differentially expressed between sun-exposed skin and non-sun-exposed skin. We then tested whether genetic variants may influence each individual’s gene expression response to sun-exposure. Our analysis revealed 10 sun-exposure-dependent gene expression quantitative trait loci (se-eQTLs), including genes involved in skin pigmentation (SLC45A2) and epidermal differentiation (RASSF9). The allele frequencies of the RASSF9 se-eQTL across diverse populations correlate with the magnitude of solar radiation experienced by these populations, suggesting local adaptation to varying levels of sunlight. These results provide the first examples of sun-exposure-dependent regulatory variation and suggest that this variation has contributed to recent human adaptation.
| Varying levels of sun-exposure across the world have significantly shaped human evolution. Previous analyses have found several skin pigmentation genes with evidence of strong evolutionary pressures throughout human history, manifesting as large differences in the frequency of genomic variants across populations. But even within populations, individuals respond differently to sun-exposure, suggesting variation in addition to the major differences in skin pigmentation across populations. Here we investigated whether genetic variants associate with response to sun-exposure within Europeans. To measure the response we analyzed gene expression in sun-exposed and non-sun-exposed skin, and identified ten genetic variants that associated with the sun-exposure response of nearby genes. One of these genetic variants, which associated with the sun-exposure response of the gene RASSF9, showed evidence of adaptation in humans in response to solar radiation. Together this evidence suggests that the regulation of gene expression is influenced by sun-exposure and that the sun-exposure dependent effect on RASSF9 expression may have had an effect on human fitness. To our knowledge, this is the first example of an environment-dependent regulatory variant with evidence of adaptation in humans.
| Despite the many successes of genome-wide association studies (GWAS), the field of human genetics is still only scraping the surface of questions encountered in day-to-day life. Questions such as why individuals respond differently to sun exposure, diet, or drugs highlight the prominent role of the environment’s varied effects on every individual. When environmental and genetic variation modify one another’s effects on phenotypes, this is known as gene-by-environment, or GxE, interaction. Many classic examples of GxE interactions exist, such as the inherited condition xeroderma pigmentosa resulting in extreme UV sensitivity [1–5]. But because the environment is both unbounded and fluid, the potential number of GxE interactions is infinite, which has hindered progress on basic questions such as the importance of GxE interactions in evolution and disease [6,7]. Indeed, GxE interactions may help explain the “missing” broad-sense heritability that has been the Achilles heel of GWAS [8–10].
Identifying new GxE interactions from genome-wide approaches is challenging due to their typically small effect sizes coupled with a formidable multiple-hypothesis testing burden [11,12]. However these challenges can be overcome by using gene expression as the trait, i.e. studying how the combination of genetic and environmental variation affects gene expression [13]. This is because cis-regulatory variants (also called cis-acting expression quantitative trait loci, or cis-eQTLs) often have large effect sizes, and identification of these cis-eQTLs requires testing only the SNPs nearby a given gene, resulting in a smaller multiple-hypothesis testing burden [14–16]. In addition, the data required to identify GxE interactions affecting gene expression regulation (hereafter referred to as GxE expression variants) are being produced at an ever-accelerating pace, in large part by consortia such as GTEx [17]. Several studies have successfully identified GxE expression variants in model organisms and in humans [18–26]. The human studies [23,24,26–29] have investigated the roles of various external immunological stimuli, statin exposure, and ionizing radiation in GxE interactions. These studies have identified a wealth of GxE expression variants associated with health-related conditions, such as Crohn's disease and drug-induced myotoxicity, suggesting the importance of GxE interactions in human health.
Here we investigate the presence of GxE expression variants associated with long-term sun-exposure. Sun-exposure is associated with diverse human pathophysiologic phenotypes, from vitamin D synthesis to cancer [30–33]. As humans migrated across the world, changes in solar radiation exposure resulted in strong signatures of local adaptation [34]. Skin pigmentation genes have repeatedly arisen as some of the strongest examples of positive selection in humans [35–37]. Notable examples include SLC45A2 and SLC24A5, which contain non-synonymous SNPs at high frequency in lighter-skin populations [34].
In addition to protein-coding changes, there is also evidence that non-coding changes have evolved in response to sun-exposure. Cis-regulatory regions of several skin-pigmentation genes show signs of selection, such as the intergenic regions of KITLG and the upstream enhancer of OCA2 [34,38,39]. Polygenic signatures of selection have also been reported: eQTLs for genes down-regulated by UV exposure show significantly higher allele frequencies of the down-regulating allele in populations with higher UV exposure [40], suggesting that the dynamic response to UV may also be "hard-wired" via adaptive changes in eQTL allele frequencies. Limitations of these studies are that genetic control of gene expression and sun-exposure are often only indirectly linked (because the SNP, gene expression, and sun-exposure were not all measured in the same study) and they do not address whether the genetic control of gene expression itself may be dynamic and depend on the environment.
To identify genetic variants whose effect on gene expression depends on sun-exposure, we analyzed RNA-seq data from 607 skin samples [17] including sun-exposed skin (357 samples, hereafter denoted as SE) and non-sun-exposed skin (250 samples, hereafter denoted as NSE), which differ in lifetime sun-exposure. We then investigated whether the sun-exposure-dependent eQTLs show signs of local adaptation. Although evidence for positive selection of skin-pigmentation genes due to sun-exposure is strong, explanations for how these adaptations improved fitness are debated [33]. Hypotheses range across a gamut of health associations, including Vitamin D synthesis, folate-dependent neurological development, cancer, dehydration, and innate immunity [32,33]. By identifying GxE expression variants, we aim to identify the genes, and ultimately the specific traits, that have been critical for our recent adaptation to local climates around the world.
We first examined the transcriptome-wide gene expression of the skin samples relative to the other tissues. Principal components of all GTEx samples were calculated. The 1st to 6th principal components primarily distinguished the samples by tissue, while the SE and NSE skin samples grouped together indistinguishably (Fig 1A, S1 Fig, Methods). Hierarchical clustering of the expression data also clustered the SE and the NSE samples together (S2 Fig, Methods). Thus, we estimate that sun-exposure has a subtle effect on gene expression in skin, when compared to the differences between tissues.
Because the principal component analysis was able to discriminate between tissues, we next investigated whether we could distinguish the exposure-type of the skin samples by performing the analysis on the skin samples alone. The first six principal components did not segregate the samples by exposure-type and the percent variance explained by each principal component was relatively small (PC1 explained 7.2% variance, PC2 explained 4.2%; percent variance cumulative in PC1-6 was 20.8%). This suggests that sun-exposure does not explain a large fraction of the variance in gene expression among these samples (Fig 1B, S3 Fig).
Despite this similarity in the SE and NSE samples, we have high power to detect differential expression because of the large sample sizes (196 NSE and 302 SE, after removal of technical replicates; see Methods). We performed differential expression analysis with DESeq2 [41], revealing 12,320 differentially expressed genes at a Benjamini-Hochberg FDR < 0.01 (out of 37,412 genes tested, S4 Fig, S1 Table). Among these genes, 522 genes had greater than 2-fold-change. We investigated the Gene Ontology enrichment of these genes using DAVID [42], identifying several enriched gene sets (S2 Table). Genes that were upregulated in SE were most strongly enriched for the GO term "Epithelial Cell Differentiation" (Bonferroni-corrected p = 1.5x10-4), whereas those upregulated in NSE were most strongly enriched for "Intermediate Filament" (Bonferroni-corrected p = 6.1x10-9).
As a critical control, we investigated whether the gene expression differences between the SE and NSE samples, which were obtained from the lower-leg and suprapubic skin of post-mortem samples, reflect effects of sun-exposure. To test this, we compared the list of differentially expressed genes with genes that were differentially expressed in human skin after repeated exposures to UVB [43]. We observed significant enrichment (Fisher's exact test p = 3.2x10-6 for differentially expressed genes at Benjamini-Hochberg FDR < 0.01, Fig 1C), indicating that sun-exposure differences in the post-mortem GTEx skin samples are broadly consistent with experimental UVB exposure. We also tested for overlap with genes differentially expressed in human skin fibroblasts in response to UV radiation [44], and found a similar trend (Fisher's exact test p = 1.6x10-6 for differentially expressed genes at Benjamini-Hochberg FDR < 0.01, S5 Fig). Investigating whether sun-exposure responses differed between ethnicities, we observed similar responses between individuals with predominantly European and African ancestry (see S1 Text).
We next investigated whether sun-exposure may affect cis-regulation, specifically at genomic loci that associate with mRNA levels (eQTLs). Previous studies of differential cis-regulation across tissues or environments have identified potential pitfalls that could lead to false positives, such as the “winner's curse” and differences in power, which we carefully aim to avoid in our analysis [45–47] (Fig 2A). To focus on the most replicable eQTLs, we first identified the strongest local eQTLs within 1 Mb of each gene. We used a Bayesian approach to map eQTLs in the SE samples and the NSE samples jointly (302 SE samples and 196 NSE samples after restricting to only samples with individual genotypes, mapped using eQTL-BMA [46]). By using only the strongest eQTL from joint-sample mapping, we remove the bias of inflated effect size estimates due to the “winner's curse”, a confounding effect that reduces the apparent replication rate when the strongest eQTL in one group is tested in another group [48]. eQTLs were selected as those that had a posterior probability greater than 0.95, resulting in 8739 eQTLs (1 SNP per gene, S6 Fig). These genes predominantly overlapped with GTEx eQTL target genes, and the posterior probabilities in our analysis were strongly correlated with the GTEx p-values, indicating that joint-sample mapping produces largely similar results (S7 Fig).
For each of the eQTLs, we then tested for significant differences in cis-eQTL effect between the SE and NSE samples. Unlike tests comparing eQTL significance between groups—which lead to false positives if low power in one group results in insignificant p-values—comparing the effect sizes (hereafter referred to as the effect-size test; [49]) does not result in false positives, since low power in one group would only decrease the ability to detect a significant difference in effect size. Applying this test to all 8739 eQTLs, we found 453 at a nominal p < 0.05, with only one eQTL significant at Benjamini-Hochberg FDR < 0.05. The most significant result was for SIM2, which was strongly affected by an eQTL in the NSE samples, but not in the SE samples (Fig 2B).
To validate these results, we analyzed allele specific expression (ASE) differences (differential ASE) between SE and NSE samples. ASE measures the effect of cis-regulation in individuals heterozygous for the regulatory variant. Thus, significant differential ASE between groups (comparing only the individuals who are heterozygous at the eQTL) indicates differential effects of the eQTL. Note that although ASE is measured using the same dataset, it is an independent source of validation because it examines the differences in expression within an individual rather than across individuals [23,25] (see simulation results in S8 Fig and S2 Text). As a control, we also confirmed that ASE identifies cis-regulatory effects in this dataset and is not biased between the SE and NSE groups (S9 Fig).
To test for validation of the SIM2 sun-exposure-dependent eQTL (Fig 2B), we compared the ASE in all individuals who were heterozygous at the eQTL. Because the eQTL effect was present only in NSE samples, we predicted that the NSE heterozygotes would exhibit higher absolute magnitude of ASE than SE heterozygotes. In addition, because the G allele is associated with higher expression in the NSE eQTL analysis (Fig 2B), we predicted that the ASE of the NSE heterozygotes would be directed towards the G allele. We confirmed both predictions (Fig 2C), validating SIM2 as the target of a GxE expression variant.
Extending this validation test to all genes, we found that results from differential ASE and the effect-size test are significantly concordant in directionality (S9 Fig). Concordance in directionality requires: 1) the same sample group (SE or NSE) must exhibit the stronger effect in both tests, and 2) the same allele must be the upregulating allele. The degree of concordance is likely underestimated by our analysis because the ASE data face several systematic limitations that reduce power, including imperfect phasing between the eQTL and SNPs inside the target transcript, and the small number of reads overlapping many heterozygous SNPs.
Because differential ASE provides independent evidence for a differential cis-regulatory effect, we combined both tests to identify additional genes with a significant effect. To combine the tests, we intersected significant SNPs by testing for differential ASE among genes with a nominally significant effect-size test (p < 0.05) and concordant ASE directionality (157 out of 289 genes concordant), resulting in seven significant differential eQTLs (Benjamini-Hochberg FDR < 0.05, Fig 2D). Four out of these seven genes show a stronger effect in the NSE samples, confirming that there is no evidence of bias due to differences in power between the groups (S3 Table). Only two of the seven genes exhibited significant differential expression at FDR < 0.05, and these two genes exhibited a stronger cis-eQTL effect in the environment with lower expression (S3 Table). This indicates that the trivial scenario of environment-specific gene expression silencing resulting in no evidence of cis-regulation for that particular environment is not occurring in these associations. Many of the genes have known roles in skin and sun exposure; for example LYPD5, the gene with the strongest effect, encodes the protein haldisin, which is associated with the late (outer) stages of skin differentiation [50,51]. NLRC3, the second strongest effect, is a gene that negatively regulates the STING-dependent innate immune pathway [52] that is activated after UV exposure [53,54].
We then identified additional GxE expression variants by testing whether any local SNPs associate with the differential expression due to the sun-exposure of each individual (Fig 3A). This approach has been used in identifying GxE expression variants responding to immunological stimuli, referred to as cis-response eQTLs (cis-reQTLs) [23,26]. The advantage of this method is that comparing the expression within the same individual will account for the inter-individual differences, such as environmental variation, that can confound typical eQTL analysis. To map the cis-reQTLs, we first calculated the fold change difference in expression (SE / NSE) for all genes in the 147 individuals with both SE and NSE data (see Methods). We then tested the fold change values for association with SNPs within one megabase from the transcription start site of the gene, in a similar manner to eQTL mapping. As expected, there is a significant concordance between the cis-reQTL p-values and the eQTL effect-size test (S10 Fig). For the cis-reQTL analysis however, we also tested all local SNPs as opposed to just the strongest eQTLs, resulting in additional associations. The three most significant cis-reQTL genes were RASSF9, NPIPL1, and SLC45A2 (Fig 3B; see Methods, p<2x10-5, Benjamini-Hochberg FDR<0.18). These cis-reQTLs were also nominally significant using the effect-size test (p = 5.0x10-6, p = 2.1x10-5, p = 2.7x10-2, respectively).
SLC45A2 is a classic example of a skin-pigmentation gene under recent positive selection [37,55,56]. The gene encodes the Membrane-Associated Transporter Protein (MATP), containing a non-synonymous SNP (rs16891982) where the ancestral allele is nearly fixed in African and Asian populations while the derived allele is at high frequency with a north-south cline in European populations [34]. The cis-reQTL found in this study (rs12653176) is not in linkage disequilibrium with the previously-identified non-synonymous SNP in any tested populations (S11 Fig), and is located within the intron of the downstream gene ADAMTS12. The fold-change association in this cis-reQTL is more significant than either the SE or the NSE expression associations (Fig 3C and 3D), and SLC45A2 had no significant cis-eQTL in skin (Fig 3E), suggesting that the cis-reQTL specifically controls the SE-response rather than baseline levels.
The strongest cis-reQTL in our analysis regulates RASSF9, a gene that is expressed highly in mouse epidermal keratinocytes [57]. RASSF9 knockout mice were previously observed to have higher epidermal proliferation and abnormal differentiation, suggesting that the gene is involved in epidermal homeostasis.
We then tested the differential cis-eQTLs and cis-reQTLs (hereafter referred to collectively as sun-exposure dependent eQTLs—se-eQTLs) for evidence of recent adaptation (S3 Table). Our hypothesis, based on previous findings [40,58], was that adaptation of human populations to local climates would lead to a correlation between the se-eQTL allele frequencies and the levels of solar radiation experienced by each population. To test this hypothesis, we used Bayenv 2.0, a Bayesian approach to test the significance of environmental associations while accounting for the relatedness between populations and the confidence in allele frequency estimates [59,60]. In a previous genome-wide analysis of 61 human populations, Bayenv identified evidence of local adaptation to solar radiation near several skin-related genes (e.g. SLC45A2, KRT77, and OCA2) [58].
Here we use Bayenv to test for association between solar radiation and the allele frequencies of the 10 se-eQTLs across the HGDP populations [60,61]. In the same manner as [58], we analyzed the winter and summer solar radiation separately. To estimate significance of the resulting Bayes factors, we calculated an empirical p-value using all skin eQTLs discovered in our study as background (see Methods). Among our 10 se-eQTLs, RASSF9 had the strongest association with winter solar radiation (Fig 4A; empirical p = 6.1x10-3, S3 Table). Hancock et al [58] also found the strongest associations in skin pigmentation genes using the winter solar radiation as the environment, as opposed to summer solar radiation. We then validated the evidence for selection in RASSF9 in another set of 92 populations while excluding HGDP individuals (empirical p = 0.040, Fig 4B, S3 Table, populations from [62]). The associations in both population sets were concordant across multiple continental regions (Spearman’s rho < -0.36, Fig 4), suggesting that this se-eQTL has been adaptive in multiple independent human lineages.
Commonly used haplotype-based tests for selection, such as EHH, did not show a significant increase for any particular se-eQTL, which may be an indication that the se-eQTLs are under very recent selection. As selective sweeps for large-effect skin-pigmentation genes have occurred within the last tens of thousands of years [37,63], we expect these small-effect events to also be recent as well. We tested this further using the singleton-density-score (SDS), a haplotype-based test of very recent selection performed on the UK10K Project data (Y. Field, N. Telis, and J. Pritchard, pers. comm., [64]). The se-eQTL for RASSF9 showed the strongest evidence of selection (Bonferroni corrected p = 0.018), lending further support to the hypothesis of recent selection acting on this locus.
Our study of sun-exposure dependent regulation of gene expression identified several examples of GxE interactions where genetic variants affect the gene expression response to sun exposure. In contrast to our differential expression analysis that identified over 12,000 transcripts—an unwieldy number that makes it challenging to uncover the causal or independent effects—focusing on GxE interactions allowed us to identify 10 independent loci directly associated with the sun-exposure response, with the strongest hits having known skin or sun-exposure related functions. This small number of significant GxE expression variants is consistent with previous observations that the number of GxE effects detected is smaller than additive effects [25].
We identified GxE expression variants using two methods, differential cis-eQTL analysis and cis-reQTLs. Although both methods uncover GxE expression variants, we found that each provided unique results in their top associations. The two methods have distinct advantages: differential cis-eQTL analysis utilized information from a larger number of samples because there was no requirement for individuals to have both the SE and NSE data, while the cis-reQTL analysis controlled for possible individual-level confounding factors (such as genetic background or environmental variation) because it compared the SE and NSE samples within the same individuals. In addition, our differential cis-eQTL analysis was based on eQTLs identified via joint-sample mapping in order to reduce the effect of the winner's curse. As a result, the differential cis-eQTL analysis preferentially analyzed SNPs with eQTL effects in both tissues, while the cis-reQTL approach agnostically considered all local SNPs. Despite these differences, overall the methods were generally concordant (S10 Fig).
Building on previous studies, we also found that allele-specific expression is an effective, orthogonal approach to identify GxE expression variants [25]. Combining the differential ASE test with the effect-size test strengthened the significance of our findings; therefore we recommend investigating both allele-specific and non-allele-specific expression whenever possible.
An important control in our analysis was confirming that the differences in gene expression of the SE and NSE samples reflected differences in sun-exposure. Because the samples were obtained from the post-mortem lower-leg and suprapubic skin, respectively, it is important to justify that the downstream analyses reflect differences caused by sun-exposure as opposed to differences unrelated to sun exposure. The relevance of sun-exposure was validated by significant concordance of the differentially expressed genes with two independent experimental approaches, one with whole skin UV-B treatment and another with skin fibroblast UV exposure. In addition, the significance of SLC45A2, a gene previously determined to be under selection due to solar radiation, also suggests that sun-exposure is likely driving this effect. Lastly, SLC45A2 and RASSF9 have both been found to be differentially expressed in response to UV-B treatment [43], thus indicating that sun-exposure is likely affecting these genes' expression.
In addition to sample sizes, statistical power can also affect GxE interaction studies via a large multiple-hypothesis testing burden. This limitation was a key factor in our choice of testing only local GxE expression variants and refraining from testing for distant trans-associations. Local eQTLs tend to be highly replicable with strong effect sizes and finding them requires testing only a small number of SNPs per gene. In contrast, trans-eQTLs are challenging to map if every SNP is tested for each gene, though it is possible to identify trans-acting GxE expression variants in well-controlled studies [18,20,29]. Larger sample sizes or testing only SNPs with a high prior probability of participating in trans-GxE effects can also boost power to map them.
Overall, the GxE expression variants did not exhibit the strong population-specific fixation patterns observed in classical examples of positive selection. This however does not necessarily indicate that GxE expression variants are in general weakly selected. Mapping GxE interactions requires variation not only in environment but also in genotype. Because strong selection will remove genetic variation, our GxE mapping approach may not detect associations with the most strongly selected SNPs. Because the GxE mapping process focuses on polymorphic variants, we may detect signs of selection that are perhaps more recent or complex.
Although the evidence for human evolution in response to sun-exposure changes is strong, hypotheses for how adaptations in skin-pigmentation genes have led to increased fitness are debated [32,33]. To our knowledge, the se-eQTL genes, aside from SLC45A2, were not previously associated with human adaptation to sun-exposure. We found evidence that the se-eQTL of RASSF9 has been subject to local adaptation using two independent population cohorts as well as from a test for very recent adaptation in Europeans. Our finding of adaptation involving RASSF9, a gene involved in epidermal homeostasis and differentiation, supports the idea that skin pigmentation is not the only trait shaped by adaptation to sun-exposure [65,66].
The approaches we employed could be applied to investigate GxE interactions involving any pair of tissues; although environment is typically considered to be extrinsic to an organism, cell type can be considered a major determinant of each gene’s intracellular environment. As more GxE interactions are discovered, we can then study their relationships with diseases or other traits to better understand their role in phenotypic variation and adaptation.
Expression and genotype data for sun-exposed (SE) and non-sun-exposed (NSE) skin samples were obtained from GTEx v6 [17]. Genotype data are from the HumanOmni5-4v1_B and the HumanOmni2.5-8v1-1_B platform with variants imputed using IMPUTE2 and the 1000 Genome Project Phase 1 version 3 reference panel (genotyping and imputation performed by GTEx consortium). To account for platform bias, we included genotyping platform as covariates in downstream analyses as described below. Sun-exposed skin samples were taken from the lower leg, and the non-sun-exposed skin samples were taken from the suprapubic area. As described in the GTEx resources, the skin samples were obtained as slices with the subcutaneous fat trimmed off, avoiding pubic hair in the suprapubic region, and subsequently fixed and stored in PaxGene tissue container. For the genotype data, we used all SNPs that passed the filters in the phased/imputed dataset (INFO > 0.4, Hardy Weinberg p-value > 10−6, call-rate > 95%). For the expression data, we used the samples selected by GTEx as adequate for eQTL analysis (which were selected based on quality). Gene models for all expression analyses were also obtained from GTEx, which combined the transcripts annotated by GENCODE v19 for each gene. Principal components in the all-sample analysis were calculated from the RPKM expression of 1000 randomly-sampled genes with a mean RPKM > 1. Random sampling was performed to reduce computation time—repeating the sampling produced similar results. Hierarchical clustering was performed with the Euclidean distance between tissues using the tissue-specific mean expression across all genes. Principal components of the skin-only analysis were obtained using all genes with mean RPKM > 1 across the skin samples.
We investigated differential expression between the sun-exposed and non-sun-exposed skin samples using the raw read counts obtained from the GTEx consortium (which used RNASeQCv1.1.8 to estimate values). Genes with zero counts across all samples were removed. Differential expression was calculated using the DESeq function in the DESeq2 package [41] using default parameters. Gene Ontology enrichment was assessed using the Ensemble Gene IDs with DAVID [42]. All genes that were tested for differential expression were used as the gene background and the Gene Ontology BP, CC, and MF collections were tested. For the ancestry-specific differential expression analysis, genotyped individuals were separated into European ancestry (377 individuals) and African ancestry (68 individuals) as determined by principal component analyses. Europeans ancestry was designated as individuals with PC1 <-0.01 and PC2 > -0.04) and African ancestry population was designated as individuals with PC1 > 0.1. This classification was concordant with the individuals' reported race (S12 Fig). We then tested each gene for evidence of ancestry-specific sun-exposure differential expression by assessing the significance of the ethnicity-environment interaction term in a linear model. The likelihood ratio test was performed within DESeq2, comparing the models (Expression ~ ancestry + exposure, and Expression ~ ancestry + exposure + ancestry:exposure).
Joint sample eQTL mapping was performed using eQTL-BMA with both SE and NSE samples [46]. To maximize the number of samples and to maintain consistency with the GTEx eQTL association mapping, we used all individuals (including individuals of varying ancestry) in the analysis. To control for effects of ancestry and other confounding factors, we used all covariates used by GTEx in their eQTL association mapping, which includes the first three genotype principal components, PEER factors, genotyping array platform, and gender. For input of the expression data, we used the residuals after regressing out covariates from the normalized expression matrix provided by GTEx (normalization used by the GTEx: RPKM data, low-expression filtered, quantile normalized, fitted to standard normal). eQTLs were mapped to SNPs within 106 bp of the TSS, with the parameters "—bfs all—error hybrid—maf 0.05—qnorm—analys join" and using the large and small grid provided by the package. This window size of 106 bp was selected to be consistent with the GTEx Consortium analysis [17]. The configuration weights were determined using the hierarchical model and pi0 was calculated using the EBF procedure as specified by the package.
For the allele-specific expression analysis, reads provided by GTEx were mapped using STAR (—outFilterMultimapNmax 1,—clip 5pNbases 6) [67] to the GRCh37.p13 human genome masked with all SNPs in dbSNP142. Duplicate reads were removed using a combination of in-house scripts and Samtools Rmdup to remove duplicate reads randomly (instead of selecting the read with the highest quality score) as suggested previously [68,69]. Allele-specific expression was calculated at each exonic SNP, with a random SNP chosen when a read overlapped multiple sites. Mono-allelic sites were removed as recommended [69] to reduce bias and the gene-level exon counts were summed using the GTEx phased SNP calls. Using these counts, the value of ASE for each gene-individual pair was calculated as allele1 / (allele1 + allele2), requiring at least 3 reads. For comparison between ASE results and differential eQTL analysis, we required at least 5 ASE measurements for SE homozygotes, SE heterozygotes, NSE homozygotes, and NSE heterozygotes.
The effect-size test was performed as described in Fraser et al. [49]. Pearson correlation for SE and NSE samples were calculated using the normalized expression matrices used by GTEx eQTL analysis. Each correlation coefficient was converted to a z-score by 0.5*log((1+r)/(1-r)). The difference in effect size was then tested by calculating the value t = (zSE—zNSE) / sqrt (1/(nSE -3) + 1/ (nNSE-3)), which is normally distributed. zSE and zNSE denote the z-scores for the SE and NSE samples. nSE and nNSE denote the number of samples used in the correlation coefficient calculation.
cis-reQTL analysis was performed using the subset of individuals analyzed in the differential eQTL analysis who had both SE and NSE data. To calculate fold change, the RPKM values for each sample were first quantile normalized across samples. Fold change (SE / NSE) was then calculated for each gene-individual pair, with values of 0 set to the minimum non-zero value of that gene. The log fold change values for each gene were then fit to a standard normal. Hidden covariates were then discovered using PEER (k = 15) [70]. In order to correct for hidden covariates most relevant to these log-ratios, PEER covariates were assessed for these directly (as opposed to using RPKM values). The first 15 PEER covariates, gender, sequencing platform, and the first 3 genotype principal components were used as covariates for mapping (Results remained consistent after also correcting for age). Mapping was performed with the linear model of Matrix eQTL using SNPs with MAF > 0.05 found within 106 bp from the transcription start site [71]. The window size of 106 bp was selected to remain consistent with the GTEx Consortium analysis for cis-eQTL mapping [17]. Significance of the SNP with the strongest association for each gene was assessed by comparing the nominal p-value with the p-values of the best-associating-SNPs from at least 103 permutations of the individual labels.
Local adaptation was tested using Bayenv2.0 using the Human Genome Diversity Panel (HGDP) and HapMap samples [60,61] and an independent set of 92 populations [62]. The Lazaridis et al. data [62] were filtered to remove African American individuals, HGDP individuals, and populations with fewer than five individuals. The population covariance matrix was estimated with 10,000 random SNPs selected from the genotyped SNPs with linked SNPs removed by PLINK v1.9 (—indep-pairwise 1000kb 1 0.2) and k = 1000. Many of the se-eQTLs were not genotyped by HGDP or Lazaridis et al., and thus the allele frequencies were estimated by first imputing the genotypes for each individual with IMPUTE 2 [72] using the 1000 Genomes Phase 3 reference panel and default parameters. Bayenv results have previously been observed to be unstable with a small number of iterations [73], so we took the median of 10 runs with k = 500,000. To estimate the significance of the Bayes factors, we performed the same analyses on all eQTLs discovered in our analysis to use as the background.
Climate data were obtained from the NCEP/NCAR Reanalysis [74] using the monthly long-term mean of the downward short-wave radiation flux to match the environmental variables used by Hancock et al. [58]. The summer radiation was calculated by averaging the months of June, July, August for positive latitudes and December, January, February for negative latitudes. The winter radiation was calculated by averaging the opposite months. Values were then standardized for Bayenv analysis.
|
10.1371/journal.pbio.0060215 | Texture Coding in the Rat Whisker System: Slip-Stick Versus Differential Resonance | Rats discriminate surface textures using their whiskers (vibrissae), but how whiskers extract texture information, and how this information is encoded by the brain, are not known. In the resonance model, whisker motion across different textures excites mechanical resonance in distinct subsets of whiskers, due to variation across whiskers in resonance frequency, which varies with whisker length. Texture information is therefore encoded by the spatial pattern of activated whiskers. In the competing kinetic signature model, different textures excite resonance equally across whiskers, and instead, texture is encoded by characteristic, nonuniform temporal patterns of whisker motion. We tested these models by measuring whisker motion in awake, behaving rats whisking in air and onto sandpaper surfaces. Resonant motion was prominent during whisking in air, with fundamental frequencies ranging from approximately 35 Hz for the long Delta whisker to approximately 110 Hz for the shorter D3 whisker. Resonant vibrations also occurred while whisking against textures, but the amplitude of resonance within single whiskers was independent of texture, contradicting the resonance model. Rather, whiskers resonated transiently during discrete, high-velocity, and high-acceleration slip-stick events, which occurred prominently during whisking on surfaces. The rate and magnitude of slip-stick events varied systematically with texture. These results suggest that texture is encoded not by differential resonant motion across whiskers, but by the magnitude and temporal pattern of slip-stick motion. These findings predict a temporal code for texture in neural spike trains.
| A fundamental problem in neuroscience is understanding how behaviorally relevant information is collected by a sensory organ and subsequently encoded by the brain. By actively moving their whiskers, rats can discriminate fine differences in textures. Little is known, however, about how whisker dynamics reflect texture properties or how the nervous system encodes this information. In one hypothesis, whisker motion over a texture produces a unique, texture-specific temporal profile of velocity, which is encoded in the temporal pattern of neural activity. In a second, alternative hypothesis, textures excite a specific subset of whiskers due to intrinsic, whisker-specific mechanical resonance frequencies. Information is then encoded by the spatial distribution of neural activity in whisker-related columns in cortex. Here, we assess these hypotheses by measuring whisker motion as animals whisk across sandpapers of varying roughness. We found that whiskers resonate in air and on surfaces, but that these resonance vibrations do not vary with, and therefore do not encode, texture. Instead, whisker motion over a textured surface produces fast, transient slip-stick events whose dynamics are dependent on texture roughness. Texture is likely to be encoded in the temporal pattern of spikes evoked by these slip-stick events.
| Rodent whiskers, like human fingertips, are tactile detectors that are actively moved through the environment to sense position, shape, and surface features of objects. A particularly salient surface feature is texture, which is more readily distinguishable using touch than vision [1]. Rats discriminate textures using their whiskers with a precision that rivals human fingertips [2–5]. How whiskers read out texture information, and how that information is encoded in the nervous system, are vigorously debated, and have important implications for sensory processing in the whisker system [6,7], which is a major model system for studying cortical function and plasticity [8,9].
Rats have an array of approximately 30 large whiskers (macrovibrissae) on each side of the face. Whisker length varies systematically across the whisker pad, with caudal whiskers being longer than rostral whiskers. Whiskers are moved rhythmically at 5–15 Hz to explore objects in the environment, including textures [5,6,10]. Two main hypotheses exist for texture discrimination by the whiskers, based on experiments using detached whiskers and in anesthetized animals. The resonance hypothesis derives from the observation that whiskers are resonant beams, with characteristic resonance frequency inversely related to whisker length [11,12]. Whisker-length variation across the whisker pad results in a spatial map of fundamental resonance frequency (FRF). In this hypothesis, whisker-tip motion across surface microfeatures causes tip vibration at a frequency that varies with texture spatial frequency. Only when textures generate tip vibration at the FRF will vibrations most effectively build up and be transmitted to the whisker follicle, where transduction occurs. As a result, each whisker is best activated by a specific range of textures, and each texture preferentially activates a subset of whiskers, leading to a spatial code for texture in the relative amplitude of vibrations across the whisker array [13]. An alternative model is that texture is encoded temporally, by unique temporal patterns of movement (“kinetic signatures”) that are induced within single whiskers scanning across surfaces. These patterns have been proposed to include both mean speed (amplitude × frequency) of whisker vibration [7,14], spectral composition of whisker vibrations [15], and the precise, irregular velocity profile of whisker motion [7]. This latter feature provides higher-resolution texture information than vibration speed or frequency alone [7,16].
To distinguish these models, it is critical to measure whisker vibrations and neural responses in awake, behaving animals voluntarily palpating surfaces. This is because the dynamics of voluntary whisker movement will critically impact the transformation of surface features into whisker-motion signatures. Whiskers are known to exhibit multiple modes of vibration during voluntary palpation of surfaces, including resonance vibration and irregular, high-velocity motion events [17]. However, which of these features correlate with, and therefore may encode, texture, is not known.
Here, we evaluated the resonance and kinetic signature models of texture by precisely measuring whisker vibrations in awake, behaving rats trained to actively whisk onto textured surfaces. Results showed that whisker resonance occurs during free whisking in air and during brief, discrete epochs while whisking onto textures. However, the magnitude of resonance vibrations did not vary across textures, as required for the resonance hypothesis. Instead, whisker resonance on surfaces primarily represented transient ringing during brief (5–10 ms), high-velocity, high-acceleration slip-stick events. Slip-stick events were a prominent component of whisker motion on surfaces, and the rate and magnitude of these events correlated well with texture. These results indicate that whisker resonance occurs in awake rats and shapes natural whisker vibrations, but that texture is not encoded by differential resonance across whiskers, at least under these behavioral conditions. Instead, slip-stick events may contribute to a kinetic signature for texture in the whisker system.
To measure whisker movement in awake, behaving rats, we trained rats to whisk in air and against textured surfaces. Two behavioral paradigms were used. In Behavior 1, six rats (N1–N6) positioned their nose in a small aperture (the nose poke) and whisked in air and onto surfaces for approximately 0.5 s to receive a water reward (Figure 1A). Textured surfaces (sandpapers of varying roughness, mounted on aluminum backing) were positioned statically in the whisking path of the right whiskers using a computer-controlled stepper motor. Whisker motion in the protraction–retraction plane (roughly rostrocaudal, parallel to the face) was measured optically from whisker shadows cast by a collimated plane of laser light onto a linear charge-coupled device (CCD) imaging array below the training cage (Figure 1C–1E). Each trial consisted of whisking either in air or onto one surface, and lasted 491 ± 179 ms. Between trials, rats moved to a separate chamber to receive a water reward, and surfaces were changed using the stepper motor. Rats performed 123 ± 43 (mean ± standard deviation [s.d.]) trials per daily session. In Behavior 2, four rats (H1–H4) were habituated to being transiently head-fixed, and whisked voluntarily in air and onto surfaces (Figure 1B). During each daily session (15–30 min), rats performed 69 ± 36 trials, with a trial defined as a 3-s epoch that included a variable duration of whisker motion. See Materials and Methods for training techniques.
In both behaviors, motion of one to four identified whiskers was tracked at 4-kHz frame rate and approximately 5-μm spatial resolution, using the linear CCD array. Because whisker shadows did not cross during whisking, up to four whiskers could be identified and tracked simultaneously using automated software (Figure 1D). Nonimaged whiskers were trimmed weekly at the base. Surfaces were presented parallel to the face, less than 5 mm from the whisker tips. Whisker motion was tracked 6–14-mm (typically 10 mm) from the face (for whisking in air), and halfway between the surface and the whisker pad (for whisking onto textures). All training and whisker measurements were performed under computer control using custom-written programs in Labview (National Instruments).
We first tested the resonance hypothesis by asking whether whiskers resonate, and whether a map of resonance frequency exists, in awake, behaving rats whisking in air. For rats whisking in the nose poke (Behavior 1), whisker motion typically included periods of regular, 5–15 Hz whisking, periods when the whiskers were held stationary, and periods of erratic motion. Examples are shown in Figure 2A. During all three types of motion, bandpass filtering (20–1,000 Hz) revealed prominent high-frequency whisker vibrations (20–150 Hz), with approximate peak-to-peak amplitude of 0.1 to 0.5 mm, that were superimposed on the low-frequency motion (Figure 2B).
These high-frequency vibrations were not apparent during motion of an isolated whisker attached to an electric motor moving sinusoidally at 8 Hz (Figure 2C), indicating that they were not due to external vibrations in the recording apparatus or to interaction between the moving whisker and air. To test whether vibrations were due to head motion, versus whisker motion relative to the head, we simultaneously measured head and whisker motion in one rat (N4) by attaching a horizontal bar to the top of the skull. The bar cast a shadow on the CCD array that could be tracked independently of the whisker shadows. Head motion and D1 whisker motion showed coherence at low frequencies (<8 Hz), but very little coherence (mean 13%) at frequencies greater than 10 Hz (Figure 2D, green trace). The same was true for head motion and D3 whisker motion (Figure 2E). Thus, head motion is not the source of high-frequency whisker vibration at greater than 20 Hz. Consistent with this conclusion, high-frequency whisker vibrations were also prominent in head-fixed rats whisking in air (unpublished data).
Despite the lack of coherence between whisker motion and head motion, neighboring whiskers exhibited high coherence in the 20–150 Hz range (Figure 2D and 2E). This suggests a common driving force for high-frequency vibrations across whiskers. Coherence between whiskers decreased slightly with whisker separation on the face, suggesting that neighboring whiskers receive the greatest common drive (Figure 2E). Together, these measurements indicate that high-frequency, coherent vibrations occur during free whisking in air, superimposed on low-frequency whisking motion.
We tested for whisker resonance during free whisking in air by examining the relationship between the frequency spectrum of whisker vibrations in air and the intrinsic resonance frequencies of the whiskers. Experiments were performed in rats N1–N4 performing the nose poke task (Behavior 1). Power spectra during whisking in air were calculated for each whisker across 44–122 trials. Power spectra were not smooth, but rather showed modest peaks and shoulders representing dominant frequencies of whisker vibration. Two such power spectra, from the D1 and D2 whiskers imaged simultaneously in rat N2, are shown in Figure 3A. Peaks and shoulders were identified precisely as minima in the second derivative of the logarithm of the power spectra, which correspond to points of negative concavity (filled circles in Figure 3A and 3B).
After whisking in air, rats were anesthetized; the intrinsic resonance frequency for each whisker was directly measured by manually delivering an impulse to the whisker, and the FRF from the resulting decaying oscillations in air was calculated, as imaged on the CCD array (see Materials and Methods). Examples are shown in Figure 3C for the D1 whisker (length: 38.7 mm, FRF: 44.1 Hz) and D2 whisker (length: 31.6 mm, FRF: 60.5 Hz) from rat N2 (same whiskers as for the power spectra in Figure 3A). Theoretical first harmonics of the FRF were calculated as (10.6/4.4) × FRF, as predicted for a conical beam model of the whisker (see Materials and Methods). For this rat, peaks in the power spectra during voluntary whisking in air (filled circles) were found to align well with the measured FRFs (open circles) and calculated first harmonics (asterisks) obtained by the impulse method in the anesthetized animal (Figure 3A).
Across eight whiskers in rats N1–N4 (n = 4 rats), the first peak in the power spectra during natural, active whisking aligned well with the measured FRFs (Figure 3D), and the second peak aligned with the predicted first harmonics (Figure 3E). Moreover, the ratio of the frequencies of the first and second peaks in the power spectra was 2.35 ± 0.14 (n = 8), close to the theoretical ratio of 10.6/4.4 = 2.41 for f1/FRF for a conical beam. Thus, during natural whisking in air, whiskers preferentially vibrated at the FRF and its first harmonic, though the magnitude of these vibrations was small.
To test whether a map of whisker resonance exists across the whisker pad in the awake, behaving rat, we compared first peaks in the power spectra during whisking in air and FRFs measured by the impulse method, to whisker length (Figure 4). We found a systematic relationship in which longer whiskers exhibited lower FRFs and vibrated preferentially at these lower frequencies during whisking in air. This indicates that a map of resonance frequency exists across the whiskers in awake, whisking animals, as predicted by the resonance hypothesis [11–13].
To confirm the role of whisker resonance in shaping high-frequency vibrations in air, we systematically altered whisker resonance by trimming whiskers in two rats (N2 and N3). Power spectra were measured daily, for 10–11 d, during whisking in air for whiskers δ, D1, D2, and D3. In rat N2, the D2 and D1 whiskers were trimmed by approximately 2 mm after each day's measurement, while δ and D3 whiskers were left untrimmed. In rat N3, δ and D1 were trimmed 2–4 mm shorter each day. Whisker length was measured daily. Results are shown in Figure 5. The power spectrum for whisking in air on each day is presented as a color plot in each vertical strip. Whisker FRF was measured daily using the impulse method, and first and second harmonics of the FRF were calculated using a model of the whisker as a truncated (i.e., trimmed) conical beam, rather than an intact conical beam ([18]; see Materials and Methods). The FRF and first and second harmonics, calculated from the impulse measurements, are plotted as open circles, asterisks, and diamonds, respectively, on each day's power spectra.
Results showed that as trimming decreased whisker length, power spectra for whisking in air shifted systematically towards higher frequencies, as expected if resonance filtering shaped whisker vibrations. Trimming shifted the shoulders of the power spectrum (black filled circles) in parallel with the resonance frequencies (FRF and harmonics) measured by the impulse method. This was particularly evident for the D1 whisker in rat N2 and the δ whisker in rat N3, where bands of amplification (shoulders) in the power spectra closely followed the resonance frequencies measured by the impulse method. In contrast, power spectra remained stable for untrimmed whiskers, measured simultaneously in the same behavioral trials.
In a converse experiment (n = 1 rat), the D2 whisker was trimmed substantially, and then allowed to regrow by 12 mm over 14 d. Power spectra for whisking in air were measured before regrowth (when the whisker was trimmed) and afterwards. Results showed that regrowth was accompanied by a pronounced shift in the power spectrum of the D2 whisker towards lower frequencies, without substantial changes in the power spectra for nearby, simultaneously measured, untrimmed whiskers, whose length did not change appreciably during the regrowth period (unpublished data).
Together, these results indicate that resonance properties of whiskers shape high-frequency (>20 Hz) whisker vibrations during natural free whisking in air. This suggests that whisker resonance may be a relevant mechanism for filtering whisker input during active whisking in awake animals.
High-frequency whisker vibration in air is not due to head movement (Figure 1D and 1E), and therefore is likely to reflect high-frequency drive by whisker muscles. High-frequency muscular drive is plausible because high-frequency (83 Hz) electrical stimulation of motor axons in the facial nerve can cause whisker movements at stimulation frequency [19]. We observed high-frequency whisker vibrations in response to facial nerve stimulation in anesthetized rats, and found that evoked vibrations can strongly drive whisker resonance (Figure S1).
To test whether whisker muscles drive high-frequency whisker vibrations in awake, whisking rats, we measured electromyogram (EMG) activity from whisker muscles while imaging whisker motion in air (n = 5 rats). EMG was measured from intrinsic muscles and the extrinsic muscle m. nasolabialis, which drive whisker protraction and retraction, respectively, during whisking [20,21]. In different rats, EMG was measured from m. nasolabialis, intrinsic muscles, or both simultaneously, together with the movement of one or two different whiskers (Table 1). In total, four m. nasolabialis EMG recordings were obtained simultaneously with movement of seven whiskers, and three intrinsic EMG recordings were made simultaneously with movement of six whiskers.
EMG activity was coherent with whisking (Figure 6A), as previously reported [20,21], with intrinsic muscles generally active during protraction and m. nasolabialis active during retraction (unpublished data). The rectified, differential EMG (|∇EMG|) power spectra revealed high-frequency muscle activity up to 50 Hz (Figure 6B). To determine whether high-frequency muscle activity drove high-frequency whisker movement, we measured the spectral coherence between |∇EMG| and whisker position in air, during all types of whisker motion (whisking, erratic, and flat). For both intrinsic and extrinsic muscles, coherence between |∇EMG| and whisker motion was generally statistically significant, with values between 0.15 and 0.65, for frequencies less than 50 Hz, and fell below significance by approximately 50 Hz (Figure 6C and 6D). This was true for both arc 1 (D1 and C1) whiskers, and arc 2 (D2 and C2) whiskers. The maximal frequency of significant coherence, termed the cutoff frequency, was defined as the frequency at which coherence magnitude fell below the p = 0.05 significance level for nonzero coherence (see Materials and Methods for confidence interval calculation). For arc 1 whiskers, cutoff frequency was less than approximately 50 Hz for six of seven measurements, and approximately 90 Hz in the remaining measurement (Figure 6E). Thus, muscle activity was significantly, but only modestly, coherent with whisker motion at the FRF of arc 1 whiskers (median measured FRF: 36.9 Hz). Coherence at the FRF was even weaker for arc 2 whiskers, which also showed cutoff frequency of less than approximately 50 Hz in six of seven cases, and had a median FRF of 57.0 Hz (Figure 6E). We conclude that whisker muscles provide some high-frequency energy that could drive whisker vibrations, but because coherence was weak at high frequencies, how muscle contractions drive high-frequency vibrations remains unresolved.
The above results indicate that resonant motion is prominent during whisking in air, and that a map of resonance frequency exists across the whiskers. To determine whether this resonance map is used to encode surface texture, we explicitly tested the two central predictions of the resonance model for texture coding: first, that whiskers resonate at distinct, characteristic resonance frequencies as they sweep across surfaces; and second, that the amplitude of resonance frequency vibrations in each whisker depends on surface texture, resulting in one preferred texture that drives the strongest vibrations. Together, these properties have been proposed to result in a spatial map of texture across the whiskers [11,22].
We measured whisker motion on sandpaper surfaces in three rats performing the nose poke task (rats N4–N6) and two rats that whisked while head-fixed (rats H1–H2). We used seven sandpapers: P150 (roughest), P240, P400, P600, P800, P1200, and P1500 (finest). These correspond to 100-, 58-, 35-, 26-, 22-, 15-, and 13-μm mean particle size. Rats can readily distinguish two coarse sandpapers [3], a smooth surface from P100 sandpaper [10], and can distinguish 60-μm differences in spacing of periodic grooves [5], suggesting that differences between these sandpapers (or at least between the roughest and smoothest sandpapers) should be discriminable using the whiskers. Up to four sandpapers were presented per measurement session, typically in blocks of five to ten trials each. Different subsets of sandpapers were presented on different days. Surfaces were placed parallel to the whisker pad, 5 mm closer to the face than the whisker length. (Because whisker tips move in an arc, this meant that approximately 5 mm of whisker tip contacted the surface at mid-whisk, and less than 5 mm contacted at maximum protraction and retraction). Because whiskers are different lengths, we measured movement of only a single whisker at a time across the surfaces. We verified continuous whisker–surface contact during whisking in each animal, by observing the presence of a consistent whisker shadow on the CCD imaging array when the array was positioned 1 mm from the surface. It was not possible to position surfaces closer to the whisker tip, given the lateral freedom of head position within the nose poke (approximately 2 mm). Whisker motion was measured halfway between the surface and the whisker pad (∼10 mm from the follicle).
An example of whisker motion across a rough (P150) sandpaper is shown in Figure 7A. The rat initially retracted the D3 whisker across the surface (negative slope in the position trace), and then protracted it (positive slope). Whisker velocity and acceleration, calculated from the position trace, revealed approximately three brief, high-acceleration, high-velocity events that occurred during whisker motion. To analyze the time-varying spectral content of whisking on the surface, we calculated the Wigner-Ville time-frequency representation (TFR), qualitatively similar to a spectrogram, for this whisker motion (Figure 7B). The TFR showed prominent, brief epochs of vibration at approximately 150–180 Hz, aligned with the rapid movement events. The integrated TFR across the entire whisking period (which is equal to the average power spectrum) revealed a broad peak at approximately 150–180 Hz (Figure 7B, rightmost trace). Similar broad peaks at 50–180 Hz were observed for D1, D2, and D3 whiskers moving across a variety of surfaces (see below).
We tested whether these broad, high-frequency peaks were consistent with whisker resonance by comparing the peak frequencies across different length whiskers. (The FRF while the whisker is pinned against a texture will not equal the FRF measured in air, because boundary conditions for vibration are changed and the whisker is effectively shortened [11,23,24]). Figure 7C and 7D show power spectra for vibrations of the D1 and D3 whiskers of rat N4, measured during palpation on five different sandpapers. Power spectra were calculated as integrated TFRs for all individual protraction and retraction epochs on a given surface, and then averaged across these epochs to obtain the average power spectrum for each surface. Consistent with the resonance model, the D1 whisker showed a high-frequency vibration peak at approximately 80 Hz, while the shorter D3 whisker showed a peak of approximately 150 Hz (Figure 7C and 7D). We repeated this analysis for ten whiskers (five D1 whiskers, four D2 whiskers, and one D3 whisker) in five rats (N4–N6, H1, and H2). We calculated the high-frequency peak of the average power spectrum for each whisker moving on each texture (identified as the first peak in the power spectrum >40 Hz).
Across all textures, the high-frequency peak for D1 whiskers was found to be between 57.7 and 91.2 Hz (mean: 71.8 Hz); for D2 whiskers, 68.5–114.0 Hz (mean: 86.6 Hz); for the single D3 whisker, 142.2–153.9 Hz (mean: 147.9 Hz) (Figure 8A). Thus, for both animals performing the nose poke behavior (open circles, Figure 8A and 8B) and head-fixed animals (asterisks, Figure 8A and 8B), measured peaks in vibration power spectra were at higher frequencies for the shorter whiskers and lower frequencies for the longer whiskers, consistent with intrinsic resonant properties of the whiskers. These data therefore suggest that whiskers vibrate at characteristic resonance frequencies when moving across surfaces, at least when distance to the surface is kept constant. Subsequent analyses assume that the high-frequency vibration peak represented the whisker's resonance frequency on surfaces. Power at the high-frequency peak during whisking on textures was 6.7 ± 3.4 (mean ± standard error) times greater than power at the resonance frequency during whisking in air (unpublished data).
Finally, we tested whether the amplitude of resonance frequency vibrations in each whisker depends on, and encodes, surface texture, as posited by the resonance hypothesis [11,13]. In isolated whiskers and anesthetized animals, prolonged, stable application of different texture or vibratory stimuli to the tip of a single whisker generates up to a 10-fold difference in steady-state power at the whisker's resonance frequency, indicating strong tuning for specific textures or vibration frequencies [11,22]. In contrast, we found that during natural whisking, the power spectrum for whisker vibrations in a single whisker was remarkably constant across different surfaces (e.g., the five sandpapers in Figure 7C and 7D). We calculated the power at the presumed resonance frequency as a function of texture for all sandpapers that were presented to each animal. Across the ten whiskers (rats N4–6, H1, and H2), no substantial or systematic relationship between sandpaper grade and vibration power at presumed resonance frequency was observed, either for nose poke or head-fixed rats (Figure 8B). On average, the maximal change in power at the presumed resonance frequency between any two textures for individual whiskers was 49 ± 29% (mean ± s.d.). This is substantially less than the 10-fold variation observed with prolonged, regular stimulation in anesthetized animals and detached whiskers. Two-way ANOVA found no significant differences in power at the presumed resonance frequencies in each behavioral trial for either whisker type (D1, D2, or D3) (F(2,1624) = 0.45; p = 0.64) or sandpaper grade (P150 through P1500) (F(6,1624) = 1.18; p = 0.32). Similar analysis of normalized power at the presumed resonance frequency (normalized to total spectral power, which controls for trial-to-trial variability in total vibration power) produced identical results (unpublished data).
These results indicate that during active whisking under our experimental conditions, whiskers resonate on textures, but resonance magnitude is independent of texture roughness. This is contrary to the expectation of the resonance model for texture coding, which predicts that sustained whisker-tip movement over texture spatial features leads to regular whisker vibrations whose amplitude builds up most effectively when vibration frequency matches whisker resonance frequency [25]. One potential explanation of the present result is that resonance frequency vibrations do not build up in a gradual, sustained manner during natural whisking, but represent transient responses (ringing) to discrete high-acceleration, high-velocity events. Such events were a prominent feature of whisker movement across surfaces (e.g., Figure 7A and 7B), and were commonly associated with transient high-frequency ringing in whisker position, acceleration, and velocity. Representative examples of this behavior measured during protraction of D1 and D3 whiskers on a P150 sandpaper are shown in Figure 9A (see also Figure 7A and 7B). TFRs of these representative events revealed postevent ringing of the D1 whisker at approximately 90 Hz, and postevent ringing of the shorter D3 whisker at approximately 175 Hz (Figure 9B).
To determine whether transient ringing induced by these discrete motion events was a significant source of overall resonance vibrations during texture palpation, we compared vibration power spectra in 0.4-s epochs centered on high-acceleration events (defined here as movement events in which acceleration magnitude exceeded mean acceleration by 2 s.d.) versus equivalent epochs of whisker retraction or protraction when no high-acceleration event occurred. Results showed that power at the high-frequency peak was, on average, 12.8 ± 2.6 times greater in epochs containing high-acceleration events versus epochs that lacked such events, and 3.8 ± 0.8 times greater versus all whisking epochs, regardless of whether they contained an acceleration event (n = 10 whiskers, 5 animals). This result is shown for D1 and D3 whiskers of rat N1 in Figure 9C, and for all whiskers in Figure 9D. Together, these results demonstrate that resonance vibrations in whiskers during texture palpation primarily represent transient ringing following discrete high-acceleration movement events, and that the amplitude of resonance vibrations does not vary across the range of sandpapers that were tested.
Discrete high-acceleration motion events were prominent on textures, but were generally absent during whisking in air. Representative whisker motion in air and on a rough (P150) sandpaper are shown in Figure 10A. In this example, large-acceleration events (acceleration > 4 s.d. in air; green dots) occurred 2.5-fold more often on the sandpaper than in air. Acceleration events of all magnitudes occurred more frequently on the texture versus air for this whisker (Figure 10B, texture: 315 trials, 340 s of whisker-movement data; air: 111 trials, 223 s), and the highest acceleration events (>0.3 mm/ms2) were detected predominantly during whisking on texture (inset). Although we use acceleration as a convenient marker for these motion events, whisker acceleration and velocity were well correlated in whisker-motion traces (Figure S2).
High-acceleration events occurred during both protraction and retraction, and could be classified into slips (events in which whisker speed suddenly increased in the direction of whisker motion) and sticks (events in which speed suddenly decreased, corresponding to sudden stopping of whisker movement). Examples of slips and sticks during protraction and retraction are shown in Figure 10C. The average kinematics of slips and sticks during protraction and retraction are shown in Figure 10D, for the D2 whisker in rat H1 moving across four sandpapers. For each type of event, separate averages were calculated for five ranges of acceleration magnitude. (High-acceleration events correspond to more abrupt slips and sticks.) The average position and acceleration traces revealed that whisker slips were followed, on average, by sticks, and sticks were preceded by slips. Thus, sequences of high-acceleration events represented slip-stick motion of whiskers along surfaces.
Slips occurred during all phases of protraction and retraction (Figure 10E). To determine the average size and time course of a slip, we compiled histograms of slip magnitudes, durations, and peak speed (|velocity|), for all rats and all whiskers (n = 10 whiskers, 5 rats), including all slip events with acceleration greater than 2 s.d. of the acceleration in air (Figure 10F). Slip duration was defined from the initial acceleration peak to the time when whisker speed returned to the average speed. An example of the calculation of slip magnitude and duration is shown in Figure 10F (upper left). Results showed that during the average slip event, the whisker traveled a mean of 1.9 mm, in a mean of 8.6 ms, and achieved a peak speed of 0.33 ± 0.24 mm/ms, before whisker speed returned to average.
We tested whether slip-stick events could provide an alternate, nonresonance-based code for surface texture. A slip-stick code is plausible since sharp, high-acceleration, and high-velocity events effectively drive spikes in somatosensory cortex [7,26,27], and thus the pattern of slip-stick events is likely to be encoded in the rat's central nervous system (CNS). We again used acceleration to identify these events. We compared acceleration events on four sandpaper textures (P150 [very rough], P400, P800, and P1200 [very smooth]) that were interleaved in blocks for each rat within a single day (five or ten trials per block). This measurement was performed for the D1 and D2 whiskers in three rats (N6: 89–103 trials per texture, H1: 52–56 trials per texture, and H2: 40–43 trials per texture). Analysis was restricted to within-day comparisons across textures to avoid complications from day-to-day variability in whisking behavior. For this analysis, an acceleration event was defined as any acceleration peak that crossed a defined threshold, with a minimum of 2 ms between events, and stick versus slip events were not distinguished.
Motion of the D2 whisker in rat H2 across a smooth (P1200) and rough (P150) sandpaper is shown in Figure 11A and 11B. (This is the same whisker whose motion in air and on P150 sandpaper was shown in Figure 10A.) Low-acceleration events (red dots, peak acceleration 0.062–0.248 mm/ms2, corresponding to 1–4 s.d. above zero on the P1200 surface) occurred on both textures, as well as in air. In contrast, high-acceleration events (green dots, >0.496 mm/ms2, corresponding to 8 s.d. above zero on the P1200 surface) occurred preferentially on the rough P150 sandpaper. This suggested that high-acceleration events may occur systematically more frequently on rougher surfaces.
We calculated the average incidence of different magnitude acceleration events on P150, P400, P800, and P1200 textures, as well as during whisking in air, for six whiskers in three rats (rat N6 performing the nose poke task, and rats H1 and H2 whisking while head-fixed; D1 and D2 whisker motion was measured in each animal) (Figure 11C and 11D). The number of acceleration events surpassing different absolute acceleration thresholds was calculated per sweep, where a sweep was defined as a single whisker protraction or retraction. Results showed that the total number of acceleration events surpassing low acceleration thresholds (e.g., 0.1 mm/ms2) was not different between whisking in air and whisking on surfaces, but the number of events surpassing high acceleration thresholds (e.g., 0.4 mm/ms2) was higher on surfaces than in air, and was systematically higher on rougher versus smoother surfaces (Figure 11C and 11D). Statistical analysis showed that low-acceleration events (with peak amplitude in the range 0.062–0.248 mm/ms2, corresponding to 1–4 s.d. above zero) were equally prevalent in air and on smooth P1200 and P800 surfaces, but were significantly less prevalent (asterisks; Mann-Whitney U-test, p < 0.01) on the rougher P400 and P150 surfaces, especially for the D2 whisker (Figure 11E). Conversely, high-acceleration events (>0.496 mm/ms2, corresponding to 8 s.d. above zero) were systematically more prevalent on rougher versus smoother surfaces, for both D1 and D2 whiskers (Figure 11F). As a result, the ratio of high to low acceleration events per sweep increased systematically and significantly with surface roughness (Figure 11G; asterisks indicate significant differences in ratio between pairs of textures). These relationships between slip-acceleration magnitude/frequency and surface roughness held true for both the nose poke rat (N6) and head-fixed whisking rats (H1 and H2) (unpublished data). These results suggest that either the frequency of high-acceleration events or the relative frequency of high to low acceleration events may contribute to a kinetic signature for surface roughness [7], independent of whisker resonance.
Sensory systems generate signals by physical interaction between sensory organs and the external environment, and the form of this interaction determines how features of the sensory environment are encoded. In the whisker system, active movement of the whiskers and whisker mechanical properties critically determine this interaction, and therefore influence neural coding [6,28]. We attempted to distinguish between two major models of whisker texture coding—the resonance hypothesis [11–13,22] and the kinetic signature hypothesis [7,10,29]. These hypotheses assume different physical interactions between whiskers and objects, different patterns of surface-induced whisker vibration, and different neural coding strategies [6]. In the resonance hypothesis, each whisker is tuned to resonate most strongly in response to a specific range of textures (those textures that drive tip vibration at the whisker's intrinsic resonance frequency). Because resonance frequency varies with whisker length, texture information is encoded spatially by the relative amplitude of resonance vibrations across whiskers, and in somatosensory cortex (S1) by relative firing rates of neurons across whisker columns [11–13,22]. In the kinetic signature hypothesis, mechanical resonance plays no special role in coding. Instead, textures generate unique, identifiable motion patterns in single whiskers, and texture information is encoded in the brain by neuronal spiking that tracks features of these patterns, including mean speed of whisker vibration [10] and irregular whisker-velocity patterns [7], which vary with texture [7].
We tested these hypotheses by measuring the physical vibrations induced in whiskers as rats actively whisked in air and across textured surfaces. Whiskers were found to resonate in air and on textures, but the amplitude of resonance vibration was equal across a wide range of textures. Thus the resonance hypothesis of texture coding is not correct, at least for the behavioral conditions and range of textures tested here. Instead, we found whiskers exhibited discrete, high-acceleration, high-velocity slip-stick events on surfaces that drove transient ringing in the whiskers. We propose that slip-stick (or slip-stick-ring) events are fundamental elements of natural whisker–surface interaction. Because the rate and magnitude of slip-stick events were correlated with texture, we propose that slip-stick events may contribute to a unique kinetic signature for textures in individual whiskers.
We found that whiskers exhibited high-frequency (>20 Hz) vibrations during active whisking in air, and that the spectral composition of these vibrations varied with whisker length, due to filtering by whisker resonance (Figures 2–5). Thus, whiskers resonate during natural whisking in air, and a map of whisker resonance exists in awake, whisking rats (Figure 4). High-frequency vibrations were coherent across neighboring whiskers, were not caused by head motion or interactions between whisker and air (Figure 2C–2E), and could be elicited by high-frequency stimulation of the facial nerve in anesthetized animals (Figure S1). This suggests that vibrations are due to neurally or mechanically coordinated drive of neighboring whiskers by whisker facial muscles. EMG recordings of extrinsic and intrinsic muscles detected high-frequency components of muscle contraction. However, when we measured spectral coherence between whisker vibrations and EMG activity, we found only modest coherence for frequencies up to approximately 50 Hz (near the FRF of arc 1 whiskers), and nonsignificant coherence for frequencies greater than 50 Hz (near the FRF of arc 2 and shorter whiskers) (Figure 6). This suggests either that (1) additional coherent, high-frequency muscular drive exists, but was not detected by the EMG recordings, or that (2) muscles drive resonance vibrations noncoherently, as could occur if sharp, pulsatile muscle contractions induced higher frequency vibrations and excited whisker ringing at the resonance frequency. This latter case is less likely because whisker motion, and muscle drive, are relatively smooth during exploratory whisking. However, sharp contractions may occur during more erratic whisker motion.
Resonance vibrations also occurred during active whisking on sandpaper surfaces, as inferred from the presence of spectral peaks in whisker vibration at specific supra-whisking frequencies, with longer whiskers vibrating at low frequencies, and shorter whiskers vibrating at higher frequencies (Figures 7 and 8). Thus, resonance filters whisker vibrations during whisking onto surfaces. However, resonance vibrations occurred primarily as transient, sporadic ringing events, rather than as sustained oscillation, and neither the amplitude of vibrations at presumed resonance frequencies nor the overall power spectrum varied with texture across a wide range of sandpaper grades (Figures 7 and 8). Thus, each whisker was not preferentially excited by a specific set of textures. We conclude that differences between sandpaper textures are not encoded by relative vibration amplitude across facial whiskers, at least in the geometrical and behavioral conditions of our study. These data argue against the resonance hypothesis for texture coding. However, they do demonstrate that whisker resonance occurs during surface palpation, and therefore may play a role in amplifying some types of whisker responses [13]. These results confirm a recent study that detected resonance vibrations on textured surfaces, but did not examine whether resonance encoded texture [17].
The critical difference between our results and the resonance hypothesis appears to be in how resonance vibrations are generated during whisker–surface interaction. Linear resonating systems can resonate in two distinct modes: In the transient mode, oscillations are triggered by discrete external impulses, and occur transiently after these impulses, in the absence of additional external vibratory forces. In this case, oscillation dynamics are determined solely by the intrinsic properties of the system, as in the case of transient resonant ringing of a tuning fork after being struck by an object. In the steady-state mode, in contrast, vibrations are produced in an ongoing manner during sustained external vibratory drive. In this case, vibratory responses occur at the same frequency as the external impulses, and vibration amplitude is much larger when external vibrations occur at the intrinsic resonance frequency of the system. The resonance hypothesis assumes that passage of a whisker over a surface generates sustained tip vibrations as the whisker interacts with surface microfeatures, and that this causes steady-state resonance to build up in the whisker. Such steady-state resonance indeed occurs when sustained vibrations are applied to isolated whiskers or to nonmoving whiskers in anesthetized animals [11,12,22]. However, our results demonstrate that voluntary whisker motion produces discrete, high-acceleration slip-stick events, rather than smooth motion across surfaces (Figure 10). These slip-stick events drive transient ringing, and this transient ringing is the major source of whisker resonance on surfaces (Figure 9). The dominance of transient resonance, as opposed to sustained resonance, explains why whisker vibrations vary with intrinsic properties of the whiskers (Figure 8A), but not with surface texture (Figure 8B). These results confirm a previous observation that sustained resonance vibrations do not appear during voluntary whisking on surfaces [12].
Together, these data indicate that whisker resonance occurs in awake animals, both during whisking in air and on surfaces, and may contribute to encoding or amplification of certain aspects of whisker input. However, differential whisker resonance does not encode texture in these behavioral conditions and using these sandpaper surfaces, which are predicted to be discriminable by rats [4,5,10]. We cannot rule out that, under conditions of behavioral discrimination, rats may adopt a different whisker exploration strategy that may enable resonance-based coding of texture. However, recent studies of texture discrimination have provided no evidence for coding by resonance [10,17], and two arguments suggest that such a coding strategy may be problematic: first, the relationship between whisker resonance frequency and effective whisker length (Figure 8A) suggests that any trial-to-trial variation in surface position or angle relative to the face will alter whisker resonance frequency, making it difficult to construct a position-independent resonance code for texture. Second, rats discriminate textures even with substantial trial-to-trial variation in whisking speed [5]. Such variation will alter the relationship between texture spatial frequency and whisker-tip vibration frequency, making it unlikely that a whisker could be “tuned” for a specific texture.
A common feature of whisker motion across sandpapers were discrete, high-acceleration slip and stick events (Figure 10). Slip and stick events occurred during all phases of whisker protraction and retraction (Figure 10E). These events often generated high-amplitude transient ringing at the whisker's resonance frequency (Figures 7A and 9). Slip-stick events were frequent: for example, 1.2 events with acceleration greater than 0.4 mm/ms2 occurred per protraction–retraction cycle for the D1 whisker, averaged across all sandpaper surfaces (Figure 11C). This corresponds to approximately 30 events when all 25 large whiskers on each side of the face are considered. These events have also been observed during whisking onto surfaces under very different geometrical and behavioral conditions [17], and are therefore likely to be basic common elements of the whisker input stream.
The average slip was 1.9 mm (measured at the whisker midpoint, ∼10 mm from the follicle), and lasted 8.6 ms before whisker velocity returned to its mean value (Figure 10F). This corresponds to a mean angular displacement of 10° and a mean velocity of 1,100°/s during slips. Peak velocity during slips was 0.33 mm/ms. This amplitude and velocity are well within the range of behavioral detectability [30] and spike encoding at primary afferent and cortical levels [26,27,30]. Thus, slip-stick events are likely to be encoded in the CNS. These slip-stick events are similar to velocity transients observed during artificial whisking onto textures in anesthetized rats [7,15]. Because high-acceleration events occur more frequently on textures than in air (Figure 11), we propose that slip-stick events may encode the presence of a surface, or surface properties, on the whisker array.
The kinetic signature hypothesis for texture coding proposes that textures generate unique, identifiable temporal patterns of whisker vibration (“kinetic signatures”) in single whiskers, and that these temporal features are encoded in neural spike trains. Candidate components of kinetic signatures for texture include the spectral composition of whisker vibration [15], the mean speed of whisker vibration [7,29], and the temporal profile of velocity transients [7]. These features vary when whiskers of anesthetized rats are artificially swept across different textures by electrical stimulation of the facial motor nerve, with rough versus perfectly smooth textures generating differences in mean vibration speed [7,29], and finer texture differences (e.g., between sandpaper grades) generating unique temporal profiles of whisker velocity [7].
Our data suggest that slip-stick events may contribute to the kinetic signature for texture. The magnitude and frequency of these events were correlated with texture, with rougher sandpapers eliciting a greater frequency of high-acceleration events (which tend to also be high-velocity events), and a higher proportion of high-acceleration versus low-acceleration events, compared to smoother sandpapers and to air (Figure 11). This relationship between slip acceleration and texture is expected from a simple model in which rougher surfaces, which have greater friction, require more forward force during whisker protraction (or retraction) to overcome static friction and move the whisker tip forward (or back). This increased forward force translates into increased acceleration during forward slips. Thus, more high-acceleration slips, and fewer low-acceleration slips, are predicted on rougher textures. This significantly extends a prior study showing more high-speed slip events on a rough surface versus a completely smooth one [17].
We propose that slip magnitude (acceleration or velocity) and frequency are components of the kinetic signature for texture in the whiskers, and that coding of these parameters by S1 neurons provides information about surface texture. In anesthetized animals, whisker deflections evoke phasic, single-spike responses in S1 neurons, with spiking probability positively correlated with whisker velocity and acceleration over the ranges of 0.02–1.0 mm/ms [31,32] and approximately 20–500 m/s2 [33], respectively. The range of slip speeds and accelerations observed here (∼0.1–0.5 mm/ms and ∼100–1,000 m/s2) fall within this dynamic range. Thus, the occurrence and magnitude of slips are likely to be encoded by time-locked spikes in S1 ensembles, with texture-related sequences of slip-stick events (Figure 10A) encoded by temporal sequences of spikes (constrained by the intrinsic dynamics of whisker circuits and synapses). The occurrence of discrete slip events related to texture, observed here under two behavioral conditions, suggests a potential temporal spike code for texture during awake, active sensation. Such a temporal code has been suggested from S1 recordings in anesthetized rats during electrically evoked whisking on texturally similar surfaces, like the sandpapers used here [7]. In contrast, active whisking onto very distinct textures (rough vs. smooth glass) evokes subtly, but significantly different, mean firing rates in S1 [10]. Slip-evoked spikes could drive such texture-specific changes in firing rate, depending on neural sensitivity to slip amplitude and velocity.
Ten rats were used in this study. All procedures were approved by the University of California San Diego (UCSD) Institutional Animal Care and Use Committee and followed Society for Neuroscience guidelines for research.
Two types of whisker behavior were studied. In Behavior 1 (whisking in nose poke, six rats), rats were trained using operant conditioning techniques to place their nose in a small port (the nose poke) and whisk for approximately 1 s in air or on textured surfaces. The behavioral apparatus, modeled after [34], consisted of an outer reward chamber containing a solenoid-gated drink port, and an inner measurement chamber containing the nose poke, texture stimuli, and whisker-motion recording system (Figure 1A). Rats received water (50 μl) as reward during behavioral training (1 h per day) and during a 1-h ad lib drinking period following each behavioral training session, but not during the remaining 22 h per day, 5 d per week. Water was freely available on weekends. Rats on this regimen were healthy and alert, and gained weight daily.
Rats (age 30 d) were initially accommodated to handling (3–5 d) and to the behavioral apparatus. Rats were then trained to drink from the drink port in response to a white noise tone (WNT). A phototransistor in the drink port signaled the rat's presence and gated water delivery. Next, rats were trained to nose poke to trigger the WNT and water delivery to the drink port. A phototransistor in the nose poke reported nose poke occupancy. Finally, rats were trained to gradually increase nose poke duration and to actively whisk while in the nose poke. Gross whisking was assessed by four phototransistors that generated voltage pulses when the whiskers passed over them. The number of phototransistor pulses required to trigger the WNT and drink port water delivery was gradually increased until rats were whisking in the nose poke for approximately 0.5 s. Each approximately 0.5-s bout of whisking in the nose poke was considered a trial, and trials were separated by the rat retreating to the reward chamber to drink. Trained rats performed 80–150 trials per day. Total training time (after accommodation) was approximately 23 d.
Whisker motion was recorded optically in trained rats whisking in air and whisking onto textures. Textures were 6 × 6-cm sandpapers of grade P150, P240, P400, P600, P800, P1200, and P1500 glued to an aluminum plate and positioned in the whisking path of the right whiskers 5 mm from the whisker tips, parallel to the face. Up to four different textures were mounted on a four-arm Plexiglas holder attached to a stepper motor (Oriental Motor, PK264B1A-SG10). Textures were rotated into place between trials while the rat was at the drink port. Surface positioning relative to the nose poke was performed as follows: first, using videography, we measured the mean position and orientation of the external edge of the whisker pad while the rat was performing the whisking behavior. Surface orientation was set parallel to the whisker pad. Next, we transiently anesthetized the rat and measured the length of the whisker to be studied (whisker movement on surfaces was measured for a single whisker at a time). We positioned the stepper motor so that the point on the surface closest to the face (i.e., the point at the intersection of the surface and of the whisker, when the whisker was normal to the face) was located 5 mm closer to the whisker pad than the whisker length. Surface positioning was verified by imaging the whisker 0.5 mm from the surface, and confirming that the whisker shadow disappeared from the imaging plane when the surface was moved approximately 5 mm from its set position. Whisking in air was measured by rotating the stepper motor into a position with no texture present. Thus, up to four textures (or three textures plus air) could be interleaved under computer control during a recording session. Training and recording procedures were controlled by custom routines in Labview (National Instruments).
In Behavior 2 (whisking while head-fixed, four rats), rats (age 30–40 d) were accommodated to handling (∼1 wk), and to being placed for 15 min in a loose fabric sack from which the head emerged (∼1 wk) [35]. Rats were habituated to being placed, while in the sack, in a 5-cm–diameter Plexiglas tube (Figure 1B). Rats then underwent surgery to implant electromyogram (EMG) recording electrodes (see below), during which a small screw was affixed to the skull with dental acrylic. After 4–6-d recovery from surgery, rats were placed again in the Plexiglas tube, and the head was stabilized via the screw (Figure 1B). Head-fixed rats naturally whisked in response to objects held in front of them. Whisker motion was measured during these whisking epochs. Recording sessions typically lasted 15–30 min. Whisker motion was recorded in 3-s trials with approximately 50–100 trials per recording session. The animal was positioned so that the head and whiskers were in the same spatial relationship to the textures and CCD imaging array as in Behavior 1. Textures (or air) were presented in blocks.
For both Behaviors 1 and 2, behavioral training was performed with all whiskers intact. The day before whisker-motion measurement, rats were transiently anesthetized with isoflurane, and all whiskers whose motion was not being studied were trimmed at the base. For Behavior 1, all but one to four whiskers (δ, D1, D2, and D3) were trimmed. For Behavior 2, all but two to three whiskers in the C row or D row were trimmed.
Whisker motion was measured in one dimension by casting shadows of the whiskers onto a linear CCD imaging array. The light source was a diode laser (670 nm), positioned above the rat and focused into a collimated line 60-mm long and 1-mm wide, using two cylindrical lenses rotated 90° from one another (Figure 1C). Below the whiskers, a third cylindrical lens focused whisker shadows onto the linear CCD array (Fairchild imaging, CCD 133AEDC, 1,060 elements, 13-μm width per element). The output of every other CCD element was sampled at 4-kHz frame rate using custom-built electronics (UCSD Physics electronics shop) and a National Instruments data acquisition card (PCI 6111). Voltage traces from the array were stored and processed offline to determine whisker position. In Behavior 1, whisker position was recorded for 1.5 s starting with nose poke onset (analysis was restricted to the epoch during which the rat remained in the nose poke). In Behavior 2, whisker motion was recorded in 3-s blocks.
The CCD array was positioned parallel to the whisker pad, either approximately 10 mm (range: 6–14 mm) from the whisker pad (whisking in air) or at the midpoint between the texture and the whisker pad (whisking on texture). Whisker contact with textures was verified for each rat by the consistent presence of whisker shadows when the array was positioned 1 mm from the texture. Each frame of CCD output was subtracted from a baseline CCD image obtained when no whiskers were present (baseline images were obtained several times during each recording session). Whisker shadows appeared as discrete voltage peaks in the baseline-subtracted CCD image, with each shadow covering eight to ten CCD pixels. Position of each whisker shadow was calculated as the weighted mean of all pixels in the whisker shadow, weighted by pixel voltage. Whisker spatial position was calculated from whisker-shadow pixel position via a calibration curve obtained using a 0.5-mm spaced wire grid held at whisker position. Repeated measurements showed that whisker position was determined with a spatial resolution of approximately 5 μm.
Whisker motion over time was computed algorithmically using custom software in Matlab. Up to four whisker shadows could be tracked simultaneously and identified unambiguously using this method. All but the imaged whiskers were trimmed weekly to the level of the skin, during transient isoflurane anesthesia (4% in 2 l/min O2, delivered via a nose cone). If a whisker transiently left the imaging plane, whisker motion was only analyzed up to that point.
To measure whisker FRF using the impulse method, rats were anesthetized with isoflurane and the head positioned in the behavioral apparatus at the standard position and angle relative to the CCD array. An impulse was delivered manually to each whisker, and the resulting decaying oscillation was measured with the CCD array. The FRF was calculated as the inverse of the average time between peaks in the oscillations [11]. For untrimmed whiskers, the first and second harmonics of the resonance frequency were calculated as: f1 = (10.6/4.4)FRF and f2 = (19.2/4.4)FRF, as predicted by theoretical models of tapered beams [23]. For trimmed whiskers, which are truncated tapered beams, the ratios f1/FRF and f2/FRF are functions of the truncated length. We calculated f1 and f2 for trimmed whiskers by interpolating f1/FRF and f2/FRF ratios that were numerically calculated by Conway et al. [18] for four different ratios of truncated to untruncated length of a thin conical beam.
Whisker length was measured while rats were anesthetized. Length was measured from the skin surface to the whisker tip, using calipers and 4× magnification under a dissecting microscope.
In some experiments, EMG activity was recorded from whisker pad muscles. EMG electrode implantation followed Berg and Kleinfeld [21]. Briefly, surgery was performed using sterile technique, under ketamine/xylazine anesthesia (90 and 10 mg/kg, respectively, i.p.). Supplemental ketamine (20 mg/kg) was administered approximately every 2 h to maintain anesthetic depth, determined by absence of limb withdrawal reflex and breathing rate of 45–60 breaths per min. EMG electrodes were made from Teflon-coated tungsten microwire (0.002” diameter; California Fine Wire; 1 mm of insulation stripped at recording tip). Microwires were implanted in pairs to record the differential EMG signal. Microwires were implanted via a midline incision at the top of the skull, and a lateral incision caudal to the mystacial pad. One electrode pair was implanted in the extrinsic muscle m. nasolabialis, by exposing this muscle and pressing the recording tips into muscle tissue. Microwires were secured at the muscle entry point using 6–0 Ethicon nylon sutures (Johnson and Johnson). To record EMG in intrinsic muscles, microwire pairs were threaded through a 26-ga targeting needle, which was used to insert wire tips into the whisker pad [21]. Wire tips were bent back at the needle tip to anchor the wires to whisker pad tissue. Wires were sutured in place where they exited the pad. Reference wires (stripped of 4 mm of insulation) were implanted in the dermis at the tip of the snout, rostral of m. transversus nasi. Microwire tip position was verified at the end of surgery by passing current to stimulate the muscles and evoke appropriate whisker and pad movements. Microwires were soldered into a ten-pin connector (Samtec) attached to the skull. Bupivicaine (0.1 ml) was administered for postoperative analgesia.
EMG recording commenced 4–6 d after EMG implantation. EMG data were collected in behaving rats in 3-s–long blocks, simultaneous with whisker-motion data. EMG signals were amplified (20× gain) and impedance buffered using an eight-channel head-mounted headstage amplifier (Plexon Instruments HST/8o50-G20). Headstage output was transmitted via twisted thin-gauge wires to a second amplifier and bandpass filter (Plexon Instruments PBX2/16sp-G50) (50× gain, 0.3–8 kHz bandpass). Amplifier output was digitized at 32 kHz (National Instruments PCI 6259). Analysis was performed on rectified, low-pass filtered (500 Hz cutoff) difference of neighboring raw EMG signals, denoted |∇EMG|. EMG and whisker data acquisition were performed on separate, synchronized data acquisition cards.
Power spectra of whisker motion in air were calculated using the multitaper estimation technique of Thomson (1982) [36]. Briefly, a whisker motion time series recorded during a single trial was first multiplied by a set of K orthogonal data tapers. The Fourier transform of each tapered time series was calculated using the Fast Fourier Transform algorithm in Matlab, and from each Fourier Transform, the power spectrum was estimated as the modulus squared of the Fourier Transform. The estimated power spectrum of the whisker motion for the nth trial,
Z, was then an average over these K power spectra
where Yn,k (f) is the Fourier Transform of the kth tapered time series of the nth trial. The average power spectrum of whisker motion for a single day's recording session was then an average over all trials performed on that day:
where SY (f) is the average power spectrum of the whisker motion and N is the total number of trials. Here, N was typically between 50 and 150 trials, and the number of tapers, K, was 5. With this method, the resulting average power spectrum of a time series of duration T is smoothed over a half-bandwidth of
The spectral coherence C(f) between the |∇EMG| and whisker motion was calculated similarly to [37],
where Z (f) is the Fourier Transform of the |∇EMG| time series and SZ (f) is the average |∇EMG| power spectrum. The theoretical confidence intervals for coherence were computed following [38], where it is estimated that the coherence magnitude will exceed
in P × 100% of measurements. Here, we take p = 0.05. Spectral estimation was performed using the Chronux (http://www.chronux.org) and signal processing toolboxes in Matlab.
The Fourier methods described above are appropriate for describing the average spectral properties of stationary signals [38]. We used the Wigner-Ville TFR to examine the brief transient ringing events during whisking onto textures. The TFR is known to provide good localization in both time and frequency, and is better suited for analyzing time series with time-varying frequencies [39]. This distribution is computed by correlating the entire time series y(t), with a time-translated version of itself and taking the Fourier transform of this locally autocorrelated function,
The color plots of the TFR were smoothed over time (windowsize = 2.5 ms) and frequency (windowsize = 20 Hz) for visualization. Power spectra of whisker motion onto textures were calculated by numerically integrating the unsmoothed TFR(t,ω) over time [39].
In artificial whisking experiments [40], an incision was made in the side of the snout posterior to the whisker pad of the anesthetized animal (urethane 1.5 g/kg). The buccal motor nerve was separated from the underlying muscle and cut to prevent antidromic activation of the motor nerve [19]. The lower branch of the buccal nerve was also cut, which generated more-natural, horizontal whisks than with this nerve intact. The distal portion of the facial nerve was isolated in a stimulating cuff with electrodes placed around the nerve. Saline was applied to keep the nerve moist. The nerve was stimulated with brief monophasic pulses (50-μs duration, 3–6 V) from a Grass Stimulator (Model S88K). Pulses were generated at 110 Hz in bursts lasting 50 ms followed by 50 ms with no stimulation. Whiskers protracted during the bursts and passively retracted during the 50 ms following the bursts, generating 10-Hz artificial whisking.
|
10.1371/journal.pgen.1007925 | Strand break-induced replication fork collapse leads to C-circles, C-overhangs and telomeric recombination | Telomerase-independent ALT (alternative lengthening of telomeres) cells are characterized by high frequency of telomeric homologous recombination (HR), C-rich extrachromosomal circles (C-circles) and C-rich terminal 5' overhangs (C-overhangs). However, underlying mechanism is poorly understood. Here, we show that both C-circle and C-overhang form when replication fork collapse is induced by strand break at telomeres. We find that endogenous DNA break predominantly occur on C-rich strand of telomeres in ALT cells, resulting in high frequency of replication fork collapse. While collapsed forks could be rescued by replication fork regression leading to telomeric homologous recombination, those unresolved are converted to C-circles and C-overhang at lagging and leading synthesized strand, respectively. Meanwhile, multiple hallmarks of ALT are provoked, suggesting that strand break-induced replication stress underlies ALT. These findings provide a molecular basis underlying telomeric HR and biogenesis of C-circle and C-overhang, thus implicating the specific mechanism to resolve strand break-induced replication defect at telomeres in ALT cells.
| 10 to 15% human cancers utilize telomerase-independent alternative lengthening of telomeres (ALT) to maintain their telomere length. Unexpectedly, we find that endogenous C-strand breaks predominantly exist in telomeres of ALT cells, which induce high frequency of replication fork collapse. While collapsed fork triggers fork regression machinery to restart the replication, leading to telomeric homologous recombination; those unresolved are converted to C-circle and C-overhang. These findings suggest that the formation of C-circle and C-overhang represents a unique manner for ALT cells to prevent chromosome instability induced by replication defect at telomeres. Moreover, multiple hallmarks of ALT are provoked during this process, demonstrating that DNA strand break at telomeres underlies ALT mechanism.
| Linear chromosome ends are capped by telomeres, which are composed of TTAGGG/CCCTAA tandem DNA repeats and a protein complex called shelterin [1–3]. Because of end replication problem [4] and possible DNA resection by Exo I (Exonuclease I) and Apollo to form single-stranded overhang [5, 6], telomeres shorten with every cell division until the critically short telomere length is reached that induces cell senescence or apoptosis [7–9]. To counteract telomere shortening, approximately 85% of human cancer cells express telomerase, while those that don’t express telomerase induce alternative lengthening of telomeres (ALT) pathway [10–12].
As a typical fragile site, telomere of ALT cells experiences a high frequency of homologous recombination (HR), which may contribute to lengthening of telomeres [13]. ALT cells are characterized by high heterogeneity of telomere length [10], an elevated frequency of telomere-sister chromatid exchanges (T-SCEs) [13, 14], the presence of APBs (ALT-associated promyelocytic leukemia nuclear bodies) [15] and abundant extrachromosomal circular telomeric DNA (C-circles) and C-rich terminal 5' overhangs (C-overhangs) [10, 16–18]. The biogenesis of C-circles and C-overhangs is not clear and their functions in cells are largely unknown. It has been proposed that telomeric DNA damage, particularly double-stranded breaks (DSBs), promotes C-circles generation in ALT cells [19, 20], and that defects in telomere replication related proteins, such as SMARCAL1 (SWI/SNF-related, matrix associated, actin-dependent, regulator of chromatin subfamily A-like 1) or the CST (CTC1/STN1/TEN1), changes the level of C-circles in ALT cells [21–23]. These results imply a potential connection between C-circles formation and DNA damage repair and/or replication defect at telomeres. Regarding to C-overhang, it appears that telomeric DNA damage is not sufficient to induce 5' C-overhangs, rather, the production of C-overhangs is associated with rapid cleavage of telomeres [24]. The question regarding whether and how the formation of C-circle and C-overhang is coordinated and their relationship with high frequency of telomeric HR and ALT remains to be elucidated.
The tandemly repeated G-rich DNA in human telomeres has a relatively high tendency to form highly compacted G-quadruplex [25]. In addition, telomeric DNA is susceptible to ultraviolet light-induced [23] and oxidative DNA damage, leading to a relatively high frequency of single- and double-stranded DNA breaks (SSBs and DSBs) and other DNA lesions in telomeric DNA [26–28]. G-quadruplex and DNA lesions frequently block replication fork progression [29–32]. In ALT cells, telomeres may experience particularly high frequency of telomeric DNA damage [17, 33, 34], leading to replication fork stalling and/or collapse. In addition, it has also been reported that ALT is linked to mutations in the ATRX/DAXX chromatin remodeling complex and histone variant H3.3, which interfere with nucleosome assembly at telomeres and likely increase replication stress [33, 35, 36]. With such replication stress, it has been interpreted that ALT cells are competent to replication defect at telomeres. Alternatively, ALT cells may develop a mechanism to cope with unsuccessfully replicated telomeres and to maintain the integrity of chromosome ends.
This study provides evidences that C-circles and C-overhangs are produced during replication of lagging and leading strand of telomeres, respectively, and that their production is associated with DNA break-induced replication fork collapse in ALT cells. Replication fork regression, which facilitates HR-dependent replication fork restart, is utilized to rescue collapsed replication fork. However, unsuccessful rescue results in the formation of C-circles and C-overhangs. Meanwhile, multiple hallmarks of ALT were raised by DNA break-induced replication fork collapse, including increased frequency of telomeric HR, formation of ALT associated PML body as well as high abundance of C-circles and C-overhangs.
C-circle is an extrachromosomal circular telomeric DNA composed of full C-rich strand and notched G-rich complementary strand that is a quantitative biomarker of the ALT mechanism [10, 37]. Here, the cell cycle dependence of the appearance of C-circles was explored in ALT-positive U2OS cells. Specifically, U2OS cells were synchronized at G1/S by double-thymidine block, released for 0, 3, 6, 9 or 12h, corresponding to G1/S, early S, middle S, late S/G2 and G1, respectively (Fig 1A), and then assayed for the presence of C-circles. Φ29 DNA polymerase-based C-circle assay was used to determine the abundance of C-circles in cells. Reliability of method was validated by experiments in which lack of Φ29 leads to no amplified product and C-circle signal is well proportional to the amount of input DNA (R2 = 0.96 in linear regression of standard curve) (S1A and S1B Fig) [37]. The results showed that the abundance of C-circles increased gradually during S phase, peaked (doubled) at late S/G2 (9h after release) and decreased when cells re-entered G1 (Fig 1B). Since telomeric DNA replicates throughout S phase [38, 39], this result suggests that C-circles may be produced during telomere replication and subsequently degraded during or after G2 [40].
It has been reported that RPA2 (replication protein A2) colocalizes with telomeric DNA in human ALT cells [40, 41]. We observed that the abundance of telomeric RPA2 foci also gradually increased during S phase and decreased during or after G2 (Fig 1C). Given that RPA2 is a sensor of single-stranded DNA that might be produced during DNA damage repair and/or replication process, the specific correlation between appearance of C-circles and telomeric RPA2 foci implies that C-circle formation may associated with DNA damage response (DDR) and/or DNA replication at telomeres.
The mechanism by which C-circles form was further explored by BrdU pulse-labeling synchronized U2OS cells for 12 h after release from G1/S [42], isolating nascent C-circle and analizing its composition by CsCl density gradient ultracentrifugation (Fig 1D). The results showed that while the leading and lagging telomeric DNA was synthesized with similar efficiency (0.96 vs 1.11 in amount) (Fig 1E), BrdU-labeled C-circles were dominently enriched in lagging strand telomeric DNA (0.88 vs 1.64 for leading vs lagging synthesized C-circles after normalizing with total amount of leading or lagging telomeres) (Fig 1F), in which C-rich strand is newly synthesized and therefore BrdU-labeled. This result is consistent with the previous observation [43], the mechanism underlying the production of C-cirlce from lagging strand would be exlored below.
A similar procedure was used to determine whether 5' C-overhangs arise preferentially during leading or lagging strand DNA replication. 12h or 6h BrdU-labeled DNA was fractionated by CsCl gradient ultracentrifugation, fractions corresponding to leading, lagging and unreplicated telomeric DNA were collected, and divided into two parts, one of which was incubated with RecJf, a 5'→3' exonuclease for ssDNA, to specifically degrade 5' overhang DNA to validate telomeric C-rich ssDNA polarity [18]. The resulting samples were analyzed by neutral-neutral 2D agarose gel electrophoresis in which DNA fragments are resolved first by size in one dimension, and then by conformation in second dimension. In combination with in-gel hybridization under native or denatured conditions, 2D agarose gel electrophoresis is able to separate and distinguish linear ssDNA (G-rich or C-rich), linear dsDNA (with or without single-stranded G/C-rich overhangs) and open circular DNA (Fig 1G). C-overhangs were sensitive to RecJf digestion, but resistant to Exo I (a 3'→5' exonuclease for ssDNA), demonstrating that C-overhang is in the 5' to 3' direction (Fig 1H, S1C Fig) [18]. Both 12h and 6h BrdU labeling experiments showed that RecJf sensitive 5' C-overhangs are preferentially generated on telomeres replicated by leading strand (leading: lagging = 1.00 : 0.36 for 12h labeling sample and 1.00 : 0.29 for 6h labeling sample) (Fig 1H, S1D Fig).
To examine a potential relationship between replication-blocking and formation of C-circle and C-overhang, U2OS cells were treated with agents that result in replication fork stalling. To this end, exponentially growing U2OS cells were treated with HU (hydroxyurea) that blocks replication fork by inducing dNTP pool deficiency or aphidicolin that inhibits B-family DNA polymerase leading to replication fork stalling [44, 45]. Interestingly, both treatments resulted in no increase of RPA2 foci or DNA damage foci (p53-binding protein 1, 53BP1 foci) on telomeres (termed as TIFs: telomere dysfunction induced foci) (S2A and S2B Fig). In addition, the number of C-circles slightly decreased (Fig 2A) and the abundance of C-overhangs and G-overhangs was not significantly changed when U2OS cells were treated with HU or aphidicolin (Fig 2B, S3A Fig). These results suggested that replication fork stalling per se is not sufficient to stimulate the formation of C-circles and C-overhangs.
When U2OS cells were exposed to zeocin, a radio-mimetic chemical that induces oxidative DNA damage including ssDNA and dsDNA breaks [46], increased level of DDR at telomeres was detected, as expected (S2B Fig). Meanwhile, we observed increased abundance of C-circles and C-overhangs (Fig 2C and 2D). In contrast, zeocin treatment led to slight decrease of G-overhangs (S3B Fig). Importantly, we also found that the increase of C-circle and C-overhang upon zeocin treatment was restricted to S-phase (when cells were treated during S-phase), and was abrogated when cells were synchronized at G1 and exposured to zeocin (S3C–S3F Fig). Altogether, these results suggested that C-circles and C-overhangs are produced in ALT cells during telomere replication encountering DNA damages.
To imitate the situation in which replication fork progress is blocked by DNA damage, U2OS cells were treated with CPT (camptothecin), a specific inhibitor of Topo I (topoisomerase I) that induces protein-linked ssDNA break at the front of replication fork, leading to fork collapse [47, 48]. Strikingly, CPT strongly stimulated the formation of C-circles and C-overhangs in U2OS cells (Fig 2E and 2F) and the abundance of G-overhangs decreased accordingly (S3B Fig). Increased C-circle and C-overhang by zeocin or CPT treatment was also observed in other ALT VA13 cells (S3G and S3H Fig).
In a similar experiment, U2OS cells were treated with inhibitors of Topo II (topoisomerase II), VP-16 (etoposide) or ICRF-187 (dexrazoxane) [49–51]. VP-16 induces protein-linked dsDNA breaks in replicating DNA leading to replication fork collapse, while ICRF-187 inhibits the cleavage activity of Topo II leading to replication fork stalling [52]. Interestingly, VP-16 stimulated formation of C-circles and C-overhangs, whereas the abundance of G-overhangs decreased. However, ICRF-187 treatment showed a limited effect on C-circle, C-overhang and G-overhangs (S4A–S4C Fig). This result further demonstrated that DNA damage-induced replication fork collapse rather than replication fork stalling promotes formation of C-circles and C-overhangs. This conclusion was further confirmed by duplicating the experiments in other ALT positive VA13 cells (S4D–S4F Fig).
Given a high abundance of C-circles and C-overhangs in ALT cells, we then asked whether intrinsic DNA strand breaks exist in telomeres that induce replication fork collapse. For this purpose, genomic DNA was digested with HinfI and RsaI, followed by digestion with Exo III (Exonuclease III), a 3' to 5' exonuclease that remove nucleotides from blunt end or break/gap in double stranded DNA to generate single stranded DNA (Fig 3A). The digested DNA were hybridized with C-rich or G-rich probe under native or denatured condition [53] (Fig 3A). The rationale was to determine whether endogenous ssDNA breaks and gaps occur in the G-rich or C-rich strand of telomeric DNA in ALT cells. If such lesions were enriched in the C-rich strand, it would be preferentially degraded by Exo III, leaving primarily G-rich ssDNA to hybridize with a C-rich probe, while the reverse specificity, or lack of specificity would be observed in the absence of preferential endogenous ssDNA breaks in the C-rich strand of the telomere (Fig 3A). We detected much more G-rich ssDNA than C-rich ssDNA (Fig 3B). This suggests that endogenous ssDNA breaks and gaps are present predominantly on the C-rich strand of telomeric DNA in U2OS cells.
MMS (Methyl-methanesulfonate) is a DNA damaging agent that preferentially creates mutagenic lesions in cytosine of ssDNA [54, 55]. Here, we showed that exposure to MMS significantly stimulates formation of C-circles and C-overhangs in ALT cells (Fig 3C and 3D). To further confirm that ssDNA breaks in the C-rich strand of telomeric DNA induce formation of C-circle and C-overhang, we expressed CRISPR-Cas9 with a D10A mutation in the RuvC nuclease domain of Cas9 (Cas9-D10A), which specifically generates ssDNA breaks in the strand complementary to sgRNA (Fig 3E) [56, 57]. Indeed, when sgRNA with telomeric G-rich sequence (sgTel) was co-expressed with Cas9-D10A in U2OS cells, we observed significant number of C-rich, but not G-rich DNA fragments (smear on gel) that are released from telomeres and detected by alkaline constant-field gel electrophoresis (alkaline plug assay, see Method for detail) (S5A and S5B Fig), indicating specific induction of DNA breaks by Cas9-D10A at C-rich strands. As expected, expression of wild-type Cas9 (wtCas9, WT) induced double-stranded breaks at telomeres [58], leading to increase of both G- and C-rich fragments (S5A and S5B Fig). However, expression of nuclease-deficient mutant Cas9 (dCas9) led to no increase of G- or C-rich fragments (S5A and S5B Fig). We observed significant increase of C-circles and C-overhangs in cells expressing wtCas9 or Cas9-D10A but not in cells expressing dCas9 (S5C and S5D Fig). Similar experiments were also performed in non-ALT human HEK 293T cells. Western blot analysis demonstrated that wtCas9, dCas9 and Cas9-D10A protein were expressed at a similar level (Fig 3F). The results showed that expression of both wtCas9 and Cas9-D10A stimulated formation of C-circles (Fig 3G). The formation of C-overhangs was strongly stimulated by expression of Cas9-D10A (Fig 3H). And the expression of wtCas9 also slightly increased the abundance of C-overhangs, consistent with our previous finding that telomeric DSB initiates homologous recombination mediated repair that produces 3′ C-rich overhang [58]. Collectively, these results suggest that endogenous breaks/gaps or extraneously induced ssDNA breaks in C-rich strand of telomeric DNA stimulates formation of C-circles and C-overhangs.
To explore how ALT cells respond to replication fork collapse, CPT-treated U2OS cells were analyzed using IF-FISH (immunofluorescence and fluorescence in situ hybridization) to determine the proteins enriched at telomeres in response to replication fork collapse induced by CPT treatment. Strikingly, the abundance of PML foci and ALT associated PML bodies (APBs) increased in cells exposed to CPT (Fig 4A and 4B). In addition, we observed a significant increase of 53BP1 foci genome-wide and at telomeres when cells were treated with CPT (Fig 4C and 4D). Moreover, RPA2 foci at telomeres increased in CPT treated cells (Fig 4E and 4F). We also found that SMARCAL1, Rad51 and SLX4 were recruited to telomeres in U2OS cells and their foci at telomeres were significantly increased when cells were challenged with CPT (Fig 4G–4L). Importantly, RPA2/Rad51/SMARCAL1/SLX4 compose a machinery termed replication fork regression [21, 31, 59–61], a mechanism that rescues collapsed replication fork.
Evidences presented above suggest that replication fork collapse in telomeric DNA is tightly linked to formation of C-circle and C-overhang structures. Therefore, it was predicted that replication fork regression, mediated by RPA2-Rad51-SMARCAL1-SLX4 axis, might rescue collapsed replication fork and thus suppress formation of C-circles and C-overhangs. To test this, U2OS cells were treated with RPA2- or SMARCAL1-targeted siRNA and the abundance of C-circles and C-overhangs was examined (Fig 5A and 5D). Indeed, C-circles and C-overhangs were more abundant in RPA2 or SMARCAL1-deficient U2OS cells (Fig 5B, 5C, 5E and 5F). Accordingly, G-overhangs were slightly decreased (S6A Fig). Moreover, when Rad51 was inhibited by B02, a specific inhibitor of Rad51 [62], the abundance of C-circles and C-overhangs also increased, while G-overhangs decreased (Fig 5G and 5H, S6B Fig). These results suggested that replication fork regression prevents the formation of C-circles and C-overhangs in U2OS cells. The same experiments were also repeated in VA13 cells. Consistently, we observed that both depletion of RPA2 (or SMARCAL1) and inhibition of Rad51 by B02 stimulated formation of C-circles and C-overhangs, but reduced the abundance of G-overhangs (S6C–S6I Fig).
Meanwhile, we observed that pDNA-PKcs (DNA-dependent protein kinase, catalytic subunit), a key sensor in the NHEJ (non-homologous end-joining) pathway, localizes to telomeres in U2OS cells, and that telomeric pDNA-PKcs (S2056) foci are more abundant in CPT-treated cells (S7A and S7B Fig). To explore whether NHEJ plays a role in production of C-circles or C-overhangs, U2OS cells were treated with DNA-PKcs inhibitor NU7441. Previous report demonstrated that inhibition of ATR by VE-821 leads to decrease of C-circles [23]. Our results showed that NU7441 treatment decreased the level of C-circles (S7C Fig), whereas the abundance of C-overhangs and G-overhangs remained largely unchanged (S7D and S7E Fig). These findings implied that NHEJ machinery promotes circularization of lagging strand DNA at collapsed telomeric replication fork, thus enabling formation of C-circles.
Previous studies suggest that replication fork regression is coupled with HR to reinitiate replication [19, 31, 61]. Consistent with this, we observed that telomere sister chromatid exchange (T-SCE), which indicates HR occurring at telomeres, was inhibited (i.e., reduced frequency) in SMARCAL1-knockdown U2OS cells (Fig 6B and 6C). In addition, when replication fork collapse was induced by CPT treatment in U2OS cells, we observed increased frequency of T-SCE (Fig 6D and 6E). These results supported the hypothesis that replication fork collapse at telomeres, which is rescued by replication fork regression-mediated process, leads to telomeric recombination.
Rad51 plays a key role in both replication fork regression and telomeric recombination [19, 31]. When U2OS cells were treated with B02, a specific inhibitor of Rad51 [62], telomeric PCNA and RPA2 foci increased (S8A–S8D Fig), likely due to the accumulation of collapsed replication fork [50]. Consistently, less fully synthesized telomeric DNA were detected (S8E Fig). After treatment with B02 for 4 days, short telomeres were accumulated in U2OS cells (S8F Fig).
This study investigates the biogenesis of C-circles and C-overhangs in ALT cells. Evidence is presented that C-circles and C-overhangs represent circularized lagging and broken leading strands of telomeric DNA, respectively, and that their formation is tightly linked to strand break-induced collapse of DNA replication forks in telomeric DNA in ALT cells. Although replication fork regression and HR-mediated replication restart can rescue replication fork collapse, the formation of C-circle and C-overhang on unresolved replication fork may represent a new manner for ALT cells to cope with unsuccessfully replicated telomeres and to maintain chromosome integrity.
The results show that replication fork collapse leading to C-circles and C-overhangs can be induced by exogenous agents that generate ssDNA or dsDNA breaks in telomeric DNA (zeocin, CPT, VP16, MMS and CRISPR-Cas9 system). In addition, we found that endogenous lesions occur preferentially in the C-rich strand of telomeres in ALT cells (Fig 3B). During replication progression, the C-rich strand templates leading strand DNA synthesis, while the G-rich strand templates lagging strand DNA synthesis [63]. Replication fork collapse, induced by a break or gap on the C-rich strand, has different consequences for leading and lagging replication (Fig 7). For leading synthesis, long single-stranded C-rich DNA remains unreplicated, forming C-overhangs at the end of the chromosome; lagging strand synthesis still proceeds, but would likely lead to “looping-out” during which the stalled replication fork is cut out and cyclized to form C-circles [50]. The evidence supporting this model include: 1) C-circles are primarily derived from the lagging strand, whereas C-overhangs are primarily derived from the leading strand (Fig 1); 2) replication fork collapse, but not conventional replication fork stalling, stimulates production of C-circles and C-overhangs (Fig 2); 3) endogenous breaks/gaps are present in the C-rich telomeric strand and agents that introduce ssDNA breaks in the C-rich telomeric strand (i.e., MMS treatment and CRISPR-Cas9 (D10A)) stimulate formation of C-circles and C-overhangs (Fig 3); 4) the formation of C-circles is NHEJ-dependent (S7 Fig).
Spontaneous telomeric DNA damage response and telomeric RPA2 foci are often observed in ALT cells [40, 64]. We found that RPA2 accumulate in telomeric DNA of CPT-treated cells (Fig 4E and 4F) but not in HU- or aphidicolin-treated cells (S2A Fig), suggesting that replication fork collapse, but not replication fork stalling, is the signal that recruits RPA2. Binding of RPA2 to ssDNA activates the ATR pathway, leading to replication fork protection and restoration [23, 65]. DNA damage-induced replication fork collapse can be rescued by replication fork regression [48]. And, SMARCAL1, which is activated by ATR-dependent phosphorylation, plays an essential role in replication fork regression [66]. Indeed, it has been previously reported that the deficiency of SMARCAL1 leads to fragile telomeres [60]. Our results also demonstrated that components of fork regression machinery including Rad51, SMARCAL1 and SLX4 are recruited to telomeres in CPT-treated cells (Fig 4E–4L). This, however, does not exclude the possibility that the mechinary other than fork regression might also be adopted to rescue collapsed replication fork. In fact, it has been reported that translesion synthesis, which is composed of FANCJ, RAD18, ubPCNA and Polη, is engaged in bypassing DNA lesions on replication fork [67] [68].
Interestingly, depletion or inhibition of each component of fork regression machinery stimulated formation of C-circles and C-overhangs (Fig 5). We thus proposed that endogenous breaks/gaps in C-rich strand of telomeric DNA in ALT cells either induce replication fork collapse, leading to C-circle and C-overhang structures, or, fork regression or other fork rescue machinary restores the collapsed replication fork, and C-circle and C-overhang formation is suppressed (Fig 7). In supporting this model, it has been recently discovered that SMARCAL1 loss-of-function mutations in cancers link to the ALT mechanism of telomere maintenance, resulting in ultrabright telomeric foci and the generation of C-circles [69]. In additon, deficiency of Polη, which is essential for translesion synthesis, increases replication stress at telomeres and stimulates the formation of C-circles [68].
ALT cells are characterized by high frequency telomeric recombination [13]. It has been found that a DNA damage at ALT telomeres triggers long-range movement and clustering, resulting in homology-directed telomere synthesis between sister and non-sister chromatids [19, 70]. Here, we demonstrated that DNA strand break in the C-rich strand of the telomere leads to replication fork collapse, followed by replication fork regression and telomeric HR (Fig 6). Whether or not this recombination contributes to telomere elongation remains to be elucidated. However, homologous searching and recombination could occur anywhere along the telomere due to its repetitive nature, creating the possibility for telomere extension (Fig 7). In the presence of Rad51 inhibitor B02, short telomeres accumulate (S8F Fig), supporting the idea that Rad51-dependent HR promotes telomere extension. Furthermore, it is worth noting that APBs accumulate in CPT-treated ALT cells (Fig 4). Therefore, replication fork collapse provokes multiple hallmarks of ALT, including telomeric HR, APBs, C-circles and C-overhangs. While C-circles and C-overhangs are associated with telomere trimming, telomeric HR might contribute to telomere elongation. Additional studies are needed to understand how these events are coordinated in order to maintain chromosome end integrity and telomere length homeostasis in ALT cells.
U2OS, HEK 293T and VA13 cells were grown in Dulbecco’s modified Eagles’ medium (DMEM) supplemented with 10% fetal bovine serum, 1% penicillin/streptomycin at 37°C in 5% CO2. Unless otherwise indicated, cell lines were treated with HU (2mM, Sigma) or aphidicolin (1μg/mL, Sigma) or zeocin (100μg/mL, Thermo Fisher) or CPT (0.25μM, MCE) or MMS (0.25mM, Sigma) or VP-16 (10μM, Sigma) or ICRF-187 (50μg/mL, Selleck) or B02 (27.4μM, EMD Millipore) or VE821 (10μM, MCE) or NU7441 (250nM, Selleck). Knockdown experiments were performed with the Lipofectamine RNAiMAX Reagent (Thermo Fisher Scientific) using following siRNA targets:
siNC/UUCUCCGAACGUGUCACGUdTdT/ACGUGACACGUUCGGAGAAdTdT; siRPA2-1/GGCUCCAACCAACAUUGUUdTdT/AACAAUGUUGGUUGGAGCCdTdT; siRPA2-2/GCCUGGUAGCCUUUAAGAUdTdT/AUCUUAAAGGCUACCAGGCdTdT; siSMARCAL1-1/CCAAGAGACACCAGCUCAUdTdT/
AUGAGCUGGUGUCUCUUGG dTdT;
siSMARCAL1-2/UUGCUAAGAAGGUCAAAGCdTdT/
GCUUUGACCUUCUUAGCAAdTdT. Cells were collected 60h after transfection for experiments.
Lenti-CRISPRv2 (Addgene plasmid # 52961) was used in this study [71]. Scaffold sequence of sgRNA was modified to
5'-NNNNNNNNNNNNNNNNNNGUUUAAGAGCUAUGCUGGAAACAGCAUA GCAAGUUUAAAUAAGGCUAGUCCGUUAUCAACUUGAAAAAGUGGCACCGAGUCGGUGCUUUUUUU-3′, as previously described [58, 72]. sgSCR (5′-TGCTCCGTGCATCTGGCATC-3'), sgTel (5'-GTTAGGGTTAGGG TTAGGGTTA-3') [58] were cloned as described previously [73]. Cas9 mutations, including D10A and dead-nuclease (D10A, H840A), were constructed by mutagenesis kit (Fast Mutagenesis System, Transgen Biotech). The transfection was carried out with Lipofectamine 2000 Reagent (Thermo Fisher Scientific).
U2OS cells were synchronized at G1/S by "double thymidine block" method as described previously [42]. Briefly, exponentially growing U2OS cells were blocked with 2mM thymidine for 19h, washed three times with prewarmed PBS and released into fresh medium for 10h, and then blocked again with 2mM thymidine for another 14h. For 5-bromo-2-deoxyuridine (BrdU, Sigma) labeling, cells were incubated with fresh medium containing 100μM BrdU for 12 h after release from G1/S. FACS analysis was carried out as described previously [42].
All genomic DNA was extracted and purified using AxyPrep Blood Genomic DNA Miniprep Kit (Axygen) according to manufacturer’s instructions. DNA concentration was measured by Nanodrop-2000. For 2D agarose gel analysis, 10μg DNA was digested overnight at 37°C with 10U HinfI (Thermo Fisher), 10U RsaI (Thermo Fisher) and 2μg/mL RNase A (Takara). The reaction was terminated with EDTA and analyzed by 2D agarose gel electrophoresis. 30U RecJf (New England Biolabs) was added for removing 5' single-stranded DNA.
For internal gaps/nicks analysis, 5μg genomic DNA was digested overnight at 37°C with 5U HinfI (Thermo Fisher), 5U RsaI (Thermo Fisher) and 1μg/mL Ribonuclease A (RNase A, Takara) and purified with QIAquick PCR Purification kit (Qiagen). Purified DNA were digested with or without 200U Exonuclease III (New England Biolabs) overnight at 37°C, and then subjected to 0.7% agarose gel electrophoresis and in gel hybridization.
CsCl gradient ultracentrifugation and DNA purification were performed as described previously [5, 42]. DNA was purified and dissolved in 60μL ddH2O. One half of each sample was incubated with RecJf prior to analysis by 2D agarose gel electrophoresis.
C-circle assay was performed as described previously [37]. The concentration of genomic DNA was determined by fluoremetry based method (Qubit 3.0 Fluorometer, Thermo Fisher Scientific). Exactly the same amount of genomic DNA was input for C-circle assay (30ng for U2OS, 100ng for HEK 293T and VA13 cells). Each assay was repeated three times to obtain the quantitative result. To determine C-circles in CsCl fractions, 1μL of each fraction was incubated in 40μL reaction containing 19μL ddH2O and 20μL C-circle amplification master buffer (0.2mg/mL BSA, 0.1% Tween 20, 1mM dATP, dGTP and dTTP each, 1× Φ29 Buffer and 7.5U Φ29 DNA polymerase (Thermo Fisher)) for 8h at 30°C, and then subjected to slot-blot and hybridization with C-probe.
Neutral-Neutral 2D agarose gel electrophoresis was performed as described previously [74, 75]. Briefly, enzyme digested DNA samples were loaded into a 0.4% agarose gel for first dimension and electrophoresis was performed at 1V/cm for 12 h at room temperature in TBE buffer. Lanes were excised, soaked in TBE containing 0.3μg/mL ethidium bromide (EB) (Sigma) for 30min, embedded in 1% agarose gel containing EB. Second dimension electrophoresis was performed at 4°Cfor 6 h at 3V/cm. The gel was then dried at room temperature by vacuum drier, and hybridized with G-/C-probe under native or denatured condition.
The telomere length assay was performed as previously described [42]. 5μg genomic DNA was digested overnight at 37°C with 5U HinfI (Thermo Fisher), 5U RsaI (Thermo Fisher) and 1μg/mL RNase A (Takara). Digested DNA samples were subjected to conventional 0.7% agarose gel in TAE buffer at 2V/cm for 16h at room temperature. The gel was dried at 42°C with vacuum drier, and hybridized with C-probe.
IF-FISH was performed as previously described [50, 58].
Fluorescent probe is Cy3-(TTAGGG)3 (Panagene). Primary antibodies include anti-53BP1 (Novus Biologicals), anti-RPA2 (EMD Millipore), anti-PCNA (Genetex), anti-PML (Santa Cruz), anti-SMARCAL1 (Santa Cruz), anti-SLX4 (Novus Biologicals), anti-pDNA-PKcs (S2056) (Abcam) and anti-Rad51 (Santa Cruz). Secondary antibodies include DyLight488 conjugated anti-rabbit (Multisciences), DyLight488 conjugated anti-mouse (Multisciences).
The telomeric sister chromatid exchange (T-SCE) was determined by CO-FISH, which is performed as described previously [58, 76]. Fluorescent probes are Cy3-(TTAGGG)3 (Panagene) and FITC-(CCTAAA)3 (Panagene).
The alkaline constant-field gel electrophoresis was performed as described previously described [77]. Briefly, 1×106 cells were rinsed with 1×PBS, resuspended in 50μl 0.7% 45°C pre-warmed agarose (made with 1×TE, pH 8.0), and solidified in 1ml decapitated injector. Agarose plugs were incubated in fresh-made lysis-buffer (30mM Tris-HCl pH8.0, 300mM NaCl, 25mM EDTA, 0.5% SDS, 0.1mg/ml Protease K, 0.1mg/ml RNase A) overnight. The plugs were then denatured in 100mM NaOH with 1mM EDTA, placed into the wells of 0.7% alkaline agarose gel (50mM NaOH with 1mM EDTA) and sealed with 0.7% alkaline agarose gel. Electrophoresis was carried out at 1V/cm for 12h at 4°C. The gel was subjected to in gel hybridization with telomeric probe.
In-gel hybridization was performed as described previously [50, 78]. For native in gel hybridization, gels were hybridized in Denhart’s hybridization buffer with 32P-labeled C-/G- telomeric probe. The telomeric probes were prepared as described previously [79]. Gels were washed 3 times with 2×SSC and 0.5% SDS, exposed to PhosphorImager screen (GE Healthcare) and scanned on Typhoon imager (GE Healthcare). Image Quant software was used for data analysis. For denatured in-gel hybridization, gels were denatured with 0.5 M NaOH, neutralized with 1 M Tris–HCl (pH 8.0) and then followed the procedure for native hybridization.
Western blots were performed with antibodies against Flag (Monoclonal ANTI-FLAG M2 antibody, F1804, Sigma), SMARCAL1 (Santa Cruz), RPA2 (EMD Millipore), or β-actin (Proteintech).
Two-tailed unpaired student’s t-test was used for statistical analysis (Graphpad Prism). Error bars represent the mean± SEM of three biological repeats/independent experiments. * P<0.05, ** P<0.005, *** P<0.001.
|
10.1371/journal.pcbi.1003622 | Slow Noise in the Period of a Biological Oscillator Underlies Gradual Trends and Abrupt Transitions in Phasic Relationships in Hybrid Neural Networks | In order to study the ability of coupled neural oscillators to synchronize in the presence of intrinsic as opposed to synaptic noise, we constructed hybrid circuits consisting of one biological and one computational model neuron with reciprocal synaptic inhibition using the dynamic clamp. Uncoupled, both neurons fired periodic trains of action potentials. Most coupled circuits exhibited qualitative changes between one-to-one phase-locking with fairly constant phasic relationships and phase slipping with a constant progression in the phasic relationships across cycles. The phase resetting curve (PRC) and intrinsic periods were measured for both neurons, and used to construct a map of the firing intervals for both the coupled and externally forced (PRC measurement) conditions. For the coupled network, a stable fixed point of the map predicted phase locking, and its absence produced phase slipping. Repetitive application of the map was used to calibrate different noise models to simultaneously fit the noise level in the measurement of the PRC and the dynamics of the hybrid circuit experiments. Only a noise model that added history-dependent variability to the intrinsic period could fit both data sets with the same parameter values, as well as capture bifurcations in the fixed points of the map that cause switching between slipping and locking. We conclude that the biological neurons in our study have slowly-fluctuating stochastic dynamics that confer history dependence on the period. Theoretical results to date on the behavior of ensembles of noisy biological oscillators may require re-evaluation to account for transitions induced by slow noise dynamics.
| Many biological phenomena exhibit synchronized oscillations in the presence of noise and heterogeneity. These include brain rhythms that underlie cognition and spinal rhythms that underlie rhythmic motor activity like breathing and locomotion. A two oscillator system was constructed in which most of the circuit was implemented in a computer model, and was therefore completely known and under the control of the investigators. The one biological component was an oscillator in which an apparently novel manifestation of biological noise was identified, dynamical noise in the period of the oscillator itself. This study quantifies how much noise and heterogeneity this simple two oscillator system can tolerate before desynchronizing. More complicated systems of oscillators may follow similar principles.
| Synchronized neural firing is a characteristic activity pattern of neural systems. Synchronized neural activity in cortical circuits [1] is thought to underlie many aspects of cognition [2], [3], including recognition [4], recall [5], perception [6], [7], and attention [8]. Phase-locked neural activity is also an essential component of central pattern generators (CPGs) located in the spinal cords of vertebrates and the ganglia of invertebrates [9], [10]. Inhibition plays a central role in oscillatory synchrony, and in this study we focus on reciprocal inhibitory coupling.
A major contribution of this paper is a distinct notion of noise in coupled oscillatory neurons, which we explore by comparing three models of noise intrinsic to the neurons (see Methods). The dominant source of noise in neurons is thought to be synaptic [11]. This thinking is shaped by studies of cortical circuits, in which neurons in a high conductance state that receive a stochastic barrage of fast and balanced excitatory and inhibitory input show fast fluctuations in membrane potential [12]. An early attempt to quantify the effect of noise on neural activity [13] examined the case of a perfect integrator with additive white noise. The output of the integrator is interpreted as the membrane potential. In the absence of noise, a baseline current produces a regular oscillator with constant angular velocity that is reset each time it reaches threshold. The noise takes the form of Gaussian current noise added to the baseline current. When this noise is integrated, it is analogous to a trajectory produced by Brownian motion, and produces a one-dimensional random walk in the membrane potential superimposed on the steady upward trend caused by the constant baseline current. In this model, membrane potential is proportional to the phase of the oscillation, so a random walk in the phase occurs. The time scale of this noise is fast, due to its theoretically flat spectrum which includes very high frequency components. The current noise has no history-dependence since the value at each time point is random and independent of all previous values. However, the membrane potential does have a memory, because at each time step, the value is a perturbed version of the value at the previous time step. The second moment, or variance, of the displacement of the membrane potential from its original value is proportional to the product of the diffusion constant and the time step. The mean squared displacement therefore grows as the square root of the size of the time step [14]. The memory of noise on the previous cycle is wiped out when the membrane potential and the phase are reset when the spike threshold is reached. Based on this model, a common way to add noise to phase models of neurons is simply to add Gaussian noise to the phase [15], [16], which is one of the noise models that we use in this study.
Real neurons have complex nonlinear intrinsic currents, and thus may not linearly integrate their extrinsic inputs. We modeled the intrinsic period as stochastic due to random fluctuations in factors that influence the period. If these factors have little history dependence, for example, variability in the number of ions passing through an open channel at any given time, then successive interspike intervals are uncorrelated and may appear to be drawn from a Gaussian distribution [13], [17], [18]. Gaussian noise added to the period is the second model used in this study. If the period of one cycle depends on the previous cycle because the stochastic fluctuations occur in history-dependent processes, then a different model must be used [19]. History dependent noise may arise from slowly changing levels of stochastic fluctuations in the numbers of open channels for adaptation currents [20] or levels of second messengers, channel phosphorylation, insertion and deletion of channels into the membrane, and other unknown factors. Instead of drawing the period from a distribution, the period itself can be made to undergo Brownian motion under the assumption that the period is equally likely to be perturbed in either direction at a given instant, and that the displacement is therefore proportional to the square root of the time step. Finally, if we assume that the mean of the noise reverts to zero, we obtain an Ornstein-Uhlenbeck [21] process added to a constant period, which is the third and final noise model used in this study. This latter model is novel, although it shares some elements with the model of Schwalger et al. [20], and constitutes a different noise model that may complement the fast noise in some circumstances. We postulate that the period of biological oscillatory neurons varies randomly but with history dependence. The direct effect on network activity of slow stochastic dynamics that cause history dependence in the period of component oscillators has not been previously investigated. This slow form of intrinsic noise may have implications for synchronization and phase locking in neural circuits.
In this study, we construct hybrid neural circuits consisting of one biological and one computationally modeled neuron. These coupled pairs exhibit different patterns of activity, which we refer to as motifs, during coupling. Understanding how and why synchronization and phase locking occur in populations of neurons is critical to understanding how neural circuits function. Phase-locking implies a constant phase relationship between neural oscillators; synchrony is a special case of phase-locking in which spikes occur in different neurons at about the same time. Another observed motif is phase slipping. In this motif, the spiking activity of the faster cell “laps” the slower one and the timing relationships are different in every cycle. Our analysis of these dynamics utilizes the phase resetting curve (PRC) measured from both biological and model neurons in response to the same stimulus pulses that the neurons receive in the circuit; an action potential in the presynaptic neuron triggers a predetermined conductance waveform in the postsynaptic neuron both in the hybrid network and in the protocol for measuring the PRC. The PRC describes how a neuron's period is shortened or lengthened depending upon at what point in the cycle a perturbation was received [22], [23]. This PRC is a useful tool for predicting synchronization and phase locking in neural systems under the assumption that the phase resetting due to an input is complete by the time the neuron receiving the input spikes next or by the time it receives another input, whichever occurs first. The PRC for biological neurons as well as the hybrid circuit activity is measured in the presence of ubiquitous biological noise. The impact of noise on PRC-based predictions is an open question.
The overall aim of this work was to assess why different dynamical motifs, such as phase locking and phase slipping, were observed in hybrid circuits and to explain how random transitions between these motifs occurred. Using PRC-based maps, we were able to predict phase locking and synchronization in two-neuron networks and describe the activity motifs observed in these circuits. By comparing the performance of three noise models in simulations of hybrid circuit activity, we were able to show that noise contributes to variability within and switching between different motifs, and that history-dependent noise in the period was necessary to mimic motif variability and transitions seen in experiments.
Aplysia californica were acquired from the Miami National Resource for Aplysia (Miami, FL) and kept in saltwater tanks at room temperature for 1–2 weeks until used. Animals were anesthetized using a solution of 71.2 g MgCl2 in 1 L 1X artificial sea water (1X ASW). 1X ASW was comprised of (in mM) 460 NaCl, 10 KCl, 11 CaCl2, 30 MgCl2, 25 MgSO4, and 10 HEPES (pH 7.6) [24]. The abdominal ganglion was dissected out of the animal and pinned in a Sylgard-lined (Dow Corning) dish filled with dissection solution (30% 1X MgCl2 solution and 70% ASW solution) for desheathing. The ganglion was desheathed under a dissection microscope. The dish solution was then replaced with a high-Mg2+ low Ca 2+ recording solution, which contained (in mM) 330 NaCl, 10 KCl, 90 MgCl2, 20 MgSO4, 2 CaCl2, and 10 HEPES, pH 7.6 [25]. Electrodes consisted of pulled (Sutter P-97 puller) glass pipettes containing 3 M potassium acetate and silver wire chlorided in bleach.
Regularly spiking neurons in the lower left quadrant of the Aplysia were used as the biological neurons in the hybrid circuits. An Axoclamp 2B amplifier with Clampex 8.2 software (Molecular Devices) was used to supply stimulus currents and record membrane potential. A Digidata 1322A Digitizer (Molecular Devices) was used to sample electrophysiological data at 10 kHz.
Wang-Buzsaki (WB) model neurons were used in the hybrid circuit experiments. The equations and parameters for the WB model neuron were the same as in [26] except that the leak reversal potential EL was set to -60.0 mV and the applied current Iapp was controlled to match the 1–5 Hz spiking frequency of the Aplysia spiking neuron. This modified WB model matches both the spike dynamics and PRC shape of experimentally measured neurons [27], [28]. Iapp for the model neuron was chosen such that the spiking frequency was similar to that of the biological neuron. Synaptic conductance values for the model neuron were selected to increase the likelihood of 1∶1 phase locking in hybrid circuits. The differential equations for the state variables of the WB model and the two virtual synapses were updated in real time. The voltage measured in the biological neuron was used to determine the time course of the conductance for the synapse onto the model neuron and the driving force for the synaptic current of the synapse onto the biological neuron.
Dynamic clamp is a real-time computational and experimental technique used to add data-driven simulated ion channel conductances to biological neurons [29]–[31]. For these experiments, we used the Model Reference Current Injection (MRCI) [32] system to construct hybrid circuits and measure phase resetting curves. The dynamic clamp system operated at a frequency of 10 kHz, which corresponds to a closed-loop sampling and computation period of 100 µs. Reciprocal inhibitory synapses were used in hybrid circuits, and inhibitory perturbations were used to measure phase resetting curves. The alpha-shaped conductance waveform was calculated using the following equations: dy/dt = −y/τ +itrig; dα/dt = −α/τ + y; Isyn = gsyn α(V- Esyn) e/τ. V corresponds to the membrane potential of the postsynaptic cell, Esyn was set to −70 mV, and τ and gsyn were varied as in Table 1. The value of itrig was zero except when an input was triggered, either because the presynaptic cell spiked in the hybrid circuit or a perturbation was needed to measure a point on the PRC, then itrig was set to amplitude 1 for 1 ms. The e/τ term normalizes the maximum amplitude of the conductance waveform to gsyn.
PRCs were measured using the dynamic clamp to apply inhibitory inputs at various times during the neuron's interspike interval (ISI). Perturbations were separated by at least 10 cycles to allow ISIs to return to pre-perturbation magnitudes. The stimulus interval ts corresponds to the time interval between the previous spike in the neuron receiving the input and the start of the applied perturbation. This interval was normalized by unperturbed period P0, which was the average of the five ISIs prior to the perturbation, to obtain the phase θ = ts/P0. Phase reset (Figure 1B) was calculated as the perturbed period P1 minus the unperturbed period P0, normalized by the unperturbed period P0 (see Figure 1A). Neuronal spikes were detected using a −40 mV threshold. Biological PRCs were fit using 3rd or 4th order polynomials to minimize least squared error and promote randomly-distributed residuals. Noiseless model neuron PRCs were spline fit. This fit was necessary in order to use the PRCs as functions in the network simulations described below.
An alternative way to present the information from a PRC is in the stimulus interval – recovery interval (ts-tr) plane (Figure 1C). Stimulus interval refers to the time interval between when the neuron last spiked and when a perturbation arrived. The recovery interval tr refers to the interval between the time of application of the perturbation and the time of the next spike in the perturbed neuron. This description preserves time information, unlike the PRC whose quantities are unitless. Very strong perturbations result in more pronounced curves on the ts-tr plane, whereas less strong perturbations manifest in the ts-tr plane as nearly straight lines. As seen in Figure 1B and C, a PRC with peak magnitude of around 0.05 (black curve), looks somewhat like a straight line on the ts-tr plane. This apparent flattening occurs because the PRC plot is scaled to the maximum PRC amplitude, whereas the scale of the ts-tr plot is determined by the maximum period of the oscillation.
Hybrid circuits of one biological neuron and one model neuron were constructed using the dynamic clamp; 13 distinct biological neurons were used to construct the 35 hybrid circuits presented here. No noise was added to the circuit, all noise was intrinsic to the biological neuron. A single biological neuron was used for multiple hybrid circuits, with different conductance and time constant values, for as long as the experiment remained viable. All synapses were inhibitory because the reversal potential for both synapses (Esyn) was set to −70 mV. In nearly all cases, PRCs of the biological neurons were measured with conductance parameters gsyn and τ that were used for the coupling experiments. In a limited number of cases, coupling experiments were performed with a weaker conductance than the one at which the corresponding PRC was measured. In such cases, the PRC was linearly scaled to calculate the curves that describe the network interactions. We previously showed that for conductance below a certain threshold, PRC shape is preserved and scales linearly with amplitude [33]. Our goal was to choose coupling values that resulted in 1∶1 synchrony; however, because the PRC measured before the experiment constrains the coupling parameters used in the experiment, but the biological neuron activity can change over time, in practice a range of effective couplings were obtained.
Dynamical motifs were defined as characteristically different episodes of network activity. Network phase φnet was defined as the position of the spike in the biological neuron within the cycle in the model neuron that contains the spike. Network phase was calculated as tsM/(tsM + trM), where tsM is the time interval between a spike in the model neuron and the following spike in the biological neuron (which perturbs the model neuron), and trM is the time interval between the spike in the biological neuron and the next spike in the model neuron. The first 10 network phases were discarded to eliminate transient effects. Network phase that remained within ±0.1 units of the network phase for 20 or more cycles was defined as phase-locked (Figure 2A). Activity in which the network phase transitioned through consecutive increasing or decreasing phases, which often resulted in one neuron spiking twice during the ISI of the other neuron, was defined as phase slipping (Figure 2B). Episodes that did not meet either criterion were categorized as other. See Text S1 for more detailed information on the algorithm used for automated characterization. In some cases, coupling was turned on and off during an experiment; this was done to determine the robustness of the hybrid circuit activity.
To measure the consistency of phase locking, we used circular statistics to find the R2 metric, often referred to as the vector strength, for each experiment [34]. In circular statistics, values are represented by a unit vector and an angle. The average vector captures the mean angle φave of all the data and the magnitude R, which is a measure of the tightness of the locking [35]. In our case, the average network phase is φave for the phase of the firing of the model neuron within the cycle of the biological neuron, and the magnitude R corresponds to how consistent the network phase is during an experiment. The strength of phase locking is represented by the length of the vector, R, where R2 = X2+Y2. As in [34], φave and R2 are calculated usingwhere atan2 is the two argument arctan function that returns a value between –π and π, Pn is the network period measured in cycle n, N is the number of network periods, and tsM,n is the nth stimulus time for neuron M, the model cell. Note that the signs of X and Y must be considered in the two argument version of arctan to put φave in the appropriate quadrant. In [34], an R2 threshold of 0.7 is used to distinguish strongly phase-locked systems, which have R2 near 1, from those with weaker locking. Higher R2 magnitudes indicate that a system does not deviate much from the phase-locked angle and has a dominant phase-locked mode, while lower R2 magnitudes indicate more variability in network phase. R2 calculations and PRC fits were performed in MATLAB (The MathWorks).
Each hybrid circuit experiment was simulated using PRC-based maps. In these simulations, the phase variable evolves at a rate determined by the intrinsic frequency, with instantaneous phase resetting applied at the time of input from the other neuron according to the measured PRCs. A key assumption is that the shape of the PRC does not change with the relatively small changes in the period of the oscillator. Simulated PRCs were constructed to mimic the shape and magnitude of biological and model neurons used during experiments. Network simulations were performed in C. Conceptually, our noiseless map [36], [37] is a modified Winfree [22] phase model in which the intrinsic phase θi ranges from 0 to 1, and is reset from 1 to 0 when a spike occurs (1)where ηi is the angular velocity in neuron i, θj is the phase in presynaptic neuron j and fi(θi) is the phase resetting due to each spike in presynaptic neuron j. We do not integrate Equation 1, instead we assume the phase changes at a constant velocity between inputs, and jumps instantaneously when an input is received. The result is a coupled nonlinear map, which was used to simulate both the PRC experiments and the hybrid circuit experiments and implemented as follows. The map requires the PRC and the initial value of the intrinsic period for each neuron, and the initial values of the phase of each neuron. The phases are only updated at the times associated with each episode of neural firing, so the first step after initialization is to determine which neuron(s) will fire next. This is accomplished by finding the shortest recovery interval (tri = Pi(1- θi)), where Pi is the current estimate of intrinsic period of the ith neuron, based on the noise models given in the main text, and θi is its phase. At the next firing time, the phase of the firing neuron is reset to zero. The recovery interval in the next neuron (j) to fire is also the stimulus interval (tsi) for the nonfiring neuron (i). The phase of neuron i is calculated as θi = tsi/Pi and the phase is decremented by the resetting fi(θi) calculated at that phase when a spike occurs in the presynaptic neuron j. The next event is again determined by finding the shortest recovery interval (trj = Pj(1- θj)) until the next spike. We added noise to Eq. 1 model in three ways, which renders it a Langevin equation in phase.
To construct hybrid circuits, one biological neuron from the abdominal ganglion of Aplysia californica was reciprocally coupled to one Wang-Buzsaki (WB) [26] conductance-based model neuron using the dynamic clamp [29], [30]. The dynamic clamp measures the potential in the biological neuron, integrates the differential equations for the WB model and the two virtual synapses, and injects synaptic current into the biological neuron. The WB model was used because it produces phase resetting curves (PRCs) that are comprised of only delays in response to an inhibitory input (Figure 1B), and because the WB PRCs resemble those measured in Aplysia neurons [27], [28]. Parameters for the hybrid circuits and maximum phase resetting values for the biological and model neurons are shown in Table 1. Notice that the maximum phase resetting is different between the biological and model neurons; this discrepancy creates a heterogeneous system. The average interspike interval of the biological neuron during coupling, which corresponds to the network period if the system is phase-locked, is different than the uncoupled biological neuron period; this provides evidence that the motifs observed in our hybrid networks result from mutual coupling effects, and do not reflect entrainment of the model neuron by the biological neuron.
All 35 hybrid circuits showed episodes of phase locking, phase slipping, or both (see Figure 2). In Figure 3, the horizontal axis represents time and each experiment is represented on one row. The experiments are ranked vertically in order of R2, a metric of the consistency of phase locking during coupling. Coupled neurons with high R2 values remain phase-locked for the entire experiment duration. As R2 decreases, more episodes of phase slipping and undefined activity occur in the hybrid circuit. Note that an experiment with motif changes can nonetheless have a higher R2 value than one that is always phase-locked, particularly when the network phase in the first case has less variability than the network phase values in the second case. Well-defined network motifs occurred in every experiment.
When two neurons are coupled, the dynamics of the resulting network can be predicted by plotting the PRC data of each neuron in the ts-tr plane. As stated in the Methods, the stimulus interval ts is the interval between the previous spike and an input from the other neuron, whereas the recovery interval tr is the interval between the arrival of an input and the next spike. We refer to these curves in the ts-tr plane as interaction curves. In contrast to the weak coupling approach [39], [40] using the infinitesimal PRC (iPRC), we do not ignore the effects of phase resetting on the network period nor do we require the relative phase of the neurons to change slowly compared to their absolute phases, however we do require that the coupling be pulsatile, meaning that the effects of an input die out quickly, before the next event occurs. In the coupled system, the stimulus interval for one neuron equals the recovery interval for the other neuron (Figure 4A) and vice versa. In a one-to-one periodic phase-locked mode, the intervals do not change from cycle to cycle, indicated by the index ∞ in Figure 4B1. For each neuron, a pair of stimulus and recovery intervals correspond to each phase at which an input is received (Figure 4A). In Figure 4B2, the stimulus interval for one neuron (magenta, model neuron) is plotted on the x-axis and the corresponding recovery interval is plotted on the y-axis, whereas the stimulus interval for the other neuron (black, biological neuron) is plotted on the y-axis and recovery interval on the x-axis. Therefore the two pairs of stimulus and recovery intervals (in two different neurons) that must be equal in a phase-locked mode are plotted on the same axes. The intersections of these curves then correspond to any possible periodic phase-locked modes of the two neuron network, as well as to fixed points of the ts-tr map in Figure 4C that is described below.
The information in the ts-tr interaction curves is not restricted to the location of the fixed points, but also provides the transient dynamics that may lead to a phase-locked mode or persist indefinitely in the absence of such a mode. The stimulus interval in one neuron determines the recovery interval in that same neuron; this leads to a map (Figure 4C1) with the following dynamics. The index n indicates successive cycles in the model neuron. The movement of the operating point from the black to the magenta curve is constrained to be horizontal because the recovery interval in the biological neuron determines the next stimulus interval in the model neuron (trB[n] = trM[n]). Similarly, the movement of the operating point from the magenta to the black curve is constrained to be vertical because the recovery interval in the model neuron determines the next stimulus interval in the biological neuron (trM[n] = trB[n+1]). For a stable fixed point that attracts nearby trajectories, the magenta curve with the coordinates listed in the order (trM,tsM) curve must have a steeper slope ([41], see also derivation in Text S1) than the black curve in which the coordinates are listed in the opposite order (tsB,trB), otherwise the point is unstable and repels trajectories. The white circle in Figure 4C2 (and B2) repels trajectories and therefore denotes an unstable fixed point, whereas the red circle in Figure 4B2 is stable because nearby trajectories would be attracted rather than repelled.
Figure 5 shows an example of stationary phase locking that occurs when there is a stable fixed point on the PRC-based map. The ts-tr interaction curves in Figure 5A were generated with the period observed in the biological neuron just prior to coupling, and intersect at two fixed points, one unstable (white) and one stable (red). The latter corresponds to the phase locking observed in both experiments and simulations. The insets reflect that a change in the intrinsic period of the biological neuron results in a shift of the ts-tr interaction curve for the biological neuron. As the neuron period gets longer, the curve shifts upward and rightward along the x-y diagonal (see left inset), and as the period gets shorter, the curve shifts inward toward the ts-tr origin (see right inset). The network phase remains relatively constant for the entire duration of coupling in this experiment (Figure 5B1), resulting in a histogram of the network phases with a distinct peak. In simulation, we can produce a similar time series of network phases and histogram in the presence of the three types of noise (Figure 5B2-B4), although only the OU noise produces a sufficiently broad peak. Figure 5C explains why the phase locking is robust to noise. The red curve shows the location of the stable fixed point in terms of tsB (and trM) as a function of the period of the biological neurons shown on the y-axis. The initial value of period (used as μ in the noise models) is shown by the lowermost dashed horizontal line labeled μ. The initial value of tsB at the fixed point is about 600 ms as shown in Figure 5A. If the period of the biological neuron decreases, the curves no longer intersect below a period of about 740 ms and a tsB of about 755 ms (rightmost vertical dashed line labeled C2) corresponding to the situation in the inset at right. Similarly, if the period of the biological neuron increases, the curves no longer intersect above a period of about 875 ms, and a tsB of about 430 ms (leftmost vertical dashed line labeled C1) corresponding to the situation in the inset at left. The ts-tr interaction curves are a snapshot of the constraints on the trajectories based on the current value of the period of the biological neuron. The variability in all three models constrains the 95% confidence interval of the intrinsic period (μ±2σeff) to lie well within the range of periods that supports phase locking, and therefore constrains the variability in the network phase observed in Figure 5B2-4, and presumably in Figure 5B1 as well.
The PRC-based map helps illustrate what happens during phase slipping. Figure 6A1 shows the ts-tr interaction curves using the period observed in the biological neuron just prior to coupling, with a trajectory around the PRC-based map indicated by dashed blue lines with arrows indicating direction. Every time one neuron spikes, a vertical or horizontal “step” is taken between the two curves. The trajectory spends more time near the point of closest approach between the two curves because it takes smaller steps in that region. Here, the ghost of a fixed point that exists at a slightly different set of parameters (at which the curves do intersect) has a significant impact on the dynamics [42], [43]. Figure 6A2 shows the sequence of network phases observed during the long episode of slipping in experiment 5, and the histogram at right shows a broad peak in the network phases. The peak and distribution of the histogram of the network phases produced by a map based on the PRC with OU noise in the period and shown in Figure 6A3 was in reasonable agreement with the experimental data in Figure 6A2. The peak of the histogram is due to the tendency to stick near a phase corresponding to the point of closest approach of the curves in Figure 6A1.
Each phase slip in the network activity is associated with a trajectory that dropped down from the upper left edge of the map and was reinjected at the lower right edge (Figure 6A1). Figure 6B displays the mechanism for dropping off the map at the upper left and returning at the lower left. This occurs when the next recovery interval in one neuron (the biological neuron in Figure 6B1) is so long that the other neuron (the model neuron in Figure 6B2) spikes twice during one biological neuron period. We defined a recovery interval tr* (see two pulse PRC protocol in Figure 6B4) that gives the interval to the next spike after two inputs separated by the intrinsic frequency of the partner neuron are received (brown curve in Figure 6B3). Therefore the trajectory is reinjected at the lower right when it falls off the upper left, and vice versa. The recovery interval tr* was not measured, but instead was calculated from the previously collected phase resetting data by assuming that the second input was received at a phase determined not only by the elapsed time but by taking into account the phase resetting from the first pulse. The only way to transition between the ends of the map is for one neuron to spike twice in a row, and the modified map can handle any firing pattern in which any single neuron does not spike more than twice in a row. Across a phase slip transition, there is a change [44] in leader-follower pattern of the neurons. It is important to note that there exists a similar analogy of ghost attractor and cycle slipping in return map of Poincare phase map of neural oscillators [45].
Phase locking and phase slipping do not always persist throughout an experiment; as seen in Figure 3, motifs can vary over time. Of the 35 hybrid experiments presented here, 12 experiments represent the case where the system is phase-locked (Figure 3, experiment 16 and experiments 24–35). In the remaining 23 cases, the coupled neurons transition between phase locking, phase slipping, and undefined phase relationships (Figure 3, experiments 1–15 and experiments 17–23). It is likely these transitions are due to fluctuations in the intrinsic spiking frequency of the biological neuron. As illustrated in Figure 5A, shifts in the ts-tr interaction curves due to drift in the period of the biological neuron can move, create, or eliminate the fixed points of the system.
Figure 7A1 shows experimentally observed phase-locked activity with a single slip. One simulated coupling experiment (Figure 7A2) with the Ornstein-Uhlenbeck noise process mimics the single slip from phase-locked activity observed in the experiment. However, the simulations are sensitive to initial conditions, and using a different random seed for the noise produces a different pattern of transitions (Figure 7A3). The ts-tr interaction curves coupled with the effective standard deviation of the period for the various noise processes, can explain why the OU process, but not the other noise models, was able to mimic the transition to slipping activity. The red curve in Figure 7B2 again shows how the location of the stable fixed point in terms of tsB (and trM) change and disappear as the period is increased or decreased. The initial value of period (used as μ in the noise models) is shown by the lowermost dashed horizontal line in Figure 7B2, and was used to generate the ts-tr interaction curves shown in Figure 7B1 that have two intersections, including a stable fixed point that predicts the initially observed phase locking. However, if the period of the biological neuron increases from the initial value of 839 ms above 845 ms, the stable intersection is lost, resulting in the ts-tr interaction curve shown in Figure 7B3 that produces phase slipping. The shaded regions in Figure 7B2 show that the OU noise model produces a larger standard deviation of the period, such that the 95% confidence interval of the intrinsic period (μ±2σeff) includes periods that correspond to phase slipping (crosshatched region). On the other hand, the 95% confidence intervals for the other two models lie well within the region of periods corresponding to phase locking, so slipping was never observed for any random seed in simulations of this particular experiment using those noise models. Figure 8 gives an example in which the same experiment (number 19) illustrated in Figure 7 was simulated with Gaussian noise added to the phase resetting. In this case, as in the case in which Gaussian noise is added to the period, there is no dependence of the noise in one cycle on that in the previous cycle. Figure 8A illustrates a typical fit to the network dynamics with low (panel A1) and high (panel A2) noise levels. The high noise levels are able to capture the network dynamics, but the low noise level fails. Conversely, Figure 8B shows that the low noise level (panel B1) faithfully captures the low variability in the PRC as measured, but the high noise level greatly overestimates the variability. Figure S1 shows that the simulations of this experiment with Gaussian noise added to the period fails in exactly the same way.
Robustness to shifts in period on the PRC-based map translates to robustness of phase-locked network activity. This robustness to period changes is represented by the shape and proximity of curves on the ts-tr plane. The network phase of a system with one curvy ts-tr neuron representation and one straight ts-tr neuron representation is more likely to stay phase-locked than a system where both neuron representations are very straight; this robustness also depends on the position of one curve with respect to the other, since intersections near the edges of ts-tr curves will be susceptible to noise-induced bifurcations. Figure 1 showed that increasing the coupling strength increases the curvature of the interaction curves for the models (and experiments, not shown) used in this study. In Figure 9, the coupling between neurons was turned off between panels, and a snapshot of the ts-tr interaction curves (top row) was generated for each coupled episode; the black curve representing the biological neuron was shifted according to the average period measured during the previous uncoupled episode. Figure 9A shows a network with stable phase locking; a change in biological neuron period did not disrupt the motif. Figure 9B shows a case where network activity transitions from locked to slipping, due to the decreasing biological neuron period and the resulting loss of the stable fixed point.
One advantage of using a hybrid circuit model is that only the biological neuron contributes noise to the system. We simulated two coupled neurons using iterated maps derived from the PRCs measured from the biological and model neurons during experiments. The model neuron was noiseless and had a constant intrinsic period. Three types of noise models were used in only the simulated biological neuron to consider different types of biological variability. Gaussian noise added to the simulated biological PRC mimics the approach in [15], [34] and describes uncertainty in the PRC itself. Gaussian noise added to the simulated biological neuron period represents uncertainty in the measurement of spikes as well as intrinsic variability in the timing of neuronal spikes. Modeling the period of the biological neuron as an OU process captures intrinsic variability in spike timing in the biological neuron and measurement errors, as well as slow, long-term trends in the intrinsic period.
Figure 10 shows the performance of the three noise models on three metrics: transitions between motifs (Figure 10A), the fraction of time spent phase-locked (Figure 10B), and the circular statistic R2 (Figure 10C). Experiments 24–35 spent 100% of the time phase-locked (see Figure 3 and Table 1) with no transitions, therefore their R2 is quite high. The vast majority of the time that the circuit was not phase-locked was spent slipping, so the performance on the metric of fraction of time spent slipping was similar and is not shown. The metrics are presented in terms of the range of values obtained for ten different simulations for each noise model for each experiment. In each of the ten simulations, the noise model was initialized with a different seed, and the pulse coupled network simulator was run for the same length of elapsed time as the original experiment. The parameters of the noise models for each experiment are given in Table 2. Since some simulations were quite sensitive to initial conditions, the best possible match would be that the values of the metrics obtained in the ten simulations bracket the value actually observed in the experiment. The OU model was calibrated to fit the bifurcation data in Figure 10A, and all experimental data points (black dots) are bracketed by the range of simulation results (yellow bars) for that model. Note that the fits given for the OU model in Table 2 are not unique (see Supplementary Figure S2), however, two consistent trends emerged. Decreasing the time constant τ for mean reversion somewhat reduced the PRC noise and the numbers of bifurcations introduced by increasing the noise intensity σ, and the variability between runs was greater for larger mean reversion time constants. The other models did a poor job of capturing the bifurcations, or transitions between motifs. In general, the transitions identified in those models were sticky regimes (meaning they exhibited a “preferred” phase) during phase slipping that the algorithm identified as phase locking episodes (see Supplementary Figure S3), and they failed to capture many transitions, such as the one illustrated in Figure 8 (see also Supplementary Figure S1). Most black data points that represent a nonzero number of bifurcations were not bracketed by the simulations for the two Gaussian models because they failed to exhibit history dependence of the period (see Supplementary Figure S4). Although the OU model was not calibrated to capture the fraction of time spent phase-locked data and R2, the OU model clearly outperformed the other models on these metrics as well. The Gaussian noise models had less effective variability in network activity than the OU model, so the OU model was better able to capture fraction of time spent phase-locked. In contrast, the Gaussian noise models had a tendency to produce simulations that were always phase-locked, or to a lesser degree that were always slipping. The OU models usually bracketed the data points for the R2 metric, but the other models in general did not.
We have shown here that multiple types of network activity occur in hybrid circuits of one biological and one computational neuron. We used PRC-based maps to explain activity observed in hybrid circuits of one biological and one computational model neuron. These maps are based on interaction curves that give a snapshot in time of the dynamics expected, assuming that the intrinsic periods of the neurons remain relatively constant during the time window in question. The fixed points, or equilibria, of the PRC-based maps presented here are given by the intersections of the ts-tr interaction curves for the two neurons, and correspond to one-to-one phase-locked modes in the circuit. However, perturbations from these fixed points are inevitable in a noisy system, and the nonequilibrium dynamics of the map as trajectories flow between the interaction curves gives the system dynamics for perturbations away from fixed points, and also in the complete absence of fixed points. The two most common dynamical motifs, phase slipping and phase locking, can occur under variable circumstances. The existence of a stable fixed point predicts phase locking, and the absence of a stable fixed point predicts continuous phase slipping. However, the interaction curves themselves can change over time because they are based on the intrinsic period of the component neurons. As the period of the biological neuron slows down, the PRC-based interaction curve for that neuron moves outward; as it speeds up, the curve moves in toward the origin. This motion can change where the curves intersect, effectively changing where the fixed points are located. This can result in a shift of the network phase if the system remains phase-locked, or a transition to phase slipping if the stable fixed point is lost.
The synaptic component of the noise is thought to be dominant compared to intrinsic noise sources, and noise is often modeled as a high conductance state [12] in which both the inhibitory and excitatory conductances are Ornstein-Uhlenbeck processes. Noise in neural systems [11] is also often modeled as a random walk in the membrane potential due to excitatory and inhibitory synaptic currents whose interevent times are generated by a Poisson process; the membrane potential is continuously pulled back toward the resting potential with a characteristic time constant. This approach may be appropriate for normally quiescent neurons, but additional considerations may apply for oscillatory neurons. An oscillatory neuron is not merely pulled towards a resting potential, but instead has a characteristic cycle period determined by the inverse of the ηi term in Eq. 1. Under our interpretation, intrinsic membrane noise in an oscillator can be modeled as perturbing the intrinsic cycle period, or 1/ηi. The stochastic form of equation 1 is referred to as a Langevin phase equation; to our knowledge we are the first to model the cycle period itself as an Ornstein-Uhlenbeck random process.
Here, we show that normally distributed noise added to cause jitter in either the PRC or the period was insufficient to capture the dynamics of the observed switches between motifs. Instead, history-dependence (see Supplementary Figure S4), presumably mediated by stochastic processes with slow dynamics that allowed the fast jitter to accumulate over time, was required in our simulations in order to replicate our experimental observations of hybrid circuit dynamics. The period of the biological neuron was modeled as an Ornstein-Uhlenbeck (OU) process with the mean reversion modeled as being on the order of 10–1000 cycle periods. The mean reversion was included because it is quite likely that period of biological oscillators is homeostatically regulated within a physiologically relevant range, but this term was not crucial (see Supplementary Figure S4) or particularly well-characterized in our data. Support for treating the period as a random process is provided by observations of slow fluctuations in oscillatory period when neurons that are nonoscillatory in a slice preparation, such as stellate cells in entorhinal cortex and CA1 hippocampal OLM cells, are made to oscillate using current injection [16], and by the successful use of an OU model to characterize the variability in period in CA1 pyramidal neurons under similar conditions [46]. There is strong support for the idea that neuronal circuits possess both intrinsic and synaptic mechanisms that operate over hours to days to maintain firing around a homeostatic stable point [47]. However, these studies were not focused on the homeostatic regulation of the intrinsic firing rate of oscillating neurons on a time scale of tens of seconds to minutes as we suggest here. The underlying biophysical mechanisms that could produce an OU process in period (or alternatively in frequency) in physiological neural oscillators are not clear. However, Schwalger [20] and Fisch et al. [48] have shown that stochastic slow ionic currents may be well-represented by OU noise in neurons.
Typically noise in the Langevin equation for phase is formulated as an additive term to the frequency [49], [50] or to the phase resetting [15], [16], [34], [51], [52] or both [53]. Another method is to convolve the noise with the infinitesimal phase resetting curve (iPRC) and then add to the phase [54] or frequency [55]. It is unlikely any of these methods could capture the transitions between modes observed in our experiments because they lack the crucial history dependence of the period from one cycle to the next. If the period of oscillatory neurons in general can be modeled as a history-dependent and likely mean-reverting random process, theoretical results to date on the behavior of ensembles of noisy biological oscillators may require re-evaluation and modification. Interestingly, theta oscillations in the local field potential of the hippocampus and prefrontal cortex also show a pattern of small frequency fluctuations over time, referred to as the microstructure of the theta rhythm [56], so the concept of the period as a random process may be extendable to network oscillations as well.
The predictions of phase locking under weak coupling assumptions [39], [40] are independent of coupling strength (as long as it is weak) for noisy identical oscillators, but clearly a minimum coupling strength is required to overcome ubiquitous biological noise. Effects of heterogeneity in period have usually been studied [26], [57], [58] assuming that the intrinsic period for each neuron is relatively constant. We examined the consequences of fluctuations in the period of neuronal oscillators in the absence of additional synaptic input using the dynamic clamp. In our hybrid circuit experiments, the biological synaptic inputs were silenced using high Mg2+, low Ca 2+ solution [25]; we found that the assumption of a slow variation in the period of the oscillatory biological neurons, even in the absence of synaptic input, produces the best fit to our data. We show explicitly the relationship between the level of variability and the tendency to remain phase-locked, as well as the effect of coupling strength on stabilizing phase locking. In our hybrid circuits, the robustness of network phase locking was related to the degree of curvature of the two interaction curves that generate the PRC-based map, as well as the amount of spatial separation between the two curves. The degree of heterogeneity in frequency largely determines the spatial separation, or difference in the x and y axis crossing points of the interaction curves. The interaction curves in Figures 5A and 7B1 have similar separation on the x and y axes, but the greater curvature of the magenta curves in Fig 5A clearly means that greater separation can be tolerated before the intersections between the interaction curves are lost. Curvature is enhanced by strengthening the synaptic coupling, and spatial separation increases with increasing heterogeneity in the intrinsic spiking frequencies. These two factors determine the amount of effective variability that can be tolerated without disrupting the locking (see Figure 5 and 7). Other investigators [16] have responded to the variability in period by utilizing a controller to stabilize the intrinsic period in order to test the predictions of weak coupling, which presumes that the coupled period is equal or very nearly equal to the intrinsic period. The direct effect of slow trends in the period of component neural oscillators on network activity has not been previously investigated. This slow form of intrinsic noise may have important implications for synchronization in neuronal networks.
These results are mainly of interest for their implications for larger networks, such as central pattern generating networks and hippocampal and cortical networks that subserve cognitive functions. There are two immediately apparent ways to generalize these results to larger networks. One is to generalize [36], [59], [60] from two neurons to two subpopulations of neurons in which the neurons in each subpopulation are different from those in the other subpopulation, but relatively homogeneous within a population. Homogeneity in frequency might be enforced by electrical coupling within but not between subpopulations for example. Another method is to directly scale up to larger networks; in this case our contribution is to suggest that slow intrinsic noise in the period that has not previously been considered may play a role in the collective dynamics of networks of coupled oscillators.
|
10.1371/journal.pbio.0060246 | The Timing of Differentiation of Adult Hippocampal Neurons Is Crucial for Spatial Memory | Adult neurogenesis in the dentate gyrus plays a critical role in hippocampus-dependent spatial learning. It remains unknown, however, how new neurons become functionally integrated into spatial circuits and contribute to hippocampus-mediated forms of learning and memory. To investigate these issues, we used a mouse model in which the differentiation of adult-generated dentate gyrus neurons can be anticipated by conditionally expressing the pro-differentiative gene PC3 (Tis21/BTG2) in nestin-positive progenitor cells. In contrast to previous studies that affected the number of newly generated neurons, this strategy selectively changes their timing of differentiation. New, adult-generated dentate gyrus progenitors, in which the PC3 transgene was expressed, showed accelerated differentiation and significantly reduced dendritic arborization and spine density. Functionally, this genetic manipulation specifically affected different hippocampus-dependent learning and memory tasks, including contextual fear conditioning, and selectively reduced synaptic plasticity in the dentate gyrus. Morphological and functional analyses of hippocampal neurons at different stages of differentiation, following transgene activation within defined time-windows, revealed that the new, adult-generated neurons up to 3–4 weeks of age are required not only to acquire new spatial information but also to use previously consolidated memories. Thus, the correct unwinding of these key memory functions, which can be an expression of the ability of adult-generated neurons to link subsequent events in memory circuits, is critically dependent on the correct timing of the initial stages of neuron maturation and connection to existing circuits.
| Previous studies have implicated adult-born hippocampal neurons in the formation of spatial and contextual memories by using mouse models where newly generated neurons are either eliminated or increased in number. Nonetheless, how new neurons are integrated in the existing circuits and contribute to memory formation still awaits clarification. Toward this end, we have developed a different approach, using a mouse model that accelerates the differentiation of the newly generated neurons without altering their number, and offers the possibility to induce the process at any chosen moment. We show that the new neurons pass through their early stages of maturation faster and, though establishing connections with the existing neuronal circuits, fail to function properly. In fact, mice are not only unable to learn new spatial information, but they are also unable to use previously acquired memories. These results demonstrate that the appropriate timing of maturation of new neurons is important for their adult performance in memory circuits, i.e., to integrate new memory traces and recall previous events.
| Observations in mammals and birds have revealed that neurogenesis continues in the dentate gyrus of the hippocampus throughout adulthood, due to the presence of progenitor cells localized in the innermost part of the granule cell layer, the subgranular zone (SGZ, [1–3]). These progenitor cells continue to proliferate and generate new dentate granule neurons for the entire life of the organism [4,5].
The process of adult hippocampal neurogenesis originates from dividing putative neural stem cells [6] and has been tentatively divided into six developmental stages [7], in which putative neural stem cells (named type-1 cells) develop into post-mitotic neurons through three consecutive stages of progenitor cells (type-2ab and type-3 cells; [8–10]). This process is thought to govern the number and the differentiation of adult-generated neurons [11].
Newborn neurons become functionally integrated into existing dentate gyrus circuitry within 3 wk, extending their axons to CA3, as indicated by morphological and electrophysiological studies [12–20]. Interestingly, recent observations indicate that new neurons of the dentate gyrus become functionally active in learning circuits at late stages of their maturation (∼4–6 postnatal weeks, [21]).
Adult hippocampal neurogenesis is required for hippocampus-dependent learning and memory [22,23]. Indeed, the almost complete ablation of neurogenesis by an antimitotic toxin, x-ray irradiation, or virus-activated pro-drugs results in profound cognitive deficits [24–28]. On the other hand, learning and/or physical exercise can enhance neurogenesis in the dentate gyrus, suggesting a two-way relationship between the generation of new neurons in the adult hippocampus and cognitive processes [9, 29–33]. Recently, it has been hypothesized that new, adult-generated neurons confer to the dentate gyrus the basic ability to encode the timing of new memories by integrating new events on the pre-existing memory circuits [23].
The strategies used so far to reveal the functional role of adult neurogenesis deeply affected the total number of new neurons, either by severely reducing neuronal progenitors [24,26] or by increasing the number of new neurons generated [29]. Nonetheless, the ablation of new neurons by toxin or irradiation in some instance did not affect certain hippocampus-dependent tasks of spatial learning, such as water maze or contextual fear conditioning [24,26,34,35]. In this context, the role of the specific differentiation steps during neurogenesis and dynamics of integration of new neurons into existing circuitry remains unknown [36].
To address this issue, we used the approach of selectively enhancing the differentiation of progenitor cells in adult dentate gyrus. We conditionally expressed the gene PC3 (also known as Tis21 or BTG2, see [37] for review; GenBank (http://www.ncbi.nlm.nih.gov/Genbank/) accession number M60921) in nestin-positive progenitor cells (type-1 and type-2; [7]) of the adult dentate gyrus. PC3 is normally expressed in neuronal precursors immediately before the last asymmetric division [38,39] and, during neurogenesis, it is known to induce their terminal differentiation in several areas of the CNS [40,41].
We found that nestin-driven expression of PC3 resulted in premature differentiation of adult-generated dentate gyrus neurons, with a reduction in the number of type-1 and type-2 neuronal progenitors. This genetic manipulation did not change the overall number of newly generated neurons; rather it altered their differentiation timing and resulted in profound changes in newborn neuron morphology. Remarkably, early PC3 expression caused a severe impairment in performance on the different hippocampus-dependent spatial learning and memory tests used and resulted in a selective reduction in synaptic plasticity in the dentate gyrus. These results indicate that early stages of adult neurogenesis are crucial for the correct functional integration of newborn neurons into cognitive hippocampal circuits.
Up-regulation of PC3 induces neural precursors to shift from proliferation to differentiation, thereby promoting the generation of new neurons [41]. Such enhancement of neurogenesis has been observed so far in neuroblasts of the neural tube and in cerebellar granule cell precursors, and follows inhibition of G1 to S phase progression and stimulation of proneural genes [41,42].
We reasoned that conditionally controlling the differentiation of progenitor cells in the adult dentate gyrus would allow a selective analysis of their involvement in spatial memory networks.
Thus, to start, we tested whether the differentiation of adult-generated newborn dentate gyrus neurons was influenced by PC3, which begins to be physiologically expressed in progenitor cells before differentiation, at the moment of their final mitosis (SF-V and FT, unpublished data). Up-regulation of PC3 was activated earlier in nestin-expressing adult hippocampal stem and progenitor cells (type-1 and type-2; [7]) of bitransgenic nestin-rtTA/TRE-PC3 mice (hereafter named TgPC3). This was achieved by conditional expression from postnatal day 30 (P30) onward of the transgene under control of the nestin promoter, through administration of doxycycline, as previously shown [41]. Sixty days after activation of the PC3 transgene, mice received five daily injections of BrdU (at P90–P94) to detect new progenitor cells and/or neurons, immediately followed by immunohistochemical analyses. Analysis of the expression of transgenic PC3 in P95 mice with the PC3 transgene active (named TgPC3 ON mice) indicated targeting to the dentate gyrus, as visualized by X-gal staining, which revealed the β-galactosidase activity of the β-geo reporter gene fused to the nestin-rtTA transgene (Figure 1A).
Among the different cell populations of the dentate gyrus, the glial fibrillary acidic protein (GFAP)-expressing astroglia are considered to be dividing putative neural stem cells from which neuronal progenitor cells originate, given their ability to repopulate the dentate gyrus after cytotoxic ablation of dividing cells [6]. These stem cells are identified primarily by the expression of GFAP, accompanied by nestin or also Sox2 [8,10,43,44] and have been defined as type-1 cells [7]. We found that, in TgPC3 ON mice at P95, the BrdU+/GFAP+/nestin+ cells—corresponding to 1–5-d-old type-1 cells—decreased significantly with respect to TgPC3 OFF control mice (about 40%; Figure 1B, p < 0.0003). The whole GFAP-positive astroglial cell population decreased only slightly, confirming the selectivity of effect on type-1 cells (Figure 1C). Since no difference between TgPC3 OFF and wild-type (WT) mice was found in these and in the following immunohistochemical analyses (unpublished data; p > 0.05), only TgPC3 OFF were considered as controls.
Next, we analyzed the transiently amplifying progenitor cells derived from type-1 cells. These progenitors express nestin but lack GFAP and astrocytic features and are divided into two subgroups based on the absence or presence of the immature neuronal marker, doublecortin (DCX) (named type-2a and type-2b, respectively; [8–10]). A further group of progenitor cells lacks nestin but expresses DCX (type-3; [9]). We observed that TgPC3 ON mice presented a significant decrease of 35% in the whole type-2a/type-2b population of 1–5-d-old newborn progenitor cells (identified as BrdU+/nestin+/GFAP–; p < 0.02, Figure 1D) and of 45% in type-2b progenitor cells (i.e., BrdU+/nestin+/DCX+; p < 0.002, Figure 1E). Moreover, TgPC3 ON mice also showed a reduction, albeit not significant, of type-3 progenitor cells (identified as BrdU+/nestin−/DCX+; Figure 1F), consistent with the fact that at this stage, the nestin promoter becomes physiologically inactive and thus ceases to drive the expression of transgenic PC3. A control of the expression of transgenic PC3 in nestin-positive type-1and type-2ab progenitor cell types, as determined by β-galactosidase expression, is shown in Figure S1A–S1D.
Type-3 progenitor cells generate early post-mitotic neurons identified by expression of the neuronal-specific differentiation marker, NeuN (stage 5 and 6; [45]). In contrast to progenitor cells type-1–type-3, in TgPC3 ON mice the number of early post-mitotic, differentiated neurons up to 5 d old expressing NeuN increased considerably—more than twice than in control mice (identified as Brdu+/NeuN+ as well as Brdu+/DCX+/NeuN+; p < 0.0001; Figure 1G and 1H, respectively). Representative images of Brdu+/DCX+/NeuN+ cells are shown in Figure 1L and 1M. These new neurons, which increased in number by activation of the PC3 transgene, were also positive for NeuroD1 (Brdu+/NeuroD1+/NeuN+; Figure 1I), whose expression begins in type-2b progenitor cells, is maintained in post-mitotic hippocampal granule neurons, and is required for differentiation [46,47]. We conclude that the nestin promoter-driven expression of transgenic PC3 accelerates the shift of adult-generated hippocampal stem and progenitors cells towards a post-mitotic, terminally differentiated phenotype, corresponding to NeuN-expressing neurons stage 5 and 6.
Notably, PC3 transgene activation did not change the total number of 1–5-d-old new progenitors generated in the dentate gyrus (BrdU+ cells analyzed at P95 after completion of BrdU injections; Figure 1J) or the total number of 4-wk-old new neurons (BrdU+/NeuN+ cells analyzed at P116; Figure 1K). The effects of transgene activation on the relative abundance of new dentate gyrus progenitor cell types and neurons is summarized in Figure 1N. In addition, no significant change was detected in the number of proliferating progenitor cells of the whole dentate gyrus in mice with active transgene with respect to control mice, as detected by Ki67 labeling (Figure S2A). Furthermore, an analysis of proliferating progenitor cell types expressing Ki67 showed that in TgPC3 mice, the number of type-1 (Ki67+/nestin+/GFAP+; p > 0.05), type-2ab (Ki67+/nestin+/GFAP–; p < 0.01) and type-2b (Ki67+/nestin+/DCX+; p < 0.01) decreased, whereas that of type-3 progenitor cells increased significantly (Ki67+/nestin–/DCX+; p < 0.01; Figure S2B–S2F).
Together, these data indicate that PC3 selectively accelerates the transition of stem and progenitor cells (type-1 to type-3) toward a post-mitotic differentiated state without affecting the following: (i) the total number of proliferating cells; (ii) the total number of newly generated neurons; or (iii) the final number of neurons differentiated. The decrease in the number of proliferating Ki67+/nestin+ type-1–type-2ab progenitor cells expressing PC3 is consistent with the notion that PC3 expression is associated to the neurogenic asymmetric type of division in neuroblasts [37–39].
To test whether activation of the PC3 transgene caused nonspecific changes, a stereological analysis of the hippocampus was conducted. Activation of the PC3 transgene after P30 did not alter the volume of the dentate gyrus or the whole hippocampus (Figure S3A and S3B) or the total cell number in the dentate gyrus (Figure S3C). Moreover, there was no evidence of reduced cell survival either in the whole hippocampus, as defined by TUNEL analysis (Figure S3D), or within different subpopulations, as defined by labeling with the apoptotic marker caspase-3 [48] (Figure S3E–S3J). Also, no evident alteration in distribution of mature granule neurons or in morphology was detected in the hippocampus or in the whole brain (unpublished data).
In conclusion, these results show that the activation of the PC3 transgene accelerates the process of differentiation of dentate gyrus progenitor cells, thus providing a tool to specifically analyze the relationship between timing of hippocampal neurogenesis and spatial learning.
We wished also to verify whether the expression of PC3 had any effect on nestin-positive neural cells in the other adult neurogenic niche, i.e., the subventricular zone (SVZ), comprising type B astrocytic-like progenitors, type C transit amplifying cells, and type A migrating neuroblasts [49–51]. We observed that the number of type B and type C cells, labeled by BrdU/GFAP and BrdU/NG2, respectively, did not change significantly in TgPC3 ON mice (p > 0.05; Figure S4A and S4B); on the other hand, the number of type A cells, corresponding to immature neurons derived from type C cells and identified by BrdU/DCX staining, was significantly lower than in control mice (40% decrease, p < 0.01; Figure S4C). Conversely, differentiated SVZ neurons up to 5 d old (BrdU+/NeuN+ cells) increased more than 2-fold in TgPC3 ON mice (p < 0.01; Figure S4D), although the final number of new neurons generated in SVZ, as analyzed measuring BrdU+/NeuN+ cells 3 wk later (at P116) in their final migratory destination, i.e., the olfactory bulb, did not differ in TgPC3 ON and OFF mice (unpublished data).
This suggests that PC3 accelerated the differentiation of SVZ cells, as observed for the cells of the dentate gyrus. It should be emphasized, however, that SVZ cells are not involved in spatial memory processes, being anatomically and functionally independent from those of the dentate gyrus, since they migrate through the rostral migratory stream to the olfactory bulb where they play a specific role in the olfactory processes [49,52].
We then investigated whether early-differentiated dentate gyrus neurons in TgPC3 ON mice showed an altered morphology in the three different stages of the morphogenetic process underlying the maturation of granule neurons from progenitors in the SGZ of dentate gyrus. We labeled the newly generated granule neurons in adult hippocampus of TgPC3 mice by infecting the dentate gyrus region with a retrovirus produced by the vector CAG-GFP [19]. Highly concentrated retrovirus (about 108 pfu/ml) was delivered through sterotaxic surgery in the dentate gyrus of TgPC3 mice at P95, two months after activation of the transgene, according to the time schedule followed for immunohistochemical analyses. CAG-GFP vector is replication incompetent; thus, only dividing cells at the moment of surgery could be infected.
The soma of the majority of GFP+ neurons detected in control mice was localized on the hilar border of the granule cell layer of the dentate gyrus at the different time points analyzed, i.e., throughout 7 −28 days post-infection (dpi). The most evident developmental feature of GFP+ neurons, detected at different dpi in TgPC3 mice with either active or inactive transgene, was the progressive growth of dendrites, whose lengths, shorter than the width of the dentate gyrus blade at 7 dpi, increased greatly at 16 dpi and further at 28 dpi, reaching far outside of the dentate gyrus. During these last two time points, the dendritic growth was accompanied by an increase in the complexity of arborization (Figure 2A). There were, however, clear differences in neuronal morphology between TgPC3 ON and TgPC3 OFF mice, which we evaluated by scoring dendritic length and branching points. Analysis of GFP+ neurons at 7 dpi revealed a significantly increased dendritic length in TgPC3 ON mice compared with TgPC3 OFF. This difference, however, was reversed in GFP+ neurons at 16 dpi and more clearly at 28 dpi, when the dendritic lengths in TgPC3 ON mice were significantly shorter than those in TgPC3 OFF mice (Figure 2B). Similarly, the number of branching points in GFP+ neurons at 7 dpi was significantly greater in TgPC3 ON than in TgPC3 OFF mice, but again this difference was transient, because in GFP+ neurons at 16 and 28 dpi, the number of branching points became equivalent in the two groups (Figure 2C).
To further evaluate the dendritic complexity, we analyzed the density of spines. The dendritic spine is the locus of connection to excitatory synaptic input, mainly glutamatergic [53]; therefore, the process of spine formation has consequences in the functionality of neurons. Normally, spine growth starts in the adult-generated dentate gyrus neurons at about 16 dpi and reaches a plateau at 56 dpi, but at 28 dpi the exponential phase of growth has already ceased, and this stage can be considered representative of the attainment of the morphological maturity of the neuron [19]. Therefore, we focused our analyses at 28 dpi, and also at 70 dpi, when the plateau of growth is attained. Quantification of spine density in dentate gyrus neurons indicated a significant reduction both at 28 dpi and at 70 dpi in TgPC3 mice with active transgene, compared to mice with inactive transgene (Figure 2D). However, at 28 dpi the decrease of spine density was substantial (60%), whereas at 70 dpi, it was mild (15% decrease; Figure 2D and 2E).
We conclude that the dendritic length of new dentate gyrus neurons in TgPC3 ON mice, albeit greater than in control mice at 7 dpi, appeared significantly shorter at 28 dpi, associated with reduced spine density. This evidence points to a decrease in dendritic growth and in spine number during the developmental stages between 7 and 28 dpi; a partial recovery of normal values of spine density may slowly occur during the following stages. The absence of alteration in the branching points would indicate that no major morphological alterations occur, other than those likely consequent to a faster attainment of the end point of the maturation process.
Spatial learning was tested in the Morris water maze [54,55]. In this task, which is mainly dependent on an intact hippocampus [54,56], mice learn across daily sessions to find a hidden escape platform using extra-maze visual cues. A first session of experiments, carried out in 3-mo-old WT, TgPC3 OFF and TgPC3 ON mice (whose transgene had been activated at P30; see time-schedule in Figure 3A), showed significant differences among groups in escape latencies both in learning (trials 1–18; F(2,38) = 32.04; p < 0.0001) and reversal learning (trials 19–30; F(2,38) = 34.15; p < 0.0001). As indicated by post-hoc comparisons (Duncan multiple range test), TgPC3 ON mice were dramatically impaired (p < 0.05) both in learning and reversal learning compared to WT and TgPC3 OFF mice, whereas no significant differences were found between the performances of TgPC3 OFF and WT mice (Figure 3B). Moreover, both in the first (p < 0.001) and the second (p < 0.01) probe trial, carried out 24 h after learning and reversal learning, respectively, TgPC3 ON mice spent a significantly smaller amount of time in the target quadrant, compared to both TgPC3 OFF and WT mice. No significant differences were found between TgPC3 OFF and WT mice (Figure 3C). In this analysis, we considered WT mice, either doxy-treated or doxy-free, as a single control sample, because no significant differences were found between the two groups. An analysis of behavioral aspects not related to learning, i.e., wall hugging behavior (thigmotaxys) and swimming speed, did not show significant differences between TgPC3 ON, TgPC3 OFF, and WT mice either during the first or the second behavioral session (Figure S5A and S5B). However, increased levels of thigmotaxys were detectable at later stages of training in TgPC3 ON mice, which may reflect their poor learning.
To further test the effect of the PC3 transgene activation on spatial memory, we submitted transgenic and control mice to a less stressful behavioral task compared to the water maze, namely the eight-arm radial maze [57,58]. In this task, mice have to search for a food pellet located at the end of each arm. Reentering a previously visited arm was considered as an error. Significant differences in the percentage of errors were found among groups (F(2,38) = 6.71; p < 0.005). As shown by post-hoc comparisons (Duncan multiple range test), the performance of TgPC3 ON mice was severely impaired (p < 0.05) in comparison to both TgPC3 OFF and WT mice, whereas no significant differences were found between these latter groups (Figure 3D, left).
Given that the above behavioral tests were conducted in TgPC3 ON mice in the presence of an active transgene, we wished to verify whether the observed behavioral phenotype could be related to indirect effects of PC3 over-expression in nestin-positive cells on neighboring, mature granule cells. Thus, we tested spatial learning in the Morris water maze using TgPC3 ON mice in which the transgene was activated at P30 and inactivated at P95, i.e., prior to behavioral tests. To this aim, doxycycline treatment was suspended at P90, to allow 5 d of metabolization of the residual levels (Figure S6A). The performance of TgPC3 ON mice was significantly impaired in learning and reversal learning, as well as in the first and in the second probe trial, compared to WT mice (Figure S6B and S6C). Thus, we can exclude indirect cell-mediated effects by nestin-positive cells expressing the PC3 transgene during the behavioral tests.
Altogether, these behavioral results indicate that, in PC3 transgenic mice, the accelerated differentiation of dentate gyrus adult progenitors dramatically impairs learning and memory of spatial information.
To verify whether genetically altered neurogenesis could impair learning performance in previously experienced tasks, TgPC3 OFF mice and their controls were treated with doxycycline (doxy; TgPC3 OFF → ON and WT-doxy, respectively) following the first experimental phase. In addition, for this set of experiments, the gene was activated for a shorter period of time (see time schedule in Figure 3A). During treatment, both TgPC3 OFF → ON and WT-doxy mice were submitted to a second session of spatial learning and memory tasks. A new piece of information was introduced in the Morris water maze by moving the hidden platform to a different location. By contrast, in the radial maze, all mice were tested with the same procedure used in the first experimental phase. This differential approach was used to test for the ability of the mice to use previously acquired information.
In the Morris water maze, TgPC3 OFF → ON mice were impaired in learning the novel position of the platform in comparison to controls (Figure 3B, trials 31–42; F(1,17) = 5.25; p < 0.05). In the probe test, TgPC3 OFF → ON mice spent a significantly smaller amount of time in the target quadrant in comparison to WT-doxy mice (Figure 3C, probe 3; p < 0.05). Similarly, in the radial maze, TgPC3 OFF → ON mice committed a significantly greater number of errors compared to WT-doxy mice (Figure 3D, right; F(1,17) = 7.79; p < 0.05).
These results indicate that the anticipated differentiation of dentate gyrus neural progenitors produces a remarkable impairment in tasks properly solved by transgenic mice before the PC3 transgene activation. Thus, the normal differentiation timing of dentate gyrus neural precursors emerges as a major constraint to learning performance, even in a previously experienced spatial setting, either in the presence or absence of new information.
We then investigated the effects on memory of the PC3 transgene activation in a fear-related learning task, namely contextual and cued fear conditioning, involving mainly the hippocampus and amygdala, respectively [59,60]. In this task, immobility (freezing), a natural reaction elicited in mice by aversive stimuli, was recorded and considered a measure of fear memory. Animals were preliminarily trained in the conditioning box (chamber A), in which a conditioned stimulus (CS) and unconditioned stimulus (US; see Materials and Methods) were paired. During the training, no significant effect of the genotype was observed among groups in the level of freezing, both in the pre-CS (F(2,34) = 1.47; p = not significant) and the post-US (F(2,34) = 0.91; p = not significant) phases, and all the groups of mice reacted alike to the US (F(1,34) = 115.52; p = 0.0001; Figure 4A). In the contextual test, analysis of the percentage of freezing showed significant differences between groups (F(2,34) = 17.68; p < 0.0001). As shown by post-hoc comparisons (Duncan multiple range test), TgPC3 ON mice spent a significantly smaller amount of time (p < 0.0001) in freezing behavior compared to both TgPC3 OFF and WT mice, whereas no significant differences were found between TgPC3 OFF and WT mice (Figure 4B, left). Two hours after the contextual test, cued memory was assessed by the administration of the CS in a different setting (chamber B). No significant differences between groups emerged (F(2,34) = 0.15; p = not significant), whereas a significant effect of the CS was observed (F(1,34) = 318.04; p < 0.0001), that is to say that, in the novel environment, all groups of mice exhibited an appreciable level of freezing only during the CS administration (Figure 4C, left).
In the following experimental phase, TgPC3 OFF → ON and WT-doxy mice were submitted to a second session of fear conditioning, with an unvaried procedure (i.e., training in the same box used in the first experimental phase, followed by contextual and cued tests; see also Materials and Methods). Transgene activation significantly impaired the contextual memory (Figure 4B, right; F(1,17) = 6.53; p < 0.05), whereas no significant differences between transgenic and control mice were observed in the cued conditioning test (Figure 4C, right; F(1,17) = 2.72; p = not significant).
As in spatial tasks, these results further indicate that, also in a fear-related task, genetically altered neurogenesis produces a memory impairment for the contextual features mice had already experienced. Both in the first and the second behavioral session, the deficit was selective for memory contents processed mainly by the hippocampus (i.e., those related to spatial information), sparing the components of the experience whose processing appears to be peculiar to the amygdala.
Long-term changes in synaptic strength have been described in several brain areas. In particular, hippocampal long-term potentiation (LTP) of glutamatergic synaptic transmission is believed to be the synaptic correlate of several forms of learning and memory (for reviews, see [61,62]). To test whether the above-described deficits in memory tasks caused by early expression of the PC3 gene correlated with a change in long-term synaptic transmission, we recorded field excitatory postsynaptic potentials (fEPSPs) in the outer molecular layer of the dentate gyrus while stimulating perforant path afferents in WT, TgPC3 OFF, and TgPC3 ON mice. Robust LTP of excitatory synaptic transmission could be evoked by high-frequency stimulations (HFS: four trains of 100 stimuli at 100 Hz) in WT animals. Overall, fEPSP slopes were 1.5 ± 0.04 larger when measured 50 min after LTP induction and compared with baseline levels (n = 13 slices, 7 mice; p < 0.001; Figure 5A). Similar levels of synaptic potentiation were observed in TgPC3 OFF animals (normalized fEPSP slope at 50 min after HFS = 1.62 ± 0.12, n = 9 slices, 7 mice, p > 0.2 when compared with WT, Figure 5A). In TgPC3 ON mice, HFS evoked LTP consistently (normalized fEPSP slope was 1.2 ± 0.04, n = 12 slices, 7 mice, p < 0.001), but we found that the level of potentiation was significantly smaller when compared to WT and TgPC3 OFF mice (p < 0.001; Figure 5A). This effect on LTP was not due to overall changes in synaptic transmission, as input-output curves (fEPSP slopes versus increasing afferent fiber volley amplitudes) were similar in WT, TgPC3 OFF, and TgPC3 ON mice (Figure 5C).
Hippocampal adult neurogenesis is selectively localized in the dentate gyrus. We thus examined whether LTP deficits recorded in the TgPC3 ON mice were selective for dentate gyrus synapses. We then recorded fEPSPs in the CA1 area of the hippocampus while stimulating the Schaffer collateral fibers in stratum radiatum. A single train of HFS induced robust LTP in WT (n = 7 slices, 2 mice), TgPC3 OFF (n = 8 slices, 2 mice), and TgPC3 ON mice (n = 5 slices, 2 mice) (Figure 5B). No significant differences were observed between all three conditions tested (p > 0.3 in all cases).
As a whole, these results indicate that the accelerated differentiation by early PC3 expression during adult neurogenesis affects synaptic plasticity selectively in the dentate gyrus, without altering CA3–CA1 hippocampal synapses.
The decrease of LTP evoked in the dentate gyrus of mice with an activated transgene prompted us to ask whether this decrease was correlated with the function of new neurons in memory circuits. It has been shown that newly generated neurons of the dentate gyrus are progressively integrated into spatial memory-related circuits after 4–6 wk of age. Such evidence was obtained by measuring c-fos expression, whose activation specifically occurs in dentate gyrus neurons of mice that have undergone spatial memory training and is thus correlated with the recruitment of new neurons into spatial memory circuits [21]. Thus, we wished to define whether the anticipated differentiation of new granule neurons of dentate gyrus had an effect on the time at which they became active in memory networks and on the extent of their activation. To this end, we measured the number of new neurons at 2, 4, and 6 wk of age that become activated by a spatial memory test. Mice, either control (TgPC3 OFF) or with the PC3 transgene activated 60 d before (TgPC3 ON), were treated with five daily injections of BrdU and then trained in different groups in the Morris water maze test 2, 4, and 6 wk later (Figure 6A). Mice were killed 1.5 h following the tests, and the number of new neurons, identified by positivity to BrdU and NeuN, was compared to the number of new neurons integrated into memory circuits, identified by c-fos, BrdU, and NeuN concomitant expression (Figure 6B). In TgPC3 OFF mice, the fraction of new neurons expressing c-fos was null 2 wk after birth (i.e., after BrdU injections) but reached about 2% after 4 wk and increased to about 5% after 6 wk (Figure 6C). These data are in agreement with those obtained by Frankland's group, and they indicate a progressive functional activation of new neurons within spatial circuits, correlated to their age and maturation [21]. In contrast, in mice with the PC3 transgene activated, virtually no 2-, 4-, or 6-wk-old new neurons expressing c-fos were detectable, clearly indicating that new neurons did not undergo activation by spatial training (Figure 6C). The lack of induction of c-fos expression by spatial training in 4- and 6-wk-old neurons of TgPC3 ON mice corresponded to a significant learning deficit of these mice in the Morris water maze task (Figure S7A and S7B).
Moreover, in the dentate gyrus, the ratio between the total number of new neurons generated (i.e., BrdU+/NeuN+ cells) and the total number of neurons (NeuN+ cells) did not significantly differ in the same TgPC3 ON and TgPC3 OFF mice groups analyzed 2, 4, and 6 wk after BrdU injection (Figure 6D). Also the absolute numbers of BrdU+/NeuN+ cells in the dentate gyrus of the 2-, 4-, and 6-wk groups were equivalent in the TgPC3 OFF and TgPC3 ON mice (unpublished data). These results are consistent with data shown above (Figure 1K) and conclusively indicate that the lack of activated new neurons observed in TgPC3 ON mice (Figure 6C) is a genuine effect, not caused by impaired generation of new neurons.
It is notable that c-fos was expressed in the neuronal population of the dentate gyrus to a similar extent in all groups of control mice, as measured by the ratio between total number of activated neurons (c-fos+/NeuN+ cells) and total number of neurons (NeuN+ cells), indicating the existence of a basal level of activation in the neuronal population. In contrast, in TgPC3 ON mice, the ratio between activated and total number of neurons was progressively reduced in the 2-, 4-, and 6-wk groups, with a significant difference in the 4- and 6-wk groups (Figure 6E). One possibility suggested by this finding is that the failure to generate new activated neurons in dentate gyrus in response to spatial behavioral training, observed above (Figure 6C), leads to a progressive reduction in the total number of activated neurons.
The transgenic conditional model used in this study offers the possibility of analyzing the hippocampus-dependent learning and memory processes after a selective enhancement of the maturation rate of new adult-generated hippocampal neurons. In contrast, previous studies eliminated new progenitor cells and neurons from the adult hippocampus, using either antimitotic toxins, x-ray irradiation or virus-activated pro-drugs [24,26,28]. Our analysis of the dentate gyrus showed that activation of the PC3 transgene driven by nestin promoter led to a great increase in fully differentiated, 1–5-d-old, adult-generated hippocampal granule neurons (stage 6; positive for NeuN and BrdU), accompanied by a parallel decrease in stem cells and putative transiently amplifying progenitor cells (type-1 and −2) of the same age. These results clearly indicate an accelerated transition toward terminal differentiation of the newborn progenitor cells expressing the PC3 transgene. Moreover, the analysis of fully differentiated hippocampal granule neurons of about 4 wk of age showed that the final number of neurons produced is the same in both TgPC3 ON and control mice. As a whole, these findings indicate that the rate of maturation of new neurons is accelerated by PC3, without any effect on the rate of neurogenesis—i.e., the birth rate of new neurons—and thus on the total number of new neurons generated.
Such PC3-dependent anticipated differentiation of progenitor cells is in line with the previously observed, intrinsic differentiative properties of PC3 on neural precursors [41] and might result from the prevalence, in progenitor cells type-1 and type-2, of the asymmetric neurogenic type of division, with which the expression of PC3 is associated [37–39,63].
The enhanced rate of maturation of the whole population of progenitor cells of the adult dentate gyrus of TgPC3 ON mice was clearly associated with a severe impairment in hippocampus-dependent learning and memory, as indicated by the performance of transgenic mice in different hippocampus-related behavioral tasks. In fact, learning and memory deficits were observed both in the Morris water maze and in the radial maze test. Furthermore, a significant memory deficit was also observed in the contextual fear conditioning test, in which both the hippocampus and the amygdala mediate the association between context and the aversive stimulus [59]. Conversely, no significant effect was observed in the cued version of the task, mainly characterized by a greater involvement of the amygdala formation (reviewed in [64–66]).
Such selective impairment of hippocampal spatial learning raises questions about the role played by functional and morphological modifications occurring in newborn neurons of our mouse model.
A relevant functional change observed in mice with maturation of progenitor cells enhanced by PC3 is the decrease of LTP evoked in the dentate gyrus, an effect that appears to be specific, because no alteration of synaptic plasticity in the CA1 region is seen. It is widely believed that LTP and depression of synaptic transmission underlie several forms of learning and memory [61,62], and are often accompanied by structural changes of dendritic spines (reviewed in [67–69]). The earliest form of excitability is shown by nestin+ non-radial precursor cells (type-2 cells; [18]), and LTP begins to be observed in adult-generated, young (1–3-wk-old) hippocampal neurons, where it is elicited more easily than in mature, post-mitotic dentate gyrus neurons [13,16]. Thus, the reduced LTP that we observe should not depend on the reduced number of progenitor cells type-2 and type-3 present in mice with activated transgene, because they are too immature to generate LTP. Rather, the reduction of LTP may depend on new neurons that are ≥2 wks old, whose numbers are unchanged in our mouse model but whose dendritic arborization and spine density are markedly decreased. Thus, bypassing the early stages of differentiation during adult neurogenesis in neurons generated following transgene activation strongly affects the ability of dentate gyrus circuits to integrate synaptic transmission. This is further suggested by the observation that basic synaptic transmission is unchanged, as input-output curve analysis shows.
A second functional change that we observe in the dentate gyrus of transgenic mice is the lack of activation of new, fully differentiated neurons, 4–6 wk of age, in response to learning, as indicated by the absence of induction of c-fos expression after training. Indeed, the activation of adult-generated dentate gyrus neurons in spatial memory circuits has been linked to the induction of c-fos expression [21,70]. This expression is elicited in adult-generated dentate gyrus neurons, 4–6 wk of age, in mice trained in the Morris water maze, indicating that at this developmental stage the dentate gyrus neuron develops functions critical for its activation and recruitment into spatial memory circuits [21]. Hence, the absence of c-fos activation in 4–6-wk-old neurons that have expressed PC3 at birth suggests that these neurons do not attain a functional state in hippocampal circuits.
The functional alterations observed by us in new, adult-generated dentate gyrus neurons that are prematurely differentiated may find a structural correlate in the morphological alterations, most evident in 4-wk-old neurons identified after infection with a GFP-expressing retrovirus, i.e., reduced dendritic length and spine density.
We can assume that the birth datings of a neuron by BrdU and retroviral infection are comparable, and we consider that a neuron analyzed 4 wk post-infection is representative of the level of mid-late maturation, when the main neuronal structures are developed [19].
Altogether, it is plausible that the faster attainment of the terminally differentiated state, observed in adult-generated dentate gyrus neurons of mice with active transgene, may (following an initial faster development of the dendritic tree seen in neurons at 7 dpi) lead to a premature end of the developmental stages relative to the establishment and growth of the dendritic tree during the second and third week after birth. Only at later stages, i.e., in 10-wk-old neurons, emerges a tendency to recover, at least in part, normal values of spine density.
Neuronal structure and functional activity are correlated [71–75]. Indeed, the reduced dendritic structure affects spine formation (the latter being spatially constrained by dendrites) and ultimately the afferent synaptic input, since each spine receives one synaptic bouton [76]. As a result, the neuron has reduced capability to activate and integrate memory patterns, as our electrophysiological and behavioral data indicate.
In conclusion, the selective anticipation of differentiation of new, adult-generated dentate gyrus neurons induced in the PC3 transgenic mouse model reveals the critical role of developmental timing, even before terminal mitosis, for the generation of neurons that are fully able to integrate new spatial memories.
As our data indicate, the number of new neurons necessary to integrate new spatial information appears relatively small. In fact, the number of new dentate gyrus neurons generated during the shortest period of activation of the transgene, sufficient to prevent new spatial learning (about 20,000, i.e., 800–1,000 per day multiplied by 22 days elapsed before behavioral session 2), should not exceed 5%–6% of the total number of neurons in the dentate gyrus (less than one half million).
It is worth noting that the progenitor cells born before or during the first session of spatial training in TgPC3 OFF mice will normally differentiate without being affected by activation of the transgene during session 2, because in progenitor cells, the nestin promoter becomes physiologically inactive within 1 wk after their birth [7]. Thus, in TgPC3 OFF → ON mice, only the new neurons born in the 3–4-wk period of transgene activation preceding and during the second behavioral session appear to be responsible for the spatial learning and memory deficits observed.
Moreover, given that TgPC3 OFF → ON mice were significantly impaired in spatial tasks since the beginning of the second session, the premature differentiation of adult-generated hippocampal progenitor cells could prevent not only learning new spatial tasks but also the use of information acquired previously in normal condition, i.e., during session 1 before activation of the transgene.
Therefore, the picture emerging from our data is that the correct differentiation and integration into existing memory circuits of a relatively small population of new adult-generated neurons not more than 4 wk old are crucial not only for spatial learning but also, notably, for the use of memories consolidated in tasks previously performed. Conversely, neurons older than 4 wk appear less efficient. This unique role of younger adult-generated neurons in the maintenance of key memory functions may derive from their proposed ability to link temporally proximal events [23], and also entails the need for continuous renewal of new neurons as soon as they terminally differentiate. The late activation (i.e., c-fos expression) observed after training in 4–6-wk-old neurons by us and by the group of Frankland [21], might thus be related to subsequent steps of integration into spatial learning circuits.
The significant impairment of the hippocampal function observed in our model, i.e., the deficit of spatial learning in the Morris water and radial maze tests accompanied by deficit of dentate gyrus LTP, is not frequently observed in studies of spatial memory (e.g., see [77]).
It is plausible that the ablation of new neurons, while eliminating their functional contribution, may at least in some case trigger compensatory processes, e.g., increased proliferation of the surviving progenitors, or a plastic reassembly of the cytoarchitecture of circuits. In our model, the new neurons exist, albeit with a reduced functionality; thus, as our data strongly suggest, no compensatory processes are activated. Moreover, these altered new neurons appear to establish connections with the existing neurons and are thus integrated in previously-formed circuits. It might be the same as expanding a circuit by interspersing new elements with lower functionality. Therefore this model, assembling within spatial circuits new neurons generated in normal number but with altered maturation, can be considered the equivalent of an anatomical dominant-negative, and may give the advantage over ablative strategies of magnifying the ability of new neurons—variable depending on their maturation state—to integrate and/or recall traces of temporally proximal events from existing circuits.
The bitransgenic nestin-rtTA/TRE-PC3 mouse line is the progeny of two mouse lines, each carrying a transgene: nestin-rtTA transgene, encoding the tetracycline-regulated TransActivator driven by the rat nestin promoter, and TRE-PC3 transgene, carrying the PC3 coding region under control of the Tetracycline Responsive Elements. The nestin-rtTA and TRE-PC3 transgenic mouse lines were generated previously [78,41], exploiting the tet-on system to control the activation of the PC3 transgene [79], and they were maintained in heterozygosity in FVB background or in homozygosity in BDF1 (C57BL/6 x DBA/2) background, respectively. The bitransgenic nestin-rtTA/TRE-PC3 mice used for experiments were isogenic, having been previously interbred for six or more generations. In bitransgenic mice, the TransActivator protein produced by the nestin-rtTA transgene binds and activates TRE-PC3 in the presence of doxycycline, consequently inducing the expression of PC3 transgene. Genotypization of bitransgenic nestin-rtTA/TRE-PC3 mice was performed as described [41]. Animals were housed in standard breeding cages under a 12-h light-dark schedule at a constant temperature of 21 °C and underwent behavioral testing during the second half of the light period (between 2:00 and 5:00 p.m.) in sound-insulated rooms. All procedures involving mice were completed in accordance with the Istituto Superiore di Sanità (Italian Ministry of Health) and current European (directive 86/609/ECC) Ethical Committee guidelines.
Bitransgenic nestin-rtTA/TRE-PC3 mice (named TgPC3 throughout this report) were treated for immunohistochemistry according to the following experimental protocol.
The PC3 transgene was activated in TgPC3 mice P30 (termed TgPC3 ON mice) by doxycycline hydrochloride (2 mg/ml; MP Biomedicals) supplied to mice in drinking water containing 2.5% sucrose; at P90–94 mice received five daily bromodeoxyuridine (BrdU; 95 mg/kg i.p.) injections to detect dividing neurons. Immunohistochemical analyses were performed after BrdU treatment, at P95 and at P116. TgPC3 mice untreated with doxycycline (termed TgPC3 OFF mice) and WT mice (i.e., the progeny of crosses between nestin-rtTA and TRE-PC3 transgenic mice, in which the TRE-PC3 transgene was not inherited) were used as controls.
Brains were collected after transcardiac perfusion with 4% paraformaldehyde (PFA) in PBS–DEPC and kept overnight in PFA. Thereafter, brains were equilibrated in sucrose 30% and cryopreserved at −80 °C.
Immunohistochemistry was performed on serial sections cut transversely at 20-μm thickness at −25 °C in a cryostat from brains embedded in Tissue-Tek OCT (Sakura). Sections were then processed immunohistochemically for multiple labeling with BrdU and other cellular markers using fluorescent methods. BrdU incorporation was visualized by denaturing DNA through pre-treatment of sections with 2 N HCl 45 min at 37 °C followed by 0.1 M sodium borate buffer pH 8.5 for 10 min. Sections were then incubated with a rat monoclonal antibody against BrdU (Serotech; MCA2060; 1:150) together with other primary antibodies, as indicated: either mouse monoclonals raised against nestin (Chemicon International; MAB353; 1:150) and NeuN (Chemicon International; MAB377; 1:100), or rabbit polyclonal antibodies against Glial fibrillary acidic protein (GFAP; Promega Corporation; G560A; 1:150) and c-fos (Chemicon International; Ab-5 PC38T; 1:500), or goat polyclonal antibodies recognizing Doublecortin (DCX) (Santa Cruz Biotechnology; SC-8066; 1:200) or NeuroD1 (R&D Systems; AF2746; 1:100). Another primary antibody used was a rabbit monoclonal antibody against Ki67 (LabVision Corporation; SP6; 1:100). These antigens were visualized with either TRITC (tetramethylrhodamine isothiocyanate)-conjugated donkey anti-rat (Jackson ImmunoResearch; BrdU), or TRITC-conjugated goat anti-rabbit (Sigma; Ki67), or Cy2-conjugated donkey anti-rabbit (Jackson ImmunoResearch; GFAP), or donkey anti-goat Cy2-conjugated (Jackson ImmunoResearch; DCX, NeuroD1) or also with donkey anti-mouse Alexa 647 (Invitrogen; nestin, NeuN, c-fos) secondary antibodies.
Terminal deoxynucleotidyl transferase-mediated biotinylated UTP nick end labeling (TUNEL) [80] was performed on cryostat sections using the in situ cell death detection kit (Roche Products), according to the instructions of the manufacturer. Apoptotic nuclei were visualized with 0.5% 3,3'-diaminobenzidine (DAB).
Images of the immunostained sections were obtained by laser scanning confocal microscopy using a TCS SP5 microscope (Leica Microsystem). All analyses were performed in sequential scanning mode to rule out cross-bleeding between channels.
About 200 transverse sections spaced 20 μm apart and comprising the entire hippocampus were obtained from each brain; about one-in-ten series of sections (200 μm apart) were analyzed to count cells expressing the indicated marker throughout the whole rostro-caudal extent of the dentate gyrus. The total estimated number of cells within the dentate gyrus, positive for each of the indicated markers or combination of markers, was obtained by multiplying the average number of positive cells per section by the total number of 20-μm sections comprising the entire dentate gyrus [81,2,21]. At least three animals per group were analyzed. c-fos was analyzed in about one-in-ten series of 40-μm free-floating sections (400 μm apart).
Stereological analysis of volumes and of the absolute number of granule cells in the dentate gyrus was performed, analyzing every sixth section in a series of 40- μm coronal sections (thus spaced 240 μm). Total cell number was obtained according to the optical disector principle, by systematic sampling of unbiased counting frames of 15-μm side in each section. Nuclei considered in the count (identified by Hoechst 33258 staining) were those appearing throughout the different focal planes of each section, excluding those nuclei that intersected the exclusion boundaries of the counting frame, as defined by the optical disector principle [82]. To obtain the absolute cell number, the average cell number per disector volume (Nv; the disector volume being 15 × 15 × 40 μm3) was multiplied by the reference volume (i.e., the total volume of dentate gyrus; [82]). The reference volume was determined by summing the traced areas of dentate gyrus (or hippocampus) and multiplying this result by the distance between the sections analyzed (240 μm).
Measurements of positive cells and areas were obtained by computer-assisted analysis using the I.A.S. software (Delta Systems).
To reveal the β-galactosidase activity of the β-geo reporter gene fused to the nestin-rtTA transgene [78], transverse sections from brains fixed as described above were post-fixed with 0.2% glutaraldehyde in PBS for 10 min, then washed three times for 5 min in lacZ wash buffer (1 M MgCl2, 1% sodium deoxycholate, 2% NP40 in PBS). Sections were then stained overnight at 30 °C with lacZ stain [lacZ wash buffer, 0.21% K4Fe(CN)6.3H2O, 0.16% K3Fe(CN)6, 25 mg/ml of 5-Bromo-4-Chloro-3-indolyl- β-D-galactopyranoside (X-Gal)], washed three times 5 min with PBS and mounted.
The murine Moloney leukemia virus-based retroviral vector CAG-GFP [19] was used to infect only dividing cells at the moment of in vivo delivery. Retroviruses were propagated by transiently cotransfecting CAG-GFP with pHCMV-G vector (which expresses the VSV-G protein; [83]) in the packaging cell line Phoenix (human embryonic kidney cell line stably expressing the gag and pol proteins of Moloney murine leukemia virus; [84]). Cells at about 90% confluence in 90-mm dishes were transfected with 11.5 μg of CAG-GFP and with 13.5 μg of pHCMV-G, using calcium phosphate precipitation. Virus-containing supernatant was harvested 36, 48, and 60 h after the start of transfection. Frozen stocks were pooled and the virus was concentrated by centrifugation for 1.5 h in a Hitachi RPS40T rotor at 25,000 rpm. The concentrated virus solution (108 pfu/ml) was infused (1.5 μl at 0.32 ml/min) by stereotaxic surgery into the right and left dentate gyrus of anesthetized P95 transgenic mice (anteroposterior = −2 mm from bregma; lateral = ±1.5 mm; ventral = 2 mm). The animal protocols were approved by the Istituto Superiore Sanità, Rome, Italy.
Dendritic analysis of GFP-positive neurons was performed by acquiring z-series of 15–25 optical sections at 1–1.5 μm of interval with a 40X oil lens, with the confocal system TCS SP5 (Leica Microsystem). Two-dimensional projections at maximum intensity of each z-series were generated with the LAS AF software platform (Leica Microsystem) in the TIFF format, and files were imported in the I.A.S. software (Delta Systems) to measure dendritic length. The number of branching points was counted manually in the same images. For each data point, 20–30 cells from two mice were analyzed (either control or with activated transgene). From the same GFP-positive neurons, the spines present on dendritic processes were imaged by acquiring z-series of 25–35 optical sections at 0.5 μm of interval with a 63X apochromatic oil lens, and a digital zoom of 3. The number of spines was counted manually on two-dimensional projections obtained by the LAS AF software. The linear spine density was then calculated by dividing the total number of spines by the length of the corresponding dendritic process.
WT, TgPC3 OFF, and TgPC3 ON mice, 95 d old, were used for electrophysiological recordings. TgPC3 ON mice were used after animals were exposed for two months to doxycycline hydrochloride. Mice were deeply anesthetized with isofluorane inhalation and decapitated, and the brains were removed and immersed in cold “cutting” solution (4 °C) containing (in mM): 234 sucrose, 11 glucose, 24 NaHCO3, 2.5 KCl, 1.25 NaH2PO4, 10 MgSO4, and 0.5 CaCl2, gassed with 95% O2/5% CO2. Coronal slices (300 μm) were cut with a vibratome from a block of brain containing the dorsal hippocampus and then incubated in oxygenated artificial cerebrospinal fluid (aCSF) containing (in mM): 126 NaCl, 26 NaHCO3, 2.5 KCl, 1.25 NaH2PO4, 2 MgSO4, 2 CaCl2, and 10 glucose, gassed with 95% O2/5% CO2; pH 7.4, initially at 35 °C for 1 h, and subsequently at room temperature. Slices were then transferred to a submersed recording chamber and maintained at 32 ± 1 °C while being continuously perfused by fresh and oxygenated aCSF at a rate of ∼2–3 ml/min. Recordings were performed in the continuous presence of 100 μM picrotoxin to block inhibitory GABAergic transmission. Monopolar stimulating and recording electrodes consisted of glass pipettes (0.5–1 Mμ) filled with aCSF and placed in the middle of the outer molecular layer of the dentate gyrus or in the middle of the stratum radiatum. Field excitatory post-synaptic potentials (fEPSPs) were amplified using a Multiclamp 700B patch-clamp amplifier (Molecular Devices). A Digidata 1320 digitizer and PClamp9 (Molecular Devices) were used for data acquisition, analysis, and generation of stimuli. Input-output (I/O) curves were generated before inducing LTP. For I/O curves, stimulus intensities were adjusted to have afferent volleys of 50, 100, 150, 200 and 250 μV in all tested slices, and the resulting fEPSP slopes were calculated and averaged across 3–5 sweeps. For LTP experiments, fEPSPs were stimulated at 0.1 Hz. If a stable baseline of at least 10 min was achieved, LTP was induced by four trains of 100 stimuli at 100 Hz repeated every 20 s for recordings in the dentate gyrus and one train of 100 stimuli at 100 Hz in the CA1 area. Results are presented as means ± SEM. Data were averaged in 0.5-min bins.
All mice were tested in an open field and a plus maze [85,86] to assess locomotor abilities and anxiety levels. No statistically significant differences were found among the groups in both tasks (unpublished data).
The behavioral experiments were carried out in two phases, according to the schedule shown in Figure 3A. In the first phase, the following groups of mice, aged 3 mo, were used: TgPC3 ON (n = 10), in which the PC3 transgene was activated at P30 by doxycycline administration; TgPC3 OFF (n = 8), in which doxycycline was not administered so that the PC3 transgene was not activated; and WT (n = 12 treated with doxycycline; n = 11 untreated). At the end of the first phase, TgPC3 OFF (n = 8) and untreated control mice (WT, n = 11) were administered with doxycycline and, after day 22 of treatment, these two mice groups (labeled as TgPC3 OFF → ON and WT-doxy, respectively) were subjected to a second behavioral testing session.
The Morris water maze [54,55] was carried out in a circular swimming pool of 1.3 m in diameter, filled with opaque water at a temperature of 25 ± 1 °C and located in a room containing prominent extra-maze cues. A hidden, 15-cm diameter platform was used. In the first experimental phase, the training consisted of 18 trials (six trials per day, lasting a maximum of 60 s, with an intertrial interval of 30 min), with the platform left in the same position. After 3 d of learning, the platform was moved to the opposite position and reversal learning was monitored for two additional days. Probe tests (60 s) were carried out 24 h after both learning and reversal learning by removing the platform from the pool. In the second phase, TgPC3 OFF → ON and WT-doxy mice were submitted to a further session of spatial learning lasting 2 d, during which the platform was located in a different position from those used in the previous experimental phase. A probe test was carried out 24 h after learning. Behavior was evaluated by EthoVision software (Noldus Information Technology).
For the eight-arm radial maze [57,58], mice were singly housed, with water provided ad libitum, and gradually reduced to 85% of their free-feeding body weight. Throughout the experiment, mice were maintained at their reduced weight by being fed with a premeasured amount of food on each day. The maze apparatus was constructed of gray plastic maze, with eight identical arms radiating 37 cm from an octagonal starting platform (side 7 cm). On each training trial, a 20-mg food pellet was placed at the end of each arm (baits were not replaced) and the animal was placed on the central platform, facing a randomly selected direction. In both the first and second experimental phases, all groups of mice were submitted to one trial per day for 10 training days; each daily trial ended when eight choices were made or 15 min had elapsed. An arm choice was defined as placement of all paws on a maze arm. An error was noted when an animal entered a previously visited arm.
For the contextual and cued fear conditioning [87], the experiments were carried out in two different chambers. In both the first and second experimental phases, all mice were trained in conditioning chamber A (26 × 22 × 18 cm), made of transparent Plexiglas with a grid metal floor and located in a sound-insulated box lighted by a white tensor lamp (60 W). After an acclimatizing period lasting 120 s, a 30-s tone was administered (CS; 3 kHz, 80 dB). During the last 2 s of tone presentation, a foot-shock was delivered (US; 0.5 mA). Both CS and US ended simultaneously. Mice were left in the conditioning chamber for a further period of 30 s and then returned to their home cage. For the contextual conditioning test, 24 h after training mice were placed in the same chamber for 5 min. Two hours after the contextual test, mice were tested for cued conditioning in chamber B, made of black Plexiglas, with a floor of triangular shape, lighted by a blue tensor lamp (60 W) and perfumed by a vanilla essence diffuser. The test lasted 6 min, with CS administered during the last 3 min.
One-way ANOVA was used to analyze the c-fos-expressing neurons, the dendritic length, branching points, and spine density, as well as the levels of freezing in the contextual and cued fear conditioning and the electrophysiological data originating from slices from the different animal groups. Morris water maze and radial maze results were analyzed by two-way ANOVA. Individual between-group comparisons, where appropriate, were carried out by Fisher's PLSD post-hoc or Duncan multiple range test. Student's t-test was used to analyze the number of neurons and hippocampal volumes as well as data obtained in the same slices after LTP-inducing stimuli.
|
10.1371/journal.ppat.1001065 | Reciprocal Analysis of Francisella novicida Infections of a Drosophila melanogaster Model Reveal Host-Pathogen Conflicts Mediated by Reactive Oxygen and imd-Regulated Innate Immune Response | The survival of a bacterial pathogen within a host depends upon its ability to outmaneuver the host immune response. Thus, mutant pathogens provide a useful tool for dissecting host-pathogen relationships, as the strategies the microbe has evolved to counteract immunity reveal a host's immune mechanisms. In this study, we examined the pathogen Francisella novicida and identified new bacterial virulence factors that interact with different parts of the Drosophila melanogaster innate immune system. We performed a genome-wide screen to identify F. novicida genes required for growth and survival within the fly and identified a set of 149 negatively selected mutants. Among these, we identified a class of genes including the transcription factor oxyR, and the DNA repair proteins uvrB, recB, and ruvC that help F. novicida resist oxidative stress. We determined that these bacterial genes are virulence factors that allow F. novicida to counteract the fly melanization immune response. We then performed a second in vivo screen to identify an additional subset of bacterial genes that interact specifically with the imd signaling pathway. Most of these mutants have decreased resistance to the antimicrobial peptide polymyxin B. Characterization of a mutation in the putative transglutaminase FTN_0869 produced a curious result that could not easily be explained using known Drosophila immune responses. By using an unbiased genetic screen, these studies provide a new view of the Drosophila immune response from the perspective of a pathogen. We show that two branches of the fly's immunity are important for fighting F. novicida infections in a model host: melanization and an imd-regulated immune response, and identify bacterial genes that specifically counteract these host responses. Our work suggests that there may be more to learn about the fly immune system, as not all of the phenotypes we observe can be readily explained by its interactions with known immune responses.
| To infect a host and survive attacks from the host immune system, bacteria express genes that allow them to counteract immune responses. By identifying these genes we can learn how hosts fight infections and how bacteria resist immune attacks. We identified Francisella novicida genes that interact with the fruit fly immune system by performing a genetic screen of bacterial mutants. We identified genes that when mutated cause the bacteria to grow poorly within the fly. Many of these genes were shown to help the bacteria survive oxidative stress, providing resistance to an immune response known as melanization. We then identified bacterial genes that interact with another branch of the immune system, the imd pathway, by performing a second screen in imd mutant flies. We identified bacterial mutants that cannot grow in wild-type flies but are rescued in imd mutants, indicating an interaction with this pathway. We followed up one example from this screen and found that mutants in the gene FTN_0869 grow normally inside cells, but cannot grow extracellularly. We found that this was due to being unable to resist previously unexplored aspects of the imd-regulated immune response that help fight off F. novicida infections.
| The outcome of any infection, whether it be clearance of the infecting pathogen, establishment of a persistent infection, or even death of the host is determined by contributions from both the host and the microbe [1]. To infect a susceptible host microbes express virulence factors, genes that allow the pathogen to invade, colonize, and survive within the host and cause essential pathology. In response, the host initiates an immune response that attempts to clear the pathogen and increase tolerance to the ensuing infection [2]. Consequently, in addition to genes that allow the bacteria to invade host cells and obtain nutrients from its host, a subset of the virulence factors expressed by the microbe must address the need of the bacteria to counteract the host immune response. Exploring this complex interplay between host and pathogen can help us to understand bacterial pathogenesis and define the contributions of the host immune system to bacterial virulence.
One way to explore the host-pathogen relationship is to apply model systems that allow us to dissect the genetics of both sides of the equation simultaneously in vivo [3]. In this study, we examine the host-pathogen interactions of Francisella novicida with an insect host, Drosophila melanogaster, and identify aspects of fly immunity that are most important for fighting F. novicida infection as well as the bacterial virulence factors that interact with each of these specific immune responses. Drosophila is used as a model of innate immunity because its simplicity and the ease at which it can be used for both forward and reverse genetics has allowed for the identification and characterization of aspects of the innate immune response that are conserved across evolution [4]–[6]. The fly immune response has three effector arms: an inducible antimicrobial peptide (AMP) response, a reactive oxygen response mediated by the activation of the enzyme phenoloxidase and the deposition of the pigment melanin, and a cellular immune response in which foreign invaders are phagocytosed by Drosophila hemocytes [7]. The humoral AMP response has been studied extensively and shown to be regulated by two pattern recognition pathways, Toll and imd which have been well-characterized and described, but the regulatory mechanisms of the melanization and cellular immune responses have only recently become the focus of increased interest and have not yet been fully elucidated [8].
Previous studies with pathogenic bacteria in the fly have shown that virulence factors that function in the vertebrate hosts of these pathogens are often required for the microbe to survive in the fly [6]. Recently, this has been shown to be true for the live vaccine strain (LVS) of the virulent pathogen Francisella tularensis, a Gram-negative facultative intracellular bacterium that is the causative agent of tularemia [9]. F. tularensis can infect a wide range of hosts that includes humans, but is more commonly associated with rabbits and small rodents. Unlike many of the pathogens used in previous fly studies, F. tularensis also has a documented arthropod vector phase in its natural life-cycle [10], [11]. While many arthropod-vectored pathogens can only be transmitted by a single specific species, F. tularensis is able to infect arthropods ranging from ticks to multiple species of mosquito to biting flies such as deerflies [12]–[14]. This makes the Drosophila model system particularly useful for studying both general F. tularensis host-pathogen interactions and insect-specific factors.
To date, the fly has primarily been used to dissect the function of known bacterial virulence factors or to demonstrate conservation between fly and vertebrate defenses. Less has been done to use forward genetic approaches in the microbe to identify virulence factors de novo. As immunologists we tend concentrate on known signaling pathways that have proven simple to study, are of interest to those working in vertebrates because they are conserved, and those that fit our idea of what the fly's immune response should be. In other words, experiments are driven by the interests of the scientists and not the pathogens. We took a more ecology-based approach and determined, from scratch, what this fly pathogen needs to kill the fly. The advantage of the fly is that it is inexpensive, rapid to use and has extensive genetic tools. The fly could be a useful tool for the identification of new virulence factors rather than a system used to study known factors.
To identify new virulence factors and examine their interactions with the fly immune system, we used the Francisella novicida strain U112 to infect flies and performed a genome-wide screen to identify factors required for growth and/or survival within the fly. Many of the genes that we identified are required for resistance to the Drosophila innate immune response, particularly to the oxidative stress produced by melanization. This is interesting in particular, because until recently, this pathway had been discarded as having no relevance to microbial infections in the fly [15]–[17]. Our work demonstrates that bacterial mutants can be used as probes of the host immune system to identify what aspects of innate immunity are most important in determining the outcome of an infection.
To identify additional interactions between the host immune system and bacterial virulence factors, we performed a second genetic screen in which we compared the ability of F. novicida mutants to grow in wild type flies to flies with an immunity defect known to affect fly survival in F. tularensis infections. These flies lack a functional imd signaling pathway, and we anticipated that this would reveal bacterial mechanisms necessary to circumvent the imd-regulated immune response. The imd pathway is primarily described as inducing antimicrobial peptides. Although we identified bacterial genes required to resist antimicrobial peptide killing in vitro, we were particularly intrigued to find a subset of genes that when mutated did not appear to change F. novicida sensitivity to the antimicrobial peptide we tested yet showed an altered phenotype in imd mutants. This suggests the possibility that there are previously undescribed immune mechanisms that are regulated by the imd pathway.
We infected flies with F. novicida strain U112, a wild type strain that causes virulent infections in its natural mouse and rabbit hosts but is not pathogenic to humans. Previous work using the Live Vaccine Strain (LVS) of F. tularensis demonstrated that F. tularensis grows to high bacterial levels within flies and causes a lethal infection [9]. Infections of the fly with the U112 strain were consistent with this result, although we found the U112 strain to be slightly more virulent in Drosophila than the LVS strain, killing the fly with a median time to death (MTD) of 5 days post-infection with 103 CFU (Figure 1A and Figure S1). As few as 5 CFU of F. novicida U112 were sufficient to kill the fly, and bacterial growth within the fly was exponential approaching 5×107 CFU per fly before they succumbed to the infection. Regardless of the initial dose, F. novicida infections quickly reached the same high levels of bacteria; colony counts in flies receiving a low dose caught up to the 108 fold higher dose within two days (Figure 1B and Figure S2).
In mammalian infections, F. tularensis is a facultative intracellular parasite that primarily grows within macrophages [18]. However, due to the extremely high bacterial levels observed within the fly we speculated that a large proportion of the bacteria were growing extracellularly. To test this idea, we performed gentamycin chase assays, infecting flies with F. novicida and subsequently injecting them with the non-cell permeable antibiotic gentamycin at various timepoints post infection. This assay kills off extracellular bacteria while leaving bacteria growing within cells intact and allowing us to determine whether the bacteria are growing intracellularly or extracellularly By 24 hours post infection, approximately 1×104 CFU per fly were found to be localized intracellularly. However, the average total population of bacteria in infected flies was at least 1×105 CFU per fly, demonstrating that a significant population of bacteria was located extracellularly. Over the course of the infection the total bacteria population increased to 1×107 CFU/fly, while the intracellular population remained steady, indicating that the extracellular population was responsible for the exponential increase in bacteria seen during fly infections (Figure 1C). This is roughly similar to what has been observed using the LVS strain, although the absolute numbers differ slightly possibly due to the differences in virulence between the two bacterial strains, differences in the host strains or environmental conditions [9].
Having demonstrated that the fly can support F. novicida growth we applied this model to the identification of bacterial genes that were important for establishing infection within the fly. Previous work has shown that many F. tularensis virulence factors that are required for growth in mammalian models are also essential in insect infections, including the Francisella Pathogenicity Island (FPI) genes iglB, iglC, and iglD and the transcription factor mglA [9], [19]. To expand upon this work, we sought to identify additional bacterial virulence factors and provide an opportunity to discover new biology using a forward genetic approach.
Using a transposon insertion library of F. novicida mutants we performed an in vivo screen for mutants with altered growth rates compared to wild-type bacteria using a Transposon Site Hybridization (TraSH) assay. Briefly, flies were infected with the pooled library and the infection was allowed to proceed for two days, at which point the bacterial populations in each fly were harvested. Genomic DNA was then purified from this population of bacteria and from the original input library that had not been subjected to the stresses found within the fly. RNA was amplified from the site of each transposon insertion and the two populations of RNAs were compared by microarray analysis. We identified mutants representing 149 F. novicida genes that were negatively selected with a false discovery rate (FDR) of 5%, indicating that these genes were essential for bacterial growth and survival within the fly (Table S1). 41 of these genes had previously been identified as negatively selected in a similar TraSH analysis performed with the same mutant library in an in vivo mouse model; this list includes the known virulence genes iglC, iglD, pdpA, and pdpB, and mglA [20]. In addition, 11 genes from our screen overlapped with data from a negative selection screen performed by Kraemer et al. using an inhalation model to observe respiratory infections in the mouse, and an additional 8 overlapped with a signature-tagged mutagenesis screen performed by Su et al. that also used an intranasal route of inoculation [21], [22]. The overlap between our TraSH assay and additional Francisella genome-wide screens is shown in Table 1 and Figure 2. Interestingly, no genes were identified in all four unique screens, although 7 genes were identified in our fly TraSH and at least two other screens. These genes were the hypothetical proteins FTN_1682 and FTN_1016, the RNA methyltransferase yibK, the ABC transporter yjjK, the amino acid antiporter FTN_0848, iglC and iglD. The degree of overlap between our fly screen and similar mouse screens both supports the hypothesis that our screen in Drosophila can be used to identify virulence factors that are conserved between insect and mammalian infections, and also presents the possibility of identifying virulence factors unique to the arthropod vector stage of the F. novicida life cycle.
To confirm the results of the TraSH screen, we tested 65 of the negatively selected mutants individually by competition assay, focusing on mutants that had particularly large decreases in abundance and/or showed homology to bacterial genes that we predicted could play a role in immune evasion or modulation. Transposon insertion mutants of each gene containing a kanamycin resistance cassette were tested to determine their ability to grow in competition with wild-type F. novicida. Each mutant was mixed with wild-type bacteria and injected into wild-type Drosophila at a 1∶1 ratio. Infected flies were incubated for 48 hours, at which point the bacteria in each fly were plated and the ratio of mutant to wild-type bacterial was determined. A competitive index of 1 represents an equal ratio of mutant to wild-type bacteria, while competitive indexes of less than one indicate that the mutant is attenuated. Mutants that were determined to be statistically significantly less than 1 by one sample t-test in a minimum of three repetitions were considered attenuated and are listed in Figure 2.
Mutants that were confirmed as negatively selected included kdpA, kdpC, kdpD, and kdpE, components of 2-component regulatory system that responds to turgor pressure, a number of genes known to be regulated by the virulence factor mglA, members of the Major Facilitator Superfamily (MFS) thought to be involved in substrate transport and drug resistance, multiple genes know to be involved in DNA repair, and a number of hypothetical proteins. (Figure 3A and data not shown) 56 of the mutants tested showed attenuated phenotypes by competition assay, with competitive indexes ranging from 0.6-0.007. The results of the competition assays indicated that the microarray data produced by the TraSH assays is useful for predicting negatively selected mutants but was somewhat non-quantitative; the degree of attenuation as measured by microarray analysis did not always correlate with the strength of the phenotype observed by competition assay.
One set of negatively selected mutants stood out as particularly interesting because they indicated a bacterial requirement for resistance to oxidative stress within the fly. These mutants included the LysR family transcriptional regulator oxyR. The homologue of this gene in E. coli has been shown to sense hydrogen peroxide and induce the transcription of downstream genes that provide protection against oxidative stress [23]. We also identified a number of genes that are essential for repairing damage to DNA such as that caused by reactive oxygen, including uvrA, uvrB, recB, ssb, mutM, and ruvC [24] (Figure 3B). This result is consistent with a screen for attenuated F. novicida U112 transposon mutants using an inhalation method of inoculation, which identified the DNA repair proteins recO and recA [21].
The fly's melanization immune response produces reactive oxygen as an effector and thus we hypothesized that these bacterial genes were involved in helping F. novicida resist melanization [25], [26]. To test this idea, we first performed in vitro disc diffusion assays to determine the sensitivity of each mutant to hydrogen peroxide (H2O2) and paraquat [27]. The oxyR mutants were extremely sensitive to both H2O2 and paraquat (Figure 3C (H2O2) and data not shown (paraquat)). In addition, all of the DNA damage repair mutants showed a significant degree of sensitivity to both reactive oxygen-producing agents (Figure 3D). Taken together, these data suggest that we identified a class of F. novicida genes that are essential for wild-type growth and survival within the fly, genes which help the bacteria to resist oxidative stress. Interestingly, 4 DNA repair genes, uvrA, recB, recO, and uvB, were identified in one screen of Francisella mutants in mice, suggesting that reactive oxygen species are a threat to the bacteria in mammalian infections as well.
To determine whether melanization is a critical factor limiting the growth of F. novicida, we performed competition assays using the oxyR mutant in CG3066 mutant flies. These flies do not induce melanization upon infection [17]. We found that the growth defect of oxyR with respect to wild type bacteria was rescued in non-melanizing CG3066 flies (Figure 3E). This supports the idea that melanization is the reactive oxygen producing immune response for limiting the growth of F. novicida in the fly. We therefore took our collection of negatively selected mutants and tested them for sensitivity to reactive oxygen. We found 25 mutants with increased sensitivity to oxidative stress and 2 with decreased sensitivity (Figure 2). Thus we were able to assign functions to these genes based on their behavior in a fly pathogenesis screen. This group of mutants makes up 45 percent of the mutants with attenuated growth phenotypes in the fly, demonstrating that oxidative stress mediated immunity is an important aspect of the fly's defenses against this pathogen and that an important class of F. novicida virulence factors in fly infections are genes that help the bacteria to counteract the effects of reactive oxygen.
Having demonstrated that the TraSH method was useful for identifying genes required for growth in the fly and that many of these mutants were involved in counteracting the oxidative stress response of the fly, we decided to expand our study to examine an additional immune pathway and look for similar interactions. We chose to focus on one of the most intensely studied aspects of the fly innate immune response, NF-κB signaling pathways. Drosophila has two well-characterized NF-κB pathways (Toll and imd) that are responsible for sensing the presence of microbes and inducing an immune response [7]. Previous work by others demonstrated that the imd pathway is an important component of the Drosophila innate immune response to the LVS strain of F. tularensis, while the Toll pathway is not [9]. To confirm this for F. novicida U112 strain, we infected flies with null mutations in Toll and imd pathway genes. Two separate alleles of imd, imd1 and the null allele imd10191, as well as mutants lacking the NF-κB homologue Relish showed significantly increased sensitivity to infections with the U112 strain; in contrast, mutants in the Toll pathway components Dif1 and dMyD88C03881 showed no significant difference compared to wild type (Figure 4A (imd alleles only, relish data not shown) and Figure S1). Therefore, we focused on F. novicida genes required to resist the fly's imd mediated response.
To identify such genes, we repeated our TraSH analysis, this time infecting imd mutant flies and compared the results to those found in wild-type flies. We identified 36 genes that appeared to be negatively selected in wild type flies and at least partially rescued in imd mutant flies (Table S2) Subsequent confirmation of these results by competition assay using transposon insertion mutants revealed a subset of 7 mutants that showed reproducible large rescue phenotypes in imd flies (Figure 4C). These genes were the orphan response regulator pmrA, the gene FTN_0889 which is a helix-turn-helix protein and putative transcriptional regulator, glpD which is an anaerobic glycerol-3-phosphate dehydrogenase, the nicotinate-nucleotide pyrophosphorylase nadC, a uridine phosphorylase udp, FTN_0649, a FAD-dependent 4Fe-4S ferrodoxin, and FTN_0869, a hypothetical protein that encodes a putative transglutaminase that is regulated by the virulence factor mglA [28].
Since the imd pathway has been well-characterized as being responsible for inducing antimicrobial peptide (AMP) mRNA levels in response to F. novicida and other bacterial pathogens, the simplest explanation for the rescue of these bacterial mutants in flies lacking an intact imd pathway is that they have increased sensitivity to AMPs. This idea is further supported by identification of pmrA in our rescue screen, as pmrA has previously been shown to be sensitive to the antimicrobial peptide polymyxin B in vitro [29]. Therefore, we wished to determine whether the other F. novicida genes identified in our rescue screen are also sensitive to AMPs. There are 7 families of AMPs in Drosophila, and more than 2 dozen individual AMPs can be expressed during an infection, producing a complex bacteriocidal cocktail. Among the characterized AMPS, four families have been implicated in killing Gram-negative microbes, attacin, cecropin, diptericin, and drosocin. The first three families contain cation rich peptides while drosocin is described as proline rich [7]. It is currently impossible to recreate in vitro, the array of AMPs brought to bear on an infecting microbe in vivo. We therefore tried testing individual AMPs for their effects on F. novicida mutants. Unfortunately, few of these Drosophila AMPs are available commercially. We tested a commercial preparation of cecropin and did not detect activity against F.novicida on plates (data not shown). We turned to the cationic antimicrobial peptide polymyxin B, which has been used to model AMP sensitivity in F. tularensis in multiple studies [29], [30].
Of the seven genes we confirmed to be rescued in imd mutant flies, we found that five of these genes, pmrA, FTN_0889, glpD, udp, and FTN_0649 were indeed more sensitive to polymyxin B in vitro. Suprisingly, mutants in the genes FTN_0869 and nadC did not show any phenotypes in these assays, suggesting that the imd rescue phenotype of these mutants may not be due AMP sensitivity, or at least not to cationic AMP sensitivity (Figure 4D). To determine how common this phenotype was, we expanded our analysis to include the entire set of confirmed attenuated mutants described in Figure 2. We found that twelve of the fourteen F. novicida mutants that were rescued in imd mutant flies on arrays showed altered sensitivity to polymyxin B, whereas this was the case with just five of the thirty eight mutants not rescued in an imd mutant (Figure 2). These five mutants were likely exceptions as they also had defects in reactive oxygen sensitivity and in the absence of an imd mediated response would still be sensitive to a melanization response. In our entire set of attenuated mutants, only nadC and FTN_0869 mutants demonstrated the unique phenotype of rescue in an imd mutant fly without showing any increased sensitivity to AMPs, so we chose to focus on one of these genes, the putative transglutaminase FTN_0869 for further analysis.
The mutation in the gene FTN_0869 was intriguing as it clearly grows better in imd mutants as compared to wild type flies yet the mutant does not demonstrate altered sensitivity to the antimicrobial peptide we tested. The fly produces dozens of AMPs at once and not all of them work by the same mechanism, therefore it is impossible and illogical to eliminate the possibility that a single untested AMP or combination of imd induced AMPs might be responsible for killing F. novicida. Regardless, the resistance of FTN_0869 mutants to an AMP raises the question that the imd pathway might be generating an immune response that was not AMP mediated. In addition, the fact that this gene is regulated by the virulence factor mglA which regulates the F. tularensis pathogenicity island and many other important virulence factors suggested that it could be particularly important to F. tularensis pathogenesis. To determine the extent of the attenuation of FTN_0869 mutants, we examined the growth and survival of these bacteria in individual infections.
We observed that with a starting dose of 5×103 bacteria the FTN_0869 mutant took significantly longer to kill wild-type flies than did wild-type F. novicida (Figure 5A). This phenotype was completely rescued in imd mutant flies, with both FTN_0869 mutants and wild-type bacteria killing the fly with a mean time to death of 7 days, consistent with the sensitivity phenotype observed for imd flies (Figure 5B and Figure S1). In wild-type Drosophila, the FTN_0869 mutant did not develop the high bacterial loads found in wild type flies; wild type F. novicida can reach titers of 5×107 CFU per fly within 4 days while the FTN_0869 mutant did not grow higher than 5×105 CFU/fly (Figure 5C). Again, this phenotype was abrogated in imd mutant flies, in which both wild type bacteria and FTN_0869 mutants were able to grow to similar high titers. (Figure 5D) This suggested that an imd-regulated immune response was preventing the FTN_0869 mutants from growing as well as wild-type U112 bacteria in the fly. We infer that the decreased bacterial population was responsible for the decreased virulence observed in terms of fly survival.
The attenuated phenotype of mglA mutants in mouse cells is due to the inability of these mutants to survive and replicate intracellularly [31]. Since FTN_0869 is regulated by mglA, we sought to determine whether the same was true for this mutant. We performed gentamycin chase assays on wild-type U112, mglA mutants, and FTN_0869 mutants. As expected, the mglA mutants showed no bacterial growth within the fly but rather were partially cleared very quickly following injection into the fly, and were completely unable to establish an intracellular population (Figure 6A). This suggests that the small intracellular population may be important, if not essential, for the establishment of a successful infection. In contrast, the FTN_0869 mutants had a robust albeit slightly reduced intracellular population as compared to wild type bacteria, but demonstrated a unique phenotype with little to no extracellular bacteria present in wild-type flies (Figure 6B). By testing sensitivity to gentamycin in vitro, we were able to show that this was due to lack of extracellular bacteria rather than an increased sensitivity of the FTN_0869 mutant to gentamycin (Figure S3). Again, loss of the imd pathway in the host animal eliminated this effect; the extracellular population of FTN_0869 mutant bacteria grew to similar levels as wild-type bacteria in imd mutant flies (Figure 6C).
This result suggested that the extracellular population of bacteria was unable to persist in the extracellular space of infected flies due to an immune mechanism that is controlled by the imd signaling pathway. Therefore, we were interested in investigating this mutant further to determine what effector arm of the innate immune system was responsible for the clearance of extracellular bacteria. We hypothesized that this clearance could be due to either an increased activation of the imd pathway by the FTN_0869 mutants, AMP activity that we were unable to test in vitro, or a novel component of the imd-regulated immune response. To determine if the imd pathway is induced more intensely by the FTN_0869 mutant, to rule out the possibility that this gene is able to downregulate the imd immune response, we measured the induction of imd-regulated AMPS as a readout for imd pathway activation. We used quantitative real-time RT-PCR to monitor the levels of Diptericin, Drosocin, Drosomycin, Attacin, Cecropin and Metchnikowin at 1,2,5 and 24 hours post-infection. We found that only Metchnikowin, Cecropin, and Diptericin were strongly induced in response to F. novicida infections and that the transcript levels of each of these highly-induced AMPs peaked at 24 hours post-infection (Figure 7A and data not shown). All of these AMPs were induced to similar levels during infections with either the wild-type or FTN-0869 mutant bacteria, with no statistically significant difference between induction by wild-type or mutant bacteria at any timepoint. This confirms that the imd pathway is indeed activated by F. novicida and that the gene FTN_0869 does not have an effect on the induction of the imd pathway.
We wished to probe the role of AMPs in clearing F. novicida further. There are more than 30 antimicrobial peptides in the fly and purified Drosophila AMPs are not readily available and the AMPs are always expressed together during an immune response; as described above, this makes it difficult to directly test the role of AMP activity on F. novicida growth in the fly. We therefore tried an indirect approach to test their importance. Recent work in the beetle Tenebrio molitor demonstrated that the majority of bacteria injected into the insect is cleared in less than an hour post-infection, much faster than antimicrobial peptides can be upregulated, transcribed, and synthesized [32]. Using this larger insect model, Haine et al. were able to conclusively demonstrate that insect antimicrobial peptide activity is induced slowly, and thus is not responsible for the bulk of the bacterial clearance. The analysis of antimicrobial peptide induction in the fly relies on the analysis of mRNA transcript levels, which are less accurate kinetically than a direct measurement of antimicrobial activity but nevertheless suggest that a slow induction with transcript levels only rising hours after infection and peaking at 6–24 hours post-infection for various AMPs [33]. To determine if the kinetics of extracellular bacterial clearance coincide with AMP induction, we performed gentamycin chase assays at early timepoints post infection. As early as 1 hour post-infection much of the FTN_0869 mutant population had already been cleared from the fly; in contrast the extracellular population of wild-type bacteria did not substantially decrease (Figure 7B). By two hours post-infection, the timepoint at which both wild-type and mutant bacteria had begun to enter cells, the wild-type bacteria now had both intracellular and extracellular populations, while in FTN_0869 mutant infections only the intracellular bacteria had survived clearance. By five hours post infection the extracellular population of U112 wild-type bacteria had begun to increase while the titer of FTN_0869 mutants did not and only the intracellular population remained. This supported the notion that imd-induced AMPs were not responsible for the clearance of extracellular FTN_0869 mutant bacteria, as the bulk of this clearance occurred within an hour post-infection before AMP activity would be upregulated.
We next sought to determine if one of the other effector arms of the fly innate immune system could be causing this phenotype. We first examined the effects of reactive oxygen species on the FTN_0869 mutants. Unlike many of the genes we isolated from our TraSH screen, the FTN_0869 mutants did not show increased sensitivity to reactive oxygen species produced by H2O2 or paraquat in vitro as measured by disk diffusion assay (Figure 2). We next examined the effects of melanization in vivo by infecting CG3066 mutant flies with wild type and FTN_0869 mutants. Unlike the oxyR mutants, the FTN_0869 mutants were just as attenuated in CG3066 mutants as they are in the wild-type control (Figure 7C) suggesting that these mutants do not have a defect in resisting reactive oxygen stress and that melanization is not responsible for the FTN_0869 imd rescue phenotype. We concluded that the imd rescue phenotype of FTN_0869 mutants not likely due to cationic antimicrobial peptides or melanization and rather suggested a third category of F. novicida interactions with the fly immune system as shown in Figure 2.
Our goal was to dissect the host-pathogen interactions between Francisella and D. melanogaster. To identify components of this complex system we used a combination of three genetic techniques that enabled us to determine the contributions of both host and microbe to the virulence of the infection. First, we identified bacterial virulence factors necessary to infect the fly using a library of F. novicida mutants. Second, we used fly immunity mutants to confirm which host immune pathways were essential for fighting F. novicida infections. Finally, we combined these two techniques to identify subsets of bacterial virulence factors that allow the bacteria to counter-respond to specific immune attacks and evade immune clearance. This paper identifies genes from both the pathogen and the host that are components of each of these aspects of the host-pathogen relationship.
To identify bacterial virulence factors, we performed an in vivo screen that identified 149 bacteria genes that are important for growth and survival within the fly. 41 of the 149 genes had previously been identified in a similar screen performed with the same bacterial library in the mouse indicating that many bacterial virulence factors are conserved between host species [20]–[22]. Genes that overlap between the Drosophila and mouse screens include known virulence factors such as mglA, iglC, and iglD, a number of various transporters, and some of the DNA repair genes we identified as helping F. novicida to survive oxidative stress. The remaining genes are unique to our screen performed in the fly model. These genes could either represent F. novicida genes that play a role specific to arthropod vectors, demonstrate a stronger phenotype in insects than in mammals, or were not identified in previous screens for experimental reasons.
We note that of the 26 F.novicida mutants identified as being sensitive to reactive oxygen, 7 (27%) had been previously identified as being important for virulence in vertebrates. In contrast, of the 16 mutants we found to be polymixin sensitive, only 1 (7%) was identified previously as being important for virulence in vertebrates. The numbers in this study are small enough that differences in representation could be due to chance and therefore future work with more pathogens will be required to confirm the trends seen here; that said, analysis of interactions with the fly's reactive oxygen based immune response seems to be useful predictor of genes that will be of interest to those studying vertebrates. In contrast, analysis of the AMP and imd sensitive mutants is not as robust a tool for identifying mutations that will are relevant in vertebrates.
Secondary screens of these mutants revealed important patterns that shed light onto what particular stresses F. novicida encounters within the fly. 25 of the 56 mutants that we confirmed to have reduced competitive indexes compared to wild-type F. novicida were also hyper sensitive to oxidative stress in vitro. This indicates that preventing or repairing damage caused by reactive oxygen species is an important survival strategy for F. novicida in insect infections. Of particular interest among the genes that were sensitive to oxidative stress was the gene oxyR, which has homology to an E. coli transcriptional regulator that senses and responds to the presence of hydrogen peroxide by inducing the transcription of catalases and other genes that can counteract oxidative stress 23. In addition, our screen identified multiple genes in DNA damage repair pathways that are also sensitive to oxidative stress [24], [25]. We expect that these genes are required to repair damage caused by reactive oxygen species to DNA as has been suggested by Kraemer et al [21].
Of the three effector arms that have been characterized in the immune response occurring within the fly's body cavity, the major producer of oxidative stress is the melanization response [15], [17], [26]. Therefore, we speculated that the large number of negatively selected bacterial mutants with oxidative stress sensitivity phenotypes suggested that the melanization response plays a large role in the fly's immune response to F. novicida. To test this hypothesis, we performed competition assays with the oxyR mutants in fly mutants that lack a melanization response. As expected, the attenuation of these mutants was rescued in flies that do not melanize and therefore would be expected to not produce toxic oxygen species. This demonstrated that melanization is an essential component of the fly immune response against F. novicida and demonstrated that we could use our characterizations of bacterial genes to learn about the fly immune system and understand the host-pathogen relationship. We note that reactive oxygen is a well-established immune effector in the Drosophila gut. Perhaps most microbes encountering the fly will face this immune barrier before encountering internal immune defenses. Thus protection against reactive oxygen is doubly important for fly pathogens [34].
Because microbes must withstand the host immune system to mount a successful infection, we were able to exploit the inherent ability of bacteria to function as metaphorical immunologists to identify which aspects of fly immunity were important to F. novicida infections. We next sought to determine if this system could be used in the reverse direction by manipulating the fly immune system to identify which bacterial virulence factors were responsible for interacting with one specific aspect of innate immunity. We did this by performing a second round of our in vivo screen for bacterial mutants in an immunocompromised fly. We focused on the imd-regulated humoral immune response, which had previously been identified as important for fighting F. tularensis infections [9]. We confirmed that the imd pathway, but not the Toll pathway, was essential in combating F. novicida infections, and performed our TraSH assay in imd mutant flies. We identified a subset of bacterial virulence factors that were important for infections of wild-type flies but not imd flies; this imd-regulated immune response has been primarily characterized for its role in the induction of antimicrobial peptides and therefore we tested these mutants for their sensitivity to a cationic, membrane active antimicrobial peptide, polymyxin B [29], [30]. As expected, twelve of the fourteen mutants were sensitive to polymyxin killing in vitro, providing another example of how resistance to host immune responses is an important component of bacterial virulence.
We identified 2 bacterial mutants that were not sensitive to polymyxin in vitro despite being rescued in imd mutants flies. This phenotype was unexpected as the majority of the literature suggests imd signaling drives antimicrobial peptide production and this is its most important job. We propose three explanations for this phenotype. First, the rescue phenotype could be due to specific sensitivity to additional antimicrobial peptides that were not tested in vitro; the bacteria show no sensitivity to polymyxin but could be sensitive to one of the 30 or more AMPs synthesized by flies. Second, the rescue phenotype could be due to the bacterial gene being an inhibitor of the imd pathway; in this case the bacteria would have wild type sensitivity to AMPs but would encounter increased concentrations of them in the fly because the bacteria could not inhibit AMP production. Finally, the rescue phenotype of these bacterial mutants could be due to an aspect of imd-regulated immunity that has not been previously described.
To differentiate between these possibilities, we chose one imd-rescue mutant, the putative transglutaminase FTN_0869 to characterize further in terms of its interactions with the fly immune system. We chose this gene because it had a strongly attenuated phenotype in wild-type flies that was significantly rescued in imd mutants and because it has previously been shown to be regulated by the virulence factor mglA, which is essential for F. novicida intracellular growth 28. More recently, the homologue of this gene in the extremely virulent Type A F. tularensis ssp. tularensis strain Shu4 was identified in a transcriptional analysis of genes that are upregulated inside mouse bone marrow-derived macrophages (BMMs) [35]. Interestingly, FTN_0869 deletion mutants in the less virulent U112 strain are unable to replicate in BMMs, but mutants of the homologue of this gene, FTT0989 in the SCHU4 strain did not demonstrate any intracellular replication defect [28], [35]. Further characterization of the FTN_0869 mutants showed that these mutants are attenuated for both lethality to the fly and bacterial growth in an imd-dependent manner. However, unlike its transcriptional regulator mglA, the FTN_0869 mutant is capable of intracellular growth within flies, but is incapable of surviving in the extracellular space. This phenotype is consistent with what is observed in mouse bone marrow-derived macrophages with the virulent Shu4 strain, but not with the phenotype of FTN_0869 deletion mutants in the U112 strain. The reason for this difference is unclear, but it is interesting to note that the ability of the putative transglutaminase deletion mutants to grow intracellularly correlates with its virulence in mammalian and insect hosts.
We found that the phenotype of FTN_0869 deletion mutants in flies is imd-dependent, and used this phenotype to investigate the role of the imd pathway in clearing the extracellular bacteria. With this mutant, we were able to show that the imd rescue phenotype of this particular mutant was not due to modulation of the imd pathway because AMP genes downstream of imd are induced to similar amounts in infections with wild-type and FTN_0869 bacteria. By examining the kinetics of the clearance of extracellular bacteria, we were able to limit the possibility that other imd-induced antimicrobial peptides that we did not test in vitro were causing the attenuation of the FTN_0869 mutant. Using non-melanizing mutants, we were able to rule out melanization as the cause of this phenotype, leaving us with the possibility that imd could either be regulating the cellular immune response or an uncharacterized effector arm of fly immunity. Thus the FTN_0869 phenotype suggested a third category of host-pathogen interactions between F. novicida and the Drosophila innate immune system. Future work with this mutant and other imd-rescue mutants identified in our screen could provide further insight into the biology of the imd-regulated fly immune response.
In summary, reciprocal studies of a pathogen, F. novicida and a host, D. melanogaster, allowed us to identify genes in the pathogen required to counteract, evade, or resist host immune responses and allow bacterial growth and survival. These studies identified two branches of host immunity that are important for fighting F. novicida infections, melanization and imd-regulated immune responses and helped us to understand how the bacteria resists these responses. By identifying the mechanism of one or two bacterial mutants based on their sequence or interaction with fly mutants we developed assays to identify the mechanism of mutants with unknown function. Our work with one of these mutants, FTN_0869, taught us that there is likely more to learn about the fly immune system as there are classes of F. novicida mutants that cannot immediately be explained by their interactions with the melanization response or AMPs. Our screen allowed us to pose directed questions and focus our investigations on particular aspects of the host immune system and the microbial strategies to evade this immune response, helping us to identify and characterize components of the host-pathogen relationship.
All experiments were performed in wild-type Oregon Red (OR) flies unless otherwise noted. The imd mutant fly line imd10191 is a null allele with a 26-nucleotide deletion at amino acid 179 that results in a frameshift mutation and has been backcrossed onto an OR background. The Toll pathway alleles tested in this study are Dif1 which is a complete loss of function mutant and MyD88C03881 The CG3066KG02818 Sp7 mutant flies are PiggyBack insertion mutants on a w1118 background (Bloomington stock number 13494), and w1118 flies are used as the wild-type control for these experiments. All experiments were performed on 5–7 day old age-matched male flies that were maintained on dextrose medium at 25°C and 65% humidity in a 12∶12h light dark cycle.
Francisella novicida strain U112 was used for all experiments described in this paper. Bacterial stocks were grown in Tryptic Soy Broth (TSB) supplemented with 0.2% L-cysteine and cultured overnight under aerobic conditions at 37°C. Cultures were grown to an OD600 of 1.5–2 and diluted in PBS to OD600 0.005–0.01 for fly infections.
Flies were anaesthetized with CO2 and injected with 50nL of bacteria using a glass needle and a Picospritzer III injector system (Parker Hannifin). Each fly was injected in the ventrolateral surface of the fly abdomen and placed into fresh vials with no more than 20 flies per vial to prevent crowding. Following infection, the flies were incubated at either 25°C or 29°C as noted. Each survival curve was performed using 3 replicates of 20 flies each for a total of 60 flies per condition and each experiment was performed a minimum of three times. The number of dead flies was monitored daily and Kaplan-Meier survival curves were generated using GraphPad Prism software, and statistical analysis was performed using log-rank analysis.
Individual infected flies were homogenized in 100µL of PBS, serially diluted, and plated onto Mueller-Hinton (MH) agar plates supplemented with 0.025% ferric pyrophosphate (Sigma), 0.1% glucose, 0.025% calf serum (GIBCO), and 0.02% Iso-VitaleX (Benton Dickinson). Plates were incubated overnight and colonies were counted to determine the number of bacterial colony forming units (CFUs) per fly. Statistical significance was determined using unpaired two-tailed t-tests. Gentamycin chase experiments were performed as described about except that 50nL of 1mg/mL of gentamycin was injected into each fly 3 hours prior to plating [36].
Three sets of 30 flies were injected with 50nL of the trash library. Each fly received approximately 2*105 CFUs of bacteria, representing approximately 2-fold coverage of the library. The infection was allowed to proceed for two days at 29°C, at which point each fly was homogenized and plated onto MH agar. Plates were incubated at 37°C overnight, and the bacteria were collected and pooled and DNA was collected by phenol-chloroform extraction. Each pool was divided in half and digested with either BfaI or RsaI (NEB) and re-pooled to be used as a template for in vitro transcription with a MegaScript T7 Kit (Ambion). The RNA was then purified and used for reverse transcription using a SuperScript III First Strand Synthesis Kit (Invitrogen) and random hexamer primers. The resulting cDNA was labeled with amino-allyl dUTP using Klenow (exo-) enzyme (NEB). The input pool was then labelled with Cy5 and the day 2 pools with Cy3 and hybridized to Francisella microarrays as has been previously described [28]. Data was normalized using the Stanford Microarray Database according to the median log2 Cy5/Cy3 and filtered using a Cy3 net median intensity of 150 and a regression correlation of >0.6. The dataset was then analyzed using SAM software using a blocked 2-class analysis to identify differences between the input and wild-type or input and imd mutant samples with a false discovery rate of 5% [37].
Genes that were selected for further analysis were knocked out of F. novicida individually to create deletion mutants. Briefly, 500bp of sequence 5′ and 3′ to the gene of interest was amplified from genomic F. novicida DNA using Phusion DNA Poylmerase (NEB), and fused onto either side of a kanamycin cassette using a sewing PCR reaction. 38 The resulting PCR products were then transformed into chemically competent F. novicida U112 as described [28] and the mutants were confirmed by PCR.
To confirm the bacterial growth attenuation phenotypes, 50nL of a 1∶1 ratio of mutant and wild-type bacteria at an OD600 of 0.01 was injected into flies. The infection was allowed to proceed for 2 days at 29°C, following which the flies were homogenized and plated onto MH agar plates with and without 30 mg/mL of kanamycin. Since only the mutant bacteria is capable of growing in kanamycin media, we were able to determine the number of wild-type and mutant bacterial CFUs for each fly by subtracting the number of mutant bacterial CFUs from the total CFUs per fly. A competitive index (CI) was determined using the formula CI = (mutant CFU day 2/wild-type CFU day 2)/(mutant CFU input/wild-type CFU input).
To determine the sensitivity of various F. novicida mutants to oxidative stress and antimicrobial peptides, disk diffusion assays were performed using protocols adapted from Mohapatra et al. and Bakshi et al. [27], [29]. Briefly, 50µL of overnight cultures of bacteria were plated onto MH agar plates to create a lawn of bacteria. Plates were allowed to dry for 10 minutes, and then 6mm Whatman filter paper disks (Fisher Scientific) were placed onto each plate and inoculated with 10µL of 100mM freshly diluted hydrogen peroxide (Sigma) or 10µL of a 10 mg/mL stock of polymyxin B. Plates were incubated overnight and the diameter of the zone of inhibition was measured for each sample. Three zones were measured for each mutant and each experiment was repeated three times.
The fold increase of antimicrobial peptide expression following infection by wild-type and FTN_0869 mutant F. novicida was determined by isolating RNA from infected flies 6 and 24 hours post-infection by trizol extraction and performing qRT-PCR analysis using an iScript One-Step RT-PCR kit with SYBR Green (Bio-Rad) and a Bio-Rad icycler. The following primer sets were used: cecropin 5′ 5″-tcttcgttttcgtcgctctc-3′, cecropin 3′ 5′-cttgttgagcgattcccagt-3′, drosomycin 5′ 5′-gacttgttcgccctcttcg-3′, drosomycin 3′ 5′-cttgcacacacgacgacag-3′, diptericin 5′ 5′-accgcagtacccactcaatc-3′, diptericin 3′ 5′-cccaagtgctgtccatatcc-3′, attacin 5′ 5′-caatggcagacacaatctgg-3′, attacin 3′ 5′-attcctgggaagttgctgtg-3, drosocin 5′ 5′-ttcaccatcgttttcctgct-3′, drosocin 3′ 5′-agcttgagccaggtgatcct-3′, metchinkowin 5′ 5′-tcttggagcgatttttctgg3′, metchnikowin 3′ 5′aataaattggacccggtcttg-3′, ribosomal protein 15a 5′-tggaccacgaggaggctagg, 3′-gttggttgcatcctcggtga.
|
10.1371/journal.pgen.1000003 | The Secreted Metalloprotease ADAMTS20 Is Required for Melanoblast Survival | ADAMTS20 (A disintegrin-like and metalloprotease domain with thrombospondin type-1 motifs) is a member of a family of secreted metalloproteases that can process a variety of extracellular matrix (ECM) components and secreted molecules. Adamts20 mutations in belted (bt) mice cause white spotting of the dorsal and ventral torso, indicative of defective neural crest (NC)-derived melanoblast development. The expression pattern of Adamts20 in dermal mesenchymal cells adjacent to migrating melanoblasts led us to initially propose that Adamts20 regulated melanoblast migration. However, using a Dct-LacZ transgene to track melanoblast development, we determined that melanoblasts were distributed normally in whole mount E12.5 bt/bt embryos, but were specifically reduced in the trunk of E13.5 bt/bt embryos due to a seven-fold higher rate of apoptosis. The melanoblast defect was exacerbated in newborn skin and embryos from bt/bt animals that were also haploinsufficient for Adamts9, a close homolog of Adamts20, indicating that these metalloproteases functionally overlap in melanoblast development. We identified two potential mechanisms by which Adamts20 may regulate melanoblast survival. First, skin explant cultures demonstrated that Adamts20 was required for melanoblasts to respond to soluble Kit ligand (sKitl). In support of this requirement, bt/bt;Kittm1Alf/+ and bt/bt;KitlSl/+ mice exhibited synergistically increased spotting. Second, ADAMTS20 cleaved the aggregating proteoglycan versican in vitro and was necessary for versican processing in vivo, raising the possibility that versican can participate in melanoblast development. These findings reveal previously unrecognized roles for Adamts proteases in cell survival and in mediating Kit signaling during melanoblast colonization of the skin. Our results have implications not only for understanding mechanisms of NC-derived melanoblast development but also provide insights on novel biological functions of secreted metalloproteases.
| Mice with black and white coat coloration patterns have long been favorites of mouse fanciers and geneticists alike. Analysis of mouse coat color mutants has yielded important insights into normal developmental pathways as well as human disease processes. In this study we have investigated how mutations in a secreted metalloprotease, Adamts20, result in mice with white belts in their lumbar region, even though Adamts20 is not expressed in the pigment producing cells. Our findings suggest that the belting pattern is due to a combination of increased pigment cell death, decreased pigment cell number in the trunk, and functional overlap of closely related metalloproteases. Adamts20 mutants have disrupted function of Kit, a protein that regulates pigment cell development, as well as alterations in the extracellular matrix that surrounds the pigment cells. These findings have implications both for our understanding of general mechanisms of pigment cell development as well as for new biological functions of secreted metalloproteases.
| A disintegrin-like and metalloprotease with thrombospondin type-1 motifs (ADAMTS) metalloproteases constitute a large family of 19 zinc-dependent proteolytic enzymes that are distantly related to both the A disintegrin and metalloproteinase (ADAM) family, and to the matrix metalloproteinases (MMPs) [1],[2]. In contrast to the ADAM proteases that are membrane anchored, ADAMTS proteases are secreted; however, some may be considered as operational cell-surface proteases as they bind to the cell surface. Some ADAMTS proteases, such as ADAMTS10 (GeneID: 224697), ADAMTS13 (GeneID: 279028), and the procollagen amino-propeptidases (e.g. ADAMTS2), are highly specialized; others process a variety of substrates within the extracellular matrix (ECM), including chondroitin sulfate proteoglycans (CSPGs), such as aggrecan (GeneID: 11595) and versican (GeneID: 13003). Mouse models harboring mutations in ADAMTS family members have demonstrated the importance of these proteases during development. In some cases mutant phenotypes can be attributed to a failure to cleave specific substrates [3]–[7]. Mutation or dysregulation of ADAMTS proteases is associated with inherited and acquired pathologies including Ehlers-Danlos syndrome VIIC (OMIM#225410), thrombocytopenic purpura (OMIM#274150), Weill-Marchesani syndrome (OMIM#277600) and arthritis [5], [8]–[10].
Among the animal mutants in Adamts proteases is a classical white-spotted mouse named belted (bt), so named because it contains white spots in the lumbar region creating the appearance of a belt [11]–[13]. Sequencing of three of the twelve known alleles of belted—btBei1, btMri1, and bt (Mouse Genome Informatics, MGI: 2660628)—revealed nonsense or missense mutations in Adamts20, thus implicating metalloproteases in skin pigmentation [12]. Analyses of white spotting mutants suggest that such phenotypes are often due to defective development of neural crest (NC)-derived melanoblasts that produce pigment of the integument (skin, hair, feathers, and scales), inner ear, and eye [14],[15].
Melanoblasts develop from a subset of NC that emigrate from the neural tube and overlying ectoderm and migrate dorso-laterally relative to the neural tube through prospective dermal mesenchyme (embryonic day (E) 8.5-E9.5) [15],[16]. Subsequently, the melanoblasts differentiate and expand (E9.5-E13.5), migrate into the epidermis and hair follicle (E13.5-E15.5), and eventually produce melanin (E15.5-P0). Several molecules, including the receptor tyrosine kinase Kit (GeneID: 16590) and its ligand, Kit ligand (Kitl, GeneID: 17311), regulate melanoblast development. Kit and Kitl act throughout melanoblast development, with independent requirements for melanoblast survival, proliferation and migration [17]–[23]. Kit is expressed on melanoblasts, and Kitl is expressed in the dermis and in dermal mesenchymal condensations and papillae [21], [24]–[26].
Similar to Kitl, Adamts20 is expressed in dermal mesenchymal cells adjacent to and in advance of migrating melanoblasts throughout their development. This expression pattern led us to initially propose that the bt phenotype was caused by defective melanoblast migration [12]. This hypothesis was also based upon the observation that the Adamts20 ortholog Gon-1 (GeneID: 177850) is essential for gonadal morphogenesis and distal tip cell migration in C.elegans [27]–[29]. The current study tests our hypothesis by performing extensive characterization of melanoblast development in bt/bt embryos. Our findings suggest that Adamts20 mutant mice exhibit white spotting in a belted pattern due to a combination of regional variation of melanoblast number, increased apoptosis, and functional overlap with Adamts9 (GeneID: 101401). Furthermore, our analyses of bt/bt embryos indicate that defective Kit signaling and versican processing may explain the failure of melanoblasts to develop properly.
Adamts20 mutant mice exhibit white spotting in the lumbar region on both dorsal and ventral surfaces, frequently resulting in the appearance of a white belt in recessive animals (Figure 1A) [11]–[13]. For these studies an allele of bt on an inbred C57BL/6 background was used, as different genetic backgrounds can affect the expressivity of the bt phenotype [30],[31]. Since an Adamts20 mutation had not been specifically demonstrated in this particular bt allele, a complementation cross was performed between these bt mice and btBei1/+ mice. Seven out of 10 animals born exhibited a bt phenotype, consistent with the expected Mendelian ratios (50% bt) and indicative of a recessive allele of bt. This allele has now been designated bt9J and is listed at MGI.
The Adamts20 gene was sequenced from bt9J/bt9J genomic DNA in order to identify the molecular lesion. A single C to T point mutation was identified at nucleotide 2451, which is predicted to cause a missense mutation (Leu761Phe) in the spacer domain of ADAMTS20. The leucine residue mutated in bt9J mice is present in both mouse and human Adamts20 genes and is highly conserved among all Adamts mouse genes (present in 15 of the 19 genes). This suggests that Leu761 may be a critical residue for ADAMTS20 folding or for enzymatic function. We designed a TaqmanTM assay to genotype bt9J mice (see Materials and Methods). We determined that bt9J is the same strain as Mutant Mouse Resource Center (MMRC) strain #183 (Figure 1B). Although the missense mutations in bt9J, btMri1, and bt affect different domains of ADAMTS20, these and other bt alleles are recessive and exhibit very similar phenotypes, strongly suggesting they act as functional null alleles.
White spotting of mouse coats is typically caused by defective melanoblast development during embryogenesis. In order to elucidate when and how Adamts20 affects melanoblast development, melanoblast distribution in bt9J/bt9J embryos was examined. We generated C57BL/6 mice containing a Dct-LacZ transgene, which marks specified melanoblasts, and crossed these onto a C57BL/6 bt9J/bt9J background [18],[32]. Since there are no phenotypes associated with bt9J/+ adult animals, this heterozygous genotype served as a control for this study.
We compared melanoblast distribution in whole mount bt9J/+;Dct-LacZ and bt9J/bt9J;Dct-LacZ embryos at E11.5 through E16.5 (Figure S1 and Figure 2). In E11.5 and E12.5 control embryos, melanoblasts had completed their initial migration from the neural tube and were distributed evenly across the dorsal and lateral surfaces of the embryo (Figures S1 and Figure 2A and 2C). A similar distribution pattern in E11.5 and E12.5 bt9J/bt9J embryos was observed (Figures S1 and Figure 2B and 2D). Quantification of melanoblasts in the head and the presumptive belt region of the trunk of E12.5 whole mount embryos showed no difference in melanoblast number between control and bt9J/bt9J embryos (Table 1, boxes in Figure 2A and 2C). These results show that initial specification and migration of NC-derived melanoblasts is not disrupted in bt9J/bt9J animals.
Since there was no defect in younger embryos, melanoblast distribution in bt9J/+; Dct-LacZ and bt9J/bt9J;Dct-LacZ E13.5-E16.5 embryos was compared (Figure 2E–2P). In E13.5 and E14.5 bt9J/+;Dct-LacZ embryos, the melanoblast population had expanded throughout the embryo and melanoblasts were distributed uniformly across dorsal and lateral surfaces of the trunk (Figure 2E, 2G, 2I, and 2K). In contrast, melanoblast distribution was specifically reduced in the trunk of E13.5 and E14.5 bt9J/bt9J;Dct-LacZ embryos (Figure 2F, 2H, 2J, and 2L). Quantitation of the head and trunk region of E13.5 and E14.5 bt9J/bt9J;Dct-LacZ embryos confirmed that the melanoblast defect was limited to the trunk (Table 1, boxes in Figure 2E, 2G, 2I, 2K) (p = 0.0000002 at E13.5, p = 0.000096 at E14.5). By E15.5 and E16.5, melanoblasts were distributed evenly on both dorsal and ventral surfaces of bt9J/+;Dct-LacZ embryos (Figure 2M and 2O). However bt9J/bt9J;Dct-LacZ embryos displayed a notable absence of melanoblasts on both surfaces in the region corresponding to the future belt (Figure 2N and 2P). Melanoblast distribution in the head and tail regions was similar between E16.5 bt9J/+;Dct-LacZ and bt9J/bt9J;Dct-LacZ embryos (Figure 2O and 2P). Close examination of the belt region in E16.5 bt9J/bt9J;Dct-LacZ embryos showed no accumulation of melanoblasts at the edges of the belt, as might be expected if migration into this region were impaired (Figure S2). Collectively, these analyses of the melanoblast distribution in bt9J/bt9J animals reveal the following: the phenotype is first apparent at E13.5, melanoblasts are reduced only in the trunk region (the location of the presumptive belt), and dorso-lateral melanoblast migration is not impaired.
The melanoblast defect is first observed at E13.5, coincident with the timing of melanoblast migration from the dermis into the epidermis, and with generalized dermal expression of Adamts20 [12],[15]. Therefore we hypothesized that Adamts20 may be required for normal distribution of melanoblasts in dermal and epidermal compartments of the skin. To address this, melanoblasts were quantified in the dermis, epidermis, and dermal-epidermal border in E13.5 trunk sections (Figure 3A and Table 2) (see Materials and Methods). In control trunk sections (n = 178) we observed an average of 33.6 melanoblasts per section, whereas in bt9J/bt9J;Dct-LacZ sections (n = 179) there was an average two-fold reduction in melanoblast number (15.9 per section, p<0.0001) (Table 2). This finding is consistent with the analyses of whole mount embryos (see Table 1 and Figure 2E–2H). However, the relative distribution of melanoblasts within each of the skin layers in bt9J/bt9J;Dct-LacZ embryos did not differ significantly from bt9J/+;Dct-LacZ control embryos (p = 0.056) (Figure 3B). These results show that while melanoblast number is significantly reduced in the lumbar region of bt9J/bt9J embryos, melanoblast migration from the dermis into the epidermis is unaffected.
At E13.5, when the defect in melanoblast development in bt9J/bt9J embryos is first apparent, melanoblasts are undergoing dramatic increases in cell proliferation [18]. To assess if the bt phenotype is caused by reduced melanoblast proliferation in the trunk region of bt9J/bt9J embryos, the mitotic index of melanoblasts at E13.5 was examined in bt9J/+;Dct-LacZ and bt9J/bt9J;Dct-LacZ trunk sections (Figures 3A and 4). Melanoblasts undergoing mitosis were identified by co-expression of β-galactosidase and phospho-histoneH3 (PH3) (Figure 4A–4C). The average percentage of dividing melanoblasts in control and bt9J/bt9J sections was not significantly different (2.94% and 3.03% respectively, p = 0.9067) (Figure 4G), suggesting that the melanoblast reduction in bt9J/bt9J embryos is not a consequence of abnormal proliferation.
The reduced number of melanoblasts in the trunk region of E13.5 bt9J/bt9J;Dct-LacZ embryos could also be due to a failure of melanoblasts to survive [33]. To examine this, the apoptotic index of melanoblasts was quantitated from cells co-expressing β-galactosidase and cleaved-caspase3 (CC3) (Figures 3A and 4D–4F). The average percentage of apoptotic melanoblasts was significantly increased in bt9J/bt9J;Dct-LacZ sections relative to bt9J/+;Dct-LacZ sections (compare 8.4% to 1.2%, p<0.00005) (Figure 4H, Table 2). Apoptotic melanoblasts were apparent in both dorsal and ventral skin in control and mutant embryos. In addition, no significant differences were seen in the distribution of apoptotic, CC3+melanoblasts in the three skin compartments, as follows: dermal (bt9J/bt9J = 77.8%; control = 82.4%), epidermal (bt9J/bt9J = 14.8%; control = 5.8%), and dermal/epidermal (bt9J/bt9J = 7.4%; control = 11.85%), (p = 0.6133). These results show that Adamts20 is required for melanoblast survival in all cell layers of the trunk at E13.5.
Adamts20 is expressed along the length of the embryo, yet the spotting phenotype in adult mice is evident only in the lumbar region. Therefore we examined melanoblast number and melanoblast apoptosis in a region outside of where the belt occurs, in the head (Table 2). In normal embryos, the density of melanoblasts is not the same in all regions, as previously described [34]–[37]. As expected, quantitation of melanoblasts in control embryos demonstrated that there were over three-fold more total melanoblasts in the head region than in the trunk (compare 110 versus 33.6, p = 0.0001). There was no difference in melanoblast number in the head of control and bt9J/bt9J;Dct-LacZ embryos (Table 2), similar to what was seen in whole mount analyses.
The apoptotic index of melanoblasts varied in different regions of control embryos as well as between control and mutant embryos (Table 2). In control embryos, melanoblast apoptosis was lower in the head than in the trunk. Comparing control and bt9J/bt9J embryos, we found that apoptosis was increased in the head although to a lesser extent than observed in the trunk. This result indicates that Adamts20 is required for melanoblast survival throughout the embryo even though it does not result in white spotting in the head.
Melanoblast survival is dependent upon Kit activation [20],[38] by Kit ligand (Kitl), which is expressed in the dermis in a similar temporal and spatial pattern to Adamts20 [12], [21], [24]–[26]. We hypothesized that Adamts20 regulates melanoblast survival through modulation of Kit signaling, and assessed the effects of altered Kit signaling upon the extent of white spotting in bt9J/bt9J animals harboring various mutant alleles of Kit (MGI: 96677) or Kitl (MGI: 96974). Kittm1Alf/+ mice contain a null mutation in Kit and have primarily ventral spotting [39]–[41]. As depicted in Figure 5, bt9J/bt9J;Kittm1Alf/+ mice exhibited dramatically wider dorsal and ventral belts in comparison to bt9J/bt9J mice or Kittm1Alf/+ mice alone (Figure 5A and 5B). The increased spotting was a synergistic effect rather than an additive one (compare 46.9% of dorsal and ventral surfaces in bt9J/bt9J;Kittm1Alf/+ with 11.2% and 9.1% in Kittm1Alf/+ and bt9J/bt9J, respectively). Synergistic increases in spotting were also observed in bt9J/bt9J mice carrying mutant alleles of Kitl. Heterozygosity for either KitlSl, a null allele, or KitlSl-d, a deletion that generates short soluble Kitl but not membrane–bound Kitl, combined with homozygosity for bt9J resulted in synergistically increased spotting (Figure 5C and 5D) [42],[43]. In contrast, bt9J/bt9J mice carrying a mutation in Mitf (bt9J/bt9J;MitfMi/+) (MGI: 104554) exhibited no synergistic spotting (Figure 5E). These results show that decreasing Kit signaling exacerbates the bt9J/bt9J phenotype.
Heterozygous KitW-v mutations reduce melanoblast number beginning at E10.5 [18], which is three days prior to the onset of the melanoblast defect in bt9J/bt9J. To determine if decreasing Kit signaling in bt9J/bt9J embryos would lead to an earlier defect in melanoblast development than seen in bt9J/bt9J alone, we examined E12.5 bt9J/+;Kittm1Alf/+ and bt9J/bt9J;Kittm1Alf/+ embryos (Figure S3). Since the Kittm1Alf mutation is a LacZ knock-in at the Kit locus, we monitored melanoblasts using β-galactosidase staining. There was no detectable difference in Kit-positive melanoblast distribution between bt9J/+;Kittm1Alf/+ and bt9J/bt9J;Kittm1Alf/+ embryos at E12.5. These results demonstrate that the synergistic defect in melanoblast development does not precede the onset of the bt phenotype at E13.5.
The synergistic interaction studies indicate that Kit signaling may be disrupted in Adamts20 mutant animals. To test if melanoblasts in bt9J/bt9J embryos could respond to sKitl we used an ex vivo embryonic skin explant assay using dorsal trunk skin from E13.5 embryos [44]–[46]. Observation of bt9J/+;Dct-LacZ and bt9J/bt9J;Dct-LacZ skin after four days in culture demonstrated that ex vivo explant culture conditions could recapitulate both normal melanoblast colonization and the bt defect (Figure 6). Melanoblasts in the control skin entered hair follicles and were distributed evenly across the explant (n = 5) (Figure 6A). While some melanoblasts survived and migrated into hair follicles in the bt9J/bt9J;Dct-LacZ skin, in a large domain of the explant the melanoblasts were reduced (n = 5) (Figure 6B), similar to the phenotype of whole mount E16.5 bt9J/bt9J;Dct-LacZ embryos (Figure 2O and 2P).
Previous studies showed that melanoblasts within embryonic skin cultures respond to soluble Kitl (sKitl) [44] and that sKitl promotes melanoblast survival in NC cultures [38]. ADAMTS20 could be required for proteolytic cleavage of Kit or Kitl, or for modifying additional molecules necessary for Kit signaling. To test if Kit signaling was defective in bt9J/bt9J embryos, we asked whether melanoblasts in bt9J/bt9J embryos could respond to sKitl and whether sKitl could rescue the bt phenotype. Similar to studies by Jordan et. al, an increase in melanoblast number and colonization of hair follicles was observed in control skin exposed to sKitl as compared to FBS alone (Figure 6C, n = 5). In contrast, exposure of bt9J/bt9J;Dct-LacZ skin (n = 7) to sKitl did not restore melanoblasts to the presumptive belt (Figure 6D). Interestingly, melanoblasts at the rostral and caudal regions of the trunk did not increase in number, even though these are outside of the expected belt. These results show that in the absence of Adamts20, melanoblasts are unable to respond to sKitl, indicating that Kit signaling is disrupted in bt9J/bt9J mutants. Furthermore, since sKitl could not rescue the bt phenotype, it suggests that the melanoblast defect in bt9J/bt9J animals is not caused by defective cleavage of Kitl.
We explored the possibility that Adamts20 could regulate melanoblast survival through additional pathways. One of the best-understood activities of ADAMTS proteases is processing of various ECM substrates. Of note, ADAMTS20 is most closely related phylogenetically to ADAMTS metalloproteases that specifically process chondroitin sulfate proteoglycans (CSPGs) [1],[2],[47]. Although no substrates for ADAMTS20 have been identified, the CSPG versican is enriched in the skin and implicated in NC development, making it an excellent candidate substrate of ADAMTS20 [48]–[52].
We examined Versican and Adamts20 expression by whole mount in situ hybridization in wild-type embryos. At E12.5, when melanoblasts are widely distributed in the trunk, both Versican and Adamts20 are expressed broadly across the trunk, overlapping in many regions of the skin and the developing mammary ducts (data not shown). Since both versican and ADAMTS20 are secreted proteins, the expression patterns of their RNAs relative to each other and to melanoblasts support the possibility of a functional relationship in vivo.
We assessed the ability of ADAMTS20 to process versican by expressing ADAMTS20 in 293 cells, and subsequently incubating these cells with versican (Figure 7A). Versican is alternatively spliced to generate four isoforms, with VersicanV1 being the dominant variant in adult skin [53]–[56]. Versican cleavage was examined using anti-DPEAEE antibody, which specifically recognizes the neo-epitope generated by ADAMTS cleavage of versicanV1 (70 KDa) [57]. While secreted media from 293 cells alone exhibited little to no cleavage of versican, versican processing was evident in secreted media from 293 cells expressing Adamts20 (Figure 7A). As a positive control, cleaved versican was also evident in secreted media from cells expressing Adamts9, which has previously been shown to process versican [47]. These results show that in vitro, versican is a substrate for cleavage by ADAMTS20.
Analysis of skin extracts from the dorsal trunk of E15.5 bt9J/+ and bt9J/bt9J embryos showed ADAMTS20 is also necessary for versican cleavage in vivo (Figure 7B). The levels of the cleaved 70 KDa band were reduced in bt9J/bt9J extracts compared to control extracts (n = 4 experiments). We also observed a reduction in levels of a 220 KDa band corresponding to the V0 isoform (data not shown). In contrast, there was no alteration in total versican, which was assessed using a GAG β antibody that recognizes the GAG β domains of full-length versicanV0 and V1 (Figure 7B). Collectively, these results show that ADAMTS20 cleaves versican in vitro and that it is required for versican cleavage in skin in vivo.
Immunofluorescence was performed to evaluate the spatial pattern of versican cleavage in embryonic skin. Versican expression was assessed on trunk sections of E15.5 C57BL/6 (+/+) and bt9J/bt9J embryos (Figure 7C–7H). In +/+ sections, cleaved versican was evident in the dermal mesenchyme and was enriched in the proliferating basal layer of the epidermis (arrow in Figure 7C) but was excluded from the outermost layer of the epidermis, in a similar pattern to that of intact versican in adult human skin [56]. In contrast to +/+ sections, bt9J/bt9J sections displayed a dramatic reduction in cleaved versican in the epidermis (Figure 7D). Inclusion of the DPEAEE peptide in the staining procedure demonstrated the specificity of the antibody for this peptide (Figure 7E and 7F). Total versican expression was similar in both +/+ sections and bt9J/bt9J sections, with high expression in the basement membrane, dermis, and dermal condensations, but lower levels in the epidermis (Figure 7G and 7H). Outside of the belt region, at the level of the forelimb, the levels of cleaved versican in bt9J/bt9J sections were reduced relative to wild-type, but not to the same extent as seen in the trunk (Figure 7I and 7J). Together these results show that in bt9J/bt9J embryos, levels of processed versican, but not total versican, are reduced in the skin of the embryonic trunk, and show that ADAMTS20 is necessary for versican remodeling during melanoblast development.
Although Adamts20 is expressed across the length of the embryo [12], only the lumbar region of the trunk in bt9J/bt9J mutants shows dramatic reduction of melanoblasts and white spotting. Given that ADAMTS20 belongs to a large family of closely related metalloproteases, we reasoned that partial redundancy with other metalloproteases might account for the higher proportion of melanoblasts surviving outside of the lumbar region. The two closest Adamts20 homologs, Adamts5 (GeneID: 23794) and Adamts9, are both expressed in the embryonic skin along the length of the embryo and are both versicanases, making them excellent candidates for regulating melanoblast development [47],[58],[59](Bon-Hun Koo and S.S. Apte, unpublished data). To test if these metalloprotease genes were functionally redundant with Adamts20, we evaluated the extent of white spotting in mutants lacking Adamts20 and either Adamts9 (Adamts9ko) or Adamts5 (Adamts5ko).
Adamts9ko homozygotes are lethal prior to gastrulation (H. Enomoto and S. Apte, submitted) therefore we used Adamts9ko heterozygotes to examine a functional overlap between Adamts9 and 20. While mice heterozygous for mutation of Adamts20 and 9 are viable, bt9J/bt9J;Adamts9ko/+ animals die shortly after birth (H. Enomoto and S. Apte, submitted). We examined pigmentation in newborn animals that were wild-type, bt9J/+;Adamts9ko/+, bt9J/bt9J, or bt9J/bt9J;Adamts9ko/+. Skin pelts containing dorsal and ventral surfaces and spanning the entire trunk region were removed from newborn pups (Figure 8A). Although not externally obvious until 4–5 days after birth, melanoblasts can be seen as clusters of pigmented cells on the dermal side of the P0 epidermis. In all skin samples, pigmentation was apparent primarily on dorsal surfaces as shown by schematic representations in Figure 8B–8E. In the skin from +/+ and bt9J/+;Adamts9ko/+ animals, pigmentation was distributed uniformly between anterior and posterior boundaries of the skin samples (n = 4 each) (Figures 8B and 8C). In the bt9J/bt9J skin samples (n = 6), a partial de-pigmentation was observed on about 50% of the anterior-posterior length, consistent with the likely boundaries of “belted” and “non-belted” regions (Figure 8D). However, in the bt9J/bt9J;Adamts9ko/+ skin samples (n = 5), de-pigmentation was apparent on 75% of the anterior-posterior length (Figure 8E). Our analyses showed a statistically significant increase in the severity of de-pigmentation in bt9J/bt9J;Adamts9ko/+ skin compared to that seen in bt9J/bt9J skin (Figure 8F). In contrast, the pigmentation phenotype in btBei1/btBei1;Adamts5ko/Adamts5ko mice was similar to that of btBei1/btBei1 animals indicating that loss of Adamts5 does not exacerbate the bt9J/bt9J phenotype and that unlike Adamts9, Adamts5 does not participate in melanoblast development (D. McCullough and S. Apte, unpublished data).
In order to assess melanoblast development in bt9J/+;Adamts9ko/+and bt9J/bt9J;Adamts9ko/+ mice, E13.5 embryos were examined using in situ hybridization with Pmel17 (GeneID: 20431), a marker specific for melanoblasts. Similar to what was observed in newborn skin, there was a significant reduction in melanoblasts in the trunk region of bt9J/bt9J;Adamts9ko/+ embryos compared to bt9J/+;Adamts9ko/+ embryos (data not shown). Importantly, there was also a significant reduction of melanoblasts in regions outside of the trunk (Figure 8G and 8H). These results indicate that Adamts9 and Adamts20 are partially redundant during melanoblast development not only in the trunk but in the head as well.
Analysis of mouse coat color mutants has yielded mechanistic insights into normal developmental pathways as well as human disease processes. In this paper we explored how mutation of Adamts20, a secreted metalloprotease that is not expressed in melanoblasts, can disrupt melanoblast development and cause white spotting. The expression pattern of Adamts20, along with studies performed on the C.elegans ortholog Gon-1, initially suggested that Adamts20 regulated melanoblast migration [12], [27]–[29]. However, rather than this predicted requirement for melanoblast migration, we have identified an unexpected role for Adamts20 in melanoblast survival. Our results have implications not only for understanding mechanisms of melanoblast development and pigmentation-associated diseases but also for the biological functions of secreted metalloproteases.
The most striking alteration observed in bt9J/bt9J embryos was a reduction in trunk melanoblasts beginning at E13.5. Early events of melanoblast specification, migration, proliferation and survival were not affected, as the number and distribution of Dct-LacZ-positive melanoblasts was similar between control and bt9J/bt9J embryos up to E12.5. This result contrasts many melanoblast mutants including Kit, Mitf, Pax3 (MGI: 97487), Ednrb (MGI: 102720), and Sox10 (MGI: 98358), in which melanoblast defects can be observed as early as E10.5-E11.5 [15].
Multiple lines of evidence also indicate that later stages of melanoblast migration were not defective in bt9J/bt9J embryos. Dct-LacZ-positive cells did not accumulate dorsally or in the dermis, suggesting that melanoblast migration along the dorso-lateral pathway and between skin layers was normal. It is also unlikely that melanoblasts undergo cell death as a result of or subsequent to failed migration, as apoptotic melanoblasts were apparent in both dorsal and ventral regions of mutant embryos and in similar proportions within dermal and epidermal compartments. Our results also exclude a defect in the second wave of melanoblast migration [37],[60], since melanoblasts did not accumulate at the belt edges in E16.5 embryos. Instead, our results show that reduction of melanoblasts is due to the seven-fold increase in apoptosis observed in the trunk region of Adamts20 mutant embryos.
We propose that the white belt occurs in Adamts20 mutants for a combination of three factors: uneven densities of melanoblasts along the length of normal embryos, regional differences in apoptosis, and functional redundancy with other metalloproteases.
In wild-type animals there are fewer embryonic melanoblasts in the trunk relative to other areas of the body, as shown in this study and by others [17], [34]–37. Related to this, many of the coat color mutants exhibit spotting in the lumbar level of the trunk suggesting that melanoblasts in this region are particularly sensitive to genetic perturbation, perhaps due to reduced numbers. This suggests that if melanoblast numbers were further reduced along the trunk in Adamts20 mutants, the resulting belt would be wider. Indeed this was the case since genetic interactions of bt9J/bt9J with Kit mutations caused a wider belt as opposed to increased spotting in the head.
Second, Adamts20 is required for melanoblast survival throughout the embryo, consistent with the expression pattern of Adamts20 along the length of the embryo. In bt9J/bt9J embryos, we found significant increases in apoptosis in both head and trunk regions. Importantly, the proportion of melanoblasts undergoing apoptosis was much more pronounced in the trunk than in the head. Thus a combination of increased apoptosis and lower initial melanoblast number in the trunk region of a bt9J/bt9J mutant depletes melanoblast numbers to levels at which it cannot generate adequate pigmentation.
Third, Adamts20 mutant animals have limited white spotting due to functional redundancy with other metalloproteases, including Adamts9. In a bt9J/bt9J background, loss of a single copy of Adamts9 results in reduced melanoblast numbers in the trunk as well as in the head. Thus, in bt9J/bt9J mice, Adamts9 can compensate for Adamts20 deficiency in regions of higher melanoblast number such as the head. The early embryonic lethality of Adamts9ko/Adamts9ko mice makes it impossible to currently assess the absolute requirement of Adamts9 in pigmentation. However, the widespread expression of Adamts9 in the skin paired with the dramatic defect in melanoblast development in bt9J/bt9J;Adamts9ko/+ animals predict an even greater pigmentation defect in a homozygous background. These future analyses will depend upon the generation of a conditional Adamts9 knock-out in the skin.
Our results suggest that Adamts20 is required for melanoblast survival at least in part through modulating Kit signaling. We show that bt9J/bt9J;Kittm1Alf/+ mutants exhibit synergistic spotting, and melanoblasts in bt9J/bt9J skin cultures are unable to respond to sKitl. This mechanism is consistent with the requirement of Kit signaling for melanoblast survival as well as the overlapping expression of Adamts20, Kit and Kitl in the dermis [12], [17], [20], [24]–[26],[61].
Kit signaling is essential for several events in melanoblast development beginning at E10.5 and including both migration and survival. Yet our findings indicate that mutation of Adamts20 modulates only a subset of Kit functions, namely survival, and only during a defined window of time, around E13.5, rather than throughout development. This is evidenced by the fact that bt9J/bt9J embryos do not show altered melanoblast development prior to E13.5, such as seen in Kit/+ embryos [17],[18]. In addition, Kit heterozygous mutations do not exacerbate the onset of the melanoblast defect in E12.5 bt9J/bt9J embryos, as would be expected if Adamts20 were required earlier. The stage at which Adamts20 disrupts melanoblast development may also explain why bt mice do not exhibit extensive ventral spotting like other white spotted mutants that impede earlier events in development.
Given these observations, it is notable that restricted disruption of Kit, during the embryonic period associated with onset of the bt9J/bt9J phenotype, results in mice with spotting phenotypes similar to bt [23],[62]. Pregnant mice injected with neutralizing Kit antibodies between E10.5 and E13.0 give birth to mice exhibiting white belts. In contrast, injection of these Kit antibodies into pregnant animals prior to E10.5 and after E13.0 results in mostly unpigmented mice.
A white belt/band is also evident in KitW-sh/+, KitW-57J/KitW-57J, and KitW-bd/+ mice that carry mutations in Kit regulatory sequences and cause altered Kit expression patterns [35],[63],[64]. In addition, Kimura and colleagues describe belted phenotypes in Kittm1Ber and Kittm2Ber mice containing targeted mutations of Kit at tyrosine residues 567 and 569, amino acids essential for proper phosphorylation and downstream signaling [65]. It is interesting that disruption of these specific amino acids results in spotting similar to that seen in belted mice, and we speculate that Adamts20 may be required for efficient Kit phosphorylation at these sites in melanoblasts in vivo.
There are several possible mechanisms by which ADAMTS20 could regulate Kit signaling. ADAMTS20 could be required directly for cleavage of either Kit and/or Kitl to produce sKitl. For example, the ADAM family member ADAM17 (GeneID: 11491) cleaves Kit in vitro, and ADAM17, 19 (GeneID: 8728) and 33 (GeneID: 110751) cleave Kitl [66]–[69]. However, sKitl did not rescue the melanoblast defect in bt9J/bt9J skin cultures as would be expected if Kitl were an essential substrate of ADAMTS20. As further evidence that the bt phenotype is not due to defective Kitl cleavage, bt9J/bt9J exhibited similar genetic interactions with both a Kitl null allele and with KitLSl-d/+ animals, which produce a short sKitl [42],[43],[70]. Kunisada et. al have also shown that transgenic mice over-expressing Kitl transgene, which greatly reduces the white spotting associated with Kit mutants, only slightly reduces the spotting in bt9J/bt9J animals [71]. Taken together this suggests that the bt phenotype is not due to defective cleavage of Kitl. Instead, Adamts20 may regulate the interaction of Kitl with Kit, the activation of Kit, or signaling downstream of the receptor. Future biochemical studies will be necessary to define the exact mechanism by which Adamts20 modulates Kit signaling.
The observation that melanoblasts in bt9J/bt9J trunk explants did not exhibit any response to soluble Kitl indicates that Kit signaling is defective, but it is important to note that this does not exclude the possibility that Adamts20 may additionally regulate other factors essential for melanoblast development. For example, Adamts20 could be required to activate other pathways that act synergistically with Kit to regulate melanoblast survival.
Another possible mechanism by which Adamts20 could modulate survival is by altering the extracellular matrix in which melanoblasts are located. We show that bt9J/bt9J mutants exhibit reduced cleavage of at least one ECM component, versican. Versican, Adamts20, and Kitl are expressed in embryonic skin at appropriate times and sites to regulate melanoblast development [12],[24],[72]. In fact, sKitl enhances melanoblast proliferation to a greater extent when primary NC cells are cultured on chondroitin sulfate, suggesting that the ECM can modulate Kit signaling [73]. Taken in context of this study, it could be melanoblasts in bt9J/bt9J explant cultures did not respond to sKitl because the ECM was defective.
Versican proteolysis could influence melanoblast behavior and Kit signaling by promoting direct interactions of Kitl with Kit receptor. Such a scenario is analogous to heparin sulfate proteoglycans, which modulate growth factor signaling by sequestering ligand and promoting its receptor binding [74]. Studies of Weill-Marchesani syndrome strongly suggest ADAMTS10 (GeneID: 81794) regulates TGF βsignaling through its interactions with the ECM component, fibrillin-1 (GeneID: 2200) [9]. Interestingly, versican binds extracellular cytokines, including the secreted growth factor, midkine (GeneID: 17242) [75]–[77]. Future studies will reveal if versican, Kit, and Kitl physically interact and if versican regulates Kit signaling directly.
Versican and its proteolytic products could also promote melanoblast survival through a pathway that is independent of Kit signaling. Versican V1 promotes survival of NIH3T3 cells and down-regulates expression of the pro-apoptotic protein Bad (GeneID: 12015) [78]. Expression of the versican G1 domain protects sarcoma cells from apoptosis, and binding of the versican G3 domain to β-Integrin promotes survival of astrocytoma cells [79],[80]. Since integrins are expressed by melanoblasts [81],[82], versican-integrin interactions are another potential mechanism by which versican could influence melanoblast development. Since versican mutants are embryonic lethal around E10.5 [83], conditional models of versican as well as transgenic animals containing constitutively cleaved and uncleaved Versican molecules will be necessary to assess the requirement of versican for melanoblast development.
Given that cell survival is an integral component of melanoma progression, ADAMTS9 and ADAMTS20 are excellent candidates for participating in melanoma. ADAMTS20 over-expression has not been carefully examined in melanoma, but has been observed in brain, colon and breast tumors [47],[84]. The related metalloproteases ADAM9 (GeneID: 8754) and ADAMTS13 are upregulated in primary melanoma tumors and in melanoma cell lines, respectively [85],[86]. Interestingly, VERSICAN and KITL overexpression are observed in primary melanomas and levels correlate with melanoma progression [87]–[91]. Given the requirement of Adamts20 for Kit signaling and versican cleavage, it is intriguing to consider how dysregulation of signaling between ADAMTS20, Kit, and versican might contribute to melanoma progression.
The following mouse stocks were used and kindly provided by: Adamts20bt-Bei1 (David Beier, Harvard Medical School), Adamts20bt9J (Lynn Lamoreux, Texas A&M University), Dct-LacZ, MiMi, Kittm1Alf, KitlSl (Heinz Arnheiter, NINDS), KitSl-d (Jackson Laboratories, Bar Harbor, ME). Adamts9ko and Adamts5ko knockout animals (both on a C57Bl/6 background) are described elsewhere (D. McCullough, H. Enomoto, S. Apte, unpublished). Quantification of white spotting was performed using Image J software (http://rsb.info.nih.gov/ij) (NIH, Bethesda, MD) and statistical significance of spotting calculated using ANOVA tests.
cDNA was sequenced by GeneDx (Gaithersburg, MD). Adamts20bt9J animals were genotyped using standard conditions on an ABIPrism7000 using a TaqmanTM assay of genomic DNA. For this assay, a region containing the point mutation is amplified by PCR. Two fluorescently labeled single-stranded oligonucleotides (probes), one complementary to the wild-type product, and one complementary to the mutant product are included in the assay. The relative amount of wild-type and mutant product is measured by fluorescence quenched upon DNA synthesis. Allelic discrimination was performed to detect the total levels of each allele at the conclusion of the PCR reaction. The following cycling conditions were used: 1. 95°C for 10 minutes 1×, 2. 92°C for 15 seconds, 3. 55°C for 60 seconds (2 and 3 repeated 40 times). The following primers and probes were used: TTCAGCACAGCTATTCTGGAAGAC (forward primer), GCACCTGAGGCAGACATACAC (reverse primer), CTTACCGAGATAGTTGTC (VIC probe, C57BL/6 allele), and CTTACCGAAATAGTTGTC (FAM probe, bt9J allele) (designed using Applied Biosystems software, Foster City, CA).
For quantitation of whole mount embryos, melanoblasts were counted from both sides of each embryo within a 1 mm×0.7 mm field surrounding the eye or the trunk. For quantitation of melanoblasts in E13.5 sections, Dct-LacZ expression was scored. The mitotic index was calculated using the average number of PH3+ melanoblasts per section divided by the average number of total melanoblasts per section and multiplied by 100. The apoptotic index was calculated using the average number of CC3+ melanoblasts per section divided by the average number of total melanoblasts per section and multiplied by 100. For quantitation of melanoblasts in different skin compartments the following criteria were used: melanoblasts above the basal membrane were classified as epidermal, melanoblasts below the basal membrane were classified as dermal, and melanoblasts crossing the basal membrane and present in both regions were classified as dermal/epidermal region. The following statistical tests were used: for average number of melanoblasts in sections and whole mount embryos (two-tailed Student's t-test), for proportion of total melanoblasts and apoptotic melanoblasts in skin layers (Chi-square test of independence), for proportion of proliferating and apoptotic melanoblasts (Fisher's test).
β-galactosidase activity staining was performed as previously described [41]. In situ hybridization was performed as previously described [92] using a 1 Kb mouse Adamts20 probe [12], a 644 bp mouse Versican V1 probe made using mouse Versican V1 cDNA (a kind gift of Andrew Copp, UCL), a Pmel17 probe made as previously described [92], and a 1 Kb mouse Adamts9 probe [58].
Frozen sections were prepared from embryos fixed overnight in 4% paraformaldehyde. Staining was performed with modifications to manufacturer's protocol (Vector Laboratories, Burlingame, CA), and sections were mounted using Hardset Vectashield with DAPI (Vector Laboratories). The following antibodies were used at 1∶200: mouse anti- β galactosidase (Promega, Madison, WI), rabbit anti-cleaved Caspase-3 (Cell Signaling, Danvers, MA), rabbit anti-phospho-Histone H3 (Upstate Biotechnology, Charlottesville, VA), rabbit anti-versican GAG β (Chemicon, Temecula, CA), rabbit anti-versican V0/V1 Neo (Affinity Bioreagents, Golden, CO), anti-mouse and anti-rabbit rhodamine-conjugated or FITC-conjugated secondary antibodies (Vector Laboratories).
Adamts20 short isoform cDNA was assembled by PCR from E17.5 mouse embryo mRNA, sequence verified, and cloned into pcDNAmyc-hisA+ plasmid (Invitrogen, Thousand Oaks, CA). The expression plasmid for full-length Adamts9 was previously described [47]. Transfected or untransfected cells were incubated with versican as previously described [47]. Extracts of embryo skin were prepared using a RIPA lysis buffer containing protease inhibitor mixture (20 µl) and 2 mM phenylmethylsulphonyl fluoride (Pierce, Rockford, IL). For analysis of total versican, cell extracts were treated with 0.1 units chondroitinase ABC (Associates of Cape Cod, East Falmouth, MA) in 0.1 M Tris, 50 mM NaAc for 10 minutes at 37°C. Skin extracts and conditioned medium from transfected cells were run on 4–12% SDS-Polyacrylamide gels (Invitrogen) and analyzed by Western blotting using the following primary antibodies: mouse anti-myc 9E10 (Invitrogen), rabbit anti-versican GAG βat 1∶1000, rabbit anti-versican V0/V1 Neo at 1∶1000, and mouse anti-α-Tubulin at 1∶1000 (Sigma, St.Louis, MO); and secondary antibodies: anti-rabbit HRP and anti-mouse HRP (Amersham, Piscataway, NJ).
Dorsal skin between the forelimb and hindlimb was isolated from E13.5 bt/+;Dct-LacZ and bt/bt;Dct-LacZ embryos with ventral tissue portions and internal organs removed. The skin explants were placed epidermal side up onto polyethylene terepthlate track-etched membranes in a cell culture insert (8.0 µm pore size, Becton-Dickinson, Franklin Lakes, NJ). Separate cultures were established for each explant. Explants were cultured for four days with 95% DMEM, 5% FBS (control medium) and as noted the medium was supplemented daily with 500 ng/ml Kitl (R&D Systems, Minneapolis, MN). After a culture period of four days, the explants were fixed in 4% paraformaldehyde in PBS (pH 7.4) for one hour and staining for β-galactosidase activity was performed as previously described [41].
Skin was removed from newborn pups, fixed in 4% paraformaldehyde overnight and washed 3× in PBS and then analyzed on the dermal side for pigmentation. Pigmentation was scored as the extent of pigmentation between anterior and posterior edges of samples, using the following scale: 1 (no pigmentation), 2 (between 0 and 25%), 3 (between 25 and 50%), 4 (between 50 and 75%), 5 (between 75% and 100%). Statistical significance was calculated using a student's t test. |
10.1371/journal.ppat.1005049 | Evaluating Human T-Cell Therapy of Cytomegalovirus Organ Disease in HLA-Transgenic Mice | Reactivation of human cytomegalovirus (HCMV) can cause severe disease in recipients of hematopoietic stem cell transplantation. Although preclinical research in murine models as well as clinical trials have provided 'proof of concept' for infection control by pre-emptive CD8 T-cell immunotherapy, there exists no predictive model to experimentally evaluate parameters that determine antiviral efficacy of human T cells in terms of virus control in functional organs, prevention of organ disease, and host survival benefit. We here introduce a novel mouse model for testing HCMV epitope-specific human T cells. The HCMV UL83/pp65-derived NLV-peptide was presented by transgenic HLA-A2.1 in the context of a lethal infection of NOD/SCID/IL-2rg-/- mice with a chimeric murine CMV, mCMV-NLV. Scenarios of HCMV-seropositive and -seronegative human T-cell donors were modeled by testing peptide-restimulated and T-cell receptor-transduced human T cells, respectively. Upon transfer, the T cells infiltrated host tissues in an epitope-specific manner, confining the infection to nodular inflammatory foci. This resulted in a significant reduction of viral load, diminished organ pathology, and prolonged survival. The model has thus proven its potential for a preclinical testing of the protective antiviral efficacy of HCMV epitope-specific human T cells in the evaluation of new approaches to an immunotherapy of CMV disease.
| Pre-emptive CD8 T-cell therapy of human cytomegalovirus (HCMV) disease in immunocompromised recipients of hematopoietic stem cell transplantation gave promising results in clinical trials, but limited efficacy and the need of HCMV-seropositive memory cell donors has so far prevented adoptive cell transfer from becoming clinical routine. Further development is currently hampered by the lack of experimental animal models that allow preclinical testing of the protective efficacy of human T cells in functional organs. While humanized mouse models with human tissue implants are technically and statistically demanding, and are limited to studying human T-cell activation and local virus control in the implants, a more feasible model for control of systemic infection and prevention of multiple-organ CMV disease is regrettably missing. Here we introduce such a model based on infection of genetically immunocompromised, HLA-A2.1-transgenic NOD/SCID/IL-2rg-/- mice with a chimeric murine CMV engineered to express the HCMV NLV-peptide epitope. Mimicking the scenario of HCMV-unexperienced donors, human T cells transduced with a human T-cell receptor specific for HLA-A.2.1-presented NLV peptide controlled systemic infection and moderated organ disease resulting in a survival benefit. The model promises to become instrumental in defining T-cell properties that determine their protective efficacy for a further development of adoptive immunotherapy of post-transplantation CMV infection.
| Reactivation of latent human cytomegalovirus (HCMV) infection is a frequent complication in patients after allogeneic hematopoietic stem cell transplantation (HSCT). Although potent antiviral drugs are available, their usage, however, is often limited by hematotoxicity and nephrotoxicity. In addition, the broad application of these drugs during pre-emptive treatment strategies is associated with a higher frequency of late-onset HCMV disease [1,2]. Preclinical research in murine models ([3–6], reviewed in [7–9]) as well as clinical phase I/II trials ([10–12], reviewed in [13,14]) have shown that the adoptive transfer of virus-specific CD8 T cells is a promising therapeutic option for preventing and treating CMV disease. However, the feasibility of HCMV-specific immunotherapy is currently impeded in clinical routine due to technical restrictions. It has also limitations in case the donor is HCMV-seronegative or carries only low frequencies of HCMV-specific memory T cells. In this situation, transduction of non-cognate T cells with virus specific T-cell receptors (TCR) may be an alternative means to transfer HCMV-specific T-cell function into HSCT recipients [15,16]. In any case, clinical protocols need to be improved before HCMV-specific cell therapy can be implemented in general clinical practice. To allow for a more reliable analysis of HCMV immunotherapies (e.g. adoptive T-cell therapy, therapeutic vaccination) animal models that mimic HCMV infections are needed.
Since HCMV replication is strictly restricted to cells and tissues of human origin ([17], reviewed in [18]), previous animal models utilized murine CMV (mCMV) as surrogate virus (reviewed in [7–9]) or mice infected with HCMV after implantation with human cells or tissues, for instance with tumor cell lines, fetal thymus, and liver biopsies ([19–23], reviewed in [24]). The implantation approach has shown that HCMV strains replicate locally with differences in pathogenicity, but fail to spread between tissue implants. To support systemic infection, Smith et al. [25] infected human CD34+ hematopoietic stem cell-engrafted mice with HCMV to establish latency and to induce virus reactivation in tissue-migrated monocytes and macrophages by granulocyte-colony stimulating factor (G-CSF) treatment. By model design, however, viral dissemination to functional organs relevant for viral pathogenesis (e.g. spleen, lungs, and liver) and transmission (salivary glands) cannot be assessed even in these advanced humanized mouse models.
We herein present a novel preclinical mouse model that allows the direct testing of HCMV-specific human T-cell products. In this, we combined the well-described murine model of mCMV infection of the immunocompromised host (reviewed in [7–9]) with the strong T-cell immunogenicity of the HLA-A*0201 (HLA-A2.1) restricted HCMV epitope pp65495-503 NLVPMVATV (briefly, NLV) [26]. We generated a chimeric recombinant mCMV expressing the NLV epitope (mCMV-NLV) during the infectious cycle to allow organ manifestation of the infection in the natural host similar to that seen in immunocompromised patients. After infection of HLA-A2.1 transgenic, constitutively combined-immunodeficient NOD/SCID/IL-2rg-/- (NSG/HHD) mice, which lack cells of adaptive immunity and are additionally deficient in natural killer (NK) cells [27], mCMV-NLV resulted in a rapid systemic infection that could be effectively combated by adoptively transferred human NLV-specific CD8 T cells as well as by human T cells transduced with an NLV-specific TCR. After migration into murine organs, these T cells not only reduced mCMV-NLV titers in an epitope-specific manner but also prolonged survival of infected mice, despite continued combined immunodeficiency in absence of hematopoietic reconstitution.
For the clinical correlate of reactivated HCMV infection associated with HSCT, our findings predict that pre-emptive CD8 T-cell immunotherapy, even though controlling CMV only transiently, can bridge the critical time of an immunocompromised state until immune reconstitution by HSCT takes over. Altogether, we provide a novel animal infection model that allows the direct evaluation and preclinical testing of HCMV-specific human T cells and potentially other HCMV immunotherapy approaches, including vaccination of HSCT recipients in parallel to hematopoietic reconstitution.
The key strategy of our model was to antigenically ‘humanize’ both the murine virus and the murine host tissue for HLA class I-restricted recognition of infected cells by human T cells in functional host organs. For construction of the chimeric virus mCMV-NLV, we integrated the coding gene sequence of the HLA-A2.1-restricted pp65 (UL83)495-503 peptide epitope NLVPMVATV (abbreviated: NLV) of HCMV, with its natural flanking amino acids for authentic proteasomal cleavage and precursor peptide sequence, into the immediate-early (IE)2/m128 gene of mCMV-Δm157 (Fig 1A). Although NSG/HHD mice are deficient in NK cell activity due to lack of the interleukin-2 receptor (IL-2R) common γ chain (CD132) causing deficient IL-2/4/7/9/15/21 signaling [28,29], mCMV-Δm157 was chosen as parental virus to also formally exclude activation of the Ly49H+ subset of murine NK cells by m157-Ly49H interaction, a feature that is only valid in the C57BL/6 genetic background and might thus interfere with a broader application of the model (for a review, see [30]). For control experiments, a lack-of-epitope mCMV-NLV derivative was generated, in which a single amino acid replacement at the C-terminal HLA-A2.1 anchor position valine (V503) by alanine (mCMV-NLVAla) prevents epitope generation in the proteasome and MHC/HLA class-I (MHC/HLA-I) presentation (Fig 1A) (for the concept, see [31,32]).
The IE2/m128 gene encodes a regulatory but non-essential protein expressed in the IE phase of viral replication [33,34] and sporadically also during viral latency [35]. To verify, in the first place, that integration of the NLV epitope into the IE2 protein does not attenuate the virus, which otherwise would render the model inapplicable for studying viral pathogenesis, we compared the in vivo growth kinetics of parental virus mCMV-Δm157 and its antigenicity variant mCMV-NLV in NSG/HHD mice. In selected organs relevant for viral pathogenesis, such as liver, spleen, and lungs (Fig 1B), the two viruses were found to replicate in a log-linear fashion with comparable doubling times within each organ (for the principle, see [36,37]), though with organ-typic numerical differences, with most aggressive growth in the lungs. In conclusion, insertion of the NLV epitope did not impair in vivo viral growth. The chimeric virus thus fulfills the model’s first prerequisite of an unaltered replicative fitness.
The second prerequisite of the model is that the NLV peptide is actually generated by processing of the chimeric IE2-NLV protein in mCMV-NLV infected cells and that it is presented in association with HLA-A2.1 at the cell surface for recognition by TCR. To directly assess presentation, an endpoint which includes the preceding processing events, we chose two independent NLV peptide-specific, cytolytic CD8 T cell lines (CTLL) as probes: (i) a published, peptide-selected murine long-term CTLL, mCD8-NLV, expressing a murine TCR [38] for serving as a murine reference line, and (ii) a freshly generated peptide-selected human short-term polyclonal CTLL, hCD8-NLV, expressing human TCRs for modeling the situation of an HCMV-experienced, seropositive cell transfer donor as the clinical correlate. While mCD8-NLV consisted of cells with effector phenotype (CD44+CD62Llow TE/EM), including effector-memory cells, early effector cells, and short-lived effector cells ([39,40], reviewed in [41,42]) the short-term hCD8-NLV line was composed of cells with central memory and effector memory phenotype, TCM and TEM, respectively (S1 Fig). The functional avidities of these two CTLL were then tested in an IFN-γ secretion-based ELISpot assay for their sensitization by HLA-A2.1 transgenic NSG/HHD mouse embryonic fibroblasts (HLA-A2.1-MEF) exogenously loaded with graded concentrations of synthetic NLV peptide (Fig 2A). Notably, when corrected for background response measured with an unrelated peptide, both CTLL were similarly sensitive with an endpoint NLV peptide concentration of 10−9 M. Pretreatment of the HLA-A2.1-MEF stimulator cells with IFN-γ, which is known to enhance MHC class-I expression [43], only modestly increased the numbers of CTLL cells responding to exogenously peptide-loaded HLA-A2.1 molecules. Contrasting with the comparable functional avidities of the two CTLL, an assay based on their physical capacity to bind NLV peptide-folded HLA-A2.1 tetramers [44,45] revealed a higher structural avidity of mCD8-NLV (Fig 2B), which predicts more efficient in vivo function [8,46]. After this basal characterization of the two CTLL, we tested their capacity to detect naturally processed NLV peptide presented by HLA-A2.1 on the surface of infected NSG/HHD HLA-A2.1-MEF (Fig 2C). In both lines, a significant, though low, number of cells had an avidity sufficient for recognizing NLV-HLA-A2.1 complexes presented on infected cells. This number, however, was significantly enhanced when NSG/HHD HLA-A2.1-MEF were treated with IFN-γ prior to infection, which contrasts to the only modest effect of IFN-γ on exogenous peptide loading. The explanation likely is the overriding of the function of viral immune evasion molecules (viral regulators of antigen presentation, vRAPs) by IFN-γ, as described previously to occur in infected cells in vitro as well as in infected tissues in vivo ([47,48], reviewed in [49]).
In this context it is worth noting that the immune evasion molecules of mCMV indeed also target the transgenic HLA-A2.1 (S2 Fig). Specifically, as shown by flow cytometry, immune evasion mechanisms of mCMV caused cell surface down-modulation of both murine H2-Kd and the transgenic human HLA-A2.1 molecules in infected (gp36.5/m164+) but not in uninfected (gp36.5/m164-) cells of the same infected cell cultures when compared to uninfected control cultures. Accordingly, H2-Kd and HLA-A2.1 expression levels remained largely unaffected upon infection with the immunoevasin gene deletion mutant mCMV-ΔvRAP (S2 Fig) [48,50,51].
Importantly, both the normal and the IFN-γ-enhanced recognition of infected cells proved to be strictly NLV epitope-specific, as shown by baseline values of response to NSG/HHD HLA-A2.1-MEF infected with virus mCMV-NLVAla (Fig 2C).
After fulfillment of the model’s prerequisites of unaltered virulence and of successful transgenic antigen presentation, both NLV peptide-specific CTLL, characterized as detailed above, were tested for their capacity to control the replication of chimeric virus mCMV-NLV in tissues of HLA-A2.1 transgenic NSG/HHD mice, which can be engrafted with components of the human immune system due to a lack of an adaptive immune system and innate immune defects [27]. Since NSG/HHD mice are devoid of endogenous murine T cells, tissue infiltration by T cells and antiviral control can be attributed exclusively to the transferred cells.
Specifically, sublethally (2 Gy) γ-irradiated NSG/HHD mice were infected by intraplantar injection with 1x105 PFU of mCMV-NLV, a dose that would be lethal after approximately 12 to 17 days. Shortly thereafter, NLV-specific CTLL were intravenously transferred in order to mimic the clinical setting of pre-emptive therapy, and viral titers in spleen and lungs as well as numbers of infected cells in livers of individual mice were determined on day 11 (for the protocol scheme, see Fig 3A). In the case of transfer of human NLV-specific CTLL, single doses of IL-2 and IL-7 were given along with the cells for a short-term growth factor supply. In the beginning of the project, a key concern had been that a cross-species adoptive transfer of CD8 T cells might fail because of xenogeneic constraints in the chemokine-driven extravasion and tissue infiltration of intravenously transferred CD8 T cells. Not unexpectedly, based on extended literature to adoptive immunotherapy by CD8 T cells in fully murine models (reviewed in [7–9]), transfer of mCD8-NLV cells controlled the infection of spleen, lungs, and liver in a dose-dependent and epitope-specific fashion (Fig 3B). As shown for the liver, dose-dependent antiviral protection correlated with dose-dependent tissue infiltration by the adoptively transferred CD8 T cells. Likewise, lack of epitope after infection with mCMV-NLVAla abolished both tissue infiltration and control of virus replication. Importantly, as a new information, the results were qualitatively identical for the human CD8 T cells, hCD8-NLV, although higher cell numbers were needed (Fig 3C). Whether this relates to the lower structural avidity (recall Fig 2B) or to xenogeneic differences remains to be investigated. The more relevant conclusion, however, is that the functionality of antiviral human CD8 T cells can, in principle, be studied in murine tissues.
Protection against disseminated CMV organ infection has a microanatomical correlate in the formation of nodular inflammatory foci (NIF) [52–57]. As shown previously in a related murine model [54] and reproduced here for the example of mCD8-NLV cells, infection by a virus expressing the cognate epitope is confined to NIF where control takes place and where infected tissue cells and infiltrating CD8 T cells co-localize. In contrast, in the absence of the cognate epitope [54] after infection with virus mCMV-NLVAla, CD8 T cells do not infiltrate infected tissue at all, can thus not form NIF, and, accordingly, allow more widespread infection visible as many foci of infection (IF), with the consequence of virally-caused tissue pathology (Fig 4A). Importantly, these rules were found to apply, in principle, also to the human CD8 T cells, hCD8-NLV, although NIF appeared as being somewhat less condensed (Fig 4B). This may relate to the lower structural avidity (recall Fig 2B) and the higher cell numbers needed for antiviral protection (recall Fig 3C). Again, although the syngeneic cell transfer system is more efficient, which is not surprising, a xenogeneic system also allows NIF formation and control of intra-tissue virus spread, despite the species barrier.
The need of an HCMV-experienced (HCMV antibody positive) donor as a source for virus-specific memory CD8 T cells is a limitation for an adoptive immunotherapy of HCMV reactivation in the highest-risk constellation of an HCMV-seronegative donor and—seropositive recipient (D-R+) [58–61], which can currently only be overcome by an HLA-matched third party memory T-cell donor [12,62,63]. An alternative approach, already successfully applied in tumor models [64,65], is the transduction of T cells derived from the HLA-matched HSCT donor with a virus-epitope specific TCR.
To model this clinical need, a first application of the here newly described HLA-A2.1 transgenic mouse CMV infection model was to test the in vivo antiviral function of human CD8 as well as CD4 T cells that were transduced by retroviral gene transfer of a modified high-affinity human T-cell receptor α/β (TCRNLV) previously shown to reprogram CD8 and CD4 T cells derived from HCMV-seronegative donors [16]. As a basal characterization of the TCRNLV transduced T cells enriched by drug selection, cytofluorometric analysis with NLV peptide-folded HLA-A2.1 tetramers for TCRNLV staining shows high transduction efficacies (Fig 5A), and Fig 5B and 5C reveal structural avidity as well as NLV epitope-specific cytolytic potential of CD8-TCRNLV and CD4-TCRNLV cells, respectively. Using NSG/HHD HLA2.1-MEF exogenously loaded with a high dose of synthetic NLV peptide as a positive standard, both transduced T-cell lines specifically lysed IFN-γ pre-treated mCMV-NLV-infected NSG/HHD cells, overcoming the presence of the mCMV immune evasion molecules (see above, [48]), and this lysis was TCR/epitope-specific as indicated by its absence after mock-transduction of the T cells as well as after infection of the target cells with the antigenicity-loss variant mCMV-NLVAla. Notably, CD8-TCRNLV CTL, but not mock-transduced CD8 T cells, lysed HLA-A2.1-expressing human primary foreskin fibroblasts infected with the HCMV immune evasion gene (US2,3,6,11) deletion mutant RV-KB6 expressing the NLV peptide [66], but not those infected with the combined US2,3,6,11 and pp65/UL83 deletion mutant HCMV-RV-KB15 lacking NLV peptide [16] (S3 Fig). This, again, showed the specificity in terms of the need for the specific TCR and the cognate epitope presented by HLA-A2.1. For completing the basal characterization, S4A and S4B Fig show the phenotyping of the human CD8 and CD4 T cells, respectively, each before (upper row) and after (lower row) TCRNLV transduction and subsequent lymphokine-driven expansion triggered by ligation of CD3 and CD28. In essence, and in accordance with previous studies on TCR transduction, the resulting CD8-TCRNLV population consisted primarily of TEM [67–70], whereas CD4-TCRNLV cells were predominantly TCM.
These cells, and the corresponding empty-vector transduction controls lacking cognate TCR, were then adoptively transferred into mCMV-NLV-infected NSG/HHD mice to test their in vivo antiviral efficacy (Fig 6). With the exception of the liver, where a tendential reduction in day 11 median virus titer did not reach statistical significance due to high variance in the therapy group, infectious virus load was found to be significantly reduced in spleen (P < 0.001) and lungs (P = 0.014) by 107 CD8-TCRNLV cells but not by 107 CD4-TCRNLV cells (P >0.05 throughout). This is important new information, as it contrasts with the cytolytic activities of both cell populations (recall Fig 5B and 5C), indicating that in vitro cytolytic activity against IFN-γ pre-treated, infected cells is not reliably predictive for an antiviral function in vivo.
Interestingly, CD4-TCRNLV cells, though not antiviral effectors on their own (see above), enhanced the antiviral function of CD8-TCRNLV cells when co-administered as a 1:4 mixture of 2 x 106 non-protective CD4-TCRNLV and 8 x 106 protective CD8-TCRNLV cells (spleen, P < 0.001; lungs, P = 0.009; liver, P = 0.049). Specifically, at least in spleen and lungs, this mixture turned out to be more efficient than the higher number of 107 protective CD8 T cells (P = 0.037 and P = 0.005, respectively), which indicates synergism in the sense of a helper effect of the CD4 T cells. Protection by the mixture was clearly NLV epitope-specific, as any antiviral function was abolished when the NSG/HHD mice were infected with the antigenicity-loss mutant mCMV-NLVAla.
Antiviral control, determined on day 11, should be preceded by migration of transferred cells into the tissues. As shown for spleen and liver (S5 Fig), CD8 and CD4 T cells could be recovered on day 3 after intravenous cell transfer from spleen leukocyte and liver non-parenchymal cell populations. Notably, after transfer of 107 CD8-TCRNLV cells in absence of CD4-TCRNLV cells, recovery of CD8-TCRNLV cells from the tissues was poor but was enhanced after transfer of the mixture. Thus, apparently, the CD4 T cells’ helper effect is by facilitating tissue infiltration, as limited longevity of transferred CD8-TCRNLV cells should not yet be an issue in the first 3 days.
By design, the model’s strength is to test immediate effector functions of the transferred T cells, as their longevity in the NSG/HHD recipient is limited by the constitutive absence of IL-2/4/7/9/15/21 signaling and cytokine heterology, except for single doses of IL-2 and IL-7 coadministered with the human T cells. Furthermore, the model of a genetically, and thus enduringly, immunodeficient recipient is highly demanding for completeness of virus eradication at an early stage, as it is established experience that, in the long run, a few infectious units (just 1–5 PFU) of ‘therapy escapee virus’ will eventually cause death in NSG/HHD and related, immunodeficient mouse strains based on tissue pathology from exponential cytopathogenic viral spread over time [71].
Despite these inherent constraints, transfer of human NLV-specific T cells led to prolonged survival, which was modest at very high initial virus burden and did not reach statistical significance (Fig 7A) but was highly significant (P < 0.001 by log-rank test as well as by Gehan-Wilcoxon test) at a moderate, more realistic, initial virus dose (Fig 7B; see Discussion). Specifically, hCD8-NLV cells delayed mortality when compared to the no transfer group (Fig 7B and 7a), and the 1:4 mixture of TCRNLV-transduced CD4 and CD8 T cells delayed mortality when compared to the mock-transduced mixture lacking an NLV-specific TCR (Fig 7B and 7b).
To demonstrate a histopathological correlate of mortality, we performed immunohistochemical (IHC) imaging for livers from the mock-transduction control group and the therapy group (corresponding to Fig 7B and 7b) at the time of onset of mortality (day 20) in the control group (Fig 8). Overview sections, identifying infected liver cells by their expression of intranuclear viral IE1 protein (Fig 8a1 and 8b1; red staining), reveal extended plaque-like lesions from the cytopathogenic effect of virus replication in the control livers, whereas foci of infection were less frequent and less extended in the therapy group, though, as predictable from exponential (log-linear) growth ([36,37] and references therein) tissue infection is likely to close up to the control group with a few days delay explaining delayed onset of mortality in the therapy group. Higher-magnification 2-color IHC images show infected hepatocytes (iHc, red staining of IE1), which are distinctive by cytomorphology, infected endothelial cells (iEC, red staining of IE1 and black staining of CD31 antigen) (Fig 8a2 and 8b2), and infected macrophages (iMΦ, red staining of IE1 and turquoise-green staining of F4/80 antigen) (Fig 8a3 and 8b3). Quantification by cell counting reveals significant reductions in the numbers of infected cells in the therapy group, with no notable cell-type specific preferences (Fig 9A).
Although therapeutic cells equipped with a transgenic virus epitope-specific TCR are not expected to cause an immunopathology, there was residual concern that xenogeneic graft-versus-host interactions, possibly by the endogenous TCRs or by TCR-independent mechanisms, might cause immunopathology and thus obscure the results of the model, in particular the survival data. We have therefore compared general histopathology and caspase 3-dependent apoptosis in liver tissue sections from the control group and the therapy group at the time of onset of death in the control group (day 20 in this specific experiment) by 2-color IHC staining of infected cells (IE1+, red staining) and apoptotic cells (caspase-3+, brown staining) (Fig 10). Notably, histopathology proved to be confined to the foci of infection, as no lesions or necrotic areas were observed in tissue regions not yet reached by the infection. Interestingly, apoptotic cells were predominantly uninfected Hc (IE1-caspase-3+), though localizing to the foci of infection. This suggests an apoptosis induction in trans, but apoptotic iHc (IE1+caspase-3+) can also be found occasionally (IHC image in S6 Fig and quantitation in Fig 9B). The relatively low number of apoptotic infected cells relates to the known fact of cell death-inhibiting viral genes being expressed in CMV-infected cells (for reviews, see [72,73]. Most relevantly, however, the number of apoptotic cells, even that of uninfected apoptotic cells, was definitively not increased but rather tended toward being reduced in the therapy group (Fig 9B), a finding that rules out any relevant immunopathology by the xenogeneic cell transfer in this immunotherapy model. As reported recently, mCMV infection as such does not induce immunopathology mediated by CD8 T cells, except if the host is combined-deficient in NK cells and perforin [74]. Finally, IHC specific for CD8 did no longer reveal presence of transferred CD8 T cells in the liver on day 16 (Fig 9C), reflecting their limited longevity due to the multiple interleukin-signaling deficiency in NSG/HHD mice. In conclusion, histopathology is undoubtedly of viral etiology and is clearly reduced by antiviral human CD8 T-cell immunotherapy.
We herein describe a new mouse model mimicking pre-emptive adoptive T-cell therapy of systemic HCMV infection in immunocompromised patients. The model is based on HLA-A2.1 transgenic NSG/HHD mice that are first infected with chimeric virus mCMV-NLV and are subsequently infused with human T cells specific for the immunodominant HLA-A2.1-restricted NLV epitope of HCMV pp65 (UL83). Despite the species barrier, transferred human T cells formed protective NIF, controlled virus spread, and limited viral pathology in classical organs of CMV disease manifestations without mediating immunopathology, thereby conveying a significant survival benefit to infected mice.
Previous mouse models for HCMV infection utilized transplantation of HCMV-infected human cells, e.g. fibroblasts embedded in agarose plugs [75], glioblastoma cell lines [76], and hepatocytes [23] or, alternatively, human fetal thymus/liver and retinal tissues infected with HCMV [19–21]. Although having contributed important information, implant models, however, are not designed for studying virus dissemination to functional organs and virus spread in these organs that causes histopathology resulting in morbidity and mortality. Therefore, life-saving therapeutic interventions could not be evaluated in such models. In contrast, systemic infection of NSG/HHD mice with the herein described chimeric virus mCMV-NLV provides a model that—in a reasonably good first approximation—resembles relevant aspects of HCMV disease in HSCT recipients, including the infection of vital organs involved in virus-associated morbidity and cause of death. It should be emphasized that the NLV peptide was genetically engineered into the non-essential IE2/m128 protein of mCMV. Thus, its synthesis and presentation occurs according to the regular gene expression kinetics of the viral IE2 carrier protein.
The NLV peptide was here just chosen as a paradigm. The model is meant as a modular system that can easily be adapted to scientific and medical needs. Specifically, by analogous strategy, the NLV peptide can be replaced with any other HLA-presented HCMV epitope of interest or might be combined with additional epitopes integrated at other positions of the mCMV genome [31]. Such chimeric viruses, together with various HLA transgenic NSG mouse strains, promise to allow analysis of the broad repertoire of HLA restricted T-cell responses against HCMV ([77], reviewed in [78,79]).
Importantly, like for HCMV infection of human cells, HLA-A2.1/NLV-antigen presentation was shown here to be susceptible to immune evasion mediated by virally encoded proteins that inhibit the trafficking of peptide-loaded MHC-I molecules to the cell surface [49,80–84]. Although the proteins involved and the precise molecular mechanisms by which they mediate immune evasion differ between individual CMV species, the common biological outcome is inhibition of antigen presentation, which has been shown to reduce but not to preclude a protective effect of adoptively transferred murine CD8 T cells in immunocompromised mice ([85,86], reviewed in [7–9,49,84]). Likewise, contrasting with earlier antigen presentation studies in transfected cells, HCMV ‘immune evasion’ genes gpUS2-11 also fail to completely prevent antigen presentation in infected cells [38,66]. In this context it is important to note that in the here presented model the species origin of the transferred CD8 T cells, secreting murine or human IFN-γ, appeared not to be qualitatively critical for controlling mCMV-NLV infection despite immune evasion mechanisms of mCMV being active that are relieved by murine IFN-γ [47,48]. For the future, the modular model is open for the option to further ‘humanize’ the conditions by replacing immune evasion genes of mCMV with those of HCMV in a next generation of chimeric mCMV-NLV viruses.
A concern affecting the validity of the model has been a potential species barrier for an efficient infiltration of infected murine tissues by human CD8 T cells. It is therefore important to recall that we observed tissue infiltration and epitope-specific formation of protective NIF by NLV-specific human CD8 T cells, reduction of viral load in representative murine target organs of CMV disease, and prolonged survival. The data in the model thus agree with previous work in humans showing that adoptive transfer of NLV-specific CD8 T cells is beneficial in patients undergoing allogeneic HSCT [87–89]. Importantly, a concern regarding a putative immunopathology by transferred xenogenic TCR-transduced T cells could be invalidated in the liver by absence of tissue necrosis and of apoptotic cells in tissue areas outside of the foci of infection.
Although previous studies have highlighted the importance of CD4 T-helper cell support in HSCT patients [87,90–93], antiviral T-cell immunotherapies so far mainly focused on cytolytic CD8 T cells. A helper contribution by CD4 T cells was proposed to be needed primarily for longevity of transferred CD8 T cells mediating enduring protection. Applying the here described new model for testing the requirements of immunotherapy by TCRNLV-transduced human CD8 T cells as a model for an HCMV seronegative HSCT donor, CD4 T cells indeed supported CD8 T cell-mediated antiviral control in mCMV-NLV-infected NSG/HHD mice, while not exerting antiviral effector functions on their own. As an interesting new information, this ‘helper’ effect corresponded to an enhanced early tissue infiltration by the CD8 T cells, indicating that CD4 T cell help is beneficial already at an early stage of immunotherapy and not only at late stages.
In the current version of the new model, the adoptive transfer of a single dose of human T cells resulted in epitope-specific tissue infiltration and tissue persistence of T cells within NIF structures for at least 11 days, and in a reduction of virus spread associated with confinement of the infection within NIF. At a very high dose of intraplantar infection, however, the survival benefit was insignificant, whereas at a moderate initial virus dose the survival benefit, though now statistically highly significant, did not result in cured long-term survivors, which corresponded to absence of tissue-infiltrating progeny of transferred T cells and relapse of virus spread at the time of onset of mortality.
For the interpretation of the relevance of only delayed but not prevented mortality, several aspects in which the model by design differs from the clinical correlate of HSCT-associated HCMV reactivation need to be considered: (i) in the model the virus doses used are much higher than to be expected in a clinical setting where acute primary infection of a seronegative HSCT recipient (R-) is a rare event, while reactivation in a seropositive, latently infected recipient (R+) likely originates from very few cells, if not just from one single cell. mCMV latency models indicated rare and stochastic reactivation events under immunocompromised conditions in vivo [94,95] as well as after tissue explantation ([96,97], reviewed in [41,42]). Indeed, it is clinical routine to start antiviral therapy, not just immunotherapy but also therapy with antiviral drugs, upon first detection of viral DNA by highly-sensitive PCR monitoring when the load of infectious virions is still minimal and mostly undetectable. This strategy is known as pre-emptive therapy, and has recently been reviewed in a clinical article by Seo and Boeckh [98]. Along the same line of argument, in approved HCMV vaccine trials, minimal doses of challenge virus (as low as 10 infectious units) were actually used to clinically evaluate vaccination efficacy (reviewed in [99]), a fact that not many are aware of. (ii) the model, by design, measures immediate antiviral effector functions of the transferred cells, as the NSG/HHD host does not support longevity and expansion of transferred cells due to multiple genetic deficiency in common γ-chain-dependent interleukin signaling, and (iii) in the clinical situation of pre-emptive immunotherapy, and likewise also of pre-emptive antiviral drug therapy, of HSCT-associated HCMV reactivation, the medical demand is only to bridge the critical phase of transient immunodeficiency until reconstitution of intrinsic host immunity in consequence of HSCT takes over. Such a temporary, transferred protection likely requires much lower cell numbers [63]. In contrast, in the model, as it currently stands, the NSG/HHD host is constitutively, and thus enduringly, immunodeficient so that even a single infectious unit of a ‘therapy escapee virus’ eventually causes death by viral tissue pathology. With these arguments in mind, the model provides a rigorous ‘stress test’ for the quality of adoptively transferred antiviral T cells, and the prolongation of host survival observed in this rigorous model predicts a good functionality of the tested human T cells in the less demanding clinical reality of low-dose HCMV reactivation followed by reconstitution of intrinsic antiviral immunity.
We wish to emphasize that the here introduced model should be understood as a basic module of a modular model system that is open to be developed further by ourselves and, hopefully, joined by other investigators. An obvious next step could be to improve the longevity and promote clonal expansion of adoptively transferred T cells by prolonged substitution with IL-2 and IL-7, and to include also IL-15 [100]. Another option for adaptation of the model to clinical needs could be the transfer of defined T-cell subsets with stem cell properties [101]. Finally, and most obviously, NSG/HHD adoptive transfer recipients could be further ‘humanized’ by human CD34+ stem-cell transplantation to reconstitute them with human immune system for the priming of intrinsic antiviral effector and helper cells [102,103] that take over for enduring viral control after the transient protection by adoptive immunotherapy has prevented life-threatening early-onset CMV disease.
In conclusion, we introduce here a promising mouse model of systemic CMV infection that allows the direct analysis of antiviral activity of HCMV-specific T-cell products in terms of infection control in functional organs and survival benefit. The model has best prospects to stimulate future research aimed at optimizing it for clinical needs, and has great potential to expedite the development of improved adoptive cell transfer as well as vaccine strategies against HCMV infection.
C57BL/6 and NOD.Cg-Prkdcscid Il2rgtm1Wjl Tg(HLA-A/H2-D/B2M)1Dvs/SzJ (NSG/HHD) [27] mice were bred and maintained under SPF conditions in the Central Laboratory Animal Facilities at the University Medical Center Mainz and at the University Hospital of Regensburg. NSG/HHD were purchased from Jackson Laboratory (Bar Harbor, ME, USA). Mice were sacrificed by CO2 inhalation or cervical dislocation. Primary mouse embryonic fibroblasts (MEF) from C57BL/6 and NSG/HHD mice were generated by standard methods [104] and maintained in minimal essential medium (MEM) supplemented with 10% fetal calf serum. HLA-A2.1+ primary human foreskin fibroblasts (HFF) were grown in MEM supplemented with 10% fetal calf serum, 2 mM l-glutamine, 50 mg/L gentamicin and 0.5 ng/ml basic fibroblast growth factor (Life Technologies, Darmstadt, Germany) [66]. Human CD4 and CD8 T cells were isolated from buffy coat products of either HLA-A2.1+ CMV-seronegative or—seropositive healthy donors using immunomagnetic bead technology (Miltenyi Biotec, Bergisch Gladbach, Germany). Bacterial artificial chromosome (BAC)-derived virus mCMV-Δm157 [105] as used as non-chimeric reference virus. Virus stocks of mCMV-Δm157, mCMV-WT.BAC [106], and mCMV-Δm04/m06/m152 (mCMV-ΔvRAP) [50,51] were prepared from infected C57BL/6 MEF by sucrose-gradient ultracentrifugation as described [107]. HCMV immune evasion gene (US2,3,6,11) deletion mutant RV-KB6 and the combined US2,3,6,11, and pp65/UL83 deletion mutant RV-KB15 were described previously ([66] and [16], respectively).
Animal research protocols of the University Medical Center Mainz were approved by the ethics committee of the Landesuntersuchungsamt Rheinland-Pfalz, permission numbers 23177-07/G09-1-004 and 23177-07/G11-1-004, according to German Federal Law §8 Abs. 1 TierSchG (animal protection law). Human blood cells were isolated from buffy coat products of healthy donors after written informed consent and approval by the ethics committee of the Landesärztekammer Rheinland-Pfalz and University Hospital of Regensburg (permission number 837.149.10 and 13-101-0240, respectively) and performed according to the Declaration of Helsinki.
For generating mCMV-NLV, 69 nucleotides of HCMV ORF UL83 (n119,567-n119,499; GenBank accession no. X17403) were integrated at nucleotide position n186,511 into ORFm128 of the mCMV genome (GenBank accession no. NC_004065). These nucleotides code for 23 amino acids including the NLV epitope and its natural flanking regions. The mutagenesis was performed essentially as previously described [50]. A linear PCR fragment containing a kanamycin resistance gene flanked by two FRT sites and viral homologies of ORFm128 was generated by PCR with primers m128_NLV_for (5′-acg tcg ggc aga aag ctg ggt tat ctc gac gtg gcg gag aag atc ctg gcc cgc aac ctg gtg ccc atg gtg gct acg gtt cag ggt cag aat ctg aag tac cag gaa ttc agg acg acg acg aca agt aa-3`) and m128_NLV_rev (5′-gga tca cgc cga gaa cct cga ggg gac cgt tgc aca tgg ggt att cct tgc gca gca gga aca ctt aac ggc tga-3′) and plasmid pKD46 [108]. This fragment was inserted into the BAC plasmid pΔm157 [105] by homologous recombination in E. coli. After subsequent FLP-mediated excision of the resistance gene, the correct nucleotide insertion was verified by sequencing (GATC, Freiburg, Germany). The generation of mCMV-NLVAla was performed as described above except using primer m128_NLVAla_for (5′-acg tcg ggc aga aag ctg ggt tat ctc gac gtg gcg gag aag atc ctg gcc cgc aac ctg gtg ccc atg gtg gct acg gca cag ggt cag aat ctg aag tac cag gaa ttc agg acg acg acg aca agt aa-3`) instead of m128_NLV_for. Recombinant CMVs were reconstituted by transfection of purified BAC DNA into C57BL/6 MEF, and high titer virus stocks were purified as described [107].
Flow cytometry was performed on FACS Canto II or FACS LSR II (BD Biosciences, Heidelberg, Germany). Fluorochrome-labeled monoclonal Antibodies (mAb) were anti-mouse H-2Kb (clone AF6-88.5), CD62L-PE-Cy7 (clone MEL-14), CD8-FITC (clone 53–6.7), anti-human CD8-PerCP (clone SK1), CD8-APC (clone RPA-T8), CD4-APC (clone RPA-T4), CD28-FITC (clone CD28.2), CD95-PE (clone DX2), CD45RA-PE (clone HI100), CD45RO-PE (clone UCHL1), CD62L-FITC (clone DREG-56), HLA-A2.1-PE (clone BB7.2) (all BD Biosciences), anti-mouse CD44-PB (clone IM7) (BioLegend, San Diego, CA, USA), and anti-human CCR7-FITC (clone 150503) (R&D Systems, Minneapolis, MN, USA). Intracellular staining of MEF cells was performed with m164/gp36.5 antiserum [51], followed by Alexa Fluor 488-conjugated mAb goat anti-rabbit IgG (Life Technologies, Darmstadt, Germany). Fluorochrome-labeled HLA-A2.1/NLV tetramer was synthesized by Beckman Coulter (Krefeld, Germany). Peptides were synthesized by PSL (Heidelberg, Germany). Analyses were performed with software FlowJo 7.6.5 (Tree Star, Ashland, OR, USA).
Human NLV-specific CD8 T-cell lines were expanded from purified CD8 T cells of HLA-A2.1+ HCMV-seropositive healthy donors by stimulation with pp65 (UL83)495-503 NLV peptide (10−6 M)-loaded and irradiated (35 Gy) autologous peripheral blood mononuclear cells (PBMC) over 2 weeks at the T cell-to-PBMC ratio of 1:1 in AIM-V medium (Life Technologies). AIM-V was supplemented with 10% human serum, recombinant human interleukin (rhIL)-2 (50 IU/mL; Proleukin, San Diego, CA, USA), rhIL-7, and rhIL-15 (each 5 ng/mL; R&D Systems) (AIM-Vcytokine). A murine NLV-specific CD8 T-cell line was generated and weekly restimulated as previously described [38]. MEF were infected with mCMV under conditions of centrifugal enhancement of infectivity [104]. HFF were infected with HCMV-RVKB6 and—RVKB15 for 24h at an MOI of 5. Standard 4h [51Cr]-release and 20h interferon (IFN)-γ ELISpot-assays were performed in duplicates as reported [109,110]. Dose-escalating equilibrium tetramer binding data were plotted in Scatchard analysis of mean fluorescence intensity (MFI)/concentration of tetramer against MFI. The dissociation constant KD equals -1/slope [44,45].
TCRNLV is a codon-optimized and affinity fine-tuned variant of the previously described TCR AV18/BV13 [16,111]. A short self-cleaving F2A sequence [112] was used to link the N-terminus of TCRα chain to the C-terminus of TCRβ chain by ligation PCR technology [113]. The antisense oligonucleotide sequence was 5’-atg gct atg gtg aag cgg aag gac ttc gtg aaa caa acg ttg aat ttt gac ctt ctc aag ttg gcc gga gac gtg gag tcc aac ccc ggg cct atg gag aag aac ccc ctg gcc gcc ccc-3’. The coding sequence of the TCRβ-F2A-TCRα gene construct was inserted into the multiple-cloning-site of the drug-selectable retroviral vector pMX (BioCat, Heidelberg, Germany). Purified CD4 and CD8 T cells from HCMV-seronegative donors were pre-stimulated with anti-CD3/CD28 Dynabeads (Life Technologies) in AIM-V medium (supplemented with 100 IU/mL rhIL-2) and then retrovirally transduced with TCRNLV as previously described [113]. Following drug-selection on TCRNLV+ cells, T cells were expanded in vitro for a period of 7-12d, using anti-CD3/CD28 beads (3–5 μL/106 cells) in AIM-Vcytokine, prior to their adoptive transfer.
Eight- to 10-week-old NSG/HHD mice were sublethally (2 Gy) γ-irradiated for cell homing conditioning, followed by intraplantar infection with 1x105 PFU of the indicated viruses. Subsequently, a single dose of up to 1x107 human or murine T cells was injected intravenously. Human T cells were co-injected with rhIL-2 (1000 IU/mouse) and FcIL-7 (20 μg/mouse; Merck, Darmstadt, Germany). At indicated times post-infection, virus replication in spleen, lungs, and liver was assessed by quantitation of infectivity from the respective organ homogenates in a virus plaque assay (PFU assay) [107]. Infected cells and T cells in liver tissue sections were visualized simultaneously in their microanatomical context, specifically in nodular inflammatory foci (NIF), and quantitated by two-color immunohistochemistry (2C-IHC) specific for the intranuclear viral IE1 protein and a conserved CD3ε epitope [107]. For testing maintenance of transferred human T cells, flow cytometric analysis was performed on splenocytes retrieved by standard methods and on non-parenchymal liver cells isolated as described [97].
To detect any type of mCMV-infected cells, one-color IHC specific for the intranuclear viral protein IE1 was performed on liver tissue sections as described in detail elsewhere [107]. For differentiating infected cells by cell type, 2C-IHCs were performed by combining IE1-specific labeling with the labeling of cell type-specific markers, such as CD31 for the identification of endothelial cells (EC) or F4/80 (Ly71) antigen for the identification of macrophages (MΦ), essentially as described recently for a 3C-IHC analysis that has included the same markers [36]. 2C-IHC specific for IE1 and active caspase 3 was described previously [114] and applied with modifications. In brief, intranuclear IE1 protein was labeled with monoclonal Ab CROMA 101 and stained in red with alkaline phosphatase-1-conjugated polyclonal goat anti-mouse IgG (AbD Serotec, Puchheim, Germany) and fuchsin substrate-chromogen kit 2 (Dako-Cytomation, Hamburg, Germany). Active caspase 3 was detected with rabbit anti-active caspase 3 IgG (Biovision, Milpitas, CA, USA) and stained in brown by using the ImmPRESS reagent: anti-rabbit-Ig-peroxidase (Vector Laboratories, Peterborough, UK) with 3,3’-diaminobenzidine (DAB) as substrate. A light blue counterstaining was achieved with hematoxylin.
Statistical significance of differences between two independent data sets was evaluated by two-sided unpaired t test with Welch’s correction of unequal variances. Comparison of survival curves was performed with log-rank test and Gehan-Wilcoxon test. Differences were considered statistically significant for P values of <0.05. Viral doubling times (vDT = log2/a) and the corresponding 95% confidence intervals were calculated by linear regression analysis from the slopes a of log-linear growth curves [36]. All analyses were performed with Graphpad Prism 6.04 (GraphPad Software, San Diego, CA, USA).
|
10.1371/journal.pgen.1001283 | A Meta-Analysis of Genome-Wide Association Scans Identifies IL18RAP, PTPN2, TAGAP, and PUS10 As Shared Risk Loci for Crohn's Disease and Celiac Disease | Crohn's disease (CD) and celiac disease (CelD) are chronic intestinal inflammatory diseases, involving genetic and environmental factors in their pathogenesis. The two diseases can co-occur within families, and studies suggest that CelD patients have a higher risk to develop CD than the general population. These observations suggest that CD and CelD may share common genetic risk loci. Two such shared loci, IL18RAP and PTPN2, have already been identified independently in these two diseases. The aim of our study was to explicitly identify shared risk loci for these diseases by combining results from genome-wide association study (GWAS) datasets of CD and CelD. Specifically, GWAS results from CelD (768 cases, 1,422 controls) and CD (3,230 cases, 4,829 controls) were combined in a meta-analysis. Nine independent regions had nominal association p-value <1.0×10−5 in this meta-analysis and showed evidence of association to the individual diseases in the original scans (p-value <1×10−2 in CelD and <1×10−3 in CD). These include the two previously reported shared loci, IL18RAP and PTPN2, with p-values of 3.37×10−8 and 6.39×10−9, respectively, in the meta-analysis. The other seven had not been reported as shared loci and thus were tested in additional CelD (3,149 cases and 4,714 controls) and CD (1,835 cases and 1,669 controls) cohorts. Two of these loci, TAGAP and PUS10, showed significant evidence of replication (Bonferroni corrected p-values <0.0071) in the combined CelD and CD replication cohorts and were firmly established as shared risk loci of genome-wide significance, with overall combined p-values of 1.55×10−10 and 1.38×10−11 respectively. Through a meta-analysis of GWAS data from CD and CelD, we have identified four shared risk loci: PTPN2, IL18RAP, TAGAP, and PUS10. The combined analysis of the two datasets provided the power, lacking in the individual GWAS for single diseases, to detect shared loci with a relatively small effect.
| Celiac disease and Crohn's disease are both chronic inflammatory diseases of the digestive tract. Both of these diseases are complex genetic traits with multiple genetic and non-genetic risk factors. Recent genome-wide association (GWA) studies have identified some of the genetic risk factors for these diseases. Interestingly, in addition to some similarities in phenotype, these studies have shown that CelD and CD share some genetic risk factors. Specifically, by comparing the results of independent GWA studies of CD and CelD, two genetic risk loci were found in common: the PTPN2 locus and the IL18RAP locus. Therefore, in order to directly test for additional shared genetic risk factors, we combined the GWA results from two large studies of CelD and CD, essentially creating a combined phenotype with anyone with CD or CelD being coded as affected. Association results were then replicated in additional cohorts of CelD and CD. It is expected that shared risk loci should show association in this analysis, whereas the signal of risk loci specific to either of the two diseases should be diluted. With this method of meta-analysis, we identified next to PTPN2 and IL18 RAP two loci harbouring TAGAP and PUS10 as shared risk loci for Crohn's disease and celiac disease at genome-wide significance.
| Crohn's disease (CD) and celiac disease (CelD) are both chronic intestinal inflammatory diseases. In CD inflammation can occur throughout the gastrointestinal tract but most commonly affects the ileal part of the small intestine. While the causative antigen(s) for this inflammation is unknown, it is thought that the disease arises as a reaction to the normal commensal flora of the bowel in a genetically susceptible individual [1], [2]. In CelD inflammation is limited to the small intestine. CelD is caused by a reaction to gluten, a dietary peptide present in wheat, barley and rye [3], [4]. In both CelD and CD contact between antigens and antigen-presenting cells (APCs) seems to be facilitated by an initial increase in intestinal permeability [5]. In both diseases the subsequent inflammatory response follows a T helper 1 pattern characterized by tumor necrosis factor beta (TNF-β) and interferon gamma (IFN-γ) production and a T helper 17 response marked by the production of interleukin 17 [5].
Although uncommon, it has been observed that CelD and CD can co-occur within families or even within individual patients; there appears to be a greater prevalence of CD among CelD patients than in the general population, although the relatively low prevalence of CD makes it difficult to establish this effect [6]. It is now well accepted that the risk for CD and CelD is partly determined by genetic factors, and recently many genetic risk factors for CelD and CD have been identified. Two genetic risk loci were previously shown to be shared between CelD and CD: a locus on 18p11 containing the PTPN2 (protein tyrosine phosphatase, non-receptor type 2) gene and a locus on 2q12 containing the IL18RAP (interleukin 18 receptor accessory protein) gene [7]–[13]. While these observations confirm the existence of shared risk loci for CD and CelD, additional such shared risk loci are likely to exist.
There are two possible approaches for identifying shared risk loci. One approach is to test known risk loci from one disease in patient-control cohorts from the other disease. This approach has already been successfully applied in a cross study between CelD and type 1 diabetes (T1D), where four shared risk loci were identified some of which were previously unknown to be associated to CelD [11]. However, this approach relies on previously identified risk alleles, indicating that there probably are many more unknown common risk loci for T1D and CelD. In addition, some of the shared loci will not have a large enough effect in the individual diseases to have been identified by previous genetic studies. A second approach that tackles this problem is to analyze genetic data from two similar diseases as a single unified disease phenotype against healthy controls. Such an analysis would be expected to dilute disease-specific genetic associations, but increase the power for finding shared genetic risk loci of small effect in the individual diseases. The availability of genome-wide association studies (GWAS) performed in both CelD and CD provides large case-control genotyping datasets that enabled us to perform a cross-disease genome-wide meta-analysis in the aim of identifying novel shared risk loci.
To identify novel shared risk loci between CelD and CD, we performed a meta-analysis of two recently published GWAS: a large meta-analysis of three CD GWAS by the International IBD Genetics Consortium and a CelD GWAS in a British population. To confirm identified risk loci, we used a combination of Italian and Dutch CD cohorts and of British, Italian and Dutch CelD cohorts.
We have performed a meta-analysis of 471,504 SNPs from genome-wide datasets of CD (3230 cases, 4829 controls) and CelD (768 cases, 1422 controls) in order to identify shared risk loci between these 2 diseases.
A quantile-quantile (Q-Q) plot of the association p-values for single-SNP Z scores from the meta-analysis was performed (Figure S1) and shows an excess of significant associations above what would be expected by chance. We observe a low inflation factor of 1.08, which is expected given the inflation observed in each of the original studies: 1.05 for CelD and 1.16 for CD. A Manhattan plot of the current study (Figure S2) highlights many strong association signals, several of which corresponding to previously reported CD and CelD loci; however, most of these show strong association in only one of the 2 diseases and have thus not been followed up due to the design of the current study. In addition, given the design of the original CelD GWAS, which included only individuals that were positive for the risk-associated allele HLA-DQ2, association to the major histocompatibility complex (MHC) region in CelD was of no relevance since it was artificially inflated. Therefore the MHC region (Chr6:22700000..35000000 from the NCBI B36 genome build) was removed from our analysis.
The meta-analysis of the CD and CelD datasets identified 25 SNPs, from 10 independent regions, that met our criteria for association (association with CelD at p-value <1×10−2 and with CD at p-value <1×10−3 in the original scans, as well as a nominal association p-value more significant than 1.0×10−5 in the meta-analysis) (Table 1 and Table S2). This is more than expected by chance, as we would expect no more than 3 independent regions to meet our criteria, which encouraged us to explore these specific loci further.
The strongest association signal identified in our scan was to the well accepted CD associated risk locus CARD15 (p-values of 3.42×10−32, 3.77×10−3 and 1.30×10−21 in the CD, CelD and scan datasets respectively). Given the strength and the width of the association signal peak at this locus in the CD dataset, our chances of detecting a false positive shared signal at this locus in the scan were artificially increased. Because of this, the CARD15 locus was not moved forward to replication. An evaluation of the association signal in the in silico CelD GWAS replication datasets confirmed that this locus did not show replication in CelD. Several of the SNPs meeting our criteria for association mapped to the known shared risk loci IL18RAP and PTPN2. Identifying these shared risk loci in the initial phase of our analysis provides proof of the effectiveness of our method. Interestingly, these two loci either reach or are very near genome-wide significance in the current meta-analysis (p-value of 8.37×10−8 for IL18RAP and of 6.39×10−9 for PTPN2), validating their previously identified role in both CD and CelD. The remaining 12 SNPs were located in seven independent regions and for each of these loci we selected the most associated SNP for testing for evidence of replication.
All SNPs selected for follow-up were genotyped in additional replication cohorts of CelD patients (n = 3149) and healthy controls (n = 4714) and of CD patients (n = 1941) and healthy controls (n = 1669). Given that these putative shared risk loci were selected through the combined analysis of our CD and CelD scan cohorts, a positive threshold for replication was therefore set at a corrected p-value of 0.0071 (Bonferroni corrected p-value of 0.05 for 7 independent tests) in the combined analysis of CD and CelD replication cohorts.
Only two of the 7 loci tested, PUS10 (pseudouridylate synthase 10; RefSeq NM_144709.2) and TAGAP (T-cell activation GTPase activating protein; RefSeq NM_054114), showed significant replication (p-values of 6.03×10−7 and 3.03×10−6 respectively) with matching direction of association between the scan and replication datasets, as well as replication p-values more significant than 0.05 in both CD and CelD replication cohorts independently (Table 1). While neither PUS10 nor TAGAP were identified as loci of genome-wide significance in the combined dataset from each disease (p-values = 1.34×10−6 and 7.00×10−7 in CelD and p-values = 6.16×10−8 and 2.13×10−6 in CD respectively), both reach genome-wide significance in a combined analysis of CD and CelD cohorts (p-values of 1.38×10−11 and 1.55×10−10 respectively). Based on the results calculated from the replication datasets, we also observe that the effects at these 2 loci are similar in size and direction for both CD and CelD (Table 1).
By performing a meta-analysis of GWAS data from CD and CelD as a single disease phenotype, we have identified four risk loci shared by these 2 diseases: PTPN2, IL18RAP, TAGAP and PUS10. This meta-analysis approach provided the power, lacking in individual disease-specific GWAS datasets, to identify shared risk loci with small effects in each single disease. This approach is a powerful and versatile way of identifying shared risk loci. In fact, two of the shared loci described here, TAGAP and PUS10, would not have reached genome-wide significance without the power gained from the combined samples (scan and replication) of these 2 diseases. As the GWAS for the individual diseases increase in power, we can expect the power of the current approach to also increase enabling us to identify further shared loci.
The TAGAP locus identified in the current study as a shared risk factor for CD and CelD is located on chromosome 6q25.3, within a 200-kb block of linkage disequilibrium (LD). This TAGAP locus was previously identified as a CelD risk locus [9] but not found in previous studies of CD. TAGAP is the best candidate of four genes in this region of strong LD [9]. TAGAP is a member of the Rho-GTPase protein family, which release GTP from GTP-bound Rho, thereby acting as a molecular switch. The gene is expressed in activated T cells and appears to be important for modulating cytoskeletal changes [14]. Little is known about the exact role of TAGAP in immune function, but it has been found to be co-regulated with IL2 and is expected to play a role in T-cell activation [14].
The current study also identifies a shared risk locus between CD and CelD in the PUS10 gene region, a locus previously described as a risk locus for CD [15]. This locus was recently identified as a risk locus in both ulcerative colitis (UC) and CelD, indicating that this locus may be a shared risk locus for these three diseases [7], [16]. This latter finding further validates the approach use in this study to identify risk factors that are shared across diseases. Interestingly, the UC study identified three independent signals in this region which seem to be shared differently across these three diseases: one signal seems to be shared only between CD and UC, a second only between CelD and UC, while the third, identified in the current study, seems shared between all three diseases. Further analysis of this locus will be necessary in order to clarify the role of these different alleles in disease risk.
In this study we aimed to find shared genetic risk factors for CelD and CD by meta-analysis of GWAS data of both diseases, defining a single phenotype for these analyses. Using readily available data, we were able to reliably establish four shared loci: PTPN2, IL18RAP, TAGAP and PUS10. For many diseases with overlapping phenotypic characteristics, GWAS data is available and joint analysis of GWAS datasets of these related diseases could lead to the identification of many new shared susceptibility loci.
For the CD aspect of the meta-analysis, we used the previously published data (available at http://www.broadinstitute.org/~jcbarret/ibd-meta/) from the International Inflammatory Bowel Disease Genetics Consortium (IIBDGC) meta-analysis of 3230 CD cases and 4829 healthy controls taken from three independent CD GWAS (Table S1) [15]. A more in depth description of these cohorts and their origin can be obtained from the original publication of this meta-analysis.
Two independent cohorts were used for the CD replication phase (Table S1). The first consisted of 1217 Dutch CD cases from three Dutch university medical centers: the Academic Medical Centre Amsterdam (n = 661), the University Medical Centre Groningen (n = 322) and the University Medical Centre Leiden (n = 234); the 804 Dutch controls used for this replication cohort were obtained from cohorts of healthy partners of IBD patients from the UMC Leiden (n = 151) and the UMC Groningen (n = 120) and from healthy blood donors recruited through the Sanquin Blood bank by the UMC Utrecht and the VUMC Amsterdam (n = 533) [17], [18]. The second replication cohort consisted of an Italian IBD case – control cohort (724 CD patients and 892 controls) collected at the S. Giovanni Rotondo “CSS” (SGRC) Hospital in Italy. This cohort has previously been used and characterized in several association reports [19], [20].
All patients and controls were of European Caucasian descent. The diagnosis of CD required objective evidence of inflammation from radiologic, endoscopic, and/or histopathologic evaluation. All affected subjects fulfilled clinical criteria for CD. Recruitment of study subjects was approved by local and national institutional review boards, and informed consent was obtained from all participants.
For the celiac disease aspect of this meta-analysis, data from a previously published genome-wide scan (768 British cases, 1422 British controls) for CelD was used (Table S1). A more in depth description of this cohorts and its origin can be obtained from the original publications [9], [12].
For the replication phase in CelD we used the genotyping results from a second celiac GWAS in three independent CelD cohorts (Table S1). From this study we received data from 3149 cases and 4714 controls from three European populations (UK, the Netherlands and Italy) genotyped on Custom Illumina Human 670-Quad, Hap550 and 1.2 M slides [13]. UK CelD cases were recruited from hospital outpatient clinics (n = 434) and directly through Coeliac UK advertisement (n = 1415) [9]; UK controls were recruited from the 1958 birth cohort and UK National Blood Service for the Welcome Trust Case Control Consortium (WTCCC) (n = 3786). Dutch CelD cases were collected by the UMC Utrecht, Leiden UMC and VUMC Amsterdam from outpatient clinics (n = 803); Dutch controls were recruited through the Sanquin Blood bank by the UMC Utrecht and the VUMC Amsterdam (n = 385). Italian CelD cases (n = 497) and controls (n = 543) were collected by a CelD referral centre (Centro per la prevenzione e diagnosi della malattia celiaca, Fondazione IRCCS Ospedale Maggiore Policlinico) in northern Italy.
All affected individuals were unrelated and were diagnosed according to the revised ESPGAN criteria (1990). The cohorts encompassed individuals that showed a Marsh II or Marsh III lesion in the initial diagnostic small-bowel biopsy specimens, or presented with dermatitis herpetiformis and were HLA-DQ2 positive. Recruitment of study subjects was approved by local and national institutional review boards, and informed consent was obtained from all participants. Some controls were shared between the WTCCC component of the CD meta-analysis and the two UK CelD cohorts (one used in the original scan and one used as replication). This was taken into account as explained in the meta-analysis description.
Imputation of the CelD datasets used in the initial and the replication phases of this study were performed with BEAGLE using HapMap phase II and HapMap phase III as reference datasets [21]. A minimum quality score for statistical certainty of the imputation of 0.98 was adhered to.
Imputation of the CD dataset for the initial CD-CelD meta-analysis was performed with the programs MACH and IMPUTE, using HapMap phase II as a reference dataset, as previously described [15], [22], [23].
For the original CD dataset, association tests were described previously [15]. Briefly, results for each SNP from three independent GWAS were summarized as Z-scores and combined in a weighted fashion into a single test statistic; imputation uncertainty was taken into account into Z-score and weight calculation using empirical variance calculated from allele dosage. For the original CelD GWAS scan, best guess imputed genotype frequency data was obtained, and association P-values were calculated using chi-square tests (1 df) of SNP allele counts.
The initial meta-analysis was performed using the statistical program R (http://www.r-project.org/). For both the CD and the CelD dataset the p-values signifying the evidence for association were converted to directional Z-scores, and an overall Z-score and two-tailed p-value for the average of the individuals was subsequently calculated. Given the fact that some controls were shared between one component of the CD meta-analysis and the CelD scan, we expect a correlation of 0.187 between CD and CelD Z-scores. We took this correlation into account in the variance term of the overall Z-score [24].
Unweighed Z-scores were used when combining the data from CD and CelD in the initial meta-analysis, since the CD cohort was substantially larger than the CelD cohort and weighing would lead to an overrepresentation of the CD signal in the meta-analysis.
A locus was selected for replication when SNPs met the following criteria: a p-value for the locus of <1×10−5 in the meta-analysis, in combination with a p-value of <1×10−2 in the CelD dataset and a p-value of <1×10−3 in the CD dataset. Different thresholds for CD and CelD for inclusion in the replication phase were used in order to reflect the difference in power between the scans for the two phenotypes. For each of the loci that met these criteria, the most strongly associated SNP was analyzed in the CD and CelD replication cohorts.
In order to evaluate the expected number of SNPs that would pass our thresholds (p-value <1×10−2, <1×10−3 and <1×10−5 for CelD, CD and meta-analysis, respectively) by chance, we first evaluated the probability for a particular SNP to reach those thresholds under the null hypothesis of no association and the expected correlation between the two datasets. This probability can be evaluated from the distribution of two correlated normal variables (correlation of 0.186), combined as described for the meta-analysis. We evaluated this probability to be approximately 6.0×10−6. If the 468,378 SNPs tested in the scan were independent, we would then expect less than 3 (468,378*6.0×10−6) independent SNPs to be selected by chance. Under a binomial model, we evaluated the probability that 9 or more independent SNPs passes our thresholds to be lower than 0.0025. Those are obviously upper bounds, as we know correlation exists among the SNPs tested.
For replication in CD, SNPs selected for testing were designed into multiplex assays, and genotyped using primer extension chemistry and mass spectrometric analysis (iPlex assay, Sequenom, San Diego, California, USA) on the Sequenom MassArray. This was performed at the Laboratory for Genetics and Genomic Medicine of Inflammation (www.inflammgen.org) of the Université de Montreal. Quality control was performed, excluding samples showing >10% missing data, as well as SNPs with >10% missing data or significantly out of Hardy-Weinberg equilibrium (p-value <0.001). The overall genotyping call rate in the CD replication dataset following quality control analyses was >99%. The CD replication datasets from the two groups (Dutch and Italian) were combined and analyzed using a weighted and directional Z-score approach.
For replication in CelD, genotype frequencies and association data for five replication SNPs were obtained from genome-wide genotyping datasets on Illumina Human 670Quad or 610Quad Genotyping BeadChips (Illumina, Inc., San Diego, CA,). Each GWAS dataset was analyzed using PLINK 1.05 and association p-values were calculated using chi-square tests (1 df) of SNP allele counts [25]. Two of the replication SNPs were not included on the Illumina Human 670Quad or 610Quad Genotyping BeadChips. For these SNPs best guess genotype frequency data was obtained by imputation as described above, and association p-values were calculated using chi-square tests (1 df) of SNP allele counts The CelD replication datasets from the three groups (UK, the Netherlands and Italy) were combined and analyzed using a weighted and directional Z-score approach.
Since selection of specific SNPs for replication was based on their association p-values in the combined CD-CelD dataset, a significant threshold for replication was set at a Bonferroni corrected p-value of 0.05 for 7 independent tests (p-value more significant than 0.0071) in the combined CD-CelD replication dataset. As for the initial meta-analysis, the data from the replication in the CD and CelD cohorts were combined through an unweighed Z-score approach. In addition, for a SNP to be replicated, both effect and direction of association trend needed to match between scan and replication within each disease.
For each SNP showing positive replication, an overall disease-specific association p-value, combining the scan and replication data, was also calculated using a weighted meta-analysis approach. Finally, an overall CD-Celiac meta-analysis of the scan and replication phases of this study was obtained, by combining these within-CD and within-Celiac datasets in an unweighed meta-analysis. Given the fact that some controls were shared between the within-CD and the within-CelD, we expect a correlation of 0.149 between CD and CelD Z-scores. As for the initial scan, we took this correlation into account in the variance term of the overall Z-score as per Lin and colleagues [24].
|
10.1371/journal.ppat.1007201 | Invasion of midgut epithelial cells by a persistently transmitted virus is mediated by sugar transporter 6 in its insect vector | Insect transmission is obligatory for persistently transmitted viruses because the vector insect is the only means of virus spread in nature. The insect midgut is the first major barrier limiting virus acquisition, but the mechanisms by which viruses are able to cross the cell membrane and then infect the midgut epithelial cells of the insect have not been elucidated completely. Here, we found that the outer capsid or nucleocapsid protein (NP) of three viruses can interact and colocalize with sugar transporter 6 that is highly expressed in the midgut of Laodelphax striatellus (LsST6). In contrast, LsST6 did not interact with the NP of rice grassy stunt virus, which cannot be transmitted by the same planthopper. LsST6 not only altered the cellular location of viral proteins and then colocalized with them in the cell membrane, but also mediated the entry of rice stripe virus (RSV) particles into Spodoptera frugiperda 9 (Sf9) cells that expressed the heterologous gene LsST6. We further showed that RSV particles initially bound to the cell membrane of midgut epithelial cells where it colocalized with LsST6, and then invaded the cytoplasm. When LsST6 expression was knocked down, viral titre, acquisition percentage and transmission efficiency of the treated insect decreased significantly, but virus replication was not affected. This work thus uncovered a strategy by which LsST6 mediates viral entry into midgut epithelial cells and leads to successful transmission by the insect vector.
| Sap/blood-feeding arthropods are major vectors of viruses that infect plants and vertebrates. Studies on the insect midgut, the first barrier for virus transmission, and its interactions with viruses and parasites are fundamental to understanding the transmission mechanism in vector insects and the epidemics caused by the vectored pathogen. Some putative receptors in arthropods have been discovered by in vitro protein interactions, but in vivo evidence is still lacking. Here, we found that the specific interaction between viral nucleocapsid protein and vector sugar transporter 6 of Laodelphax striatellus (LsST6) determines whether the virus can invade midgut epithelial cells or not. These results provide direct evidence that LsST6 is an essential and key factor in crossing the midgut infection barrier for viruses, especially for RSV. This vector protein may be a promising target for blocking transmission of diverse plant viruses. Our discovery has important implications for better understanding the interaction among host–virus–insect vector and disease epidemics caused by plant and animal arboviruses.
| Many viruses persistently transmitted by arthropods cause serious diseases in plants, animals and humans. More than 76% of plant viruses and 40% of mammalian viruses are transmitted to the hosts by specific arthropods, mainly planthoppers, aphids, mosquitoes, and ticks [1, 2]. Frequent epidemics of viral diseases in rice, wheat and vegetables are largely attributed to high populations and viral transmission efficiency of the insect vectors [3–6]. Similarly, viruses that cause diseases in humans and animals such as dengue fever, Zika fever and Japanese encephalitis, are vectored by different species of Aedes mosquitoes and are endemic in many areas of the developing world [7–10]. Understanding the virus–insect vector interaction and transmission mechanisms will provide important information on the epidemics of the diseases caused by plant and animal arboviruses and lead to the development of better control strategies.
Plant viruses transmitted in a persistent propagative manner and animal arboviruses follow a similar circulative route within their insect vectors. After they are acquired from plant sap or blood ingested by the insect, the virions must first cross the cell membrane of the midgut epithelial cells where the viral particles multiply [11]. They must then leave the midgut to disseminate to other tissues including the salivary glands, from where they can be transmitted to new hosts [12]. During the circulative process, arboviruses must overcome multiple barriers, including the infection and dissemination barriers of the midgut, salivary gland escape barrier, and transovarial barrier [13, 14].
Previous studies showed that Aedes aegypti cannot be infected by eastern equine encephalomyelitis virus after ingesting viruliferous blood; however, it can transmit this virus after a virus suspension is directly injected into the insect’s abdomen [15]. Many plant viruses can also be transmitted by an insect that is not a natural host after the virus is injected into the hemocoel of the insect [16, 17], thus bypassing the midgut infection barrier, the first major barrier that viruses encounter and an important factor limiting virus transmission [1, 18, 19]. To overcome the midgut barrier, viruses have evolved different strategies. The entry of rice dwarf virus into cultured cells of its vector insect and of tomato yellow leaf curl virus midgut in its vector Bemisia tabaci, is mediated by clathrin-dependent endocytosis [20, 21], whereas southern rice black-streaked dwarf virus (SRBSDV) induces the formation of tubules as a vehicle for viral spread in infected epithelial cells of Sogatella furcifera [22].
The small brown planthopper, Laodelphax striatellus (Hemiptera: Delphacidae), is an important vector because it transmits numerous viruses that cause serious diseases of staple crops such as rice stripe virus (RSV), rice black-streaked dwarf virus, maize rough dwarf virus, northern cereal mosaic virus and barley yellow striate mosaic virus [4, 6, 23]. These plant viruses infect and replicate in L. striatellus and are retained by the vector insect throughout their life, as are vertebrate-infecting arboviruses [24–26]. In most experiments when feeding on RSV-infected plants, less than 30% of the insects acquire the virus [27–29]. A high affinity line of L. striatellus attained an acquisition level of about 50–60% after 4 days of acquisition feeding on RSV-infected rice plants, but four other lines reached a level of less than 10% after 8–11 days of acquisition feeding [30]. However, once acquired, the virus will replicate and be transmitted by vector insects at a moderate to very high rate [24, 31]. Although various biotic and abiotic factors affect virus acquisition by the vector insect, the epithelium, intercellular junctions, and basal lamina of the midgut present further barriers to viral entry and dissemination [13, 32].
The genome of RSV consists of four single-stranded RNAs (RNA1-4), which can encode at least seven proteins including the major nucleocapsid protein (NP) encoded by the ORF at the 5′ half of the viral complementary RNA3 [33]. NP is considered the key viral component for specifically interacting with the vector components and may play an important role in persistent transmission process. In previous studies, 66 proteins (including LsST6) were identified as being able to interact with the NP of RSV. Among these proteins, we chose several proteins according to molecular function and biological pathway to further investigate their function in virus transmission. CPR1 was demonstrated to stabilize the viral particles in the hemolymph [34], while vitellogenin, the precursor of a yolk protein in the insect, mediates virus entry into the ovary [35]. However, the proteins involved in the ability of RSV to overcome the midgut infection barrier were not identified. Because the sugar transporter Glut1 of human acts as a receptor for human T cell leukemia virus (HTLV) infection [36], and because a membrane protein named sugar transporter 6 of L. striatellus (LsST6) is highly expressed in the midgut, we selected LsST6 for further study. Our results showed it is an essential and key factor for RSV to cross the midgut infection barrier in vector insects.
The full-length LsST6 (GenBank accession: MG589412), amplified from total RNA of L. striatellus using RT-PCR and 5’RACE, contained a 1470-bp open reading frame (ORF) encoding a predicted 489-amino-acid (aa) protein, which had 85.6% identity with sugar transporter NlST6 in the brown planthopper (Nilaparvata lugens) (S1 Fig). LsST6 belongs to the major facilitator superfamily (MFS) of membrane transport proteins because it has two symmetrical six-TMS (transmembrane spanner) units within a single polypeptide chain and a GRK domain conserved between TM2 and TM3 (S2 Fig).
A yeast two-hybrid (Y2H) assay based on split ubiquitin was used to verify whether LsST6 interacts in vivo with the NP of RSV and rice grassy stunt virus (RGSV) or the p10 of RBSDV and SRBSDV. In nature, L. striatellus can acquire RSV, RBSDV and SRBSDV, but not RGSV. The Y2H results showed that the yeast cells cotransformed with LsST6 and RSV NP, RBSDV p10 or SRBSDV p10, respectively, grew on the selective plates, but those with RGSV NP did not (Fig 1A). A similar result was obtained in a β-galactosidase activity assay (Fig 1A). Further, we exploited the cells of Spodoptera frugiperda 9 (Sf9) to coexpress LsST6 and the respective viral proteins for in vitro coimmunoprecipitation (Co-IP). LsST6 was labeled with a His tag and viral proteins with Myc tags at their C termini. The result showed that the anti-Myc antibody coimmunoprecipitated LsST6 that was coexpressed with RSV NP, RBSDV p10 or SRBSDV p10 in Sf9 cells, but not with RGSV NP (Fig 1B). All these results demonstrated that LsST6 only interacted with the proteins encoded by viruses that can be acquired by L. striatellus.
Sf9 cells were transfected using a recombinant baculovirus that produces LsST6 or different viral proteins. Laser scanning confocal microscopy (LSCM) revealed that LsST6 (green) was mainly localized in the cell membrane, whereas viral proteins (red) localized in the cytoplasm of Sf9 cells (Fig 2A). Interestingly, RSV NP and p10 of RBSDV and SRBSDV moved from the cytoplasm to the cell membrane and colocalized with LsST6, while RGSV NP remained in the cytoplasm, separated from LsST6 when Sf9 cells coexpressed the respective viral protein and LsST6 (Fig 2B). The result suggested that the expression of LsST6 only changed the position of each interacting viral protein in Sf9 cells, and then colocalized with the protein on the cell membrane of Sf9 cells.
The salivary gland, gut, hemolymph and ovary were excised separately from adult insects, then total RNA was extracted from individual tissues to quantify LsST6 by RT-qPCR. The results showed that LsST6 had the highest expression in gut tissues, followed by the hemolymph, salivary glands and ovary (Fig 3A). The alimentary canal of the planthopper, which mainly comprises the esophagus (es), anterior diverticulum (ad), midgut (mg), hindgut (hg) and malpighian tubules (mt) (S3A Fig), was excised, and then incubated with anti-LsST6 antibody labeled with Dylight 488 (green). The confocal image showed that LsST6 was located in the cell membrane and cytoplasm of the midgut epithelial cells (Fig 3B).
The distribution of RSV particles in the alimentary canal over time was also visualized by LSCM using anti-RSV antibody labeled with Dylight 549 (red). At 2 d after a 2-d acquisition access period (AAP), a few RSV particles were observed in several epithelial cells of the midgut. By 4 d after the AAP, RSV particles had replicated and spread to the neighboring epithelial cells, and by 8 d after the AAP, RSV particles were observed in the entire alimentary canal (S3B Fig). Thus, the original RSV infection site was the epithelium of the midgut. Interestingly, at 2 d after the AAP, LsST6 had colocalized with RSV in the cell membrane of epithelial cells (Fig 3C). At 4 d and 8 d after the AAP, viral particles were still colocalized with LsST6 in the cell membrane of the epithelial cells, and some particles had already invaded the cytoplasm, but the fluorescent signals indicative of the viral titre were stronger at 8 d (Fig 3D and 3E). When we used immunoelectron microscopy to examine the virus-infected midgut epithelium, RSV particles also colocalized with LsST6 on the microvilli of cell membranes and cytoplasm (Fig 4). This evidence strongly suggested that LsST6 is a key factor for enabling RSV entry into the midgut epithelium of the planthopper.
RSV particles were added to a liquid culture of Sf9 cells that expressed the heterologous gene LsST6; by 7 h, they had colocalized with LsST6 on the cell membrane of Sf9 cells. Few viral fluorescence signals were found in the cytoplasm by 15 h, but more signals were detected in the cytoplasm by 20 h (Fig 5A). Immunoelectron microscopy also consistently showed the same results; RSV particles colocalized with LsST6 in the cell membrane at 7 h (Fig 6A) and were observed in the cytoplasm of Sf9 cells at 15 h and 20 h (Fig 6B and 6C).
We then quantified the mRNA level for RSV and LsST6 at various times using northern blots and RT-qPCR. The RSV mRNA level was obviously higher at 20 h than at 7 h when LsST6 was expressed (Fig 5B). The RT-qPCR data also supported the northern blot results: at 15 h and 20 h, the level of viral RNA was 1.5 and 3.5 times higher, respectively, than at 7 h (Fig 5D), whereas the level of LsST6 mRNA was approximately 1.2 and 1.3 times higher, respectively, at 15 h and 20 h (Fig 5C). That viral particles passed though the cell membrane, mediated by LsST6, revealed that LsST6 played a critical role in RSV entry Sf9 cells.
When dsRNA of LsST6 (dsLsST6) was injected into third-instar nymphs, LsST6 mRNA level in the midgut had declined by 61% and 87% at 1 d and 2 d compared with the control (insects injected with dsGFP). Subsequently, LsST6 mRNA level remained 20–40% lower than in the control group injected with dsGFP (Fig 7A). We then allowed third-instar nymphs that had been injected with dsLsST6 or dsGFP to feed on RSV-infected plants for a 2-d AAP, then quantified RSV and LsST6 mRNA levels in treated insects after different times using RT-qPCR and northern blots. The RSV mRNA level in insects injected with dsLsST6 was 22% to 35% of that in SBPH injected with dsGFP (Fig 7C), indicating significant interference with LsST6 expression (Fig 7B). The northern blot also showed that the quantity of RSV RNA was evidently lower than in the group injected with dsGFP after 4 and 8 d (Fig 7D). In addition, viral acquisition and transmission by dsLsST6-injected SBPHs decreased by nearly 80% compared with those injected with dsGFP (Fig 7E, S3 Table). Confocal images also showed fewer RSV infection sites and fewer RSV particles in a single epithelial cell than in the control at 2 and 4 d (Fig 7F). Overall, these results demonstrated that RSV initial infection of the midgut epithelial cells was inhibited when LsST6 expression was knocked down.
We also injected viruliferous insects with dsLsST6 or dsGFP. Compared with the RSV mRNA level in the insects injected with dsGFP, the level in the dsLsST6-injected insects only decreased by 10–20% at 2, 4 and 8 d after injection, while the LsST6 mRNA level decreased by 65–70% (Fig 8A and 8B). The northern blot assay showed that the change in RSV mRNA levels followed a similar trend in insects after injection with dsLsST6 or dsGFP (Fig 8C). Confocal images also suggested that the quantity of RSV in epithelial cells was similar to the control (Fig 8E). In addition, two groups of treated viruliferous insects had a similar transmission efficiency (Fig 8D, S4 Table). Therefore, LsST6 has no significant effect on virus replication.
The insect midgut consists mainly of a single layer of epithelial cells, with extensive microvilli on the lumen side and a porous basal lamina on the hemocoel side [37, 38]. The midgut absorbs the nutrients necessary for insect survival and provides an environment for the development and multiplication of viruses and parasites [39, 40]. The midgut epithelial cells have been identified as the initial infection site and the first barrier to virus invasion [1, 13, 41]. Our results on the distribution of RSV particles in the alimentary canal over time also demonstrated that the midgut epithelial cells of L. striatellus served as the initial infection site of RSV. After successful invasion of the midgut, RSV began its replication process, then spread into neighboring cells.
Most viruses invade insect epithelial cells via specific interaction between the structural proteins of the virus and the cell surface receptor complexes in vectors, similar to their infection of host cells [18, 19, 42, 43]. Viral surface components have been well demonstrated to play an important role in virus infection and transmission [44, 45], and putative surface receptors including glycans and glycoconjugates for flaviruses have been found in host and insect cells [46, 47]. A 32-kDa laminin-binding protein and 35-kDa prohibitin that mediate entry of Venezuelan equine encephalitis virus and Dengue virus-2, respectively, into mosquito cells have been identified [48, 49]. Membrane alanyl aminopeptidase N has been identified in the pea aphid as responsible for the entry of the pea enation mosaic virus into the aphid gut [50]. Most of these putative receptors in insects were discovered by in vitro interactions; in vivo evidence is still lacking.
We thus used cellular and molecular biological techniques to advance our understanding of the interaction between virus and insect vector. Here, we found that LsST6 not only strongly interacts with RSV NP in vitro and in vivo, but also alters the cellular location of NP and then colocalizes with it in the cell membrane of Sf9 cells. Moreover, LsST6 mediates the entry of RSV particles into Sf9 cells that expressed the heterologous gene LsST6. In the vector insect body, RSV initially binds to the cell surface of midgut epithelial cells where it colocalizes with LsST6 in the cell membrane. When expression of LsST6 was knocked down in healthy insects injected with dsLsST6, viral titre and acquisition subsequently decreased significantly. Therefore, LsST6 plays an important role in facilitating virus invasion in both Sf9 model cells and the midgut epithelial cells of the vector insect, L. striatellus.
Our previous bioinformatic analysis revealed that LsST6 has 85% similarity with NlST6, a facilitative glucose/fructose transporter in brown planthoppers (N. lugen) [51]. They both belong to the major facilitator superfamily (MFS) of transporters, which are ubiquitous among organisms and enable the import and export of essential nutrients and ions (not just sugars), the excretion of metabolic end products and deleterious substances and communication of the cells with the environment [52, 53]. Some members of the MFS are also exploited by viruses to invade host cells. Glut1, a receptor for HTLV [36], was recently found to mediate glucose transport, which regulates human immunodeficiency virus (HIV) infection in human T cell lines [54]. Glut1 of shrimp is thought to be a putative cell surface receptor for white spot syndrome virus [55]. Feline leukemia virus subgroup C receptor (FLVCR), another member of the MFS, is considered to be the cell surface receptor for feline leukemia virus [56, 57]. HTLV and HIV infection of host cells are all regulated by Glut1-mediated glucose metabolism, via an increase in Glut1 expression and to a change in the conformation of the protein [53,54]. These studies thus provide a precedent for the involvement of another MFS member, LsST6, in virus invasion of midgut epithelial cells in L. striatellus using a similar transport mechanism, rather than receptor-/clathrin-dependent endocytosis or membrane fusion and/or actin-based tubular structures to overcome the cell barriers. Knock down of the expression of LsST6 in healthy insects of L. striatellus, resulted in a decrease in virus acquisition after they fed on RSV-infected rice plants compared with the control insects with the functional gene. The viral titre in viruliferous insects that were similarly treated by injection with dsLsST6 had no significant changes. These results suggest that LsST6 mediates RSV entry into cells but is not involved in virus replication. On the basis of our results, we propose that LsST6 on the cell membrane of epithelial cells in the midgut mediates RSV invasion during facilitative transport of glucose/fructose from the phloem sap of rice plants across the cell membrane.
Interestingly, we found that LsST6 not only interacted with the outer capsid of RBSDV and SRBSDV, but also colocalized with each in the cell membrane of Sf9 cells, but it did not colocalize with NP of RGSV. Because L. striatellus can transmit both RSV and RBSDV but not RGSV, we consider that the two transmitted viruses may also require LsST6 to mediate their entry into midgut epithelial cells. Although the planthopper cannot transmit SRBSDV efficiently, previous evidence has shown that this virus does invade midgut tissues, but it does not spread into the hemolymph or other organs of SBPH [58, 59]. SRBSDV cannot break through the release barrier of the midgut because it cannot replicate enough to reach the threshold required for further spread, and/or the siRNA antiviral pathway has a direct role in controlling viral dissemination from the midgut [60, 61]. RGSV cannot invade the midgut epithelium of L. striatellus, because LsST6 did not interact or colocalize with the NP of RGSV. Therefore, LsST6 might specifically mediate initial infection by the numerous viruses that are transmitted by L. striatellus. Based on all the data obtained, we propose a model by which RSV overcomes the midgut infection barrier in vector planthopper. After entering the alimentary canal of the vector insect and arriving in the midgut, intact RSV particles can bind to the midgut epithelial cells, where the NP of RSV interacts specifically with sugar transporter 6 on the cell membrane and is transported into the epithelial cells, where it replicates and finally disseminates to other parts of the vector (Fig 9). This model should also be applicable to other viruses transmitted by L. striatellus.
In conclusion, our results provide direct evidence that LsST6 is essential for RSV invasion of the midgut epithelial cells in its insect vector. The fact that LsST6 can also interact and colocalize with the outer capsid or NP of other viruses transmitted by L. striatellus suggests that numerous arboviruses might use a similar vector protein to invade the midgut epithelium of the insect vector. This key vector protein could be used as a target for blocking virus transmission and lead to a new strategy to control outbreaks of diseases caused by arboviruses.
Nonviruliferous and viruliferous L. striatellus with a high affinity for rice stripe virus were reared on healthy and RSV-infected rice seedlings (cv. Wuyujing 3), respectively [29, 34]. Every 3 months, the offspring of viruliferous insects were confirmed as RSV-positive using RT-PCR.
Viruses (RSV, RBSDV and SRBSDV)-infected rice leaves collected from the field were stored at −80°C in the lab (61) and RGSV-infected rice plant was provided by Prof. Taiyun Wei (Fujian Agriculture and Forestry University). RSV particles were extracted from RSV-infected rice leaves as described previously [62], and purified virions were stored at −80°C.
The mouse monoclonal anti-RSV antibody was kindly provided by Prof. Jianxiang Wu (Zhejiang University). The rabbit anti-RSV antibody was the kind gift of Yan Huo (Chinese Academy of Sciences). An anti-LsST6 monoclonal antibody against the LsST6 peptide SKGDHNTEAALP was produced by Abmart (Shanghai, China). The following antibodies were obtained from the sources indicated: mouse monoclonal anti-Myc tag (cat. 66004, Proteintech), rabbit polyclonal anti-His tag (cat. 2365, Cell Signaling Technology), Dylight 488 goat anti-rabbit IgG (cat. A23220, Abbkine), Dylight 488 goat anti-mouse IgG (cat. A23210, Abbkine), Dylight 549 goat-anti-mouse IgG (cat. A23310, Abbkine), goat anti-mouse IgG+HRP (cat. 32430, Thermo), goat anti-rabbit IgG+HRP (cat. 32460, Thermo). Alexa Fluor 633 phalloidin was obtained from Invitrogen and 4’,6-diamidino-2-phenylindole (DAPI) was purchased from Sigma.
The genes including RSV NP, RBSDV p10, SRBSDV p10, RGSV NP and LsST6 were amplified using specific primers (S1–S2 Tables) and were subsequently cloned into the bait plasmid pDHB1 or prey plasmid pPR3-N (Dualsystems Biotech) to generate pDHB1-RSV NP, RBSDV p10, SRBSDV p10 or RGSV NP or pPR3-LsST6. Genes including RSV NP, RBSDV p10, SRBSDV p10 and RGSV NP linked with a Myc tag sequence were cloned into plasmid pFastBac (Invitrogen), while LsST6 was inserted into the BamHI/XbalI sites of pFastBacHTB (Invitrogen) vector containing a 6×His tag.
Sf9 cells were incubated in Sf-900 III SFM Serum Free medium containing 5% newborn calf serum at 27°C.
To confirm any interaction between LsST6 and RSV NP, RBSDV p10, SRBSDV p10 or RGSV NP, we used the DUALhunter starter kit, a yeast two-hybrid system based on the reconstitution of ubiquitin. A clone of yeast strain NMY51 was selected and incubated in 25 ml yeast peptone dextrose adenine agar (YPDA) at 30°C with shaking until the culture reached an OD546 of 0.6–0.8. The culture was collected by centrifugation, and the pellet was suspended in 1.5 ml water. Then, 1.5 μg bait vector plasmid (pDHB1-NPs/-p10s) and 1.5 μg prey vector plasmid (pPR3-LsST6) were added to 100 μl culture resuspended in 300 μl PEG/ lithium acetate Master Mix (50% PEG, 1 M lithium acetate and 125 μl single-stranded carrier DNA) and incubated in a 42°C water bath for 45 min. Finally, the mixture was collected by centrifugation at 700 × g for 5 min, and the pellet resuspended in 100 μl 0.9% NaCl (wt/vol), then plated onto selection plates of DDO (SD-trp-leu) and QDO (SD-trp-leu-his-ade) medium with 20 mM 3-aminotriazole (3-AT) and incubated for 3–4 days at 30°C. For distinguishing false-positive interactions, the clone grown on DDO was incubated on 1 ml liquid DDO overnight at 30°C with shaking until the culture reached an OD546 of 0.5–0.8 to check for β-galactosidase activity. The culture was then collected by centrifugation, and the pellet was added to 100 μl of freshly prepared lysis mixture (9.95 ml of one-step lysis and assay reagent with 50 μl of dye stock solution) provided in the HTX High-throughput β-galactosidase kit with a brief vortex.
Sf9 cells were transfected using Cellfectin II according to the manufacturer’s instructions. Briefly, 2 × 106 cells were added to each well of a 6-well culture dish for at least 30 min. Then 2 μg recombinant bacmid–plasmid DNA encoding LsST6, RSV NP, RBSDV p10, SRBSDV p10 or RGSV NP was mixed with 8 μl Cellfectin II and incubated for 20 min. Then each mixture was added to a well of the 6-well culture dish and incubated at 27°C for 5 h. The transfection mixture was then removed and replaced with growth medium. Transient expression was measured 72 h later using LSCM and western blot.
Genes including RSV NP, RBSDV p10, SRBSDV p10 and RGSV NP linked with a Myc tag sequence were cloned into plasmid pFastBac (Invitrogen), while LsST6 was inserted into the BamHI/XbalI sites of pFastBacHTB (Invitrogen) vector containing a 6×His tag. Recombinant LsST6 DNA was used to cotransfect Sf9 cells with recombinant RSV NP, RBSDV p10, SRBSDV p10 or RGSV NP DNA during a 72-h incubation at 27°C, then cultures were collected and lysed in the lysis buffer (20 mM Tris-HCl pH 7.6, 150 mM NaCl, 0.5% NP-40, 5 mM EDTA, and complete protease inhibitor cocktail tablets) for 1 h on ice. After the cultures were centrifuged at 10,000 × g for 20 min at 4°C, 50 μl protein A/G plus agarose beads was added to the supernatant for 1 h at 4°C to decrease any nonspecific binding of proteins. Then the supernatant was incubated with 2 μl (1 μg/μl) anti-Myc antibody for 1 h, followed by incubation with protein A/G plus agarose beads at 4°C with end-over-end agitation for overnight. The protein A/G plus agarose beads were washed with washing buffer (20 mM Tris-HCl pH 7.6, 150 mM NaCl, 5 mM EDTA) 5 times, then mixed with 1× loading buffer (0.08 M Tris pH 6.8, 2.0% SDS, 10% glycerol, 0.1 M dithiothreitol, and 0.2% bromophenol blue) and finally boiled for 5 min. The cell lysates and IP cultures were separated on SDS-PAGE gels and transferred to nitrocellulose membranes. Membranes were incubated with antibody of anti-His or anti-Myc (1:3000) for 1.5 h at room temperature, then incubated with the secondary antibody-alkaline phosphatase-conjugated goat anti-mouse or rabbit IgG (1:5000) at 37°C for 1 h. Membranes were then incubated with a chemiluminescent substrate mixture and imaged using the imageQuant LAS 4000 mini biomolecular imager (GE Healthcare Life Sciences, USA).
RSV particles (1.5 μg/μl) were added to the Sf9 cells transfected by recombinant bacmids LsST6 or empty bacmids at 48 h, then the sf9 cells were rinsed with PBS buffer 3 times and collected at 7, 15 and 20 h after RSV particles were added to the culture medium. To avoid the influence of residual inoculum, the cells are collected by centrifugation to remove the inoculum. Then they were washed with double-distilled water for 3 times to avoid residual inoculum, finally they were prepared for RNA extraction. The relative quantity of RSV particles RNA was determined using RT-qPCR and DIG-northern blot.
A partial sequence of LsST6 and GFP was amplified using specifically designed primers as templates to synthesize dsRNA using the protocol for the T7 RiboMAX Express RNAi System. Approximately 400 third-instar and nonviruliferous nymphs of L. striatellus were injected with 23 nl dsLsST6 (2.5 μg/μl) or GFP (2.5 μg/μl) using an Auto-Nanoliter Injector (Drummond, USA) and then allowed to feed on the RSV-infected rice plant for a 2-day acquisition access period (AAP). At 2, 4 and 8 days after the AAP, RNA was extracted from 50 insects to quantify the LsST6 and RSV mRNA levels by RT-qPCR and DIG-northern blot, and the midgut tissues were excised from 50 insects for immunofluorescence assay. The remaining insects were tested for RSV by RT-PCR with RSV-specific primers. To identify the influence of injection on RSV titre and transmission in insects, we injected 400 third-instar and viruliferous nymphs with 23 nl dsLsST6 (2.5 μg/μl) or dsGFP (2.5 μg/μl) and then transferred them to healthy rice seedlings. Fifty insects were collected for RNA extraction and detection at 2, 4 and 8 days after injection. One hundred insects were transferred to healthy seedlings (1/seedling) for a 10 h inoculation access period, and the seedlings were then grown in a greenhouse for 3 weeks. The infection status of each seedling was assessed by RT-PCR using specific primers for RSV NP. Midguts of the remaining insects were excised for immunofluorescence assay. RT-qPCR and northern blot were used to quantify any changes in RNA levels for RSV and LsST6, and LSCM was used to visualize RSV particles in the excised midgut of L. striatellus. Each set of experiments was repeated three times.
cDNA was synthesized from 1 μg of total RNA using a FastQuant RT Kit according to the manufacturer’s instructions at 4°C for 3 min, 42°C for 15 min, and 95°C for 3 min. The RT-qPCR was performed using a SuperRealPreMix Plus (SYBR Green) kit, a reaction volume of 20 μl (10 μl of PCR buffer, 0.6 μl of each primer [10 μM/μl], 3 μl of template cDNA, and 5.4 μl of DEPC H2O and 0.4 μl 50× ROX Reference Dye) and ABI-7500 thermocycler (Applied Biosystems). The thermocyling program was 94°C for 15 min, followed by 40 cycles of 95°C for 10 s and 60°C for 32 s. Fluorescence was measured at the end of every 60°C extension phase. Beta-actin of SBPHs or ecdysoneless (ECD) of Sf9 cells were used for normalization as housekeeping genes in respective experiments. RT-qPCR data were analysed using the Livak method (2−ΔΔCt) [63]. The experiments were repeated 3 times independently.
RNA probes for LsST6 and RSV detection were labeled with DIG using the DIG Northern Starter Kit according to the manufacturer’s instructions. Total RNA extracted from SBPH or Sf9 cells was separated in 1.2% formaldehyde agarose gel and then transferred to nylon membranes using a vacuum regulator (Bio-Rad, USA) for 3–4 h. The membranes were then incubated with the respective RNA probes for 16–20 h in a 65°C hybridization oven, incubated with anti-digoxigenin-AP for 30 min, then in CDP-Star solution for 5 min and imaged with the imageQuant LAS 4000 mini (GE, USA).
Sf9 cells previously fixed on cover slips were incubated in 4% (wt/vol) paraformaldehyde for 30 min at room temperature. After being washed 3 times with PBS buffer, the samples were then incubated in osmotic buffer (2% [vol/vol] TritonX-100 in PBS) for 15–30 min at room temperature, then incubated with anti-Myc monoclonal antibody (MAB) (1:400), anti-His MAB (1:400), anti-His rabbit polyclonal AB (1:400) or anti-RSV MAB (1:500) in PBS containing 3% (wt/vol) BSA at room temperature for 1 h and then with goat anti-mouse (1:400) or goat anti-rabbit (1:400) secondary antibody labeled with Dylight 488 or Dylight 549 in PBS for 1 h at 37°C after extensive washing with PBS buffer. The nucleus was stained with 50 nM 4′,6-diamidino-2-phenylindole (DAPI) in PBS at 37°C for 10 min. Midgut tissues excised from the planthoppers were fixed in 4% (vol/vol) paraformaldehyde for 2 h at room temperature and incubated in osmotic buffer for 30 min at room temperature. Then the samples were incubated in anti-RSV MAB labeled with Dylight 549 (red) or anti-ST6 labeled with Dylight 488 (green) MAB for 1.5 h at room temperature. All samples were visualized with LSCM (Zeiss LSM880, GER), and the images saved in ZEN 2011 blue light.
Dissected midguts or Sf9 cells were fixed for 2 h in 2% (vol/vol) paraformaldehyde and 2% (wt/vol) osmium tetroxide in PBS, and after sequential dehydration in 30%, 50%, 70%, 90%, 100% and 100% alcohol, midguts or Sf9 cells were embedded in LR Gold Resin (cat. 62659, Sigma). Sections (70–90 nm) of the midguts or Sf9 cells were cut using an ultramicrotome (Leica, GER), then blocked for 30 min in blocking buffer. The sections were then incubated at room temperature with the antibodies in the following order: anti-RSV rabbit serum (1:300) for 1.5 h, 10-nm gold-conjugated goat-anti-rabbit IgG for 1 h, anti-LsST6 mouse serum (1:50) for 1.5 h and 5-nm gold-conjugated goat-anti-mouse IgG for 1 h with a wash in distilled water after each antibody incubation. They were then stained in 2% neutral uranyl acetate (w/v in distilled water) for 10 min. The sections were viewed with a transmission electron microscope at 80 kV accelerating voltage.
Means ± SEM of three independent experiments (one-way ANOVA, least significant difference test) were statistically analysed using Prism 6 software (GraphPad); *P < 0.01 was considered statistically significant.
|
10.1371/journal.pcbi.1001047 | Detecting Remote Evolutionary Relationships among Proteins by Large-Scale Semantic Embedding | Virtually every molecular biologist has searched a protein or DNA sequence database to find sequences that are evolutionarily related to a given query. Pairwise sequence comparison methods—i.e., measures of similarity between query and target sequences—provide the engine for sequence database search and have been the subject of 30 years of computational research. For the difficult problem of detecting remote evolutionary relationships between protein sequences, the most successful pairwise comparison methods involve building local models (e.g., profile hidden Markov models) of protein sequences. However, recent work in massive data domains like web search and natural language processing demonstrate the advantage of exploiting the global structure of the data space. Motivated by this work, we present a large-scale algorithm called ProtEmbed, which learns an embedding of protein sequences into a low-dimensional “semantic space.” Evolutionarily related proteins are embedded in close proximity, and additional pieces of evidence, such as 3D structural similarity or class labels, can be incorporated into the learning process. We find that ProtEmbed achieves superior accuracy to widely used pairwise sequence methods like PSI-BLAST and HHSearch for remote homology detection; it also outperforms our previous RankProp algorithm, which incorporates global structure in the form of a protein similarity network. Finally, the ProtEmbed embedding space can be visualized, both at the global level and local to a given query, yielding intuition about the structure of protein sequence space.
| Searching a protein or DNA sequence database to find sequences that are evolutionarily related to a query is one of the foundational problems in computational biology. These database searches rely on pairwise comparisons of sequence similarity between the query and targets, but despite years of method refinements, pairwise comparisons still often fail to detect more distantly related targets. In this study, we adapt recent work from natural language processing to exploit the global structure of the data space in this detection problem. In particular, we borrow the idea of a semantic embedding, where by training on a large text data set, one learns an embedding of words into a low-dimensional semantic space such that words embedded close to each other are likely to be semantically related. We present the ProtEmbed algorithm, which learns an embedding of protein sequences into a semantic space where evolutionarily-related proteins are embedded in close proximity. The flexible training algorithm allows additional pieces of evidence, such as 3D structural information, to be incorporated in the learning process and enables ProtEmbed to achieve state-of-the-art performance for the task of detecting targets that have remote evolutionary relationships to the query.
| Using sequence similarity between proteins to detect evolutionary relationships—protein homology detection—is one of the most fundamental and longest studied problems in computational biology. A protein's function is strongly correlated with its 3D structure, and due to evolutionary pressure, protein structures diverge much more slowly than primary sequences. Because protein sequence data will always be far more abundant than high-quality 3D structural data, the computational challenge is to infer evolutionarily conserved structure and function from subtle sequence similarities. When the evolutionary distance is large and the sequence signal faint—so-called remote homology detection—this problem is still unsolved.
Stated in purely computational terms, remote homology detection involves searching a protein database for sequences that are evolutionarily related (even remotely) to a given query sequence. Most work in this area has focused on developing more sensitive pairwise comparisons between the query and target sequences, including sequence-sequence local alignments (BLAST [1], Smith-Waterman [2]); profile-sequence (PSI-BLAST [3]) and HMM-sequence comparisons (HMMER [4]); and, most recently, profile-profile [5] and HMM-HMM (HHPred/HHSearch [6]) comparisons. From a machine learning point of view, these recent methods involve building a model of the neighborhood of the query and of the target in protein sequence space and using the local neighborhood models to compute a better similarity measure. However, recent advances in massive data domains such as web search and natural language processing suggest that the global structure of the data space can also be exploited. For example, motivated by the success of Google's PageRank algorithm, we previously developed RankProp [7], an algorithm that uses graph diffusion on the protein similarity network, defined on a large protein sequence database, in order to re-rank target sequences relative to the query and substantially improve remote homology detection.
In the current study, we are motivated by large-scale learning of language models in recent work in natural language processing (NLP) [8]. This NLP work exploited large online text data sets (e.g., Wikipedia) to learn an embedding of words into a low-dimensional semantic space, inducing an embedding of sentence fragments. The embedding algorithm iteratively pushes pairs of real sentence fragments together and pulls pairs of real and randomized sentence fragments apart. Thus, at the end of training, words that are near each other in the embedding space are likely to be semantically related. Moreover, the embedding representation can be leveraged to simultaneously train models to solve multiple NLP tasks, using the framework of multitask learning [9].
Here, we present an algorithm called ProtEmbed that learns an embedding of protein domain sequences into a semantic space such that proximity in the embedding space captures homology relationships. After this large-scale training procedure, remote homologs of a query sequence can be detected by mapping the query to the embedding space and retrieving its nearest neighbors. Furthermore, as in the NLP case, we can use multitask learning to incorporate auxiliary information, where available, to improve the embedding, including structural class labels from databases such as SCOP [10] or structural similarity scores for pairs of training examples where both 3D structures are known. It is important to note that our embedding is defined naturally on protein domain sequences rather than multidomain sequences. In particular, inclusion of multidomain sequences in the training data can lead to incompatible distance relationships in the semantic space due to lack of transitivity, resulting in a worse embedding. At testing time, it may be possible to resolve the domain structure of a multidomain query sequence using the learned embedding (see Discussion); however, we only evaluate performance on domain sequence queries in the current study.
We show that ProtEmbed achieves state-of-the-art performance for remote protein homology detection, outperforming our previous algorithm RankProp, which also exploits global structure but uses a fixed weighted similarity network rather than a learned embedding. Our procedure also yields statistical confidence estimates and enables a visualization of the learned protein embedding space, giving new intuition about the global structure of the protein sequence space.
The main idea of our approach is to learn a mapping of protein domain sequences into a vector space that captures their “semantic similarity”, i.e. closeness in the semantic space should reflect homology relationships between sequences.
In order to learn an embedding of protein sequences into a semantic space, we need to define (i) a feature representation for proteins, (ii) a training signal that determines whether a given pair of training sequences are similar and should be pushed together by the algorithm, or dissimilar and should be pulled apart, and (iii) an algorithm that learns an appropriate embedding.
Let us denote the set of proteins in the database as and a query protein as , where is the set of all possible sequences of amino acids. We then choose a feature map to represent proteins as vectors. This map is necessary so that we can perform geometric operations on proteins. We use the following representation for a protein :where is the E-value returned by a surrogate protein alignment algorithm, such as PSI-BLAST, suitably transformed. Following Rankprop [7], we use the following transformation:where is the PSI-BLAST E-value assigned to protein given query and where we set the parameter . This transformation yields a stochastic connectivity matrix; i.e., the value can be interpreted as the probability that a random walk on the protein similarity network will choose to move from protein to protein . Note that, because most protein pairs exhibit no detectable similarity according to an algorithm such as PSI-BLAST, most feature values are zero. (Specifically, PSI-BLAST assigns a large maximal E-value to all database sequences for which no homology to the query is detected, and the exponential transfer function converts these values to zero.) The sparseness of the feature vectors will be important for computational reasons.
Next, we again use a surrogate protein alignment algorithm, this time as a teacher to provide a noisy training signal. We construct a training set of tuples , where each tuple contains a query , a related protein and an unrelated (or lower ranked) protein . The tuples themselves are collected by running PSI-BLAST in an all-versus-all fashion over the database of proteins. Taking any given protein as the query, we consider any protein with an E-value lower than 0.1 to be a similar protein (instance of a ); in the current implementation, instances of are chosen randomly from all training examples and with high probability will be dissimilar to . We can then, in principle, construct all possible combinations (tuples) from which we sample randomly during online training.
Given the feature vectors and the training tuples, our aim is to learn a feature embedding that performs well for protein ranking and classification tasks. We will learn an embedding functionwhere is an matrix, resulting in an embedding . Typically, is chosen to be low dimensional, e.g. . The learning procedure consists of finding a matrix such that similar proteins have close proximity in the embedding space. Specifically, we would like to choose such that, for all tuples ,expressing that should be ranked higher than , relative to an appropriate distance measure in the embedding space. We define this distance measure using the -norm (which is defined as ):
After training, given a query protein , we will rank the database using the ranking score:where we consider smaller values of to be more highly ranked.
The training objective employs the margin ranking loss [11], which has been used successfully in the field of information retrieval to rank documents given a query [12]–[14]. That is, we minimize:(1)which encourages to be smaller than until a margin constraint of is satisfied. Intuitively, the algorithm tries to push and together while pulling and apart, until the difference in distances achieves a margin of . For an equivalent formulation, we can introduce a slack variable for each tuple and enforce the constraintsfor all tuples while minimizing the objective function
This optimization problem is solved using stochastic gradient descent [13]: iteratively, one picks a random tuple and, if , makes a gradient step for that tuple as follows:(2)where denotes that the sign function is applied componentwise to the vector to yield a vector of values. Pseudocode for training the ProtEmbed embedding is given in Algorithm 1 in Text S1.
One can exploit the sparsity of and when calculating these updates to make them computationally cheap. To train our model, we choose the (fixed) learning rate that minimizes the training error, i.e. the loss defined by equation (1). We initialize the matrix randomly using a normal distribution with mean zero and standard deviation one. Overall, stochastic training is highly scalable and is easy to implement for our model, and learning can scale to millions of proteins.
After training, we precompute the embedding for every protein in the database. At test time, given a query protein , we compute its linear embedding once. Then we are left with only operations per protein in the database to perform when retrieving results for that query.
In general, recognizing remote homology relationships among protein structures is easier than recognizing remote homologies based only on protein sequences. Although structural information is available for only a subset of the proteins in the database, we would like to ensure that our embedding captures this structural information in addition to the sequence-based information provided by PSI-BLAST. We consider two sources of structural information: (1) category labels for a given protein and (2) similarity scores between pairs of proteins. For the the category labels, we use the Structural Classification of Proteins (SCOP) [10]. For pairwise similarity scores, we use pairwise structure alignments of known 3D structures using MAMMOTH [15].
We incorporate this auxiliary information using the framework of multitask learning: in addition to the main embedding task, we simultaneously learn models to solve additional tasks using appropriate subsets of the training data. The tasks share internal representations learned by the algorithm, in this case, the embedding function . In particular, we pose an auxiliary classification task using SCOP categories, and we pose an auxiliary ranking task using either SCOP category relationships or using MAMMOTH similarities. In all cases, the multitask objective function is simply the sum of the original ProtEmbed objective function and of that of the auxiliary task. We consider these two task types in turn.
Class-based data. For auxiliary data in the form of a class label for protein we train an auxiliary classification task that is multitasked with the original ProtEmbed objective, sharing the same embedding space. For each fold and superfamily class we create a vector , , which can be thought of as a set of class centroids. We then would like to satisfy the constraints:That is, proteins belonging to some class should be closer to that class centroid than proteins that do not belong to that class. We train this model using the margin ranking loss as before, and multitask this problem with the original objective using the following updates:(3)Here is a matrix containing the centroid vectors as columns, and (resp. ) is the bit vector of length whose two non-zero entries are placed at indices for the fold and superfamily of the labeled training example (resp. ). Pseudocode for training the ProtEmbed embedding with class-based auxiliary data is given in Algorithm 2 in Text S1.
Ranking-based data. For auxiliary data in the form of similarity scores between pairs of proteins, we simply add more ranking constraints into the set of tuples . That is, we consider additional tuples of the form where and are similar SCOP proteins based on auxiliary data—i.e., a similarity score comparing these proteins is above a cutoff value—while is chosen at random from all of SCOP and with high probability will be structurally dissimilar to . Then we require these additional tuples to satisfy constraints of the formanalogous to the constraints in the main optimization problem. Two examples of the use of such auxiliary constraints are given by using SCOP superfamily labels or MAMMOTH. For SCOP labels, if two proteins are in the same superfamily, we say they are similar. For MAMMOTH, we choose a cutoff value of 2.0, and a pair of proteins that has a structural alignment scoring above this cutoff is deemed to be similar. Pseudocode for training the ProtEmbed embedding with ranking-based auxiliary data is given in Algorithm 3 in Text S1.
For labeled data—namely, proteins with structural category labels and 3D structures from which to compute pairwise similarity scores—we used proteins from the SCOP v1.59 protein database. We used ASTRAL [16] to filter these sequences so that no two sequences share greater than 95% identity. This filtering resulted in 7329 sequences. Our test set consists of 97 proteins selected at random from these SCOP sequences. These test sequences were excluded entirely from the training data.
For unlabeled data, i.e. protein domain sequences without category labels or structural information, we used sequences from the ADDA domain database version 4 [17] (http://ekhidna.biocenter.helsinki.fi/downloads/adda). This database contains 3,854,803 single-domain sequences. We removed from the database sequences comprised entirely of the ambiguity code “X,” sequences shorter than 6 amino acids and sequences longer than 10,000 amino acids. We then randomly selected sequences from the remaining sequences until we had picked 3% of the original sequences. This left us with an unlabeled single domain database of 115,644 sequences.
We ran PSI-BLAST version 2.2.8 on the combined SCOP+ADDA database using the default parameters, allowing a maximum of 6 iterations. For a second and more powerful pairwise sequence similarity method based on HMM-HMM comparisons, we also ran HHSearch version 1.5.0, using default parameters. HHPred/HHSearch is considered a leading method for remote homology detection [6]. When searching for homologs to the test set domains, we added the HHSearch options “-realign -mact 0,” which uses local Viterbi search followed by MAC to realign the proteins globally on a local posterior probability matrix. Similarly, MAMMOTH was run with its default settings.
We first trained embeddings on SCOP+ADDA (with SCOP test sequences held out) using PSI-BLAST or HHSearch as the pairwise sequence comparison method to serve as “teacher” for producing tuples. In this setting, we did not make use of the category labels or structural information for the SCOP training examples. We then trained embeddings using ADDA as unlabeled data and SCOP as labeled data, where the labeled data was used in (i) an auxiliary classification task based on SCOP category labels or (ii) an auxiliary ranking task based either on SCOP category relationships or on MAMMOTH similarity scores.
As an initial proof-of-concept test of the ProtEmbed algorithm, we created an embedding of protein domains into a two-dimensional space. This embedding is necessarily underfit, because two dimensions does not provide very much capacity to learn a good embedding. However, a two-dimensional space has the advantage of being easy to visualize. We trained the embedding using the 7329 SCOP proteins from the training set, and then calculated the locations of the all SCOP proteins from all superfamilies with 25 or more members. Figure 1 shows these locations. Proteins are colored and labeled according to their SCOP superfamilies. The embedding generally places members of the same superfamily near one another.
To investigate the ability of ProtEmbed to rank homologous proteins above non-homologs, we used a gold standard derived from the SCOP database of protein domain structures. We then used PSI-BLAST, Rankprop, HHSearch and ProtEmbed to rank a collection of 7329 SCOP domain sequences with respect to each of 97 test domains. To provide a rich database in which to perform the search, we augmented the SCOP data set with 115,644 single-domain sequences from the ADDA domain database. In our evaluation, protein domains that reside in the same SCOP superfamily as a query domain are labeled positive, and domains in different folds than that of the query are labeled negative. The remaining sequences—from the same fold but different superfamilies—are ignored, because their homology to the query is uncertain. For each query, traversing the ranked list of labeled sequences induces a receiver operating characteristic (ROC) curve, which plots the percentage of positives as a function of the percentage of negatives observed thus far in the ranked list. We measured the area under this curve up to the first false positive () or the 50th false positive (). Both scores are normalized such that perfect performance corresponds to a score of 1.0.
Before training our embedding, we ran a series of cross-validation experiments within the training set to select hyperparameters; i.e., parameters that are not subject to optimization. Based on these experiments, we used, for PSI-BLAST, a learning rate of 0.05 and an embedding dimension of 250; and for HHSearch, a learning rate of 0.02 and an embedding dimension of 100. In each case, the training was run for 150 epochs, where one epoch corresponds to 20,000 tuples. We used the same hyperparameters when training with or without the auxiliary, structural information.
Figure 2 compares the performance of PSI-BLAST, RankProp, HHSearch and various versions of the ProtEmbed algorithm. The performance of each algorithm is summarized by the mean or score. To establish the statistical significance of the observed differences, we used a Wilcoxon signed-rank test with a 0.05 significance threshold. For both of the performance metrics that we considered, the ranking of the three previously described methods is the same: HHSearch outperforms Rankprop, which outperforms PSI-BLAST. Also, the standard ProtEmbed algorithm, with no auxiliary data, outperforms PSI-BLAST when it is trained using PSI-BLAST and outperforms HHSearch when it is trained using HHSearch, although for the latter comparison, the difference is only significant for the performance metric. Figure 2 in Text S1, which plots the number of queries for which the or score exceeds a given threshold, shows that the differences among methods are not traceable to queries with particularly high or low ROC values; on the contrary, the improvements from one method to the next span the entire range of ROC values.
Figure 2 shows that adding auxiliary, structural information during ProtEmbed training significantly improves the quality of the resulting rankings. Adding structural information to ProtEmbed improves the mean score by 0.038–0.170 and improves the by 0.083–0.180. Perhaps most strikingly, if we consider ProtEmbed trained from HHSearch, the initial embedding is 0.154 away from a perfect score, whereas the embedding learned using SCOP rankings is only 0.025 away from a perfect score. Thus, in this case, structural information removes 83.7% of the residual error. In general, using SCOP information leads to better rankings than using MAMMOTH. This is not surprising, because we are using a gold standard based on SCOP. Between the two modes of representation, the SCOP ranking appears to give better results than using SCOP class-based structural information. This result is somewhat surprising, because our gold standard is based explicitly on SCOP classes and perhaps suggests that the ranking representation is more resistant to overfitting.
In evaluations of remote homology detection algorithms, some researchers prefer to ignore members of the same family as the query, since these family members are presumably easy to identify [18]. To ensure that our results are not dependent on family-level information, we repeated the ROC calculations above, but we skipped target proteins that fall into the same family as the query. Figure 3 in Text S1 shows that the conclusions above remain unchanged in this setting: ProtEmbed outperforms HHSearch, RankProp and PSI-BLAST, and using structural information significantly improves ProtEmbed's performance.
Next, we evaluated how well ProtEmbed scores are calibrated between queries. We say that our scores are well calibrated if pairs of query and target sequences at similar distances from each other in embedding space also have similar degrees of homology, regardless of where the query embeds. If this property holds, then the scores generated by ranking database sequences relative to different queries can be compared to each other and modeled to assign statistical significance.
The experiment reported in Figure 2, in which ROC scores are computed separately for each query and then averaged, only measures how well the target sequences in the database are ranked relative to each query sequence. To measure the calibration of the scores among queries, we sorted all of the scores from all 97 test queries into a single list. The resulting ROC curves are shown in Figure 3. The overall ranking of methods is the same as in Figure 2, in order of improving performance: PSI-BLAST, Rankprop, HHSearch, ProtEmbed. To obtain calibrated scores, PSI-BLAST, Rankprop and HHSearch include specific calibration procedures—calculation of E-values for PSI-BLAST and HHSearch, and calculation of superfamily probabilities for Rankprop. ProtEmbed, in contrast, requires no explicit calibration procedure; instead, the scores are naturally calibrated because they all correspond to distances in a single embedding space.
To be useful, a homology detection algorithm must provide scores with well defined semantics. For example, PSI-BLAST reports an expectation value, or E-value, that corresponds to the number of scores as good or better than the observed score that are expected to occur in a random database of the given size [3]. Rankprop reports for each query-target pair the probability that they belong to the same SCOP superfamily [19]. To convert ProtEmbed distances to an interpretable score, we employed a simple empirical null model in which protein sequences are generated by a third-order Markov chain, with parameters derived from the SCOP+ADDA database. We randomly generated decoy protein sequences according to this null model, and we embedded these proteins into the PSI-BLAST ProtEmbed space. Empirical analysis of the resulting sets of scores (Figure 1 in Text S1) shows that the left tail of the null distribution is well approximated by a Weibull distribution. To compute a p-value, we select the null distribution based on the length of the given query sequence. Further details are given in Text S1.
We cannot use these p-values directly, because we must correct for the large number of tests involved in searching a large sequence database. To do so, we employ standard false discovery rate-based multiple testing correction procedures. In particular, for a given query, we first estimate the percentage of the observed scores that are drawn according to the null distribution [20]. We then use the Benjamin-Hochberg procedure [21] to estimate false discovery rates, including the multiplicative factor . Finally, we convert the estimated false discovery rate into a q-value [20], which is defined as the minimum FDR threshold at which an observed score is deemed significant.
For many users of alignment tools such as PSI-BLAST, the multiple alignment produced with respect to a given query is as useful as the rankings and accompanying E-values, because the multiple alignment provides an explanation of the ranking. However, a method like ProtEmbed does not rely solely on multiple alignments. Therefore, although it would certainly be feasible to create, in a post hoc fashion, an alignment of the ranked proteins up to, e.g., a specified ProtEmbed q-value threshold, such a multiple alignment is not likely to accurately reflect the semantics of the ProtEmbed embedding space. Instead, we propose to use a multidimensional scaling approach to project the top-ranked protein domains into an easy-to-visualize 2D representation.
To illustrate how effective such a visualization can be, we systematically generated 2D maps of the neighborhood for all 97 test set domains, using a q-value threshold of 0.01. Thumbnail versions of all 97 neighborhoods are provided in the supplement.
Here, we focus on a single example. Figure 4 shows the structure learned by the embedding near a particular query, the C-terminal domain of Staphylococcal enterotoxin B (PDB ID 3seb). Figure 4(A) shows the neighborhood of the query relative to the initial PSI-BLAST based feature embedding of the domain sequences, projected into 2D for easier visualization. This mapping corresponds to the initialization of the embedding algorithm, before any training. We see that the other members of the query's family—the superantigen toxins, C-terminal domain (SCOP 1.75 ID d.15.6.1), shown in green—are generally near the query in the initial embedding, but these true positives are intermingled with members of a functionally related but structurally distinct superfamily, the bacterial enterotoxins (SCOP 1.75 ID b.40.2, shown in blue) as well as several members of unrelated superfamilies. When we map the query sequence into the final embedding space (Figure 4(B)), we now find that it lands in a tight cluster of its family members, which is near but separated from the cluster of related bacterial enterotoxins. Meanwhile, unrelated superfamilies are appropriately separated into distinct clusters distant from the query. In this example, the homology detection performance improves from an score of 0.091 ( of 0.716) relative to the initial embedding to a perfect (and perfect ) of 1.0 after training.
We have shown that ProtEmbed learns an embedding of protein domain sequences such that proximity in the embedding space reflects homology relationships. Due to efficient stochastic gradient descent methods, the training algorithm can scale to millions of sequences. A flexible multitask framework also enables the use of additional label or ranking information, e.g. protein structural classes or pairwise structural similarity scores, where known, to improve the embedding. Given a test query sequence, its embedding can be computed in the same time that it takes to run the underlying pairwise sequence alignment method. The query's homologs can then be efficiently retrieved by determining the nearby database proteins based on their precomputed embedding coordinates. Moreover, using a faster but less accurate pairwise alignment method, such as PSI-BLAST, together with ProtEmbed, when supplied with labeled data through an auxiliary task, leads to better performance than state-of-the-art but slower pairwise alignments methods, such as HHSearch, used on their own. Moreover, use of more sensitive PSI-BLAST parameters rather than the default choices could potentially further improve the performance of the embedding.
While alignment-based pairwise sequence similarity scores are used as features for calculating the embedding, ProtEmbed does not produce multiple sequence alignments for query sequences as an output of its computation. Instead, the embedding neighborhood of the query can be visualized for insight into the relationship between the query and its homologs. For further sequence-based analysis of query-homolog similarities, hits from the ProtEmbed neighborhood could be used to compute an alignment using standard methods [22] or newer graph algorithm approaches [23].
The ProtEmbed algorithm learns its embedding on domain sequences rather than full-length protein sequences, because the embedding only makes sense when transitivity relationships hold. For example, a multidomain sequence will have sequence similarity to its constituent domains, which will typically also be represented as entries in the database; if these domains are dissimilar from each other, then the set of pairwise relationships lead to conflicting constraints during training. Nonetheless, it is possible to process a multidomain query sequences using ProtEmbed by first applying an existing domain decomposition algorithm [24] and then embedding each domain separately. Alternatively, one could potentially use the embedding to help resolve the domain structure: first, one could run a pairwise alignment method such as PSI-BLAST to determine the start and end positions of all the hits, and then these subsequences could be embedded separately as candidate domain sequences. The p-value for the score between the embedded candidate sequence and its nearest neighbor in the database should generally favor candidates with boundaries similar to those of the true domains.
Protein sequence analysis is one of the oldest subfields of computational biology, with mature and specialized tools designed to describe the local structure of protein sequence space. By adapting new techniques from massive data domains such as natural language processing and web search, we have demonstrated that the global structural of protein sequence space can be exploited for classical problems like homology detection.
|
10.1371/journal.pbio.1002512 | Spatial Embedding and Wiring Cost Constrain the Functional Layout of the Cortical Network of Rodents and Primates | Mammals show a wide range of brain sizes, reflecting adaptation to diverse habitats. Comparing interareal cortical networks across brains of different sizes and mammalian orders provides robust information on evolutionarily preserved features and species-specific processing modalities. However, these networks are spatially embedded, directed, and weighted, making comparisons challenging. Using tract tracing data from macaque and mouse, we show the existence of a general organizational principle based on an exponential distance rule (EDR) and cortical geometry, enabling network comparisons within the same model framework. These comparisons reveal the existence of network invariants between mouse and macaque, exemplified in graph motif profiles and connection similarity indices, but also significant differences, such as fractionally smaller and much weaker long-distance connections in the macaque than in mouse. The latter lends credence to the prediction that long-distance cortico-cortical connections could be very weak in the much-expanded human cortex, implying an increased susceptibility to disconnection syndromes such as Alzheimer disease and schizophrenia. Finally, our data from tracer experiments involving only gray matter connections in the primary visual areas of both species show that an EDR holds at local scales as well (within 1.5 mm), supporting the hypothesis that it is a universally valid property across all scales and, possibly, across the mammalian class.
| It was recently shown that the network of connections between different areas of the macaque cortex has strong structural specificity in terms of the strength of connections as a function of the distance between areas. This has led to a model of cortex connectivity that predicts many observed architectural features, including the existence of a strong core-periphery organization. When viewed across species, increases in brain size are accompanied by a relative decrease in connectivity, and thus an important question is whether there are architectural commonalities in the cortical networks within the mammalian branch. Here, based on tract tracing data from the folded macaque brain and the smooth mouse brain, we introduce a common model framework that allows network comparisons between species. We show that despite important differences in size, the cortices of both species share several network invariants, suggesting that the mammalian cortex exhibits universal architectural principals. This framework also captures differences between the two brains, including the fact that, unlike the macaque, the mouse core includes primary areas and that there is a relative decrease in the frequency of long-distance connections in the large macaque cortex compared to mouse. This approach allows network architectural extrapolations to the human cortex.
| Understanding brain networks is arguably one of the major challenges of the 21st century [1]. The mammalian cortex is an extraordinary computational device, and analysis of its network properties with 107–1010 neurons and 1011–1015 synaptic connections is still largely unresolved. In the brain, activity of a single neuron encodes relatively little information; instead, that is achieved via population coding, through spatially distributed temporal activity patterns of cell assemblies. This contrasts with packet-switching information technology (IT) networks, which encode information directly into the packets and the network merely ensures routing between any two nodes. Since the spatiotemporal activity of cell populations is strongly determined by their connectivity and physical layout, cortical network structure and its spatial embedding play a significant role in the brain’s processing algorithm, in sharp contrast with IT networks.
A purely bottom-up approach to deriving global brain function from local circuitry is currently intractable [2]. In contrast, a meso-scale approach is more feasible, focusing on the network of interactions between the elements of a mosaic of distinct areas representing the loci of function-specific computation (visual, auditory, somatosensory, motor, etc.). As the mammalian brain is shaped by evolution, morphological and areal network level inter-species comparisons will help identify those features that are conserved across species from those that are species-specific. This will lead to a better understanding of network structural properties and provide valuable clues to the evolution of brain function [3]. However, progress in this direction has been hindered due to the absence of (i) the necessary data to address the physical properties of the network between areas and (ii) adequate theoretical network comparison methods.
Published connectivity maps using consistent interareal tract tracing studies, first in the macaque [4] and more recently in the mouse [5,6], allow consideration of the network as a directed, spatially embedded and weighted graph (weights representing neuronal connection densities projecting between areas). The absence of full homology between the nodes (areas) and edges (projections) of the networks of the two species makes it difficult to determine commonalities and similarities between them. However, if generic, global organizational principles exist (constraining the adaptation and growth of cortical connections in similar ways), then we expect to see similarities at the statistical level between the network features in the two species.
Here we show that the cortical networks in the macaque and the mouse in fact do exhibit a common organizational principle despite their very different evolutionary trajectories and large differences in brain size. Supplemented by partial tract tracing data in the microcebus (the mouse lemur) we suggest that this principle and the associated network model is a universal determinant of the interareal network across mammals, allowing tentative predictions for the human brain.
Expansion of the cerebral cortex is accompanied by an increase in the proportion of white matter relative to brain size [7–10]. However, this increase is not rapid enough to maintain a constant neuronal connection density (defined as the fraction of neuron-to-neuron connections compared to all possible ones). Thus, an increase in brain size is expected to result in a reduction in the long-distance connectedness of cortical areas [11–14]. The reduction of the fraction of connections with cortical expansion and the minimization of the metabolic costs are important design features of the cortex [4,15–26]. One can hypothesize that this wire minimization constitutes a critical constraint for the optimal placement of areas in the cortex, serving to increase communication efficiency in larger brains [11,27–29], and is supported by recent evidence suggesting reduction of long-distance connectivity with increases in brain size [28].
Recent retrograde tract tracing data in macaque [30] provides supporting evidence precisely of such a wiring constraint, in the form of an exponential decay of the wiring probability p(d) with projection distance d: p(d)~e−λd, with a decay length (~1/λ) that is short relative to hemispheric dimensions (in the macaque λ ≅ 0.19 mm−1, corresponding to a decay length of 1λ≅5.2 mm). A simple way to think of the decay length 1λ is that every increase by 1λ in projection length leads to a decrease in the number of projections by a factor of 1e≅0.37 (i.e., 37%). Note that using the base of the natural logarithm is convenient, as in this case 1λ is equal to the average projection length, providing a simple, intuitive interpretation. We refer to this decay property of connection density with distance as the Exponential Distance Rule (EDR). Retrograde labeling using fluorescent tracers (see Materials and Methods section) is an accurate labeling method that reveals all incoming connections j→i to an injected (target) area i by labeling the cell bodies of the neurons in source area j whose axons make connections in area i. Importantly, there is no transneuronal labeling, so the retrograde labeling method used yields only one-step incoming connections to the injected nodes of the network.
Note that the EDR is purely a property of the distribution of the physical lengths of individual axons, without regard to any network topological structure. The EDR states that there are many fewer long-range axons than short ones and quantifies this: the number of axons of length d that we find in the cortex is proportional to e−λd. In general, to experimentally establish the EDR, we do not need to work with brain areas as nodes of a network; we only need to be able to count neurons and measure the corresponding axon lengths. In this sense, the EDR is a more basic and general property than the description of cortical connectivity as a network at some coarse-grained (e.g., mesoscale) level. Once the level of description is defined (e.g., areal), the network properties are, however, consequences of the distribution of the axonal lengths connecting the vertices. Since connectomes are embedded in physical space, the EDR property effectively constrains the topological structures that connectomes can form across different levels, ranging from the single neuron to the areal level [31].
In addition to the discovery of the EDR in the macaque, the consistency and completeness of this tract tracing data [32] has led to a deeper insight into the interareal network properties of the macaque cortex [30,33]: it revealed a much denser (ρ = 0.66) interareal cortical graph than previously reported (network density is defined as ρ = MN(N−1), where N is the number of areas and M is the number of connected ordered area pairs, see glossary). High density graphs have low specificity at the binary level (areas connected or not), so that what distinguishes one area from another is the particular combination of areas it is connected to, combined with the weights of the connections, i.e., their connectivity profile or fingerprint [33–36]. Because the range of weights spans many orders of magnitude (five in the macaque), the specificity of individual connectivity profiles is actually very high [5,30,37].
We first give a schematic description of EDR-based network models (Fig 1) before developing a formal methodology for comparing EDR model graphs with experimentally obtained graphs, thereby allowing a quantification of the predictive power of the EDR network model for a given brain. This sets the stage for empirical measurements in the mouse brain, which are required for the construction of a mouse EDR model (Fig 2) and to examine how well the mouse EDR graphs fit with selected local and global mouse network properties obtained from empirical data (Fig 3). We next identify the core-periphery organization in the mouse network and show that it is well captured by the mouse EDR model. The following section is dedicated to a comparison of the capacity of the mouse and macaque EDR models in predicting empirically measured motif distributions (Figs 5 and 6) [38]. Analysis of network motif distribution is a recognized method of capturing the functional features of a network. The motifs analyses suggest the existence of common architectural features in the networks of both species; the following section analyses these structural commonalities by investigating the connection similarity index profiles between all node-pairs as a function of their spatial separation. However, in order to be able to perform comparisons involving distances in brains of very different sizes, we first introduce a common spatial template by an appropriate dimensional rescaling of the two brains. This allows us to show that, effectively, there is a common distribution of similarity indices as a function of adimensional separation in both brains (Figs 7 and 8). The finding that similarity changes across the cortex are only relatively consistent in the two species naturally leads us in the following section to consider the differences in cardinal features governing functional layout and to relate these differences to species characteristic properties of the cortex such as size and cortical folding (Fig 9). We conclude with a Discussion (Fig 10) in which we hypothesize that the EDR is a universal property across scales, i.e., valid also locally (through the gray matter), not just globally (through the white matter), and across the mammalian branch. As preliminary evidence supporting this hypothesis, we present results of tracer experiments (Fig 11) involving local connections only within the gray matter in three species—mouse, macaque, and microcebus—and quote results from other experiments in the rat. We conclude with mathematical arguments that further support the universal character of the EDR and speculate on the importance of these findings for understanding the human brain.
To what extent does the EDR, as a connectivity constraint, determine the properties of the interareal network? To address this issue, one needs (i) a family of EDR-based network models and (ii) a method of comparison between the model-generated networks and the experimental data network. The exponential decay rule ~e−λd in the macaque was obtained from collating all the labeled neurons (over 6.4 million) following tract tracing experiments in different areas and constructing an interareal distance matrix, the latter estimated as the distances between the area barycenters through the white matter (WM), along the shortest paths. Here axonal p(d) should be interpreted as an average property (see Fig 1A), the probability that an axonal bundle projects to a distance d, independently of the specific functional nature of the areas. At this level of description, the strength of the connection between areas, expressed as the fraction of labeled neurons (FLN), depends uniquely on their geometrical separation. Thus, the network is viewed as a spatial, directed, and weighted graph dependent on the matrix D = {dij} of interareal distances dij. We emphasize here that the EDR arises from the estimated probability distribution of axon lengths. Although the strength-distance relation is consistent with the EDR, the probability distribution of axons lengths provides a more compelling demonstration of the property and leads naturally to the parametric EDR model described below. The probability density function, q(d), of the distances in the matrix D is typically a unimodal distribution (Fig 1B), which, when combined with the exponential decay p(d), leads to a log-normal distribution of edge weights, confirmed by the empirical FLN data [4,39].
The EDR distribution with the corresponding distance matrix D in a given brain naturally defines a parametric family of random graphs, called EDR random graphs (Fig 1C), parameterized by the decay rate λ. For these model graphs we make the choice p(d) = λe−λd, where now λ is the (only) model parameter. To distinguish the decay rate parameters in these models from the experimentally measured ones, we denote the latter as λexp, e.g., for macaque λexpmac = 0.19 mm−1. We also employ, as a null model, the constant distance rule (CDR) family of random graphs, where there is no dependence of connection probability on distance, corresponding to the λ→0 limit, i.e., to the choice p(d) = const.
The EDR family of random graphs is defined via a simple algorithm [4] in the spirit of the Maximum Entropy Principle, i.e., it is based only on the given information (p(d) and D), while all else is uniformly random. The algorithm proceeds as follows: First, we randomly draw a connection length d from the distribution p(d). Second, we choose uniformly at random an area pair whose separation distance in the matrix D falls in the same distance bin as d, according to some binning criterion (bin sizes used in this study were typically 5 mm for the macaque and 0.4 mm for the mouse) and finally, insert a randomly oriented connection between them. Multiple connections between the same area pair in the same direction generate the weights for the directed edges with a log-normal distribution. These steps are then iterated until the graph density in the model reaches the observed value in the experimental network.
We denote the data network obtained from the experiments by Gexp (e.g., for the macaque we use Gexpmac, and for the mouse Gexpmac). Our goal is to compare the properties of the EDR model networks with the properties of Gexp. Since the model networks are only based on distance-dependent connection probabilities, one cannot expect perfect agreement (edge-by-edge) with the biological connectivity graph Gexp, however, if the distance rule is a strong determinant of the interareal network, the model graphs should be statistically similar to Gexp. The comparisons are performed via parameter matching of network properties [4]: for a given network property P, the interareal distance matrix D and parameter λ is used to generate a large ensemble of EDR graphs GEDR(λ). By varying λ we determine the value λP via minimizing the deviation |P(Gexp) − 〈P(λ)〉|, with the average 〈∙〉 taken over at least 103 EDR graph realizations from GEDR(λ). Thus the model parameter is determined so that the average of P in the model is as close as possible with the value of P observed in the data network. We then compare the fitted value λP with λexp, the decay rate obtained directly from the experiments. If the two are close, then the EDR is a strong determinant for the measure P of the cortical network. Thus, the extent a particular measure in the EDR model and in the data network agree, i.e., |P(Gexp) − 〈P(λP)〉| with respect to the same comparison with the CDR model, i.e., with |P(Gexp) − 〈P(λ = 0)〉|, expresses the degree to which the EDR influences that particular measure in the cortical network. This analysis is repeated with several local and global network measures. The more measures for which there is an agreement between λP and λexp, the stronger the effect of the EDR in shaping the interareal network. This method also has the added advantage of identifying those network properties that are not well described by the EDR, and thus, based on the nature of these measures, providing us with clues for additional network mechanisms.
In the macaque, the EDR model predicts very well many local, global and weighted network properties of the interareal network (see [4] for details), and thus it is a strong determinant for the large-scale network organization of the macaque cortex. It also captures its pronounced core-periphery organization (i.e., a densely connected set of areas—core, with feedback and feedforward links to/from a more loosely connected set of peripheral areas), with the core strongly dominated by associative areas [4,40].
The EDR network model of cortical connectivity represents a radical departure from previous, purely topological models of cortical networks, which do not take into account their physical, i.e., weighted and spatially embedded nature, and this has now been well documented in the recent literature [41,42]. The spatial clustering and geometrical positioning of the nodes in the EDR model in the macaque is observed to strongly echo the functional layout of the cortex as revealed by numerous physiological and anatomical studies [36,43].
To determine whether a similar description is valid for the mouse (Mus musculus) cortex, we first conducted retrograde tracer experiments in the mouse neocortex (S2 and S3 Figs), in order to determine the projection length distribution p(d), which, indeed, shows a clear exponential decay (Fig 2A). The decay rate, λexpmus, was determined from an exponential fit as λexpmus = 0.78 mm−1, with a 95% confidence interval of (0.72, 0.83) (see Fig 2A, and inset). This exponential decay is to be compared with the same distribution for the macaque from Fig 2B in Ref [4] (see Table 1 for the λ parameter estimates). The distance matrix Dmus was determined from flattened cortex measurements. The corresponding distance distribution q(d) is unimodal, as shown in Fig 2B, which is to be compared with the same distribution for the macaque from Fig 2C in reference [4] (the consequences of the differences in these distributions are discussed in more detail in the section “Functional Layout in Terms of Spatial Clustering of Cortical Areas”).We then applied the network analysis described in the section above to the largest available edge-complete graph (the status of connectivity between all pairs of nodes is known) of 33 areas in one hemisphere of the mouse neocortex [5,6], denoted by Gexpmus from here on (S1 Fig). This mouse dataset contains 719 directed pathways and has an interareal network density of ρmus = 0.68, similar to that reported in the macaque.
Fig 3 shows the proximity of λPmus obtained using the parameter matching method to the decay rate λexpmus for several network measures including the number of area pairs connected uni- (M1) or bidirectionally (M2), 3-motifs, clique distributions and the second largest eigenvalue of the symmetrized form AAT of the adjacency matrix A. These measures have been selected in part because they probe graph properties from local to global scales, and are of varying complexity. Additionally, these measures (see glossary for definitions), and in particular the deviations from their values in random graphs carry functional significance: unidirectionally connected areas depict an asymmetric role in information processing (driver versus driven nodes), the 3-motifs have extensively been studied as building blocks of functional organization in complex networks [38], cliques identify maximally connected network regions usually representing activity-specific strongly correlated communities or clusters, and the second largest eigenvalue is related to the rate of spreading processes (e.g., epidemics or information) on the network [4].
The different comparisons and fits based on these measures are highly consistent and indicate λ to be in the range 0.78–0.93 mm−1 (purple vertical band in Fig 3A–3D). The broader range in mouse of |λP−λexpmus| compared to macaque might be due in part to the fact that the mouse connectivity matrix was generated by anterograde tracing [5,6]. Other statistical network properties, such as degree distributions, are likewise well captured by the EDR network model with λ = 0.78 mm−1, (see S4 Fig).
A clique (see glossary in S1 Text) is a complete subgraph of a network, i.e., it carries the maximum number of possible edges between its nodes. In dense graphs (thus with many cliques) the size (number of nodes) distribution of the cliques provides insight into the network’s heterogeneity [4]. The largest cliques in dense graphs can be used to define the cortical network core [4,40]. As in macaque [4,40], the clique distribution analysis in the mouse (Fig 4A) reveals a distinct core-periphery structure. The mouse connectome, Gexpmus, includes a dense core of 12 nodes organized into the two largest cliques each of size 11, plus a periphery of 21 nodes. There are a total of Mcc = 131 links within the core, Mcp = 190 links from the core to the periphery, Mpc = 170 from periphery to core, and Mpp = 228 links within the periphery. Densities for the mouse are the following: core 99% (versus 92% in macaque), periphery 54% (versus 49%) and the links between the core and periphery, 71% (versus 54%). The likelihood of a core having 12 nodes in a random graph on 33 nodes with the same density ρ = 0.681 as in Gexpmus is vanishingly small: (3312)(1321)p131(1−p)1 = 2.07×10−12 (versus 10−17 in the macaque). To see how well the EDR model reproduces the clique distribution, we define a scalar deviation measure σcl(λ) between the clique-size distributions in the data and the EDR model as the root mean square (RMS) of the clique-count log-ratios. The best agreement between the two distributions is achieved at λclmus = 0.93 mm−1 (Fig 3C) and the clique distributions in the model and data are rather close at this value (Fig 4A).
Anatomically, the cortical core in the mouse shows significant differences with that previously reported in macaque, the most striking being that the mouse core includes portions of primary somatosensory cortex (SSp-ll and SSp-tr) and primary motor cortex (MOp) (Fig 4B). While additional injections may well expand the core membership in macaque, primary areas in the macaque core are extremely unlikely, given the rarity of connections linking primary areas [30]. This contrasts with the mouse where the inter-primary area subgraph has a density of over 80% [5,6]. In agreement with the presence of primary areas in the mouse core, the two-dimensional map of the flattened cortex (Fig 4D) shows that the mouse cortical core might be spatially more widespread across brain regions compared to that of the macaque, where the core appears concentrated in frontal and parietal areas [4]. Note that in both mouse and macaque, the core areas have overall, higher in-degrees than non-core areas (Fig 4B for the mouse). The wider spatial spread of the mouse compared to the macaque core may reflect the relative expansion in primates of higher-level association cortex with respect to the primary areas [3]. These differences in the cortical core of the mouse and macaque need to be considered in light of the proposal that in primates at least, the core is related to cognitive architectures such as the global workspace, thought to be involved in consciousness [40,44].
Network motifs refer to the different possible connectivity patterns of a small, fixed number of nodes. For example, in the 33-node mouse cortical network there are (333) = 5456 triplets of nodes, each of which has one of the 16 connectivity patterns shown in Fig 5A. Three-node motifs have been proposed as the building blocks of network circuits and their pattern of variation in frequency to reflect functional properties of the networks [38]. For instance, motif 10 (oriented 3-cycle) is significantly under-represented in the cortex, while motif 3 is significantly over-represented (lone, bidirectional link) when compared to a random network, in both species (see Fig 5D).
As in macaque [4], the mouse EDR predicts the observed motif frequency distributions in this species significantly better than does the CDR (Fig 5B). Despite the marked quantitative differences in motif distributions between mouse and macaque (Fig 5C), there could, however, be qualitative similarities. Testing this requires comparing the observed motif profiles to that of a randomized null model [38] consisting of an ensemble of random networks having the same degree sequence as the data. Graphs were uniformly sampled from this ensemble by repeatedly rewiring edges [45].
Fig 5D shows how the motif counts of the empirical connectomes differ from such a randomized null model. As similar patterns are observed in both species, these findings suggest that they are part of the same class of large-scale networks with similar architectural and functional constraints. Repeating this analysis for networks generated by the EDR model (see Fig 6) we find a remarkable similarity in the motif profiles not just between the networks of the real data network and the EDR model but also between the model networks of the two species (Fig 6). This confirms the existence of a common network architectural invariant in these two species. This is unexpected, insofar as the decay rates and the distance matrices are very different between the two cortices. Since the motif profiles are binary measures, these findings indicate graph structural similarity between the two brains. In order to test whether there are significant similarities in the large-scale connectomes of the two species beyond the constraints imposed by the EDR, we used the EDR model as a null model [46]. S5 Fig shows that motif counts continue to look similar between the mouse and macaque, although their similarity is now less pronounced. To further probe the wiring similarity between the mouse and macaque connectomes, we next study the connectivity similarity profile measure.
Elsewhere we have demonstrated that a quantitative measure of the similarity of the connectivity profiles of target cortical areas decreases in a regular fashion with increasing distance between them, i.e., the closer two target areas are, the more their source areas overlap [4,33]. We have also shown that changes in similarity reflect the functional layout of the cortex [33], and thus it is natural to compare the behavior of this measure between the mouse and macaque.
A similarity index [4] can be defined for both incoming (in-link similarity) and outgoing (out-link similarity) connections. In order to compare macaque and mouse similarity indices, we focus here on the incoming connections, as those are the ones fully specified for all the injected areas in the macaque dataset. Next, we analyze the similarity between the connectivity profiles, for all possible target area pairs. The in-link similarity index for any two (target) areas is a measure describing the extent to which both targets receive/or do not receive in-links from the same source areas, compared to a fully randomized state of the network (see Materials and Methods section for details). Fig 7 shows the distribution of in-link similarity indices as function of the distance between all area pairs in the mouse (Fig 7A) and the macaque (Fig 7B). In both species, in-link similarity decreases with increasing distance between the area pairs, i.e., areas that are further apart on the cortical sheet have increasingly dissimilar in-link connectivity profiles on average, while the opposite is true for areas that are closer to one another. The colored regions in both Fig 7A and 7B are the probability densities of in-link similarity indices generated by the corresponding EDR models, with red corresponding to higher, and blue to lower probabilities; in both cases the EDR model captures the average behavior rather well.
In order to compare distance-dependent quantities between brains of very different sizes, all distances are rescaled by the average interareal distance in each species (〈d〉mus = 4.54 mm and 〈d〉mac = 26.35 mm). Interestingly, as the largest distances are dmaxmus = 10.1 mm and dmaxmac = 58.2 mm respectively, this fits both brains onto the same adimensional template, as dmaxmusdmus = 2.22 mm and dmaxmacdmac = 2.21 mm. Fig 7C shows the corresponding distribution of adimensional distances q(d/〈d〉). When plotting the in-link similarity indices against the rescaled distances (Fig 7D), we find a remarkable overlap between the clouds of points in the two species. This is rather surprising given the fact that they have very different decay rates λ. They also have rather different interareal distance matrices as the macaque cortex is folded, resulting in it having a more peaked distance distribution than the mouse (Fig 7C). Fig 8 shows the sensitivity of the in-link similarity indices using the EDR models in both the mouse (panels 8A–8D) and the macaque (8E–8H). For a given distance matrix, the point clouds are observed to rotate as a function of λ in both species, and hence there is no a priori reason for the overlap in Fig 7D. This overlap, however, is an indication of the existence of a network architectural invariant, present in both species, also reflected in the motif profiles discussed earlier. Further explanation for the significant, overall overlap between the similarity distributions for the two species is provided in the Discussion section.
With the help of the common adimensional template defined above we now discuss species-specific characteristics in our comparison of cortical networks. The EDR decay p(d) can simply be recast in terms of adimensional distances, by writing (d)~e−λd = e−γ dd, where γ = λ〈d〉 is the adimensional (or normalized) decay rate. Accordingly, γmus = 0.78 × 4.54 = 3.54 and γmac = 0.19 × 26.35 = 5, showing that on the common template, the mouse has a shallower connectivity decay than the macaque. The distribution of distances in the mouse is broader compared to the macaque (Fig 7C), which when coupled with the shallower connectivity decay contributes to the mouse cortex experiencing a less constraining effect of the EDR than does the macaque. This difference in the EDR between the two species explains some of the differences in the functional layout of the cortex in mouse and macaque. In Fig 9A and 9B we show the same similarity indices for all area pairs as before but also indicate which area pairs are connected (black circles) and which are not (white circles) and provide smooth estimates (colored regions) of connection probability as a function of similarity and adimensional distance. Comparing Fig 9A and 9B we see that in macaque, spatially clustered, presumably functionally related neighboring areas are heavily interconnected and share similar connectivity profiles, while more distant areas show weaker probability of connectivity and similarity index. This relationship between probability of connectivity, spatial separation and similarity is, however, weaker in the mouse.
In both species, connection probability changes as a function of distance. Fig 9C and 9D show how this relationship differs in the two species. Consistent with a steeper EDR in the macaque, neighboring areas exhibit 100% connectedness, and the probability of connections (density) decreases smoothly and consistently with distance to around 10% density at maximum distances [47]. This contrasts with the mouse, in which neighboring areas do not quite reach densities of 100% and widely separated areas have densities in the region of 50% to 80% (Fig 9C). Hence, these results show that compared to macaque, in the mouse, widely separated areas are more likely to be interconnected. These differences in the probability of being connected as a function of distance between the two species appear highly significant (smooth curves in Fig 9C and 9D).
Numerous studies point to the cost of long-distance connections as an inherent design challenge associated with differences in brain size [48]. One way to define total wire length is: Λ = Σi,jAijDij, where A denotes the binary adjacency matrix and D is the interareal distance matrix. Yoked permutations of the rows and columns of the adjacency matrix reassign the distances to each pair of areas while maintaining the connectivity unchanged. As in macaque [4], the total wire length of the mouse inter-areal network is significantly shorter than a random permutation of the areas (S6 Fig). Simulated annealing methods [4] showed that optimization of area placement can lead to a 12% reduction in total wire length in the mouse, significantly higher than the 5% reduction obtained in macaque [4].
Next, we address the strength of connections with the expectation that long-range connection strengths (expressed as FLNs) would decrease in the larger brain. Due to the EDR, the FLN clearly decreases with distance. Distinguishing interareal association and canonical connections allows an improved understanding of the effect of distance on connection weight (for definition of associative and canonical connections see [3]) (Fig 9E and 9F). This suggests that the decline in FLN is steeper in canonical cortex compared to association cortex, so that the long-distance association cortex connections are one to two orders of magnitude stronger than the connections between canonical cortex areas with the same separation (see [3]). However, the results suggest that the decline in weight with distance is steeper in the macaque compared to the mouse. Together these findings show that compared to the larger macaque cortex, in the smaller mouse brain long-distance binary interareal connections are marginally more numerous. By contrast there is a highly significant increase in the weight of the long-distance connections in the mouse, and this species difference is more pronounced in the projections of association than in the canonical connections of the primary areas.
The present meso-scale network investigation of the neocortex, with appropriate network comparisons, provides detailed information on a common organizational principle that explains numerous network features in two widely separated species, with distinct evolutionary histories. Based on phylogenic considerations, and the fact that evolution is essentially a tinkerer [49], one expects to find evolutionarily preserved features embedded in these networks, i.e., architectural invariants. Evolutionarily preserved features, however, often are expected to manifest themselves as organizational principles tied to biophysical constraints.
The success of the mammalian class includes adaptation to diverse habitats and lifestyles, which is in part attributed to the behavioral flexibility ensured by the neocortex [50]. The modulation of corticogenesis [51] has led to extant mammals exhibiting a five-orders of magnitude range of brain size [52], going from small-brained mammals that include miniaturization of ancestral forms to the expansion and additional arealization that characterize primates, especially humans. The present results suggest that the EDR plays a key role across the mammalian order to optimize the layout of the inter-areal cortical network allowing larger-brained animals to maintain communication efficiencies combined with increased neuron numbers. Our results indicate that the EDR and the associated network model provide a unifying framework to capture common network properties but also some of the differences across the mammalian branch and thus allow network comparisons between species. The EDR decay rate λ and cortical geometry (interareal distances) significantly impact on the structural heterogeneity of the cortical network with important consequences for the general functional layout and core-periphery structure, that we speculate, could be involved in higher cognitive processes [40]. The limitation of the EDR model stems from the fact that the EDR describes an overall, or average property. At this level, without additional determinants, it should not be used as a generative model of individual connections as we have emphasized elsewhere [4,40]. If we plot the decay of the probability of connections for several target areas, as shown in Fig 10A for the mouse, we find significant variability. The black line in Fig 10A, the average decay, is the same as that in Fig 2A. The fluctuations for a given target, however, are not noise, but rather they are part of a signal. This we illustrate in macaque: Fig 10B shows the consistency of fluctuations following repeat injections in area V1, in five different individuals. There are numerous factors that one might need to take into account to better understand this variability. For example one may need to consider the observed systematic variation in neuron numbers across the cortex [53,54], the anisotropy of axon outgrowth distributions [55] and possibly diverse developmental factors [56]. Overall, however, these considerations emphasize that the EDR network serves as a framework, upon which other details are imposed.
Note that in order to assess the ability of the EDR model (or any connectome model) to reproduce properties of empirical network data, it is crucial that the data is as edge-complete as possible, i.e., that the connectivity between any two nodes is known. Otherwise the lack of fit between model and data cannot be used to discard the model [57]. This holds for two reasons: (a) the EDR network model produces complete connectivity information between its nodes, it cannot generate “untested” connections, by default, and (b) many network measures can be sensitive to the absence or presence of an even a small fraction of connections in the network. It is also important to emphasize the roles of cortical geometry [58] and that of areal segmentation in shaping the network properties of the connectome. Since the connection probability depends on distance, network properties are influenced by the relative proximity of areas. In turn, the strength of connections between functionally defined areas correlate with the amount of signaling activity between them and therefore with their functional roles within the information processing hierarchy in the brain. Ad-hoc segmentations, however, will generate ad-hoc distance matrices for the EDR model, and accordingly, the model networks would no longer be interpretable from a functional circuitry point of view, and in this sense, it is important to use optimally defined functional parcellation of the cortex.
Our comparative analysis of motifs and connectivity similarity indices demonstrates the existence of network architectural invariants, which in turn imply that the EDR parameter λ and areal positioning (geometry) are not independent parameters: while both change during evolution, the changes are orchestrated in such a way as to ensure that certain network/circuitry properties are preserved. As argued in the introduction, the network, i.e., the graph connectivity (form) must play a significant role in the information processing algorithm itself (function), and thus these network invariants are a reflection of common processing dynamics in the cortex.
Our use of a normalized or adimensional distances facilitates comparisons across brains of different sizes. Fig 10C shows directly the fingerprint of such universal principles in neocortical organization: it shows the connection probability decay on the adimensional template brain from a common target area (area V1) in macaque (data from reference [30]), mouse as well as microcebus. At short to medium distances where the vast majority of neurons are located, decays are identical, but are observed to change in a species dependent fashion for the long-range connections. Microcebus belongs to a group that contains the smallest existing primates, with a brain under 2 cm in length. Although the microcebus data is only for V1, it remarkably fits to the same adimensional template, with a decay rate λ between that of mouse and macaque, suggesting that the quantitative differences that distinguish the species are due to both brain size, and primate-rodent differences.
The EDR could be the expression of the consequence of a universal information processing principle implemented in the cortex across several scales, specifically to include single neurons in the local circuit, which present over 80% of the cortical connectivity [31,39]. Hence, the two major ingredients for the EDR are found in the local circuitry, the log normal distribution of synaptic weights [59] and an exponential decay of connection distances as reported here. Further, the experimental evidence presented in Fig 11, shows that p(d) follows a nearly identical, exponential decay out to within 1.5 mm for both mouse and macaque, with λexplocal≅4.54±0.08 mm−1. These are gray matter, non-myelinated connections, and are observed to have a very different decay rate than white matter connections. Thus, at least in area V1, the decay of connectivity with distance seems to behave in a very similar fashion in both mouse and macaque, and therefore surprisingly the decay rate in the gray matter does not appear to be related to brain size. Using the reported data in [59] for the rat visual cortex obtained from quadruple whole cell recordings, the local decay rate in the rat can be determined to be λexplocal≅4.96 mm−1, a value consistent with the one found above in mouse and macaque, above. Table 1 summarizes the EDR related parameters in the mouse and macaque, for both white matter and gray matter connections.
The universal character of the EDR is further supported by mathematical arguments. The exponential distribution (EDR) is memoryless (Markov property), i.e., in our case, the probability that an axon of some length grows by an additional amount is independent of its current length (within cutoff limits). In this way, it has the property that f(d+l) = f(d)f(l), where f(l) = ∫l∞p(l′)dl′ = e−λl is the probability of an axon growing to a length beyond l. The exponential distribution is the only continuous distribution with this property [60]; for all other distributions, growth depends on the current length, i.e. on past growth history. This also implies that the EDR is the maximum entropy probability distribution for axonal lengths with given expectation value (= 1/λ), see [61]. These properties are evolutionarily advantageous, conferring maximum adaptability during cortical expansion. Moreover, as more neurons are added, the probability distribution of the shortest connection among an arbitrary number of other connections also obeys an exponential distribution [62], making the EDR an invariant property locally as well, supported by the experimental data quoted above.
The present findings could have important consequences for understanding the human brain. The recognized limitations of current tractographic analysis of diffusion MRI data [63,64], means that direct observation of long-distance connections in the human brain is not presently feasible. Given the specificity of long-range cortico-cortical connectivity [47], this technical limitation has important consequences for understanding the human connectome, and we believe that comparative connectomics as developed in the present study will be a necessary step for determining universal principles of cortical connectivity. Hence, an in-depth understanding of the influence of changes in brain size will play an important role in better understanding the human brain. Since the EDR leads to a decrease in the strength of long-range connections in macaque compared to mouse, we may hypothesize that increase in brain size leads to increased reductions of weight in long-range projections for the whole mammalian branch. In the human brain the small number of fibers in such long distance connections will pose an acute problem for detection for some time. This could constitute an important limitation. For example, one could speculate that the low weight of human long-range connections may contribute to an increased susceptibility to disconnection syndromes, such as have been proposed for Alzheimer disease and schizophrenia [65–67].
Experiments were performed in male and female PV-Cre [68] (Jax: 008069), x Ai9 reporter mice (Jax: 007905), harboring the loxP-flanked STOP cassette, which prevented the transcription of the tdTomato protein driven by the chicken β-actin (CAG) promoter [69]. The crossing produced Cre-mediated recombination, which resulted in the expression of the red fluorescent protein in the subset of parvalbumin (PV)-positive GABAergic neurons. All experimental procedures were approved by the institutional Animal Care and Use Committee at Washington University and conformed to the National Institutes of Health guidelines.
Injections were made in Microcebus murinus in area 10 and area V1. Surgical and experimental procedures were in accordance with European requirements 2010/63/UE and approved by the ethics committee CELYNE (ref 00439.02).
For tracer injections, mice were anesthetized with of a mixture of Ketamine (86 mg · kg−1) and Xylazine (13 mg · kg−1, i.p) and secured in a head holder. The body temperature was maintained at 37°C. Intracortical connections within the left hemisphere were retrogradely labeled by inserting a glass pipette (20 μm tip diameter) into the brain and injecting Diamidino Yellow (50 nl, 2% in H2O; EMS-Chemie, Gross-Umstadt, Germany) by pressure (Picospritzer, Parker-Hannafin). Injections were performed stereotaxically 0.35 mm below the pial surface, using a coordinate system whose origin was the intersection between the midline and a perpendicular line drawn from the anterior border of the transverse sinus at the posterior pole of the occipital cortex. The injections were made in the following areas: V1, RL, AL, LM, P, RSD, ACAd, MOs, SSp-bfd, SSs, Au. Area AM was injected twice (S3 Fig). The parcellation (names and locations of the areas) is based on Wang et al. [70] (S2 Fig) but differs from those used in Fig 4D and associated analyses.
Four days after the tracer injection, mice were deeply anesthetized with an overdose of Ketamine/Xylazine and perfused through the heart with phosphate buffered saline, followed by 1% paraformaldehyde (PFA) in 0.1 M phosphate buffer (PB, pH 7.4). Immediately after, the cortex was dissected from the rest of the brain, completely unfolded, flat-mounted and post fixed overnight in 4% PFA at 4°C. Next, the tissue was cryoprotected in 30% sucrose and cut at 40 μm on a freezing microtome in the tangential plane.
To survey the injection site and the distribution of labeled neurons across cortical areas, sections were wet-mounted in PB and imaged in St. Louis under a dissection scope equipped for UV- and red-fluorescence illumination. For plotting DY labeled neurons, the sections were permanently mounted onto glass slides and stored at 4°C. The distribution of DY-labeled neurons was analyzed in Bron. Plots of DY neurons were made at 20× under a fluorescence microscope equipped for UV illumination (excitation: 387–398 nm, emission: 435–475 nm), using the Mercator software package running on ExploraNova technology. Labeled neurons were contained in 12–16 sections per hemisphere. Digital charts of the coordinates of DY labeled neurons across each section were stored in the computer. Next, the regional pattern in the density of PVtdT expression was imaged under fluorescence optics. Finally, the sections were stained for Nissl substance, imaged under bright field illumination and superimposed onto the digital maps of DY and PVtdT fluorescence. In Bron, all the images were acquired using MorphoStrider software (ExploraNova).
The digital charts were saved in PDF files and were scaled in Adobe Illustrator. The charts and the corresponding images were brought to a common scale, allowing reconstruction of the sections. Sections were stacked in order, and then aligned. The landmark for the alignment of the sections was the injection site, followed by rotation around this point, allowing a 3-D reconstruction of the flattened brain. The injected area was delimited, as were the borders of the neocortex.
Automated processing was performed using in-house software, written in Python. For each case, the positions of labeled neurons inside neocortex, but outside the injected area (i.e., extrinsic neurons) were extracted in digital format, The fraction of labeled neurons per area (FLN) was estimated as the number of labelled neurons extrinsic to the injected area expressed as a fraction of the total number of labeled neurons in the cortical hemisphere [39]. Unlike the template matching procedure used in previous studies [5,6] we parcellated each cortex individually based on multiple markers expressed across different tangential sections. In a stepwise procedure we first used density differences in the expression of PV-tdT- labeled cell bodies and processes to delineate borders of single areas such as V1, S1, S2, Au, PD, UF, PV, GU, ORBI, MM. RSD, MOp, MOs, and ENTm (S2 Fig). Next, the PVtdT-expression pattern was used to outline regions which from previous studies are known to include multiple areas. The list includes (1) LM, LI, P, POR, 36p; (2) AL, LLA, RL, A, AM, PM; (3) TEp, TEa, ECT, PERI; (4) AIp, AId, AIv; and (5) ACAd, ACAv, PL, ILA, ORBm, FRP (S2 Fig). Each of these regions was further partitioned into areas based on the topographical distribution of DY labeled neurons, the size and location relative to readily identifiable areas, the rhinal sulcus, the crest of the medial wall [70–74] and the cytoarchitecture revealed by Nissl staining [5,75,76].
The segmentation shown in Fig 4D was carried out in Adobe Illustrator, combining Allen Brain Atlas boundary criteria (visualized with Brain Explorer 2) with photos of PVtdT and Nissl staining, for one section of the flattened mouse brain. The contours of the cortical areas are non-self-intersecting closed polygons; therefore, computing their centroids is straightforward. The distances between areas were considered as the distances between their respective centroids.
The Allen Brain Institute (ABI) dataset was collected on their website, offered as link in their original research publication [5]. The University of Southern California (USC) matrix was, on the other hand, extracted directly from their original article [6]. The ABI mouse atlas possesses 40 isocortical areas according to their Supplemental Table 1. Out of these 40 areas, 2 did not correspond to any line or column in the data as structured in S1 Fig (i.e. the connectivity matrix). An additional 4 areas were not considered as primary target of an injection by the authors, leading to our decision to exclude them from our analysis. We then extracted from the ABI data a 34 × 34 weighted and directed connectivity matrix.
The larger USC dataset has a finer grained parcellation than that of ABI, although based on the same fundamental scheme. We contracted the USC final matrix down to a level of 42 × 42 by merging areas together in both rows and columns so as to obtain a squared matrix.
At this point, the ABI and USC matrices had 33 areas in common, which corresponds to 97% and 79% of their full respective matrices. A similar parcellation scheme was extracted out of the two datasets, allowing complete, connection-by-connection comparison between the two matrices, see S1 Fig for the final connectivity matrix.
The database in the mouse has been generated following tracer injections in all cortical areas. The macaque data, however, was obtained from 29 injections using a 91-area atlas. Because in macaque we are using an edge-complete subgraph, the statistical features are predicted to reflect those of the, as yet unavailable, fully connected graph. However, the presently available dataset cannot give complete information on detailed areal relationships, such as for example the full membership of the cortical core.
Fig 7C shows the histograms of connection distances for mouse and macaque after normalization by the mean distance for each species. By construction, both distributions have mean equal to 1 and can be reasonably well described by truncated normal distributions. When fitting the distributions by maximum likelihood using functions from the truncnorm package [77] in R [78], the variance of the macaque normalized distances appears smaller than for the mouse data with a ratio of 0.608. Is this significant? First, we examine this question with an F-test on the ratio of variances. The test is two-sided because we do not specify a priori which variance is greater. This is a more conservative approach. The F-statistic is the variance ratio with degrees of freedom (405, 527) giving a highly significant p = 1.58 × 10−7. The test assumes normality, however. To verify the conclusion, then, we performed a permutation test that does not make the normality assumption [79]. In short, we permute the macaque and mouse labels a large number of times and recomputed the variance ratio for each new permutation. Under the null hypothesis that both distance distributions are the same, we expect a large number of variance ratio estimates on the permuted datasets that are more extreme than the variance ratio computed on the data. The p-value is computed from the proportion of ratio estimates more extreme or equal to the obtained value. For ratios, the definition of more extreme is based on the values that are less than the estimate and greater than its reciprocal. We include the ratio estimate from the dataset in the distribution of permutation estimates. S7 Fig shows the value of the ratio of variances for 100,000 permutations of the two datasets. The vertical line indicates the value obtained from the data, which is lower than all of the other values of the permutation distribution, indicating that the obtained ratio is highly unlikely under the hypothesis that both distributions are the same. The p-value is indicated in the graph. The p-value is smaller than 10−5, which is the resolution of the test for 100,000 permutations. Thus, the width of the distribution of distances for macaque is significantly narrower than that for mouse.
To analyze the density of connectivity with distance, we estimate the probability of a connection with distance. This can be done with a logistic regression. By performing the analysis on the binary connectivity (that is, presence/absence of a connection) at each distance, no binning is involved. Standard logistic regression implemented via a Generalized Linear Model with a binomial family [80] specifies that the expected value of the connection probability is related to a linear predictor through a link function that is often taken to be the log of the odds ratio or logit function. The model fit would be
g(E(Y = 1)) = β0,i+β1,iDistance
where Y is a binary variable indicating whether a connection is present between two areas, g is the link function, here log(p/(1 − p))with p the expected value or probability of a connection, and β0,i and β1,i are intercept and slope, respectively, of the linear predictor, with i varying with the species. There is no a priori reason to suppose, however, that the sigmoid function of distance that this model implies will provide an adequate description of the change in probability with distance. To allow for a more flexible description of this relation, we fit the data with a Generalized Additive Model (GAM) using a binomial family [81]. The GAM model is given by
g(E(Y = 1)) = Imousefmouse(Distance)+Imacaquefmacaque(Distance)
where fi are smooth functions of the covariates constructed from sums of spline curves with increasing complexity and Ii are indicator variables taking on the value of 1 for i = mouse (or, respectively, macaque) and 0 otherwise. The complexity (or wiggliness) of the fitted model is controlled by including a penalty in the fitting criterion based on the integrated square of the second derivatives of the f’s. The choice of degree of penalization (or smoothness) is controlled by minimizing a criterion related to prediction error (i.e., fitting some of the data and calculating the error on the remaining portion) called the un-biased risk estimator (UBRE) that is closely related to Aikake’s Information Criterion (AIC). Like AIC, UBRE favors a model that maximizes the predictability of future rather than the actual data and serves to minimize the tendency to overfit the data. The fits were performed with functions from the mgcv package [81] in R [78]. The estimates of the smooth curves for macaque and mouse are plotted in Fig 9C and 9D for macaque and mouse, respectively, with twice the estimated standard errors of the fits. To estimate the significance of the differences between the two estimates, we also fit the simpler nested model in which a single smooth curve described connectivity dependence with distance for both species. A likelihood ratio test of the nested models gave a χ2(2.26) = 54.04 with p = 3.05 × 10−12, strongly supporting that the differences in the curves are significant. Note that the generalization of degrees of freedom in the case of GAM fits are not necessarily integer valued. We extended this analysis to consider the connection probability as a smooth function of both distance and similarity. The GAM framework is used again but now to model surfaces of two variables, here giving the log-odds ratio of the connection probabilities as function of similarity and normalized distance. The model is given as
g(E(Y = 1)) = Imousefmouse(Distance,Similarity)+Imacaquefmacaque(Distance,Similarity),
where fi are now smooth 2D functions of the covariates constructed from sums of spline surfaces with increasing complexity and Ii, as before, are indicator variables taking the value of 1 for i = mouse (or, respectively, macaque) and 0 otherwise.
Contour plots of the estimates of the connection probability as a function of normalized distance and similarity are shown in Fig 9A and 9B. The color gradient indicates connection probability, passing from high (yellow, near 1) to low (green, near 0) connection probability. The curves indicate estimates of contours of constant connection probability (notated on the curves as probability values).
To evaluate the significance of the species difference displayed in Fig 9A and 9B, we also fit the simpler nested model in which a single smooth surface described connectivity dependence with distance and similarity for both species. A likelihood ratio test of the nested models gave a χ2(8.4) = 61.6 with p = 3.5 × 10−10, strongly supporting that the difference in the surfaces are significant. Note that as above the generalization of degrees of freedom in the case of GAM fits are not necessarily integer valued.
The method used to compute binary similarity indices with macaque data has been described previously [4]. Our published macaque database is made of 29 injected areas for a 91 parcellation scheme, thus giving a 91 × 29 connectivity matrix. In this context, only the in-degrees of injected areas are completely known, the out-degrees of source areas remaining incomplete. Therefore, if one wants to compare macaque and mouse using a degree-based binary similarity measure, one has to restrict oneself to in-degrees, in order to use complete data. For this reason, we detail here only the in-degree based similarity measurement calculations. The union of ABI and USC databases used here provides information about all 33 areas in terms of the in and out-going connections between 33 areas. We compared the similarity of the input pattern of pairs of areas by evaluating the number of sources areas from which both receive projections or neither do (i.e., similarity implies both projections exist or are absent; dissimilarity implies one is absent and the other is present). We define a normalized in-link similarity measure, Sxyin as follows: For any pair of areas (x, y), let nxyin denote the number of projecting areas from which either both x and y or neither x nor y receive an incoming link. Because nxyin≤33, we compute the ratio nxyin/33 for every area pair (x, y). Clearly, this number will depend on the in-degrees of x and y, denoted by kxin and kyin ( 0≤kx(y)in≤33). We define the in-link similarity as:
Sxyin = nxyin33−pxyin,
where pxyin is the expected value of the ratio (nxyin/33) if the incoming connections of x and y were distributed uniformly at random across the 33 source areas. Thus:
pxyin = (kxin33)(kyin33)+(1−kxin33)(1−kyin33),
where the first term is the probability that both x and y receive a link from a given source, and the second term is the probability that neither of them receive a link from a given source.
|
10.1371/journal.pntd.0001908 | Forecast of Dengue Incidence Using Temperature and Rainfall | An accurate early warning system to predict impending epidemics enhances the effectiveness of preventive measures against dengue fever. The aim of this study was to develop and validate a forecasting model that could predict dengue cases and provide timely early warning in Singapore.
We developed a time series Poisson multivariate regression model using weekly mean temperature and cumulative rainfall over the period 2000–2010. Weather data were modeled using piecewise linear spline functions. We analyzed various lag times between dengue and weather variables to identify the optimal dengue forecasting period. Autoregression, seasonality and trend were considered in the model. We validated the model by forecasting dengue cases for week 1 of 2011 up to week 16 of 2012 using weather data alone. Model selection and validation were based on Akaike's Information Criterion, standardized Root Mean Square Error, and residuals diagnoses. A Receiver Operating Characteristics curve was used to analyze the sensitivity of the forecast of epidemics. The optimal period for dengue forecast was 16 weeks. Our model forecasted correctly with errors of 0.3 and 0.32 of the standard deviation of reported cases during the model training and validation periods, respectively. It was sensitive enough to distinguish between outbreak and non-outbreak to a 96% (CI = 93–98%) in 2004–2010 and 98% (CI = 95%–100%) in 2011. The model predicted the outbreak in 2011 accurately with less than 3% possibility of false alarm.
We have developed a weather-based dengue forecasting model that allows warning 16 weeks in advance of dengue epidemics with high sensitivity and specificity. We demonstrate that models using temperature and rainfall could be simple, precise, and low cost tools for dengue forecasting which could be used to enhance decision making on the timing, scale of vector control operations, and utilization of limited resources.
| Without effective drugs or a vaccine, vector control remains the only method of controlling dengue fever outbreaks in Singapore. Based on our previous findings on the effects of weather on dengue cases and optimal timing for issuing dengue early warning in Singapore, the purpose of this study was to develop a dengue forecasting model that would provide early warning of a dengue outbreak several months in advance to allow sufficient time for effective control to be implemented. We constructed a statistical model using weekly mean temperature and rainfall. This involved 1) identifying the optimal lag period for forecasting dengue cases; 2) developing the model that described past dengue distribution patterns; 3) performing sensitivity tests to analyze whether the selected model could detect actual outbreaks. Finally, we used the selected model to forecast dengue cases from 2011–2012 week16 using weather data alone. Our model forecasted for a period of 16 weeks with high sensitivity in distinguishing between an outbreak and a non-outbreak. We conclude that weather can be an important factor for providing early warning of dengue epidemics, long term sustainability of forecast precision is challenging considering the complex dynamics of disease transmission.
| Dengue fever is a rapidly spreading viral infection that is endemic in more than 100 tropical and subtropical countries in Africa, the Americas, and the Asia Pacific regions. It is caused by any one of the four serotypes of dengue virus, and infection of one serotype of dengue virus does not provide cross immunity against the other three serotypes. Dengue viruses are spread by female Aedes mosquitoes through blood-feeding on human hosts. Patients suffering from dengue fever experience sudden onset of fever, rashes, muscle aches, joint pain, and leucopenia. A dengue patient usually recovers within 14 days. Nevertheless, some patients develop severe dengue which is a potentially lethal complication characterized by hemorrhagic manifestations, severe plasma leakage, and severe organ impairment [1]. Globally, about 500,000 severe dengue cases with 12,500 deaths have been reported annually [2].
Singapore has recently experienced an upsurge of dengue incidence with a 5–6 year cyclical epidemic pattern since 1980. The frequency of epidemics has increased in recent years and the nation has experienced four outbreaks over the past eight years (2004–5, 2007, and 2011). During the period 2000–2010, the annual incidence rates of dengue cases per 100,000 populations in Singapore increased from 17 in 2000 to 332 in 2005 before declining to 106 in 2010. Dengue was mainly detected in the eastern region of Singapore before 2004. Subsequently, dengue cases have been reported island-wide with the highest incidence rates in the Central and Southeast dengue zones as demarcated by the National Environment Agency of Singapore (NEA). Over the last decade, the Central and Southeast zones contributed 31% and 25% of total national reported dengue cases, respectively. Reasons for dengue expanding into western region of the island could be complex. Studies have reported that herd immunity among Singapore residents has declined from 47% in the early 90 s to about 29% by 1998; this implies a rise in susceptible populations [3], [4]. A recent seroprevalence study in Singapore showed a ratio of 23 asymptomatic cases to each reported clinical case [5]. These asymptomatic cases can possibly infect the Aedes mosquitoes and so form a reservoir of infection. The two main vectors of dengue in Singapore are Aedes aegypti and Aedes albopictus and studies have shown that they are able to disperse up to the 21st floor of a residential building [6]. All four serotypes of dengue virus (DENV 1–4) have been detected simultaneously in Singapore during the study period, except 2001. DENV 1 was the predominant circulating serotype during the outbreaks in 2004–2005 and DENV 2 was the predominant serotype in 2007 [7]. A study by Lee et al. (2012) has suggested that clade replacement in a predominant dengue serotype could also increase dengue incidence in Singapore [8].
Generally, dengue epidemiology is influenced by a complex interplay of factors that include rapid urbanization and increase in population density, capacity of healthcare systems, effectiveness of vector control systems, predominant circulating dengue serotypes, herd immunity, and social behavior of the population. Most dengue endemic countries in Asia Pacific have limited resources and/or lack of preparedness to contain dengue epidemic [9], [10]. Rising international and domestic trade and population movement contribute to the increases in domestic and cross border dengue transmission. As a result, the region is experiencing dengue epidemics with increasing frequency and magnitude. Until a vaccine or drug for dengue is available, vector control operations that eliminate adult mosquitoes and their larvae through breeding-source reduction remain the only effective method to curb dengue transmission. However, vector control can be resource and labor intensive, which poses an economic burden on nations with limited resources.
An early warning system is an essential tool for pre-epidemic preparedness and effectiveness of dengue control. In recent decades, weather variables such as temperature and rainfall have been widely studied for their potential as early warning tools to fend off climate-sensitive infectious diseases such as Malaria, Dengue, and West Nile Virus [11], [12], [13], [14].
Numerous studies have revealed the influence of weather variables on the magnitude of dengue distribution [15], [16], [17], [18], [19], [20] through the effects on life cycle development, biting rates, infective and survival rates of vectors and on the incubation period of dengue virus [21], [22], [23]. As temperature increases, Aedes mosquitoes display shorter periods of development in all stages of the life cycle leading to increased population growth; the mosquito feeding rate also increases; and dengue viruses in Aedes adult mosquitoes require shorter incubation periods to migrate to salivary glands [21], [22], [24]. Conversely, high temperatures above 35°C or heavy rainfall possibly lower dengue transmission by reducing the survival rate of Aedes [21], [23], [25]. Heavy rainfall creates abundant outdoor breeding sources for Aedes in the long run, but dry spells in some settings trigger an increase in water storage containers which can serve as breeding habitats.
In recent years, the National Environment Agency of Singapore has been using rising ambient temperature as an indicator of increase in dengue cases. During periods with median ambient temperatures above 27.8°C, the national vector control unit increases surveillance and control operations and the community are urged to increase efforts to reduce mosquito breeding habitats in the relevant residential areas [26]. Nevertheless, a more comprehensive weather-based forecasting tool is required to obtain precise information on the correlation between risk of dengue epidemic and weather conditions favorable for Aedes mosquitoes, so that dengue control efforts in the nation can be made more effective in the future.
Previous study by Hii et al. (2009) has shown that elevated weekly mean temperature and cumulative rainfall influence the risks of dengue cases in Singapore at lag times up to 20 weeks with higher relative risks of dengue cases at time lag of 3–4 months [15]. Also, a recent study by Hii et al. (2012) has suggested that a dengue early warning issues about 3 months in advance could provide sufficient time for an effective mitigation [27]. Based on previous findings, this study aims to develop a simple, precise, and low cost early warning model to enhance dengue surveillance and control in Singapore. Hence, our objectives were first to develop a weather-based dengue forecasting model to project dengue cases or potential outbreak that would allow sufficient time for local authorities to implement preventive measures and second to validate and report the performance of the forecast.
Singapore is a highly urbanized island state nation situated at 1°.17′N and 103°.50′E of the equator with a land size of about 700 km2. As of 2011, the island accommodates a population of around 5.2 million with about 93% of the population residing in either government or private high rise residential buildings [28]. As a tropical country, Singapore experiences high temperature, rainfall, and humidity year round. Weather in Singapore is influenced by the monsoon rain-belt with highest rainfall between December and early March [29].
Weekly dengue cases from 2000 to 2011 were obtained from the weekly infectious diseases bulletins of Communicable Diseases Division, Ministry of Health (MOH) Singapore [30]. The Infectious Diseases Act in Singapore stipulates mandatory disease notification within 24 hours of diagnosis by all medical clinics and laboratories.
Daily mean temperature and rainfall recorded by the Changi Airport meteorological, southeast of Singapore, for the period of 2000–2011 were extracted from the National Climatic Data Centre, National Oceanic and Atmospheric Administration (NOAA), USA [31]. Weather data were provided to the NOAA by the Meteorological Department of National Environment Agency, Singapore under the regional data collaboration agreement. The daily mean temperature was based on 24 hours average temperature; while daily rainfall was the summation of 24 hours rainfall collected using rain gauges.
We developed a dengue forecasting model using time series Poisson multivariate regression that allowed over-dispersion of data. Mean weekly predicted cases were estimated through regression on multiple independent variables that include retrospective dengue cases, weekly mean temperature, weekly cumulative rainfall, trend, epidemic cycles and seasonal factors. The forecasting model was developed using three processes: 1) model construction and training using data from 2000–2010; 2) model validation by forecasting cases in 2011–2012; and 3) sensitivity tests on outbreak diagnoses. Our statistical analysis was conducted using R [32] and STATA 11 (2009 StataCorp LP, Texas) based on 95% confidence interval.
We modeled dengue distribution patterns using retrospective data and then extrapolated the patterns several weeks ahead. We developed dengue forecasting models based on assumption that the past dengue distribution patterns will, to a large extent, continue in the future [33]. Bivariate equation (Dx) for each independent variable was first formulated using quasi Poisson regression and subsequently combined to form a multivariate model that takes multiple factors into consideration.where represents weekly average number of predicted dengue cases as a function of independent variable x.
One characteristic of infectious disease is the influence of past cases on the number of current cases. Therefore, autoregression was included in the model to account for the serial relationship between past and current cases. We derived possible lag time of serial correlation through data analysis using Autocorrelation Function (ACF), Partial Autocorrelation Function (PACF), and prior knowledge on dengue transmission. ACF analysis on dengue data showed gradual decreasing spikes that indicated strong autocorrelation between past and current cases; whereas, PACF cut off after the 4th spike suggesting a lag time of 4 weeks. However, previous studies have shown possible autocorrelation of dengue cases for longer period due to complex reasons that influence the dynamic of dengue transmission [34]. Thus, we examined lag times ranging from 4–12 weeks and selected the optimal lag order using model selection and validation tests.
We denote DAR as the autoregression of dengue cases k weeks before forecast in week t. The effects of autoregression on dengue cases are computed as:(1)where = dengue cases at lag week k,
= the constant number of dengue cases,
= parameter of autoregression at lag week k.
We examined the time gap between exposure to weather conditions and subsequent occurrence of dengue cases using cross correlation function and literature reviews. Correlation between temperature and dengue showed sine wave oscillating at about 24-weeks cycle or interval with stronger positive association between lag week 9 and 17. While correlation between rainfall and dengue revealed different length of time cycles with a negative relationship from week 0 to 22. It is possible for dengue transmission to occur several months after favorable weather conditions as mosquito eggs can withstand desiccation for several months with an average egg survival time of 18.3 weeks for Aedes aegypti [35]. We identified the optimal lag term and weather time cycle for forecasting by testing lag terms 1–20 weeks with various data cycle periods of weather variables ranging from 12 to 24 weeks. Piecewise regression was used to consider a non-linear relationship between weather and dengue cases. Thus, we partitioned weather data into 4 equally spaced percentiles with knots at 25th, 50th, and 75th percentiles using spline function.
The impact of weekly weather on dengue cases is estimated as follows:
Let depicts the number of dengue cases as a function of weekly mean temperature:(2)where is the baseline number of dengue cases; = parameter of mean temperature at lag term f in p range of mean temperature; f = t - (L+n); t = week; L = lag term in week; n = data cycle period of weekly mean temperature; p = temp11 to temp14 derived from piecewise spline function.
Let denotes number of dengue cases as a function of weekly cumulative rainfall:(3)where β0 is the baseline number of dengue cases; = parameter of rain at lag term g in q range of weekly cumulative rain; g = t - (L+m); t = week; L = lag term in week; m = data cycle period of weekly cumulative rainfall; q = rain11 to rain14 derived from piecewise spline function.
To account for non-climatic factors such as vector control, circulating serotypes of dengue virus, and other factors that influence the number of dengue cases, we performed graphical examination on the trend, cycle, and seasonal distribution patterns of dengue cases over the period 2000–2010. The trend of dengue cases increased with cyclic variation from 2000 to peak at 2005 before declining thereafter. Increases in dengue cases were generally observed in the second half of each year; while major epidemics occurred in 2004–5 and 2007. We included a curvilinear or parabola and sine function to account for trend, epidemic cycle and seasonal influence on dengue cases during the study period, respectively.
Let represents dengue cases influenced by trend over the study period:(4)whereas,
= constant,
= parameter measures the trend,
t = week,
= point in time where maximal impact of trend is reached.
Let depicts cyclical and seasonal impacts on dengue cases:(5)where = constant or baseline contribution of cycle and season,
= parameter that gives rise to cyclical and seasonal effects,
t = week.
Dengue cases are subject to interactions of multiple complex factors. Thus, we composed a Poisson multivariate regression model by combining equations (1) to (5) to account for influences of multiple factors on dengue cases. We also adjusted our findings for population change by offsetting midyear population (offset = log (pop)) during the study period.
Now we summarize our model as follows:and(6)where
is the average predicted dengue cases at week t,
is the constant derived from multivariate model, and
if all the independent variables remain constant.
Model selection was based on lowest Akaike's Information Criterion (AIC) or Bayesian Information Criterion (BIC) and standardized Root Mean Square Errors (SRMSE) of prediction. Residuals diagnoses were performed to examine and validate a good fit of the model using sequence plots to ensure sufficiency of model and constant variation of errors, and residual normality plots to examine normal distribution of errors. Furthermore, plots of fitted versus reported dengue cases were also examined for good fit of the model.
Upon selection of a model that best described the data based on 2000–2010 dengue cases, we used the model to forecast cases for years 2011 and 2012. In the first 16 weeks of 2011, we used data in the last quarter of 2010 to forecast dengue incidence from January–April 2011. Subsequently, we input only weather data for January–December 2011 and prompt our model to forecast dengue cases from week 17 of 2011 to week 16 of 2012. Only weather data that were known at the time of issuing the 16 weeks forecast were used. Forecasted dengue cases in each period were then computed as autoregression for subsequent 16-week forecast. The forecast was repeated iteratively over time to generate the forecast for 2011–2012. Finally, we analyzed forecast precision by comparing forecasted cases against real-time clinical and laboratory-confirmed dengue cases (external data) reported by the MOH in each week. We also performed sensitivity tests on these data.
An effective dengue forecast provides accurate information and minimizes false alarms so as to reduce unnecessary wastage of limited resources. We therefore further identified the optimal model using C-statistics or a Receiver Operating Characteristics (ROC) curve to evaluate and compare the sensitivity of the selected model in detecting true dengue outbreaks during both the model development and forecasting periods. The ROC curve analyzes the sensitivity or true positive rate of a model to predict outbreaks versus the false positive rate (1-specificity). The area of the ROC curve is the proportion of accurate prediction and this measures overall ability of a model to distinguish between a true outbreak and non-outbreak. We obtained annual outbreak or epidemic thresholds that were available for 2004–2011 from epidemiological reports published by the MOH Singapore. The local authorities computed warning level and epidemic threshold annually and dengue epidemic would be declared if total weekly cases exceed the epidemic threshold. We computed binary outcome of positive or negative outbreaks in each year based on given epidemic threshold values.
During the study period, the heaviest rainfall occurred in December (max = 394 mm, mean = 70 mm, std dev = 74 mm); whereas the highest temperature occurred in May (max = 30.3°C, mean = 28.7°C, std dev = 0.7°C). As shown in Figure 1, the average weekly mean temperature increased from week 1 and peaked at week 21 before declining gradually to the end of the year, whereas rainfall, which had less distinctive pattern generally demonstrated a wide ‘U’ pattern with the lowest amount of rainfall during weeks 19–42. Dengue incidence was generally higher during June–October period or between week 23 and 43, except in 2005 when Singapore experienced a dengue outbreak in early 2005 which was a spillover from the end of 2004. From 2000–2010, dengue outbreaks occurred in years 2004, 2005, and 2007.
Our findings showed that dengue cases with 6 weeks serial relationship best fitted the selected model. The cross correlation between temperature and dengue cases showed a symmetrical sine wave oscillating about the zero line at a time frame of about 24 weeks per cycle. The symmetrical pattern suggested a consistent and stable relationship between mean temperature and dengue incidence; indicating that mean temperature could be a strong predictor for dengue forecast. Simultaneously, cross correlation between weekly cumulative rainfall and dengue revealed asymmetrical oscillation at less consistent time cycles. Our findings showed that a model using weather time cycle of 20–24 weeks at lag term of 16 weeks performed with consistency during both training and forecast periods compared with models with other lag terms and time cycles. We selected the model that exhibited consistency in performance, high prediction precision, and lowest SRMSE in the forecast period. Standardized prediction errors (SRMSE) of the selected model were 0.3 and 0.32 of the standard deviation of reported dengue cases during the model development period (2000–2010) and forecast in 2011–2012, respectively. The SRMSE can be interpreted as the average error in the forecast of weekly dengue counts. Weather time cycles included in the selected model were 24 weeks for mean temperature and 20 weeks for rainfall.
According to our findings, the autoregressive term (k) in equation (1) v = 6; lag term (L) in equation (2) and (3) = 16; time cycle of mean temperature (n) in equation (2) = 24; and time cycle for rainfall (m) in equation (3) = 20.
The R2 (0.84) of our model suggests that mean temperature, rainfall, past dengue cases, season and trend explained 84% of the variance of weekly dengue distribution. The time series of fitted cases against actual reported cases as shown in Figure 2 exhibited a good fit of the model. The model was able to predict the peaks of the outbreaks that occurred in years 2004, 2005, and 2007. A scatter plot of fitted versus reported dengue cases illustrated that most of the fitted cases are scattered about the zero value with constant variance; suggesting no violation of model assumptions. Residual histograms exhibited a single modal and almost symmetrical pattern, the residual normal probability plot presented a reasonably straight line, and a residual sequence plot showed that distribution is consistent about the zero value and within upper and lower limits of +/−100. Thus, suggesting approximate normal distribution of residuals.
During the forecast for 2011–2012, the optimal model forecasted cases versus actual clinical reported dengue cases gave an average error of 0.32 of the standard deviation of reported cases. As shown in Figure 3, the model forecasted cases with lower errors against actual reported cases in the 2nd half of the year. In 2011, reported clinical cases exceeded the epidemic threshold for 5 consecutive weeks between weeks 27 and 31. Our model forecasted all the cases above the epidemic threshold with one false positive case at week 32. We have matched our forecast against external data or the real-time reported weekly cases from MOH up to week 12 of 2012; thus far, the model forecasted dengue incidence within the estimated range of errors.
ROC analysis suggested that our model performed with sensitivity ranging from 98–99% during outbreaks in 2004, 2005, and 2007. Estimated ROC areas for the period 2004–2010 indicated that the selected model performed at about 96% (CI = 93%–98%) sensitivity in distinguishing between outbreaks and non-outbreaks (Figure 4: Graph A), and in 2011 forecast with 98% (CI = 95%–100%) sensitivity in detecting a true outbreak (Figure 4: Graph B). ROC curves as shown in Figure 4 suggest a sensitivity for diagnosing true outbreaks between 90% and 98% during years 2004–2010 corresponding with a 10% to 20% risk of false alarm; whereas, in 2011 the forecasting model showed 100% sensitivity with less than 3% risk of false positive. Overall, the ROC suggested that the selected model performed consistently at above 90% during both model development and forecast periods.
Our model forecasted dengue cases up to 16 weeks ahead using retrospective weekly mean temperature and cumulative rainfall. It showed a consistent ability to separate weeks and years with epidemic and non-epidemic transmission in the training data, as well as outside the training time period in 2011. Based on lagged weather data and dengue counts the model predicted 5 out of the 5 epidemic weeks in 2011 correctly, using a 16 week lead time, thus, allowing sufficient time to prepare and potentially curb the epidemic. During the forecasting period in 2011, forecast precision based on prediction error (SRMSE) and sensitivity (ROC) tests suggested that the model forecast cases with high sensitivity for detecting outbreaks with a low risk of false alarms. The tests results during both training and forecast periods showed small discrepancy in SRMSE with absence of over fitting; thus demonstrating the stability of the model since the forecast in 2011 was performed without using actual reported cases as autoregression.
In recent years, the ability to predict local and regional weather in terms of accuracy and lead times has rapidly been improved due to advances in technology. This had allowed a better understanding of the interaction between climate and the temporal-spatial distribution of infectious diseases as well as stimulating research interest on epidemic prediction modeling [36]. We developed the weather-based dengue forecasting model based on scientific evidence that temperature and rainfall has significant influence on vectors and dengue viruses [21], [22], [23], [24], [35], [37], [38]. Dengue cases are influenced by complex interactions of ecology, environment, human, vectors, and virus factors. The lag time between weather and dengue cases could be partly accounted for by the impact of weather conditions on the biological development of the mosquito vector including long egg hatching periods and high possibility of Aedes' eggs to survive waterless for several months [21], [22], [23], [24], [35].
Several studies have documented relationship between weather variables and dengue cases in Singapore. In the late 90 s, a study that examined the links between dengue cases and Aedes mosquito population as well as weather conditions in Singapore shows that escalating temperature precedes rising dengue incidence by 8–20 weeks [16]. A recent study on the association between weather variables and dengue cases in Singapore using data from 2000–2007 has suggested that minimum and maximum temperature are strong weather predictors for the increase of dengue cases; whereas, rainfall and relative humidity have little correlation with dengue cases [39]. Using a different approach in study design, Hii et al. (2009) have quantified the effects of weekly mean temperature and cumulative rainfall on the risks of dengue cases across lag times up to 20 weeks [15], [27]. In their study they considered lag relationship between weather and dengue cases, impact of previous outbreaks on current number of cases, and influences of non-climatic factors. In addition, they applied smoothing functions to allow non-linear relationship between exposures (mean temperature and rainfall) and responses (risk of dengue cases) as well as adopted quasi-Poisson to permit over dispersion of data. Their findings show impacts of mean temperature and cumulative rainfall on risks of dengue cases vary according to each unit change in weather predictors in different lag windows (1–20 weeks). Overall, higher relative risks of dengue cases were identified at lag weeks 9–16. Evidence that weather is also a driver of dengue epidemics and trends of dengue has recently been confirmed by Descloux et al. (2012) in a study in New Caledonia [40]. It therefore seemed reasonable to assume that weather would be a precipitating factor in dengue epidemics in Singapore.
This study demonstrates that weather variables could be important factors for the development of a simple, precise, and low cost functional dengue early warning. A weather-based dengue early warning system could benefit local vector surveillance and control in several ways. First, an early warning system enhances efforts of dengue control to reduce the size of an outbreak which in turn decreases disease transmission, averts possible mortality, and lowers healthcare burden and operating costs incurred during an outbreak. Second, the use of publicly available weather variables removes the necessity for financial investment in weather-based predictive methods and further allows vector control units to focus their operations on high risk period; thus, maximizing limited vector control resources. Third, reports and study have suggested that local authorities require a maximum 3 months to curb a localized dengue outbreak [7], [27]. The forecast window of 16 weeks shown in this model offers ample time for local authorities to mitigate a potential outbreak effectively. Finally, high precision and sensitivity of a forecast minimizes the use of resources and prevents unnecessary vector control operations triggered by false alarms. Vector control can be resource and capital intensive; hence, high operating costs and unnecessary psychosocial stress in the population subsequent to false alarms could possibly hamper the decision to adopt a dengue early warning. Thresholds for true or false positive rates could vary according to scale of operational complexity and its consequences. We recommend an economic study on cost-effectiveness analysis to identify thresholds of true and false positive rates of forecast to serve as yardstick for decision making as well as to evaluate the long term benefits of an early warning against operating costs.
Nevertheless, a dengue forecasting model faces the challenge of long term sustainability of forecast precision since it assumes that a historic distribution pattern will be repeated in the future; while dengue epidemiology is influenced by a combination of factors which are dynamic and possibly evolving over time. Implementation of a new vector control policy could exert direct impact on the size of the vector population and dengue incidence rate in the locality. These changes are likely to influence the trend and epidemic cycle in the long run. Though changes of dengue distribution in the long term are inevitable due to the dynamics of disease transmission and changes of relevant policy, forecast errors can be minimized by making appropriate adjustment of the model through anticipating 1) changes in risk factors and 2) changes in related fields that will eventually influence the disease transmission. Therefore, current knowledge of factors influencing dengue distribution patterns can be used to re-calibrate the model in the future to maintain long term forecast precision.
A weather-based dengue forecast is often geographically fixed due to variability of local weather conditions. Likewise, the dynamics of dengue disease transmission in a community can be influenced by risk factors unique to that local context. Therefore, a locality based dengue forecast is usually applicable only to a specific study area. Nevertheless, the methodology of a weather-based dengue forecasting model could be extrapolated to other geographical areas. Partly due to an exponential growth of regional travels and trades, the Asia Pacific region has experienced an upsurge of dengue incidence in recent years. This suggests that a dengue endemic nation such as Singapore will no longer be able to curb or eliminate dengue without wider regional efforts. A regional dengue early warning system could signal risk of epidemic to all neighboring countries and help to prevent the regional chain effects of dengue outbreaks and so reduce the burden of dengue disease in neighboring countries. Therefore, a regional dengue forecast using weather anomaly such as El Nino index or sea surface temperature will inevitably complement and enhance the success of both national and regional dengue control.
In recent years, local authorities in Singapore heighten alert for the risk of increase in dengue cases as ambient temperature increases. Our study results demonstrate that a weather-based dengue forecasting model could provide more precise information on occurrence, timing, and size of dengue epidemics. A forecast that diagnoses outbreaks accurately and simultaneously gives about a four months window for implementing control measures could be invaluable in making control or even elimination of the cyclical dengue epidemic in Singapore a feasible possibility. We recommend a further study to analyze the possibility of incorporating a weather-based dengue early warning into the national dengue surveillance system. Further studies to improve long term sustainability of forecast precision will help to maintain the performance of a forecasting model. Moreover, a research to transform the forecasting model into a user-friendly or non-technical operational instrument comprehensible by users without specialist knowledge would encourage widespread adoption of such a dengue early warning system.
|
10.1371/journal.ppat.1007262 | Protective antigenic sites in respiratory syncytial virus G attachment protein outside the central conserved and cysteine noose domains | Respiratory syncytial virus (RSV) is the major cause of lower respiratory tract disease in infants. Previously, we elucidated the antibody repertoire following primary RSV infection in infants. Whole genome-fragment phage display libraries (GFPDL) expressing linear and conformational epitopes from RSV bound 100-fold more phages within attachment protein (G) following primary RSV infection. The G-reactive epitopes spanned the N- and C-termini of G ectodomain, in addition to the central conserved domain (CCD). In the current study, we examined the contribution of antigenic regions of G outside of the CCD to RSV-specific immunity. We evaluated the immunogenicity, neutralization and protective efficacy of all RSV-G antigenic sites identified following primary RSV infection using recombinant E. coli expressed G ectodomain (REG), CCD-deleted G ectodomain (REG ΔCCD), N- and C-terminal G subdomains, and antigenic site peptides. The REG ΔCCD, N- and C-terminal subdomains and peptides generated antibody titers in rabbits and mice that bound fully glycosylated Recombinant Mammalian expressed G ectodomain (RMG) and intact RSV virion particles but minimal in vitro neutralization titers compared with the intact G ectodomain. Vaccinated mice were challenged intranasally with RSV-A2 Line 19F. Viral replication in nasal cavity and lungs was significantly reduced in vaccinated animals compared to unimmunized controls. Control of viral loads post-RSV challenge correlated with serum antibody binding to the virus particles. In addition, very low Th2/Th1 cytokine ratios were found in the lungs of REG ΔCCD vaccinated mice after challenge. These data demonstrate the presence of multiple protective sites in RSV G protein outside of the CCD that could contribute to the development of a bacterially produced unglycosylated G protein as safe and protective vaccine against RSV disease.
| A vaccine against RSV that provides protection without potential for disease enhancement is required. The G attachment protein represents an important candidate for inclusion in an effective RSV vaccine. However, the contribution of different antigenic sites to protection against RSV is not completely understood. We evaluated the protective efficacy of recombinant unglycosylated RSV-G protein vaccine produced in E. coli (REG) vs. CCD-deletion (REG ΔCCD). We also investigated immunogenicity and protective efficacy of all antigenic sites identified in post-primary infection infant sera using GFPDL that includes N- and C-terminal G subdomains, and linear peptides. The REG ΔCCD, N- and C-terminal subdomains and peptides generated antibody titers in rabbits and mice. Vaccinated mice challenged intranasally with RSV demonstrated significant reduction of viral replication in the nasal cavity and lungs. Our study highlights the safety and immunogenicity of recombinant G protein as economical protective vaccine against RSV disease.
| Respiratory syncytial virus (RSV) is the major cause of lower respiratory tract disease among infants and children globally [1] [2] [3]. Hospitalizations for respiratory tract disease among young children, especially in less than one year old, is most often attributed to RSV infection[4] [5]. Furthermore, despite the development of immunity following RSV infection during childhood, individuals remain susceptible to RSV upper respiratory tract reinfection life-long[6, 7] [8].
RSV isolates can be classified into two antigenically distinct groups (A and B) with genetic differences occurring mostly in the attachment glycoprotein G (47% heterogeneity at the amino acid level) and to a lesser degree in the fusion protein F (9%) [9]. In addition, continuous evolution of RSV generates diversity primarily in the G gene[10] [11]. Heterologous RSV strains are the main cause of re-infections, and homologous RSV strains are observed less frequently [12] [13]. However, while there are instances of evolution, e.g. the RSVs with duplications in the G gene, there are also cases of same genotype reappearing over many years. Even though F specific antibodies have been reported to contribute to majority of virus neutralization measure in vitro PRNT assays, the relative contribution of F and G specific antibodies to protection in vivo is not completely understood. A recent study by Capella et al found that higher concentrations of pre-F and G antibodies (but not post-F antibodies) were associated with lower clinical disease severity in infants and young children (< 2yr) [14].
In a previous study, we generated gene fragment phage display libraries (GFPDL) for the RSV surface proteins F and G to elucidate the complete antibody epitope repertoire in serum samples from infants either prior to (<9 months) or after primary and early RSV infection (15–18 months) [1]. Primary RSV infection predominantly resulted in an increase of G specific binding antibodies as observed by 100-fold increase in the number of phages that bound to the post-RSV infection sera compared with pre-infection sera. Bound phages displayed epitopes that spanned most of the ectodomain of RSV-G with two large conformationally dependent antigenic regions flanking the CCD motif in addition to the CCD [1].
In the current study, we evaluated the immunogenicity of the G-ectodomain lacking part of the central conserved domain (CCD) and the cysteine noose, as well as the individual G-subdomains (N- and C- termini), and G-derived peptides previously identified using GFPDL analysis of post-RSV infection infant sera by immunization of rabbits and mice using virus plaque reduction neutralization test (PRNT), several binding assays including recombinant mammalian cell produced G ectodomain RMG), RSV A2 virions, and recombinant CX3CR1 competition assay. The protective efficacy of these antigenic sites was determined in mouse challenge studies with RSV-A2 line 19F expressing firefly luciferase [rRSV-A2-L19-FFL]. Viral loads in the nasal cavity and lungs were inferred using fluorescence measurements obtained via whole body live imaging as previously described [15] in addition to plaque forming units (PFU) in the lungs. We found that animals vaccinated with REG ΔCCD, G-subdomains, and G-peptides had significantly lower viral loads after RSV challenge than unimmunized controls. Lung viral loads inversely correlated with RSV A2 virion binding antibody titers (but not with in vitro neutralization titers). Low lung pathology and low Th2/Th1 cytokine ratios was observed in all vaccinated-challenged groups. Therefore, several antigenic sites apart from the CCD motif in the G protein provide protective immunity against RSV.
Our earlier findings with post primary infection in plasma from infants, demonstrated very broad epitope repertoire spanning the entire G ectodomain. In the current study, we investigated the contributions of antigenic regions outside of the conserved central domain (CCD; aa residues 172–186) of G to RSV-specific immunity. As a comparator, we used a G-ectodomain protein (residues 67–298) of RSV-A2 containing the CCD motif produced using E. coli (REG 67–298), which was previously found to generate protective immunity in mice and cotton rats [16, 17]. We evaluated the immunogenicity of a CCD-deleted G-ectodomain [REG ΔCCD; with residues 172–186 replaced by a (Gly4Ser)2 linker], and two large G subdomains covering the N-terminus (REG 67–163) and C-terminus (REG 187–298) upstream and downstream of the CCD, respectively, that were identified as immunodominant in the epitope profiling of post-RSV primary infection human sera (Fig 1 panels A-B). Size exclusion chromatography (SEC) profiles of the four REG proteins using Superdex 200 gel filtration are illustrated in Fig 1 panels C-D. All recombinant G proteins ran as two distinct peaks likely representing tetramers and higher molecular weight oligomers.
Rabbits were immunized three times intramuscularly (i.m.) with 100 μg of purified REG protein combined with Emulsigen adjuvant, bled 8 days after each immunization, and sera evaluated in plaque reduction neutralization assay (PRNT) in the presence of guinea pig complement (GPC). GPC was added to increase the sensitivity of the neutralization assay, as previously we have shown that GPC specifically promotes RSV-A2 virus neutralization by anti-G but not anti-F specific antibodies in vitro [18]. As expected, the intact ectodomain (REG 67–298) was highly immunogenic, with peak neutralization titers measured after the second vaccination (Fig 1E). The REG ΔCCD and the two large fragments elicited modest or low neutralization titers after the second or third vaccinations, respectively (Fig 1E).
Since in vitro PRNT assays may not detect all antibodies, we also evaluated the binding of immune sera (post-third vaccination) to fully glycosylated G protein from RSV-A2 strain produced in mammalian 293 cells (RMG-A2) [16]. All four recombinant proteins generated antibodies in rabbits that bound to fully glycosylated G protein (Fig 1F). The REG 67–298 immune sera demonstrated the highest peak binding (2,500 RU), followed by REG ΔCCD immune sera (1,600 RU), and by the immune sera generated against the N-terminal and C-terminal domains. The hierarchy of binding to the glycosylated G recapitulated the PRNT virus neutralization titers for the same sera. These findings suggested that while the CCD region is key for generation of measurable neutralizing antibodies in the PRNT assay, there are immunological targets outside the CCD that elicit antibodies that can bind to the fully glycosylated G attachment protein.
Mice were immunized intramuscularly (i.m.) at day 0 and day 20 with 20 μg of purified REG proteins combined with Emulsigen adjuvant. Blood was collected from the tail vein on days 0, 14, and 30. On day 34, mice were inoculated intranasally (i.n.) with 106 PFU of RSV rA2-Line19F-FFL containing homologous RSV-A2 G protein sequence identical to the immunizing REG protein [15] (Fig 2A). Previously, we described the applicability of live imaging for following RSV replication and dissemination from the nasal cavity to the lungs, and demonstrated a strong correlation between bioluminescence flux units and viral loads measurements by PFU in the lungs of infected animals [15]. All animals vaccinated with REG 67–298 (intact G-ectodomain) completely controlled virus replication in the lung as measured by either live imaging (Fig 2B) or plaque assay (Fig 2C) compared to sham (PBS) vaccinated animals. Surprisingly, most the animals immunized twice with REG ΔCCD were also capable of controlling virus replication in the lungs by day 5 post-challenge (Fig 2B and 2C). Unexpectedly, some of the animals immunized with the N-terminus REG 67–163, or the C-terminus REG 187–298 also showed significant protection against viral replication in lungs as determined by either Flux or PFU measurements (Fig 2B and 2C). In addition to lungs, the live imaging allowed us to measure viral loads in the nasal cavities. All groups of vaccinated animals showed significant reduction in nasal viral loads compared with the PBS control group (Fig 2D).
In search for correlates of protection, we evaluated the immune sera (pre-challenge) in virus neutralization and virus particle–binding ELISA. Only sera from mice vaccinated with the entire G-ectodomain (REG 67–298) had detectable virus neutralization activity in vitro (Fig 2E). As such, the mechanism of protection was not apparent for the REG proteins without CCD and for the N- and C- domains. Several anti-G specific MAbs including 131-2G have been defined that do not neutralize in classical A549 based virus neutralization assay however provide significant protection in animal challenge studies [19, 20]. Therefore, we evaluated total antibody binding to RSV particles by ELISA. We have confirmed in earlier studies that the virus ELISA express relevant protective epitopes (including conformational epitopes) and can capture monoclonal and polyclonal antibodies with protective titers. However, it is possible that coating of virus particles (virions) on microtiter plates in ELISA may not accurately represent some aspects of structurally intact virions. Antibody binding to RSV A2 virions was highest in immune sera from animals vaccinated with the entire RSV-G ectodomain (REG 67–298) (Fig 2F). However, some binding to virus particles was also observed with sera from REG ΔCCD, and to a lesser degree with sera from animals vaccinated with the C-terminus domain (REG 187–298) and even weaker binding with sera from the N-terminal domain immunized mice (REG 67–163). The binding antibody endpoint titers to RSV A2 virions are shown in Fig 2H. We also measured binding to RSV B1 virion particles. As expected, sera from REG 67–298 vaccinated animals bound RSV B1 virions in ELISA although to a lower level compared to RSV-A2 virions, in agreement with our earlier study [16] (Fig 2G). In contrast, the sera from the other three groups did not show significant binding to RSV B1 virions.
On day 5 post-challenge, the viral loads (as measured by plaque assay) in the lungs (Fig 2C) and nasal cavity (measured by live imaging) (Fig 2D) inversely correlated with virion binding antibody endpoint titers (Fig 2I and 2J) (r = -0.5622 and r = -0.4043, respectively) that reached statistical significance for the lungs (p < 0.0152). The results show that in vivo challenge studies are more sensitive than in vitro assays for detection of protective activity of RSV G targeting immune mechanisms.
In addition to viral loads, lung pathology (bronchiolitis, perivasculitis, interstitial pneumonia, and alveolitis) was evaluated for all mice on day 5 post-RSV challenge (Fig 3A). The histology scores for vaccinated animals were not significantly different from the sham vaccinated (PBS) control group. To further explore the possibility of perturbation to the local cytokine milieu, the lung extracts were tested against a cytokines/chemokines multiplex panel (Fig 3B–3I). Levels of most cytokines/chemokines were similar between the vaccinated animals and were not significantly different from the sham vaccinated (PBS) control group following RSV challenge. However, a non-significant elevation of Eotaxin in the vaccinated groups was observed compared with the control group (Fig 3G), and an elevated MIP-1α levels in the group vaccinated with the C-terminus antigenic domain (REG 187–298) (Fig 3I). But there were no significant differences in the levels of Th1 and Th2 cytokines and the Th2/Th1 ratio (Fig3J) [16].
Next we synthesized peptides representing linear antigenic sites that were identified in human post-primary RSV infection in infants [1]. These peptides were conjugated to KLH and mixed with Emulsigen for vaccination of rabbits and mice (Fig 4A).
The rabbit polyclonal sera obtained after two peptide vaccine doses bound to glycosylated RMG protein in SPR. However, antibody binding titers were 3 to 10-fold lower than antibody binding observed following immunization with REG ΔCCD (Fig 4B vs. Fig 1F). Interestingly, the peptide that elicited the highest RMG-binding titer was found in rabbits immunized with the N-terminal G peptide (aa 66–90). However, only peptide G169-207 that overlapped the CCD region elicited neutralizing antibodies after three immunizations (Fig 4C).
Murine immunization with KLH-conjugated G peptides (Fig 5A) elicited antibodies that bound RSV A2 virions at various levels, but did not reach the binding titers elicited by the REG 67–298 ectodomain (Fig 5B and 5C)[21]. Interestingly, following challenge with RSV rA2-Line19F-FFL, the majority of immunized animals showed reduced viral loads in the lungs on day 5 compared with sham-vaccinated animals as measured by either Fluxes (Fig 5D) or PFU (Fig 5E). The reduction in lung viral loads measured using live imaging ranged between 40–78% of the sham (PBS vaccinated) control group, and reached statistical significance for most groups of immunized animals, apart from animals immunized with CCD peptide G 169–207 or the G 236–263 peptide (Fig 5D), suggesting less consistent control of virus replication after challenge in these two peptides vaccinated groups. However, in the PFU assay, all vaccinated groups had statistically lower viral loads compared with the control group (Fig 5E). The viral loads in the nasal cavity were similar between the control and vaccinated groups, but two animals in the group vaccinated with G 169–207 had elevated fluxes (Fig 5F). These were the same animals that also had elevated lung fluxes (Fig 5D), suggesting less consistent control of virus replication after challenge in this specific group. Correlation of the binding antibody ELISA end-point titers to RSV A2 virions with the viral load measurements in the lungs and nasal cavities (Fig 5G–5I) demonstrated significant inverse correlation only between ELISA endpoint titers and the lung viral loads measured in the plaque assay (Fig 5H, r = -0.4835, p = 0.0053).
In the same study, lung sections were evaluated for histopathology (Fig 6A) at day 5 post-viral challenge. The only notable finding was an increased bronchiolitis score in animals following RSV challenge that were vaccinated with G peptide 169–207, and to a lesser degree in some of the animals vaccinated with G peptide 66–90. Cytokine/chemokine profiling (Fig 6B–6I) revealed significantly higher levels of eotaxin in the lung extracts from the same two (G 66–90 and G 169–207) groups (Fig 6G). In addition, the Th2/Th1 cytokine ratio was 1 for the G 66–90 peptide vaccinated group, but well below 0.5 for all other vaccinated groups following RSV challenge (Fig 6J).
Since the in vitro neutralization assay does not reflect infection of human broncho-epithelial cells, which involves interaction between the RSV G protein and surface CX3CR1 [22–25], an SPR-based assay was performed to measure direct antibody-mediated blocking of recombinant CX3CR1 binding to glycosylated G protein produced in 293T cells (RMG) (Fig 7A).To that end, rabbit pre-vaccination and post 2nd boost sera (at 10-fold dilution) were run on the chip captured with RMG prior to addition of rCXC3R1, and % inhibition was calculated for each serum sample. In this real-time binding assay, the strongest inhibition (90%) of CX3CR1-RMG interaction was observed with antibodies against complete REG ectodomain (67–298) followed by G 169–207 (85%), both of which contain the CX3C motif required for CX3CR1 binding[22]. Importantly, polyclonal antibodies elicited by the REG ΔCCD (aa 172–186 deleted), inhibited CX3CR1-RMG binding by 70%. In addition, short peptide (G 148–178) and the C-terminal antigenic domain (187–298) inhibited CX3CR1 binding by 48% and 55%, respectively (Fig 7A).
The degrees of conservation of the individual peptides used in our study among RSV A and RSV B strains are depicted in S1 Fig. and in S1 Table. The percentage homology was high among all RSV A strains but dropped significantly for RSV B strains. However, RSV A2 G peptides aa148-178 and aa169-207 showed ~70% conservation with RSV B1 and circulating B strains (S1 Table). Therefore, a G based protective vaccine may require a combination of G proteins from diverse RSV strains to protect against all RSV strains.
In summary, our study demonstrated that in addition to the central conserved domain (CCD) (aa 164–176), there are several protective antigenic sites within RSV-G, including the G- derived N- and C- sub-domains and linear peptides that were recognized by post-RSV infection human sera as previously reported [1]. Vaccination with these recombinant proteins and peptides provided at least partial protection in mice as measured by reduced lung viral loads on day 5 post challenge with RSV-A2 virus, with no significant lung pathology (Fig 7B).
The contribution of immune response against the RSV attachment protein G to either protection or potential enhanced disease have been well documented [8]. The G protein undergoes constant diversification in circulating RSV strains, and may contribute to the ability of the virus to reinfect throughout life [9–11]. Several monoclonal antibodies targeting the G protein were demonstrated to have protective activity against severe disease in animal models as well as anti-inflammatory effects [19, 26–29]. These MAbs bound to either linear or conformational epitopes overlapping and surrounding the CCD/cysteine noose CX3CR1 binding regions. Furthermore, levels of anti-G antibodies, in addition to antibodies against pre-F, were associated inversely with disease severity in RSV-infected infants and young children (<2yr) [14]. Therefore, dissecting the immune response to the G protein is important for better understanding of RSV viral immunity and the design of RSV G based vaccines [30].
In an earlier study, we used F- and G- phage display libraries (RSV-GFPDL) to dissect the antibody repertoire prior to and following primary RSV infection in infants. This analysis identified a large number of epitopes spanning the entire G ectodomain [1]. In this current study, we evaluated the ability of these linear and conformational antigenic sites within the RSV G attachment protein for their immunogenicity and ability to elicit protective immunity against RSV.
To address the contribution of G-CCD for protection against RSV, we compared full length recombinant RSV-G protein with and without the central conserved domain (CCD) (aa 172–186) in rabbits and mice. While virus-neutralizing antibody response to the CCD-deleted protein was significantly lower compared with the intact ectodomain, both proteins significantly protected mice from RSV challenge as measured by reduced viral loads in the lungs and nasal cavity. Lung viral loads were inversely correlated with ELISA titers of binding antibodies to RSV A2 virion particles. Furthermore, a low pathology scores and very low Th2/Th1 cytokine ratio was measured in the lungs of vaccinated animals on day 5 post RSV challenge, not significantly different from the sham–vaccinated control animals (Fig 3), suggesting avoidance of the severe lung pathology caused by FI-RSV vaccination [31].
It was previously shown that the CX3C motif downstream of CCD interacts with the CX3CR1 that serves as a receptor for RSV G protein on primary human airway epithelial cells [23]. Several MAbs targeting the CCD motif as well as MAbs 131-2G that binds to a motif upstream of the CCD were shown to block this interaction and neutralize RSV infection of airway broncho-epithelial cells, but not in traditional plaque neutralization assays [20, 23, 25, 26] and to protect mice from RSV challenge [19, 27]. In agreement, the antibodies generated against REG ΔCCD [with CCD residues 172–186 replaced by a (Gly4Ser)2 linker] in the current study, did not demonstrate significant virus neutralization in vitro, but provided significant in vivo protection.
We also investigated the immunogenicity of N- and C- subdomains of RSV-G and linear G antigenic sites that were identified by GFPDL screening of post- RSV exposure infant sera [1]. We found that the N- and C-subdomains (aa 67–163 and aa187-298) flanking the CCD motif generated antibodies that bound to fully glycosylated recombinant G protein produced in mammalian cells (RMG) using SPR, as well as to RSV virions in ELISA, but did not neutralize the virus in a PRNT assay. However, after RSV intranasal challenge of vaccinated mice, partial protection (i.e. reduction in lung and nasal virus loads on day 5) was observed (Fig 2). Furthermore, synthetic G-peptides containing mostly non-conformational antigenic sites, upstream and downstream of the CCD, also generated antibodies that bound the RMG [16] as well as intact RSV virions in ELISA (Figs 2 and 5). Since RMG protein and the G protein expressed on RSV virion particles (propagated in mammalian cells) are highly glycosylated, the antigenic sites outside the CCD are expected to be shielded from antibody recognition. Yet RSV-infected children clearly generated antibodies covering both conserved and less conserved sites in the G protein irrespective of predicted glycosylation levels, that could bind to the fully glycosylated G protein on virus particles as shown in our previous study[1] and was confirmed in the current study. Importantly, most animals in the challenge studies demonstrated mild lung pathology, and low Th2/Th1 cytokine ratios in lungs post-RSV challenge (Figs 3 and 6)
In a newly developed SPR-based real-time kinetics assay, we demonstrated inhibition of interaction between recombinant CXC3R1 protein and glycosylated RSV-G protein with sera from rabbits vaccinated with the G ectodomains, sub-domains, and linear peptides (Fig 7). The percentage of inhibition ranged between high (70–90%) for REG (67–298) REG ΔCCD, and G 169–207 peptide, medium (48–50%) for REG C-terminus domain (187–298) and G 148–178 peptide, or low (<30%) for all other peptides.
Surprisingly, G peptide (aa 169–207) spanning the cysteine noose and the conserved CCD motif, which was the immunodominant region recognized by post-RSV infection plasma from humans across multiple age groups [1], was not very protective in the current study (only 40% protection with large intragroup variability) (Fig 5D and 5F and Fig 7B). Interestingly, animals vaccinated with this peptide showed an increase in perivasculitis score and higher levels of eotaxin in lung extracts on day 5 post viral challenge (Fig 6A and 6G). It was earlier reported that atypical eosinophilia in RSV infected BALB/c mice was triggered by vaccination with G peptide (aa 184–198) in a CD4—dependent mechanism[32]. Another study using recombinant vaccinia virus (rVV) expressing G inserts identified residues 193–205 to be responsible for G-induced weight loss and lung eosinophilia in mice [33]. Both sequences are included in our G (169–207) peptide that was recognized by convalescent sera from infected children. Therefore, it is possible that this immunogen has the potential to induce both protective antibodies inhibiting interaction of RSV-G with the CX3CR1 and to promote proinflammatory environment in the form of high eotaxin secretion and eosinophilia. The role of CD4 T cells will need further investigation and may vary between mice and humans.
Taken together, these data suggest that in addition to the highly conserved CCD region, other antigenic sites in the G protein may contribute to protection against RSV in animal models and possibly humans[21]. The mechanisms of protection mediated by G antibodies needs to be further investigated. The PRNT assay in A549 cells is clearly not optimal for detection of all protective anti-RSV G antibodies. It is likely that antibodies elicited by some of the G domains and linear peptides would block RSV infection in vivo by directly blocking the G protein-CX3CR1 receptor interaction on lung broncho-epithelial cells. In addition, antibodies to other regions of RSV-G could mediate protection by other effector mechanisms including antibody dependent cellular cytotoxicity (ADCC) and antibody-dependent cellular phagocytosis (ADCP) that could contribute to removal of RSV infected cells in vivo [20]. We also cannot exclude contribution of cell mediated immunity by RSV-G specific CD8+ T cells. However G protein lacks MHC class I-restricted epitopes and has not been shown to elicit CTL responses in mice or humans [34, 35]. It was reassuring that the REG ΔCCD construct elicited a favorable ratio of Th1/Th2 cytokines, different from the formalin inactivated killed RSV vaccine that skewed the immune system towards Th2 responses and showed enhanced disease in naïve humans and animals after virus exposure.
In earlier studies, Powers et al. described the immunogenicity of bacterially produced fusion protein, BBG2Na that contained the central conserved domain of RSV-A2 G (aa 130–230) (G2Na) fused to the albumin-binding domain of streptococcal protein G (BB) formulated with aluminum adjuvant [36, 37]. Three of our peptides overlap with the G sequence in BBG2Na and our data are in general agreement with these studies. However, our REG immunogen encompass the entire G ectodomain with no “foreign/fusion” sequence that may re-direct the immune responses following vaccination.
In summary, the current study evaluated the immunogenicity of multiple antigenic sites within the RSV G protein and shows for the first time the presence of in vivo protective epitopes outside the CCD conserved motif. This information can help to explain findings in RSV exposed individuals and could contribute to further development and evaluation of safe and effective G protein based vaccine against RSV.
A549 cells (Cat. No. #CCL-185) were obtained from the American Type Culture Collection (ATCC, Manassas, VA, USA) and were maintained in F-12K medium supplemented with 10% fetal bovine serum, 1X penicillin streptomycin (P-S), and L-glutamine. Cells were maintained in an incubator at 37°C under 5% CO2. RSV rA2-Line19F-Firefly Luciferase (rRSV-A2-L19-FFL) expressing the firefly luciferase gene upstream of the NS1 gene was prepared by infecting sub confluent A549 cell monolayers in F-12K medium supplemented with 2% FBS and 1X Penicillin-Streptomycin (infection medium) [15]. To generate a challenge stock, at 5 days post infection (dpi), cells were freeze-thawed twice and virus was collected. Harvested viruses were cleared of cell debris by centrifugation at 3,795g for 15 min. Virus stocks used in challenge studies were pelleted by centrifugation at 10,509g overnight. Pelleted virus was resuspended in TEN buffer and purified by sucrose-gradient ultracentrifugation. Virus titers were determined by plaque assay on A549 cells. The optimal challenge dose (106 PFU intranasally) was determined in earlier study in which viral loads were measured by traditional plaque assay, by qRT-PCR and by live imaging (flux) and gave comparable results in terms of viral kinetics and peak values [15]. Codon-optimized RSV G coding DNA for E. coli was chemically synthesized. NotI and PacI restriction sites were used for cloning the RSV A2 G ectodomain coding sequence (amino acids 67 to 298) into the T7-based pSK expression vector for bacterial expression. DNA coding REG 67–163 and REG 187–298 were amplified by PCR using primers containing NotI and PacI restriction sites. DNA coding REG ΔCCD with residues 172–186 deleted and replaced with a (G4S)2 linker was prepared by a two-step overlapping PCR. The deleted sequence contains the cysteine noose in addition to the CX3CR1 binding motif present in all RSV G proteins. All amplified DNA was digested with NotI and PacI and ligated into the T7-based pSK expression vector for bacterial expression.
Recombinant RSV G 67–298 (REG 67–298), REG 67–163, REG 187–298, and REG ΔCCD proteins were expressed in E. coli BL21(DE3) cells (Novagen) and were purified as described previously [16, 17]. Briefly, REG proteins expressed and localized in E. coli inclusion bodies (IB) were isolated by cell lysis, denatured and renatured in redox folding buffer followed by dialysis. The dialysate was purified through a HisTrap FF chromatography column (GE Healthcare). The protein concentrations were analyzed by bicinchoninic acid (BCA) assay (Pierce), and the purity of the recombinant G proteins from E. coli (REG) was determined by SDS-PAGE. Linear peptides were synthesized chemically using Fmoc chemistry, purified by HPLC, conjugated to KLH, and dialyzed.
The 293-Flp-In cell line (Cat. No. #R75007; ThermoFisher Scientific) stably expressing the RSV A2 G protein with secretory signal peptide from IgG kappa chain was developed as described previously (15). Briefly, 293-Flp-in cells were co-transfected with the plasmids expressing Flp-in recombinase and the RSV A2 G ectodomain in DMEM media (Invitrogen). Twenty-four hours after transfection, culture medium was replaced with fresh DMEM containing 150 μg/mL of hygromycin for selection of stably transfected cells. For protein expression, cells were maintained in 293-Expression media (Invitrogen), and culture supernatant was collected every 3–4 days. Supernatant was cleared by centrifugation and filtered through a 0.45 μm filter before purification through a His-Trap FF column (GE healthcare).
Proteins at a concentration of 2 mg/ml were analyzed on a Superdex 200 Increase 10/300 GL column (GE Healthcare) pre-equilibrated with phosphate-buffered saline (PBS), and protein elution was monitored at 280 nm. Protein molecular weight (MW) marker standards (GE Healthcare) were used for column calibration and for the generation of standard curves to identify the molecular weights of each purified protein.
For the plaque reduction neutralization test (PRNT), heat-inactivated serum was diluted 4-fold and incubated with RSV-A2 virus (diluted to yield 20–50 plaques/well) containing 10% guinea pig complement (Rockland Immunochemical; Philadelphia, PA, USA) and incubated for 1 h at 37°C. After incubation, 100 μl of the antibody-virus mixtures were inoculated in duplicate onto A549 monolayers in 48-well plates and incubated for 1 h at 37°C. Inoculum was removed prior to adding infection medium containing 0.8% methylcellulose. Plates were incubated for 5 to 7 days at which time the overlay medium was removed and cell monolayers fixed with 100% methanol; plaques were detected by immunostaining with rabbit RSV polyclonal anti-F sera (14), followed by addition of alkaline phosphatase goat anti-rabbit IgG (H+L) (Jackson) antibody. The reactions were developed by using Vector Black Alkaline Phosphatase (AP) substrate kit (Vector Labs, Burlingame, CA). Numbers of plaques were counted per well and the neutralization titers were calculated by adding a trend line to the neutralization curves and using the following formula to calculate 50% endpoints: antilog of [(50+y-intercept)/slope].
All animal experiments were approved by the U.S. FDA Institutional Animal Care and Use Committee (IACUC) under Protocol #2009–20 (mice) and #2008–10 (rabbit). The animal care and use protocol meets National Institutes of Health (NIH) guidelines.
Female New Zealand white rabbits (KBL(NZW)BR strain from Charles River Labs) were immunized three times intramuscularly (i.m.) with 100 μg of each purified REG protein combined with Emulsigen adjuvant, or with KLH-conjugated RSV-G peptides combined with Emulsigen, every 28 days. Blood was collected 8 days after each immunization.
Four- to 6-week-old female BALB/c mice (BALB/cAnNCr strain code #555) were obtained from the Charles River Labs. Mice [N = 6–8 per group] were immunized intramuscularly (i.m.) at day 0 and day 20 with 20 μg of purified REG protein or with 25 μg of KLH-conjugated peptides combined with Emulsigen adjuvant at total volume of 100 μl. Blood was collected from the tail vein on days 0, 14, and 30. On day 34, mice were anesthetized with isofluorane through inhalation according to mouse body weight and infected intranasally (i.n.) with 106 PFU of rRSV-A2-L19-FFL as previously described [15]. Mice were sacrificed by CO2 asphyxiation 5 days post-RSV challenge (previously determined to be the day with peak viral load), and blood and lungs were collected. For determination of the viral load and cytokine analysis, the left lobe of the lung was collected.
Whole body live imaging of infected mice was performed using IVIS imaging system as previously described [15]. In brief, mice were anesthetized in an oxygen-rich induction chamber with 2% isoflurane and administered 20 μl of RediJect D-Luciferin bioluminescent substrate (Perkin Elmer) intranasally. After a 5-min interval, mice were placed in the IVIS 200 Imaging systems (Xenocorp) equipped with the Living Image software (version 4.3.1.). Bioluminescence signals were recorded for 2 min for whole body and for 1 min for lungs and nasal cavities, respectively. Images were analyzed with the LivingImage 4.5 software (PerkinElmer) according to manufacturer’s instructions.
Lungs (unperfused) were weighed and homogenized in F-12K-2% FBS-1X P-S (5 ml medium/g of lung) using an Omni (Kennesaw, GA) tissue homogenizer. The supernatant was cleared by centrifugation at 3,795 xg for 10 min and was used immediately for viral titration by plaque assay in A549 cells as described above.
All lungs were weighed and homogenized in 5 ml of medium/g of lung, as described above, to normalize the amount of lung tissue used per sample. Homogenized lungs were further diluted in infection culture medium containing a 2X concentration of Complete EDTA Free protease inhibitor cocktail (Roche, Basel, Switzerland) and were used in a Bio-Plex Pro mouse cytokine 23-plex assay according to the manufacturer’s recommendations. Plates were read using a Bio-Plex 200 system (Bio-Rad, Hercules, CA).
Immulon 2 HB 96-well microtiter plates were coated with 100 μl of purified RSV rA2-Line19F-FFL or RSV B1 virus in PBS (104 pfu/well) per well at 4°C overnight. After blocking with PBST containing 2% BSA, serial dilutions of mouse serum in blocking solution were added to each well, incubated for 1h at RT, followed by addition of 2,000-fold dilution of HRP-conjugated goat anti-mouse IgG-Fc specific antibody, and developed by 100 μl of OPD substrate solution. Absorbance was measured at 490 nm.
Steady-state equilibrium binding of post-vaccination animal sera was monitored at 25°C using a ProteOn surface plasmon resonance (SPR) biosensor (Bio-Rad). The recombinant G protein from 293T cells (RMG) was coupled to a GLC sensor chip via amine coupling with 500 resonance units (RU) in the test flow channels. Samples of 100 μl of freshly prepared sera at a 10-fold dilution or MAbs (starting at 1 μg/ml) were injected at a flow rate of 50 μl/min (contact duration, 120 seconds) for association. Disassociation was performed over a 600 seconds interval. Responses from the protein surface were corrected for the response from a mock surface and for responses from a buffer-only injection. Pre-vaccination animal sera were used as a negative control. Total antibody binding and data analysis results were calculated with Bio-Rad ProteOn Manager software (version 3.0.1).
The recombinant G protein from 293T cells (RMG) was captured on a HTG sensor chip via Histidine tag with 500 resonance units (RU) in the test flow channels. Samples of 500 μl of freshly prepared post-2nd immunization rabbit sera at a 10-fold dilution were injected at a flow rate of 50 μl/min (contact duration, 300 seconds) for association. Following antibody binding, recombinant CX3CR1 (Abnova; 5 μg/mL) was injected at a flow rate of 50 μl/min (contact duration, 120 seconds) for association. Responses from the protein surface were corrected for the response from a mock surface and for responses from a buffer-only injection. Pre-vaccination animal sera were used as a negative control. Total CX3CR1 binding and data analysis and % inhibition by immune sera were calculated with Bio-Rad ProteOn Manager software (version 3.0.1).
The statistical significances of group differences were determined using one-way analysis of variance (ANOVA) and a Bonferroni multiple-comparison test. Correlations were calculated with a Spearman two-tailed test. P values less than 0.05 were considered significant with a 95% confidence interval.
|
10.1371/journal.pntd.0005841 | Predictive factors for a one-year improvement in nontuberculous mycobacterial pulmonary disease: An 11-year retrospective and multicenter study | Nontuberculous mycobacterial pulmonary disease (NTM-PD) has become an emerging infectious disease and is responsible for more deaths than tuberculosis in industrialized countries. NTM-PD mortality remains high in some series reportedly ranging from 25% to 40% at five years and often due to unfavorable evolution of NTM-PD despite established treatment. The purpose of our study was to search for early factors that could predict the favorable or unfavorable evolution of NTM-PD at the first year of treatment.
In this retrospective and multicenter study, we selected 119 patients based on clinical, radiological and microbiological data from 2002 to 2012 from three French university hospitals (Guadeloupe, Martinique, Montpellier) with definite (meeting the criteria of the American Thoracic Society and the Infectious Disease Society of America in 2007; ATS/IDSA) or probable (one positive sputum culture) NTM-PD. We compared two patient groups: those who improved at one year (clinical symptoms, radiological lesions and microbiology data) and those who did not improve at one year. The data were analyzed for all patients as well as for subgroups by gender, HIV-positive patients, and Mycobacterium avium complex (MAC) infection.
The average patient age was 50 years ± 19.4; 58% had respiratory comorbidities, 24% were HIV positive and 19% had cystic fibrosis. Coughing concerned 66% of patients and bronchiectasis concerned 45%. The most frequently isolated NTM were MAC (46%). 57% (n = 68) of patients met the ATS criteria and improved status concerned 38.6% (n = 46). The improvement factors at one year of NTM-PD were associated with the duration of ethambutol treatment: (Odds ratio adjusted [ORa]: 2.24, 95% Confidence interval [CI]; 2.11–3.41), HIV-positive status: (ORa: 3.23, 95% CI; 1.27–8.45), and male gender: (ORa: 2.34, 95% CI; 1.26–8.16). For the group with NTM-PD due to MAC, improvement was associated with the duration of macrolide treatment (ORa: 3.27, 95% CI; 1.88–7.30) and an age <50 years (ORa: 1.88, 95% CI; 1.55–8.50).
In this retrospective multicenter study, improvement at one year in patients with definite or probable NTM-PD was associated with the duration of ethambutol treatment, HIV-positive status and male gender. For the group of patients infected with MAC, improvement was associated with the duration of macrolide treatment and an age <50 years. Identifying predictors of improvement at one year of NTM-PD is expected to optimize the management of the disease in its early stages.
| Early predictive factors for a favorable development of nontuberculous mycobacterial pulmonary disease (NTM-PD) are important to improve management due to the high mortality of this infection at 5 years. The purpose of this study was to search for early factors that could predict at the first year, the favorable or unfavorable evolution of NTM-PD. This multicenter and retrospective study shows the importance of the duration of use of certain antibiotics (e.g. ethambutol and macrolides) in combination with other drugs in the one-year improvement of patients with NTM-PD. It also confirms the favorable prognosis at one year of NTM-PD patients with HIV-positive status. Identifying predictors of improvement at one year of NTM-PD is expected to optimize prognosis of the disease in its early stages.
| Infection with nontuberculous mycobacteria (NTM) preferentially affects the lungs and occurs by inhalation of aerosols containing mycobacteria [1, 2]. NTM are ubiquitous environmental bacteria found in soil, but also in other sources such as contaminated water taps. The frequency of NTM species can vary from region to region in the world [1, 3]. NTM pulmonary disease (NTM-PD) has today become an emerging infectious disease in industrialized countries. Its increasing prevalence is estimated at more than 50 cases per 100,000 persons in some demographic groups in the US [4]; while its incidence in Europe ranges from 0.2 to 2.9 / 100,000 inhabitants [1]. Remarkably, all NTM species are not likely to cause NTM-PD; only a few species such as Mycobacterium avium complex (MAC), M. abscessus, M. xenopi and M. kansasii are frequently involved [5]. Indeed, the clinical relevance of NTM differs by species since they are not endowed with the same virulence [6].
The diagnostic criteria of the American Thoracic Society and the Infectious Disease Society of America in 2007 (ATS/IDSA) [7] have established the diagnosis of NTM-PD based on clinical symptoms, radiological lesions and microbiology data. During this decade, real progress has been made in the understanding of this disease [4]. We know for example that besides immunosuppression by HIV or cystic fibrosis, NTM-PD occurs in lungs whose architecture is already weakened by chronic respiratory diseases such as primarily chronic obstructive pulmonary disease (COPD) and bronchiectasis [1, 5]. The establishment of NTM-PD in impaired lungs can cause the destruction of the pulmonary parenchyma [8] and eventually lead to death due to the evolution of NTM-PD [9]. Patients with NTM-PD are not all treated because current treatments are often long, expensive and not without side effects [10]. NTM-PD mortality remains high in some series ranging from 25% to 40% at five years [1, 9, 11]. The main factors of poor outcomes identified in mortality studies at five years corresponded to an advanced age, the existence of respiratory comorbidities, radiological cavity lesions, and some mycobacteria such as M. xenopi [11,12,13].
Given the deteriorating respiratory status of patients due to the evolution of NTM-PD despite established treatment and the relatively high mortality at five years, it seemed important to search for early factors that could predict from the first year the favorable or unfavorable evolution of NTM-PD, and thus improve prognosis. Hence, the main purpose of this study was to identify factors that contribute to the clinical, radiological and microbiological improvement at one year of a cohort of 119 patients with definite (meeting the criteria ATS/IDSA) or probable (one positive sputum culture) NTM-PD, regardless of their immune status or their respiratory history. The secondary goal of this study was to report for the first time, a clinical, radiological and microbiological description of NTM-PD in a population of Afro-Caribbean patients in the French West-Indies.
This observational study received approval from the Institutional Review Board of the French learned society for respiratory medicine (Société de Pneumologie de Langue Française; No: 2015–003). All the participants gave their written consent. The parents/guardians provided written informed consent on behalf of participants below 18 years of age.
This study was carried out in accordance with the principles of the Helsinki Declaration.
This study was a retrospective, multicentric, observational study over a 11-year period between 2002 and 2012 in three French university hospitals (CHU), two of which are located in the French West-Indies (University Hospital of Fort de France, Martinique; and University Hospital of Pointe-à-Pitre, Guadeloupe), and the 3rd in Metropolitan France (University Hospital of Montpellier, France). From the computerized databases of the bacteriological laboratories of these three institutions, we searched all patients over 13 years old with at least one positive culture for NTM between 2002 and 2012. A total of 119 patients were therefore finally retained for this study regardless of their immune status. The exclusion criteria were an age below 13 years and the absence of patient consent.
This was a composite endpoint defined by the disappearance at one year of respiratory symptoms and/or initial symptoms, regression or normalization at one year of the initial radiological lesions, and negative bacteriological cultures at one year. Negative bacteriological cultures were defined as at least three consecutive negative respiratory culture specimens at the end of one year. Patients were classified as having an improved status at one year only if all the three criteria were met (vs. unimproved status if this was not the case).
Statistical analyses were designed to determine the parameters related to the primary endpoint, i.e., an improved status at one year. Univariate analysis was first conducted to study the independent variables related to the primary endpoint. Statistical tests used for categorical variables were the Chi-squared test or the Fisher exact test and for quantitative variables, the Student’s t-test or the Wilcoxon-Mann-Whitney test. For all statistical tests, the significance level was set at 5% and a power >90%.
Independent variables with a p-value less than 0.2 determined by univariate analysis were retained for the multivariate model. Multivariate analysis consisted of logistic regression analysis. The dependent variable was the binary variable (improved / unimproved status); independent variables were introduced into the model using a backward regression approach. Variables with a p-value less than 0.05 were selected. The results were produced as odds ratios with 95% confidence intervals.
The choice of multivariate logistic regression was dictated:
A subgroup analysis was performed for the population infected by MAC for the HIV-positive population and by gender.
Processing and statistical analysis were performed using version 3.3.2 of the R software. The libraries used in the statistical analysis with R included: base-package, stats-package, BioStatR-package, MASS-package and pwr-package.
Patients with an improved status represented 38.6% (n = 46). The regression of clinical symptoms at one year concerned 56.3% (67/119), the disappearance or regression of radiological lesions at one year concerned 38.6% (46/119) of patients and negative bacteriological cultures at one year were obtained for 51.2% (61/119). A statistically significant difference was revealed between the two groups for age (p<0.05), place of residence (p<0.01) and the percentage of patients with HIV-positive serology (p<0.02).
No difference was found between the two groups for the ATS/IDSA diagnostic criteria (58.6% vs. 56.1%, p = 0.93).
A statistically significant difference was found between the two groups (improved / unimproved status) in the circumstances of the disease discovery (p<0.006). There was no statistically significant difference between the two groups in terms of initial respiratory symptoms and initial radiological lesions.
In Guadeloupe, the main NTM encountered in decreasing order were MAC, M. simiae and M. fortuitum, in Martinique, M. fortuitum followed by MAC, then M. gordonae; and in Montpellier, MAC then M. abscessus complex, followed by M. xenopi. No statistically significant difference was found between the improved / unimproved status groups for the mycobacterial species. The ATS/IDSA criteria were met for 62% of patients with MAC, 82% with M. abscessus, 50% with M. fortuitum and 45% with M. simiae. For bacteriological samples, 76% met the ATS/IDSA microbiological criteria. There was no statistically significant difference between the two groups for the ATS/IDSA microbiological criteria. The positive predictive value (PPV) of the ATS/IDSA microbiological criteria for definite NTM-PD was 89% (68/76) CI 95% (83%-94%). Lastly, patients who did not meet the ATS microbiological had a four-fold increased risk of death at one year (OR = 4.01, 95% CI; 1.40–14.51, p<0.01).
No statistically significant difference was found between the two groups for treated patients, as well as in the total duration of treatment. There was a statistically significant difference in the duration of ethambutol treatment between the two groups (p<0.001, effect size: 0.81, power: 0.99). Side effects related to treatment concerned 10 of 63 patients (15.8%), five had minor side effects (digestive disorders) and five had major side effects (three cases of drug-induced hepatitis, one case of eye damage and a kidney failure). These 5 patients with major side effects had to stop their therapy. No patient in our cohort benefited from associated surgical treatment.
There was a statistically significant association between the absence of negative cultures and mortality at one year (p<0.001). The conversion rate of bacterial cultures was 60% (33/55) for MAC, 35% (6/17) for M. abscessus complex, 37% (6/16) for M. fortuitum and 72% (8/11) for M. simiae.
The total number of mortalities at one year was 14.2% (n = 17), all belonging to the unimproved group. The average age of deceased patients (13 men and four women) was 60 years±12.7. We recorded 52% tobacco smokers, and 44% COPD, 29% HIV-positive and 5% cystic fibrosis patients. NTM of deceased patients were MAC (9/17; 52.9%), M. abscessus complex (4/17; 23.5%), M. kansasii (2/5; 40%) and M. fortuitum (2/16; 12.5%). Eight patients died of unfavorable NTM-PD evolution, one patient from pulmonary embolism and two patients from COPD exacerbations. An association between the mortality rate and mycobacterial species in the study (p = 0.86) was not found.
Factors associated with an improvement at one year were the male gender (OR = 2.34), HIV-positive serology (OR = 3.23) and duration of ethambutol treatment (OR = 2.24). For the population meeting the ATS/IDSA diagnostic criteria, the factor associated with improvement was the duration of ethambutol treatment (OR = 2.45). For the group of patients infected by MAC, improvement factors were associated with age under 50 years (OR = 1.88) and duration of macrolide treatment (OR = 3.27).
For the group of non HIV-positive patients, improvement factors were associated with Male (OR = 3.54) and duration of ethambutol treatment (OR = 1.90).
A total of 29 patients were included. The CD4 count was below 200 for 23 patients (79%). The discovery of NTM revealed an HIV-positive status for 98% of patients. The most frequently found species was MAC (58%). A statistically significant difference between the two groups (improved / unimproved status) was found for age (p<0.04), percentage of treated patients (p<0.04), negative bacteriological cultures at one year (p<0.001), and percentage of deaths at one year (p<0.04).
Sixteen of the 29 HIV–positive patients (55%) were treated. The percentage of patients improved on treatment was 81% (13/16).
A statistically significant difference was found between men and women for age (p<0.04), chronic respiratory diseases (cystic fibrosis, COPD and bronchiectasis; p<0.004), the incidence of sputum for women (p< 0.02), the type of radiological lesions and improved status at one year (p<004) in favor of men.
In this work, HIV infection, treatment duration with ethambutol in combination with other antibiotic molecules, and the male gender were independent factors associated with a favorable outcome at one year of definite or probable NTM-PD. The composite endpoint adopted in our study could have a prognostic value as it effectively allows to discriminate patients who survived at one year from those who did not. When clinical symptoms, radiological lesions and negative microbiology data were observed in function of patient outcome at one year, only the absence of negative cultures was associated with death.
The diagnostic criteria for NTM-PD proposed by the ATS/IDSA were not associated with an improved status in our study. These results were similar to those found by other authors [11,14] on the relationships between ATS/IDSA diagnostic criteria and patient prognosis. In our study, the ATS/IDSA microbiological criteria had an excellent PPV of 89% for definite NTM-PD. Winthrop et al. [15] reported a PPV of 86% for NTM-PD. Jankovics et al. [16] showed that the PPV for NTM-PD varied from 64% to 94% depending on the clinical relevance of the NTM.
The majority of NTM-PD patients in this study (52%) were from the French West-Indies. In Guadeloupe, the species that predominated among these patients was MAC (45%), while in Martinique, it was M. fortuitum (47%). These frequencies are in agreement with Streit et al. [17] based on an epidemiological study in Guadeloupe, Martinique and French Guiana. Likewise, NTM-PD isolates from Montpellier showed a predominance of MAC (57%) followed by M. abscessus (32%), which paralleled NTM epidemiology in metropolitan France [18]. Finally, NTM-PD patients with M. abscessus in our study met more often the ATS/IDSA criteria for pulmonary disease (82%) than patients with MAC (62%). It should however be underlined that clinical characteristics and outcomes for the cohort with MAC were not further categorized by genetic sequencing to discriminate with distinct MAC species as reported recently [19], hence not allowing to conclude on a possible variability as regards to ATS/IDSA criteria for pulmonary disease depending on infection with M. avium, M. intracellulare, M. chimaera, or other MAC species.
In our study, no association between improved status and NTM species could be distinguished at one year. Indeed, the relatively shorter observation period in our study may not have allowed to perceive a species dependent association; e.g., Andrejack et al. [11] did not find significant differences in mortality based on the mycobacterial species in the first year of their study, although M. xenopi was associated with increased mortality as compared to MAC at the longer term. In this respect, mortality due to M. kansasii in our study (40%) seemed high as compared to published data [20], although our cohort (n = 5) was too small to draw conclusions. Furthermore, these patients had significant comorbidities, cavity lesions (3/5) and poor prognosis [12]. For NTM-PD with MAC, two conventional radiological presentations were the cavity radiological form that preferentially affected male smokers with a respiratory history, poor prognosis at one year, and bronchiectasis in women with no deaths at one year [12, 21]. The age <50 years and duration of macrolide treatment was associated with an improvement at one year. Advanced age was reported to be a poor prognostic factor [11,12].
Macrolides are an essential treatment of NTM-PD due to MAC [1, 4, 22]. In our group with MAC, the duration of macrolide treatment was predictive of an improvement at one year. Macrolide pharmacodynamics could explain the more effective action of these molecules over time; it was shown that the maximum eradication kinetics of MAC was slow with clarithromycin compared to amikacin [23]. Hence our observation that the duration of macrolide treatment was an improvement factor at one year for NTM-PD due to MAC was not unexpected, nonetheless it remarkably corroborates the positive effect of the duration of macrolide treatment in a clinical study. Significantly, the duration of ethambutol treatment in combination with other molecules in our study emerged as an independent factor of NTM-PD improvement at one year. Although the current treatment of NTM is largely empiric [24], ethambutol, an inhibitor of arabinogalactan synthesis, is known to significantly boost M. avium drug susceptibility in vitro [25], and was shown to enhance activity both of clarithromycin and rifampin against MAC in extra- and intracellular assays at serum level concentrations [26]. Indeed, the tripartite cell-envelope architecture of M. avium [27], is partially responsible for exclusion of antimicrobial agents, leading to its observed multiple drug resistance [28, 29]. With regard to the M. avium lipid bilayer at the surface of the cell-wall skeleton which decreases the permeability for hydrophilic molecules, ethambutol acts by decomposing not only the skeleton but also the lipid layer, thereby facilitating the diffusion of antibiotics [30]. These observations corroborate our main hypothesis that more the action of ethambutol is maintained over time, more the antibiotics are able to access and concentrate intracellularly with higher concentrations in the bacterium. This assumption is further substantiated by observations on synergistic effects of ethambutol with several molecules with intracellular action: rifampin, quinolones and macrolides [31, 32]. Undeniably, ethambutol is an important component in the current multidrug regimens for treatment of patients with MAC lung disease [7, 33, 34]; the microbiological response being significantly related to the duration of its use [35].
In our study, an HIV-positive status was an independent factor in improving NTM-PD at one year. Other studies have shown a dramatic improvement in HIV-positive patients with NTM-PD due to immune restoration by anti-HIV therapies [36]. Restoration of immunity, monitoring of these patients in specialized structures and treatment of NTM-PD have played an important role in improving prognosis.
Men with NTM-PD have been reported as having a worse prognosis in terms of mortality than women [11, 12]. In our study, men showed greater improvement at one year than women. The men were younger and sputum was more common among women (79% vs. 37%), as was bronchiectasis (64% vs. 36%). There was no difference between the sexes on the negativity of bacteriological cultures. Women showed little improvement clinically and radiologically compared to men due to the persistence of clinical symptoms. However, mortality at one year remained higher in men.
The strengths of our study are based on its multicenter approach, subgroup analyses and a large cohort constituted by the analysis of clinical and radiological records of patients rather than on discharge data. However, the limitations of this study are its retrospective design and the short observation period (one year), which does not allow to know if the improvement observed will persist over time. Another limitation of this study is the inclusion of probable cases of NTM-PD, which nonetheless translates a reality faced by clinicians. However, the current diagnostic criteria for NTM-PD does not have perfect sensitivity and specificity [37], and a fraction of these probable cases may evolve during monitoring to definite cases, although there is no data to confirm this assumption. Furthermore, at times one cannot conclude definitely, for example for the HIV subgroup, since the numbers in each sample (improved / unimproved) were too small to be able to carry out a multivariate analysis.
In this retrospective multicenter study, improvement at one year in patients with definite or probable NTM-PD was associated with the duration of ethambutol treatment, HIV-positive status and male gender. For the group of patients infected with MAC, improvement was associated with the duration of macrolide treatment and an age <50 years. Identifying predictors of improvement at one year of NTM-PD is expected to optimize the management of the disease in its early stages.
|
10.1371/journal.pcbi.1000813 | Combinatorial Gene Regulation Using Auto-Regulation | As many as 59% of the transcription factors in Escherichia coli regulate the transcription rate of their own genes. This suggests that auto-regulation has one or more important functions. Here, one possible function is studied. Often the transcription rate of an auto-regulator is also controlled by additional transcription factors. In these cases, the way the expression of the auto-regulator responds to changes in the concentrations of the “input” regulators (the response function) is obviously affected by the auto-regulation. We suggest that, conversely, auto-regulation may be used to optimize this response function. To test this hypothesis, we use an evolutionary algorithm and a chemical–physical model of transcription regulation to design model cis-regulatory constructs with predefined response functions. In these simulations, auto-regulation can evolve if this provides a functional benefit. When selecting for a series of elementary response functions—Boolean logic gates and linear responses—the cis-regulatory regions resulting from the simulations indeed often exploit auto-regulation. Surprisingly, the resulting constructs use auto-activation rather than auto-repression. Several design principles show up repeatedly in the simulation results. They demonstrate how auto-activation can be used to generate sharp, switch-like activation and repression circuits and how linearly decreasing response functions can be obtained. Auto-repression, on the other hand, resulted only when a high response speed or a suppression of intrinsic noise was also selected for. The results suggest that, while auto-repression may primarily be valuable to improve the dynamical properties of regulatory circuits, auto-activation is likely to evolve even when selection acts on the shape of response function only.
| Bacteria adjust which proteins they make, and how many copies of each kind, depending on their environment. The production rate of each regulated protein is controlled by a special class of proteins called transcription factors. The rate at which a certain protein is produced usually depends on the cellular concentrations of a few such transcription factors. When circumstances change, the concentrations of these transcription factors alter too and consequently the production rates of all proteins regulated by them are adjusted. Interestingly, many transcription factors also regulate their own synthesis rate. This suggests that this self-regulation must have one or more important functions. In this article we study one possible function. In order for cells to function properly each protein concentration has to respond in a particular way to changes in transcription factor concentrations. We have studied how bacteria can optimize and fine-tune these responses. To this end, we formulated a physical model of the regulation by transcription factors and performed computer simulations. These simulations show that self-regulation—and in particular self-activation—is often a useful tool to achieve the prescribed response. Therefore we conclude that natural selection on the regulation of protein levels could naturally lead to self-regulation.
| Many transcription factors (TFs) in Escherichia coli regulate the transcription rate of their own gene. In fact, 59% of the TFs are known to auto-regulate and the list is growing [1], [2]. Negative auto-regulation (auto-repression) occurs more frequently than positive auto-regulation (auto-activation), but both are very common: 71 auto-repressors and 34 auto-activators are found in the current databases (including 9 TFs that have binding sites for auto-activation as well as for auto-repression). This suggests that auto-regulation has one or several important functions [3], [4]. In this paper, one possible function is explored. In general, the expression level of a gene is a function of the concentrations of the TFs that regulate its transcription rate. We propose that auto-regulation can naturally be used to optimize the shape of this response function.
Auto-regulating transcription factors are typically regulated by other TFs too. In fact, 23 auto-regulating TFs in E. coli are known to respond to at least two additional regulators [1], [2]. In such cases, the response of the regulated TF to changes in the “input” TF concentrations must reflect an interplay between regulation and auto-regulation. Conversely, this suggests that auto-regulation could emerge as a result of natural selection on the shape of these responses.
In the past years, several other functions of auto-regulation have been proposed. Negative auto-regulation has been shown to decrease the sensitivity of expression levels to intrinsic fluctuations in the transcription rate under certain conditions [5]–[7] and to mitigate variations due to changes in the bacterial growth rate [8]. In addition, auto-repression can speed up the response of expression levels after a sudden change in conditions [9], [10]. In the presence of time delays, it can also create oscillations [11]. On the flip side, negative auto-regulation tends to reduce the sensitivity of the expression level to input signals [12], [13]. The effects of positive auto-regulation are usually opposite to those of auto-repression: it slows down responses and tends to amplify intrinsic fluctuations. At first sight, these qualities may not seem very desirable. Yet, a slow response can be beneficial if a sensitive response to persisting signals is desired while fast fluctuations in the input signal should be ignored [12]. Occasionally the fact that auto-activation can provide bi-stability may also be useful [14]. Each of these qualities could be relevant in some cases; our new suggestion does not contradict or replace any of them.
To study the benefits of auto-regulation we use a computational approach that we developed recently [15]. In this approach, an evolutionary algorithm and a physical–chemical model of transcription regulation are integrated to design in silico cis-regulatory regions with predefined response functions. The evolutionary algorithm subjects a population of model cis-regulatory regions to rounds of mutation and selection. The mutations are introduced at the level of base-pair sequences while the selection step is based on the emerging network properties calculated using the model of transcription regulation. In the course of these simulations complex promoter designs develop that perform the desired function; these designs often reveal new design principles. In earlier work, auto-regulation was not included in this method. In contrast, we now use an extended version of the method to design cis-regulatory constructs that can exploit feedback.
Many cis-regulatory regions in real cells essentially implement logical decisions [15]–[17]. We therefore study the class of response functions that can be interpreted as analogue equivalents of logic gates. Gates are computational devices that produce an output signal depending on one or more input signals; logic gates are gates that implement a binary (Boolean) decision rule. For example, a transcriptional AND gate would be a gene whose expression (the “output”) is regulated by two TFs (the “inputs”, TF1 and TF2) such that it is transcribed only if both TF1 and TF2 have a sufficient expression level [16]. We refer to Table 1 for the definitions of other logic gates. Even though it has proven fruitful to think of promoters as analog approximations of logic gates, we stress that gene expression levels are of course not actually binary and that we do not treat them as binary in the models below.
We analyze the cis-regulatory sequences resulting from the simulations by calculating DNA footprints for the resulting transcription factors and promoter sequences. These footprints show that auto-regulation—in particular auto-activation—is often used in these cis-regulatory regions; indeed, further analysis shows that auto-activation can be used to construct “better” transcriptional logic gates by allowing for more switch-like, “steep” response functions. However, the use of auto-regulation in shaping response functions is not limited to creating switch-like functions. To demonstrate this, we also applied our method to the design of cis-regulatory constructs that respond in a linear fashion to input concentrations. Again we find that auto-activation emerges spontaneously in the results.
Finally, we also performed simulations in which we selected for designs with desirable dynamical qualities. First, we adjusted the method to select for gates with a short response time. Second, we performed selection against intrinsic fluctuations. In agreement with earlier results [5]–[7], [9], [10], auto-repression evolved in both cases, demonstrating how auto-repression can be used to speed up response times or to reduce intrinsic fluctuations.
Before describing the results we first provide a detailed description of the model and the algorithm used.
We combine a model of transcription regulation and an evolutionary algorithm to design in silico cis-regulatory regions with a predefined function. The model of transcription regulation is constructed such that all properties of a model regulatory network follow entirely from the sequences of TFs and cis-regulatory regions. Binding sites of TFs are therefore not specified beforehand but appear gradually in the course of the simulations. The model is an extension to the one described in detail in our earlier publication [15]. The main innovation is that auto-regulatory interactions are now also included, so that auto-regulation can evolve if this is beneficial.
We consider one “output” gene, tf3, and at most two “input” transcription factors, TF1 and TF2. The gene tf3 codes for another transcription factor called TF3. All three TFs can regulate the transcription rate of tf3 by binding to its cis-regulatory region. (See Fig. 1 for an illustration of the model.)
The cis-regulatory region and the TFs are represented as nucleotide sequences and amino-acid sequences respectively. All TFs can bind anywhere on the cis-regulatory region, but the affinity of a TF for a particular site depends on the sequences of the TF and the site. For our purpose, it is sufficient to only model the DNA-binding domains of the TFs explicitly. We assume that amino-acids in these domains are responsible for the DNA-binding specificity and therefore represent each TF as an amino-acid sequence of length . We choose in our simulations because known binding sites in E. coli typically have length 6 to 15 and usually one base pair interacts with amino acid in TF–DNA binding [2]. The cis-regulatory region of tf3 is a base-pair sequence of length ; we take because in E. coli most transcription factors bind within from the start of transcription [18]. By the rules specified below all interactions between TFs, RNA polymerase (RNAP) and the cis-regulatory region can be deduced from these sequences; therefore each transcriptional gate is completely specified by them.
The various molecules interact in the following ways:
We model the dynamics of the concentration of TF3, , by the following ordinary differential equation:(4)Here is the maximal production rate of TF3, and is the degradation rate constant of TF3. The function was defined above. In this simplified description, transcription and translation are concatenated and translational regulation is not included.
The concentrations and of TF1 and TF2 are considered the inputs of the gate. Assuming that the system is mono-stable (bi-stability is discussed below) equation 4 defines a unique steady state for each set of input concentrations in which has a value . This steady-state concentration is considered the output of the gate. Because time delays between transcription initiation and translation are ignored, oscillations are excluded and can be calculated by propagating the dynamics numerically from any initial condition until the steady state is reached. (If a gate has only one input, the dependence on is simply dropped.)
We choose the constants and such that ; this ensures that stays within the range . Apart from this ratio the values of and are irrelevant because in this work we are not interested in absolute time scales of the dynamics; for simplicity we use a time unit such that .
In order to design networks with a prescribed function an evolutionary algorithm was used. A population of 200 transcriptional gates was subjected to 1000 cycles of mutation, selection and replication. Initially, all gates had random sequences. Auto-regulation was not imposed, but the system was free to exploit it by evolving binding sites for TF3. The details of the evolutionary algorithm were chosen to combine an effective optimization of the gates with computational efficiency; we emphasize that we do not intend to faithfully mimic biological evolution.
Several types of mutations were included. First, a base substitution could occur in cis-regulatory sequences (with probability per cis-regulatory region). If this happened, a base pair was selected at random from the cis-regulatory sequence and substituted by a randomly chosen nucleotide. Second, insertions or deletions of a random base pair occurred in cis-regulatory regions (with probability ). Third, we applied point mutations to the sequences of the TFs (with probability per TF), in which case one randomly chosen amino acid in the sequence was replaced by a random alternative. The exact mutation rates are not crucial for the results, as long as the rates are (i) high enough to generate significant variation and (ii) low enough to allow high-quality gates to persist in the population.
For the selection step a fitness score was used. Here (where RF stands for Response Function) measures the deviation of the response function from a predefined goal function (the desired response function). It was computed as follows. When evaluating gates with two inputs, and , the output level was computed for 16 combinations of the input concentrations: (see the red dots in Fig. 2B). Next, the differences between these output levels and the goal function were computed. was defined as the sum of the squares of these deviations. If the gate had only one input, the definition was analogous, except that seven input values were used, equally spaced in the interval . The constant is required to make dimensionless (for simplicity, ) and is an arbitrary constant large enough to ensure . Based on the fitness scores, the top 20% of the population were selected and the remaining gates discarded. Subsequently the population was brought back to its initial size by duplicating gates randomly chosen from the survivors of the selection process.
If auto-activation evolved, the system could become bi-stable. In bi-stable systems, given the inputs two different values of are stable under the dynamics of the system (equation 4) so that the output concentration is not uniquely defined by the input concentrations. Even though bi-stability is likely to occur in some real transcription networks we decided that such systems do not qualify as gates, since gates by definition map input states to a uniquely defined output state. The fitness function therefore contained an additional term that was designed to penalize bi-stability. When evaluating the fitness of a gate, we always computed the steady-state value twice for input values : once by propagating the differential equation 4 using initial condition and once using . If the results were different, the difference squared was added to the fitness function, which was sufficient to assure that the particular gate was eliminated by the selection process. However, because this method did not exclude bi-stability for all possible input values we also checked afterward whether the results were bi-stable.
All gates are defined for input concentrations in the domain only. The logic gates are specified in Table 1. In addition we define LACT, LIN, MEAN and NMEAN gates. A LACT (linear activate) gate has one input, , and the output responds as . A LIN (linear inhibit) gate also has one input, but responds according to . A MEAN gate is linear in two inputs, and , and obeys . Lastly, we define NMEAN to have the following linearly decreasing response function: .
We repeated the simulations for each of the gates 20 times with different random seeds.
In order to quantify the importance of auto-regulation in a particular design we defined the measure (referred to as the “feedback measure”). First, we calculated the response function for the particular design. Then we artificially removed all possible binding sites for TF3 by setting the affinity of TF3 for all sites on the cis-regulatory region to zero and calculated the response function again; we call the result . In the absence of auto-regulation one should find , but if auto-regulation does play a role the two functions differ. Therefore the difference between these functions is a measure of the degree of auto-regulation; we define as the mean of the squared differences over 16 combinations of the input concentrations (again, the red dots in Fig. 2). If auto-regulation is not being exploited by a certain design is generally small (). If, on the other hand, auto-regulation is used the resulting value is typically in the range . (For instance, if all 16 points shift by 100nM when auto-regulation is removed .)
During the simulations, binding sites emerge in the initially random cis-regulatory promoter sequences. However, since the equilibrium binding constants have continuous values there is no fundamental distinction between binding sites and non-binding sites. Recognizing binding sites is further complicated by the fact that, in particular in the presence of cooperativity, weak binding sites can be important. Nevertheless, in order to understand the design principles of a particular gate we wish to identify which binding sites are necessary and sufficient to explain the observed promoter response. This problem is not an artifact of our models: the exact same conceptual problems occur whenever one tries to identify the binding sites of real TFs by experiment.
Since a direct cutoff in terms of the equilibrium constants would eliminate possibly important weak sites we use computational “DNA footprints” (analogous to experimental techniques such as DNase I footprinting) to select those sites that are likely to be important. For each TF and each site , we calculate the steady-state occupancy for four sets of input concentrations . Sites that influence the response of the gate should have a significant occupancy in at least one of these digital footprints.
We define to be the maximal occupancy of site by TF over the four conditions. Figure 4 in Text S1 shows a histogram of these occupancies for all TFs and all sites using data gathered from the results of 200 simulations. This histogram is bi-modal. The vast majority of the maximal occupancies have negligible values but a second peak occurs at . This peak is the result of selection pressure and is associated with functional binding sites. Based on this histogram, we use a rather stringent cut-off at to separate binding sites from pseudo binding sites. Simplified models that only take into account these selected binding sites and assume that all other binding affinities are zero usually accurately reproduce the response function of the full, unsimplified system. In rare cases where this is not the case the threshold can be lowered to obtain more accurate but more complex models; this was not necessary for the examples presented below. (See the Text S1 for more details and examples of footprinting profiles.)
Below, we describe the results of the simulations. We regularly compare the results to our previous work in which auto-regulation was excluded (Ref. [15]). In some cases the simulations presented here resulted in the same designs as before. But in other cases the designs exploited auto-regulation to arrive at novel and often superior designs. We identified several mechanisms that are used repeatedly in the results and present them below. Details about several of the analyses below can be found in Text S1.
The first mechanism that our scheme elucidated, we called conditional auto-activation. This mechanism occurred in AND and ACT (activation) gates (see Table 1 for the definitions), in which cooperative activation plays a key role. In those gates, conditional auto-activation is used to create a steep, switch-like response. As an example, we first discuss the design of AND gates.
In simulations in which auto-regulation was excluded by the method, the resulting AND gate designs always consist of a tandem array of binding sites to which TF1 and TF2 bind cooperatively (see Fig. 2A) [15], [16]. We called this a hetero-cooperative module. This design functions as follows. Crucially, the binding site from which RNAP is recruited (the site directly next to the core promoter) is too weak to considerably activate transcription on its own. As a result, only when TF1 and TF2 are both present at sufficient concentrations they bind cooperatively and activate transcription, as the definition of an AND gate requires. In the new simulations, in which auto-regulation can evolve, this design still emerged in 14 out of 20 simulation runs. Each of these gates has a feedback measure , proving that auto-regulation does not play any role.
The remaining 6 simulation runs resulted in conditional auto-activation. In these gates the feedback measure was high, in the range . The new design looks very similar to the old one (see Fig. 2A and B). However, the hetero-cooperative module now also contains a binding site for TF3, which leads to a positive feedback loop. Importantly, TF3 bound at its binding site cannot recruit RNAP directly; instead, it interacts with the hetero-cooperative activation module. As a result, the auto-activation is conditional on the presence of TF1 and TF2. As the concentrations of TF1 and/or TF2 increase, the auto-activation is gradually turned on, leading to a sudden (steep, switch-like) response.
The exact same mechanism is exploited by some ACT gates. Out of the 20 simulations of ACT gates, 3 resulted in conditional auto-activation ( values were , and ), while the other 17 did not use auto-regulation ().
The basic mechanism can be studied in minimal models inspired by the simulation results. In Fig. 2C and D, we compare three activation mechanisms. The first scenario is conventional activation by a single TF1 binding site. In the second scenario only a homo-cooperative activation module is present, consisting of two binding sites for TF1. In the third scenario, the auxiliary TF1 site is replaced by a binding site for TF3, introducing conditional auto-activation. In all designs we chose the binding site affinities such that they maximize the fitness function for the ACT gate. Conditional auto-activation indeed produces a response that is steeper than the one resulting from the design with a single activator binding site (Fig. 2D). However, the conventional cooperative design with two binding sites gives an even steeper result. The results imply that, after one binding site has evolved for the activator TF1, the response can be improved in two ways: by adding an additional site for TF1 (leading to cooperative activation) or by adding a site for TF3 (resulting in conditional auto-activation). Which design emerges therefore depends critically on the actual sequences and mutations occurring in the population. This explains why cooperative activation and conditional auto-activation show up as alternatives in the simulations.
The effect of conditional auto-activation can be understood quantitatively by studying the minimal model mathematically. The response function for the minimal model follows from the condition and is given by(5)withHere and denote the dissociation constants for TF1 and TF3 binding to their respective operator, is the concentration of RNAP normalized by the dissociation constant of RNAP binding to the promoter, and . We first consider the limit of and . In this limit is small as long as . The numerator of 5 can then be approximated by . As a result we can distinguish two regimes depending on the sign of :(6)where is the border between the two regimes, implicitly given by :(7)Note that provided ; under this condition approximation 6 holds around the transition . As both and are linear functions of , the second regime has the form of a Hill function with Hill coefficient . Therefore equation 6 shows that equation 5 behaves like a sharp threshold response. This threshold effect is responsible for the increased steepness of the response due to conditional auto-activation.
The maximal expression following from equation 5, at full activation, is . This demonstrates that in the limit of the maximum expression level becomes very low (for a given value of ). On the other hand, if is increased the term becomes more and more significant and the transition between the two regimes in equation 6 becomes more and more gradual. Consequently, in the optimized functions plotted in Fig. 2 the values of reflect a compromise between the opposing requirements of having a high maximal expression (requiring a large ) and a sharp threshold response (requiring a small ).
The steepness of a function in the point can be formalized by the sensitivity, defined by . The sensitivity of a Hill function is limited by the Hill coefficient , which is equal to the number of cooperatively interacting binding sites for the input TF. We therefore ask if a similar limitation applies to the minimal model of conditional auto-activation. From equation 5 the sensitivity function can be derived straightforwardly. The result is rather cumbersome and therefore an exact expression for the maximal sensitivity is hard to obtain. However, since the most sensitive part of the function is in the region where (i.e., close to ), the maximal sensitivity can be approximated by . In the limit of large this expression converges to(8)Importantly, since we evaluated the sensitivity in a point close to but not exactly at the maximum, this approximate result is a conservative estimate: the true maximum cannot be lower than this. In the limit of small the maximum sensitivity diverges as , which proves that the sensitivity of response function 5 does not have a theoretical upper limit, unlike those of Hill functions.
So far we have neglected the dynamical properties of the designs because the current model only considers the steady-state response of the system. As we mentioned, auto-activation tends to slow down the response time of the system. Therefore, in systems where the speed of response is of great importance cooperative regulation is expected to outperform conditional auto-activation. Selection on response speed is discussed in more detail below.
A second feedback pattern emerges in logic gates in which repression is important, notably the NAND, NOR and IN (inhibition) gates. As it turns out, whenever steep repression is required, we also find strong auto-activation; this occurred in every simulation run for our NAND, NOR and IN gates (20 repeats each), with in all cases. We present the NAND gate as an example.
Fig. 3A shows the cis-regulatory region of a typical NAND gate using auto-activation. The corresponding response function plotted in Fig. 3B indeed shows an excellent NAND-like behavior. As quantified below, in fact it performs better than the design without auto-activation reported earlier and reproduced for comparison in Fig. 3A and B [15].
The design that resulted when auto-activation was excluded is composed of a hetero-cooperative repression module (a tandem series of repressor sites to which both input TFs bind cooperatively). The function of this module is to repress transcription only when both TF1 and TF2 are present in sufficiently high concentrations, as required of a NAND gate. In Ref. [15] we pointed out that in the simulation results such a repression module was always accompanied by strong activation sites for both input TFs. This counter-intuitive feature turned out to enhance the sharpness of the response. At low TF concentrations, the activation sites counter-act the repression module, so that the expression stays high. At higher TF concentrations, however, the repression module dominates and represses transcription. Under the parameters used designs of this type reached a modest fold-change of and a deviation measure (see section “Evolutionary algorithm” in the methods section for the definition of ).
In the new results (Fig. 3) the activation sites for TF1 and TF2 have disappeared, but instead we find auto-activation. In the absence of TF1 and TF2, tf3 is highly expressed, aided by auto-activation. As the concentrations of TF1 and TF2 are increased, the repression module starts to compete with the auto-activation module. Quite suddenly, the repression module wins this competition and displaces RNAP from the promoter. The strong, cooperative repression module now leads to a rather complete inhibition. The new design can lead to fold-changes of and a deviation measure of .
To study the mechanism responsible for the steepness of the response function we again analyzed a minimal model. In Fig. 3C and D two scenarios for an IN gate are compared. In the first scenario, a transcription factor TF1 cooperatively binds to a pair of repressor sites to inhibit the gene tf3. In the second scenario we use the same configuration, but add an activator site for TF3. Thus, auto-activation competes with cooperative repression. The fitness of each design is optimized using the fitness function for the IN gate. As can clearly be seen in Fig. 3D the second scenario, using conditional auto-activation, results in a steeper and more complete repression. Figure 1 in Text S1 shows plots of the sensitivity as a function of for the response plots in Fig. 3D and clearly demonstrates that auto-activation enhances the sensitivity.
Does the sensitivity of the response function 9 have an upper bound, as is the case for Hill functions? To answer this question we again study the minimal model mathematically.
The response function of the minimal model is given by(9)with(10)(11)(12)We first describe the limit in which and . The form of this equation is obviously similar to equation 5 for conditional auto-activation. However, here is large and negative () when . In this regime, , so that the numerator is approximated by . As increases, decreases while the denominator increases; therefore the expression is rapidly repressed. This regime ends suddenly as reaches zero, at ; at this point the expression is almost fully repressed.
The sensitivity function can be derived from equation 9. The exact expression is again too cumbersome to derive the maximal sensitivity analytically. However, the most sensitive region of the response plot is again expected around so that we estimate the maximal sensitivity as(13)For large this converges to(14)which for large approaches . Numerical tests demonstrate that this conservative approximation becomes excellent for (data not shown). In the absence of auto-regulation, the sensitivity cannot exceed 2, the number of repressor sites. Equation 14 demonstrates that in the presence of auto-regulation the sensitivity can easily exceed 2 but is nevertheless limited given .
We note that the sensitivity is optimal for . Hence, unlike the case of conditional auto-activation the requirements of a high maximal expression and a high sensitivity do not contradict. In the simulations as well as in reality, however, the promoter strength is bounded by other factors. Clearly the binding affinity of RNAP for the promoter is bounded by the physics of RNAP–DNA binding. A less obvious constraint follows from the fact that the expression switches from high to low around ; if the repression is to occur at reasonable TF1 concentrations (the simulations impose the interval ) high values of require low values of the dissociation constant (i.e., strong repression). Finally, a high sensitivity in one point does not guarantee that the response function switches from high to low in a narrow interval as is required by the fitness function; this explains why in the plots in Fig. 3 the maximal sensitivity is not optimal ().
In both previous cases, auto-regulation was used to obtain the steep or switch-like behavior required to approximate the binary responses of logic gates. Indeed, sharp responses are observed and probably required in many real examples; nevertheless many genes respond in a more gradual manner to their input signals (Ref. [17] provides examples of both sharp and gradual responses). Is auto-regulation also useful in cases where a gradual response is required? To test this, we now turn to the results of simulations with linear goal functions.
Indeed, simulation results for linear repression (i.e., the LIN and NMEAN gates) always use auto-regulation, with in the range . As can be seen in Fig. 4, approximately linear repression can be obtained when repression is combined with auto-activation; the deviation measure for the simulation result shown is . The same figure shows results of simulations in which auto-regulation is excluded. In that case a large cooperative repression module results, which leads to a less linear result ().
Again, we analyzed the mechanism through a slightly simplified model presented in the same figure. The promoter design of the simplified model is identical to the one presented in Fig. 3C, where it was used to demonstrate how auto-activation can provide sharp responses. In essence, the difference between the two cases is that in the IN gate the two repressor sites have the same affinity, whereas in the LIN case they do not: one of them is many times weaker than the other ( vs. . As a result, the repression is introduced gradually as the repressor concentration increases.
Linear repression requires that in the domain . Since is defined as the solution of , we can take the total derivative of this relation to arrive at(15)In the absence of auto-regulation the denominator equals 1. In this case is a Hill-type function of and therefore its derivative is not constant. In the presence of auto-regulation the denominator can be used to correct some of the variation in the numerator. (See Text S1.)
In contrast, in the simulation results for linear activation (both the LACT and the MEAN gates) auto-regulation is never used. To test if these results are an artifact of the algorithm, we studied a series of models (see Text S1). Each model is a possible layout of transcription factor binding sites and includes auto-regulatory sites. For each of the models, we optimized the affinities of all binding sites with respect to the fitness score, using a standard Nelder–Mead optimization routine. Consistent with our simulations, in the solutions for all models the affinities of the auto-regulatory sites vanished. Even though the list of models tested is not exhaustive, this suggests that auto-regulation is not helpful in constructing LACT or MEAN gates.
To illustrate an important difference between linear activation and linear repression we provide the following general argument. Suppose that an accurate LACT gate can be constructed using auto-regulation. By definition the response function should then be in the interval . Consequently, in this interval. Interestingly, this shows that if all TF3 binding sites in the cis-regulatory region are replaced by binding sites for TF1—resulting in a gate without auto-regulation—the exact same response function should be obtained. Even though this argument does not prove that auto-regulation cannot be used to construct LACT gates, it does show that if a high-quality LACT can be constructed with auto-regulation a similar response can always be obtained without it as well. This is in stark contrast with the linear repression case, where .
Surprisingly, auto-repression does not show up in any of the simulations described so far, whereas auto-activation features regularly. As we mentioned in the introduction, previous studies have shown that negative auto-regulation can be used to diminish intrinsic noise and to speed up response times. In the simulations presented so far, such qualities were not rewarded. Therefore we asked if auto-repression would emerge if we did select for such dynamic properties on top of our usual selection criteria.
First, we used a heuristic measure (where RT stands for Response Time) to select for a quick response to changes in the input parameters; it was computed as follows. For 16 combinations of input concentrations (corresponding to the red dots in Fig. 2) we numerically solved the differential equation 4 with two different initial conditions: and . The solutions were used to measure the time it took for the system to approach the steady-state value up to a small distance . The measure was defined as the sum of all 32 response times. The total fitness function, combining selection on the response function with selection on the response time, was , where the factor was used to tune the relative strength of the selection on the response time. Again, is an irrelevant constant ensuring that .
In fact, the initial condition of the simulations, in which the gate is completely dysfunctional, is a local optimum of this fitness score. This is because initially the steady-state expression level is negligible so that the response time for initial condition is practically zero. Even though mutations that increase the constitutive promoter activity improve both the response function and the response time for initial condition , the concomitant increase in the response time for initial condition dominates. To ensure that the simulation was not trapped in this local optimum was increased slowly from 0 to in the course of the simulations, according to:(16)where is the simulation time (i.e., the cycle number of the evolutionary algorithm), and .
Second, we selected against intrinsic noise, i.e., fluctuations in the concentration of due to the stochasticity of the processes involved in the production and degradation of TF3. This type of noise should be contrasted with extrinsic noise, which here is understood to be the noise due to fluctuations in the input concentrations and or due to changes in RNAP concentrations [12], [27]. Even though extrinsic noise is generally important too [28], the treatment of extrinsic fluctuations involves subtleties that are beyond the scope of this work, such as the question which changes in and should be considered changes in the input signal and which should be considered noise. We therefore only consider intrinsic noise.
In order to treat intrinsic fluctuations in a tractable manner we now replaced the ordinary differential equation 4 by the following stochastic differential equation:(17)The term represents Gaussian white noise and is characterized by (see Text S1):(18)The first term on the right-hand side describes the noise in the production of TF3 while the second term describes the stochasticity in the degradation of TF3. Both terms depend explicitly on the volume because, at constant concentration, the copy number of TF3 scales with which affects the variance in .
In Text S1 we show that the standard deviation of the concentration can be approximated as:(19)with(20)where and are the first and second partial derivatives of with respect to , and is the steady-state solution of the deterministic equation 4. We computed the right-hand side of equation 19 numerically for 16 input values (again, corresponding to the red dots in Fig. 2) and treated the sum of the results as an additional fitness measure (where N stands for Noise). The strength of selection against noise was again increased gradually during the simulations (analogous to equation 16). The total fitness function thus became .
We performed simulations with several values for and : (where is the arbitrary unit of time, see Methods), and . Indeed, in these simulations auto-repression emerged.
In activating gates (ACT, AND, OR) auto-repression resulted in all simulation runs with or . The auto-repression was invariably strong, with , and mediated by multiple cooperative binding sites. If or were further increased eventually the resulting cis-regulatory regions became completely dysfunctional; this can be understood from the fact that both the response time and the noise reduction can be optimized by abolishing expression altogether. Figure 7 in Text S1 demonstrates how the properties of resulting OR gates changed as a function of . As is increased the deviation of the response from the ideal OR gate, measured by , increases, while the noise, measured by , decreases.
As we explained, the response functions of the NAND, NOR and IN gates benefit from auto-activation; in those gates auto-activation occurred unless the selection pressure on the dynamical properties dominated (i.e., if or were large), in which case the quality of the response functions was negatively affected. Fig. 4 shows results from simulations selecting for NAND gates at various values of . As the selection pressure on response time was increased the response functions became more and more compromised. Auto-activation resulted for and ; in the former case the promoter designs were of the type shown in Fig. 3A, while in the latter case only one auto-activation site remained. Interestingly, in most of the simulation runs (18 out of 20) at a weak auto-repression site shows up in conjunction with the auto-activation site. These weak auto-repression sites are incorporated in the hetero-cooperative repression module and have a high occupancy only at high concentrations of TF1 and TF2; analogous to conditional auto-activation, this effect could be called conditional auto-repression. At the auto-activation was replaced by strong auto-repression mediated by a single or multiple binding sites and the sites for the input TFs were very weak. Finally, at the resulting gates became completely dysfunctional and no significant binding sites remained.
The results above suggest that auto-activation and auto-repression have very different functions. We therefore wondered whether in the known transcription regulatory network of E. coli the auto-activators and auto-repressors have different statistical properties.
Surprisingly, we found that auto-activators are more often regulated by other TFs than auto-repressors. According to the data in RegulonDB [2], 18 of the 25 auto-activating TFs in E. coli are regulated by at least one additional TF (72%) versus 30 out of 62 auto-repressing TFs (48%); this indeed suggests that auto-activators are more likely to have additional inputs (). The difference becomes more convincing if we look at the total number of inputs for the two sets. The 25 auto-activators have, in total, 52 inputs (i.e., an in-degree of 2.08 on average; the auto-regulation is not counted as an input) while the 62 auto-repressors have 50 inputs in total (0.81 on average). Evidently, auto-activators have significantly more inputs than auto-repressors ().
Since auto-regulation can potentially have many functions and most of the auto-regulators are poorly characterized, we can only speculate about the origin of this difference. One possible explanation would be the following. If a common function of auto-activation is to shape response functions, as suggested by our analysis, then auto-activation should evolve preferentially for TFs that are regulated by one or more input TFs. In that case one would expect the average in-degree for auto-activators to be high. The same argument does not hold for auto-repression: our results suggest that auto-repression typically evolves for different reasons. Some of the functions of auto-repression suggested in the literature, such as its tendency to decrease intrinsic noise and to mitigate the effect of changes in the bacterial growth rates on gene expression, do not require additional input TFs. It is therefore not too surprising that for many auto-repressors (32 out of the 62) no input TF is known.
Our results shed new light on the use of auto-regulation. We described three situations in which auto-activation can be used to improve the response function of promoters. First, if auto-activation is conditional on the presence of other TFs, it can give rise to sensitive responses that otherwise require multiple cooperative binding sites of the input TF. Presumably, not all input TFs can bind cooperatively to multiple binding sites; in those cases conditional auto-activation can serve as an alternative. Secondly, auto-activation can strongly contribute to the sensitivity of the response of repression systems. Whenever sharp repression is required, auto-activation can have a selective advantage. Thirdly, we showed that auto-activation is also useful if a linearly decreasing response function is desired. Together, such mechanisms may help explain the large number of auto-activators present in E. coli.
Auto-repression never appeared in the simulation results if selection was based on the response function of the gates only. Most likely, the limited use of auto-repression in shaping response functions is due to its general tendency to decrease the fold-change and sensitivity of the response. A low fold-change or sensitivity can typically also be achieved without auto-repression by tuning both the promoter strength and the affinities of the TF binding sites. Nevertheless, we cannot exclude the possibility that auto-repression would show up in simulations selecting for response functions different from the ones presented here.
If the fitness function was altered to favor a fast response or suppression of intrinsic transcriptional noise, auto-repression did emerge. It has been suggested before that the function of negative auto-regulation is to regulate such dynamic properties [5], [9], [10]; our results support this view.
In retrospect, the emergence of auto-regulation is hardly surprising. The evolution of cis-regulatory regions can be perceived as adaptive curve fitting. Allowing for auto-regulation gives gene-regulatory systems additional degrees of freedom to optimize their performance, and it would perhaps be more surprising if this freedom were not exploited. We therefore expect that the conclusions based on the idealized gates studied in this work are also relevant for real biological systems requiring more complex response functions.
We have seen that in some cases the advantage of using auto-regulation is large (e.g. when sensitive repression is required) whereas in other cases there is only a small difference between the quality of the response function for designs with or without auto-regulation. This leads one to wonder whether in the latter case natural selection on the shape of the response would be large enough to evolve and maintain auto-regulation, in particular in the presence of noise. This is largely an open question; yet, the fact that some E. coli promoters contain a large number of TF binding sites many of which contribute only marginally to the expression (see for instance [29]) suggests that, at least in some cis-regulatory regions, natural selection is strong enough to fine-tune the response function in great detail.
The results presented are quite insensitive to the parameter values chosen. The value of influences important properties such as the maximum fold change in activation systems, but as long as it is chosen within the biological range 10–100 the designs of the gates do not seem to depend qualitatively on the value chosen. To verify this, we performed simulations with for AND, NAND, NOR and OR gates (without selection against noise or response speed) and found the results to be qualitatively the same as those presented. The value of influences the spacing of binding sites within a module, but not the basic designs properties, as long as so that overlapping modules can be constructed that bind independently. The results are also insensitive to the length of the binding sites (we tested this with simulations for AND, NAND, NOR and OR gates with ) and the matrix elements of the binding energy matrix; essential is only that the evolutionary algorithm can tune the dissociation constants of the binding sites to a wide range of values (1–10000nM), as in reality. The length of the cis-regulatory region, , determines the maximum number of tandem binding sites that fit on the regulatory region; larger values of therefore ultimately lead to larger tandem arrays. However, since tandem arrays of five or more binding sites can form in the simulations, we believe that is large enough to accommodate typical E. coli promoters.
Even though in eukaryotes the mechanisms of gene regulation are generally different and various additional layers of regulation exist, recent work has shown that many basic principles of prokaryotic gene regulation—in particular the interplay between cooperative binding and competitive inhibition—are equally important in eukaryotes (see for instance [30] about repression and inhibition in yeast and [31] about enhancers in Drosophila). Auto-regulation is also widespread in eukaryotes [32]; therefore, our findings could also be relevant for gene regulation in eukaryotes.
As we mentioned, auto-activation is known to reduce the response speed in some situations and to increase the amplitude of fluctuations. Clearly, those issues may be problematic in some real-life situations. On the other hand, a slow response can be a positive feature as well if it is applied as a filter of high-frequency noise (a low-pass filter). Fluctuations may in some cases be beneficial or even necessary. For instance, when cells respond to a fluctuating environment via the strategy of stochastic switching, fluctuations are essential [33]. But even when cells cope with a fluctuating environment via the strategy of deterministic switching, fluctuations may be beneficial, since they can increase the population's growth rate when the response function is suboptimal [34]. Indeed, the fact that auto-activation is found so often in E. coli demonstrates that the associated reduction of the response speed and the amplification of fluctuations can apparently be circumvented, tolerated or put to use.
|
10.1371/journal.pbio.1001610 | Production of α-Galactosylceramide by a Prominent Member of the Human Gut Microbiota | While the human gut microbiota are suspected to produce diffusible small molecules that modulate host signaling pathways, few of these molecules have been identified. Species of Bacteroides and their relatives, which often comprise >50% of the gut community, are unusual among bacteria in that their membrane is rich in sphingolipids, a class of signaling molecules that play a key role in inducing apoptosis and modulating the host immune response. Although known for more than three decades, the full repertoire of Bacteroides sphingolipids has not been defined. Here, we use a combination of genetics and chemistry to identify the sphingolipids produced by Bacteroides fragilis NCTC 9343. We constructed a deletion mutant of BF2461, a putative serine palmitoyltransferase whose yeast homolog catalyzes the committed step in sphingolipid biosynthesis. We show that the Δ2461 mutant is sphingolipid deficient, enabling us to purify and solve the structures of three alkaline-stable lipids present in the wild-type strain but absent from the mutant. The first compound was the known sphingolipid ceramide phosphorylethanolamine, and the second was its corresponding dihydroceramide base. Unexpectedly, the third compound was the glycosphingolipid α-galactosylceramide (α-GalCerBf), which is structurally related to a sponge-derived sphingolipid (α-GalCer, KRN7000) that is the prototypical agonist of CD1d-restricted natural killer T (iNKT) cells. We demonstrate that α-GalCerBf has similar immunological properties to KRN7000: it binds to CD1d and activates both mouse and human iNKT cells both in vitro and in vivo. Thus, our study reveals BF2461 as the first known member of the Bacteroides sphingolipid pathway, and it indicates that the committed steps of the Bacteroides and eukaryotic sphingolipid pathways are identical. Moreover, our data suggest that some Bacteroides sphingolipids might influence host immune homeostasis.
| While human gut bacteria are thought to produce diffusible molecules that influence host biology, few of these molecules have been identified. Species of Bacteroides, a Gram-negative bacterial genus whose members often comprise >50% of the gut community, are unusual in that they produce sphingolipids, signaling molecules that play a key role in modulating the host immune response. Sphingolipid production is ubiquitous among eukaryotes but present in only a few bacterial genera. We set out to construct a Bacteroides strain that is incapable of producing sphingolipids, knocking out a gene predicted to encode the first enzymatic step in the Bacteroides sphingolipid biosynthetic pathway. The resulting mutant is indeed deficient in sphingolipid production, and we purified and solved the structures of three sphingolipids that are present in the wild-type strain but absent in the mutant. To our surprise, one of these molecules is a close chemical relative of a sponge sphingolipid that is the prototypical ligand for a host receptor that controls the activity of natural killer T cells. Like the sponge sphingolipid, the Bacteroides sphingolipid can modulate natural killer T cell activity, suggesting a novel mechanism by which Bacteroides in the gut might influence the host immune response.
| Sphingolipids and their breakdown products modulate a variety of eukaryotic signaling pathways involved in proliferation, apoptosis, differentiation, and migration (Figure 1). Although sphingolipids are ubiquitous among eukaryotes, few bacteria produce them [1]. The genus Bacteroides and its relatives are an important exception; 40%–70% of the membrane phospholipids of these prominent symbionts are sphingolipids [2],[3]. While the structures of several Bacteroides sphingolipids have been solved, the full repertoire of these molecules has not yet been defined [1]–[19]. Here, by systematically exploring the sphingolipid repertoire of Bacteroides fragilis, we show that this gut commensal unexpectedly produces an isoform of α-galactosylceramide, a sponge-derived sphingolipid that is the prototypic ligand for the host immune receptor CD1d.
To gain insight into the potential role of Bacteroides sphingolipids in mediating microbiota–host interactions, we set out to define the complete set of sphingolipids produced by Bacteroides fragilis NCTC 9343 [20], a genome-sequenced, genetically manipulable human gut isolate. Reasoning that a chromatographic comparison of lipid extracts from wild-type B. fragilis and a sphingolipid-deficient mutant would reveal the complete set of B. fragilis sphingolipids, we began by attempting to identify genes involved in B. fragilis sphingolipid biosynthesis. We took a candidate gene approach, hypothesizing that the Bacteroides sphingolipid pathway would harbor homologs of the eukaryotic pathway [17]. BLAST searches of the B. fragilis genome using the Saccharomyces cerevisiae sphingolipid biosynthetic enzymes as queries yielded two hits encoded by adjacent genes: BF2461, a putative serine palmitoyltransferase, and BF2462, a putative sphinganine kinase.
Bioinformatic analysis suggested that BF2461, like its yeast homolog, is a pyridoxal-phosphate-dependent α-oxoamine synthase that conjugates serine and a long-chain acyl-CoA to form 3-dehydrosphinganine. In eukaryotes, this serves as the first committed step in the sphingolipid biosynthetic pathway. We therefore predicted that a Δ2461 mutant would be completely deficient in the production of sphingolipids. The eukaryotic homolog of BF2462, sphingosine kinase, phosphorylates sphingosine to form sphingosine-1-phosphate (S1P). Given that this reaction diverts the flux of the sphingosine base away from ceramide and toward S1P, we hypothesized that a Δ2462 mutant would produce a higher titer of mature sphingolipids than the wild-type strain.
We constructed a mutant harboring a deletion of BF2461 (Δ2461) (see S1.8 in Supporting Information S1). Although we obtained co-integrates for the BF2462 mutant, double crossover mutants were never obtained despite repeated attempts to screen through thousands of colonies, suggesting that BF2462 may be essential for Bacteroides viability. An interesting alternative comes from the observation that dihydrosphingosine, the putative substrate of BF2462, is toxic to Bacteroides melaninogenicus at 4 µM [11]; the absence of BF2462 could therefore lead to the buildup of a toxic intermediate.
Nevertheless, since the yeast homolog of BF2461 constitutes the entry point to the sphingolipid pathway, we hypothesized that the Δ2461 mutant would be sphingolipid-deficient, providing an ideal starting point for enumerating the B. fragilis sphingolipids. To test our hypothesis, we used comparative HPLC-ELSD to analyze alkaline-stable lipid extracts from the wild-type (WT) and Δ2461 strains. Our analysis revealed three primary peaks that were present in the WT but not the Δ2461 extract (Figure 2). Preparative thin layer chromatography was used to purify multimilligram quantities of these compounds, and HPLC-MS analysis of the purified material revealed that each peak consists of a mixture of co-migrating compounds that vary in mass by 14 Da. Measured in negative mode, the most abundant mass ions for peaks 1, 2, and 3 were 677.5 Da, 554.5 Da, and 716.6 Da, respectively.
To solve the chemical structures of the sphingolipid species, we first subjected the purified compounds to high-resolution MS. The mass of peak 1 was consistent with ceramide phosphorylethanolamine (CPE) (C36H74N2O7P; [M-H]− m/z: calculated 677.5234, observed 677.5221), a sphingomyelin isoform previously found to be the principal B. fragilis sphingolipid, while the mass of peak 2 was consistent with the corresponding dihydroceramide base (C34H68NO4; [M-H]− m/z: calculated 554.5148, observed 554.5156) (Figure 1A; Figure S1 in Supporting Information S1). A set of 1D and 2D NMR experiments on the purified compounds from peaks 1 and 2 yielded resonances and couplings consistent with these assignments (see S4.1 and S4.3 in Supporting Information S1).
In contrast, peak 3 was not a known compound. High-resolution MS analysis of the purified material from peak 3 was consistent with an empirical formula of C40H79NO9 ([M-H]− m/z: calculated 716.5682, observed 716.5698). 2D NMR analysis indicated that this compound and CPE harbor an identical dihydroceramide base (C34H68NO4), suggesting that the difference (C6H11O5) corresponded to a distinct head group. Four lines of evidence suggest that this head group is an α-configured galactose: (i) The molecular formula is consistent with a glycosphingolipid bearing a hexose as a head group. (ii) MS/MS analysis reveals a fragment that is consistent with the elimination of a hexose head group from a ceramide base ([M-H]− m/z: calculated 536.5048, observed 536.5055). (iii) The 1H NMR spectrum shows an anomeric proton with a chemical shift of 4.64, consistent with an α-linkage. (iv) Chemically synthesized α-galactosylceramide, prepared by selective α-galactosylation of the B. fragilis dihydroceramide base (see S1.10 in Supporting Information S1), has a 1H NMR spectrum indistinguishable from that of peak 3 (see S4.2 in Supporting Information S1). We term this novel glycosphingolipid B. fragilis α-galactosylceramide (α-GalCerBf) (Figure 1B). α-GalCerBf, CPE, and the ceramide base were each purified as an inseparable mixture of varying lipid chain length. This inseparable mixture of alpha-galactosylceramides, hereafter “purified α-GalCerBf,” was the material used for the immunological experiments described below.
α-GalCerBf is a close structural relative of the sponge-derived α-galactosylceramide agelasphin-9b (Figure 1B) [21]; aside from α-GalCerBf and the sponge-derived agelasphins, no naturally occurring α-galactosylceramides have ever been discovered. Substantial data have accumulated suggesting that α-GalCer is a ligand for a subset of human and mouse T cells, termed invariant natural killer T cells (iNKT), which express a conserved T cell receptor (TCR) that recognizes glycolipids presented by the major histocompatibility complex class I-like molecule, CD1d [22]. A synthetic derivative of agelasphin-9b termed KRN7000 (Figure 1B) is the prototypical agonist of iNKT cells and has become a critically important reagent for studying NKT cell biology both in vitro and in vivo. Indeed, iNKT cells are often identified or isolated by flow cytometry on the basis of their ability to bind a synthetic tetramer of CD1d loaded with a derivative of KRN7000. A variety of iNKT cell ligands have been described. One class consists of low-affinity host-derived self-ligands such as isoglobotrihexosylceramide and β-glucopyranosylceramide [23],[24]. Another class includes glycolipids from bacterial species including GSL-1 from Sphingomonas, BbGL-II from Borrelia, and a family of diacylglycerol-containing glycolipids from Streptococcus pneumoniae, all of which have been postulated to be naturally occurring ligands for CD1d [25]–[27]. It has also been proposed that liver infection by Novosphingobium aromaticivorans, a close relative of Sphingomonas that produces CD1d-binding sphingolipids, results in an NKT-cell-dependent autoimmune response against the liver and bile ducts [28].
Based on the striking chemical similarity of α-GalCerBf to KRN7000, we reasoned that α-GalCerBf might serve as an endogenous ligand for CD1d and stimulate iNKT cell activity. To test our hypothesis, we began by loading synthetic mouse CD1d tetramers with purified α-GalCerBf and determining the ability of the sphingolipid/CD1d-tetramer complex (hereafter “tetramer”) to stain two iNKT-cell-derived hybridomas [29],[30]. As with KRN7000, the α-GalCerBf-loaded tetramer (but not an empty tetramer) bound both hybridomas but not a CD4+ MHCII restricted hybridoma reactive to GFP (GFP-36) (manuscript in preparation, Yadav and Bluestone), indicating that the tetramer staining was ligand- and TCR-specific (Figure 3A; Figure S2 in Supporting Information S1). The iNKT cell hybridomas tested produced IL-2 in response to both the marine-sponge-derived and B. fragilis-derived sphingolipids in a dose-dependent manner and in absence of antigen presenting cells (APCs). These results suggested that α-GalCerBf is a stimulatory ligand that directly activates iNKT cells in vitro (Figure 3B–C; Figure S3 in Supporting Information S1).
We next examined the ability of purified α-GalCerBf to stimulate freshly isolated mouse and human iNKT cells in vitro and in vivo. Liver mononuclear cells, 30%–50% of which are NKT cells, were incubated with splenocytes as APCs in the presence of increasing doses of α-GalCerBf and examined for IFN-γ production. α-GalCerBf induced IFN-γ in a dose-dependent and CD1d-dependent manner. The response was inhibited completely by anti-CD1d antibodies (Figure 3D), consistent with our previous result that NKT cell stimulation required ligand presentation by CD1d (Figure 3B).
To explore whether the response of NKT cells to α-GalCerBf is conserved in humans, we determined whether Vα24+ cells could be expanded in vitro with purified α-GalCerBf as previously described for KRN7000 [31]. We cultured peripheral blood mononuclear cells (PBMCs) from six independent donors with 0.1 µg/ml KRN7000, 1 µg/ml α-GalCerBf, or 1 µg/ml ceramideBf for 13 d and assessed the presence of CD3+Vα24+ cells by flow cytometry (Figure 3E–F). PBMCs cultured with KRN7000 or α-GalCerBf showed an expansion of a population of CD3+Vα24+ cells, while PBMCs left untreated or treated with ceramideBf did not show an expansion of this population. Importantly, this result shows that the activity of α-GalCerBf is specific and not due to a contaminant of the lipid purification process since ceramideBf, which was purified in a similar manner, did not exhibit this effect. These results demonstrate that α-GalCerBf has similar activities in murine and human NKT cells and binds human CD1d.
To test whether α-GalCerBf can activate iNKT cells in vivo, mice were immunized with BMDCs pulsed with LPS alone or LPS + purified α-GalCerBf [32]. Consistent with activation, iNKT cells isolated from the liver showed upregulation of the cell surface markers CD25 and CD69 (Figure 3G), 15% of these liver-resident iNKT cells expressed IFN-γ after treatment (Figure 3H), and elevated IFN-γ levels were observed in the serum of these mice (Figure 3I). Anti-CD1d blocking antibodies inhibited liver iNKT cell activation and IFN-γ production, demonstrating the specificity of iNKT cell activation (Figure 3G–I). We therefore conclude that α-GalCerBf is capable of stimulating iNKT cell activation and cytokine production in vivo.
The marine sponge-derived agelasphins and the nonphysiological CD1d ligand KRN7000 have been the basis for numerous studies over the last two decades implicating iNKT cells in immunity (“α-galactosylceramide” has 3,290 citations in Google Scholar, 5/29/12). Unlike the pathogens from which CD1d ligands have previously been isolated, Bacteroides is extraordinarily prevalent in the human population, comprising >50% of the trillions of cells in the gut community of a typical human [33]. By showing that B. fragilis produces the only known α-galactosylceramide other than the sponge-derived agelasphins, and demonstrating that α-GalCerBf binds to CD1d and activates iNKT cells in vitro and in vivo, our results suggest a physiological basis for the activity of KRN7000. It is tempting to speculate that CD1d and iNKT cells function in the context of a microbiota–host interaction, especially in light of a recent report showing that neonatal colonization of germ-free mice by a conventional microbiota downregulates the level of iNKT cells in the colonic lamina propria and lung [34]. Indeed, it has been hypothesized that the agelasphins are not produced by Agelas mauritianus, but instead by a bacterial symbiont that inhabits the sponge [22].
In an attempt to determine the in vivo effect α-GalCerBf on NKT cells, we colonized germ-free (GF) mice with WT or sphingolipid-deficient B. fragilis by gavage and measured the percent and activation status of NKT cells in the liver and spleen. Colonization was confirmed by fecal cultures and PCR. We varied the length of colonization (1, 3, 4, and 14 d), the mice's age at the time of colonization (4 and 8 wk old), sex, and strain (Swiss Webster and C57BL/6). Several of these experiments indicated an expansion of NKT cells mice colonized by WT but not mutant B. fragilis. However, the effect was inconsistent and the levels of NKT cells in our control mice—germ-free (GF) and specific-pathogen-free (SPF)—fluctuated widely. As a percentage of total liver lymphocytes in the GF mice, NKT cells (CD3+tetramer+) varied between 8% and 48%, making it difficult to draw any conclusions about differences in NKT cell number or activation markers between our experimental data points.
Blumberg and coworkers recently showed that GF mice have increased levels of NKT cells in the colon compared to SPF mice and that colonization of neonatal, but not adult, GF mice with microbiota from SPF mice can reverse this effect [34]. Interestingly, neither the increase nor the reversal after colonization is seen in the liver or the spleen and there were no changes in the activation status of NKT cells. Taken together, our results suggest that the microbiota may affect NKT cells in the colon but not the liver or spleen, and that interventions to change the numbers of NKT cells must occur very early in life and may take weeks to be evident. Although Blumberg and coworkers showed the effects of the microbiota on NKT cell numbers and morbidity in models of IBD and allergic asthma, they did not identify the strain or the molecular pathway responsible for these effects; our results raise the possibility that α-GalCerBf, produced by B. fragilis, may be at least partially responsible for the results seen in their models.
There are subtle but important differences between KRN7000 and α-GalCerBf, indicating that the natural ligands for CD1d may be less potent than KRN7000. The principal structural differences between α-GalCerBf and KRN7000 are (i) a shorter N-acyl chain bearing a hydroxyl group on the β- rather than the α-carbon, (ii) the absence of a hydroxyl group at C4 of the sphinganine base, and (iii) iso-branched lipid termini (Figure 1B). Synthetic derivatives of KRN7000 that either have shorter N-acyl chains or lack a C4 hydroxyl group have been shown to have less potent activity and/or an altered cytokine response, an effect that might be due to a change in the conformation of the CD1d–lipid complex [35]. Notably, one of the iso-branched lipid termini of α-GalCerBf is shared with agelasphin 9b. Since iso-branched lipids are commonly associated with specific bacterial genera (for example, comprising 55%–96% of the total fatty acid pool in Bacteroides) [36], their presence in agelasphin 9b is consistent with a bacterial origin for these sponge-derived sphingolipids.
The absence of CPE, dihydroceramide, and α-GalCerBf from the Δ2461 mutant confirms that BF2461 is involved in B. fragilis sphingolipid biosynthesis, marking the first known member of the Bacteroides sphingolipid pathway (Figure 4). BF2461 is widely conserved among human-associated genera of Bacteroidales including Bacteroides, Parabacteroides, Porphyromonas, and Prevotella (known sphingolipid producers) but absent from Alistipes (a nonproducer), supporting its role in the bacterial sphingolipid pathway. Our inability to construct a deletion mutant of BF2462 prevents us from exploring its potential role in the pathway, though it is tempting to speculate that it generates dihydrosphingosine-1-phosphate from dihydrosphingosine. Although the later steps of the pathway remain unclear, the intermediacy of dihydroceramide is supported by the fact that CPE and α-GalCerBf share a common C34 scaffold and by our direct observation of dihydroceramide production by B. fragilis. On the basis of these observations, we propose a model of Bacteroides sphingolipid biosynthesis that closely mirrors the eukaryotic pathway (Figure 4). Given that sphingolipids comprise ∼30% of total cellular lipids and Bacteroides lacks an endoplasmic reticulum (the site of eukaryotic sphingolipid synthesis), the regulation of this pathway in the context of lipid metabolism and the localization of its biosynthetic enzymes will be important areas to explore.
Detailed methods are provided in Supporting Information S1.
Primer sequences are listed in Table S1 in Supporting Information S1. DNA fragments flanking BF2461 were PCR amplified from B. fragilis NCTC9343 using the following primers: LF_5′; LF_3′; RF_5′; RF_3′. These fragments were digested with SstI and MluI and cloned into the SstI site of pNJR6. The resulting plasmid was introduced into B. fragilis NCTC9343 by conjugation, and cointegrates were selected using erythromycin. Cointegrates were passaged, plated on nonselective medium, and replica plated to medium containing erythromycin. Erythromycin-sensitive colonies were screened by PCR to detect those acquiring the mutant genotype.
B. fragilis NCTC9343 was cultured under standard conditions, and harvested cells were extracted with CHCl3∶MeOH (2∶1). The organic extract was subjected to alkaline hydrolysis, neutralized, and extracted with CHCl3∶MeOH (2∶1). The crude extract was purified by preparative TLC (CHCl3∶MeOH∶H2O, 65∶25∶4) to give α-GalCerBf (Rf = 0.6). For complete experimental details, including yields and full characterization (NMR, high-resolution mass spectrometry) of all compounds, see Supporting Information S1. α-GalCerBf was isolated in five independent batches, and the in vitro and in vivo experiments were repeated with different batches of purified compound.
α-GalCerBf, CPE, and the ceramide base were each purified as an inseparable mixture of varying lipid chain length. Mass spec analysis of the methanolyzed long chain base (LCB) (S4.6 in Supporting Information S1) suggests that this portion of the structure carries the variation (see next paragraph). The inseparable mixture of alpha-galactosylceramides (>95% pure), referred to as “purified α-GalCerBf,” was the material used for the immunological experiments.
Methanolysis of ceramideBf produced a mixture of three LCB amines that could be separated and analyzed by HPLC-MS (S4.6 in Supporting Information S1). Analysis of each by HRMS indicated that they are structural variants that differ in tail chain length. These data suggest the major parent α-GalCerBf variants (m/z 716.57, m/z 730.58, and m/z 744.60) also differ in chain length of the LCB.
For dose titration experiments, BMDCs and DN3A4-1.2 and N38-2C12 NKT hybridomas (M. Kronenberg) and GFP36 CD4+ hybridoma were cultured at a 3∶1 hybridoma∶BMDC ratio and the indicated doses of KRN7000 or α-GalCerBf in the presence of 1 µg/ml LPS. Supernatants were harvested after 24 h and IL-2 production was measured by ELISA. For APC-free experiments, CD1d monomers were coated on a 96-well plate for 1 h, and wells were blocked with PBS/10% FBS. The indicated amount of α-GalCerBf was added to each well and incubated at 37°C for 3 h. After washing unbound α-GalCerBf, hybridomas were added. Supernatants were harvested after 16–18 h and IL-2 production was measured by ELISA. For in vitro CD1d blocking experiments, α-GalCerBf pulsed BMDCs were cultured at a 3∶1 hybridoma∶BMDC ratio in the presence of 10 µg/mL anti-CD1d antibody (Clone 1B1, BD Pharmingen). Supernatants were harvested after 16–18 h and IL-2 production was measured by ELISA.
For blood draws from healthy donors, informed consent was obtained in accordance with approved University of California, San Francisco IRB policies and procedures (IRB 10-02596). PBMCs were cultured for 13–14 d in RPMI containing 10% autologous serum plus lipids as described in Figure 3. On day 1 of culture, 100 U/ml hIL-2 was added. Cultures were harvested on day 13 or 14 and the percentage of CD3+Vα24+ NKT cells was determined by flow cytometry after staining with CD3 and 6B11 antibodies.
Mice were sacrificed 16–18 h after transfer of 0.4×106 mature CD86hiMHCIIhi BMDCs. Livers were cut into small pieces and passed through a stainless mesh. Cells were resuspended in 40% Percoll solution (GE Healthcare), underlaid with 60% Percoll solution, and centrifuged at 2,300 rpm for 20 min at room temperature. All isolations were performed in the presence of brefeldin A (Sigma). After cell surface staining, cells were fixed in Cytofix/Cytoperm (BD Biosciences) according to the manufacturer's instructions and stained for intracellular cytokines. Serum IFN-γ was measured by ELISA.
|
10.1371/journal.pbio.2004874 | mDia1/3 generate cortical F-actin meshwork in Sertoli cells that is continuous with contractile F-actin bundles and indispensable for spermatogenesis and male fertility | Formin is one of the two major classes of actin binding proteins (ABPs) with nucleation and polymerization activity. However, despite advances in our understanding of its biochemical activity, whether and how formins generate specific architecture of the actin cytoskeleton and function in a physiological context in vivo remain largely obscure. It is also unknown how actin filaments generated by formins interact with other ABPs in the cell. Here, we combine genetic manipulation of formins mammalian diaphanous homolog1 (mDia1) and 3 (mDia3) with superresolution microscopy and single-molecule imaging, and show that the formins mDia1 and mDia3 are dominantly expressed in Sertoli cells of mouse seminiferous tubule and together generate a highly dynamic cortical filamentous actin (F-actin) meshwork that is continuous with the contractile actomyosin bundles. Loss of mDia1/3 impaired these F-actin architectures, induced ectopic noncontractile espin1-containing F-actin bundles, and disrupted Sertoli cell–germ cell interaction, resulting in impaired spermatogenesis. These results together demonstrate the previously unsuspected mDia-dependent regulatory mechanism of cortical F-actin that is indispensable for mammalian sperm development and male fertility.
| Paternal genetic information is transmitted to the offspring via sperm. The unique cell morphology of the sperm plays essential roles in sperm transport through the female reproductive tract and in fertilization with oocytes. Sertoli cells are somatic cells located in the seminiferous tubules of the testis and are known to contribute to the development of sperm. While many studies have analyzed sperm development, the mechanisms underlying its morphogenesis remain obscure. In this work, we showed that the interaction between developing sperm and Sertoli cells is critical for sperm morphogenesis. We further unraveled that this interaction is strongly dependent on the cortical F-actin meshwork and contractile actomyosin bundles of Sertoli cells, and that two actin polymerization and nucleation factors of the formin family, mDia1 and mDia3, are involved in the generation of both actin-based structures. Loss of these formins in mice result in disrupted Sertoli cell actin structures, abnormal sperm morphology, and male infertility. We conclude that mDia1 and mDia3 play a role in sperm development through the regulation of the actin cytoskeletal architecture of Sertoli cells and that defects in these proteins might contribute to male infertility.
| Filamentous actin (F-actin) nucleation and polymerization are controlled by actin nucleators, including actin-related protein2/3 (Arp2/3) and formin, in mammalian cells. Previous studies in typical cultured cell lines showed that Arp2/3 generates branched actin filaments, while formin generates straight actin filaments [1]. However, F-actin structure and dynamics regulated by formin in a variety of mammalian cell types and their physiological roles in vivo remain largely unknown. Mammalian diaphanous homolog (mDia) proteins, mammalian homologues of Drosophila diaphanous, belong to the formin family of proteins and consist of three isoforms in mammals, namely mDia1, mDia2, and mDia3 [2]. To unravel mDia-dependent F-actin structures in the mammalian body and explore their physiological functions in vivo, we generated mice deficient in each isoform and analyzed their phenotypes [3–5]. These studies showed the functional redundancy between mDia1 and mDia3 isoforms and demonstrated that mDia1/3-mediated F-actin is critical for neuroblast migration [5] and neuroepithelium integrity [6] in the developing brain, and mediates presynaptic plasticity of mature neurons in the adult brain [7]. However, the contribution and function of mDia-mediated F-actin in other systems remain an open question.
Spermatogenesis is a process by which spermatogonial stem cells give rise to spermatozoa through spermatocytes, round spermatids, and elongated spermatids in seminiferous tubules. Besides these germ cells, the seminiferous tubule contains a somatic constituent of the seminiferous epithelium, the Sertoli cells. Sertoli cells make adhesion with developing germ cells and provide them with structural and functional supports [8], which are indispensable for normal spermatogenesis. Adhesion between Sertoli cells and round spermatids is mediated by the adherens junction (AJ) [9]. In addition, a special form of cell–cell adhesion, the apical ectoplasmic specialization (ES) junction, is formed between Sertoli cells and spermatids in association with the elongation of the round spermatid [10,11]. In contrast to the AJ, which is a structure associated with contractile actomyosin bundles [12], the apical ES junction is typified by a layer of densely packed noncontractile F-actin bundles [13] concentrated immediately beneath the Sertoli cell plasma membrane. It is known that these noncontractile F-actin bundles are associated with actin bundling protein espin1 [14] and support elongated spermatid adhesion and orientation through the regulation of the apical ES junction [15,16]. Although the structure and function of AJ and ES junctions have been extensively investigated, how F-actin structures associated with these junctions in Sertoli cell are formed, interact with each other, and are maintained during spermatogenesis are largely obscure.
F-actin makes several different forms of architectures beneath the cell membrane, which are collectively termed cortical F-actin [17] and are determined by the environment surrounding the cell. For example, in cells cultured in suspension under steady state, F-actin in the cortex makes a highly dense and fine mesh structure with a homogeneous pore size of approximately 30 nm [18]. It was previously shown that the Arp2/3 complex dominantly nucleate these meshlike cortical actin filaments [18,19], although about 10% of the cortical F-actin of these cells was shown to be formin dependent [19]. On the other hand, cortical F-actin structures in adherent cells are more complicated because dynamic remodeling of cortical network results in additional F-actin architectures, such as the contractile stress fibers [20]. Stress fibers are mostly localized at the basal plane of the cell and are more visible than those finely organized filaments, making conventional imaging of cortical F-actin very challenging. Indeed, the structure and dynamics of cortical F-actin of the adherent cell remain largely unexplored to date, and how and which actin nucleators regulate them are unknown.
In this study, we show that both mDia1 and mDia3 are dominantly expressed in the Sertoli cells of mouse testes and are indispensable for sperm development and male fertility. Utilizing superresolution microscopy, we resolved the nanoscale F-actin architecture of Sertoli cells attached to the substrate and observed a previously unsuspected F-actin meshwork with a large mesh size of about 100 nm. Live imaging of F-actin has further revealed the highly dynamic nature of this cortical F-actin meshwork of the Sertoli cell. Moreover, genetic and pharmacological experiments have shown that this Sertoli cell cortical F-actin meshwork depends on actin nucleation and polymerization activity of formins, including mDia1/3, but not Arp2/3. Intriguingly, these mDia1/3-dependent actin filaments are required for the generation of contractile actomyosin bundles in Sertoli cells. Consequently, loss of mDia1/3 in Sertoli cells results in severe reduction of cortical F-actin meshwork and actomyosin bundles and, instead, unexpectedly causes ectopic formation of espin1-associated noncontractile F-actin bundles in the cell. Finally, we demonstrate that mDia-dependent cortical F-actin meshwork and contractile actomyosin in Sertoli cells are together critical for interaction with germ cells, which is required for their competency to support spermatogenesis.
When breeding mice in our colony, we found that, while mDia1 knockout (KO) male mice [3] and mDia3 KO male mice [5] were fully fertile, the mating of mDia1/3 double knockout (DKO) male mice [10] yielded no offspring upon mating. The fertilization rate upon in vitro fertilization (IVF) of wild-type (WT) oocytes with mDia1/3 DKO sperms was also greatly reduced, suggesting abnormality in the sperm (S1A and S1B Fig). Analysis of the epididymis of 8-wk-old male mDia1/3 DKO mice further revealed reduced number, abnormal morphology, and impaired motility of mDia1/3 DKO sperm compared with the control WT mice (S2A–S2I Fig). These results together suggest that the male infertility phenotype of mDia1/3 DKO mice is likely due to impaired sperm development.
To investigate how mDia1/3 double deficiency causes sperm abnormalities, we next performed histological analysis on spermatogenesis in the testes of adult WT and mDia1/3 DKO mice. Paraffin sections of testis with Periodic acid-Schiff (PAS) staining revealed that, while numerous elongated spermatids aligned at the adluminal compartment of the seminiferous tubule in WT testes, the number of elongated spermatids in mDia1/3 DKO testes is extremely low compared with WT testes (Fig 1A, red boxes). Notably, the heads of mDia1/3 DKO sperm were not properly oriented toward the basal lamina of the seminiferous tubule (Fig 1A, black arrows in the lower red box). In addition, a fraction of mDia1/3 DKO elongated spermatids exhibited ectopic localization in seminiferous tubules, as evidenced by their presence at the proximity of basal lamina (Fig 1A, black boxes). Moreover, we found a significant increase in apoptotic cell number in the mDia1/3 DKO seminiferous tubule by TUNEL staining (S3A and S3B Fig). These results together suggest that mDia1/3 is indispensable for spermatogenesis.
We next attempted to figure out whether the mDia1/3 deficiency in germ cells or Sertoli cells is responsible for the above abnormal spermatogenesis phenotype observed in mDia1/3 DKO mice. We first examined the expression of mDia1 and mDia3 in the seminiferous tubule by immunofluorescence staining. We found positive signals for mDia1 and mDia3 in vimentin-positive Sertoli cells [21] in the seminiferous tubule from adult WT mice (Fig 1B). These mDia1 and mDia3 signals were absent in testis sections from mDia1 or mDia3 knockout (KO) mice (S4A and S4B Fig), confirming the specificity of these signals. We also examined the expression of mDia3 in the WT seminiferous tubule during the spermatogenic cycle [22] and compared its localization with F-actin (S5A Fig). We found that mDia3 is expressed in Sertoli cells throughout the spermatogenic cycle, though its staining intensity and localization varies in different stages. For example, in stages IV–V, although the majority of mDia3 staining localizes in the cell body, a fraction of mDia3 staining with relatively low intensity (white arrows) colocalizes with F-actin bundles associated with apical ES. Similarly, in stages VI–VII, a fraction of mDia3 staining localizes close to F-actin bundles associated with basal ESs (white arrowheads), and the mDia3 staining becomes weak in stages X–XI. Specificity of these mDia3 signals was confirmed as they were mostly absent in testis sections from mDia3 KO mice (S5B Fig).
We then examined the functional requirement of mDia1 and mDia3 expressed in Sertoli cells for spermatogenesis by germ cell transplantation (Fig 1C–1H), according to our previously established protocol [23]. We first transplanted testicular germ cell suspensions from WT or mDia1/3 DKO mice to germ cell–free and Sertoli cell–only testes of W/Wv mice (Fig 1C), and examined spermatogenesis by hematoxylin–eosin (HE) staining of histological sections of the recipients (Fig 1D). We found that W/Wv mouse testes transplanted with mDia1/3 DKO testicular cell suspensions had normal spermatogenesis, as did W/Wv mouse testes transplanted with WT testicular cell suspensions (Fig 1D). Quantification results revealed no significant difference of the spermatozoa number per seminiferous tubule in these transplanted mice (Fig 1E). Thus, mDia1/3 expression in germ cells is dispensable for spermatogenesis. As the reciprocal approach, we depleted germ cells of WT or mDia1/3 DKO mice with busulfan treatment and transplanted testicular germ cell suspensions from acrosin/actin-enhanced green fluorescent protein (acro/act-EGFP) transgenic mice to each Sertoli cell–only testis (Fig 1F). We found that acro/act-EGFP testicular germ cells transplanted to busulfan-treated mDia1/3 DKO testes mimicked abnormal sperm phenotypes seen in mDia1/3 DKO testes, including a reduced number of elongated spermatid (Fig 1G, red box), ectopic localization near the basal lamina of the seminiferous tubule (Fig 1G, black box), and abnormal shape of the head (Fig 1G, black box, yellow arrowhead). Quantification revealed a significant decrease in the spermatozoa number per seminiferous tubule of busulfan-treated Sertoli-only mDia1/3 DKO testes compared with WT testes (Fig 1H). Therefore, mDia1/3 function in Sertoli cells is required for normal spermatogenesis.
Given the critical function of mDia1/3 expression in Sertoli cells, we then used confocal microscopy and examined F-actin structure in primary Sertoli cells cultured on gelatin-coated cover glass. We found that actin filaments in control WT Sertoli cells are composed of thick F-actin bundles and faintly stained actin filaments between these bundles (S6A Fig, left, and magenta line scanning). On the other hand, in mDia1/3 DKO Sertoli cells, F-actin bundles were apparently retained, but the staining intensity of faintly stained actin filaments between these bundles was greatly reduced (S6A Fig, right, and green line scanning). To further resolve the architecture of faintly stained actin filaments in Sertoli cells at the nanoscale level, we next examined phalloidin-stained control WT cells with total internal reflection (TIRF) three dimensional-N-stochastic optical reconstruction microscopy (3D-N-STORM) superresolution microscopy [24]. We found that the faintly stained actin filaments between thick F-actin bundles observed in the cortices of WT cells by confocal microscope are actually a meshwork of actin filaments with a relatively large pore size of about 100 nm (Fig 2A, upper, white box). We then next examined the actin cytoskeleton organization of mDia1/3 DKO Sertoli cells and found that the above cortical F-actin meshwork was markedly sparser, and there were areas devoid of actin filaments in these cells (Fig 2A, below, white box). Quantitative analysis of STORM images revealed that the occupancy of F-actin meshwork was significantly reduced in mDia1/3 DKO Sertoli cells (Fig 2B). These results suggest that mDia1/3 contribute to the formation and maintenance of cortical F-actin meshwork in Sertoli cells.
We next sought to analyze the relative contribution between formins and Arp2/3 on cortical F-actin meshwork in Sertoli cells. To this end, we utilized pharmacological inhibitors for formins and Arp2/3. We found that treatment of WT cells with small molecule inhibitor of formin homology 2 domain (SMIFH2), a formin inhibitor [25], reduced cortical F-actin meshwork similarly as was observed in mDia1/3 DKO cells (Fig 2C, third row). Notably, a fraction of SMIFH2-treated Sertoli cells showed total suppression of cortical F-actin meshwork with concomitant decrease in F-actin bundles (Fig 2C, fourth row). On the other hand, treatment of WT cells with CK-666, an Arp2/3 inhibitor [26], did not suppress the cortical F-actin meshwork (Fig 2C, second row). Quantitative analysis (Fig 2D) confirmed that treatment with SMIFH2 significantly suppressed cortical F-actin meshwork in Sertoli cells and additionally revealed that treatment of CK-666 resulted in a slight but significant increase of cortical F-actin meshwork. We speculate that the latter might be due to the increased availability of actin monomers for formins when Arp2/3-mediated actin polymerization is suppressed. These results together therefore suggest that the cortical F-actin meshwork in Sertoli cells is a structure dependent on the actin polymerization activity of formins, including mDia1 and mDia3, but not Arp2/3.
We next examined the dynamics of actin filaments in living primary cultured Sertoli cells. To this end, we introduced LifeAct-EGFP, a probe for F-actin [27], into primary cultured Sertoli cells. Utilizing spinning disk superresolution microscopy (SDSRM) [28], which allows fast image acquisition at the spatial xy-axis resolution of approximately 120 nm, we found that, while thick F-actin bundles are a relatively static structure, the cortical F-actin meshwork is an extremely dynamic structure, in which actin continually polymerizes and depolymerizes underneath the cortex of WT control cells (S1 Movie). Quantification of the dynamics of cortical meshwork actin filaments of WT Sertoli cells revealed that the F-actin elongated for a distance of several micrometers, and their extension rate was very fast, with a mean of 0.81 ± 0.06 μm/s (about 300 globular actin (G-actin) subunits incorporated per second) (Fig 3A and 3B, S2 Movie). The speed distribution further revealed at least two subpopulations of cortical meshwork actin filaments with different polymerization rates; the first subpopulation has a peak polymerization rate around 0.4 μm/s and the other subpopulation around 1.3 μm/s (Fig 3B). It should also be noted that F-actin elongation was highly straight (Fig 3C). On the other hand, in mDia1/3 DKO cells, although the straightness of elongated F-actin was not affected (Fig 3C), the subpopulation of nascent actin filament with polymerization rate peak around 1.3 μm/s was absent (Fig 3B), and the number of elongation events was significantly reduced (Fig 3D). These results together suggested that cortical F-actin meshwork is a highly dynamic structure and its formation is dependent on mDia1 and mDia3.
Given the extremely rapid rate of actin polymerization observed in the cortical F-actin meshwork area and its impairment in mDia1/3 DKO Sertoli cells, we hypothesized that actin filament polymerization in the meshwork is driven by mDia1 and mDia3. To test this hypothesis, we carried out single-molecule speckle imaging [29] of EGFP-mDia3 in living primary cultured WT Sertoli cells. The EGFP-mDia3 used in the experiment rescued the reduced cortical actin filament meshwork phenotype of mDia1/3 DKO Sertoli cells (S7A–S7C Fig). Single-molecule speckle imaging with TIRF microscopy revealed fast and directional molecular movement of EGFP-mDia3 in the cortex of Sertoli cells (Fig 3E, S3 Movie and S4 Movie). Kymograph (Fig 3F) and quantification analyses of EGFP-mDia3 single-molecule speckles further indicated that the average speed of EGFP-mDia3 was extremely fast, with an average rate of 1.38 ± 0.06 μm/s (Fig 3G). In addition, the EGFP-mDia3 molecular movement was highly linear, similar to cortical meshwork actin filament (Fig 3H). Moreover, a single molecule of EGFP-mDia3 traveled for long distance—on average, 4.46 ± 0.40 μm (Fig 3I). Given that the spatial localization of a single molecule of EGFP-mDia3 in the cell cortex was similar to the cortical F-actin meshwork and the linear directional movement, long travel distance and fast movement speed of an EGFP-mDia3 single molecule were correlated with LifeAct-EGFP dynamics, and that the latter was diminished in the absence of mDia1 and mDia3, we concluded that cortical F-actin meshwork is an actin filament structure directly generated by mDia1/3 in primary cultured Sertoli cells. These data further substantiated that cortical F-actin meshwork is a mDia1/3-dependent structure.
In addition to the actin meshwork, thick actin bundles were seen in WT Sertoli cells and apparently similar actin bundles were observed also in mDia1/3 DKO Sertoli cells (Fig 2A). We therefore examined the origin and composition of these actin bundles. Higher magnification of TIRF-N-STORM images revealed that thick F-actin bundles of Sertoli cells are structurally continuous with the cortical F-actin meshwork both in WT and mDia1/3 DKO Sertoli cells (Fig 4A, left and middle). However, we noted that whereas these thick F-actin bundles were mostly intact in WT cells (Fig 4A, left), those in mDia1/3 DKO cells were occasionally bent (Fig 4A, right). Utilizing confocal microscopy, we found that most of F-actin bundles in primary cultured WT Sertoli cells were co-stained for phosphorylated myosin light chain (pMLC) (Fig 4B, top) and thus represent contractile actomyosin bundles [30]. On the other hand, the pMLC staining intensity in F-actin bundles was greatly reduced in mDia1/3 DKO Sertoli cells (Fig 4B, bottom). Quantification of pMLC staining showed a significant decrease in mDia1/3 DKO cells (Fig 4C). To identify actin bundles observed in mDia1/3 DKO seminiferous tubules, we next performed immunostaining for espin1, an actin bundling protein associated with noncontractile F-actin bundles [31]. We found that most of F-actin bundles in primary cultured Sertoli cells from mDia1/3 DKO mice were associated with espin1 (Fig 4D). Quantification further revealed that F-actin bundles associated with espin1 comprised about 50% of total F-actin bundles in mDia1/3 DKO Sertoli cells but were rarely observed in WT Sertoli cells (Fig 4E). These results together indicate that mDia1/3 are indispensable not only for the formation and maintenance of the cortical F-actin meshwork but also for the generation of contractile actomyosin bundles in Sertoli cells.
To examine how our findings on F-actin structures in Sertoli cells cultured in vitro may reflect those in vivo, we next investigated the F-actin structures in Sertoli cells in intact seminiferous tubules. To this end, we selectively fluorescently labeled Sertoli cells in vivo by microinjection of lentivirus-expressing LifeAct-EGFP to seminiferous tubules, according to our method reported previously [32]. We then dissected and untangled the lentivirus-injected seminiferous tubules, fixed and prepared them for imaging ex vivo [33]. Deconvoluted Z-stack images with spinning disk confocal microscopy showed that, although the morphology of intact Sertoli cells was different from that of cultured Sertoli cells, their F-actin structures imaged with LifeAct-EGFP consisted of at least two different F-actin structures, the thick F-actin bundles and thin F-actin meshwork (S8 Fig, S5 Movie). These results suggest that the overall F-actin architectures of Sertoli cells in intact seminiferous tubules are largely similar to those found in cultured Sertoli cells.
Recent studies suggested an important role of contractile actomyosin bundles in the regulation of cell–cell AJ [34,35]. Given that actomyosin bundles were impaired in primary cultured mDia1/3 DKO Sertoli cells, we next utilized an in vitro reconstitution experiment and investigated the impact of the loss of mDia1/3 on the formation of the AJ between a Sertoli cell and a round spermatid cell (Fig 4F). In this system, germ cells isolated from the approximately 3-wk-old EGFP transgenic mice testes containing mostly round spermatids were added onto the primary cultured Sertoli cells and further cocultured for 24 h to allow the formation of Sertoli cell–round spermatid AJ. Immunocytochemistry of an AJ protein, N-cadherin, revealed that while approximately 50% of the round spermatids on WT primary Sertoli cells formed N-cadherin positive cell–cell adhesion, less than 20% of germ cells formed N-cadherin positive cell–cell adhesion with mDia1/3 DKO primary Sertoli cells (Fig 4G and 4H). In addition, we found that the continuity of F-actin on the round spermatid–Sertoli cell interface was often disrupted in the mDia1/3 DKO Sertoli cell (Fig 4I and 4J). Quantitative analysis further revealed that the density of the cortical F-actin meshwork beneath the germ cell in WT Sertoli cells was compromised in mDia1/3 DKO Sertoli cells (Fig 4I and 4K). These results together suggest that mDia1/3-mediated F-actin is critical for formation of the AJ between Sertoli cells and round spermatids.
Because the above findings showed the role of mDia1/3 in the formation of AJ between Sertoli cells and germ cells in vitro, we next examined if the mDia1/3 deficiency in Sertoli cells affects these cell–cell interactions in the seminiferous tubule in vivo. To this end, we performed immunostaining for nectin-2, an adhesion molecule exclusively expressed in Sertoli cells [36,37]. Composite Z-stack images of nectin-2 in WT seminiferous tubules showed nectin-2 signals at the AJ between a Sertoli cell and a round spermatid (see the area shown by the white two-way arrow in Fig 5A, WT) and concentration of nectin-2 at the apical ES junction between Sertoli cells and elongated spermatids and basal ES between Sertoli–Sertoli cells (Fig 5A). On the contrary, nectin-2 signal at the boundary between Sertoli cells and round spermatids was significantly reduced in mDia1/3 DKO seminiferous tubules (Fig 5A, below). We also stained for nectin-2 in WT and mDia1/3 DKO seminiferous tubules throughout the spermatogenic cycles and observed consistent results (S9 Fig). Moreover, high magnification images (Fig 5B) revealed that the accumulation (Fig 5C) and the continuity (Fig 5D) of concentrated nectin-2 signals at the apical ES junction were severely impaired in the mDia1/3 DKO seminiferous tubule. These results indicate that mDia1/3 are indispensable for both AJ and apical ES junction in the seminiferous tubule.
It is known that while conventional AJs are associated with contractile actomyosin bundles [34,35], ES junctions are typified by the presence of highly dense noncontractile F-actin bundles in the Sertoli cell [13,14]. To further resolve the F-actin architecture of WT and mDia1/3 DKO seminiferous tubules, we next conducted phalloidin staining and examined F-actin structures by confocal imaging (Fig 5E). Composite Z-stack images of F-actin in WT seminiferous tubules show that, while thick F-actin bundles specifically localize to the apical ES junction along the elongated spermatid head (Fig 5E, above, red box) and basal ES at this stage (Fig 5E, above, yellow box), thick F-actin bundles were intriguingly increased in number and intensity and aberrantly distributed over the whole region of the mDia1/3 DKO seminiferous tubule (Fig 5E, below). Moreover, whereas the background F-actin staining in the area where round spermatids are located as shown by the white two-way arrow, is relatively high in WT seminiferous tubules (Fig 5E, above), they were strongly reduced between the ectopic thick F-actin bundles in mDia1/3 DKO seminiferous tubules (Fig 5E, blue box). Line scanning of the F-actin intensity in the round spermatid area (Fig 5E, blue box, magenta and green lines) clearly showed the stark difference between the faintly stained F-actin density in WT and mDia1/3 DKO Sertoli cells (Fig 5F). Based on these observations, we summarized our results in a simplified diagram (Fig 5G). Quantification revealed that the total F-actin intensity per seminiferous tubule was significantly reduced in mDia1/3 DKO seminiferous tubules compared with that of WT seminiferous tubules (Fig 5H).
To further clarify the identity of abnormal thick actin bundles specifically observed in mDia1/3 DKO seminiferous tubules, we next performed immunostaining of espin1. We found that whereas espin1 staining is relatively weak and confined to the area around the head of elongated sperm in control WT seminiferous tubules at stages X–V of the spermatogenic cycle, it showed persistent and abnormally strong signals in mDia1/3 DKO seminiferous tubules, which are mostly colocalized with ectopic F-actin bundles (Fig 5I and S10 Fig). Quantification showed that F-actin bundles associated with espin1 comprised about 60% of total F-actin bundles in mDia1/3 DKO seminiferous tubules but were rarely observed in WT seminiferous tubules (Fig 5J). These findings together suggested that the loss of mDia1/3 impaired normal F-actin structures in both AJs and apical ES junctions and induced the formation of ectopic aberrant espin1-containing F-actin bundles in seminiferous tubules.
Despite advances in our understanding of formins’ biochemical activity, how and where formins generate F-actin and what forms of F-actin structures they make in a variety of cell types and tissues remain largely unknown. Technically, the presence of abundant F-actin bundles in adherent cells makes the observation of the fine cortical actin network challenging. In this work, we found that although mDia1 KO male mice and mDia3 KO male mice were fully fertile, mDia1/3 DKO male mice were infertile, suggesting the functional redundancy between the two mDia isoforms. Starting with the Sertoli cell–intrinsic defects in spermatogenesis in mDia1/3 DKO mice, we examined how the formins mDia 1 and mDia3 contribute to actin cytoskeleton structure and dynamics in Sertoli cells and spermatogenesis. We employed superresolution microscopy, 3D-N-STORM, to unravel the nanoscale F-actin architecture in Sertoli cells and found that the actin cytoskeleton in these cells is strongly dependent on formins. We observed at least two types of actin cytoskeleton structure coexisting in primary cultured WT Sertoli cells, the thin cortical F-actin meshwork and the thick actomyosin bundles. Notably, two types of actin filaments were also seen in Sertoli cells in situ in testis tubules. High magnification superresolution images further revealed that these two types of F-actin are of continuous structure. In contrast to the previous findings on cortical F-actin meshwork in cells cultured in suspension [18, 19], we found that the cortical meshwork of Sertoli cells on the substrate has a larger pore size and is largely dependent on formin activity, including mDia1/3, but not on Arp2/3. Utilizing spinning disk superresolution live imaging [28], we further found that the cortical F-actin meshwork is a highly dynamic structure. Quantitative analysis of the cortical F-actin network in cultured Sertoli cells revealed two subpopulations in the cortical F-actin meshwork, the first subpopulation with the polymerization rate peak at about 0.4 μm/s and the second subpopulation with polymerization rate peak at about 1.3 μm/s. In mDia1/3 DKO Sertoli cells, the latter subpopulation was totally absent. Therefore, the subpopulation of newly polymerized actin with faster speed is dependent on mDia1/3 activity. Moreover, we found that the dynamic of a single mDia3 molecule in living Sertoli cells on a substrate is similar to that of the second subpopulation of cortical F-actin meshwork, with polymerization rate of about 1.3 μm/s. Given this correlation, it is likely that mDia1 and mDia3 generate a particular fraction of cortical F-actin meshwork with fast dynamics in Sertoli cells. We speculate that the other fraction of cortical F-actin meshwork is mediated by other formins expressed in Sertoli cells, such as formin1 [38], because cortical F-actin meshwork was strongly suppressed by treatment with the formin inhibitor, SMIFH2 [25].
Intriguingly, we also found that actomyosin bundles are impaired in mDia1/3 DKO Sertoli cells. In the absence of mDia1 and mDia3, noncontractile actin bundles containing espin1 replaced contractile actomyosin bundles in Sertoli cells. The mechanism by which mDia1/3-generated actin filaments are preferentially incorporated to actomyosin bundles and not to espin1-containing actin bundles remains unclear. We previously demonstrated that mDia1 helically rotates at the barbed end of the actin filament [39] and proposed that it might cause the twist of F-actin [40]. Because it was reported that the twisting of an actin filament affects its conformation and influences subsequent binding of actin binding proteins (ABPs), including myosin II [41], we speculate that mDia1 might affect the conformation of an actin filament that allows efficient binding of myosin II. It should also be noted that filament spacing in actin bundles is an architectural feature dependent on the actin bundling proteins. It was reported that the incorporation of small-sized (about 8 nm) espin1 molecules into F-actin results in compact F-actin bundling, which is segregated from F-actin containing contractile myosin of the larger size [42]. We speculate that long and straight actin filaments generated by mDia1/3 in the cortical meshwork may not be favorable for espin1 binding and instead may allow larger protein cross-linking to the formation of actomyosin bundles. A previous study reported that espin-associated F-actin bundles assemble into highly ordered, densely packed bundles [13]. On the other hand, the actomyosin bundle is a structure of parallel filaments of opposite polarity spaced apart by myosin [43,44]. Therefore, it is possible that espin-based actin bundles contain a larger number of actin filaments per bundle than actomyosin in mDia1/3 DKO Sertoli cells. Nevertheless, they are apparently more fragile than actomyosin bundles as they are occasionally bent, as observed in mDia1/3 DKO Sertoli cells in this study.
Actomyosin is a contractile structure generating force in various cellular contexts, and its importance in maintaining the cell–cell AJ has been recently reported [34,35]. Consistently, we found that formation of the AJ was impaired in the mDia1/3 DKO seminiferous tubule, as evidenced by suppressed cell adhesion between round spermatids and Sertoli cells in vitro and reduced nectin-2 staining in Sertoli cell–germ cell junction in vivo. It is plausible that mDia1/3 supply cortical actomyosin bundles to encircle spermatids that may limit the diffusion of AJ proteins and promote their clustering in Sertoli cells, as reported as a prerequisite event for cell–cell junction formation in other systems [17,45]. Notably, a small fraction of germ cells that could form cell–cell junctions with mDia1/3 DKO Sertoli cells in vitro showed impaired continuity of F-actin in the proximity of cell–cell junctions. Therefore, mDia1/3-mediated actomyosin bundles in Sertoli cells also play a role in stabilization and maintenance of cell–cell junctions. It should also be noted that in addition to AJs, apical ES junctions were also severely impaired in mDia1/3 DKO mice. As the ES junction is formed subsequent to the AJ during sperm development, we speculate that proper AJ is a prerequisite for ES junction formation. One intriguing question is how the actomyosin bundles generated by mDia1/3 localize specifically to the AJ and exert their function there. Several models were previously proposed for the generation of actomyosin bundles at the AJ [34,46]. Because we observed mDia1/3-dependent actin polymerization throughout the Sertoli cell cortex, our results support a model that mDia1/3-depedent actomyosin bundles are generated in the cortex but the localization to the AJ is mediated by other side-binding ABPs, such as α-catenin [47].
Sertoli cells make adhesion with developing germ cells and contribute to spermatogenesis by providing structural and functional support [8]. Thus, proper adhesions between Sertoli cells and germ cells are indispensable for normal spermatogenesis. Involvement of F-actin in sperm development was also previously reported [48]. However, the detailed molecular mechanism of F-actin action in spermatogenesis, especially in the supporting Sertoli cells, was largely unknown. In this work, we demonstrated that mDia1/3-dependent cortical F-actin meshwork and contractile actomyosin in Sertoli cells are critical for spermatogenesis through the regulation of Sertoli cell–germ cell adhesion.
In summary, here we have combined genetic and pharmacological manipulation of formins with superresolution and single-molecule imaging, and revealed a previously unsuspected requirement for the two formins, mDia1 and mDia3, in sperm morphogenesis and male fertility. Importantly, mDia1/3 catalyze the formation of cortical F-actin meshwork and actomyosin bundles indispensable for the formation and function of AJs and apical ES junctions essential for proper spermatogenesis (S11 Fig). The function of mDia unraveled here thus sheds a light on the importance of the F-actin structure and dynamics in Sertoli cells for spermatogenesis and paves a way for molecular dissection of its regulatory mechanisms.
Primary antibodies used were rabbit anti-mDia1 polyclonal (LifeSpan BioSciences, Seattle, WA), rabbit anti-mDia3 polyclonal (SIGMA-Aldrich, St. Louis, MO), chicken anti-vimentin polyclonal (Millipore, Burlington, MA), rabbit anti-GFP polyclonal (MBL, Nagoya, Japan), goat anti-nectin-2 polyclonal (Santa Cruz Biotechnology, Dallas, TX), rabbit anti-espin1 polyclonal (#PB538, a gift from Dr. Bechara Kachar, NIH) [49], mouse anti-espin1 monoclonal, mouse anti-N-cadherin monoclonal (BD Bioscience, San Jose, CA), and mouse anti-phospho-myosin light chain2 (Ser19) monoclonal (Cell Signaling, Danvers, MA). Inhibitors used were SMIFH2 (Tocris, Bristol, United Kingdom) and CK-666 (Tocris, Bristol, UK).
mDia1 KO mice were generated and backcrossed to C57BL/6N mice for more than 10 generations, as previously described [3]. mDia3 KO mice and mDia1/mDia3 DKO were generated and backcrossed to C57BL/6N mice for more than 10 generations, as previously described [5]. C57BL/6-Tg (CAG-EGFP) mice were purchased from the local supplier (Japanese SLC, Hamamatsu, Japan) and acro/act-EGFP [TgN(acro/act-EGFP)OsbC3-N01-FJ002] mice were kindly provided by Dr. Masaru Okabe (Osaka University). All animal care and use were in accordance with the United States National Institutes of Health Guide for the Care and Use of Laboratory Animals and were approved by the Institutional Animal Care and Use Committee of Kyoto University Graduate School of Medicine.
Testes from 12-wk-old WT or mDia1/3 DKO mice were collected in cold Hanks’ balanced salt solution (HBSS; Invitrogen, Carlsbad, CA) and placed on ice. After removal of the tunica albuginea, seminiferous tubules were dissociated from the testis and transferred immediately into 1 mL of 1 mg/mL collagenase in HBSS. Tubules were incubated for 15 min at 37°C in a water bath, then washed with HBSS three times. Collagenase-treated tubules were then dissociated in a mixture of 0.8 mL of 0.25% trypsin and 0.2 mL of 7 mg/mL DNase in HBSS for 10 min at 37°C in a water bath. A total of 5 mL of 10% FBS-contained Iscove’s modified Dulbecco’s medium (IMDM; Invitrogen, Carlsbad, CA) were subsequently added to the trypsinized cells and mixed by gently pipetting. The mixture was then centrifuged at 400g for 5 min. Cells were resuspended in 10% FBS-contained IMDM, counted, and then plated at a concentration of 8 × 105 cells/mL per well on a 0.2% gelatin-coated cover glass. Cells were allowed to settle for 24 h in 5% CO2 at 37°C. Nonattached cells were washed out with serum-free IMDM four times. Finally, 1 mL/well of IMDM containing 10% FBS was added the culture and cells were cultured for an additional 48 h in a CO2 incubator, as above.
In vitro fertilization was performed as described previously [50]. In brief, female B6D2F1 mice were superovulated by intraperitoneal injections of equine chorionic gonadotropin (eCG; Teikoku Zoki, Tokyo, Japan) and human chorionic gonadotropin (hCG; Teikoku Zoki, Tokyo, Japan) at 48-h intervals. Ovulated eggs were recovered 13 h after the hCG injection and placed in a 200-μL drop of modified Krebs–Ringer bicarbonate solution (TYH medium) [51] containing glucose, sodium pyruvate, bovine serum albumin, and antibiotics (0.05 mg/mL penicillin, 100 IU/mL streptomycin). Fresh spermatozoa from the cauda epididymis were dispersed in a 200-μL drop of TYH medium, diluted to 1×106 spermatozoa/mL, and incubated for 1.5 h to induce capacitation. An aliquot of capacitated spermatozoa from control WT mice and mDia1/3 DKO mice were then added to the eggs at 2×105 spermatozoa/ml for IVF. The mixture was incubated for 8 h at 37°C under 5% CO2 in air. Fertilization was confirmed by pronuclear formation. For observation of sperm–zona binding, egg masses were treated with bovine testicular hyaluronidase (175 U/mL; Sigma-Aldrich, St. Louis, MO) for 5 min to remove the cumulus cells. The cumulus-free eggs were placed in a 200-μL droplet of TYH medium and inseminated as described. After 30 min of incubation, sperm binding to the zona pellucida of the eggs was observed under an IX-70 microscope (Olympus, Tokyo, Japan).
The cauda epididymis was collected and minced with a razor blade in 1 mL PBS. The suspension was filtered through 70-μm-nylon mesh (Corning, Corning, NY). The number of spermatozoa was counted using a hemocytometer.
Cauda epididymal spermatozoa were suspended and incubated in TYH medium. Sperm motility was then measured using the CEROS sperm analysis system (software version 12.3; Hamilton Thorne Biosciences, Beverly, MA). Analysis settings were as described previously [52]. The percentage of hyperactivated spermatozoa was analyzed as described previously [53].
Mature sperm were collected from cauda epididymis by pricking with a fine needle and rinsed twice with PBS. Sperms were then plated on slide glasses, air-dried, and fixed with 4% PFA/PBS for 15 min at room temperature. Sperm morphology was observed on a laser scanning confocal imaging system (Leica SP5) with an oil immersion objective lens (100×1.4 numerical aperture).
To enrich spermatogonial stem cells in donor testicular cell suspension, experimental cryptorchidism was surgically induced in donor mice [54]. Three to six weeks after the surgery, testis cells were dissociated into single-cell suspensions by two-step enzymatic digestion using collagenase type IV and trypsin (SIGMA-Aldrich, St. Louis, MO), as described previously [23]. The dissociated WT or mDia1/3 DKO cells were then transplanted into W/Wv recipient mice (Japan SLC, Hamamatsu, Japan) through the efferent duct. For reciprocal transplantation, testis cells from cryptorchid acro/act-EGFP transgenic mice were transplanted into busulfan-treated control WT or mDia1/3 DKO mice through the efferent duct. Approximately 5 or 10 μL was introduced into the testes, respectively. Each injection filled 75%–85% of the seminiferous tubules.
Testes or epididymides were fixed with 4% PFA at 4°C overnight. For paraffin sections, the testes or epididymides were embedded in paraffin and cut into sections at 4-μm thickness. For frozen sections, the testes were cryoprotected in 0.1 M PB containing 30% sucrose, frozen in Tissue-Tek OCT compound (Sakura-finetek, Tokyo, Japan) in dry ice, and then were cut into sections at 12-μm thickness using cryostat. HE and PAS staining were performed according to standard protocol. For immunohistochemistry, cryosections were washed with PBS three times for 5 min each. Antigen retrieval for mDia1, nectin-2, espin1, and vimentin staining was carried out by boiling the sections in 10 mM citrate buffer, pH 6.0, with a pressure cooker. Sections were incubated with the blocking buffer (PBS containing 1% normal donkey serum (Jackson laboratory, Bar Harbor, ME) or 1% normal goat serum (Jackson laboratory, Bar Harbor, ME) and 0.3% TritronX-100) for 1 h at room temperature. Sections were then incubated with the primary antibody diluted in blocking buffer. After three washes with 0.3% Triton X-100 in PBS, sections were incubated with appropriate secondary antibodies conjugated to Alexa Fluor 488, Alexa Fluor 555, Alexa Fluor 594, or Alexa Fluor 633 (Invitrogen, Carlsbad, CA). Hoechst 33342 (Invitrogen, Carlsbad, CA) or DAPI (Molecular Probes, Eugene, OR) was used for nuclear staining. Phalloidin conjugated with Alexa 488, 546, 555, or 647 (Invitrogen, Carlsbad, CA) was used for the staining of F-actin. Peanut agglutinin (PNA) conjugated with TRITC (SIGMA-Aldrich, St. Louis, MO) was used for the staining of round and early elongated spermatids acrosome. TUNEL staining was performed according to the manufacturer’s instructions using ApopTag Fluorescein Direct In Situ Apoptosis Detection Kit (Millipore, Burlington, MA). Fluorescent images were acquired with a laser scanning confocal imaging system (Leica SP5) equipped with 100× NA 1.4 HCX PL APO CS oil immersion objective lens (Leica, Wetzlar, Germany). Images were processed using Adobe Photoshop CS5 software (Adobe, San Jose, CA).
For quantification of nectin-2 intensity, nectin-2 staining signals associated with the apical ectoplasmic specialization junction were selectively visualized by setting a threshold, and then the average intensity per pixel was analyzed by ImageJ software (http://rsbweb.nih.gov/ij). For coefficient of variation of nectin-2, a line scan of the nectin-2 staining signal along the apical ectoplasmic junction was performed using ImageJ software by drawing a line, and the intensity values along the line were obtained using the Plot Profile. Coefficient of variation of nectin-2 staining intensity along the apical ectoplasmic junction were then calculated as the ratio of the standard deviation and average staining intensity. Line scan for measurement of F-actin intensity was performed using ImageJ software by drawing a line on the seminiferous tubule, and the intensity values along the line were obtained using the Plot Profile. Plots were generated based on the data obtained by ImageJ using Prism software (GraphPad Software, San Diego, CA). For quantification of seminiferous tubule F-actin intensity, region of interest (ROI) was defined to specifically include the seminiferous tubule but not the myoid cell layer, and average F-actin intensity per pixel was then analyzed by ImageJ software. Averages of F-actin intensity per pixel of WT and mDia1/3 DKO seminiferous tubules were then generated based on the data obtained by ImageJ using Prism software (GraphPad Software, San Diego, CA). For quantification of the proportion of espin1-positive actin bundles, a binary image of F-actin bundles and espin1 were generated by setting a threshold. The two images were then merged and the pixel number of total F-actin bundles and espin1 signals that are colocalized with F-actin bundles was analyzed. The proportion of espin1-positive F-actin bundles were then calculated as a ratio of the pixel number of espin1-associated F-actin bundles and total F-actin bundles for seminiferous tubules. Averages of the proportion of espin1-positive actin bundles of WT and mDia1/3 DKO seminiferous tubules were then generated using Prism software (GraphPad Software, San Diego, CA).
Sertoli cells were cultured for 72 h on 25 mm round, No. 1.5 fiducialated cover glass (Hestzig, Leesburg, VA; #600-100AuF) for TIRF 3D-N-STORM superresolution imaging. For the fixation, the cells were briefly washed with 2.5 mL of prewarmed (37°C) PBS and fixed and permeabilized with 0.3% glutaraldehyde (EM Grade; Electron Microscopy Sciences, Hatfield, PA) and 0.25% Triton X-100 in 1 mL of cytoskeleton buffer (10 mM MES, pH 6.1, 150 mM NaCl, 5 mM EGTA, 5 mM glucose, and 5 mM MgCl2) for 1.5 min. The second fixation step was performed with 2% glutaraldehyde in 1 mL of cytoskeleton buffer for 10 min. Samples were quenched with 0.1% NaBH4 (SIGMA-Aldrich, St. Louis, MO) in 2.5 mL of freshly prepared PBS for 7 min on ice to reduce autofluorescence. Finally, fixed cells were washed with PBS twice, each time for a 10-min incubation, and then kept in PBS overnight at 4°C.
Before imaging, F-actin was probed using 0.33 μM of Alexa Flour 647-conjugated phalloidin (Life Technologies, Waltham, MA) and incubated for 30 min at room temperature. The samples were imaged on a Nikon N-STORM microscope (Nikon, Tokyo, Japan) in TIRF mode. The microscope is equipped with a 100× NA 1.49 Apo TIRF objective lens, a back-illuminated EMCCD camera (Andor Ixon3, Belfast, UK), a Cy5 (excitation, 620/60; emission, 700/75) filter set (Chroma, Bellows Falls, VT), and a cylindrical lens for 3D imaging. During imaging, a 100-mW 641-nm laser (Coherent, Santa Clara, CA) and a 100-mW 405-nm laser (Coherent, Santa Clara, CA) were used to excite and photoswitch Alexa Fluor 647, respectively. In each image acquisition, a 5–10-min period of full-powered illumination by 641-nm laser was first performed to turn most Alexa Fluor 647 fluorophores into dark state to achieve a sparse distribution of single molecules. Image frames were acquired when single-molecule blinking could be observed. The intensity of 405-nm laser was periodically tuned during imaging to reactivate fluorophores to maintain sufficient fluorophore density.
For each image, 40,000 raw frames were acquired in the frame-transfer mode, with an exposure time of 50 ms, EM gain of 200, and a readout speed of 10 MHz. Fluorophore detection and localization were carried out by PeakSelector (courtesy of Harald Hess, Howard Hughes Medical Institute), a customized software developed in IDL (Exelis Vis, Boulder, CO). The centroid position of each detected fluorophore was calculated via 2D-Gaussian nonlinear least-square fitting, as described earlier [55]. The localization precision of each centroid coordinate was computed using the Thompson-Webb formula depicted earlier [56]. Fluorophore peaks with localization precisions smaller than 20 nm were rejected from subsequent analysis. The fiducials pre-embedded on the glass coverslips were used to correct drift. The N-STORM images were then reconstructed by representing each localization coordinate as a normalized Gaussian function whose widths depend on the localization precision [57]. Three-dimensional data were rendered with colors encoding the z positions of fluorophores.
The N-STORM image of F-actin in Sertoli cells contains a mixture of fine meshwork and high-density bundles. To separate these two structures and analyze the density of the fine F-actin meshwork, the image was first reconstructed with a pixel size of 160 nm (an image size of 256 × 256 pixels) to blur regions of F-actin meshwork, leaving F-actin bundles as more prominent structures. The cell region was then anisotropically enhanced [58, 59] to highlight F-actin bundles, and then segmented using Otsu's method [60]. Next, another image was reconstructed with a pixel size of 10 nm, which is sufficiently high for the F-actin network, generating an image size of 4,096×4,096 pixels. The previously segmented regions of F-actin bundles were scaled up by 16 times and used as masks to isolate regions between thick F-actin bundles from the high-resolution image. The F-actin network was segmented from the isolated areas using Otsu's method. F-actin occupancy was then computed as the ratio between the number of F-actin pixels and the total number of pixels in the binarized regions of the F-actin network.
For immunocytochemistry, cells were fixed with 4% PFA in PBS for 15 min at room temperature. The fixed cells were permeabilized with 0.1% Triton X-100 in PBS for 15 min at room temperature and then blocked with 5% skim milk in PBS for 1 h at room temperature. The cells were then stained with anti-pMLC antibodies (CST, Danvers, MA) and anti-espin1 antibodies (BD Biosciences, San Jose, CA). Secondary antibodies were Alexa Fluor 488-conjugated goat anti-mouse IgG (Molecular Probes, Eugene, OR) and Alexa Fluor 488-conjugated goat anti-rabbit IgG (Molecular Probes, Eugene, OR). F-actin was stained with Alexa Fluor 555-conjugated phalloidin (1:100; Molecular Probes, Eugene, OR). DAPI (Molecular Probes, Eugene, OR) was used for nuclear staining. Fluorescent images were acquired on a laser scanning confocal imaging system (Leica SP5) equipped with 100× NA 1.4 HCX PL APO CS oil immersion objective lens (Leica, Wetzlar, Germany). Images were processed using Adobe Photoshop CS5 software (Adobe, San Jose, CA).
ROI including a cell was defined, and average pMLC intensity per pixel was then analyzed by ImageJ software. Averages of pMLC intensity per pixel of WT and mDia1/3 DKO cells were generated based on the data obtained by ImageJ using Prism software (GraphPad Software, San Diego, CA). For quantification of the proportion of espin1-positive actin bundles, binary images of F-actin bundles and espin1 were generated by setting a threshold. The two images were then merged and the pixel number of total F-actin bundles and espin1 signals that are colocalized with F-actin bundles were analyzed. The proportion of espin1-positive F-actin bundles were then calculated as a ratio of the pixel number of espin1-associated F-actin bundles and total F-actin bundles for each cell. Averages of the proportion of espin1-positive actin bundles of WT and mDia1/3 DKO cells were then generated using Prism software (GraphPad Software, San Diego, CA).
Primary cultured Sertoli cells were transfected with pCAG-LifeAct-EGFP [5] or pCAG-EGFP-mDia3 [5] by electroporation using Neon Transfection System (Invitrogen, Carlsbad, CA) according to the manufacturer’s protocol. Briefly, 2×105 cells were transfected with 1.5 μg of plasmid DNA. Electroporation was performed at a single pulse of 1,350 V for 30 ms. After electroporation, cells were plated in IMDM medium containing 10% FBS for 24 h in 5% CO2 at 37°C. For observation of LifeAct-EGFP, primary cultured Sertoli cells were maintained in IMDM without phenol red containing 2% FBS. Time-lapse imaging of LifeAct-EGFP was carried out at 37°C using a SD-OSR IX83 inverted microscope (Olympus, Tokyo, Japan) equipped with a 100× NA 1.4, UPLS APO oil immersion objective (Olympus, Tokyo, Japan) and Yokogawa W1 spinning disk unit (Yokogawa, Musashino, Japan) and controlled by MetaMorph software (Universal Imaging, San Jose, CA). An area near the cell periphery was selectively illuminated. Images were recorded at a rate of 2 s/frame. Speed, straightness, and frequency measurements were performed by tracking individual actin filaments manually using ImageJ software. Graphs were generated based on the data obtained by ImageJ using Prism software (GraphPad Software, San Diego, CA). Single-molecule speckle imaging of EGFP-mDia3 was acquired using a microscope (IX81, Olympus, Tokyo, Japan) equipped with 100-W mercury illumination, a Plan-Apo 100×, 1.4 NA oil-immersion objective (Olympus, Tokyo, Japan), and a cooled EMCCD camera (Evolve 512; Photometrics, Tucson, AZ). Cells expressing a low level of EGFP-mDia3 were observed. An area near the cell periphery was selectively illuminated. Images were recorded at a rate of 200 ms/frame. Measurements of speed, straightness, and travel distance of single-molecule movement were performed by tracking individual single molecules manually using ImageJ software. Graphs were generated based on the data obtained by ImageJ using Prism software (GraphPad Software, San Diego, CA). Kymograph analysis was performed using MetaMorph software (Universal Imaging, San Jose, CA).
Testes were isolated from the 3-wk-old C57BL/6-Tg (CAG-EGFP) mice, and the tunica and blood vessels were dissected and removed by fine forceps. Isolated testes were minced in PBS using scalpel blades. The minced tissue was then transferred into 50-mL plastic tubes and centrifuged at 100g for 1 min. Supernatant was discarded and the pellet was washed twice in PBS and centrifuged at 100g for 1 min to recover germ cells. Cells were then resuspended in PBS and filtered through the 40-μm nylon mesh to remove tissue debris and cell clumps. Germ cells were then centrifuged at 500g for 10 min. The resultant cell pellet was finally resuspended in Ham's F12/DMEM containing 10% FBS [61].
To remove residual germ cells contaminated in the primary culture of Sertoli cells prepared as described above, the primary cultures, after 3 d of incubation, were subjected to hypotonic treatment for 2 h at room temperature with 20 mM Tris-HCl buffer, pH 7.4, for 2 min [62]. Purified Sertoli cells were then incubated with 10% FBS in IMDM for 2 h in 5% CO2 at 37°C to recover from the hypotonic treatment. Germ cells (5×105 cells/well) isolated from testes of 3-wk-old C57BL/6-Tg (CAG-EGFP) mice were then plated onto the Sertoli cells. These cells were cocultured for 1 d, fixed with 4% PFA/PBS for 15 min at room temperature, and subjected to immunocytochemistry. Fluorescent images were acquired on a SD-OSR IX83 inverted microscope (Olympus, Tokyo, Japan) equipped with a 100× NA 1.4, UPLS APO oil immersion objective (Olympus, Tokyo, Japan) and Yokogawa W1 spinning disk unit (Yokogawa, Musashino, Japan) controlled by MetaMorph software (Universal Imaging, San Jose, CA).
Living GFP-positive germ cells with round nuclear shape, as determined by DAPI staining and continuous N-cadherin signal around the germ cells, were defined as adhesive cells. Living GFP-positive germ cells with round nuclear shape, as determined by DAPI staining and continuous phalloidin staining surrounding the germ cells, were defined as germ cells with continuous F-actin cup. Cortical F-actin meshwork density and the continuity of the F-actin cup were quantified manually using ImageJ software. Graphs, including the plot of correlation, were generated based on the data obtained by ImageJ using Prism software (GraphPad Software, San Diego, CA).
Sperms prepared from the cauda epididymis and suspended as above were fixed with an equal volume of 2% PFA and 2% GA in 0.1 M cacodylate buffer. Thereafter, they were fixed with 1% GA in 0.1 M cacodylate buffer, pH 7.4, at 4°C overnight. The samples were further fixed with 1% tannic acid in 0.1 M cacodylate buffer, pH 7.4, at 4°C for 1 h. After the fixation, the samples were washed four times with 0.1 M cacodylate buffer for 30 min each, followed by postfixation with 2% OsO4 in 0.1 M cacodylate buffer at 4°C for 2 h. The samples were dehydrated through a series of graded ethanol (50%, 70%, 90%, 100%). The samples were substituted into tert-butyl alcohol at room temperature and then frozen. The frozen samples were vacuum dried. After drying, the samples were coated with a thin layer (30 nm) of osmium by using an osmium plasma coater (NL-OPC80NS, Nippon Laser & Electronics Laboratory, Nagoya, Japan). The samples were observed by a JSM-6340F scanning electron microscope (JEOL, Akishima, Japan) at an acceleration voltage of 5.0 kV.
We constructed the pLV-LifeAct-EGFP (LV, lentivirus) by inserting a LifeAct-EGFP fragment amplified by PCR from pCAG- LifeAct-EGFP [5] with primers LifeAct-EGFP forward (5′-GCTCTAGAATGGGCGTGGCCGACCTGAT-3′) and LifeAct-EGFP reverse (5′-CTCTCGAGTTACTTGTACAGCTCGTCCATGCC-3′) into pLV-CAG1.1 (a gift from Dr. Inder Verma, Salk Institute). Lentivirus were generated as previously reported [32]. We injected recombinant lentivirus vectors into male C57BL6/N adult mice at 6 wk of age. Mice were anesthetized by i.p. injection of Avertin before the operation. Approximately 10 μL of lentivirus vector solution containing 0.04% trypan blue was injected into the right seminiferous tubules via the efferent ductules, according to the method described, and Sertoli cells were selectively labeled [32]. After 1 wk, mice were humanely killed and the injected right testis was dissected and the tunica albuginea were removed. Preparation of seminiferous tubules for imaging was then performed according the previous report [33], with slight modifications. Seminiferous tubules were disentangled and then fixed with 4% PFA for 1 h at room temperature. Fixed seminiferous tubules were then cut into small pieces and attached to MAS-coated slide glass (Matsunami, Kishiwada, Japan) by half-drying. Specimens were mounted in antifade prolong diamond (Thermo Fisher Scientific, Waltham, MA). Samples were observed under SD-OSR IX83 inverted microscope (Olympus, Tokyo, Japan) equipped with a 100× NA 1.4, UPLS APO silicone immersion objective (Olympus, Tokyo, Japan), and Yokogawa W1 spinning disk unit (Yokogawa, Musashino, Japan) and controlled by MetaMorph software (Universal Imaging, San Jose, CA). Deconvolution processing of stacked images was performed by cellSens Dimension software (Olympus, Tokyo, Japan). Three-dimensional reconstruction of stacked images was generated by Volocity software (PerkinElmer, Waltham, MA).
Prism (GraphPad Software, San Diego, CA) and Excel (Microsoft, Redmond, WA) were used for statistical analyses. Data are presented as mean ± SEM, and were analyzed by one-way factorial ANOVA or unpaired Student t test. P < 0.05 was considered statistically significant.
|
10.1371/journal.pntd.0007375 | Animal model of arthritis and myositis induced by the Mayaro virus | The Mayaro virus (MAYV) is an endemic arbovirus in South American countries, where it is responsible for sporadic outbreaks of Mayaro fever. Clinical manifestations include fever, headache, ocular pain, rash, myalgia, and debilitating and persistent polyarthralgia. Understanding the mechanisms associated with MAYV-induced arthritis is of great importance due to the potential for its emergence, urbanization and dispersion to other regions.
15-day old Balb/c mice were infected by two distinct pathways, below the forelimb and in the rear footpad. Animals were observed for a period of 21 days. During this time, they were monitored every 24 hours for disease signs, such as weight loss and muscle weakness. Histological damage in the muscles and joints was evaluated 3, 7, 10, 15 and 20 days post-infection. The cytokine profile in serum and muscles during MAYV infection was evaluated by flow cytometry at different post-infection times. For pain analysis, the animals were submitted to the von Frey test and titre in different organs was evaluated throughout the study to obtain viral kinetics.
Infection by two distinct pathways, below the forelimb and in the rear footpad, resulted in a homogeneous viral spread and the development of acute disease in animals. Clinical signs were observed such as ruffled fur, hunched posture, eye irritation and slight gait alteration. In the physical test, both groups presented loss of resistance, which was associated with histopathological damage, including myositis, arthritis, tenosynovitis and periostitis. The immune response was characterized by a strong inflammatory response mediated by the cytokines TNF-α, IL-6 and INF-γ and chemokine MCP-1, followed by the action of IL-10 and IL-4 cytokines.
The results showed that Balb/c mice represent a promising model to study mechanisms involved in MAYV pathogenesis and for future antiviral testing.
| The Mayaro virus, although restricted to some regions of Latin America, has great potential for emergence, which makes it of great medical-scientific interest. Therefore, pathogenesis study of the MAYV in an animal model has fundamental importance for the determination of viral and host factors that contribute to disease development. In addition, it will allow develop and evaluate the effectiveness of possible antiviral agents. In this study, the authors were able to standardize a disease model for the MAYV in BALB / c mice. From the obtained data it is possible to observe the induction of acute arthritis and myositis, accompanied by the reduction of physical strength of the animals. As described for other alphaviruses in both animal models and patient, the proinflammatory mediators TNF-α, IL-6, INF-γ and MCP-1 were elevated in the serum of MAYV-infected animals and therefore appeared to be mediators which also play an important role in the pathogenesis of MAYV.
| Mayaro virus (MAYV, genus Alphavirus, Togaviridae family) is an emergent arbovirus, responsible for sporadic cases, outbreaks and small epidemics of acute febrile disease in South American countries, particularly in the Amazon basin. However, serological data have shown the circulation of the MAYV in Central American countries, such as Guatemala, Costa Rica, Panama and more recently in Haiti, where a new strain was isolated [1,2]. MAYV infections mainly affect people living or working near forest areas where MAYV is kept in an enzootic cycle, which primarily involves the Haemagogus janthinomys mosquito as a vector and nonhuman primates as natural hosts, however other vectors (Culex sp, Sabethes sp, Psorophora sp, Coquillettidia sp and Aedes sp mosquitoes) and wild vertebrates (marsupials, rodents and birds) may be important in the transmission cycle and spread of the virus [3].
The MAYV as well as the chikungunya virus present a high potential for urbanization and emergence due to the ability of these viruses to mutate and adapt to new transmission cycles [4]. Recent studies in Brazil have already demonstrated the occurrence of Mayaro fever in urban areas of Manaus city and Mato Grosso state [5,6,7]. In Cuiabá, capital of Mato Grosso state, an entomological surveillance study identified the species of Culex quinquefasciatus and Aedes aegypti as mosquitoes naturally infected by MAYV, corroborating with the occurrence of urban transmission in Cuiabá and possibly in other cities of central region of Brazil [7]. This shows that Mayaro fever should be included in the differential diagnosis with dengue virus (DENV), chikungunya virus (CHIKV) and Zika virus (ZIKV) in areas where there is co-circulation of these arboviruses [8]. The recent MAYV case in Haiti poses a potential threat for spreading through Central and North America. Due to its geographical proximity, vector species and population flows, the United States is the country with the highest risk for MAYV emergence [9,10]. In addition, reports of European tourists who visited the Amazon region and returned to their countries with symptoms characteristic of Mayaro fever, highlights concerns regarding MAYV as a potential emerging disease in Europe, where Aedes albopictus is present and experimentally able to transmit the MAYV [11,12].
Mayaro fever is characterized as a self-limiting disease, which can range from mild to moderately severe. Clinical manifestations include fever, headache, ocular pain, rash, photophobia, joint edema, myalgia, and arthralgia. Hemorrhagic phenomena such as petechiae, gingival bleeding and epistaxis have also been observed [5]. Like other arthritogenic alphaviruses, MAYV also induces disabling and long-lasting polyarthralgias, with the joints of the wrist, ankle and small joints of hands and feet being predominantly affected [13].
As the pathophysiology of the MAYV is not fully elucidated and represents an important public health issue, the objective of this research was to establish an animal model for the study of the pathogenesis induced by the MAYV. The results demonstrated that Balb/c mice at 15-days developed acute disease with musculoskeletal damage, which included myositis, arthritis, tenosynovitis and periostitis. A strong inflammatory response was induced on initial days of infection, which was controlled through the participation of IL-4 and IL-10 cytokines. These results present an animal model to study the pathogenesis of the Mayaro virus.
The inoculation of the animals was performed via two distinct pathways, below the forelimb and in the rear footpad, determined as FLP and FPP, respectively. Clinical signs such as ruffled fur, hunched posture and eye irritation were observed in both groups (Table 1). Eye irritation was characterized by tearing, blepharospasm, and may present unilaterally or bilaterally (Fig 1). Only FLP animals presented weight loss and gait alteration (Fig 2A, 2B and 2C). Weight loss in FLP begins at day 1 PI, while in the other groups, we observed a gradual increase in body weight. A significant difference between FLP and the control group is evident on day 7 PI and remains up to day 19 PI. Some signs such as hind limb drag, tremors, and falling and circulating spontaneously were observed in a single experiment in the FLP animals. The change in the animals′ gait was more pronounced from day 10 PI. Tremors were observed only in three animals and only one animal showed the signal falling and circulating spontaneously. In addition, 33% of the animals succumbed to infection in the same experiment (Fig 3).
In the physical test, it was found that both infected groups showed significant reduction in time differences when compared to controls (Fig 4A, 4B and 4C). In the Figure, it can be observed that in the first five days, all groups had very close and low time averages (Fig 4C). Up to the tenth day post-infection, an increase in the time of all four groups was observed. However, from the tenth day a significant difference was observed between the infected and uninfected groups, which continued up to day 19 PI. On day 16, FPP began to increase the permanence time adhered to the grid, while in the FLP animals, this increase occurred from day 17.
The analysis of MAYV-induced morphological changes was performed on skeletal muscles and joints, on 3, 7, 10, 15 and 20 days PI. Myositis occurred symmetrically for both infection pathways, affecting the skeletal muscle of the pelvic limbs in a similar way. In FLP, little or no inflammation was observed on day 3 PI, with the muscle tissue being very similar to the control (Figs 5B and 6B). However, on days 7 and 10 PI, an increased number of inflammatory infiltrates were observed (Fig 5C, 5D and 5J; Fig 6C, 6D and 6J), concentrated in specific points, but disperse in all tissue. Additionally, there is also the presence of degenerating fibers (Fig 5C, 5D, 5G and 5H; Fig 6C and 6H), which are characterized by a pale color, swollen appearance and central nucleus. Collagen deposition (Fig 5C) and presence of vasculitis were also observed (Fig 5I). With 15 and 20 days PI there were still small foci of inflammatory infiltrate. In the joints and bones of the foot, inflammation started on day 3 PI, and the presence of inflammatory cells in the synovium, tendon, ligament, periosteum and extensive inflammation and destruction of associated muscle was observed (Fig 7). In FPP, the inflammation in the muscle began on day 3 PI, with the presence of degenerating fibers (Figs 8B and 9B). On days 7 and 10, the inflammatory process continued (Figs 8C and 8D, 9C and 9D), but there was no difference in the histological score between post-infection times (Figs 8G and 9G). At day 15 PI there were still few infiltrate cells (Figs 8E and 9E), with full tissue recovery on day 20 PI (Figs 8F and 9F). In both groups, tissue recovery occurred between 15 and 20 days PI, which is indicated by the presence of fibers with rows of centralized nuclei (Figs 5E, 6H and 8F).
MAYV infection initially activates the Th1 pattern, inducing a strong inflammatory response, which is followed by cytokine induction of the Th2 response. Both groups showed similar cytokine profile in serum. On the third day after infection, high levels of proinflammatory cytokines TNF, INF-γ and IL-6 and chemokine MCP-1 were detected in the serum of both infected groups. TNF and INF-γ levels remained high until day 10, when an increase in the anti-inflammatory cytokine IL-10 was also observed. In FPP at 30 days PI, IL-4 and TNF were detected again. In FLP, the IL-4 cytokine was present at 10 days PI and on 30 days PI no cytokines were detected (Figs 10 and 11). In muscle tissue, MCP-1 was detected at high levels at 3 days PI in FLP (Fig 12A), while in FPP it was present on days 7 and 10 PI (Fig 12B). Cytokines IL-9, IL-13 and IL-12p70 were not found in either serum or muscle tissue.
Peak titers of infectious virus were detected at 24 to 48 hours post-infection (hpi) in both infection pathways (Figs 13 and 14). In animals from FLP the highest viral titers were detected in the left and right ankles with 109 PFU/g and 1010 PFU/g, respectively (Fig 13E and 13F), and live, spleen and muscles reached a peak of 108 PFU/g (Fig 13A, 13B, 13G and 13H). In FPP animals, viral titers of up to 1010 PFU/g were detected in the right ankle (Fig 14D), and 108 PFU/g in the muscles and left ankle (Fig 14F, 14G and 14H). MAYV has also been found in other organs such as the heart and eye. The viremia persisted for up to 144 hpi in both groups.
Until now, we have shown that MAYV infection induced clinical, morphological and inflammatory changes. Our next question was if these inflammatory indices were associated with articular dysfunction. For that, mice with 28 days were used to evaluate the intensity of hypernociception. Our results revealed that at 2 and 3 days PI, both mice from FLP and FPP groups, showed a decrease in their nociceptive thresholds in comparison to MOCK-infected groups. From day 7 PI the nociceptive thresholds were already similar to baseline levels (Fig 15). These animals did not show any sign of disease besides pain.
The arthritogenic alphaviruses include chikungunya (CHIKV), Sindbis (SINV), Ross River (RRV), Mayaro (MAYV), Barmah Forest (BFV) and O′nyong′nyong (ONNV) viruses, which are associated with arthritis and/or debilitating arthralgia, and in some cases with bone pathologies [14,15]. These viruses are considered an important socioeconomic problem, since they have the potential to cause large epidemics with a significant impact on human health and the economy, due to their debilitating effects [16]. The current understanding of the mechanisms involved in alphaviral pathogenesis has been based on studies with animal models and clinical observations, mainly with RRV and CHIKV. With the risk of emergence and urbanization of the MAYV in South America and other continents, it has become necessary to understand the pathology and immune response of this virus. Therefore, this research presents an animal model of arthritis and myositis induced by the MAYV.
In the identification of murine as animal models for alphavirus, the age of the animal has been considered a determining factor in the susceptibility and severity of the disease. As neonatal and young mice are more susceptible, they have been the most used to reproduce alphaviral disease [17,18]. The resistance of adult mice to alphavirus infection is due to type I interferons, which quickly provide an antiviral state [17,19]. It has been shown in experiments with adult IFN-α/βR+/- and IFN-α/βR-/- mice that develop moderate or severe disease, respectively [19]. Over time several animal models for arthritogenic alphaviruses have been tested in different strains of neonatal or young mice, including CD-1, Swiss, Balb/c and C57BL/6 [17]. In the present study, it was shown that the Balb/c mice at 15 days old and, regardless of the infection pathway, were susceptible to the disease induced by MAYV.
Clinical signs such as ruffled fur, hunched posture and eye irritation were common and frequent in both infection pathways. Eye irritation is a frequent symptom in patients with Mayaro fever, who report photophobia and retroocular pain [5,13]. This clinical sign in the animal model was described only for sindbis virus, but the infection pathway used was intracranial, a localized infection pathway [20]. However, in other animal models for alphavirus, such as CHIKV and RRV, which also used peripheral infection pathways, no eye irritation was reported [18]. In addition, it has been reported that CHIKV can infect humans’ cornea and be transmitted by the ocular pathway in animals [21], and here we demonstrate that the MAYV is able to replicate in the eye of the animals. Therefore, it is important to investigate their potential for transmission by this pathway. Signs of disease such as gait alteration, weight loss and death occurred only in animals infected below the forelimb and have been described for pathway-independent RRV [18]. Edema in paw, which is a characteristic sign of alphavirus infection, has also been reported for MAYV in models using the C57BL / 6 and A129 lines, however, it was not observed in this work [22,23]. As for the signs of tremors, hunched posture and animals falling and circulating spontaneously, no reports were found in the literature. The kinetics of MAYV replication showed a similar pattern in the two infection pathways studied, with the virus persisting longer in the joints than in the other tissues. In addition, both pathways were efficient at spreading the virus. The presence of alphavirus in the heart and eye in other animal models was not reported in the literature, making it necessary to investigate the tropisms of other alphaviruses by these tissues.
The infection also resulted in several histopathological changes, such as myositis, tenosynovitis, synovitis and periostitis. These tissue changes were responsible for the lower strength of infected animals in the physical test. This resistance loss was more intense between 10 and 15 days PI, corresponding to the tissue damage and the beginning of the regeneration process. Both infection pathways could cause myositis, which was characterized by presence of infiltrates and destruction of muscle fibers. Tissue damage described for MAYV is similar to that found in animal models for CHIKV and RRV, with extensive degeneration and necrosis of skeletal muscle, inflammation of the periosteum and tissues associated with the joint being reported [18,24].
New, potentially aggressive lesions have been identified in CHIKV infections in animals, such as necrosis of cartilage and bone proliferation in the periosteum. In patients with CHIKV, imaging studies have revealed the presence of tenosynovitis, synovial thickening, periostitis, periosteum proliferation and bone erosion. These clinical findings show that the initial acute disease may progress to chronic erosive arthritis [15,24]. In this study, identification of tenosynovitis and periostitis in animals infected by MAYV emphasizes the importance of imaging tests in clinical studies of patients with prolonged polyarthralgia.
As with other alphavirus infections, Mayaro fever causes persistent, often debilitating joint pain [25]. Joint pain is due to the action of inflammatory cytokines, such as IL-6, TNF-α and IL-1β, as well as tissue damage, which sensitizes nociceptors [26]. The presence of pain at only 2 and 3 days PI shows that induced-MAYV pain is associated with the inflammatory response, which is characterized by elevated IL-6 and TNF-α levels on day 3 PI.
The evaluation of the cytokine profile showed that MAYV induces a strong inflammatory response in early days of infection, mediated by cytokines TNF-α, IL-6 and INF-γ and chemokine MCP-1. These mediators have also been expressed in animal models for RRV and CHIKV, and more recently it has been demonstrated in patients with Mayaro Fever [27,28]. High levels of IL-6 and TNF-α during early stages are probably responsible for inducing pain observed in animals on day 3 PI. In patients with chronic arthralgia in MAYV infection, these factors are also elevated [28]. Initial inflammatory response in animals was rapidly controlled by anti-inflammatory action of IL-4 and IL-10, and in later stages by IL-4. Besides the control of immune response, IL-10 also plays an important role in muscle regeneration by inducing the exchange of macrophages from the M1 to M2 phenotype in the injured muscle [29]. The action of IL-10 on tissue recovery is indicated by the presence of fibers with centralized nuclei, which is a characteristic of the muscle regeneration process. The presence of INF-γ and IL-4 suggests the participation of adaptive immunity. A second increase in TNF-α levels in animals FPP after 30 PI may indicate the persistence of viral RNA in the tissue, as has been demonstrated for CHIKV [30].
Patients with prolonged arthralgia during MAYV infection show that levels of IL-1Ra, IL-6, IL-7, IL-8, IL-13, IL-17, G-CSF, IFN-γ, PDGF- α, VEGF and IL-12p70 remained elevated up to 12 months after infection. The acute phase was marked by a high expression of MCP-1, while IL-9 and IL-2 were predominant during the convalescent phase. These data showed that a robust early inflammatory response is associated with the development of chronic arthralgia [28]. Cytokines IL-13, IL-9, IL-12p70 and IL-17 were also evaluated in animals but showed no differences in relation to the control. The major anti-inflammatory mediators induced in humans were IL-1Rα and PDGF-BB [28], different to the results found in this study using animals, in which IL-10 was the major agent responsible for inflammation control.
These results show that the animal model for Mayaro virus has been able to reproduce the acute disease that occurs in humans, regardless of the infection pathway, and is adequate to understand most events resulting from MAYV infection, such as tissue damage and inflammatory response. The presence of the cytokines MCP-1, TNF and INF-γ serum, as well as MCP-1 in the muscle, indicate the action of macrophages in target tissues of infection, which makes this cell and MCP-1 important elements common in therapy of arthritis caused by RRV, CHIKV and MAYV [31,32]. It is important that further studies are conducted to evaluate the role of these immune factors in the pathogenesis induced MAYV. Additional efforts need to be made in cohort studies to obtain more information about the course of the disease in humans, since we presented evidence here for ocular infection and severe tissue damage.
Animals were anesthesiaded with ketamine/Xylazine in all procedures and all experiments were done with agreement of Ethics Committee on Animal Use (Comitê de Ética no Uso de Animais—CEUA) of the Federal University of Viçosa, protocol number 63–2013, established by the National Animal Experimentation Control Council (Conselho Nacional de Controle de Experimentação Animal—CONCEA), an organ that is part of the Brazilian Ministry of Science, Technology and Innovation (MCTI).
The Mayaro virus (ATCC VR-66, strain TR 4675) was kindly provided by Dr. Davis Ferreira, Federal University of Rio de Janeiro—UFRJ. This strain corresponds to the first Mayaro virus isolate in Trinidad, 1954. The passages history of this strain, includes passages in the mouse brain [33], and the MAYV stock used in this work was propagated in C636 cells whenever necessary for viral titre increase. Vero cells were propagated at 37 °C in minimum essential media (MEM) supplemented with 10% fetal bovine serum (FBS), which were selected and plated in transfer to 24-well plate (3 X 105 cells per well) and used for virus titrations. The plates were incubated at 37 °C and, upon reaching ∼80% confluence, the medium MEM was removed. 100μl of each 10-fold virus dilution was then added, and the control wells received phosphate-buffered saline (PBS). After 1 hour incubated in the rotary shaker, 1.5 mL of CMC 3% (carboxymethylcellulose 3%, supplemented with 2% FBS) was added. Plates were returned to incubation at 37 °C for more 48 h. Throughout this time, cytopathic effects were observed. After incubation, 500μL of CMC 3% from each well was removed, 1 mL of formaldehyde 20% was added and maintained for 1 hour under these incubation conditions. Subsequently, the plates were washed and 500μL of 5% crystal violet (CV) was added. The plates were incubated for 30 minutes with the dye. After incubation, the plates were washed for lysis plate count. To determine viral titers in tissues, mice were sacrificed by exsanguination and perfused with 1X PBS. The right and left ankles, right and left quadricep muscles, spleen, liver, eyes and heart were removed by dissection at 24, 48, 72, 96, 120 and 144 hours post-infection. The tissues were weighed and then homogenized in MEM media supplemented with 10% FBS and stored at 80°C until viral load. To perform a 10X serial dilution after maceration of the tissue, the weight of each tissue was multiplied by nine, and the final value corresponds to the volume of medium required for each tissue, weight of organ as 10% of the total volume. All experiments were performed 5 times.
Balb/c mice obtained from the animal facility of the Federal University of Viçosa (Central Bioterium), with agreement of Ethics Committee on Animal Use (Comitê de Ética no Uso de Animais—CEUA) of the Federal University of Viçosa, protocol number 63–2013. The 15 day-old, female and male mice were subcutaneously inoculated with high doses of MAYV in the right rear footpad and in the thorax, below the right forelimb. The animals infected in the thorax, below the right forelimb (FLP) received 50 μL (1.25 x 107 PFU) of viral suspension, while the animals infected in the right rear footpad (FPP) received 10 μL (2.57x106 PFU) of virus suspension in PBS. Mock-infected mice received the PBS diluent alone. Cohort: 12/group for evaluation of clinical signs and physical test; 8/group for evaluation of pain; 3/group for evaluation of tissue damage and cytokine profile.
Animals were observed for a period of 21 days. During this time, they were monitored every 24 hours for the appearance of signs of disease, such as weight loss and muscle weakness. The physical test used to evaluate muscle weakness was wire-hanging, modified from the method described by Sango, McDonald [34]. This method uses a metal grid and a box lined with wood shavings; the mice are placed on the grid and it is shaken to make the animal hold on. The grid is then inverted, placed over the box and maintained at a height of 23 cm. The tendency to fall is measured using a stopwatch, and a maximum time of 2 minutes is established.
At 3, 10, 15 and 30 days post-infection, the serum was collected, centrifuged at 514xg for 10 minutes and stored at -20 °C. The levels of serum cytokines were determined using the CBA mouse Th1/Th2/Th17 kit (BD Biosciences), which simultaneously quantifies IL-2, IL-4, IL-6, IL-10, IL-17A, IFN-γ and TNF and CBA flex IL-9, IL-13, MCP-1 and IL12-p70, following manufacturer′s instructions. The level of cytokines was also evaluated in muscle tissue at 3, 7, 10, 15 and 30 days post-infection. For this, the muscle was macerated in PBS, centrifuged at 514 xg and stored at -80 °C. The data were acquired using by Flow cytometer BD FACS Verse.
For histology, the mice were sacrificed and the quadricep muscles and ankle joints were removed and fixed in Karnovsky fixative (paraformaldehyde 4% and glutaraldehyde 4%, pH 7.3). The bone-associated tissues were decalcified with EDTA 10%, pH 7. Further, the tissues were embedded in paraffin and a 5 μm section obtained by microtomy, stained with hematoxylin and eosin (H & E) and observed by light microscopy. The quantification of the inflammatory cell infiltration was performed based on the following score: 0 without infiltrates; 1 light infiltrates; 2 moderate infiltrates; 3 severe infiltrates. The score considers the number of infiltrates, the size and their distribution in the tissue. For each time the best slide of each animal (N = 3) was chosen to quantify the inflammatory infiltrate. After analysis the scores were added.
The electronic method of von Frey is based on the use of an apparatus (anesthesiometer electronic) that has a pressure transducer connected to a digital force counter expressed in grams (g). The pressure transducer is adapted to a polypropylene tip, which is brought into contact with the animal′s leg [35]. Nociceptive thresholds were determined by exerting a linearly increasing pressure on the center of the animal′s paw until a flexion reflex occurred, followed by a flinch response upon withdrawal of the paw. The pressure-meter automatically records the pressure value when the foot is removed. To perform the test, the animals were placed in acrylic boxes with non-malleable wire floor for 15 minutes for aclimatization. A sloping mirror below the railing provides a clear view of the animal hindpaw during measurements. The baseline measurements of all groups were performed before infection, and the other measurements were taken at 3, 7, 10, 15 and 21 days postinfection. There were always two measures per animal at a time. The intensity of hypernociception was quantified by the difference between baseline and the other evaluated times (delta = basal—other times). In this experiment, animals with 28 days of age were used, since it is the minimum age required to perform the test.
|
10.1371/journal.pgen.1000160 | Genetic Variation in an Individual Human Exome | There is much interest in characterizing the variation in a human individual, because this may elucidate what contributes significantly to a person's phenotype, thereby enabling personalized genomics. We focus here on the variants in a person's ‘exome,’ which is the set of exons in a genome, because the exome is believed to harbor much of the functional variation. We provide an analysis of the ∼12,500 variants that affect the protein coding portion of an individual's genome. We identified ∼10,400 nonsynonymous single nucleotide polymorphisms (nsSNPs) in this individual, of which ∼15–20% are rare in the human population. We predict ∼1,500 nsSNPs affect protein function and these tend be heterozygous, rare, or novel. Of the ∼700 coding indels, approximately half tend to have lengths that are a multiple of three, which causes insertions/deletions of amino acids in the corresponding protein, rather than introducing frameshifts. Coding indels also occur frequently at the termini of genes, so even if an indel causes a frameshift, an alternative start or stop site in the gene can still be used to make a functional protein. In summary, we reduced the set of ∼12,500 nonsilent coding variants by ∼8-fold to a set of variants that are most likely to have major effects on their proteins' functions. This is our first glimpse of an individual's exome and a snapshot of the current state of personalized genomics. The majority of coding variants in this individual are common and appear to be functionally neutral. Our results also indicate that some variants can be used to improve the current NCBI human reference genome. As more genomes are sequenced, many rare variants and non-SNP variants will be discovered. We present an approach to analyze the coding variation in humans by proposing multiple bioinformatic methods to hone in on possible functional variation.
| Characterizing the functional variation in an individual is an important step towards the era of personalized medicine. Protein-coding exons are thought to be especially enriched in functional variation. In 2007, we published the genome sequence of J. Craig Venter. Here we analyze the genetic variation of J. Craig Venter's exome, focusing on variation in the coding portion of genes, which is thought to contribute significantly to a person's physical make-up. We survey ∼12,500 nonsilent coding variants and, by applying multiple bioinformatic approaches, we reduce the number of potential phenotypic variants by ∼8-fold. Our analysis provides a snapshot of the current state of personalized genomics. We find that <1% of variants are linked to any known phenotypes; this demonstrates the dearth of scientific knowledge for phenotype-genotype associations. However, ∼80% of an individual's nonsynonymous variants are commonly found in the human population and, because phenotypic associations to common variants will be elucidated via genome-wide association studies over the next few years, the capability to interpret personalized genomes will expand and evolve. As sequencing of individual genomes becomes more prevalent, the bioinformatic approaches we present in this study can be used as a paradigm to pursue the study of protein-coding variants for the genomes of many individuals.
| Genetic variation in the protein-coding portion of genes is of significant interest in the study of human health. The focus on coding exons, or the ‘exome’, is due to the common belief that the exome harbors the most functional variation [1]. This is based on the observation that mutations that cause Mendelian diseases occur primarily in genes [1],[2]. Mutations that cause amino acid substitutions, including changes to nonsense codons, in their respective genes are the most frequent type of disease mutation (∼60%) [1]. In addition, small indels in genes account for almost a quarter of the mutations in Mendelian disease [1],[2]. Meanwhile, less than 1% of Mendelian disease mutations have been found in regulatory regions. For complex diseases, such as Alzheimer's, obesity, and heart disease, it is unknown how much variation in genes will contribute to disease, compared to variation in regulatory regions [3],[4].
There have been many efforts to re-sequence genes to identify and characterize gene variation in humans [5]–[11]. One of the proposals of the 1000 Genomes Project, an international collaboration that aims to sequence one thousand genomes, focuses specifically on re-sequencing coding exons [12]. Additionally, many groups are developing technology for high-throughput resequencing of exons [13]–[15].
Because there has been much progress in sequencing individual human genomes [16]–[20], our understanding of functional variation is an important step towards an era of personalized medicine, where a doctor could inform patients' of their disease susceptibilities based on their genome sequences. If the exome harbors much of the functional variation responsible for a person's phenotype, then identification and characterization of the individual's variation in the exome could enable individualized genomics.
In this study, we focus our analysis on the exome of an individual human by providing a detailed characterization of the variants in protein-coding regions. We present the analysis of the coding variants in an exome from a diploid human genome assembly, which was termed HuRef [16]. This paper analyzes the different types of nonsilent coding variants. There are ∼12,500 coding variants that change protein sequence in the HuRef genome. We show that most of the variation in this individual is common and appears to be functionally neutral. Furthermore, we are able to reduce the ∼12,500 coding variants down to ∼1,600 variants that potentially affect protein function and may be involved in phenotypic effects. To the best of our knowledge, this is the first analysis of the exome of an individual human, and may serve as a benchmark for future studies on the variation in human exomes.
There are several aspects to this study. The first is to describe a snapshot of what personalized genomics means today. If a person was to have his genome sequenced today, we show what insights about the individual could be gleaned from the protein coding component alone. Another aspect is that we can use the HuRef variants to improve the current NCBI reference genome, which will simplify future analysis of additional genomes. The final aspect is how one could characterize the variation obtained from sequencing many individual human genomes, and the approaches that could be developed to mitigate these studies. This is our first glimpse of an individual's human genome and as additional genomes are sequenced, many rare variants and non-SNP variants will be discovered. From this first exome, we can see what challenges one might encounter and propose approaches to face these challenges. Thus we present one approach for the analysis of coding variation in a human by detecting different trends for each variant type and demonstrating what phenotypes can be interpreted with our current knowledge.
This individual has 10,389 nsSNPs, of which 5,604 are heterozygous and 4,785 are homozygous (Table 1), where homozygous variants are loci where the alleles differ from the NCBI reference genome, but are the same within the HuRef assembly. It has been previously estimated that the number of heterozygous nsSNPs in an individual ranges from 24,000 to 40,000 [6]; our observed value of 5,604 is much less than this. This estimate was based on the nonsynonymous substitution rate based on a small number of genes, and extrapolating this value across all genes. The overestimate is partly due to the assumption of the human genome having 45,000–100,000 genes, but even if we assume the human genome has 20,000–30,000 genes, the estimate remains 1.5–2× higher than what we report. Possible explanations for the discrepancy is that the substitution rate in genes is extremely variable due to differences in local rates of mutation and recombination [6]. Thus we believe our number to be more accurate because it examines all genes rather than extrapolating from a small gene set.
The nsSNPs account for a little more than half of the coding SNPs in the diploid genome (Table 1). The 1∶1 ratio of nonsynonymous SNPs to synonymous SNPs agrees with previously published reports [6],[8]. Approximately 7% of the nsSNPs were not found in dbSNP and are thus novel. We expect novel SNPs to be rare [21]–[23] and hence observed on a single chromosome in an individual. This was affirmed with the observation that 72% of the novel nsSNPs are heterozygous.
We wanted to find nsSNPs in this individual that may be undergoing negative selection. We use allele frequency in the human population as an indicator that a variant might be under negative selection. According to the theory of natural selection, functionally neutral mutations can reach high minor allele frequencies, whereas deleterious mutations will be selected against and remain rare in a population. Rare variants do not necessarily have to be deleterious; they can be recent mutational events. Variants that are neutral, slightly deleterious, or under positive selection can become common in a population.
To see what proportion of nsSNPs may be undergoing negative selection, we retrieved the allele frequencies of these nsSNPs from the HapMap Project [24],[25] (Figure 1). The majority of HuRef nsSNPs with known allele frequencies are common (> = 0.05). For 79% of the homozygous nsSNPs, the NCBI human genome has the minor allele (MAF<0.5). Therefore, the homozygous nsSNPs in HuRef tend to represent the major alleles in the human population and it is likely that the majority of these homozygous nsSNPs are neutral because they have reached such high frequencies. Also, 19% of the homozygous alleles in HuRef have an allele frequency of 1, which suggests that NCBI contains a rare or erroneous allele at these positions. The majority of HuRef heterozygous SNPs are also common with only 9% of the nsSNPs with known allele frequencies being rare (MAF<0.05).
Since we do not have allele frequencies for all the HuRef nsSNPs, we must estimate the proportion of rare nsSNPs (MAF<0.05) in this individual. A previous simulation has estimated that ∼28% of heterozygous SNPs in an individual are rare, but that study made assumptions about human population size and its growth [26]. For 67% of the nsSNPs with known HapMap allele frequencies, we know the exact number of rare nsSNPs (56 homozygous and 326 heterozygous). For the remaining 33% of nsSNPs with unknown allele frequencies, we can estimate the proportion of rare nsSNPs based on the sequencing of a subset of heterozygous novel nsSNPs and the fraction of rare homozygous SNPs with known allele frequencies (see Methods). Using this approach, we estimate ∼1,600–2,000 rare nsSNPs in this individual's genome, the lower bound takes into account the ∼25% false positive rate for novel SNPs (see Methods). We conclude that ∼15–20% of the nsSNPs in an individual are rare, and ∼95% of the rare nsSNPs are heterozygous (see Methods). The number of rare variants found in this individual may guide our expectations when we sequence additional genomes in the future.
We wanted to identify the nsSNPs that may affect protein function and possibly be involved in human health and undergoing negative selection. Algorithms exist that predict whether an amino acid substitution affects protein function based on sequence conservation and/or structure [27]–[34]. When applied to human nsSNPs from re-sequencing projects, 0–30% of nsSNPs are predicted to affect function [9], [31]–[33],[35]. This range is based on datasets containing a relatively small number of nsSNPs (∼50–600) in a small number of genes (∼100–200).
What distinguishes our analysis from previous reports [9], [31]–[33],[35] is that we examine a single individual, rather than a population of individuals – thus we are establishing a benchmark for individualized genomics, as opposed to population genetics. Furthermore, we study all genes, unlike the previous studies that focused on certain classes of genes that were selected for their possible relevance in human health. For our study, we use the algorithm, SIFT (Sorting Intolerant From Tolerant) to determine if a nsSNP may affect protein function [33]. SIFT takes into account whether the amino acid change resulting from a nsSNP lies in the conserved region of the protein and the type of physiochemical change, and outputs a prediction to whether a nsSNP may affect protein function. We note that SIFT and other amino acid substitution prediction algorithms [27]–[34] only predict whether a nsSNP affects protein function. These algorithms do not predict whether a variant alters the processing or stability of transcripts.
Approximately 75% of the HuRef nsSNPs had SIFT predictions (see Methods), and 14% were predicted to impact protein function (Figure 2). This suggests that the majority of nsSNPs in this individual are functionally neutral. It also indicates that an individual has ∼1,500 (14% of 10,389) nsSNPs that affect protein function with deleterious effects, and we are able to confirm a previous estimate [34]. This previous estimate was obtained by taking the average nonsynonymous nucleotide substitution rate (based on a small number of genes) and extrapolating it for all genes. Meanwhile our estimate is based on the actual number of observed nsSNPs in an individual.
The ∼1,500 nsSNPs predicted to affect protein function are deleterious in the evolutionary sense with each nsSNP having a selection coefficient s≈10−3, on average [34]. A small selection coefficient suggests that the deleterious nsSNP either has negligible effects on health, has effects late in life after reproduction has occurred, or causes a disadvantage in certain environments. We term the nsSNPs predicted to affect protein function as predicted-protein-affecting nsSNPs. The other ∼9,000 nsSNPs are effectively neutral mutations with 0<s<10−4, assuming an effective population size of 10,000 [36]. Thus, the effects of nsSNPs span a spectrum ranging from neutral to mildly deleterious, and the predicted-protein-affecting nsSNPs tend to be more detrimental.
Heterozygous nsSNPs are two times more likely to be predicted as protein-affecting compared to the homozygous nsSNPs (p<0.001; Figure 2). We reason that predicted-protein-affecting nsSNPs are expected to be selected against and therefore, less likely to both reach high allele frequencies and be observed in homozygote form. Rare nsSNPs are also two times more likely to be predicted as protein-affecting compared to common nsSNPs (p<0.001; Figure 2) and this trend has been reported previously [9],[10],[35]. This suggests that a higher proportion of the rare nsSNPs are deleterious, and are more likely undergoing negative selection. Also, a higher percentage of novel nsSNPs and nsSNPs with unknown allele frequencies are predicted to be protein affecting compared to all the nsSNPs (Figure 2), but this difference only holds for the heterozygous SNPs and not the homozygous nsSNPs (Figure S1; Text S1).
Therefore, novel, rare, and heterozygous nsSNPs are more likely to affect protein function and cause phenotypic effects. Yet rare and novel nsSNPs are difficult to characterize because they are underpowered in whole-genome association studies [4]. Thus, one of the major challenges in the future of genomics is how to correlate rare and novel nsSNPs with phenotypes.
There are 105 HuRef SNPs that result in premature termination codons, or stop codons, in 103 genes, hereafter referred to as PTC-SNPs. This corresponds to 0.5% of coding SNPs. These SNPs are expected to result in the loss of their respective proteins and hence be under strong negative selection. Yet when we retrieved allele frequencies for PTC-SNPs, all of the 36 PTC-SNPs with known allele frequencies were common, which shows that not all PTC-SNPs are under strong purifying selection.
We investigate possible reasons for why some PTC-SNPs are not under strong negative selection and are able to reach high allele frequencies in the human population. Thirty percent (31/105) of the PTC-SNPs occur in segmental duplications in the human genome compared to 9.8% for synonymous SNPs. We assume gene redundancy would rescue these mutations, although loss of one copy of a gene can still have quantitative effects [37]. It is also possible that these PTC-SNPs have been mistakenly mapped due to the difficulty of assembling highly duplicated regions.
We remove the PTC-SNPs in segmentally duplicated regions from consideration and 74 PTC-SNPs in 73 genes remain. A substantial fraction (42%) of the remaining genes with PTC-SNPs are hypothetical. Hypothetical genes containing common PTC-SNPs may not be important to the human population, and using these variants may improve annotation of the human genome. If the PTC-SNPs in hypothetical genes are removed from consideration, 43 PTC-SNPs in 42 genes remain, and we sought to characterize these further.
Because we saw that three times as many PTC-SNPs occur in segmental duplications than expected, we postulated that multiple copies of a gene may permit the existence of a PTC-SNP. We examined the size of the gene family for the remaining 42 genes, and found that the median gene family size is 6, which is higher than the median gene family size for all genes, which is 2 (p<0.001). Thus, PTC-SNPs tend to occur in genes that have other homologues present in the genome, which may rescue the full or partial loss of a related gene [38]. There are only 9 PTC-SNPs in 8 genes that are non-hypothetical and unique members of their gene family.
In general, PTC-SNPs tend to occur in hypothetical proteins, segmentally duplicated regions, and gene families with multiple members. In the future, we may be able to use these trends to prioritize which PTC-SNPs are most likely to have functional consequences. All the PTC SNPs in HuRef can be found in Table S1, and none are in genes known to be involved in disease.
Indels are the second most abundant type of genetic variation, following single nucleotide substitutions and account for almost a quarter of the genetic variation implicated in disease [2]. Coding indels can significantly impact their corresponding genes if they introduce frameshifts that lead to unfinished protein products.
The HuRef genome contains a total of 739 coding indels, which consists of 281 heterozygous indels and 458 homozygous indels. To the best of our knowledge, this is the largest set of human coding indel variants identified to date [39].
We find an enrichment of indels that have sizes that are multiples of 3 in the HuRef coding indel set (Figure 3). We will refer to indels that have lengths divisible by 3 as 3n indels and indels with lengths not divisible by 3 as non-3n indels, where n is an integer. In coding regions, a non-3n indel would cause a frameshift that usually leads to a truncated protein product whereas a 3n indel would cause deletion or insertion of amino acid(s). By comparing the diversity rates between coding indels and indels genome-wide (Figure S2), we find that 94% of non-3n coding indels have been eliminated by natural selection. In contrast, only 46% of 3n coding indels have been eliminated. This signifies that 3n coding indels are not as strongly selected against as non-3n indels.
Many of the indels are due to polymorphism in tandem repeat sequence. Only 6% of coding regions are classified as tandem repeats, yet 52% (381/739) of all coding indels occur in tandem repeats. The majority (73%) of the tandem repeats in coding regions have a periodicity of 3 and account for 66% (252/384) of the 3n bp coding indels. This suggests that local regions in a protein coding sequence can be prone to mutation that will either remove or insert amino acids into the protein product.
In contrast to SNPs, many indels are not validated and their allele frequencies are unknown due to difficulty in their ascertainment using either sequencing or genotyping technologies [39],[40]. To validate these indels, we verified if the indel was confirmed in the chimpanzee genome sequence [41]. We determined that 24% (181/739) of the coding indels correspond to the chimpanzee sequence. These indels are likely to be common in the human population if the indels occurred before the divergence of chimpanzee and human, although an alternative possibility is that some of these mutations are recurring events. This signifies that at least 24% of the HuRef coding indels are real.
We sought explanations of how an individual human genome could have 739 indels affecting 607 genes, and yet the individual appears to lack severe phenotypic effects. Small coding indels (< = 30 bp) account for 84% (621/739) of the indels. In the following section, we analyze the 621 small coding indels (< = 30 bp). Large indels are discussed in a later section.
Many of the small coding indels were located at the N- and C-termini of their respective proteins. We calculated the relative position of the indel in the protein by dividing the indel's position by the total protein length. With this metric, one would expect that the indel's position would be uniformly distributed across the protein. Instead, indels tend to occur at the N- and C-termini of their proteins (Figure 4). If a coding indel occurs at the C-terminus of the protein, it may not affect the function of the protein because most of the protein product has been translated successfully. If a coding indel occurs at the N-terminus of the protein, this may be rescued by a downstream start codon in the coding region (see Figure S3 for an example). This suggests that indels at the N- and C-termini of their proteins are functionally neutral and a future study using these indels to propose alternate start and stop sites could improve human gene annotation.
Furthermore, a high proportion of homozygous coding indels were located near an exon boundary. A large proportion of the small homozygous coding indels were within 10 bp of the exon boundary: 27% (101/344) compared to 12% (34/277) for small heterozygous coding indels. Close inspection of the homozygous indels near exon boundaries showed that these indels were near small introns and the HuRef allele corrects the NCBI reference genome to provide a better gene model. Figure 5 shows an example where a 1-bp homozygous coding insertion borders a 2 bp intron, so rather than causing a frameshift, the small intron is replaced by an amino acid. Incorporating the homozygous indel from HuRef likely produces the correct protein sequence. Therefore, it is very likely that the HuRef assembly has the correct sequence, and could potentially be used to correct gene structures that were based on the NCBI human genome. Many of these small homozygous indels within 10 bp of an exon boundary were also found in chimp (61% (62/101)), further evidence that these indels are the ancestral alleles and likely to be accurate.
We re-sequenced seven homozygous non-3n indels that were either near exon boundaries (4), and/or confirmed by chimp (4), and/or at the N-terminus of a protein (1) (see Methods). These non-3n indels would supposedly cause frameshifts, yet all seven indels were found to be common (MAF>0.05) and four were determined to have an allele frequency of 1 (Table S2). This indicates that NCBI has a rare or erroneous allele at these positions and suggests that homozygous indels located at certain locations may correct the NCBI genome sequence.
We assume that indels near protein termini and near exon boundaries are functionally neutral. Removing these indels from the indel set reduces the entire set by 45%, with 342 indels remaining. Of the remaining indels, the fraction of indels that have length 3n increases from 49% (303/621) to 60% (205/342). This suggests that while the termini of a protein may be able to tolerate the presence of non-3n bp indels, elsewhere in a protein there is a greater preference for 3n indels.
We categorize the remaining indels by their 3n and non-3n lengths. Whereas 35% (71/205) of 3n indels are found in hypothetical proteins, 56% (77/137) of non-3n indels are found in hypothetical proteins (p<0.001). This suggests that non-3n indels occur in genes that can tolerate deleterious mutations, which may be pseudogenes or genes under weak selective constraints and we may be able to use these variants to identify genes that are likely not important for human health. We also noticed that the 3n coding indels that are not in tandem repeats tend to avoid regions of the protein that are highly conserved (Figure S4).
In summary, many of the indels are located at exon boundaries or protein termini and these are likely to be functionally neutral. The remaining coding indels typically have sizes of 3n, and those that do not tend to occur in hypothetical proteins. To the best of our knowledge, these trends have not been previously reported, and this suggests that a substantial fraction of coding indels are functionally neutral. In the future, we can use the HuRef indels to improve gene annotation and our observations to develop a method that distinguishes between functional and neutral indels.
We sought to understand genes missing large regions of coding sequence, leading to a gene and/or exon deletion that would render the gene non-functional. Therefore, we focused on deletions of coding sequence in this analysis. We will discuss newly observed genes and insertions of coding sequence in existing genes in a future manuscript.
We identified 1,454 exons in 1,046 genes where at least half of the coding exon's sequence was missing from the HuRef assembly. Further investigation showed that genes with “missing” exons are most likely due to low coverage or assembly issues with repetitive regions rather than the human individual truly missing part of a gene (see Methods and Text S1; Figure S5). This was confirmed by resequencing a subset of the missing exons in the HuRef sample and validating that in fact, most of the “missing exons” are present in the HuRef sample (see Methods and Text S1).
We examined the HuRef variants in genes known to be involved in disease (based on the OMIM database [42]) to make correlations between the HuRef variation and possible phenotypes. There are a total of 682 nsSNPs in 443 disease genes (Table S3). The allele frequency distribution and fraction of predicted-protein affecting SNPs are similar to non-disease genes (Figure S6).
We examined the nsSNPs from dbSNP [43] that were found in the disease database OMIM [42] (see Text S1). Using this approach, seven HuRef nsSNPs were found to be associated with disease (Table 2). The HuRef individual is heterozygous for all seven SNPs and most of these disease-associated SNPs are common in the population. It may be considered surprising that these nsSNPs are common since they were found in the OMIM disease database. This is due to the fact that we looked for overlap with nsSNPs that are in both OMIM and dbSNP, and dbSNP tends to be biased for common SNPs [21].
From these seven variants, one could simplistically infer that the HuRef individual has an increased risk for eating disorder (BDNF), 1.5-fold reduced risk to multiple myeloma (LIG4), an increased risk to prostate cancer which can be rescued by taking vitamin E supplements (SOD2), and allergic tendencies (SPINK5) (Table 2). Some of the SNPs have known interactions with environmental factors (SOD2 and BDNF in Table 2). However, the published risks for these variants are from population-based studies, and may not apply to this specific individual because it does not take into account other interacting genetic loci and his environment [44]–[47]. Therefore, Table 2 does not show exact risks for this individual and predicting the phenotype or lack of phenotype of this individual is premature. These variants, like many of the risk variants being uncovered by genome-wide association studies, have low risks and we may not have a clear understanding of their clinical utility until all the relevant factors (both genetic and environmental) for a particular disease have been elucidated [44]–[47].
These seven well-studied examples demonstrate the complexity of trying to interpret variants and their impact, but there are still many variants in this individual that are uncharacterized. For the 682 nsSNPs in 443 disease genes, 27 are rare (MAF<0.05), 18 are novel, and 81 are predicted to affect function (Table S3). Interpretation of these variants is difficult because of the absence of literature for many of the observed variants. One challenge is that these variants, even if they affect protein function, could be phenotypically neutral in certain contexts (see SOD2 and BDNF in Table 2 and [48]). Also, even if there is evidence that a nsSNP is under negative selection (e.g. predicted to affect function and/or rare), it is not straightforward to interpret a possible phenotype because mutations at different locations in the same gene can have different effects [49]. The difficulty of inferring phenotypic consequences from a variant is depicted in the following example. rs562556, which is homozygous in HuRef and has an unknown minor allele frequency, introduces the amino acid substitution V474I in PCSK9, and this amino acid substitution is predicted to affect protein function. Because defects in PCSK9 cause familial hypercholesterolemia (OMIM∶607786), one could speculate that this SNP could affect the donor's cholesterol levels. However, extensive functional studies of this variant and others are necessary before any conclusions can be made. Because only 1% (7/682) of the nsSNPs in disease genes in this individual human have been well-characterized, this indicates that we are only at the beginning of relating genotypes to phenotypes, even for the well-characterized disease genes.
There were 28 indels in 26 disease genes (Table S4). Only 3 out of the 28 indels have lengths that are not multiples of 3n, and would cause frameshifts. Two of these indels appear to be annotation issues and are likely to be functionally neutral (see Text S1). The third indel is in ACOX2. This protein is involved in lipid metabolism, and patients with Zellweger syndrome lack this protein [50]. Because the indel is heterozygous in the HuRef individual and one functional copy is present, the individual may not be adversely affected by this mutation.
For the remaining 25 indels with length 3n in 23 disease genes, 84% were in tandem repeats. Some of these genes are known to cause disease due to polyglutamine and polyalanine repeat expansions (AR, ATXN2, ATXN3, HD, TBP). For these genes, we confirmed that the number of repeats in the HuRef genome falls within the range of what is observed for unaffected individuals. In addition, this individual is heterozygous for a 24-bp duplication in the CHIT1 gene, which activates a cryptic 3′ splice site and causes chitinase deficiency [51]. Even though the indel produces a nonfunctional protein, the indel is observed at an allele frequency of 23% in the general population. One possibility for its high incidence is that this indel may provide a selective advantage against fungal pathogens [51].
To examine selective constraints on different regions across the genome, we estimated the diversity rate θ (see Methods). We calculate the diversity rates for regions throughout the entire genome as well as calculating the diversity values for genic regions (Table 3). Values of θ tend to be negatively correlated with the strength of selection, where low values of θ tend to indicate strong selective pressures, while high values do not. The order of θ is: coding in disease genes<coding<conserved noncoding intronic<splice sites<3′UTR≈5′UTR≈conserved noncoding intergenic<promoter<introns<repeats. This indicates that coding regions in disease regions are the most selectively constrained regions and repeats are the least. Some diversity values and trends have been reported in previous publications [6]–[8],[52],[53], and our results are in agreement with these values. However, to our knowledge this is the first study with such an extensive list of regions. We also find indels are significantly under-represented in coding regions; there is a 43∶1 SNP∶indel ratio for coding regions compared to a 7∶1 SNP∶indel ratio genome-wide. It is reasonable to observe stronger selection against indels in coding regions where they can introduce frameshifts.
We further explored if the observed HuRef variation could indicate whether certain genes were under stronger selective constraints than others. We identified 538 genes containing common nsSNPs located in conserved regions of the protein and are predicted to affect protein function. We called this set of genes the Commonly-Affected genes. In parallel, we identified 79 genes containing rare nsSNPs predicted to affect protein function, and we termed this set the Rarely-Affected genes.
We hypothesized that Commonly-Affected genes would be under weaker selective constraints since predicted-protein-affecting SNPs, common in the human population, were found in these genes. Ka/Ks, is a metric used to quantify selection of a gene and is calculated as the ratio of the nonsynonymous (amino-acid affecting) substitution rate to the synonymous substitution rate. A high Ka/Ks ratio indicates that a gene is undergoing weak selection, although positive selection is also a possibility if Ka/Ks is >1 across the entire gene or part of the gene [54]. We observe that the Commonly-Affected genes tend to have higher Ka/Ks ratios than the Rarely-Affected genes (p = 0.09; Figure 6). This suggests that Commonly-Affected genes may not be under strong selective constraints. Presumably, mutations affecting gene function in the Commonly-Affected genes will not have significant consequences on human health, which is why these predicted-protein-affecting nsSNPs can rise to high allele frequencies. In support of this hypothesis, the Commonly-Affected genes are dominated by olfactory receptors (78/538) which is consistent with the previous observation that humans do not depend on olfaction to the same extent as other species [55].
In the future, extending this type of analysis to nsSNPs from HapMap or additional human genomes may allow identification of additional Commonly-Affected genes. This could help improve the scientific community's knowledge of the human genome by identifying which genes may not play an important role in human health if mutated.
We also compared the genome properties of Dr. James Watson [20] to those of Dr. Craig Venter. SNP diversity rates based on Dr. Watson's genome are slightly elevated compared to Dr. Craig Venter's (Table 3), but in general, follow a similar trend as stated previously. The one exception is simple repeats, which have a 2-fold higher SNP diversity rate in Dr. Watson compared to Dr. Venter; this may be due to differences between sequencing technologies.
Of the 3.1 million SNPs in Dr. Venter's genome that map to the NCBI genome, 56% are shared with Dr. Watson. There are 606,719 and the 288,723 novel SNPs that map to the NCBI human genome in Dr. Watson's and Dr. Venter's genomes, respectively. One possible reason for the smaller number of novel SNPs in Dr. Venter's genome is because Dr. Venter's genome is partially represented in the Celera human genome assembly [16],[56], and variants have been mined from the Celera assembly and subsequently deposited in dbSNP [57]. Of the novel SNPs, only 32,528 SNPs are shared between the two individuals. This demonstrates the value of sequencing additional human genomes to discover novel variants.
We compare the number of nsSNPs and coding indels between the two exomes (Table 4). Similar numbers of nonsynonymous variants are detected in both individuals. However, 14% and 20% of nsSNPs are predicted to affect function in Dr. Venter's and Dr. Watson's exomes, respectively. One possible reason for this difference could reside in the different prediction algorithms employed [20],[30],[33]. Future analyses that use a standardized approach may clarify this apparent difference. The number of indels in Dr. Watson's genome is less than half of the number of indels we observe in HuRef (Table 4), most likely because 1 bp indels were discarded [20]. As more individuals are sequenced, scientists will be able to establish general trends and the average values for metrics that characterize a human individual's genome.
Coding exons are believed to be rich in functional variation because many coding mutations have been found to cause phenotypic effects [58]. As a result, re-sequencing projects tend to focus on coding regions [5]–[11], multiple technologies that specifically target the exome are being developed [13]–[15],[58], and the 1000 Genomes Project may specifically target the exome [12]. This motivated us to focus on the variation in an individual's exome to see what insights could be gleaned. Protein-coding exons are thought to amount to a little less than half of the functional portion of the genome; noncoding highly conserved regions constitute the other half [59]–[62]. We did not analyze noncoding variants even though they can be involved in disease [3], [63]–[66] because exons are the best characterized regions that correlate to phenotypes and it is difficult to characterize the impact of noncoding variants at this time. As our understanding of non-genic regions increases, so can we expand our interpretation since the data for this individual's entire genome is available.
This is the first study of an individual's exome and we establish what one may expect to observe from the variation in the exome of an individual. We show that the majority of coding variants in a human are neutral or nearly neutral. This is not unexpected, since we know that this genome creates an individual who has survived past 60 years of age. We also find that within an individual, the basic principles of genetics are followed. Additionally, we examined the variation in genes known to be involved in disease, and found no indication that the individual should have a severe disease, which matches the phenotype currently known.
Despite having a human's complete genome sequence, we are only at the tip of the iceberg for understanding how an individual's genotype and phenotype are related. One significant challenge is that the phenotypic effects of the majority of genes are unknown. Currently, only 7% of genes are annotated with OMIM disease associations so that it is difficult to predict the phenotypic effects of variants for a large proportion of genes. If one is to rank genes by importance and effect on phenotype, then based on the results of this study, one might consider that genes containing PTC-SNPs, frameshifting indels, and damaging nsSNPs that are common in the human population to be under weak selection, and variants in these genes may not be as relevant to human health. Several groups have used gene ontology, literature, and other sources to predict potential disease genes [67],[68], we propose that one can also use observed human variation to increase our understanding of the human genome.
Even if a gene is known to be involved in disease, it is difficult to understand if a variant in the gene will have a phenotypic effect. We found that 99% of the nsSNPs in disease genes could not be characterized by current literature. Different mutations in the same gene can cause different phenotypic effects [49], thus making it difficult to interpret possible phenotypes. Furthermore, some variants have phenotypic effects only under certain environments (see SOD2 and BDNF in Table 2 and [48]). Also, when looking at complex phenotypes, multiple variants in coding and non-coding regions are likely to be involved [63]–[66]. This genetic complexity, as well as exposure to various environmental factors, will need to be taken into account in assessing risk for various diseases.
How can geneticists start to grasp the significance of phenotype-genotype correlations? This question is especially relevant to companies offering personalized genomics to their consumers (e.g. 23andMe, Navigenics, deCODE Genetics). When looking amongst the human population, there are many rare SNPs, but when looking at a single human individual, the majority of the SNPs are common [6],[69],[70], and in this study we estimate that >80% of the nsSNPs in an individual are common. Therefore, understanding which common variants are involved in common disease will greatly benefit an individual, because common variants account for a significant fraction of the variation in each human. Recent genome-wide association studies have identified common variants implicated in disease [71]–[73] and these studies will continue to find common disease-associated variants in the near future. These discoveries will be valuable for interpreting a large proportion of an individual's genome. However, one should be cautious in the interpretation of these variants because variants with low associated risks may not necessarily have good predictability in the clinical setting [44],[45].
In contrast, rare variants are harder to study because genome-wide association studies are insufficiently powered to detect rare variants [4]. We found that a higher fraction of rare nsSNPs were predicted to affect protein function compared to common nsSNPs, in agreement with previous studies [9],[10],[35]. This suggests that a small proportion of a large number of common variants and a larger proportion of a small number of rare variants will contribute to the health of a human individual. Genome-wide association studies tend not to have the power to detect rare etiological variants [4] so that predicting whether a rare mutation in an individual causes disease without any other phenotypic information is extremely difficult. Therefore, one of the major future challenges in personalized genomics is the interpretation of the effects of rare variants found in an individual, especially if this information will be relayed back to the individual and could impact the person's lifestyle.
In addition to interpretation and analysis, much effort was expended in ensuring that a variant was authentic (see Methods). There could be unintended negative consequences for telling a person that they have a disease variant, when in actuality the variant could be detected in error. Common reasons for our false positive variant calls were technical sequencing error, low sequence coverage, and low-complexity sequence. When interpreting an individual's genome that can potentially impact a person's lifestyle, manual curation, editing, and verification by other technologies seems prudent.
The human genome sequence has had a significant impact on research since its availability in 2001 [56],[74], but our analysis suggests that rare or erroneous alleles may have been incorporated into the NCBI human genome sequence, and that this sequence can be corrected and improved. Because one of the goals of the 1000 Genomes Project is to improve the human reference sequence [75], our study points to where such improvements can be made. We find that ∼80% of the homozygous SNPs in HuRef tend to be the major allele in the population (Figure 1), and ∼20% of the homozygous HuRef nsSNPs had allele frequencies equal to 1. Thus, at the majority of HuRef homozygous positions, the NCBI human reference sequence has the minor allele. If the scientific community sequenced many individuals, it could determine the major allele at each position in the genome. If the major allele was incorporated into the coding sequence, and this was used in subsequent gene prediction models, then the predominant form of the protein in the human population would be represented instead of a rarer form. Also, by using the common allele instead of the sometimes rarer NCBI allele, the number of perceived variation would be reduced when comparing human genomes. For example, if the NCBI human genome sequence were to incorporate the major allele and HuRef was then compared to this modified NCBI genome sequence, then we estimate the new number of HuRef homozygous SNPs genome-wide would be ∼300,000 (0.2 * 1.45 million [16]) and the number of total variants between HuRef and the modified NCBI would reduce from 4.1 million [16] to 3 million. Similarly, if we remove the homozygous nsSNPs that correspond to the common allele in European population (AF>0.5), then the fraction of rare nsSNPs increases by 6%. This demonstrates the importance of the human genome reference sequence for the evaluation of variation in individual human genomes.
The scientific community could also make use of coding indel variation to correct and improve gene annotation. Indels near exon boundaries appear to provide the correct gene models to give the appropriate protein product (Figure 5, for example) [76]. Gene annotation for translation starts and stops can be further refined based on our observation that coding indels are found frequently at the N- or C-termini of their proteins. Frameshifting-indels at the N-termini of proteins could indicate that a translation start site further downstream may be the true start codon, or at least an alternative start codon can still yield a functional protein. Indels such as these may be polymorphic and so accounting for these indels could simplify future analyses on exomes as they could be quickly regarded as functionally neutral and reduce the total number of indels that need to be analyzed.
We also observed a trend where common predicted-protein-affecting nsSNPs, PTC-SNPs and frameshift-inducing indels tend to occur in hypothetical genes. This suggests that these genes are not under strong selective pressures and mutations in these genes may not be relevant in the human population. Future studies with more human sequences identifying additional nonfunctional mutations in genes would help us confirm whether these genes are essential.
Our exome analysis is currently limited to one individual. However, there will be significant benefits from sequencing many individuals whose phenotypes are known. One can envision collecting the genetic variation from these genomes and grouping individuals based on their respective phenotypes. Then for each phenotype, one may discover the genes which are involved in disease by looking collectively at the rare and common variants [11],[77],[78]. The analysis can also be strengthened by analyzing pathways instead of individual genes [78].
Furthermore, whole genome sequencing means we need not be limited to the exome. Using whole genomes, one could look for clustered mutations in conserved regulatory elements, especially since many association studies have found disease-associated loci in non-coding regions [63]–[66]. To assess the role of noncoding variation, we examined HuRef variation in and around genes involved in the melanoma pathway because the HuRef donor has reported a case of melanoma (Figure S7). We found that the majority of the variants occurred in conserved noncoding regions. This suggests that it may be insufficient to sequence just the exome and it is important to understand all types of variation, coding and noncoding, as well as interactions with the environment when studying phenotypes.
We have filtered the initial set of ∼12,500 coding variants that affect protein sequence to a substantially smaller set that are most likely to have major effects on gene function (Figure 7). The trends that we have detected suggest we can reduce the number of putative functional coding variants by ∼8-fold and will provide a future guide for how one can analyze coding variants when additional human genomes are sequenced in the future. Additionally, the variants found here and in future studies may be used to improve our understanding of the human genome by correcting gene annotation and identifying genes not likely to be relevant to human health. We anticipate that this study will help guide the scientific community's expectations and experimental design in future genome sequencing projects.
We used the filtered variant set as described in [16]. We define homozygous variants as loci where the alleles differ from the NCBI reference genome, but are the same within the HuRef assembly. This variant set was used to generate the diversity values in Table 3. For all other sections in this manuscript, quality inspection of variants was performed. To assure the quality of novel coding variants, we inspected manually the sequencing traces of novel heterozygous nsSNPs, all heterozygous PTC-SNPs, and all coding indels less than 20 bp in length. The sequence traces for these coding variants were extracted and three people independently reviewed the traces, by examining the quality of the traces and determining whether the variant was correctly called. If at least two people confirmed the existence of the variant, the variant was deemed acceptable, otherwise the variant was discarded. 35% (424/1196) of the novel heterozygous nsSNPs, 12% (9/73) of the heterozygous PTC-SNPs, and 33% (355/1088) of the coding indels were discarded. This may suggest that a significant fraction of the variants reported in [16] are dubious. However, this analysis is restricted to coding variation which is known to be under strong selection compared to the rest of the genome. Hence, there will be fewer real variants in coding regions and a higher proportion of the novel coding variants will be false positives. 20% (123/611) of the homozygous indels were reclassified as heterozygous because there was trace evidence for a second allele.
The importance of filtering is demonstrated with the following observation. Manual inspection reduces the number of novel nsSNPs by a third, but especially filters out a higher proportion of predicted-protein-affecting nsSNPs. Prior to filtering, the number of novel heterozygous predicted-protein-affecting nsSNPs is 195, after filtering this is more than halved to 89 novel nsSNPs predicted to affect protein function.
SNPs not found in dbSNP v. 126 [43] were designated as novel. All allele frequencies were based on the CEU samples genotyped from the HapMap Project [24],[25], unless otherwise stated. Allele frequencies for 72% (3429/4785) of the homozygous nsSNPs and 63% (3544/5604) of the heterozygous nsSNPs were obtained. For heterozygous SNPs, we report the minor allele frequency (MAF). For homozygous SNPs, we report the CEU allele frequency of the allele observed in the HuRef genome. We define common SNPs as SNPs with allele frequencies > = 0.05 and rare SNPs with allele frequencies <0.05.
The amino acid changes resulting from coding variants were determined by SNPClassifier, an internally developed software tool. The HuRef variants, their alleles, and positions in genomic coordinates, are provided as input into SNPClassifier. Annotation is automatically retrieved from Ensembl and is used to assign variants to defined gene categories. Variants in or near genes can be subtyped as: promoter (1 kb upstream of the transcription start site), intronic, 5′ UTR, 3′UTR, coding, or downstream of the transcript (1 kb). Coding SNPs are designated either as synonymous or nonsynonymous and coding indels are designated as either frameshift or amino acid insertions/deletions. The resulting protein product from coding indels that introduce frameshifts is also output.
SIFT predictions for nonsynonymous SNPs were obtained by using SIFT 2.1.1 [33]. The protein sequences containing nonsynonymous SNPs were searched against SwissProt-Trembl 54. Confidence in predictions is measured by the median sequence information, we used a cutoff of 3.5 for confidence. Approximately 75% (7,781/ 10,389) of the nsSNPs had SIFT predictions, the remaining 25% did not have a sufficient number of homologous sequences that are needed for prediction.
We estimate the number of rare nsSNPs with allele frequency (AF) <0.05 in an individual. For the 67% nsSNPs with known AFs from the HapMap Project, there are 56 rare homozygous nsSNPs and 326 rare heterozygous nsSNPs. For the 1,356 homozygous nsSNPs with unknown AFs, the percentage predicted to affect function is similar to that seen for homozygous SNPs with known AFs (Figure S1). If the homozygous nsSNPs had a higher proportion rare SNPs, then a higher fraction should be predicted-protein-affecting but because they are similar, we assume that the homozygous nsSNPs with unknown AFs have a similar proportion of rare SNPs as the homozygous nsSNPs with known AFs. Because 1.6% (56/3429) of the homozygous nsSNPs with known AFs are rare, we estimate ∼22 (1.6% * 1,356) of the homozygous nsSNPs with unknown AF are rare, so in addition to the 56 rare homozygous nsSNPs with known AF, there is a total of ∼80 rare homozygous nsSNPs in this individual. For heterozygous nsSNPs, there are 326 heterozygous rare nsSNPs with known MAF, and 2,060 heterozygous nsSNPs with unknown MAF. From sequencing, as much as a quarter of the heterozygous nsSNPs with unknown MAF could be false positives, although this estimate is likely to be an upper bound (see Sequencing for Variant Validation in Methods). Therefore the range of novel heterozygous nsSNPs falls within ∼1,550–2,060. We also ascertained from sequencing that ∼75% of the heterozygous novel nsSNPs are rare. Therefore, we estimate that ∼1,200–1,550 of the heterozygous nsSNPs with unknown MAFs are rare and in total, there are ∼1,500–1,900 rare heterozygous nsSNPs. Thus, we estimate ∼1,600–2,000 rare nsSNPs in this individual's genome, and ∼95% of the rare nsSNPs are in heterozygous state.
For an indel's location, we calculated the relative position of the indel in the protein by taking the first amino acid position affected by the indel. We divided the position by the total length of the protein, so that a relative protein position value close to 0 indicates that the indel affects the N-terminus of the protein, and a relative protein position value close to 1 indicates that the indel affects the C-terminus of the protein. We designate that an indel affects the N-terminus of a protein if the relative protein position is between 0 and 0.1; an indel affects the C-terminus of a protein if the relative protein position is between 0.9 and 1.0. Thus, an indel is said to affect the N-terminus or C-terminus of the protein if it lies within the first 10% or last 10% of the open reading frame, respectively.
To examine whether an indel occurs in a conserved region of the protein, the sequence alignment of the protein sequence with homologues from other organisms were retrieved from Ensembl. At every position in the protein alignment, sequence conservation was calculated [79]. The conservation value at the indel's position is compared with all other positions, and the percentile rank is calculated. If the number of sequences in the alignment was less than 10, the data point was removed.
The HuRef assembly was mapped by an assembly-to-assembly comparison to the NCBI build 36 human reference genome [16]. Regions in NCBI reference that were missing in the human diploid assembly were identified. We intersected the missing regions with coding exons greater than 50 bp in length and ensured that at least 50% of the exon was missing from the HuRef assembly in order to consider the exon. To double-check that the missing sequence was not in unassembled sequence, we searched the exonic sequence using MEGABLAST [80] against the HuRef assembled sequence and the unassembled singletons. MEGABLAST hits greater than 95% identify and with 50 bp minimum length were kept. We decided exons were not truly missing if >90% of its length were covered by these MEGABLAST hits. The final set consisted of 1,454 exons in 1,046 genes.
We removed the genes located on sex chromosomes because the sex chromosomes are known to have low coverage [16]. After removing these genes, there were 719 genes with 880 missing or partial exons. To investigate read depth for this set of exons, we re-mapped all untrimmed reads from [16] to the set of exons using ‘snapper’ (http://kmer.wiki.sourceforge.net/), a seed-and-extend mapper. All 20-mer seeds were extended, and any alignments over 94% identity were reported. As a control, we also remapped reads to a set of exons that were randomly selected from all exons. Whereas the control exons show a normal distribution with the median number of reads centering at 7.6, the missing exons show a bimodal distribution with either very few reads or many reads (Figure S5b). This reflects that genes with “missing” exons are most likely due to assembly issues with repetitive regions or low coverage. 66% of the “missing” exons have an average read depth of less than 2 reads, which emphasizes the importance of adequate coverage in a human genome.
We generated PCR primers to 15 regions in 12 genes with ‘missing exons’, 9 PTC-SNPs, 15 coding indels, and 26 novel heterozygous nsSNPs (Table S2). These 65 PCR primers consistently amplified their cognate genomic regions in 46 unrelated CEU individuals and the HuRef sample. The DNA for the HuRef sample was extracted from whole blood (see Methods in [16]). We sequenced the PCR products using Sanger dideoxy sequencing (see [16] for sequencing protocol). The 46 unrelated CEU individuals were part of the HapMap CEU panel, and their Coriell identifiers are provided in Table S2.
Of the 26 heterozygous novel nsSNPs, 6 failed to be confirmed in the HuRef sample and instead matched the NCBI allele. There was also 1 nsSNP that failed to be confirmed in the HuRef sample but was observed in other samples and this was considered to be a false negative. This suggests that the false positive error for HuRef's novel nsSNPs is ∼25% (23% = 6/26). This estimate is likely to be an upper bound due to the following reason. The 26 nsSNPs occurred in non-hypothetical genes, and nsSNPs in hypothetical genes may be under little or no selective pressures compared to nsSNPs in non-hypothetical proteins and the former can reach high allele frequencies. Hence, this false positive error may be inflated. For the novel nsSNPs that we could confirm in HuRef, the mean MAF of the novel SNPs was 0.09 and 74% (14/19) of the SNPs were rare (MAF<0.05).
For the PCR products spanning missing exons, 14 regions from 11 genes were successfully amplified in the HuRef sample and this confirmed that HuRef is not missing exons for these genes (Table S2). In the 12th gene PRED58, a 66 bp coding deletion in HuRef was observed but this was seen in all other DNA samples, suggesting that NCBI has the rare or erroneous allele.
All genome coordinates are with respect to NCBI build 36 and all gene designations are with respect to Ensembl v. 41. A gene was considered hypothetical if in its gene description, it had no description or was described as an “open reading frame”, an “orf” which signifies an open reading frame, a cDNA clone, putative, probable, uncharacterized, “similar to” another protein, a pseudogene, a fragment, hypothetical, a novel protein, novel transcript, or if it was invalid as described in Clamp et al. [81]. Under this classification, there were 20,561 non-hypothetical genes and 10,624 hypothetical genes consisting of 29,401,727 bp and 6,176,706 bp respectively. Ka/Ks is the ratio of the nonsynonymous substitution rate to the synonymous substitution rate. Ka/Ks values based on human-mouse orthologous gene pairs were retrieved from Ensembl Biomart. In Table 3, constitutive exons are those coding exons that are expressed in 100% of the transcripts for its given gene. If a coding exon was present in <50% of the transcripts, it was designated as an alternative exon. Splice sites include the 20 bp within exon boundaries (10 bp intronic, 10 bp exonic for each exon boundary).
Segmental duplication regions were taken from the UCSC genomicSuperDups file (>1 kb length, >90% identity). TandemRepeatFinder [82] was used to designate tandem repeats using the parameters match = 2, mismatch = 5, delta = 5, PM = 75, PI = 20, minscore = 35.
Non-genic conserved regions were taken from phastConsElements17way that were > = 50 bp in length. If any part of the PHAST region intersected with coding, 5′UTR, and/or 3′UTR, the PHAST conserved element was removed. Therefore, the conserved regions in Table 3 were not overlapping or bordering coding, 5′UTR, or 3′UTR regions.
We estimate diversity θ [83] as θ = K/aL, where . K is the number of variants identified, L is the number of base pairs, and n is the number of alleles. For indels, K is the number of indel events. In the case of a single diploid genome, n = 2, so a reduces to 1. Then θ = K/L which is simply the number of heterozygous variants divided by the length sequenced. The 95% confidence interval for θ is [0, θ+2θ] or [0, 3θ], as calculated in [16].
Diversity values were calculated for the various types of regions listed in Table 3 and the counts for these values can be found in Table S5. We also attempted to look at the diversity values for gene ontology categories, but were unable to do so because of the low numbers of coding variants per gene (data not shown).
Diversity values for Dr. James Watson's genome were calculated using the 1.86 million heterozygous SNPs reported in [20]. For the denominator L, we assumed the entire chromosome was covered by reads and used the chromosome lengths from the UCSC genome browser. If this assumption is not true, then an inflated L will underestimate θ. Diversity values for indels were not calculated because indel data was not available for Dr. Watson's genome.
|
10.1371/journal.pntd.0006918 | Efficacy comparison between long-term high-dose praziquantel and surgical therapy for cerebral sparganosis: A multicenter retrospective cohort study | Sparganosis is a parasitic infection caused by the plerocercoid larvae of Spirometra mansoni in East and Southeast Asia. The plerocercoid larvae sometimes invade the encephalon, resulting in severe cerebral sparganosis. Surgical removal of the larvae is considered a standard therapy for cerebral sparganosis. In contrast, the efficacy and safety of long-term, high-dose praziquantel treatment for cerebral sparganosis have not been explored.
In this multicenter retrospective study, we assessed the records of 96 patients with cerebral sparganosis who consulted at three medical centers from 2013 to 2017. Forty-two patients underwent surgical lesion removal, and the other 54 patients received long-term, high-dose praziquantel (50 mg/kg/day for 10 days, repeated at monthly intervals). The primary outcome was the complete disappearance of active lesions on cerebral magnetic resonance imaging. The secondary outcomes included the modified Rankin scale score at 90 days, incidence of seizure, eosinophil count, and serological Spirometra. mansoni antibody titer. The efficacy of praziquantel treatment was similar to that of surgical lesion removal for cerebral sparganosis with respect to both the primary outcome and secondary outcomes. Although binary logistic regression models also supported the primary outcome after adjustment for age, sex, lesion location, and loss to follow-up, some unavoidable confounders might have biased the statistical power. No significant clinical complications or laboratory side effects occurred in the praziquantel group with the exception of a relatively benign allergic reaction.
In this small-sample, nonrandomized, retrospective exploratory study, some patients with cerebral sparganosis were responsive to long-term, high-dose praziquantel with an efficacy similar to that of surgical lesion removal. These findings increase the treatment flexibility for this serious infection.
| Sparganosis is most prevalent in developing countries in East and Southeast Asia, probably because public health strategies have not prioritized its prevention. The plerocercoid larvae of Spirometra mansoni sometimes invade the brain parenchyma, resulting in cerebral sparganosis. In general, surgical removal of the larvae is considered a standard therapy for cerebral sparganosis. One alternative treatment for sparganosis is short-term, low-dose praziquantel, which has had limited success. However, the efficacy and safety of long-term, high-dose praziquantel treatment for cerebral sparganosis have not been explored. In this study, we evaluated the effectiveness of long-term, high-dose praziquantel for treatment of cerebral sparganosis in China. We conducted a retrospective exploratory study using routinely recorded data from 96 patients at three medical centers. Forty-two patients underwent surgical lesion removal, and the other 54 patients received long-term, high-dose praziquantel. Treatment of cerebral sparganosis by long-term, high-dose praziquantel showed an efficacy similar to that of surgical lesion removal with respect to the primary outcome (complete disappearance of the active lesions on cerebral magnetic resonance imaging). However, this was a small-sample, nonrandomized retrospective study, and the results should be further confirmed by a large-sample prospective study or other studies.
| Sparganosis is a type of parasitic zoonosis associated with infection by the larval cestode of Spirometra mansoni [1]. Most cases have been reported in East and Southeast Asian countries, especially in China [2], South Korea [3], Japan [4], and Thailand [5]. Humans become infected with Spirometra mansoni by drinking water contaminated with procercoid-infected copepods, eating undercooked meat of snakes or frogs infected with Spirometra mansoni, or applying the flesh or skin of an infected frog or snake to poultice open wounds [1, 6]. The plerocercoid larva usually affects the subcutaneous tissue or muscle in a human host [7]. However, it can also sometimes invade the encephalon, resulting in severe cerebral sparganosis [8, 9], which can be manifested as headache, seizure, limb paralysis, aphasia, cognitive disorder, and other focal neurological deficits [10, 11]. Characteristic magnetic resonance imaging (MRI) findings such as aggregated ring-like enhancement, the tunnel sign, and wandering lesions are very useful in the diagnosis of cerebral sparganosis [12].
Although surgical removal of the parasite has been considered standard therapy [13], failure of this treatment in some cases has also been reported [14, 15]. Surgical lesion removal has been considered the first-line therapy for cerebral sparganosis just because treatments with anthelmintics, including praziquantel, have been described as ineffective [16, 17]. Since praziquantel was first introduced as a broadspectrum antiparasitic drug in 1975, it has been proven to be a successful treatment for the majority of human infections by trematodes and cestodes including schistosomiasis, clonorchiasis, paragonimiasis, taeniasis, and cysticercosis [18]. However, the efficacy of praziquantel therapy for cerebral sparganosis remains controversial. Some patients who received conventional-dose praziquantel (25mg/kg/day for 3 days) alone experienced a worsening of their clinical condition [19].
In 2012, our group reported that three patients with cerebral sparganosis showed good therapeutic responsiveness to high-dose praziquantel (50 mg/kg/day for 10 days) [9]. In 2013, Roman et al. also found that high-dose praziquantel treatment (75 mg/kg/day for 7 days) was efficacious in a patient with inoperable cerebral sparganosis [20]. Cerebral sparganosis is a severe and disabling disease, but public health strategies have not prioritized its prevention and treatment. Therefore, investigation of the efficacy and safety of high-dose praziquantel treatment for cerebral sparganosis is important because the drug treatment may be convenient and cost-effective. In this study, we compared the efficacy and safety of long-term, high-dose praziquantel with surgical removal in the treatment of cerebral sparganosis. Our results showed that most patients with praziquantel therapy achieved favorable outcomes during follow-up.
All patient data were anonymized in the manuscript and database. Because of the uncertainty of the efficacy and safety of long-term, high-dose praziquantel treatment for cerebral sparganosis, the routine clinical procedure was evaluated and approved by the institutional ethics review board of the First Affiliated Hospital of Nanchang University (approval number: 2013–06).
This was a multicenter retrospective study using data collected by physicians monitoring patients with cerebral sparganosis in three academic medical centers: the First Affiliated Hospital of Nanchang University, Peking University People’s Hospital, and Jiangxi Provincial Institution of Parasitic Diseases. The primary objective of this research was to evaluate the efficacy and safety of long-term, high-dose praziquantel treatment (50 mg/kg/day for 10 days, repeated at monthly intervals) for cerebral sparganosis compared with surgical therapy. Because of the successful experience in our clinical centers before 2012 [9], we established a routine clinical procedure for cerebral sparganosis under the supervision of Drs. Hong, Xie, and Wan in the three medical centers at the end of 2012. Therefore, the clinical charts and imaging data were relatively complete in the retrospective review, but a quantitative assessment of the pretreatment disease severity was absent.
All consecutive patients with cerebral sparganosis from January 2013 to December 2017 were retrospectively retrieved from the database. The patient list was compiled by searching the electronic medical records using the International Classification of Diseases Tenth Revision (ICD10) discharge code B70.151. The inclusion criteria included: (1) patients showed cerebral symptoms associated with at least one structural lesion; (2) patients had definite evidence of sparganum infection that was proven by immunopositivity to Spirometra mansoni antibody in both serum and cerebrospinal fluid (CSF) tests and/or pathological evidences; (3) patients underwent follow-up cerebral MRI and serological immunological tests in our centers. The exclusion criteria included: (1) patients lost to follow-up; (2) patients with severe cardio-pulmonary dysfunction, resulting in contra-indication for surgery; (3) patients with severe liver and/or renal dysfunction, resulting in contra-indication for surgery or praziquantel treatment; (4) patients with surgical lesion removal, who had initially received praziquantel; (5) patients with praziquantel formula treatment after surgical lesion removal.
The patients were classified into a surgical therapy group and long-term, high-dose praziquantel therapy group according to the following principles. All patients initially received a detailed clinical assessment, followed by medical education about cerebral sparganosis in order to enable an informed decision regarding the choice of either praziquantel treatment or surgical treatment themselves. The physicians did not make a decision regarding treatment options for them, except patients with multiple lesions who would directly receive praziquantel treatment. The final treatment plan was chosen according to the willingness of the patients or their legal guardians. The disease severity and high-risk lesions located at some important brain regions might have been overemphasized in some communications, but the proportion could not be determined in the retrospective study. However, high-risk lesions were not considered an operative contra-indication because our neurosurgeons had excellent technical skills and experience for the operation. The main reasons for opting for long-term, high-dose praziquantel treatment were refusal to undergo surgery, lesions located at important functional areas of the brain, and initial tentative therapy (i.e., some patients wanted to try two cycles of praziquantel treatment, and subsequently underwent surgical lesion removal if a positive therapeutic outcome was not obtained).
The clinical variables evaluated in this study were age, gender, epidemiological history, headache, seizure, hemiparesis, and aphasia. The epidemiological history was judged by whether the patients had been infected with Spirometra mansoni by drinking water contaminated with procercoid-infected copepods, eating undercooked meat of snakes or frogs infected with Spirometra mansoni, or applying the flesh or skin of an infected frog or snake to poultice open wounds. The laboratory variables were the cerebral MRI characteristics (aggregated ring-like enhancement, the tunnel sign, lesion migration, high-risk lesions, and multiple lesions), blood eosinophil percentage, and serological and CSF levels of antibodies for a panel of parasitic infections including spirometra mansoni, schistosoma japonicum, cysticercosis, paragonimiasis, clonorchiasis, toxoplasmosis, and echinococcosis. Aggregated ring-like enhancement refers to conglomerate ring-shaped enhancing lesions on MRI, usually three to six bead-shaped rings (S1A Fig). The tunnel sign is about 4 cm in length (usually 2–6 cm) and 0.8 cm in width (usually 0.5–1.5 cm), and exhibits marked enhancement on coronal and sagittal contrast MRI (S1B Fig). Lesion migration indicates the presence of new and old lesions in different cerebral locations due to the migration of larva (S1C and S1D Fig). Multiple lesions are defined as the existence of at least two active lesions located at different encephalic regions. A high-risk lesion is a lesion located at an important functional area of the brain, including the brain stem, thalamus, and precentral gyrus. The Spirometra mansoni IgG antibody titer was expressed as the optical density value on microplate enzyme-linked immunosorbent assay. The cut-off value of the optical density was 0.30 as determined by normal human serum in our laboratory.
Surgical removal of lesions included two types of clinical procedures: craniotomy and CT-guided stereotactic aspiration. After the surgery, the patients were routinely administered with low-dose praziquantel (50 mg/kg/day for 4 days) to prevent possible residual infection. In the praziquantel treatment group, the patients were initially treated with 50 mg/kg/day in three divided doses for 10 days, and the treatment cycle was then repeated at monthly intervals, until the active lesions had completely disappeared on MRI. The MRI scan was performed at the beginning of the next treatment cycle. The maximum number of repetitive cycles of praziquantel administration could not exceed eight cycles. If more than eight were needed, the patient would be treated by surgery. If a patient developed an allergic or hypersensitive reaction during therapy course, 5 mg/day of dexamethasone was intravenously administered for no more than five days.
The clinical manifestations of cerebral sparganosis were closely related to the site of the lesions, and the clinical prognosis was significantly associated with the recovery of granulomatous lesions. Active ring-like or tunnel-like enhancements represented the direction of active inflammatory tunnels in different dimensions. Therefore, the primary efficacy endpoint was defined as the disappearance of active lesions on contrast MRI. All patients underwent an MRI scan before the treatment. The surgical patients underwent a follow-up MRI at one month postoperatively. The patients treated by praziquantel underwent an MRI scan at the beginning of each treatment cycle. If the active lesions disappeared, the praziquantel treatment was discontinued. If a patient developed neurological symptoms at any time after completion of the treatment regimen, cerebral MRI was performed immediately. Development of a new lesion after the active lesions had disappeared was defined as treatment failure.
The secondary efficacy endpoints were the clinical outcome assessed by the modified Rankin scale (mRS) score 90 days after the end of the treatment, the incidence of seizure, the eosinophil count, and the serological Spirometra mansoni antibody titer. The mRS is a 7-point scale ranging from 0 (no symptoms) to 6 (death). A score of ≤2 indicates functional independence. A seizure was defined as a clinical event occurring after the end of treatment regardless of whether seizures existed before treatment. The eosinophil count was determined at the end of treatment, and the number of patients with an eosinophil percentage of >5% was counted. Serological titers were only compared between the beginning and the end of treatment.
The clinical symptoms assessed in this study were vital signs, headache, dizziness, sleepiness, abdominal pain, diarrhea, and any reported adverse events. The laboratory indices were hematology parameters, alanine aminotransferase, aspartate aminotransferase, blood creatinine, blood urea nitrogen, urine parameters, and electrocardiography. An allergy to praziquantel treatment for cerebral sparganosis usually manifested as fever, chills, pruritus, and urticaria occurred in the first or second treatment cycle. Adverse events were defined as clinical symptoms with onset or worsening severity at or after the first dose of praziquantel until the end of the safety follow-up (day 30).
All statistical analyses were performed using the Statistical Package for the Social Sciences 17.0 software (SPSS, Inc., Chicago, IL, USA), and a p-value of <0.05 was considered statistically significant. Categorical variables were presented as count (percentage). Continuous variables were reported as mean ± standard deviation. The statistical significance of intergroup differences was assessed by pooled-variance and separate-variance Student’s t-test, chi-squared test, or Fisher’s exact test as appropriate. To adjust for the confounders in this retrospective study, several binary logistic regression models were established in a sub-analysis to identify the outcomes.
The patient disposition and analysis were depicted in Fig 1. In total, 108 patients were screened by the discharge code, among whom 99 patients met the inclusion criteria and nine were excluded due to loss to follow-up. Initially, 42 patients underwent surgical lesion removal, and 57 patients received long-term, high-dose praziquantel. Three patients who initially received tentative praziquantel treatment underwent surgery when no therapeutic effects were observed after two treatment cycles. Therefore, the surgical group comprised 42 patients and the praziquantel group comprised 54 patients. There were no significant differences in demographics, epidemiological history, clinical manifestations, or relevant laboratory data between the two groups (Table 1 and S1 Table).
All 42 patients in the surgery group initially underwent surgical lesion removal (24 by craniotomy and 18 by CT-guided stereotactic aspiration) and achieved complete lesion recovery one month after surgery. However, three patients developed new lesions during the postoperative follow-up: one had a new lesion in the second month after stereotactic aspiration, one reoccurred in the third month after craniotomy, and one reoccurred in the fifth month after stereotactic aspiration. The treatment effects between craniotomy and CT-guided stereotactic aspiration were not significantly different (95.8% vs 88.9%; absolute difference, 6.9%; 95% confidence interval [-23.9%, 36.9%]; p = 0.567; Fisher’s exact test).
Fifty-one patients who underwent long-term, high-dose praziquantel treatment achieved complete lesion recovery at the end of treatment. Three patients still had active lesions after eight cycles of praziquantel treatment, and then underwent surgical removal. Among the 51 patients, three had recurrent lesions at the third, fourth, and sixth month after recovery of the active lesions, respectively. Among the patients who achieved successful praziquantel treatment, two underwent one treatment cycle; five underwent two treatment cycles; 19 underwent three treatment cycles; 13 underwent four treatment cycles (Fig 2); and nine required five to eight treatment cycles (S2 Table and S2 Fig). The long-term, high-dose praziquantel treatment for cerebral sparganosis showed an efficacy similar to that of surgical lesion removal with respect to the primary efficacy endpoint (88.9% vs. 92.9%; p = 0.727) (Table 2). Even when the cut-off of the number of treatment cycles was set at five, the primary efficacy endpoint still showed no significant difference between the praziquantel and surgical groups (81.5% vs 92.9%; p = 0.106). No patients in either group had a mRS score of more than or equal to 5. The number of patients with a mRS score of ≤2 was similar between the praziquantel and surgical groups. The incidence of seizures after the end of treatment was similar between the praziquantel and surgical groups. The numbers of patients with an eosinophil count of >5% and the serological Spirometra mansoni antibody titer were similar between the two groups at the end of treatment (Table 2).
Several logistic regression models were configured to further evaluate the confounders that may have introduced bias into this retrospective study. After adjustment for age, sex, multiple lesions, and high-risk lesions, the long-term, high-dose praziquantel treatment showed an efficacy similar to that of surgical lesion removal (Table 3 and S4 Table). Because the lesion locations was considered to be a possible major confounder causing selection bias, the patients with multiple lesions and high-risk lesions were excluded from a sub-analysis, which showed that the long-term, high-dose praziquantel treatment still had efficacy similar to that of surgical removal (Table 3). Of the nine patients who were lost to follow up and thus initially excluded from the study, six had not received any treatments after diagnosis, but the other three had initially received praziquantel for two to three cycles. Even when these three patients plus the three retreated patients were counted as a negative primary outcome in the praziquantel group, the sub-analysis adjusted for age, sex, multiple lesions, and high-risk lesions showed no significant difference between the praziquantel and surgical groups (Table 3). Considering that allergic reactions were only observed in the praziquantel group, the patients with allergic reactions were excluded from the sub-analysis, which still showed that the praziquantel treatment had an efficacy similar to that of surgical removal (Table 3).
Six patients developed allergic reactions in the praziquantel group, but no patients developed allergic reactions in the surgical group. Although a difference was observed in the number of patients with allergic reactions between the two groups (Table 4), the clinical course of the allergic reactions was relatively benign and rapidly resolved after the administration of 5mg/day of dexamethasone. Headache, dizziness, abdominal pain, diarrhea, and sleepiness were also reported by the patients. The incidence of these symptoms was similar between the praziquantel and surgical groups (Table 4). No differences in vital signs or electrocardiography findings were identified between the two groups. Increases in the aspartate aminotransferase (AST) and alanine aminotransferase (ALT) indicated a possibility of liver dysfunction, which occurred in a small proportion of the patients in the first or second treatment cycle. The dysfunction was characterized by two- to four- fold elevation of the AST/ALT without jaundice and was resolved by symptomatic treatment. Overall, no clinically meaningful differences in these laboratory abnormalities were observed between the two groups (Table 4). No patients withdrew from the study because of adverse events. Binary logistic regression models adjusted for age, sex, multiple lesions, and high-risk lesion further showed that these safety variables were not different between the two groups (S5 Table).
Cerebral sparganosis is a severe disease when the plerocercoid larva of Spirometra tapeworm targets the central nervous system [21]. The larva damages brain tissue and gives rise to neurological function deficits caused by inflammatory attacks and the migration process [22, 23]. The natural lifetime of the larva is several decades long in some cases [24]. Therefore, eradication of the larva is the main therapeutic strategy [25]. Surgical removal is considered the first-line treatment for cerebral sparganosis because treatment with anthelmintics, including praziquantel, has been described as ineffective [16, 17]. However, both craniotomy and CT-guided stereotactic aspiration may result in incomplete removal, especially when remnants of the scolex segment are left behind after CT-guided stereotactic aspiration, resulting in larva regeneration [16]. Additionally larva lesions may localize at some important functional structures, in which cases the surgical procedure will cause severe neurological dysfunction. Thus, the surgical removal of larva has some limitations.
In this study, long-term, high-dose praziquantel resolved the cerebral lesions in 88.9% of patients with cerebral sparganosis. Most patients with cerebral sparganosis achieved effective treatment by praziquantel at 50 mg/kg/day for 10 days at the third or fourth treatment cycle. The absolute efficacy rate of praziquantel was relatively lower than that of surgical removal, but there were no significant differences in the primary efficacy endpoint (no active lesions on MRI) and secondary efficacy endpoints (mRS score at 90 days, seizure, eosinophil count, and serological titer) between the praziquantel and surgical groups. Quantitative assessments of disease severity before therapy were unavailable in this retrospective study. Naturally, it was possible that patients with milder severity whose clinical course showed little progression and who may not need invasive surgery might select praziquantel treatment. In addition, patients with complicated clinical course or multiple or high-risk lesions might not select surgical removal. As a result, although there were no significant differences in the available baseline data between the praziquantel and surgical groups, some confounders could have biased the primary outcome. To compensate for this bias, several logistic regression models were established to adjust for age, sex, lesion location, and loss to follow-up. These regression models showed that praziquantel still had an efficacy similar to that of surgery, but some of these results actually showed borderline significance and indicated a tendency toward a surgical benefit.
Praziquantel can significantly damage and destroy the whole body of the plerocercoid except the scolex and neck. The spargana may regenerate from the remnants of the scolex and neck, which might be responsible for resistance to praziquantel [26]. Therefore, the therapeutic efficacy of praziquantel for human sparganosis remains controversial. In one study, a patient with subcutaneous sparganosis failed to recover after being treated with two praziquantel regimens of 3×25mg/kg×5 days at one month interval [27]. In contrast, patients with pleural or pericardial sparganosis were successfully cured with a regimen of 60 to 75 mg/kg/day of praziquantel for three days [27,28]. For sparganosis in the central nervous system, praziquantel treatment might be more difficult because only 1/7 to 1/5 of the drug within the plasma can diffuse into the tissue near the larva lesions [29]. Previous studies also supported the finding that praziquantel treatment dose not seem to have a killing effect on live worms [19]. In a recent study, however, the larva in a patient with cerebral sparganosis was successfully eradicated by high-dose praziquantel with a regimen of 3×25 mg/kg/day for seven days combined with cimetidine (3×400mg daily) and a high-carbohydrate diet [20]. As early as 2012, we modified our praziquantel treatment formula to 50 mg/kg/day in three divided doses for 10 days, and then repeated this cycle at monthly intervals until the active lesions completely disappeared on enhanced MRI. Our clinical data showed that most patients with cerebral sparganosis were effectively treated with three or four treatment cycles. The host’s immune cells (e.g., granulocytes, histiocytes) and antibodies might act synergistically with praziquantel in the treatment of tissue-invading sparganosis infections [30]. The repeated administration of praziquantel might increase the probability of antigen exposure on the surface of the regenerative worm, which might produce more immunoreactivity to attack the regenerative larva [31]. However, the detailed pharmacological mechanism needs to be investigated in further studies.
It is generally acknowledged that praziquantel is highly safe without serious adverse reactions. Common adverse reactions to praziquantel are usually mild and include abdominal pain, diarrhea, dizziness, sleepiness, and headache [32]. These clinical complications were not significantly different between the praziquantel and surgical groups of the present study. In rare situations, however, praziquantel might cause allergic or hypersensitive reactions in some patients [33]. In the present study, these allergic reactions usually occurred in the first or second treatment cycle with a benign clinical course, and were sensitive to short-term, low-dose dexamethasone. Although some patients showed a tendency toward mild liver dysfunction in the first two praziquantel treatment cycles, the laboratory indices showed no significant differences between the two groups. Importantly, follow-up of the patients treated with long-term, high-dose praziquantel revealed no delayed adverse events.
This study had some limitations that need to be explicitly acknowledged. First, it was a small-sample, nonrandomized retrospective study; thus, the statistical power was low. Cases of cerebral sparganosis are rare, so it is necessary to include data from more medical centers. Second, the fact that the patients made the treatment decision themselves undermined the reliability of the results. Confounders such as the disease severity, multiple lesions, high-risk lesions, and doctor-patient communication skills might have produced some bias with respect to the treatment decision. Third, the clinical data were quite heterogeneous because the treatment with praziquantel was different in each patient based on the lesion site and response. It was impracticable to assess the outcome only once and at a fixed time to avoid biasing the results because of ethical considerations and the pharmacological properties of praziquantel. Overall, the outcomes of this study should be cautiously interpreted that long-term, high-dose praziquantel had an efficacy similar to that of surgical removal, though the logistic regression models supported these results. Indeed, several baseline outcomes and safety indices showed borderline significance with respect to the benefit of surgical removal of larva lesion. However, the findings in this exploratory study involving real-world practitioners with a high level of experience add treatment flexibility for this serious infection, and provide a basis to promote a large-sample, randomized, prospective study of long-term, high-dose praziquantel treatment for cerebral sparganosis in the future.
|
10.1371/journal.pbio.0060141 | Two-Dimensional Patterning by a Trapping/Depletion Mechanism: The Role of TTG1 and GL3 in Arabidopsis Trichome Formation | Trichome patterning in Arabidopsis serves as a model system to study how single cells are selected within a field of initially equivalent cells. Current models explain this pattern by an activator–inhibitor feedback loop. Here, we report that also a newly discovered mechanism is involved by which patterning is governed by the removal of the trichome-promoting factor TRANSPARENT TESTA GLABRA1 (TTG1) from non-trichome cells. We demonstrate by clonal analysis and misexpression studies that Arabidopsis TTG1 can act non-cell-autonomously and by microinjection experiments that TTG1 protein moves between cells. While TTG1 is expressed ubiquitously, TTG1–YFP protein accumulates in trichomes and is depleted in the surrounding cells. TTG1–YFP depletion depends on GLABRA3 (GL3), suggesting that the depletion is governed by a trapping mechanism. To study the potential of the observed trapping/depletion mechanism, we formulated a mathematical model enabling us to evaluate the relevance of each parameter and to identify parameters explaining the paradoxical genetic finding that strong ttg1 alleles are glabrous, while weak alleles exhibit trichome clusters.
| Trichomes, the specialized hair cells found on plant leaves, represent a model system to study how cellular interactions coordinate the development and arrangement of a collection of initially equivalent cells into regularly placed specialized cells. It was assumed that a regulatory feedback loop of positively and negatively acting factors governs these decisions. In this work, we show that trichome spacing also is controlled by the local depletion of the trichome-promoting protein TTG1. We provide evidence that binding of TTG1 to a second trichome-promoting protein, GL3, causes a depletion of TTG1 in the neighborhood of cells with elevated GL3 levels. We postulate that this leads to trichome fate determination in cells containing high GL3/TTG1 levels and prevents trichome formation in surrounding cells because of the reduced TTG1 levels. We show by theoretical modeling that this mechanism alone is capable of creating a spacing pattern and has properties that can explain even apparently paradoxical genetic observations.
| During the development of animals and plants, specific cell types need to be placed in a regular pattern within a field of cells. In the simplest scenario, this occurs in a two-dimensional sheet of cells. Mathematical modeling of such a spacing pattern has uncovered two general principles. Both rely on the assumption that the factor promoting the formation of the specific cell type is autocatalytic. In the “activator–inhibitor” mechanism autoactivation is counteracted by the production of an inhibitor. In contrast, in the “substrate-depletion” mechanism, a substrate is consumed by the autocatalysis of the cell type promoting factor. A common requirement of both principles is significantly reduced mobility of the autocatalytic species compared to that of the inhibitor and the substrate, respectively [1].
The activator–inhibitor system is thought to generate the regular spacing pattern of leaf trichomes in Arabidopsis [2–4]. Trichomes are regularly distributed on the leaf surface without any reference to morphological landmarks, and clonal analysis indicated that cell lineage is not involved [5,6]. Therefore, trichomes are an ideal model system to study how single cells become regularly spaced within a sheet of equivalent cells.
Current models assume that the R2R3 MYB transcription factors GLABRA1 (GL1) and MYB23 [7–9], the bHLH factors GLABRA3 (GL3) and ENHANCER OF GLABRA3 (EGL3) [10–12], and the WD40 repeat protein Transparent Testa Glabra1 (TTG1) [13,14] form a trichome-promoting trimeric complex due to the binding of one R2R3 MYB factor and TTG1 to a bHLH factor. Formally, this complex acts as the activator described in the theoretical models [1]. The activity of this complex is thought to be counteracted by the single R3 repeat MYB-like transcription factors TRIPTYCHON (TRY) [15], CAPRICE (CPC) [16], ENHANCER OF TRY and CPC1 (ETC1) [17], ETC2 [18], TRICHOMELESS1 [19], and CAPRICE LIKE MYB3 (CPL3) [20] through competition for binding of the R2R3 MYB factors to the bHLH protein [21]. The single R3 repeat MYB proteins are collectively considered to represent the inhibitor in the theoretical models. The active complex (AC) is postulated to activate the inhibitors, which can move into neighboring cells, where they repress the activators. This type of model is generally consistent with most data though several aspects have not been confirmed experimentally [3,4,6,22,23].
The role of TTG1 in trichome patterning is obscure, as the glabrous phenotype of strong alleles suggests that it promotes trichome development, whereas the formation of trichome clusters in weak alleles suggests that it is involved in the inhibition of trichomes [5,24]. This dual function of TTG1 suggested to us that TTG1 has a central function in the patterning process. In this work, we identified TTG1 as the key component of a newly discovered depletion mechanism, likely to act in parallel to the above-described activator–inhibitor mechanism. We demonstrate that TTG1–YFP depletion depends on GL3, suggesting an underlying trapping mechanism, such that GL3 captures TTG1 in trichomes. Finally, we provide a mathematical model to evaluate the properties of this new GL3/TTG1 trapping/depletion mechanism.
TTG1 is expressed in most tissues of the plant [14,25]. To determine the TTG1 expression in young leaf parts, where trichome initiation takes place, we created transgenic plants, in which the β-glucoronidase (GUS) reporter gene was driven by a 2.2 kb promoter fragment including the 5′ UTR of TTG1 (pTTG1:GUS). This fragment is sufficient to rescue completely the ttg1–13 null-mutant phenotype when driving the TTG1 cDNA (Table 1). pTTG1:GUS is ubiquitously expressed in young leaves with slightly elevated levels in incipient trichomes, and expression ceases in more mature leaf parts (Figure 1A and 1B). To determine the localization of TTG1 protein, we created a C-terminal fusion of TTG1 with yellow fluorescent protein (YFP) and an N-terminal fusion with green fluorescent proten (GFP), which both rescued all aspects of the ttg1–13 mutant phenotype, including the seed coat mucilage, transparent testa, and trichome number when expressed under the TTG1 promoter (unpublished data; Table 1 and Figure 2A–E). We further substantiated the functionality of this rescue construct by demonstrating that protein–protein interactions of TTG1–YFP with GL3 are indistinguishable from TTG1 in yeast two-hybrid interaction assays (unpublished data). Both fusion proteins were found in the nucleus and in the cytoplasm (Figure 2F). The integrity of the TTG1–YFP fusion protein was confirmed by western blot analysis (Figure 2G).
The distribution of the pTTG1:TTG1–YFP fusion protein differed strikingly from pTTG1:GUS expression. Initially, in very young leaf regions, in which trichomes are not yet initiated, TTG1–YFP is detected in all cells reflecting the gene expression pattern (Figure 1C). In slightly older leaf regions, TTG1–YFP accumulates in incipient trichomes (Figure 1C, 1D, and 1E). In the cells adjacent to young trichomes, TTG1–YFP levels are the lowest, and fluorescence gradually increases with the distance from the trichome (Figure 1D and 1E). This initial observation was confirmed by quantifying the fluorescence intensity, using the Leica Confocal software (Figure 1F). On average, cells next to a trichome showed 39% of the fluorescence of that in the trichome, the cells in the second tier around a trichome 76%, and cells of the third tier 93% (n = 31). As a control, we measured the distribution of fluorescence of a nuclear-localized GFP under the control of the TTG1 promoter (pTTG1:GFP-NLS, Figure 1G and 1H). The fusion to the nuclear localization signal (NLS) reduces or completely prohibits the movement of proteins [26–28], and therefore the distribution of GFP–NLS should reflect the expression pattern of the TTG1 promoter in this assay system. Consistent with the pTTG1:GUS lines, TTG1 expression is elevated in trichome initials and ubiquitously distributed in the surrounding cells (first tier 74%, second tier 76%, third tier 77%, n = 30). Depletion next to the trichome cell was not found, demonstrating that the relative distribution of TTG1–YFP differs significantly from its expression pattern. Using the Mann–Whitney U test, the strong fluorescence reduction in the first tier is highly significant (p < 0.0001).
The difference between the homogeneous TTG1 reporter expression and the non-homogeneous protein distribution could be explained in principle by two mechanisms. First, the protein stability could be controlled spatially, such that TTG1 is more stable in trichomes than in the neighboring cells. Second, the uneven distribution could result from TTG1 movement from neighboring cells into trichomes.
To determine whether TTG1–YFP depletion around trichomes is regulated by protein degradation, we treated whole pTTG1:TTG1–YFP plants with epoxomicin, a specific and irreversible inhibitor of the proteasomal degradation machinery [29]. The TTG1–YFP protein depletion around trichomes was not affected by epoxomicin treatments, suggesting that uneven distribution of the TTG1–YFP is not caused by a difference in TTG1 stability in trichome initials and its adjacent cells (Figure 3E–G). As a control to show that TTG1 is an actual target of the 26S proteasome and that the proteasomal inhibitor was active, we used cotyledons of the same plants analyzed for the depletion of TTG1–YFP around trichomes on rosette leaves. TTG1 is expressed in cotyledons ([25], our own observation); however, TTG1–YFP protein is not detectable in cotyledons of untreated plants or control plants (Figure 3A and 3C). In plants treated with 20 μM epoxomicin for 24 h, TTG1–YFP protein could be detected in cotyledons, showing that the epoxomicin treatment was effective (Figure 3B and 3D). Control plants treated with the solvent DMSO showed no YFP-specific fluorescence in cotyledons (Figure 3C).
The concept that TTG1 moves from neighboring cells into trichomes was proved by the following series of experiments. First, we demonstrated movement of the TTG1–YFP fusion protein from non-trichome cells into trichome cells, using the #232 activation tag line from the Poethig collection (http://enhancertraps.bio.upenn.edu/default.html, line #232). This line was identified as a line, driving the expression of the GAL4/VP16 activator, triggering expression of a UAS promoter driven mGFP5-ER, a GFP form localized to the endoplasmatic reticulum (ER) as a cell-autonomous marker. GFP-ER was expressed in an apparently random pattern but never in trichomes at any stage of development (Figure 4A–C). In contrast, the TTG1–YFP fusion under the control of the UAS promoter in this enhancer trap line showed additional YFP-specific fluorescence in initiating trichomes next to epidermal cells expressing the GAL4/VP16 activator (Figure 4A–C). This suggests that the TTG1–YFP fusion moved from the trichome neighboring cells, where it was expressed, into the trichome.
Second, we asked whether TTG1 exerts its function in a non-autonomous manner. We used the Cre-LoxP recombination system to create ttg1 mutant sectors in plants, where wild-type-expressing cells were marked by GUS expression [30]. This was achieved by cloning the TTG1 and the GUS genes, each under the control of the CaMV 35S promoter, between the two LoxP recombination sites and by introducing this construct into ttg1–13 mutants, containing the Cre recombinase under the control of a heat-shock inducible promoter (Figure 4D). These plants showed a wild-type trichome pattern due to the rescue of ttg1 by 35S:TTG1 (Table 1) and ubiquitous expression of GUS. Heat shocks were applied when the first two leaves emerged. After a saturating heat treatment of 1–2 h, no GUS staining and no trichomes were detected on leaves three and four (unpublished data). Heat-shock conditions (5–15 min) were chosen such that a recombination event excising the 35:TTG1/35:GUS occurred rarely. These cells subsequently developed into large clonal sectors on leaves number three and four. As shown in Figure 4E and 4F, GUS-negative and therefore ttg1 mutant sectors were found that clearly exhibited trichomes. This shows that TTG1 can rescue the ttg1 mutant in a non-cell-autonomous manner.
Third, we analyzed whether TTG1 protein can actively move between cells. It has been shown that soluble GFP, 2 × GFP, and 3 × GFP (27, 54, and 81 kDa, respectively) move passively between cells with higher capacity at early stages and restricted mobility later in development [31,32]. Therefore, the size of a protein is not the main criterion for its ability to move between cells. Transport of molecules between plant cells is mainly regulated through plasmodesmata (PDs), plant-specific channels that span the cell wall and connect plant cells with each other. In recent years, several proteins have been shown to move between cells, most likely by using the PD pathway [33,34]. Hence, the potential of TTG1 to act non-cell-autonomously and to move between cells raises the question whether the 38 kDa TTG1 protein moves by actively opening the PDs. To test this general biological property of TTG1, we used microinjections in tobacco mesophyll cells (Figure 5 and Table 2). This system can be used to monitor changes in the symplasmic connectivity after injection of proteins [35]. Each set of experiments on a given leaf includes four steps. First, the injection of the small fluorescent tracer molecule acridine orange and lucifer yellow confirmed that the leaf tissue was healthy and that cells were symplasmically connected (Figure 5A). Second, 11-kDa rhodamine–dextran or 12-kDa F-dextran were injected to show that molecules larger than the plasmodesmatal size exclusion limit (SEL) for this tissue do not move into the neighboring cells (Figure 5B and Table 2) [36].
Third, the coinjection of the normally cell-autonomous 12-kDa F-dextran and TTG1 protein was done to test whether TTG1 can increase the SEL for this tracer. As shown in Figure 5C and 5D, the F-dextran moved out of the injected cell into neighboring cells in these coinjection experiments, suggesting that TTG1 increases the SEL. Fourth, to test directly whether the 38-kDa TTG1 protein can move, it was labeled with fluorescein isothiocyanate (FITC) or rhodamine. After injection, the fluorescent signal emitted by labeled TTG1 protein appeared within minutes in adjacent cells (Figure 5E and 5F and Table 2). GST–rhodamine and NtMPB2C–FITC were used as negative controls in these experiments [37,38]. Both proteins did not move and did not trigger movement of the tracer, indicating that the injection procedure as such did not change the movement behavior of the tracer or proteins in general. Thus, recombinant TTG1 protein shows an equivalent behavior in microinjection assays as the non-cell-autonomous KN1 protein [36]. These data indicate that TTG1 similar to KN1 increases the plasmodesmatal SEL and moves actively to neighboring cells via the intercellular transport pathway established by PDs.
Finally, we tested the movement ability of TTG1 between cell layers. For subepidermal expression studies, we used the mesophyll-specific phosphoenolpyruvate carboxylase promoter from Flaveria trinervia (ppcA1) [39]. To corroborate the specificity of the promoter in Arabidopsis, we used it to express a GFP–YFP fusion, which does not move between leaf tissue layers in Arabidopsis [40]. The GFP–YFP signal was exclusively detected in subepidermal tissue from early primordia stages on (Figure 6A and 6B). In contrast, lines expressing TTG1–YFP under the ppcA1 promoter showed additional fluorescence in the epidermal layer, showing that TTG1–YFP moved from mesophyll to epidermal tissue (Figure 6E and 6F). Consistent with this, cDNA expressed under the ppcA1 promoter rescued the ttg1 mutant trichome phenotype equally well as under the endogenous TTG1 promoter. Also the TTG1–YFP fusion rescued the ttg1 mutant phenotype, though less efficiently (Table 1). In young leaves, the TTG–YFP signal was found in all epidermal cells (Figure 6E), whereas in older leaves it was found only in trichomes (Figure 6F). This finding is consistent with the earlier observation that TTG1 is expressed only in subepidermal tissues during embryo development but is required in the protodermal tissue (the embryonic epidermis) [41]. To test whether trichomes can generally attract proteins or whether this is a specific property of TTG1, we also expressed YFP under the control of the ppcA1 promoter (Figure 6C and 6D). The YFP protein was observed in all cell layers in young tissues (Figure 6C). However, YFP did not accumulate in trichomes (Figure 6D). These data indicate that trichome-specific localization is a property of the TTG1 protein rather than due to trichome characteristics, such as a larger SEL of PDs or generally higher import rates of molecules.
To understand the mechanism leading to the depletion, we tested the hypothesis that TTG1–YFP might be trapped by GL3 in trichomes. This seemed reasonable because GL3 expression is increased in trichomes relative to the surrounding cells and because GL3 strongly binds to TTG1 in yeast two-hybrid assays [12]. If the hypothesis is correct, then one would expect that TTG1–YFP would not show depletion in gl3 mutants. As shown in Figure 1I and 1J, TTG1–YFP is ubiquitously distributed in the epidermis in plants lacking functional GL3. The quantification revealed elevated fluorescence in trichome initials and ubiquitously similar levels in the surrounding cells (first tier 79%, second tier 77%, third tier 79%, n = 40). These data strongly suggest that TTG1–YFP is depleted through trapping in trichome cells by GL3.
We used mathematical modeling to evaluate the properties of a patterning mechanism solely based on GL3/TTG1 depletion. Therefore, we neglected the influence of additional inhibitors on the patterning mechanism. The model is based on the following assumptions: (i) TTG1 is constantly and ubiquitously expressed (shown in this work). (ii) TTG1 moves nondirectionally between cells. Although we show that TTG1 can actively open the PDs, there is no evidence for regulated transport affecting the actual rates. (iii) TTG1 forms a dimer with the GL3 protein as indicated by yeast two-hybrid results [12]. (iv) The AC enhances the expression of GL3 cooperatively. This is assumed because nonlinearity of the positive feedback is absolutely necessary for pattern formation. The data toward this end are not clear. At the whole plant level, it appears that GL3 is involved in a negative feedback loop [42]; however, at the current experimental resolution, these data do not contradict our assumption. Moreover, the GL3 homolog TT8 was shown to act in an autoactivation [43]. (v) GL3 and the AC are cell-autonomous. This assumption is based on the observation that GL3 protein does not move in the leaf (unpublished data). (vi) All components are degraded by first-order kinetics. The corresponding interaction scheme is shown in Figure 7A.
Because the model parameters are unknown, we employed a two-step approach. First, a rescaling of model variables allowed the confinement of the parameters to relevant ranges. Second, we fitted the resulting model to the experimentally obtained relative fluorescence intensities of TTG1 in the vicinity of the trichomes. Fitting of the parameters also took into consideration the mean trichome density in the initiation zone. For parameter values and details of the optimization, see the Materials and Methods section.
A typical simulated concentration pattern of total TTG1 (i.e., TTG1 + AC) is presented in Figure 7B. The highest TTG1 levels are found in the trichomes where it is completely bound to GL3. In cells adjacent to trichomes, the level of unbound TTG1 is significantly lowered by depletion, while the level increases with distance from the trichomes.
Our rescaling and fitting procedure enabled us to estimate the model parameters and in turn to judge their relevance. We focused on the dependence of trichome density and clustering on parameters related to TTG1 function (Figure 7C). Here, trichome density is defined as the ratio of trichome cells to the total number of epidermal cells in the initiation zone of the young leaf. A decrease of the degradation rate λ3 of the AC (cyan line, circles) or of the transport rate d of TTG1 (green line, squares) results in an elevated trichome density/clustering. Conversely, an increase in the complex formation rate β (blue line, diamonds) raises the trichome density/clustering. Surprisingly, the trichome density/clustering is unaffected by a decreased degradation rate λ1 of TTG1 (red line, triangles). The increase of trichome density is correlated with a corresponding change of the percentage of the trichomes found in clusters (Figure 7C, inset). Note that blunt ends correspond to a loss of the trichome pattern; e.g., a decreased complex formation rate leads to glabrous plants.
These data provide for the first time an explanation for the apparently paradoxical observation that strong ttg1 alleles are glabrous (suggesting a positive function) and weak ttg1 alleles show clusters (suggesting an inhibitory function). While it is trivial that the absence of TTG1 in this model causes a glabrous phenotype, surprisingly, simulations of the depletion mechanism revealed that alterations of all parameters, except for the protein degradation rate, can lead to clusters.
In this study, we focus on the functional analysis of TTG1 in trichome patterning on Arabidopsis leaves. We show that TTG1 is ubiquitously expressed with slightly higher levels in developing trichomes. The distribution of TTG1–YFP differs from the expression pattern such that the signal is strongly reduced in cells immediately next to the trichome. In showing that the proteasome inhibitor epoxomicin does not affect the protein distribution, we exclude the possibility that differential protein degradation results in the local depletion of TTG1–YFP around incipient trichomes. We demonstrate that TTG1 acts non-cell-autonomously by clonal analysis and that the TTG1–YFP protein can move within the epidermis into trichomes by using a GAL4-based expression system. Further, we show that TTG1–YFP can move between cell layers and that the TTG1 protein can open actively PDs in a heterologous system. Together these data suggest that TTG1 is redistributed from neighboring cells into the trichome by intracellular movement. What is the underlying mechanism of the observed depletion/attraction of TTG1? One possibility is that TTG1 moves freely and becomes trapped in trichomes. Alternatively, the redistribution could be achieved by directional movement into the trichomes, although both mechanisms do not necessarily rule out each other. The latter scenario is similar to that proposed for the function of auxin in the positioning of primordia in the meristematic region [44]. In this system, directional transport of auxin by the transporter PIN1 leads to an accumulation of the hormones in primordia and a reduced level of auxin in the neighborhood [44,45]. A directional transport similar to auxin is unlikely for TTG1 because TTG1–YFP can move from the cells, expressing it not only into trichomes but also into other epidermal cells (Figure 4A–C). We therefore hypothesized that TTG1 accumulates in trichomes, because it binds to another protein, as suggested for SHORT ROOT (SHR) in the root [46]. SHR is expressed in the stele and moves specifically into the endodermis, where it is required and sequestered in the nucleus due to interaction with SCARECROW [46]. In support of this hypothesis, we find no depletion of TTG1–YFP in gl3 mutants, indicating that TTG1 binding to GL3 causes the depletion.
Current models explaining trichome patterning on Arabidopsis leaves are based on the activator–inhibitor-like mechanisms described above [2–4,47]. These mechanisms can explain the generation of a pattern in the absence of pre-existing positional information. However, not all aspects of the model have been shown experimentally. The mobility of the inhibitors was shown for CPC in the root system [48], but nothing is known about the mobility in leaves. Moreover, the theoretical requirement that the activators can autoactivate lacks experimental proof. Another problem with the current models is that various genetic data cannot be explained [3].
Our finding that in addition to the activator–inhibitor mechanism a substrate-depletion-like mechanism is operating during trichome patterning may provide some missing clues. In general, a substrate-depletion mechanism is superficially similar to the activator–inhibitor mechanism. Instead of producing an inhibitor that laterally suppresses trichome development in cells next to a developing trichome, a factor necessary for trichome development is removed from these cells. When simulating this type of mechanism, however, it turned out that the system properties are different [1,49]. In particular, it was noted that new peaks are formed at the maximum distance by the activator–inhibitor mechanism and by splitting already existing ones by the substrate-depletion mechanism [1,49].
To understand the properties of the GL3/TTG1 trapping mechanism, we formulated a mathematical model and fitted it to our experimental data to obtain a biologically relevant parameter range. This strategy enabled us to test how parameter changes affect patterning. In particular, we aimed to simulate the weak ttg1 cluster phenotype as this genetic finding was the most confusing, because the lack of trichomes in strong ttg1 mutants suggested that TTG1 functions as a trichome-promoting factor and the cluster phenotype in weak ttg1 mutants pointed toward a role as a negative regulator [5,13,24,50,51]. The simulations of the GL3/TTG1 trapping mechanism revealed that changes of several parameters related to TTG1 function can result in a clustering phenotype. Thus, we can offer for the first time explanations for the apparently paradoxical genetic results on TTG1 with our new GL3/TTG1 trapping/depletion model. However, our reduced model can only partially capture the experimental observations. For example, the simulated mean trichome density as predicted by the optimal parameter set is still substantially larger than that in the wild type. We expect that more complex models involving additional patterning genes will improve the agreement between theory and experiment.
As GL3 is also a central component of all activator–inhibitor-based models, it is conceivable that the two models act in concert. We can recognize TTG1–YFP depletion at the earliest stages of morphologically recognizable trichome development. This would suggest that the trapping/depletion mechanism becomes relevant after the activator–inhibitor mechanism already has started the selection of trichomes. However, it is well possible that more sensitive microscopic techniques and more sophisticated imaging analysis tools will reveal the depletion much earlier, so we consider the relative timing of the two processes to be elusive at the moment. It will be a future challenge to combine both principles in a single model. To operate in biologically reasonable parameter ranges, it will be crucial to base such a model not only on qualitative but also on quantitative data.
In this study, the wild-type ecotypes Landsberg erecta (Ler) and RLD were used. The ttg1-1, -9, -10, and -13 and gl3-1 mutant lines have been described previously [14,24,52]. The Poethig activation tag line #232 (Columbia ecotype) was a kind gift from Scott Poethig, University of Pennsylvania (http://enhancertraps.bio.upenn.edu/default.html). The heat-shock inducible HSP:CRE3 line containing the pCGNHCN construct in a Nossen ecotype background [30] was crossed into the ttg1-13 mutant line (RLD background), and plants homozygous for both the transgene and the ttg1-13 allele were isolated and crossed to TTG1-Lox lines. The TTG1-Lox construct is a descendant from the pCGNLox2a construct [30], introducing a 35S:TTG1:NOSpA cassette into the PmeI site of pCGNLox2a. The resulting plants of these crossings were used for heat-shock treatments. Plants were grown on 1 × Murashige Skoog agar (1% sucrose) plates for approximately 10 d at 20 °C under 16 h light/8 h dark conditions. Heat shock was performed by placing the plates into an illuminated incubator at 41 °C for 10–15 min. All transgenic lines were produced using the floral dip method [53].
The TTG1 promoter (position −2227 to −1 from the start codon and includes the 110 nucleotide of the 5′ UTR) was isolated from Arabidopsis thaliana ecotype Ler by PCR (forward primer, 5′-AAAGCTTAACCGAGAATGTCTCCCGACTTCTAT-3′; reverse primer, 5′-AGTCGACTCAAACTCTAAGGAGCTGCATTTG-3′) and cloned into pGEM-T vector (Promega Corporation) (pTTG-pGEM). An AscI restriction site was added by adapter ligation (5′-CTAGAATGGCGCGCCATT-3′) into the SpeI site of the vector. To generate the pTTG:GUS construct, the pTTG-pGEM was digested with AscI and SalI, and the resulting fragment was cloned into the binary gateway vector pAM-PAT-GW-GUS (GenBank accession AY02531) to replace the existing CaMV 35S promoter between the AscI and the XhoI sites.
To create the pPPCA1-pAMPAT binary vector, the 2117 bp promoter fragment of the phosphoenolpyruvate carboxylase 1 gene (ppcA1) from Flaveria trinervia (GenBank accession X64143) [39] was removed from ppcA1-pBS 5′ with HindIII and religated using an oligonucleotide linker to generate an AscI restriction site. The resulting AscI–XhoI fragment was inserted into pAM-PAT-GW using the same restriction enzymes.
The yeast UAS promoter was PCR-amplified with the attachments of AscI for the forward primer and XhoI for the reverse primer. The corresponding fragment was ligated into pAMPAT-GW by exchanging the existing CaMV 35S promoter using AscI and XhoI, giving rise to pUAS-pAMPAT.
The TTG1 cDNA (GenBank accession AT5G24520.1) was PCR-amplified with attB1 forward and attB2 reverse linker primers for Gateway BP recombination with the pDONR201 vector (Invitrogen). To create the TTG1–YFP fusion, the TTG1 cDNA was PCR-amplified again to add a SalI site at the 5′ and a XhoI site at the 3′ of the coding sequence deleting the stop codon (forward primer, 5′-AGTCGACATGGATAATTCAGCTCCAGA-3′; reverse primer, 5′-ACTCGACAACTCTAAGGAGCTGCATTT-3′). The digested fragment was ligated into the SalI site of pUC18, then a XbaI–SacI EYFP fragment (Clontech) was fused C-terminally to TTG1 using the same sites. The fusion was isolated using XhoI and EcoRI and ligated into pEN1a SalI–EcoRI fragment. The resulting construct was called TTG1–YFPpEN. pEYFP (Clontech) was digested with SalI and NotI and ligated into pEN1a to create EYFPpEN. The GFP–YFP fusion was constructed using an NcoI fragment of mGFP4, which was ligated in frame into the NcoI site of EYFPpEN. All constructs were sequenced. To create all of the binary constructs or yeast two-hybrid vectors, the Gateway LR Reaction System was used according to the user's manual (Invitrogen).
GUS activity was assayed as described previously [54]. After adding the X-Gluc-solution (5-bromo-4-chloro-3-indolyl-β-d-glucuronic acid), plants were vacuum-infiltrated for 15 min and then incubated at 37 °C overnight. The tissue was cleared by an ethanol series (15%, 30%, 50%, and 70% EtOH solutions at 37 °C for several hours).
Seed coat mucilage staining was done with a 0.01% ruthenium red solution for 15 min.
Light microscopy was performed using a Leica DMRE microscope using differential interference contrast optics. Images were taken using a KY-F70 3-CCD JVC camera and DISKUS software (DISKUS, Technisches Büro). Confocal laser scanning microscopy was done with a Leica TCS-SP2 confocal microscope equipped with the Leica software Lite 2.05 (LCS, Leica Microsystems). Z-stacks in steps of 1 or 2 μm were taken and processed using deconvolution tools of the Leica software. Quantification of fluorescence was performed using the same software. Plants were incubated for 10–15 min with a 10 μg/ml propidium iodide solution to visualize cell walls. Transverse sections were generated by embedding the tissues in 4% low-melting-point agarose and by hand sectioning using a razor blade as described by [55]. Images were assembled and processed using GIMP 2.2 software (http://www.gimp.org).
Recombinant TTG1 protein was produced in Escherichia coli, labeled, purified, and microinjected as previously described [38,56]. The protein concentration used for microinjection was 2 μg/μl. A Leica SP2 AOBS UV confocal microscope was employed to detect the fluorescent probes after microinjection. Tissues were scanned in sequential mode to excite and detect fluorescence probes in their specific wavelengths, and the resulting Z-stack (5 μm distance) images were merged using the NIH image software ImageJ (version 1.32j) (http://rsb.info.nih.gov/ij/).
pTTG1:TTG1–YFP plants were grown on Murashige Skoog agar plates containing 1% sucrose at 22 °C for 6 d under 16 h light/8 h dark conditions and then transferred into liquid ½ MS medium containing 1% sucrose. The medium contained either 2% DMSO (control) or 20 μM epoxomicin (Sigma-Aldrich, stock solution in DMSO). The samples were vacuum-infiltrated for 15 min and incubated under the same growth conditions as previously for 24 h. After being washed with ½ MS (1% sucrose), plants were analyzed using confocal laser scanning microscopy (see above).
Yeast two-hybrid interaction assays were performed as described previously [9]. Fusions with the GAL4 activation domain and GAL4 DNA-binding domain were performed in the pACT and pAS plasmids (Clontech).
TRY, GL3, and a truncated version of GL3 lacking 96 amino acids at the N-terminus were fused to the GAL4 activation domain in the pACT vector. TTG1 and TTG1–YFP were fused to the GAL4 DNA-binding domain of pAS. None of the constructs or empty vectors showed any self-activation in yeast.
Fifteen 10 d old plants (long day conditions, 24 °C) were harvested without roots, frozen in liquid nitrogen, and afterwards ground. The powder was mixed and boiled in 300 μl of sample buffer (50 mM Tris/HCl. pH 6,8, 2% (w/v) SDS, 8 M urea, 30% (v/v) glycerol, 5% (v/v) β-mercaptoethanol, and 0.5% (w/v) bromphenol blue) for 15 min followed by centrifugation (16,000g at 4 °C) for 15 min. Approximately 25 μl of the supernatant was analyzed by 12% SDS-PAGE, which was followed either by Coomassie staining or by western blotting and subsequent immunodetection with anti-GFP monoclonal IgG mouse antiobody (Roche). Detection was done by electrochemiluminescence.
On the basis of the interaction diagram presented in Figure 7A, a system of coupled ordinary differential equations was derived that describes the temporal evolution of the protein concentrations of TTG1, GL3, and the AC inside each cell. The model was formulated on a two-dimensional grid of hexagonal cells with the cell index j = (y,x), where 1 ≤ y ≤ N and 1 ≤ x ≤ M. N and M denote the number of cells in the y and x directions, respectively. Periodic boundary conditions were chosen for model simulation and analysis.
The nondirectional transport of TTG1 between cell j and its six neighboring cells is characterized by the coupling term
The model includes parameters αi for the expression of TTG1 and GL3 and parameters λi for the degradation of the single proteins and the active complex. The parameter d is the transport rate of TTG1 between neighboring cells and the parameter β is the rate of active complex formation. To allow an assignment of reasonable parameter ranges and to reduce the number of model parameters a rescaling of the model variables was applied. All concentrations were multiplied by the factor β/λ3, and the new dimensionless time was expressed as τ = tλ3. The transformed, but mathematically equivalent, dimensionless equations are
The relation between the dimensional and the dimensionless parameters k1 to k5 is given in Table 3. Let ν0(i) = ([ttgl]0(i),[gl3]0(i),[ac]0(i))T denote the ith uniform steady state. Equations 1–3 have three uniform steady states given by
where f = ((k1k4 – 1)2 – 4k2k4k5)1/2. For biological relevance, all three steady states must be real and positive, which restricts the range of possible values for parameters ki.
In a pioneering work, Turing introduced the concept of pattern formation from homogeneous conditions by a diffusion driven instability; a uniform steady state that is stable for a single cell can be driven unstable by the interaction between cells [57]. On the basis of the idea of Turing, the criteria for pattern formation from a uniform steady state were derived in two steps: (i) criteria for the stability of the steady state without TTG1 mobility and (ii) criteria for an instability of the uniform steady state when adding TTG1 mobility. The stability of the steady in the absence of TTG1 mobility was analyzed by a linearization of equations 1–3 leading to ∂τΔν(i) = J(i)Δν(i). Here, Δν(i) = ν(i) – ν0(i) are small deviations from the ith steady state, and J(i) is the Jacobian matrix evaluated at steady state ν(i). A steady state is stable if small deviations from it decay with time. This is the case if all eigenvalues of the Jacobian matrix have negative real parts [58]. The eigenvalues of J(i) are the roots of the characteristic equation λ3 + a1(i)λ2 + a2(i)λ + a3(i) = 0 with the coefficients
All three roots have negative real parts if the following three necessary and sufficient criteria for the coefficients of the characteristic equation are fulfilled [58]
Next, we considered the stability of the steady state ν0(i) including the mobility of TTG1. The temporal evolution of small spatially inhomogeneous deviations Δνj(i) = ν0(i) – νj from the uniform steady state ν0(i) are again described by a linearization of Equations 1–3, now including the cellular coupling ∂τΔνj(i) = J(i)Δνj(i) + D〈Δνj(i)〉. The matrix of transport coefficients is D and has a single entry for [ttg1] at D11 = k3. Fourier analysis was used to study the temporal evolution of spatially periodic solutions of the form Δνj(i) = ∑s=1N∑r=1Mφs,r(i) e2πisy/N e2πirx/M. The transformed linear equations read ∂τφs,r(i) = (J(i) – 4Dg(s,r)) φs,r(i) with the function g(s,r) = sin2(πs/N) + sin2(πr/M) + sin2(π(s/N – r/M)). The uniform steady state ν0(i) becomes unstable to small spatial variations if any of the eigenvalues of the matrix J(i) – 4Dg(s,r) has a positive real part. The eigenvalues of J(i) – 4Dg(s,r) are roots of the characteristic equation λs,r3 + b1(i)(s,r)λs,r2 + b2(i)(s,r)λs,r + b3(i)(s,r) = 0 with the coefficients: b1(i)(s,r) = (a1(i) + 4Dg(s,r), b2(i)(s,r) = a2(i) + 4Dg(s,r))(1 + k5 + [ttgl]0(i)), and b3(i)(s,r) = a3(i) + 4Dg(s,r)(k5 + [ttgl]0(i) – 2k4[ttgl]0(i)[ac]0(i)). If any of the three necessary and sufficient criteria
are violated, then the ith steady state gives rise to a Turing instability. For the analysis, we restricted all parameters ki to be real and positive. Analysis of steady state ν0(1) revealed that conditions C1a–C1c and C2a–C2c are always fulfilled. Furthermore, if both steady states ν0(2) and ν0(3) are real and positive, then only ν0(3) fulfills conditions C1a–C1c. Therefore, only steady state ν0(3) was considered in the following. For a given parameter set, all six conditions were verified numerically. Here, it is sufficient for Turing instability if conditions C2a–C2c are violated at the maxima of g(s,r).
The parameter optimization was confined to the region in parameter space that gave rise to a Turing instability of steady state ν0(3) as defined by the criteria given above. Additionally, parameters were restricted to the biological reasonable ranges given in Table 3. Parameters were estimated by fitting the model Equations 1–3 to the experimentally determined relative fluorescence intensities of TTG1 in the vicinity of the trichomes as well as the mean trichome density in the initiation zone of the young leaf. The optimized function was
with k = (k1,k2,k3,k4,k5). The trichome number T(k) was determined from a numerical solution of Equations 1–3. The uniform steady state ν0(3) plus a small inhomogeneous perturbation were used as the initial conditions. The average total [ttg1] level of the cells in tier j around trichome i is Pi,j(k). It was normalized by the total [ttg1] level in trichome i; i.e., Pi(k). Rj is the experimentally determined average relative TTG1 level in tier j, and σR,j is the corresponding standard deviation. The levels are R = (0.387,0.765,0.935), and the standard deviation is σR = (0.14,0.22,0.183). For the mean trichome density in the initiation zone, we used μD = 0.075 with the corresponding standard deviation σD = 0.035. Both values reflect the experimental observation that the mean trichome distance in the initiation zone is between 3 and 5 cells. Because the numerical solution of T(k) and Pi,j(k) depends on the initial conditions, the optimal parameter set also depends on the initial conditions. Therefore, optimal parameters were averaged across 10 optimizations to determine the mean and standard deviation given in Table 3. For each of the 10 optimizations, a different random perturbation of the initial conditions was chosen. Parameters k4 and k5 cannot be determined simultaneously from the data. To resolve this nonidentifiability, we fixed k5 = 1. Global optimization was performed using an algorithm based on adaptive simulated annealing (Lester Ingber, http://www.ingber.com) in combination with the MATLAB interface ASAMIN by Shinichi Sakata (http://www.econ.ubc.ca/ssakata/public_html/software/). All numerical analysis was performed with MATLAB from Math Works, Inc. The predicted mean trichome density and mean percentage of the trichomes in clusters given in Figure 7C were determined from an average over 100 simulations for each parameter set.
Accession numbers for genes mentioned in this paper from the National Center for Biotechnology Information (http://www.ncbi.nlm.nih.gov) are: pAM-PAT-GW-GUS (AY02531), ppcA1 (X64143), and TTG1 (AT5G24520.1).
|
10.1371/journal.pgen.1005453 | The Drosophila bag of marbles Gene Interacts Genetically with Wolbachia and Shows Female-Specific Effects of Divergence | Many reproductive proteins from diverse taxa evolve rapidly and adaptively. These proteins are typically involved in late stages of reproduction such as sperm development and fertilization, and are more often functional in males than females. Surprisingly, many germline stem cell (GSC) regulatory genes, which are essential for the earliest stages of reproduction, also evolve adaptively in Drosophila. One example is the bag of marbles (bam) gene, which is required for GSC differentiation and germline cyst development in females and for regulating mitotic divisions and entry to spermatocyte differentiation in males. Here we show that the extensive divergence of bam between Drosophila melanogaster and D. simulans affects bam function in females but has no apparent effect in males. We further find that infection with Wolbachia pipientis, an endosymbiotic bacterium that can affect host reproduction through various mechanisms, partially suppresses female sterility caused by bam mutations in D. melanogaster and interacts differentially with bam orthologs from D. melanogaster and D. simulans. We propose that the adaptive evolution of bam has been driven at least in part by the long-term interactions between Drosophila species and Wolbachia. More generally, we suggest that microbial infections of the germline may explain the unexpected pattern of evolution of several GSC regulatory genes.
| Animals need to make gametes–sperm or eggs–in order to reproduce. Gametes are produced from a specialized tissue called the germline that is found within the testes or ovaries. These organs contain a small population of stem cells that are able to both self-renew and differentiate to generate gametes and are thus essential for maintaining gamete production throughout the reproductive lifespan of most animals. Surprisingly, some of the genes that control this process evolve rapidly between Drosophila species. We find for a key germline stem cell regulatory gene, bag of marbles (bam), that its rapid evolution affects only female but not male functions. We further report that the endosymbiont bacterium Wolbachia that infects insects and other species interacts with bam and may be contributing to the wider pattern of rapid evolution of germline stem cell regulatory genes.
| Population genetic and comparative analyses in diverse taxa have shown that many genes involved in reproduction are evolving under adaptive evolution [1–3]. Various selective pressures have been hypothesized to drive the adaptive evolution of those reproductive genes including sexual conflict, sexual selection, pathogen resistance, and avoidance of interspecific fertilization [2,4,5]. While population genetic and comparative approaches have been valuable in identifying adaptively evolving genes [4,6–11], a combination of population genetic and functional approaches is needed to identify the adaptive phenotypes and to determine the contribution of these selective pressures.
The gene bag of marbles (bam) is an intriguing example of a rapidly evolving reproduction gene, having experienced recurrent, adaptive evolution in D. melanogaster and D. simulans [12,13]. Unlike many other reproductive genes that have experienced positive selection, however, bam functions early in gametogenesis, making it unlikely that many of the selective pressures mentioned above could act on it. Surprisingly, genes involved in germ cell development and cystoblast division are over-represented genome-wide among those adaptively evolving in both D. melanogaster and D. simulans [7,14].
bam regulates germline stem cell (GSC) differentiation and germline cyst development in both males and females. GSCs are present in a niche environment that is required to maintain their stem cell state [15,16]. When a stem cell asymmetrically divides, the daughter cell, a cystoblast, moves away from the niche, which relieves repressive mechanisms and allows it to differentiate [15–17]. The cystoblast then undergoes four synchronous mitotic divisions to generate an interconnected, 16-cell cyst. In females, one of these cells will become the oocyte and enter meiosis while the remaining 15 nurse cells will become polyploid and provide nutrients to the oocyte. In males, all 16 cells will enter meiosis and give rise to mature sperm [18].
In females, bam is the key factor for inducing GSCs to differentiate and is thus transcriptionally repressed in the GSC and derepressed in the cystoblast [19–21]. Bam expression is transient, as its protein is present only in late cystoblasts, and 2-, 4-, and 8-cell cysts (Fig 1A) [22]. In males, bam is not required for GSC differentiation, as bam mutant GSCs differentiate but continue undergoing mitotic divisions and never enter meiosis [23–25]. As in females, Bam protein is expressed transiently in males, as it is present only in 4-, 8-, and 16-cell cysts (Fig 1B) [25].
Bam also functions downstream of GSC differentiation in both males and females. Bam also localizes to the fusome, an ER-like organelle that interconnects the cells of a cyst, mediates the synchrony of the mitotic divisions, and likely determines the future oocyte [22,26]. This localization requires the gene benign gonial cell neoplasm (bgcn) [27], and bam mutants show a reduction in fusome vesicles [22]. Bam also has a role in counting cyst divisions in females [22,28,29]. This function is more clearly established in males, where the accumulation of Bam to a critical threshold is required for cysts to cease mitotic divisions and initiate spermatocyte differentiation [25,30].
The molecular function of bam is not fully understood, but Bam physically interacts with and requires the function of bgcn [27,31–33] and Sex lethal (Sxl) [34–36] in GSC differentiation in females. Sxl has been shown to bind nanos mRNA, downregulating it and allowing for GSC differentiation [34–36]. Additionally, Bgcn is related to the DExH-box family of ATP-dependent RNA helicases, leading to the hypothesis that Bgcn functions together with Bam to repress translation [31]. This has been shown directly in males for the target gene mei-P26 [30].
Because bam is essential for fertility yet is involved in the early stages of reproduction, theories of sexual conflict and sexual selection that apply to many other rapidly evolving reproductive genes do not readily explain the adaptive evolution of bam. We therefore explore here interactions between bam and the bacterial endosymbiont, Wolbachia pipientis. Wolbachia is maternally inherited and manipulates host reproduction in a variety of organisms [37–40]. One report found that Wolbachia infection partially rescues the oogenesis defects of Sxl mutants in Drosophila melanogaster [41,42]. This result is an important motivation for examining possible interactions between Wolbachia and bam because a subsequent study showed that bam requires Sxl to function in GSC differentiation [34].
Wolbachia localization and activity are highly dynamic among Drosophila species and are controlled by both host and bacteria [43–48]. For example, in D. melanogaster, Wolbachia is present throughout the germline of females but preferentially accumulates at the somatic stem cell niche, a microenvironment required to maintain somatic stem cells that, when differentiated, produce follicle cells [49]. In contrast, Wolbachia preferentially localizes to the germline stem cell niche in D. mauritiana [46,49]. Transinfection and introgression studies have shown this trait to be primarily controlled by Wolbachia strain, rather than host background [48]. Wolbachia can rapidly spread through a population using a reproductive manipulation known as cytoplasmic incompatibility (CI), where Wolbachia causes the death of offspring from matings of Wolbachia-infected fathers with uninfected mothers [39]. When CI-inducing Wolbachia from D. simulans are transferred to D. melanogaster, their ability to induce CI decreases dramatically [43]. Conversely, when strains that do not induce strong CI in D. melanogaster were transinfected into D. simulans, they induced high levels of CI [50]. Additionally, some strains of Wolbachia do not cause CI, suggesting that both Wolbachia and its host control the occurrence/penetrance of CI [50].
These studies suggest that Wolbachia may be inducing species-specific adaptations, yet no studies to our knowledge have identified host genes that are candidates for mediating an adaptive response to Wolbachia. The critical function of bam in GSC differentiation and the striking consequences of bam divergence in females that we document in this study motivated us to explore interactions between Wolbachia and bam.
To identify the functional consequences of bam’s divergence, we developed a transgenic system to assay the ability of a bam ortholog from D. melanogaster or D. simulans to rescue the female and male sterility of a D. melanogaster bam mutant. We generated strains of D. melanogaster containing transgenic copies of either D. melanogaster bam (mel-bam-yfp) or D. simulans bam (sim-bam-yfp) (Fig 1C). Each bam ortholog was C-terminally tagged with Yellow fluorescent protein (YFP) and driven by the native D. melanogaster regulatory region which has been previously defined [21,23]. This approach was designed in an effort to attribute any phenotypic differences to coding sequence divergence. Each transgene was integrated separately in the same position of the D. melanogaster genome at two different attP sites on chromosome 2 (attP16a or attP40), and then crossed into a D. melanogaster bam transheterozygous, null mutant background. PCR using primers designed to the Wolbachia wsp gene confirmed that Wolbachia was not present in the transgenic or bam mutant stocks (see Materials and Methods). The nomenclature used throughout this study is described in Table 1.
qRT-PCR analyses from ovarian cDNA provided two unexpected results. First, bam expression levels in mel-bam-yfp; bam−(mel-bam-yfp/+; bamΔ86/bamΔ59, see Table 1) ovaries are 13–15-fold less than in controls with a single D. melanogaster bam allele (bam heterozygote of bamΔ59/+) generated from the same cross (Fig 2A). To determine if the unexpectedly low bam expression in mel-bam-yfp; bam−is due to a mutation caused during transformation or to a background effect, additional qRT-PCR was performed in which we found that the results are consistent in different bam mutant backgrounds (Fig 2A) and across different transgene insertion sites (Fig 2B). We also determined that bam expression in the stock from which the bam allele in mel-bam-yfp was cloned is similar to the D. melanogaster heterozygote (+/bamΔ59), demonstrating that the particular allele we chose is not defective in expression (Fig 2C). Additionally, we found that bam expression in the heterozygous genotype used as a reference is not an outlier as it is similar across several genetic backgrounds (Fig 2C). Finally, we compared bam expression in mel-bam-yfp; bam−to that of another bam transgene, bam-α, previously reported to fully rescue both female and male sterility of D. melanogaster bam mutants [23]. We found that the bam-α transgene is similarly under-expressed relative to the D. melanogaster bam heterozygote (Fig 2A). We attempted to perform similar qRT-PCR analyses of bam expression in males, but could not generate reliable results due to its low level of expression. Overall, these results demonstrate that mel-bam transgenes do not express at a wildtype level in females. This is likely caused by the lack of some regulatory sequences, although we cannot eliminate the possibility that bam transgenes are particularly sensitive to position effects. We therefore designed the genetic assays below to assess whether mel-bam-yfp is fully functional.
The second unexpected result is that bam expression in sim-bam-yfp; bam−(sim-bam-yfp/+; bamΔ86/bamΔ59, see Table 1) ovaries is similar to the D. melanogaster bam heterozygote and ~13–15-fold higher than mel-bam-yfp; bam−(Fig 2A and 2B), despite the fact that both transgenes contain the same D. melanogaster bam regulatory region. We examined protein levels by Western blots and found that sim-Bam-YFP accumulates ~2–3-fold higher than mel-Bam-YFP which is considerably less than the difference in RNA levels (Fig 2D). We conclude that bam coding sequence (CDS) divergence affects both RNA and protein levels. We were unable to assess how protein levels from each transgene compare to wildtype levels as anti-Bam antibodies did not work well on Western blots under our experimental conditions (monoclonal mouse Anti-BamC, rabbit Anti-Bam) [22,51,52]. The difference in expression levels between the transgenes does complicate the ability to attribute phenotypic differences between the orthologs to coding sequence divergence. We therefore expanded our analyses to include the D. melanogaster bam heterozygote as a control, since its expression level is not significantly different from bam levels in sim-bam-yfp; bam–, resulting in a three-way comparison: mel-bam-yfp; bam−vs. bam heterozygote, mel-bam-yfp; bam−vs. sim-bam-yfp; bam–, and sim-bam-yfp; bam−vs. bam heterozygote (See S1 Fig for crossing diagrams).
To assay transgene function, we crossed each into a D. melanogaster bam transheterozygous, null mutant background. Sibling flies that were heterozygous for bam but did not carry a transgene were used as a control for comparison in fertility experiments (S1 Fig). We found that mel-bam-yfp fully rescues D. melanogaster bam female sterility to the level of the D. melanogaster bam heterozygous control (Fig 3A), suggesting that this transgene is fully functional in females despite having a reduced expression level relative to wildtype bam alleles. However, sim-bam-yfp; bam−females were significantly less fertile than mel-bam-yfp; bam−at every time point in the experiment for both insertion sites tested (Fig 3B), demonstrating the sim-bam-yfp cannot fully rescue D. melanogaster bam female sterility.
In contrast to female fertility assays, sim-bam-yfp; bam−males were as fertile as their mel-bam-yfp; bam−or D. melanogaster bam heterozygous counterparts (Fig 4A). To test for more subtle differences in male fertility, we used a sperm exhaustion mating assay by providing the males with two new, virgin females every day over a five-day period. Surprisingly, mel-bam-yfp does not fully rescue male sterility, suggesting that this transgene is not fully wildtype in function (Fig 4B). Under sperm exhaustion conditions mel-bam-yfp; bam−males become sterile quickly which we also found when using the bam-α transgene previously reported to fully rescue bam male sterility (S2A Fig) [23], suggesting that D. melanogaster bam transgenes are unable to fully rescue male sterility. We therefore compared sim-bam-yfp; bam−to bam heterozygotes under sperm exhaustion conditions and found that sim-bam-yfp fully rescues male sterility. While we were unable to accurately quantify bam RNA expression in males due to its low expression, we found that Bam-YFP protein expressed from both transgenes localizes in testes (S2B and S2C Fig) in a manner similar to published reports [25]. These data demonstrate that sim-bam divergence strongly affects females yet causes no observable defects in males.
To determine the cause of the reduced fertility of sim-bam-yfp; bam−females, we stained mel-bam-yfp; bam−and sim-bam-yfp; bam−ovaries with antibodies to the germline marker Vasa, the fusome marker Hts-1B1, and the YFP tag in Bam-YFP. The ovaries of flies with mel-bam-yfp; bam−show wildtype morphology (Fig 5A and 5B). GSCs were identified by their spherical fusome (i.e. the spectrosome) and their location within the germarium. mel-bam-yfp; bam−ovaries had 2–3 GSCs per germarium, which is comparable to wildtype levels, and Bam was properly localized [22,53]. Furthermore, the vast majority of egg chambers underwent the proper number of cyst divisions giving rise to 16-cell cysts (S1 Table). In contrast, ovaries from sim-bam-yfp; bam−flies showed multiple ovarian defects that increased as the flies aged (Fig 5C and 5D). First, they exhibit stem cell loss, with an average of only 1.5 GSCs per ovariole when young (days 1–5; Fig 5F). Second, as the flies age (days 6–15) they appeared to have a reduction in the number of ovarioles containing mature egg chambers as a consequence of GSC loss, though we did not quantify this effect. Third, many of the egg chambers (24/100) that are present have an improper number of cyst divisions and show mitotic synchrony defects (S1 Table). Mitotic synchrony defects are typically seen with fusome mutants (e.g. hts [54] and α-spectrin [55]) suggesting that sim-bam-yfp; bam−flies may have fusome defects. However, sim-bam-yfp; bam−ovaries have both reduced and increased numbers of cyst divisions while fusome mutants have only reduced numbers, suggesting instead that sim-bam-yfp cannot properly regulate the number of cyst divisions, independently of potential fusome defects. Despite these multiple ovarian defects, it is important to note that sim-Bam-YFP shows a proper localization pattern (Fig 5C and 5D). It is absent in GSCs and present in mitotically active cysts, suggesting that the defects are not due to gross misregulation of Bam. Furthermore, sim-bam-yfp; bam−flies never show the D. melanogaster bam null mutant phenotype of tumorous ovaries [23] (e.g see Fig 5E), suggesting that sim-bam-yfp is capable of rescuing the GSC differentiation defect in D. melanogaster bam mutant females.
The above experiments suggest that sim-bam-yfp does not function properly in a D. melanogaster background and may be acting in a gain-of-function manner as observed by the loss of GSCs. We further explored this by asking if adding additional copies of the mel-bam-yfp or sim-bam-yfp transgenes either improve or worsen the fertility phenotypes. We found no significant differences in fertility when comparing mel-bam-yfp; bam−(one transgene copy) to 2x mel-bam-yfp; bam−(two transgene copies, see Table 1) (Fig 6A). However, 2x sim-bam-yfp; bam−(two transgene copies) flies showed a significant decrease in fertility when compared to sim-bam-yfp; bam−(one transgene copy) and were nearly sterile by day 15 (Fig 6B).
Ovarioles from 2x sim-bam-yfp; bam−flies showed accelerated rates of stem cell loss, even in young (1–5 day old) flies (Fig 6D and S2 Table), as compared to sim-bam-yfp; bam−(Fig 6C). They typically lacked GSCs and in some cases no longer contained any germline cells, as seen by lack of Vasa staining (Fig 6D). This phenotype contrasts with sim-bam-yfp; bam−flies, where GSCs were almost always present in every ovariole though often reduced in number (see Fig 5C and 5D).
We performed qRT-PCR comparing the ovarian RNA expression levels of bam from flies with one or two copies of the transgene. As expected, doubling the dose of the transgenes results in an approximate doubling of expression for both mel-bam-yfp and sim-bam-yfp (S3A Fig). Notably, however, bam RNA levels of 2x sim-bam-yfp; bam−are not greater than in D. melanogaster wildtype flies (S3A Fig). Additionally, sim-Bam in 2x sim-bam-yfp; bam−ovaries does not show aberrant localization when present (S3B and S3C Fig). Thus, we conclude that the 2x sim-bam-yfp; bam−defects are specifically due to increased dosage of the functionally diverged D. simulans bam, rather than to a general effect of increasing bam dosage or gross mislocalization.
We further explored the nature of sim-bam-yfp-mediated defects by asking how they are modulated in the presence of a wildtype D. melanogaster bam allele. We envisioned 3 possible outcomes. The first is that if the effects are purely due to increased dosage then they should become worse with the addition of wildtype D. melanogaster bam. The second is that if the defects are purely neomorphic as a consequence of D. simulans bam divergence, then they should be unchanged. In other words sim-bam-yfp will be dominant over D. melanogaster bam. And the third is that if the defects are due to a failure of sim-bam function due to divergence, then they should be rescued by D. melanogaster bam and thus be recessive.
We assayed our transgenes with the addition of an endogenous copy of D. melanogaster bam. We found that sim-bam-yfp; bam–-dependent defects are mostly alleviated by the addition of even a single endogenous copy of D. melanogaster bam (Fig 6E and S2 Table). This result suggests that D. melanogaster bam is dominant over sim-bam-yfp, but it is unlikely that sim-bam-yfp is simply a loss-of-function allele as the sim-bam-yfp; bam−phenotypes do not match bam loss-of-function alleles in D. melanogaster. We therefore suggest that sim-bam-yfp has both loss and gain of function attributes. Several hybrid incompatibility alleles, alleles that when expressed in a hybrid background result in sterility or lethality, show similar properties [56,57].
In D. melanogaster, Bam and Bgcn physically interact [30,32,33], and like bam, bgcn is also evolving under rapid, adaptive evolution in both D. melanogaster and D. simulans [12]. One might expect that if substitutions occurred that reduce their interaction, compensatory mutations would be selected for to re-establish a strong interaction. Therefore, independent and compensatory substitutions occurring at Bam and Bgcn within each species might render the protein partners incapable of, or less efficient at, interacting when brought together with the heterospecific protein. To determine if sim-Bam and mel-Bgcn interact with one another, we performed immunoprecipitation assays from Drosophila S2 cells. Cells were transiently transfected with either mel-Bam::HA or sim-Bam::HA, and with mel-Bgcn::MYC transgenes. We found that in reciprocal immunoprecipitation experiments both the conspecific and heterospecific Bam coimmunoprecipatated with mel-Bgcn::MYC, indicating that sim-Bam can interact with mel-Bgcn (Fig 7A–7D).
These assays involve gene over-expression and cannot discriminate whether the protein interactions are reduced in efficacy. Ohlstein et al. [31] showed that bgcn acts as a dominant enhancer of partial female sterility caused by D. melanogaster bam hypomorphic mutants. Reducing bgcn dosage exacerbated the bam phenotype, causing sterility and giving rise to completely tumorous ovaries. We reduced the copy number of bgcn by half (bgcn1/+) in sim-bam-yfp; bam−flies and found no exacerbation of the sim-bam-yfp phenotype (Fig 7E–7H). Additionally, adding a copy of sim-bam-yfp rescued the bgcn-induced sterility of the bam hypomorph (S4 Fig). Together the co-immunoprecipitation and genetic-interaction experiments strongly suggest that sim-bam-yfp; bam−ovarian defects are not due to an inability of sim-Bam to interact with mel-Bgcn.
Our transgenic rescue experiments suggest that sim-bam-yfp has diverged specifically in regards to its role in the female germline. The bacterial endosymbiont Wolbachia pipentis is maternally inherited and manipulates its host to ensure transmission [39] and could thus provide selective pressures on genes in the female germline such as bam. To explore possible interactions between bam and Wolbachia, we crossed a naturally occurring strain of D. melanogaster Wolbachia, wMel, into a heteroallelic combination of bam alleles (used above in bam genetic-interaction assays) that results in a hypomorphic phenotype [31,32]. bamBW/bamΔ59 flies lacking wMel Wolbachia are weakly fertile, giving rise to a mix of tumorous and wildtype egg chambers [31]. Thus, the number of nurse-cell positive egg chambers (i.e. non-tumorous egg chambers) can be counted to look for enhancers or suppressors of bam activity [31–33,58]. We compared bamBW/bamΔ59 flies infected with wMel, denoted as "bam +wMel", to hypomorphic flies cured of Wolbachia using tetracycline, denoted as "bam-Tet". We found that the ovarioles of bam +wMel flies contain significantly more nurse-cell-positive egg chambers than the bam-Tet flies (S3 Table).
We then assayed the fertility of the bam +wMel and bam-Tet females. We found that the presence of Wolbachia increases the fertility of bam +wMel females to high levels (Fig 8A; compare to Fig 3, days 1–5). The fertility increase was only observed in bam hypomorphs and not in combinations of bam null alleles that result in complete female sterility (bamΔ86/bamΔ59+wMel, N = 20). The fertility increase is not due to effects on bam mRNA levels, as expression is not significantly different between bam +wMel and bam-Tet females (Fig 8B). Fertility assays were also performed in males. However, bam hypomorphic males were completely sterile, and the presence of Wolbachia had no rescuing effect (bam +wMel, N = 20; bam-Tet, N = 20).
To ensure that the fertility rescue of the bam hypomorph was not due to a difference in the gut microbiota caused by tetracycline treatment, we repeated the experiment by controlling for the gut microbial composition (see Materials and Methods). The female fertility assay was repeated and produced very similar results showing that bam +wMel females are significantly more fertile than bam-Tet females (S5 Fig). This experiment demonstrates that fertility rescue of bam hypomorphs is specifically due to Wolbachia infection.
Female sim-bam-yfp; bam−flies have reduced fertility (Fig 3B). We therefore compared the fertility of mel-bam-yfp; bam−and sim-bam-yfp; bam−females with and without Wolbachia (wMel) and found that mel-bam-yfp; bam−fertility was neither enhanced nor diminished in the presence of Wolbachia. In contrast, we found a significant increase in the fertility of young sim-bam-yfp; bam−females (days 1–5) infected with Wolbachia, a result which was consistent across multiple insertion sites (Fig 9A).
If Wolbachia has co-evolved with bam, one possibility is that Wolbachia levels will be influenced by the species-specific ortholog of bam that is present in females. To test this, we used qPCR to measure wMel Wolbachia titer in ovaries and found that Wolbachia levels are reduced in sim-bam-yfp; bam−compared to mel-bam-yfp; bam−ovaries (Fig 9B).
One possible explanation for this reduced titer is that Wolbachia does not localize properly in sim-bam-yfp; bam–. While Wolbachia is present in low levels throughout the germarium, it preferentially accumulates in the somatic stem cell niche (SSCN) in D. melanogaster [46,48,49]. As germline cysts pass the SSCN, high Wolbachia titer and prolonged exposure via somatic cells that encapsulate the cyst may allow it to efficiently infect the cyst and ensure vertical transmission [48]. We examined Wolbachia accumulation using an antibody to Hsp60 which cross-reacts with Wolbachia [44,59,60]. We found that as in mel-bam-yfp; bam–, Wolbachia accumulates normally within the SSCN in sim-bam-yfp; bam−flies (Fig 9C).
A detailed comparison of bam function in D. melanogaster versus D. simulans is not possible due to the lack of available bam mutations in D. simulans. More importantly, such an approach might be insensitive to functionally important amino-acid changes if compensatory mutations have occurred in other genes in either lineage. We therefore designed a transgenic construct of D. simulans bam and transformed it into D. melanogaster, along with a parallel D. melanogaster control construct transformed into an identical place in the genome using the phiC31 transformation system [61].
We designed our constructs to have more non-coding DNA than a previously used bam transgene [21,23], yet found that both sets of D. melanogaster bam constructs have lower mRNA expression in females than a wildtype bam allele. Despite this expression difference, we found that our mel-bam-yfp construct fully rescues a bam null mutation in females. One possible explanation is that female flies are indifferent to large differences in bam levels. Alternatively, Bam protein levels may be controlled by a feedback loop that can compensate for differences in mRNA levels. This hypothesis is supported by the fact that differences in protein level between mel-bam-yfp; bam−and sim-bam-yfp; bam−genotypes are considerably smaller than the corresponding mRNA level differences (compare Fig 2A with 2D). We were not able to reliably quantify mRNA levels in males, but the inability of mel-bam-yfp to fully rescue male sterility suggests that it also under-expresses in males. If so, it would also suggest that males are more sensitive than females to lower levels of bam or that a feedback loop involving bam in females is not present in males.
Our goal in this study was to compare the effects of bam coding sequence divergence, and therefore we made the sim-bam-yfp construct using the untranslated regions (UTRs) and non-coding DNA from D. melanogaster, expecting that it would express similarly to mel-bam-yfp. Surprisingly, we found that sim-bam-yfp significantly overexpresses relative to mel-bam-yfp. One possible explanation is that sim-bam contains diverged regulatory sequences within its coding sequence or introns that affect transcription initiation. A second possibility is that these regions affect mRNA stability. Finally, it is possible that our sim-bam-yfp construct contains an intragenic incompatibility affecting mRNA stability between the D. melanogaster and D. simulans portions of its transcript. If true, then a D. simulans bam genomic transgene might have been more effective than a chimeric gene composed of sequences from both species. That alternative, however, is not a panacea because even genes that have similar expression levels between D. melanogaster and D. simulans can mis-express when placed in a foreign species due to "cis x trans" regulatory divergence [57,62,63].
We have performed several controls to minimize the complications arising from the differential mRNA expression levels of the mel-bam-yfp and sim-bam-yfp transgenes. First, we used the endogenous D. melanogaster bam locus as an additional control because its expression is not significantly different from sim-bam-yfp expression in ovaries (Fig 2A). Second, we have shown that the YFP protein localization patterns in both ovaries and testes are similar for both transgenes and resemble wildtype Bam (Figs 5A–5D, S2B and S2C).
We also note that female fertility levels do not appear to be highly sensitive to bam expression level. mel-bam-yfp; bam−and the bam heterozygote are not significantly different in their levels of female fertility even though they express at different levels. Furthermore, sim-bam-yfp fertility rescue is significantly lower than both genotypes despite having a similar expression level to the bam heterozygote. These findings provide confidence in our conclusion that sim-bam-yfp has functionally diverged in its female germline function.
Reproductive genes are strongly affected by sexual selection, adaptive divergence, and intra- and inter-sexual conflict. Many lines of evidence suggest that these forces affect males more strongly than females. For example, hybrid male sterility evolves much more rapidly than hybrid female sterility, demonstrating that functionally relevant divergence between species is more likely to occur in males [64–66]. Gene expression of male-biased genes diverges more between species than does the expression of female-biased genes [67,68]. Finally, genes encoding male reproductive proteins are among the most rapidly evolving classes of genes [2,4,5,69,70].
GSC regulatory genes also are over-represented among adaptively evolving gene classes [7,14], which is surprising considering that there is no obvious role for sexual selection or sexual conflict to operate at such early stages of germline development. Selection to increase gamete production could occur in either sex, but would perhaps be stronger in males where energetic investment in gametes is less than for females. We were thus surprised to see how clearly sim-bam-yfp divergence affects female but not male fertility, even when males were assayed under stringent sperm exhaustion conditions.
Only in females does bam function in GSC differentiation [24,71]. Forced expression of a bam transgene in GSCs results in their differentiation only in females and not males [71]. Only after males are exposed to a longer duration and occurrence of heat shock are GSCs lost in males [72–74]. Instead, bam’s primary role in males is regulating cyst divisions and entry into meiosis [24,25,30]. Elegant studies have shown that increased or decreased levels of bam result in cysts with either less or more cells per cyst, respectively, which give rise to elongating spermatids and presumably mature sperm [25]. Therefore, males may be less affected by sim-bam-yfp divergence because either they are less sensitive to bam expression differences or females have additional sex-specific functions of bam.
In our fertility assays, we found that the bam trans-heterozygous mutants used in the female fertility assay resulted in reduced rescue in male fertility assays, presumably due to the accumulation of background mutations that affect male fertility (see Materials and Methods). Therefore, transgenic experiments in males were performed using a different combination of bam alleles. We consider it unlikely that the different allelic combinations underlie the sex-specific differences we see in the ability of mel-bam-yfp or sim-bam-yfp transgenes to rescue. mel-bam-yfp expression level in females is not significantly different in these two bam mutant combinations, arguing that the different genetic backgrounds do not cause a general difference in bam expression (Fig 2A).
sim-bam-yfp; bam−ovaries display a range of defects but never the "bag-of-marbles" phenotype seen in D. melanogaster bam loss-of-function mutations. The increased severity of phenotypes with increased sim-bam-yfp dosage also argues against a loss-of-function effect. Furthermore, the presence of D. melanogaster bam does not fully rescue sim-bam-yfp; bam−defects, suggesting that it may have both loss and gain of function properties.
Since Bam and its interacting partner Bgcn are both adaptively evolving, we hypothesized that these ovarian defects might be due to an inability of sim-Bam to interact with mel-Bgcn. We provide three lines of evidence against this. First, bgcn is required for bam’s role in GSC differentiation. If this interaction were eliminated or reduced, we would expect to see tumorous ovaries but never do in sim-bam-yfp; bam−flies. Second, sim-Bam::HA and mel-Bgcn::MYC reciprocally co-immunoprecipitate with one another in S2 cells. Third, removing one copy of bgcn does not exacerbate sim-bam-yfp; bam−ovarian defects nor does it cause tumorous ovaries. This combination of biochemical and genetic data strongly suggests that sim-bam-yfp; bam−defects are due to incompatibilities with D. melanogaster genes other than bgcn.
GSC loss is one of the most striking phenotypes we discovered in sim-bam-yfp; bam−flies, a phenotype that was enhanced with additional copies of sim-bam-yfp transgenes (Fig 6). While bam is transcriptionally repressed in the GSC in wildtype D. melanogaster, there is a small amount of Bam protein present in the GSC which must be kept inactive (i.e. not cytoplasmic) [22,32,75]. One hypothesis to explain Bam silencing is that all Bam protein present in GSCs is localized at the spectrosome (i.e. round fusome), rendering it inactive in promoting differentiation [22,76]. This hypothesis is supported by data in which a subset of antibodies show Bam localized to the fusome. Bam itself is required for function of the fusome, and in bam mutants, the spectrosome shows a reduced amount of vesicular material [22]. A second hypothesis suggests that there is a small amount of cytoplasmic Bam present in the GSC, but that other proteins antagonize its activity [32,75]. Only after Bam accumulates to high levels can it titrate away antagonizing proteins and bind to other partners to promote differentiation.
Based on our data, we suggest that sim-bam-yfp; bam−GSC loss results from sim-Bam-YFP either (1) failing to localize to the spectrosome, thus leaving it active in the GSC cytoplasm, and/or (2) preventing other proteins from localizing to the spectrosome. Fusome-protein components change during fusome growth and assemble in a hierarchical manner [27,76,77]. Based on our dominance study, we also hypothesize that the fusome cannot properly form in sim-bam-yfp; bam−ovaries but can when D. melanogaster bam is added, thus allowing proper fusome localization of sim-Bam-YFP and/or other proteins. We favor this hypothesis since sim-bam-yfp flies also show mitotic synchrony defects, a hallmark of improper fusome function. Moreover, proper endocytic recycling of the fusome is required for GSC maintenance, as rab11 mutants show GSC loss and have defects similar to bam mutants [77]. We have been unable to fully test this model though as Bam-F antibodies which show fusome localization [22] are no longer available and anti-Gfp antibodies used with our bam-yfp transgenes do not show fusome localization (see Fig 5), a result seen previously with different epitope-tagged transgenes [21].
Although bam is essential for fertility of both sexes, we only detected fertility defects in female sim-bam-yfp; bam−flies. We cannot of course exclude the possibility that an unexamined aspect of male reproduction is impaired; nevertheless, it seems highly implausible that bam divergence is being driven by a selective force operating in males if the functional consequences of that divergence are so clearly deleterious in females. We therefore sought to identify selection pressures that could potentially drive female-specific functional divergence of bam. We examined the bacterial endosymbiont, Wolbachia pipientis, due to its maternal transmission and its ability to manipulate the reproduction of the hosts that it infects [39,40]. We found that Wolbachia infection increases the fertility of two different bam mutant genotypes: D. melanogaster bam hypomorphs, and sim-bam-yfp; bam−females. It might be unexpected for a D. melanogaster strain of Wolbachia to partially rescue the female fertility defects of sim-bam-yfp. However, sim-bam-yfp at least partially maintains many of the same functions of wildtype D. melanogaster bam: promoting GSC differentiation, regulating cyst divisions, and interacting with bgcn. Therefore, an interaction with Wolbachia could potentially be maintained as well. We did though find that a D. melanogaster-specific strain of Wolbachia cannot accumulate to high levels when only sim-bam-yfp is present, suggesting an incompatibility between D. melanogaster Wolbachia and D. simulans bam. The lower Wolbachia titer might also explain why the level of rescue seen in sim-bam-yfp; bam−(Fig 9A) was not to the level seen in the D. melanogaster bam hypomorph (Fig 8A).
The gene Sex lethal (Sxl) is required for bam’s function in GSC differentiation [35]. Intriguingly, Wolbachia partially rescues the female sterility of Sxl mutants in D. melanogaster. This interaction is allele-specific, suggesting that suppression is unlikely due to a general increase in germline Sxl expression [41]. Additionally, microarray studies showed no significant increase in Sxl expression when infected with Wolbachia [42]. Sxl is expressed in both GSCs and cystoblasts, while bam expression is repressed in GSCs and is active in cystoblasts and mitotically-active cysts, though each requires the other to promote differentiation [34,35]. Therefore, it has been proposed that Sxl partners with newly-expressed Bam in cystoblasts to promote differentiation by antagonizing nanos and likely other genes required to maintain GSCs [34–36]. Since bam itself provides cell-type specificity [78], we suggest that the increased fertility of Wolbachia-infected Sxl mutants is a result of increased bam activity driving the differentiation of GSCs, rather than a direct effect on Sxl activity.
bam has experienced recurrent, adaptive evolution in both D. melanogaster and D. simulans [12,13]. There is evidence that current Wolbachia infections in D. simulans have been present for at least 8.8x105 generations [79] and possibly predating the speciation of D. simulans and D. sechellia, which would be > 2.4x106 generations (assuming 10 generations/year) [80]. For D. melanogaster, however, the association appears more recent, 2.2x104-8.0x104 generations [81,82]. Therefore it is difficult based on current evidence to propose that Wolbachia has been the sole driver of bam divergence for D. melanogaster. It is possible, however, that the species has experienced recurrent infections resulting in the replacement of old infections not currently sampled today. Wolbachia can provide fitness advantages to its hosts; for example viral pathogen protection in Drosophila [83–86]. Therefore fitness benefits combined with cytoplasmic incompatibility can result in rapid displacement of less beneficial Wolbachia strains, an observation that has been reported for both D. melanogaster and D. simulans [81,87–89].
We therefore propose two models for how an interaction with Wolbachia may have driven the adaptive evolution of bam, while acknowledging that other factors may also have contributed. The first model assumes a mutualistic interaction between bam and Wolbachia and is inspired by research on the parasitic wasp, Asobara tabida, where Wolbachia is required for oogenesis to occur properly [90,91]. Pannebakker et al. [92] proposed that the initial introduction of Wolbachia infection suppressed normal host apoptosis that occurs during oocyte production, causing the wasp to adapt by upregulating apoptosis. This response, while beneficial in the presence of Wolbachia, results in hyperactive apoptosis and oogenesis inhibition in its absence [92]. Thus in this host, Wolbachia has transitioned from facultative parasite to obligate mutualist. While the precise mechanism underlying the Wolbachia effect is unknown, Wolbachia infection in insects alters the expression levels of numerous RNAs and proteins [93–96]. Thus in D. melanogaster and D. simulans, initial introduction of Wolbachia may have changed bam expression. Because these expression changes could affect fertility, strong directional selection would then act on bam to restore its proper expression in the presence of the bacteria. The result would be a mutualistic interaction between Wolbachia and Drosophila where Wolbachia provides a constant benefit to host GSC differentiation.
Our second model assumes an antagonistic interaction between bam and Wolbachia. In the ovary, GSCs continually divide, and a host must receive cues such as nutritional status and age to balance GSC division rates and GSC differentiation throughout its lifetime [97–101]. As a reproductive parasite, Wolbachia is reliant upon host oogenesis for transmission and wants to ensure that oogenesis is continually occurring. One way in which Wolbachia may increase oogenesis is to override host cues and cause GSCs to continually divide and differentiate by increasing bam activity. Wolbachia may act either directly on bam, or indirectly on antagonists of bam activity or on downstream differentiation factors. However, having too much bam activity would be deleterious to the host, as forced expression of bam in GSCs results in premature GSC loss [71]. Therefore, the host would respond by limiting overactive bam activity caused by Wolbachia infection. This conflict between host and endosymbiont over bam activity could lead to an evolutionary arms race.
The first model predicts that bam RNA and/or protein levels would be different in the presence of Wolbachia. The second model makes at least two predictions. The first is that both host and endosymbiont proteins involved in this interaction would adaptively evolve. A second prediction of the antagonistic interaction model is that each Wolbachia strain will have coevolved with its species-specific bam ortholog and that the transmission success of Wolbachia will be reduced in the presence of a heterospecific bam ortholog.
We examined the predictions of each model. For model 1, we found no evidence of altered bam expression at the RNA level, but we were unable to examine protein levels. In examining the predictions of model 2, it has already been shown that bam is adaptively evolving in both D. melanogaster and D. simulans [12,13]. While we do not know which Wolbachia genes are responsible for this interaction, the Wolbachia genomes of D. melanogaster strains (wMel) and D. simulans strains (wRi) differ dramatically. Ankyrin-repeat-domain-containing genes have extensively diversified between the two strains [102,103], which is intriguing because ankyrin repeats are known to mediate protein-protein interactions [104]. Thus this divergence may allow the different Wolbachia strains to target different host molecules [103]. In examining the second prediction of model 2, we found that the titer of D. melanogaster-specific Wolbachia is reduced in sim-bam-yfp; bam−ovaries. It is important to note that sim-bam-yfp; bam−ovaries show a range of defects, and thus could have an altered Wolbachia titer due to cellular differences from the control strain rather than a specific interaction with Wolbachia. We specifically used young flies to minimize such effects, but are unlikely to have fully eliminated them.
Further support of model 2 comes from our experiments testing Wolbachia-bam interactions. First, we find evidence of Wolbachia increasing bam activity (either directly or indirectly) in the bam hypomorph experiment where, when bam is not fully active, Wolbachia infection results in increased bam activity and thus increased fertility. We would expect the host to try to limit Wolbachia manipulation of bam and find evidence of this in our transgenic experiments where, in mel-bam-yfp; bam−flies (with wildtype fertility), Wolbachia infection is incapable of further increasing bam activity (i.e. no increase in fertility). Our data suggest that D. melanogaster has responded to Wolbachia manipulation by utilizing or perhaps developing a feedback system to regulate bam activity. The feedback structure limits the ability of Wolbachia to overactivate bam activity, thus limiting deleterious effects on the host while still allowing increased bam activity when beneficial to the host. It should be noted in regard to GSC differentiation that this interaction does not suggest that mutualism has been established because the wildtype host shows no decrease in fitness without Wolbachia. It is only under specific bam mutant conditions that we see a fitness benefit to the host. Such conditions are unlikely to be common in nature, thus limiting any fertility benefit of Wolbachia infection.
Overall, our data are more consistent with the predictions of model 2. We note, however, that the predictions of each are not mutually exclusive. While altered bam RNA/protein levels are a prediction of model 1, this prediction is not incompatible with model 2. Similarly, the predictions of model 2, adaptive evolution of the genes involved and incompatibilities between Wolbachia and host proteins are also consistent with model 1.
Our discovery of interactions between Wolbachia and bam from D. melanogaster and D. simulans suggests that bam and Wolbachia have been interacting (either mutually or antagonistically) for an extensive period. We speculate that this history of association of Wolbachia with D. melanogaster and D. simulans has had major consequences on the evolution of bam in these species. Furthermore, infection with germline parasites may explain the more widespread pattern of adaptive evolution of early acting germline development genes [7,12,14,105].
All stocks were cultured at room temperature on standard yeast-glucose media. The bamΔ86, bamBW, bamBG, and bgcn1 stocks are described in FlyBase [106]. The bamΔ59 allele was generated through a P-element excision of bam1 (D. McKearin, pers. comm.). We sequenced this allele and discovered that the excision deletes all but the 31 amino acids from the C-terminal end of the protein. All five stocks were kindly provided by Dr. Dennis McKearin (HHMI). All stocks (including CS, y w, and transgenic stocks described below) were confirmed to be free of Wolbachia infection by PCR using primers wsp81F/wsp691R [107]. The wMel-infected strain of D. melanogaster, w; Sp/CyO; Sb/TM6B +wMel, was kindly provided by Dr. Bill Sullivan.
The wMel strain of Wolbachia was introgressed by crossing wMel-infected females into bamΔ59/TM3, generating bamΔ59/TM3 +wMel. The bamΔ59/TM3 +wMel stock was then cured of Wolbachia by feeding the flies on media supplemented with 0.03% tetracycline for three generations, generating bamΔ59/TM3 Tet. Females of the bamΔ59/TM3 +wMel stock were then backcrossed to males of the bamΔ59/TM3 Tet strain for at least six generations to generate genetically similar backgrounds including the mitochondria.
ΦC31-mediated transformation was used to generate transformants in D. melanogaster [61] and was performed by Genetic Services, Inc. Correct integration was assayed using a PCR-based assay developed by Venken et al. [109]. For the attP40 site at cytological position 25C6, the primer pair 949/1125 was used to check docking-site specificity. We discovered that the attP16 stock contains at least two attP docking sites at unknown locations. Southern blots using a probe designed to the white locus present on p{Casper4}\attB were used to determine that p{mel-bam-yfp} and p{sim-bam-yfp} both integrated into the same attP site (S6 Fig). We refer to this attP site in the attP16 stock as attP16a. All transformants were then outcrossed for at least six generations to a y w strain that had been inbred for 10 generations, to make the genetic backgrounds similar.
All crosses were performed at room temperature (22–23°C). Prior to crossing all flies were aged for 2–3 days post-eclosion on media supplemented with yeast. In female fertility experiments, single transgenic females were crossed to two wildtype Canton S (CS) males. The trio of flies were transferred to a new vial every five days for a total of 15 days and then discarded. Progeny from each vial were counted for 8 days after the first flies eclosed. In male fertility experiments, single males were mated to two wildtype CS females as described above. In sperm exhaustion assays, single males were mated to two wildtype CS females. The males were aspirated without anesthetizing into new vials containing two fresh CS females every day for 5 days. The females remaining in the vial were transferred to a new vial every five days for 10 days, and fertility was assessed by scoring the number of progeny that eclosed over 8 days.
For female fertility assays, the transgenes were crossed into the bam mutant background bamΔ86/bamΔ59. For male fertility we found that use of bamΔ86/bamΔ59 resulted in reduced fertility of mel-bam-yfp flies relative to the D. melanogaster bam heterozygous control, suggesting that background mutations in these mutants reduce male fertility. It is also likely that combinations of bamΔ86 or bamΔ59 with bam1, the chromosome from which they were derived, will share these background effects. Therefore, all male fertility experiments were done with the transheterozygous combination bamΔ86/bamBG, which are independently-derived mutations of bam. In this background we found no reduction of fertility of mel-bam-yfp; bam−relative to the D. melanogaster bam heterozygote under normal fertility assays.
To control for effects of tetracycline treatment on the gut microbiome in the bamΔ59/TM3 Tet line, axenic versions of the bamΔ59/TM3 +wMel and the bamΔ59/TM3 Tet lines were generated and the gut microbiota from conventional (i.e., non-axenic) bamΔ59/TM3 +wMel males were introduced to both lines. To generate axenic lines, embryos (less than 18 hour old) from the bamΔ59/TM3 +wMel and the bamΔ59/TM3 Tet lines were collected and dechorionated with 0.6% sodium hypochlorite. Sterile embryos were then seeded onto standard sterile yeast-glucose media. Embryos were allowed to develop into adults, and to ensure the lines were microbe-free, 5 adults from each line were homogenized and all were plated onto MRS agar [110]. Axenic flies of each line were allowed to mate for one generation. To introduce a homogenous population of gut microbiota to the two lines and to control for genetic background, axenic virgin females were backcrossed for three generations to conventional males of the bamΔ59/TM3 +wMel line collected from a single bottle. BC3 virgin females were then crossed to conventional bamBW males to generate the bamBW/bamΔ59 hypomorphic genotype.
Fecundity of these bam hypomorphs with and without Wolbachia was then assayed as follows. Prior to crossing all flies were aged for 3 days post-eclosion. Single bamBW/bamΔ59 +wMel or bamBW/bamΔ59 Tet females were crossed with two wildtype Canton S males. The trio of flies was removed from the vial after 6 days and adult progeny were counted every other day for a total of 8 days. To ensure that Wolbachia infection status was accurately maintained, each mated female was homogenized at the end of the experiment and Wolbachia status was assayed by PCR with primers designed to wsp (wsp440F/wsp691R) and dprA genes (dprA483F/dprA663R). Female fertility was only analyzed for females whose Wolbachia status was consistent with the status of the original stock as determined by typing with PCR.
Flies were aged 2 days on media supplemented with yeast. Ovaries were dissected in 1XPBS and total RNA was isolated from 10 ovaries using Trizol reagent (Invitrogen) following the manufacturer’s protocol. Samples were treated with 20 units DNaseI at 37°C for 2 hours (Roche) and purified using RNeasy columns (Qiagen) following the manufacture’s protocol. cDNA was generated from 2μg of total RNA using the Superscript III First Strand Synthesis kit (Invitrogen) and oligo-dT primers following the manufacturer’s protocol. Quantitative RT-PCR was performed on a Biorad MyiQ cycler using iQ SYBR Green Supermix (Biorad). For bam, primer pair 1160/1170 amplified bam from both species with high efficiencies. For rpl32, primer pair 844/845 from Maheshwari and Barbash [57] was used. The standard curve method was used to estimate bam and rpl32 levels. Three technical replicates were performed from at least three biological replicates for each sample.
To assay levels of Wolbachia, qPCR was performed on genomic DNA as in [44,111]. Females who eclosed on days 1–2 were aged on media supplemented with yeast for 2 days post-eclosion. DNA was isolated from 10 ovaries using phenol-chloroform extraction followed by 2 rounds of ethanol precipitation and rehydration in water.
For Wolbachia, primer pair wsp440F/wsp691R was used [111]. For rpl32, primer pair 844/845 was used. The standard curve method was used to estimate levels of each product. Three technical replicates were performed from at least three biological replicates for each sample.
D. simulans bam was amplified from w501 ovarian cDNA using primers 662/661, cloned into pENTR/D-TOPO vector (Invitrogen), verified by sequencing, and recombined into destination vectors using LR-Clonase II (Invitrogen) following manufacturer’s directions. D. simulans bam was recombined into pAFHW containing both Flag and HA epitope tags (http://emb.carnegiescience.edu/labs/murphy/Gateway%20vectors.html). D. melanogaster bam in pAFHW and D. melanogaster bgcn in pAFMW were kindly provided by D. McKearin [33].
Combinations of pAFMW-Bam and pAFHW-Bgcn or empty vectors were co-transfected into Drosophila S2 cells, cells incubated for 3 days, and then lysed in lysis buffer (50mM Tris-HCl pH7.8, 150mM NaCl, 0.1%NP-40). Anti-HA (Roche, 3F10) or anti-Myc (Roche, 9E10) antibodies were conjugated to 50 μl of Protein G Dynabeads (Invitrogen) in 200ul of PBST (0.01% Tween 20) at 4°C overnight with rotation. Antibody-conjugated beads were then added to cell lysate (80μg total protein) in 200μl in lysis buffer containing 1X protease inhibitor (Roche) and 1mM PMSF and incubated at 4°C overnight. Washes were performed following manufacturer’s directions and Dynabeads were boiled in 1X SDS sample buffer to elute protein.
25–35 ovaries from females aged 2–3 days post-eclosion on media supplemented with yeast were homogenized in lysis buffer (50 mM Tris-HCl pH 7.5, 10 mM EDTA, 1.25% TritonX-100, 1X protease inhibitor, Roche) and centrifuged at 14000 rpm at 4°C for 5 minutes. Total protein in the supernatant was estimated using the Bradford assay (Biorad) and samples were boiled in an equal volume of 4X SDS sample buffer for 5 minutes. 10–20 μg were loaded on 10% SDS-PAGE gels. Primary antibodies were anti-GFP Jl-8 (Clontech, 1:2000) and mouse anti-tubulin T5168 (Sigma; 1:120,000). Secondary antibodies were HRP conjugated goat anti-mouse (Jackson; 1:1,000 for anti-GFP and 1:60,000 for anti-tubulin) and were detected with ECL Western blotting substrate (Pierce).
Immunostaining was performed as in Aruna et al. [112]. Primary antibodies were: anti-GFP (Invitrogen A6544, 1:200), anti-vasa (DSHB, 1:25), anti-1B1 (DSHB, 1:4), monoclonal anti-Bam (1:100). Anti-Bam antibody was provided by D. McKearin. Secondary antibodies including goat anti-rat, anti-rabbit, or anti-mouse were conjugated with Alexa fluor dyes (Molecular Probes, 1:200–1:500). Samples were mounted in Vectashield containing DAPI (Vector Laboratories) and analyzed using the Leica SP2 confocal microscope at the Cornell University Core Life Sciences Microscopy and Imaging Facility. Images were resized in Photoshop (Adobe, version 11.0).
|
10.1371/journal.pntd.0006663 | Schistosoma haematobium effects on Plasmodium falciparum infection modified by soil-transmitted helminths in school-age children living in rural areas of Gabon | Malaria burden remains high in the sub-Saharan region where helminths are prevalent and where children are often infected with both types of parasites. Although the effect of helminths on malaria infection is evident, the impact of these co-infections is not clearly elucidated yet and the scarce findings are conflicting. In this study, we investigated the effect of schistosomiasis, considering soil-transmitted helminths (STH), on prevalence and incidence of Plasmodium falciparum infection.
This longitudinal survey was conducted in school-age children living in two rural communities in the vicinity of Lambaréné, Gabon. Thick blood smear light microscopy, urine filtration and the Kato-Katz technique were performed to detect malaria parasites, S. haematobium eggs and, STH eggs, respectively. P. falciparum carriage was assessed at inclusion, and incidence of malaria and time to the first malaria event were recorded in correlation with Schistosoma carriage status. Stratified multivariate analysis using generalized linear model was used to assess the risk of plasmodium infection considering interaction with STH, and survival analysis to assess time to malaria.
The overall prevalence on subject enrolment was 30%, 23% and 9% for S. haematobium, P. falciparum infections and co-infection with both parasites, respectively. Our results showed that schistosomiasis in children tends to increase the risk of plasmodium infection but a combined effect with Trichuris trichiura or hookworm infection clearly increase the risk (aOR = 3.9 [95%CI: 1.7–9.2]). The incidence of malaria over time was 0.51[95%CI: 0.45–0.57] per person-year and was higher in the Schistosoma-infected group compared to the non-infected group (0.61 vs 0.43, p = 0.02), with a significant delay of time-to first-malaria event only in children aged from 6 to 10-years-old infected with Schistosoma haematobium.
Our results suggest that STH enhance the risk for P. falciparum infection in schistosomiasis-positive children, and when infected, that schistosomiasis enhances susceptibility to developing malaria in young children but not in older children.
| Despite the progress made in the last decade, malaria remains a serious public health issue in sub-Saharan region, where it overlaps with helminths infections. The interactions between both are manifold and complex, requiring further investigation. We report here that Trichuris trichiura or hookworm associated with schistosomiasis increase the risk of Plasmodium falciparum infection; whilst schistosomiasis is independently associated with a malaria increase in young children, but not in the older children. Our finding is an additional evidence that optimizing helminth control mainly in children contributes to overcoming malaria in areas endemic for both parasitic infections where children are those who bear the highest burden of these infections.
| Over the past fifteen years, morbidity and mortality due to malaria have globally decreased. However, sub-Saharan Africa, where 90% cases of malaria and 92% of deaths related to malaria have occurred in 2015, still bears the highest burden of the disease [1]. Most of these cases remain confined to rural and semi-urban areas [2–4] where helminths are co-prevalent [5–7]. In these areas where malaria significantly overlaps with helminth infections, several studies have reported interactions between the two parasitic infections at both immunological [8–12] and epidemiological [13–16] levels. Studies have reported an effect of helminths on the cellular and humoral immune responses to the malaria parasites mainly in children [8–12,17]. Some authors have reported that this effect leads to the aggravation of clinical manifestations of malaria. Indeed, it has been shown that Trichuris trichiura infection was associated with increased malaria prevalence, while an increased helminth burden was associated with increased Plasmodium falciparum or Plasmodium vivax parasitemia [18], or enhanced anemia in co-infected children [19]. Another author has reported a positive effect of helminths on malaria outcomes. Indeed, Nacher et al found that helminths, particularly Ascaris, may have a role in the establishment of malaria tolerance in Thai patients [20]. However HIV co-infection complicated the picture further. Indeed, high prevalences of helminth infections and malaria have been reported in HIV positive people, particularly in pregnant women under anti-retroviral therapy (ART) in Rwanda [21–23] with 10% co-infections with both [23]. These high infection prevalences are found to be associated with a low CD4 counts and moreover, each of these infections is a risk factor for the other [22].
The situation is similarly unclear when it comes to Schistosoma spp. infections. It has been reported that the effect of schistosomiasis on malaria may depend on the Schistosoma species [24], or may be conflicting even for the same species [15,25–27]. Indeed, some reports have indicated that infection with S. haematobium can confer protection against severe malaria in children [25], reducing the risk of progression to symptomatic disease in long-term asymptomatic carriers of P. falciparum [15], or can delay the occurrence of a malaria episode in children [26]; whereas others found that S. haematobium may increase the prevalence of P. falciparum parasites in co-infected children [27]. In contrast, Schistosoma mansoni was reported to significantly increase the malaria incidence rate in children [28]. These finding provide evidence of the effect of schistosomiasis on Plasmodium infection. Current results on schistosomiasis and plasmodial co-infection are conflicting as reviewed recently by Adegnika et al. [24]. Most studies conducted to address these co-infections are cross-sectional in nature and could be limited in their capacity to precisely examine interactions between schistosomiasis and malaria. In this study, we conducted a longitudinal survey in order to address this issue in an area where S. haematobium and P. falciparum are the main prevalent species of schistosomiasis and malaria [13,14]. We thus assessed the effect of S. haematobium on clinical and parasitological aspects of P. falciparum infection in school aged children living in this co-endemic area, including the effect of soil-transmitted helminths (STH) in this association.
The study was approved by the institutional ethics committee of CERMEL, reference number: CEI-MRU 002/2012. Parents or legal representatives of the participant gave a written informed consent. The study was conducted in line with the Good Clinical Practice (GCP) principles of the International Conference on Harmonisation (ICH) [29] and the Declaration of Helsinki [30].
The study took place at CERMEL (Centre de Recherches Médicales de Lambaréné). Data and samples were collected from May 2012 to December 2014 in Bindo-Makouké villages (BM) and Zilé-PK villages, two settlements in the vicinity of Lambaréné [14] situated approximately at 60 km and 120 km, respectively, to the South of the Equator. The rainfall is perennial except for the long dry season (from June to September) with a mean of 1,216 mm per year [31]. The region is irrigated by the Ogooué River and its tributaries, with many ponds, lakes and streams constituting favourable conditions for fresh-water snail habitation. Recent published data demonstrate that the prevalence in the area for S. haematobium range from 15% to 75% [9,13,14]. Water supply, fishing, household work, fetching water and playing are some activities which expose the local population to schistosomiasis. Malaria transmission is perennial and the dominant malaria parasite species is P. falciparum [32,33].
The study was designed as a prospective longitudinal study.
School-age children living in two vicinities of Lambaréné (Nzilé-PK villages and Bindo-Makouké villages) were invited to participate in the study. Volunteers without any known chronic diseases other than possible helminth co-infections, and living in the study area for at least one year before inclusion were eligible to take part to the study. During the survey, participants found with a recurrent or severe disease other than malaria or helminth infection were excluded from the study.
A previous study conducted in the vicinities of Lambaréné reported a 42% prevalence for plasmodium infection among school age children [9]. To be able to detect a minimum of 12.5% prevalence difference of plasmodium infection between children with schistosomiasis and those without, with a minimum of 80% power, we needed to include in the study at baseline at least 249 children for each study group, giving a total of 498 volunteers school age children.
Field-workers went to each house and school of both villages to invite through their parents or legal representatives school-age children to participate in the study. Eligible and consenting volunteers were included. At baseline, demographics (age, sex and location) and anthropological (weight, height) data were collected. Axillary temperature was recorded. S. haematobium infection, P. falciparum infection and soil-transmitted helminths (STH) status were assessed. Participants were treated if they were found to harbour either of those parasitic infections. The follow-up consisted of two kinds of visit: active visits consisted of monthly home visits for any malaria-like symptoms assessment and recording of any medication intake; and passive visits were ad-hoc presentations of participating children at the research centre for any health issues, including flu-like symptoms.
Malaria status was defined as positive thick blood smear (TBS) associated with fever, or history of fever in the past 48h from the time of visit. Fever was defined as an axillary temperature of 37.5°C or higher. In case malaria was diagnosed, urine filtration was performed to assess evidence of co-infection with urogenital schistosomiasis. Urine filtration was also performed during the follow-up every time the children had visible haematuria.
Study groups were determined based on the schistosomiasis status. This was done differently for baseline analysis and for longitudinal analysis. At baseline and for baseline analysis, participants found infected with S. haematobium were assigned to the ‘Schistosoma-positive’ (S+) group and the others to the ‘Schistosoma-negative’ (S-) group. For longitudinal analysis, study groups were formed at the end of the follow-up period, and we assigned any participants found with infection at baseline and at any time point of the study course to the ‘Schistosoma-infected’ (SI) group. Those found negative at baseline and who did not experienced schistosomiasis during the study course were assigned to the ‘Schistosoma-uninfected’ (SU) group. Time of exposure to malaria infection for incidence calculation did not include the first 28 days after each malaria treatment.
In accordance with the national guidelines, treatment of schistosomiasis consisted of the administration of 40 mg of praziquantel per kilogram body weight once; asymptomatic P. falciparum parasitemia and malaria episodes were treated with tablets of 20/120mg of artemether-lumefantrine combination therapy given according to the body weight twice a day, in three consecutive days. Treatment of STH was a once-daily dose of 400 mg of albendazole for three consecutive days [34]. For any other cause of a disease episode, the participant was referred to the appropriate health centre.
Detection of malaria parasites was done microscopically by TBS using the Lambaréné method as described elsewhere [35–37]. Detection of S. haematobium eggs was done by filtration of 10ml of fresh urine using a 12μm pore-size filter as previously described [38–40]. For the diagnosis of urogenital schistosomiasis, urine samples were collected over three consecutive days, unless the participant was found positive with at least one parasite egg in any sample before the second or the third day. The Kato-Katz technique was performed to assess the presence of A. lumbricoides, T. trichiura and hookworm in fresh stool samples [41]. For each time point of STH assessment, one stool sample was collected. For each stool sample, two slides were performed and each slide was independently read by two readers.
Data were captured on the patient report form (PRF), entered in Access 2013 software and transferred to R software version 3.2.4 for analysis. Univariate and multivariate analysis were performed applying the Generalized Linear Model (GLM). For multivariate analysis, first we considered the interaction between asymptomatic P. falciparum infection as the main variable and each explanatory variable. In case of effect measure modification, the analysis was stratified on the variable, and the Breslow test was done to assess the homogeneity of the strata. Otherwise, the variable was evaluated as confounding factor to be include in the final model. Ten per cent (10%) or more difference of estimated measure of association before and after adjustment was used to define confounding factors. The effect of STH infection was assessed separately with respect to the species. Incidence of malaria was estimated in person-year according to each variable. A Kaplan Meier curve was drawn to assess time-to-malaria occurrence. The Log-rank test was used to compare the curves and the Cox model was used for adjusted analysis.
Among the participants who were invited to participate in the study, informed consent was granted for 754 children by their parents or their legal representatives. Of those, a total of 739 children with schistosomiasis and P. falciparum status available were included at baseline (Fig 1). Among participants enrolled, 68 (9%) children were not able to provide sample stool at baseline and from the others, 31% [95%CI: 27%-35%] were infected with STH. The most prevalent infection was trichuris with 21% [95%CI: 18%-24%] followed by ascaris and hookworm with 19% [95%CI: 16%-22%] and 6% [95%CI: 5%-8%], respectively. Mean age of these study population was 10.4 (SD = 3.1) years, the boy-to-girl sex ratio was 1.1:1 (Table 1). Of participants included, 586 (79%) agreed to be followed-up for malaria incidence.
The prevalence of P. falciparum and S. haematobium at baseline was 23% [95%CI: 20%-26%] and 30% [95%CI: 27%-34%], respectively; with 67 (9%, [95%CI: 07%-11%]) participants co-infected with both parasites. Eight per cent of participants infected with P. falciparum had fever. As shown in Table 2, both infections were more prevalent in Zilé-PK villages compared to Bindo and Makouké villages with 29% [95%CI: 24%-34%] vs 18% [95%CI: 14%-22%], respectively, for P. falciparum and 45% [95%CI: 40%-51%] vs 19% [95%CI: 15%-23%], respectively, for S. haematobium. The prevalence of both infections was similar for age and sex groups.
Among the participants followed-up for malaria incidence assessment, 216 (37%) were found positive for Schistosoma infection during the study period, including 176 cases on inclusion and additionally 40 new cases during follow-up and assigned to SI group. The 368 (63%) others participants who remained negative during the whole study course were thus assigned to the SU group (Fig 1). As shown in Table 3, the two study groups were comparable for all parameters except for location and P. falciparum parasite carrier status. Indeed, 168 (78%) of SI children came from Zilé-PK villages while 253(69%) of SU children came from Bindo and Makouké villages (p<0.001). Additionally, prevalence of P. falciparum parasite carriage was significantly higher in the SI group compared to the SU group (31% vs 20%, p = 0.004).
At crude analysis as given in Table 4, Schistosoma infection (p = 0.002) and location (p<0.001) were associated with P. falciparum infection. Children infected with S. haematobium had a 1.8 [95%CI: 1.2–2.5] times odds of being co-infected with P. falciparum parasites compared to their non-infected counterpart. After adjustment (Table 4), P. falciparum infection remains associated only with location (p = 0.02) while a trend of association with schistosomiasis infection was found (p = 0.06).
In the following analysis, we found effect modification of Trichuris or hookworm infections on P. falciparum and S. haematobium infections association. As presented in Table 5, analysis stratified on those two infections showed that among study participants without T. trichiura and hookworm infections, there is no effect of S. haematobium on P. falciparum parasite carriage; while among those infected with either hookworm or T. trichiura or a combination, children co-infected with S. haematobium had a 3.1 ([95%CI: 1.5–6.4], p = 0.002) time odds of being infected with P. falciparum. This finding remained statistical significant when adjusted for age, sex, ascariasis and location (aOR = 3.9 [95%CI: 1.75–9.19], p < 0.001). Age, sex and Ascaris infection were forced in the final model of the GLM analysis.
During the 19 months follow-up phase for P. falciparum malaria incidence assessment, 210 (36%) participants had developed a total of 318 new cases of malaria (Table 6). The overall incidence was 0.51 [95%CI: 0.47–0.55] per person-year. Taking into account the study groups, participants in the SI group had 1.4 [95%CI: 1.1–1.8] times the risk of developing malaria compared to their counterparts in the SU group.
The time-to-first malaria episode was assessed for the first twelve months of follow-up of each participant and those who did not develop malaria before the end of that time were censored. Our results show that in the SI group, among the 101 (46%) participants developed malaria, the median time to first malaria episode was 52 weeks. For their counterparts of SU group where 109 (30%) participants developed malaria, the median time to first malaria episode was not reached at the end of 52 weeks of follow-up. As presented in Fig 2, we found a significant delay to malaria in the SU group compared to SI group (Log-Rank test: p = 0.00037). Assessing the delay until development of malaria according to schistosomiasis status, we found that SI group participants had a 1.6 (Cox model: p = 0.0004) times increased risk of early development of malaria as compared to their counterparts of the SU group. This association remains significant when adjusted for location, age and sex (aRR = 1.9, p = 0.000034). Stratifying for age yielded a significant delay in time-to-first-malaria episode in the SU group (median time not reach) compared to the SI group, where the median time was 51 weeks (p = 0.00003) for 6–10 year old children. Children with schistosomiasis had a 2.1 (Mantel-Cox test: p = 0.00004) times increased risk of developing malaria compared to children without schistosomiasis. On the contrary, there was no difference in terms of delay in time-to-first malaria episode between the two study groups (p = 0.41) in children aged 11–16 years.
In area where schistosomiasis is endemic, the question of its effect on malaria outcome is a growing concern. Our study area is endemic for both infections [14], and our study reveals that up to 11% of school-age children could be co-infected with P. falciparum and S. haematobium parasites, comparing to up to 15% of pregnant women co-infected in Cameroon [42] or 23% of children co-infected with S. mansoni in a co-endemic area of North-Western Tanzania [43]. In our study area, poly-parasitism is evident [8,9,13,14,24]. The prevalence of STH species ranged from 32% to 48% among children infected with S. haematobium. This finding is not surprising since STH is commonly reported to be prevalent in rural areas [43]. Therefore, the risk to be infected by multiple parasites including S. haematobium is high [44]. Since these intestinal parasites are known to be able to modulate the immune system of the host, it would be necessary to assess the effect of these infections on Schistosoma-P. falciparum association.
Everything else equal, we found a trend of association between risk of P. falciparum carriage and Schistosoma infection. In univariate analysis, Adedoja et al. found that children infected with S. heamatobium have equal chances of being infected with P. falciparum as children with no worm infection [45], while Ateba and collaborators found a significant increase of plasmodium asexual parasite prevalence among Schistosoma infected children in comparison to the uninfected [9]. Conflicting results found in the literature on schistosomiasis-malaria co-infection issue suggest that there are potential confounding factors not yet established which need to be taken into account. In this study, we found an effect-measure modification of Trichuris and Hookworm infections on the association between S. haematobium and P. falciparum. As well, location was identified as confounding factor. Some authors have previously shown that T. trichiura [18] and hookworm [45,46] individually can affect the association between schistosomiasis and malaria co-infection by increasing the risk of being infected with P. falciparum parasites. Our analysis was stratified for those two STH infections and adjusted for age and sex which could affect malaria infection [47], and for location found in our analysis as confounding factor. The result shows that when considered only schistosomiasis infection, there is a trend on the risk of being infected with P. falciparum. But, in combination with trichuriasis or hookworm infection, schistosomiasis clearly increases the risk of being infected with P. falciparum. This result shows that S. haematobium alone does not predispose to P. falciparum infection in children instead of combined effect. We hypothesize that the cumulative effects of Schistosoma, Trichuris and Hookworm infections on P. falciparum parasite carriage acts at the immunological level. A potential immuno-modulation effect of a poly-parasitism not measured in our study could explain the combined effect of helminths we have observed. Indeed, there is evidence that Schistosoma infection can modulate the immune system in response to P. falciparum [9,11]. There is also evidence that T. trichiura can exert an influence on the immune response, and for instance negatively affect the antibody response to malaria vaccine candidate in children [17]. However, these potential cumulative effects of helminth infections on plasmodium infection need to be properly investigated at immunological level.
The overall incidence of malaria was 0.51 per person-year. This incidence was higher among people infected with S. haematobium compared to the uninfected, suggesting that schistosomiasis infection increases the risk of developing malaria. Children infected with S. haematobium infection had 1.4 times the risk of developing malaria than uninfected.
Regarding time-to-first malaria infection, we found that malaria occurred earlier in participants infected with Schistosoma than those uninfected even after adjustment for age, sex and location. The main symptom we have considered to define malaria was fever, which is one of the results of some endogenous pyrogen molecules activities, notably pro-inflammatory cytokine TNF-α during the infection. Some authors reported that during malaria infection, the production of pro-inflammatory cytokines as well as of anti-inflammatory cytokines can be affected by co-infection with schistosomiasis infection in an age-dependent manner [11,48]. In our study population, the assessment of the delay-to-malaria in relation to age group shows that there is no difference in terms of delay in time-to-first malaria in children aged from 11 to 16 years, while the difference was significant in children from 6 to 10 years. Children aged from 6 to 10 years infected with S. haematobium developed malaria earlier than those without S. haematobium infection. This finding could support the possible effect of age on the immune responses of malaria in co-infected subjects. However, since the finding is based on statistical significance, biological assessment is suitable for confirmation.
We can retain that exposure to schistosomiasis enhances incidence of, and susceptibility to develop malaria in our study population. This finding corroborates with previous reports like the one by Sokhna and collaborators who reported an increased in susceptibility to developing malaria in co-infected children, even though it was only in children excreting high S. mansoni eggs loads [28]; supporting therefore the hypothesis that schistosomiasis negatively affects the outcome of malaria. This stands in opposition to the idea that schistosomiasis possibly improves the outcome of malaria. Indeed, it was reported, for instance, that protection from malaria is conferred by asymptomatic P. falciparum infection or co-infection with S. haematobium in a Malian study cohort [15]. In the study presented here, we assessed the effect of having been S. haematobium-infected on malaria instead of becoming infected at time of malaria, which could affect our conclusion compared to the studies mentioned above. On the other hand, it has been shown that STH can affect susceptibility to malaria infection by acting at the immunological level [49,50]. Not having considered the STH status of participants in our analysis could have affected our results; however, since all participants were assessed and infected ones were treated for STH at inclusion, we assume that the effect of STH was minimized. The prevalence of STH was similar between the both study groups at baseline, and the STH treatment effect was considered as equally distributed between groups.
We have assessed the effect of schistosomiasis on clinical and parasitological aspects of P. falciparum infection based on prevalence of P. falciparum parasite carriage and malaria incidence. We therefore grouped our study population in relation to the schistosomiasis status. If it was easy at baseline to discriminate children infected or non-infected with S. haematobium, the problem we faced during the follow-up phase was to appropriately group our population in accordance with schistosomiasis status. Subjects infected at baseline or during the follow-up phase were treated systematically. However, they were considered as Schistosoma-infected for the whole follow-up phase and those who were not found positive throughout the survey were considered as non-infected. This approach was sustained by the fact that schistosomiasis is known as a chronic infection and, in areas where schistosomiasis is prevalent, the risk factors as playing habits, swimming, taking baths, washing clothes, distance from river are usually constant [51,52] and therefore the probability to be re-infected after treatment is high [53]. Thus, we have assessed the effect of schistosomiasis infection on P. falciparum parasite carriage at baseline and on malaria infection during the follow-up study phase.
An earlier conducted study in the same population showed that PCR has a better sensitivity than microscopy for the detection of P. falciparum parasites [9]. We have used the light microscopy Lambaréné method for the detection of P. falciparum parasites as it is the clinical gold standard. However, this may lead to potential misclassification of the participants regarding P. falciparum status at baseline. We think that if prevalence of P. falciparum carriage could be underestimated, this potential misclassification of participants would have been equally distributed in both groups and would not therefore affect the trend of our results.
This study confirms that the transmission of schistosomiasis is not evenly distributed in the vicinity of Lambaréné. Schistosomiasis infection is present in many villages but the prevalence varies significantly from one point to another. For example, we found a moderate prevalence for Bindo and Makouké villages where 19% of our study participants were found to be positive when compared to Zilé-PK villages, where 45% of our study participants were found to be positive. This corroborates with a previous pilot study conducted in the same population in 2012. The earlier-indicated prevalences of 15% and 43% for Bindo and Zilé-PK villages, respectively [14], suggest that prevalence of schistosomiasis infection is stable on each location. It was suggested that the difference observed could be explained by the fact that in Zilé-PK villages, streams represent the first source of water compared to Bindo village. The same observation could be applied to Makouké village where piped water is available for the majority of the population. Indeed, the lack of pipe water supply observed in the PK area promotes daily open freshwater contact by the population for household activities, bathing and playing, using the streams well known as schistosomiasis foci. In addition to humans, other ecological factors influence Schistosoma host snail density [54], which affect schistosomiasis prevalence. Therefore, we can assume that such factors may also sustain the difference of prevalence for schistosomiasis observed between the both locations, which requires further research. On the other hand, we have observed that areas where S. haematobium prevalence is high, a high prevalence of P. falciparum carriage was also found. Indeed, the difference in prevalence observed in favour of the PK area for S. haematobium infection was also observed for P. falciparum. This observation suggest a correlation of factors affecting both infections as either a consequence of the presence of same environmental risk factors. Another explanation could be indeed the effect of S. haematobium infection on P. falciparum infection, as demonstrated above. However, this need to be more investigated.
In summary, this study demonstrates that S. haematobium infection alone does not increase the risk of being infected with P. falciparum parasite but when associated with STH particularly with T. trichiura and hookworm, the risk does increase. On the other hand, in people exposed to schistosomiasis infection, risk and susceptibility of developing a malaria event increase in an age-dependent manner. Our results suggest that Schistosoma and probably STH co-infections in general cumulatively impact on malaria outcome in school-age children and therefore need to be accounted for when designing malaria control programs. Thus, in areas of co-endemicity and in support of higher efficiency, STH and schistosomiasis control should be considered as an additional tool of malaria control.
|
10.1371/journal.pcbi.1003444 | An Allosteric Signaling Pathway of Human 3-Phosphoglycerate Kinase from Force Distribution Analysis | 3-Phosphogycerate kinase (PGK) is a two domain enzyme, which transfers a phosphate group between its two substrates, 1,3-bisphosphoglycerate bound to the N-domain and ADP bound to the C-domain. Indispensable for the phosphoryl transfer reaction is a large conformational change from an inactive open to an active closed conformation via a hinge motion that should bring substrates into close proximity. The allosteric pathway resulting in the active closed conformation has only been partially uncovered. Using Molecular Dynamics simulations combined with Force Distribution Analysis (FDA), we describe an allosteric pathway, which connects the substrate binding sites to the interdomain hinge region. Glu192 of alpha-helix 7 and Gly394 of loop L14 act as hinge points, at which these two secondary structure elements straighten, thereby moving the substrate-binding domains towards each other. The long-range allosteric pathway regulating hPGK catalytic activity, which is partially validated and can be further tested by mutagenesis, highlights the virtue of monitoring internal forces to reveal signal propagation, even if only minor conformational distortions, such as helix bending, initiate the large functional rearrangement of the macromolecule.
| 3-Phosphoglycerate kinase (PGK) is an essential enzyme for living organisms. It catalyzes the phospho-transfer reaction between two catabolites during carbohydrate metabolism. In addition to this physiological role, human PGK has been shown to phosphorylate L-nucleoside analogues, potential drugs against viral infection and cancer. PGK is a two domain enzyme, with the two substrates bound to the two separate domains. In order to perform its function the enzyme has to undergo a large conformational change involving a hinge bending to bring the substrates into close proximity. The allosteric pathway from the open non-reactive state of PGK to the closed reactive state as triggered by substrate binding has only been partially uncovered by experimental studies. Here we describe a complete allosteric pathway, which connects the substrate binding sites to the interdomain hinge region using Molecular Dynamics simulations combined with Force Distribution Analysis (FDA). While previously identified key residues involved in PGK domain closure are part of this pathway, we here fill the numerous gaps in the pathway by identifying newly uncovered residues and interesting candidates for future mutational studies.
| 3-Phosphoglycerate kinase (PGK) is a key enzyme in glycolysis that catalyzes phospho-transfer from 1,3-bisphosphoglycerate (BPG) to ADP producing 3-phosphoglycerate (PG) and ATP [1]. It has been shown that human PGK (hPGK) also phosphorylates L-nucleoside analogues, which are potential antiviral and anticancer drugs [2]–[6]. PGK is a monomeric enzyme composed of two domains of approximately equal size with the C-terminus of the protein bending back to the N-terminal domain, constituting an integral part of it. BPG binds to the N-terminal domain while the ADP binding site is located on the C-terminal domain (for the nomenclature of the secondary structure elements see Suppl. Table S1.)
Crystal structures of PGK from numerous species have shown the enzyme in two distinct conformations: the open conformation 7–12, where the substrate binding sites are too far from each other (12–15 Å) for the phosphoryl transfer reaction to occur (Figure 1), and the closed conformation [13]–[15], where the substrates are proximal enough to allow nucleophilic attack. Thus, these end states are experimental evidence for a hinge bending motion of the enzyme that brings the substrates of this bimolecular reaction together. From combined crystallographic data and small angle X-ray scattering, it has been hypothesized that a spring loaded trap and release mechanism regulates the opening and closing of the domains [12]. By normal mode analysis of the open structure of PGK, Guilbert et al. [16] described three types of interdomain motions: hinge bending, twisting and a shear motion. Our previous Molecular Dynamics (MD) simulations showed that both the apo and the complex enzyme exhibit a small amplitude hinge bending type of motion on nanosecond scale, with the substrate binding changing the character of the motion and restraining the hinge bending [17]. These data put forward the notion of PGK being able to exhibit fast small amplitude hinge bending motions even in the apo state. However, full closure of the enzyme upon substrate binding to adopt the active conformation requires a large and directional hinge bending motion. This raises the question of how the signal of substrate binding penetrates to the interdomain region, where the conformational change happens, leading to closing/opening of the enzyme. Szabo et al. [18] in their combined mutational and kinetic experimental work identified amino acids that play a role in the communication between the substrate binding sites and the hinge region of the enzyme, in particular reported Arg38, Lys219, Asn336 and Glu343 as residues essential in domain closure.
A coherent signaling pathway from the binding sites to the hinge region of hPGK has, however, remained elusive. Computational methods that have been developed to describe allosteric communication are mostly based on coordinates, among others by following the correlated motions as observed in Molecular Dynamics (MD) trajectories [19]–[22] or Elastic Network models [23], [24]. However, typical timescales of allosteric transitions are in order of millisecond to seconds, while atomic simulations cover the femtosecond to microsecond range. Also, these methods focus on large-amplitude motions, while allosteric signal propagation is likely to involve pathways through the rather stiff protein core. Force Distribution Analysis (FDA) recently developed in our group [25], [26] is based on inter-atomic forces instead of coordinate changes and has proven useful in revealing the intramolecular communication pathways of the allosteric proteins MetJ [26] and Hsp90 [27]. FDA monitors the change in pairwise atomic forces within the protein structure upon ligand binding or other perturbations. The advantage of this force-based method over coordinate-based approaches such as Elastic Network Models [28]–[30] or Principle Component Analysis of MD trajectories is three-fold. First, forces within a protein upon perturbation equilibrate relatively fast as compared to the conformational change they trigger [31]. Secondly, pairwise forces are based on internal coordinates and thus do not require any fitting as opposed to coordinate-based observables such as atomic fluctuations. Thirdly, secondary structure elements such as helices or beta-sheets in the protein core are relatively rigid as compared to outer loops of a domain. Consequently, they typically feature small fluctuations hidden in an analysis of (correlated) motions or normal modes, and yet can propagate large forces upon slight hinging or twisting [26], [27]. We here show a connected allosteric pathway of hPGK originating from the two substrate binding sites and extending into the interdomain region involved in hinge bending. Our force distribution analysis identified two primary hinge points in a helix and a loop of the interdomain, the straightening of which causes closing and activation of hPGK. Residues previously identified to be crucial for the regulation of hPGK [18] are part of our computed force network, which also comprises other interesting candidates for point mutations to test our allosteric mechanism.
To characterize the strength of substrate binding to the hPGK complex, residue-wise forces, , between the substrates and all protein amino acids were calculated by summing up forces Fij for all pairs of atoms i and j in residues u and v, where atoms i∈u, j∈v. The results presented in Figure 2 identify amino acids subjected to high forces upon ADP or BPG binding, which are in a very good agreement with the experimentally found binding residues [9], [12], [32], [33]. By comparing the absolute force magnitudes for the substrates, we observe that BPG exerts stronger forces on its binding residues. According to experimental binding measurements, BPG also shows a higher binding affinity for hPGK as compared to ADP [34]. Thus, here, the higher BPG-hPGK interaction forces, which are mostly attractive and of electrostatic nature, reflect steeper binding potentials, i.e. stronger binding. Indeed, we obtained a potential energy between BPG and hPGK of −835±3 kJ/mol, as opposed to −306±50 kJ/mol between MgADP and the protein from our MD simulations (Note that this is an only qualitative comparison as solvent and entropic contributions are missing in this rough estimation for a direct comparison to experiment).
The difference in atomic pairwise forces, ΔFij between the apo hPGK and the complex was calculated to analyze for the effect of perturbation caused by substrate binding. Summing up pairwise force changes sensed by an atom (), the obtained atomic punctual stresses, ΔFi, were mapped on the three dimensional structure of hPGK (Figure 3A), with blue showing minimal and red maximal punctual stresses. By examining the nature of these forces we found that electrostatic interactions are playing the dominant role (see Suppl. Figure S1).
We observed an overall asymmetric pattern of stress in the two domains of hPGK. The BPG binding site shows greater punctual stresses than the ADP binding site, which can be interpreted as a consequence of the tighter binding of BPG compared to that of ADP (see above). Regions distant from the binding sites remain unperturbated by substrate binding. An exception is the interdomain region – in particular Glu192 in helix 7, Ser398/Glu400 in helix 14, and Gly394 in loop L14 – which exhibits a high degree of stress.
We initiated the simulations of both the apo and ligand-bound state from the same experimental structure (complex, 2XE7), and observed the spontaneous opening of the two domains in the apo state, as expected, validating our MD setup. To study whether the environment of residues with high stress feature significant conformational changes by acting as a hinge point, angles formed by Cα atoms were calculated in the two interdomain regions (helix 7 and its adjacent loops F7 and 7G; helix 14 and its adjacent loops L14 and 14–15) for the apo and the complex systems (see Figure 3A right and bottom inserts). Table 1 shows the angle values averaged for the 9 MD trajectories and their differences between apo hPGK and complex. Upon substrate binding, loop L14 and helix 7 exhibit significant changes in their bending angles, namely −23 and −7.2 degrees, correspondingly, which indicates a flattening of these secondary structural elements. The other neighboring structural elements show only minor changes.
The superposition of the open structure to the known catalytically active closed hPGK X-ray structures (PDB entry 2WZB [15] or 2WZC [15] or 2X13) corroborates our results: in all cases, the experimental structures feature a similar flattening of these secondary structural elements in the closed form with respect to the open form. As an example, Figure 3B shows the superposition of the open structure and 2WZB. This similarity also verifies the capacity of our simulations to capture those conformational rearrangements of the interdomain region, which facilitate the domain closure.
It can be seen in Figure 3B that by the “inverse bending” of helix 7, the C-domain moves towards the N-domain. Similarly, flattening of loop L14 provides space for the C-domain to get closer to the N-domain. According to Figure 3B, hinges of these movements are located at about residues Glu192 and Gly394, which are identical with the high-stress “hot spots” determined by FDA. The results of our angle analysis point out that these high stresses are accompanied by extensive conformational changes even in rigid secondary elements such as helix 7. Based on the available PGK sequences found in the ExPASy Molecular Biology Server [35], both residues are highly conserved, which underlines their likely role in domain allostery. In previous studies, Szabo et al. [36] showed that the side-chain of Glu192 is involved in the hinge motion, since its mutation led to substantial decrease of the catalytic efficiency. Furthermore, their calorimetric and thiol-reactivity studies showed that Glu192 is a key residue in maintaining the structural stability of the whole PGK molecule.
Instead of Glu192, previous studies – based on visual comparison of open and closed structures from different species [13], [37] or on MD simulations of these systems [17] – identified the N- and C-terminal loop of helix 7 as hinge points. The other hinge point, Gly394 loop L14, however, has been successfully identified [37], by visual inspection of the two X-ray structures of different species. Here FDA allowed to reveal the allosteric hot spot in the middle of helix 7, which was overlooked in previous studies based on conformational analyses of high amplitude motions.
The FDA results, however are coherent with the DynDom comparisons [39] of hPGK crystallized in open and closed conformations where the following hinge regions were suggested: Tyr195-Leu200 (helix 7), Leu211-Asp228 (helix 8) and Ser392-Gly394 (sheet βL-loopL14).
Based on crystal structures of the complex and solution small angle x-ray scattering data Zerrad et al. [12] suggested a “spring loaded mechanism” for the domain movement of hPGK that is driven by the hydrophobic residues of the interdomain region. The open conformation is stabilized and thus favored by the burial of these hydrophobic residues, while a closing of the domains entails their exposure, counteracted by a stabilization through ionic interactions to allow catalysis. We can extend this scenario by the additional role of the hydrogen bond network in the hinge region. Figure 4 shows both the hydrophobic amino acids (in red) of the interdomain region and the hydrogen bonds (enlisted also in Suppl. Table S2) formed by the hinge points for apo and complex forms of the enzyme. Upon complexation, the hydrogen bond network weakens between Glu192 and adjacent residues (Ser392, Thr393, Gly394 and Ser398) and is only partially substituted (Ser392-Gly394 and Gly 394-Ser398), which we hypothesize to cause the subsequent extensive flattening of helix 7. This suggests the hydrophobic area exposure and the weakened H-bond network of the interdomain residues to work in concert, i.e. the burial of hydrophobic residues in the apo form and the stronger hydrogen bonds jointly “pull” the hinge of helix 7 towards the interdomain region and curve the helix. As our analysis cannot identify causalities of the observed effects, similarly, the helix hinging could reversely effect substrate binding sites and substrate or product release.
The distance between the substrates and the identified hinges poses the question of the pathway of the long-range communication between these two regions, i.e. the signaling pathways through which stress propagates from binding residues towards the functionally important interdomain region.
Figure 5 displays the largest connected network of pronounced atomic pairwise force changes. Edges connect Cα atoms of two residues with an inter-atomic force difference |ΔFij|>90 pN. While all of the BPG binding residues are involved in the signal transmission, only two of the ADP binding residues transmit the perturbation. Since BPG exerts higher forces on its binding residues, these residues are capable of propagating the perturbation in different directions throughout the structure of the protein. While the weaker binding of ADP to the protein results in a low number of binding residues being involved in signal transmission. Figure 5 indicates that the two hinge points, Glu192 and Gly394, are included in the network, suggesting the remote binding of the substrates to perturb the force distribution at the hinges, a clear evidence of allostery. The hinges are perturbed via several force distribution pathways stemming from binding residues of both substrates, indicating that hinges detect the effects of both substrates. Furthermore, the hinges act in concert, since they are directly connected in the path. The indirect communication between the two substrates of hPGK through this stress transmission path, jointly with a direct electrostatic repulsion between the substrates of 92±38 pN (between MgADP and BPG, obtained from FDA), can explain the cross-talk of the two molecules in terms of their affinity for hPGK [6] and in terms of their joint action to trigger the hPGK conformational change [38], [40].
Tracking down the transmission path of BPG, we can note that the BPG binding residues form a tight interaction network around the substrate and are further passing down the force signal to Arg38 and Arg170. From Arg38, stress propagates through Thr393 to both hinges, Glu192 and Gly394. Thus, the signaling pathway of the BPG effect towards the hinges is: BPG→BPG binding residues (Arg122, Arg65, Arg170, His62, Asp23, Asn25, Arg38)→Thr393→hinge residues (Glu192, Gly394). Similarly, tracking down the stress propagation path due to ADP binding, we can note that only two binding residues, Asn336 and Thr375 transmit the effect of ADP. From Asn336 the stress propagates through Gly371, Ser398 and Ser392 to reach the two hinges, Glu192 and Gly394. The stress originating in Thr375 is transmitted by Gly372 (which in our previous study, based on conformational analysis, was identified as a hinge point) and then follows the same path. Thus, the signaling pathway of the ADP effect towards the hinges is: ADP→ADP binding residues (Asn336, Thr375)→Gly371, Gly372→Ser398, Ser392→hinge residues (Glu192, Gly394).
Previous structural and dynamical analyses of PGK allostery have allowed unprecedented insight into the large conformational changes required for phosphoryl transfer catalysis. We here could reconcile and complete the picture of a signal transduction pathway from the ADP and BPG binding sites to the interdomain hinges Glu192 and Gly394, using our novel force distribution analysis. Our allosteric pathway overlaps with a previously identified transmission path, which however was restricted to a single hinge in sheet L [18]. While previously identified key residues involved in PGK domain closure [1], [18], [34], [36] are part of this pathway, underlining their functional relevance, residues Gly371, Gly372 and Gly394 are residues newly uncovered as critical allosteric spots. We hypothesize that an addition of a sidechain to these residues would abolish or measurably alter hPGK allostery, and thus suggest them as interesting candidates for future mutational studies.
MD simulations were carried out on apo hPGK and on its natural ternary complex: ADP*BPG*Mg*hPGK (complex). The starting structures were derived from the crystal structure of hPGK (PDB entry 2XE7) complexed with ADP and 3-phospho-glycerate (PG). The apo structure was built by removing ADP and PG from the 2XE7 structure. Starting all simulations from the 2XE7 structure instead of two different experimental structures for complex and apo states allowed to validate our MD setup by tracking the unbiased relaxation from the complex to the more open apo state and to measure force differences (see below) between the two states only due to this allosteric conformational change and not due to any potential additional differences in starting structures because of crystallization conditions or crystal packing. For the ternary complex, an extra phosphate group and a Mg-ion were placed into the 2XE7 structure based on the MgADP bound structure (PDB entry 1PHP) [32] of B. Stearothermophilus PGK using the Schrödinger-Maestro program [41]. The extra phosphate group was covalently attached to the 1-carboxyl group of PG and coordinated by Arg38 and a water molecule via non-bonded interactions.
The initial three N-terminal residues missing in the crystal structure were modeled using the Schrödinger-Maestro program. Coordinates for the N-terminal loop and loops showing extraordinarily high B-factors were optimized by the Modloop web server, version r181M [42], [43]. The water molecules present in the crystal structure were retained around the binding sites.
MD simulations were performed with GROMACS 4.5.1 [44] using the CHARMM all-atom parameter set 27 [45]. Both systems, apo hPGK and the ternary complex, were immersed in a rhombic dodecahedron box of TIP3P water [46] with vector length 100 Å with a distance of 12 Å between the protein surface and the box face. The boxes were replicated by periodic boundary conditions. Sodium and chloride ions corresponding to a physiological ion strength of 120 mM were added. Additional chloride or sodium counterions were added to achieve a neutral net charge of the apo and complex systems, respectively.
The real space summation of electrostatic interactions was truncated at 12 Å, and the Particle Mesh Ewald (PME) [47] method was used to calculate the electrostatic interactions beyond 12 Å with a maximum grid spacing of 1 Å and an order of 6. The width of Gaussian distribution was set to 0.34 Å−1. Van der Waals interactions were calculated using a cut-off of 12 Å.
The solvated systems were energy minimized by the following procedure: the steepest descent algorithm was used first with harmonic constraints applied to heavy protein atoms to achieve smooth minimization. The harmonic force constant was decreased every 100 steps, adopting the values 40 000, 10 000, 1 000 and 100 kJmol−1 nm−2. Then unconstrained minimization was applied for 100 steps with steepest descent, followed by 1000 steps of conjugate gradient algorithm.
The energy minimization was followed by a 5 ns MD simulation with harmonic constraints on protein heavy atoms with a force constant of 1000 kJmol−1 nm−2 to equilibrate water and ions around the protein. An unconstrained MD simulation of 3 ns length was performed to equilibrate the whole system.
For each system, 9 independent 50 ns production MD simulations were performed totaling 900 ns of simulation time. An integration time step of 2 fs was used. The coordinates of the trajectories were saved every 50 ps. For each simulation new random velocities were generated and different initial frames were used to enhance conformational sampling. Starting frames were chosen from the second half of the first MD simulation for subsequent simulations. Simulations were carried out in the NPT ensemble. The temperature was kept constant at 300 K using the Nosé-Hoover thermostat [48] with a time constant of 0.4 ps. The pressure was set to 1 bar using isotropic coupling to the Parrinello-Rahman barostat [49] with a time constant of 1 ps and a compressibility of 4.5*10−5 bar−1. All bonds were constrained using the LINCS [50] algorithm. The stability of the simulation is shown by Suppl. Figure S2.
We used the FDA extension [51] for GROMACS 4.5.3 to write out forces, Fij, between each atom pair i and j as calculated during our MD simulations. Forces were computed between all atom pairs within the cut off range for each frame of the MD simulation, and included all interaction types (bonds, angles, dihedrals, electrostatic and Lennard Jones interactions). Residue-wise forces, , were introduced by summing up forces Fij for all pairs of residues u and v, where atoms i∈u, j∈v, which in our case were used only for the characterization of the binding pocket.
Since pair-wise force vectors are affected by rotation and translation of the system, for subsequent analysis, the norm of the force was used, with opposite signs assigned to attractive and repulsive forces [26]. Forces for each system were averaged over all nine equilibrium trajectories, each 50 ns in length, to achieve converged averages. Differences between averaged pairwise forces (ΔFij) of the apo and complex form of hPGK were then calculated to describe the perturbation of substrate binding. While we observed positive and negative pair-wise forces to similar extent, for both, Fij and ΔFij, implying a balance of repulsion and attraction within the protein at equilibrium, we could not infer any relevance of the sign of the forces for the allosteric mechanism, so that we restricted the analysis presented in the Results section to the absolute values only for the sake of clarity. Atomic punctual stress was defined as the absolute sum of force differences sensed by a single atom: . Here, absolute forces are summed up, because the sum of positive and negative ΔFij values is by definition zero over time, i.e. are balanced, resulting in the absence of any net translation of the system during the MD simulation time. The atomic punctual stress has been defined previously [51] and here serves as an easily accessible measure of the perturbation upon ligand binding. It is similar but not equal to the virial stress and other more complex definitions of local stresses. [52], [53]. The convergence of force are shown on Suppl. Figure S3. For the identification of pathways of stress propagation, the largest network of connected atomic pairwise force differences between apo and complex hPGK beyond a cut off value was determined using a vertex count algorithm [26]. The cut off for force differences was chosen such that a consecutive path originating from the substrate binding sites was obtained [27]. A larger cut off than 90 pN resulted in only small vertex counts, i.e. a broken path, while a smaller cutoff resulted in additional networks at distant unconnected sites, indicative of noise in the forces.
|
10.1371/journal.pbio.0060002 | A Mechanism Regulating the Onset of Sox2 Expression in the Embryonic Neural Plate | In vertebrate embryos, the earliest definitive marker for the neural plate, which will give rise to the entire central nervous system, is the transcription factor Sox2. Although some of the extracellular signals that regulate neural plate fate have been identified, we know very little about the mechanisms controlling Sox2 expression and thus neural plate identity. Here, we use electroporation for gain- and loss-of-function in the chick embryo, in combination with bimolecular fluorescence complementation, two-hybrid screens, chromatin immunoprecipitation, and reporter assays to study protein interactions that regulate expression of N2, the earliest enhancer of Sox2 to be activated and which directs expression to the largest part of the neural plate. We show that interactions between three coiled-coil domain proteins (ERNI, Geminin, and BERT), the heterochromatin proteins HP1α and HP1γ acting as repressors, and the chromatin-remodeling enzyme Brm acting as activator control the N2 enhancer. We propose that this mechanism regulates the timing of Sox2 expression as part of the process of establishing neural plate identity.
| During early development, when the embryo has three layers of cells (ectoderm, mesoderm, and endoderm), a region of the ectoderm called the neural plate becomes specified to generate the entire nervous system. One of the earliest molecular markers for the neural plate is the transcription factor Sox2, which is critical for cells to acquire their neural fates and also defines neural progenitor character. We know very little about the intracellular mechanisms by which the neural plate cells acquire these fates. Here, we show that recruitment of transcriptional repressors to chromatin-remodeling complexes regulate the onset of Sox2 expression. Competitive interactions between three proteins, ERNI, BERT, and Geminin, modulate the choice of repressors and regulate Sox2 expression. During gastrulation, when the three embryonic cell layers form, ERNI recruits the repressor HP1γ to prevent Geminin from activating Sox2 prematurely. By the end of gastrulation, this repression is counteracted by competitive binding of BERT to ERNI and Geminin, causing activation of Sox2. We propose that this mechanism regulates the timing of Sox2 activation in the very early neural plate and thus helps to define the domain that will give rise to the nervous system.
| Sox2 is a transcription factor that plays multiple critical roles during embryonic development in vertebrates. In embryonic stem (ES) cells, as well as in adult central nervous system (CNS) stem cells, Sox2 expression is required for the maintenance of multipotency and for the ability of cells to self-renew [1]. Sox2 is also expressed in cells that retain their ability to proliferate and/or acquire glial fates, whereas it is down-regulated in cells that become postmitotic and differentiate into neurons [2–4]. In addition, it is also transiently expressed outside the CNS in cranial sensory organs derived from the placodes and in subsets of peripheral nervous system (PNS) cells [5,6].
In all vertebrates studied to date, Sox2 is also a general marker for the very early developing neural plate. In the chick, for example, Sox2 expression starts at the late primitive streak stage (stages 4–4+ [7]) in the future neural territory [8,9]. A morphologically recognizable neural plate only becomes visible after the beginning of Sox2 expression [8]. Importantly, Sox2 function is required for development of the neural plate [10]. Time-course experiments have shown that induction of Sox2 requires the same period of exposure to organizer-derived signals (the tissue responsible for inducing the neural plate in the normal embryo [11–13]) as is required to induce a mature neural plate [14–17]. For these reasons, Sox2 is considered to be the earliest definitive marker for the neural plate [18,19].
The complex expression profile of Sox2 is controlled by multiple regulatory elements, each responsible for directing expression to a specific subset of expression sites. A very compelling analysis of the noncoding regions of Sox2 in the chick embryo [20] revealed as many as 25 distinct conserved enhancers, of which two account for the expression of this gene in the early neural plate at stages 4+–5. One of these enhancers, named N2, is responsible for the initial expression (stage 4–4+) and is activated in a large domain corresponding to the entire forebrain/midbrain and most of the hindbrain. The other, N1, drives expression in the future caudal hindbrain and spinal cord and is activated a little later (around stage 5) [20,21]. To understand the processes that define the neural plate, it is essential to understand how the activity of these two elements, and especially N2, is regulated in the embryo.
Analysis of the N2 enhancer reveals multiple putative binding sites for known transcription factors [20,21]. However, the spatial and temporal expression patterns of these factors do not provide an obvious explanation for the time of onset of Sox2 expression in normal development (unpublished data). Furthermore, to date, no single secreted factor or any combination thereof has been found to induce either Sox2 expression or a neural plate in competent cells not normally fated to form part of the neural plate [13,19]. We therefore directed our attention to nuclear factors that might regulate this enhancer. Here, we provide evidence that a group of coiled-coil proteins interact with each other and with chromatin-remodeling factors and heterochromatin proteins to regulate the activity of the N2 enhancer. We propose that this is part of a mechanism that regulates the time of onset of expression of Sox2 in the nascent neural plate.
A recent study [22] using the P19 cell line demonstrated that the chromatin-remodeling enzyme Brahma (Brm) can activate Sox2 by binding directly to the N2 enhancer. Is Brahma also involved in regulating Sox2 expression in the normal embryo? To test this, we introduced a mutated version of Brahma (BrmK755R, which does not bind ATP and is therefore unable to remodel chromatin [23]) by electroporation into the prospective neural plate of embryos at stage 3–3+. This resulted in strong inhibition of Sox2 expression in the electroporated domain (Figure 1A and 1B; 5/6), unlike controls electroporated with green fluorescent protein (GFP) (Figure 1C and 1D; 0/5 expressing).
However, Brm is expressed ubiquitously in the embryo [24]; what mechanisms prevent premature expression of Sox2? A good candidate is the transcriptional repressor HP1α, which binds directly to Brahma-related proteins at a highly conserved site [25] and which is also ubiquitously expressed in early embryos (Figure 2). Consistent with this, overexpression of HP1α in the neural plate represses Sox2 (Figure 1E and 1F; 3/3). Could HP1α be an endogenous inhibitor of Sox2 expression? To address this, we took advantage of the fact that both the chromoshadow domain and the chromodomain are necessary for the function of HP1 proteins [26,27]: targeting to chromatin requires interaction of the chromoshadow domain with a chromatin-tethered partner, as well as binding of the chromodomain to a methylated Lys9 of histone H3 [28]. We therefore made a dominant-negative form of HP1α (ΔHP1α) consisting of its isolated chromoshadow domain (which can bind to Brahma-related proteins but lacks repressor activity [25]). When ΔHP1α is misexpressed as a line extending from the neural plate into the peripheral, nonneural ectoderm (see Materials and Methods, “Design of assays”), Sox2 is induced (Figure 1G and 1H; 6/7), whereas similar electroporation of GFP has no effect (Figure 1I and 1J; 0/8). This suggests that HP1α activity is required to prevent expression of Sox2 in the nonneural ectoderm.
In embryos in which ΔHP1α was expressed as a line, we observed that Sox2 was up-regulated, not only in the embryonic nonneural ectoderm (prospective epidermis), but also in the more peripheral area opaca epiblast (extraembryonic ectoderm) (Figure 1G). We were surprised by this observation because until now, various factors (such as bone morphogenetic protein [BMP] antagonists [16,17,19]) have been described that can expand the neural plate domain, but never as far as the extraembryonic epiblast, and none can induce a separate domain of Sox2 expression in this region. The only treatment described to date that can induce neural markers in the area opaca is a graft of the organizer, Hensen's node, which is able to generate a complete, patterned nervous system in this region [13,29–32]. These observations define the area opaca as a particularly rigorous location in which to test for the neural inducing ability of factors (see Materials and Methods, “Design of assays”). Electroporation of ΔHP1α in this region dramatically induces Sox2, showing that HP1α normally represses Sox2 expression (Figure 1K and 1L; 8/8). In contrast, electroporation of GFP in the same region has no effect (Figure 1M and 1N; 0/10). As an additional control, since HP1α may have more general activity as a transcriptional repressor, we also tested whether ΔHP1α can also induce other early embryonic genes using Brachyury, a marker for mesoderm expressed at this stage of development. It did not (Figure 1O and 1P; 0/3). This result also confirms that the induction of Sox2 is direct, rather than a consequence of prior induction of mesoderm by ΔHP1α. Likewise, electroporation of BrmK755R or Brm had no effect on neural or mesoderm markers (0/5 for each; unpublished data). To test whether the inducing activity of ΔHP1α requires Brahma, we introduced ΔHP1α together with BrmK755R. This combination fails to induce Sox2 (Figure 1Q and 1R; 0/12), suggesting that HP1α normally inhibits Sox2 expression through a Brm-dependent mechanism (Figure 1S).
In Xenopus, the gene encoding the coiled-coil protein Geminin is expressed in the early prospective neural plate, and its misexpression induces neural markers [33]. More recently, it has been shown that Geminin interacts genetically with Drosophila Brahma, that it binds directly to its vertebrate homologs Brg1 and Brm (at the same site as does HP1α), and that Geminin knock-down abolishes Sox2 expression [25,34]. Could Geminin be responsible for releasing the repression of Brm activity by HP1α? To test this, we cloned the chick homolog of Geminin. Before and during early gastrulation, Geminin is expressed in a large domain, which then (from stages 4–4+) becomes restricted to the neural plate (Figure 3). The early expression of chick Geminin resembles that of “pre-neural” genes (such as Sox3, ERNI, and Churchill), which precede the initiation of Sox2 expression and which are induced by fibroblast growth factor 8 (FGF8) [8,9,35,36]. We therefore tested whether FGF can also induce Geminin. Indeed, Geminin can be induced by FGF8-soaked beads (9/10; Figure 3H, arrow), but not by control beads (0/10; Figure 3H, arrowhead).
When misexpressed as a line extending laterally from the neural plate, Geminin strongly induces ectopic Sox2 (Figure 4A and 4B; 8/8). Geminin can also strongly induce Sox2 expression when introduced into the extraembryonic epiblast (Figure 4C and 4D; 20/20). To test whether this induction requires the chromatin-remodeling activity of Brm, we cointroduced Geminin and BrmK755R: the mutated chromatin remodeler abolishes the induction of Sox2 by Geminin (Figure 4E and 4F; 0/15), suggesting that Brm activity is required for Sox2 induction by Geminin. Together, these results suggest that early in development Sox2 expression is constitutively repressed by the presence of HP1α bound to Brm at the N2 enhancer of Sox2, and that later in development, FGF may release this inhibition through induction of Geminin, which competes HP1α away from the protein complex (Figure 4G).
Geminin is already expressed at the beginning of gastrulation (Figure 3), long before Sox2 (which appears at stage 4 [8,9]), suggesting that an additional mechanism must exist to prevent premature Sox2 activation. A good candidate for this repression is ERNI, which is broadly expressed in the epiblast at early stages but is rapidly down-regulated from the prospective neural plate at stage 4+ [36], around the time when Sox2 starts to be expressed. To test whether ERNI can inhibit the induction of Sox2 by Geminin, we cointroduced them into the area opaca: ERNI does indeed inhibit the induction of Sox2 by Geminin (6/14 with very weak induction, 8/14 with no induction; Figure 5A–5C).
The above findings are consistent with the idea that ERNI normally functions to repress Sox2 expression at very early stages of development. However, it is unlikely that down-regulation of ERNI transcription is sufficient to relieve this inhibition because Sox2 expression begins at stages 4–4+ [7], when some ERNI transcripts can still be detected within the prospective neural plate [36]. Therefore, an endogenous inhibitor is likely to exist whose expression should begin at around this time (stage 4–4+). To identify such an inhibitor, a two-hybrid screen was performed using ERNI as bait and a library of cDNAs from stage 3–6 chick embryos (Figure 9A). Only one partner was found, encoding a small coiled-coil domain protein which we named BERT (Figure 9B). An equivalent human protein (SCOCO, corresponding to the fragment shown in bold in Figure 9B) was previously isolated as a partner of human ARL1, a component of the Golgi apparatus [44], and a nematode homolog (unc69) was found to be essential for neural development [45]. BERT is expressed ubiquitously at low levels at all stages, but is up-regulated specifically in the prospective neural plate from stage 4–4+ (Figure 10), just prior to when Sox2 expression appears [8,9]. When misexpressed as a line across the nonneural epiblast, BERT acts like the dominant-negative ERNI constructs: it induces strong expression of Sox2 (Figure 9C and 9D; 15/15), which also acquires a neural plate-like morphology (Figure 9D′: note that the thickened ectoderm characteristic of the neural plate has greatly expanded on the electroporated side; see arrowhead on right). Mesodermal markers (Brachyury, Chordin, and BMP4; 0/5, 0/4, 0/4, respectively; unpublished data) are not induced, showing that this expansion of the neural plate is a direct effect. These findings indicate that BERT has an activity compatible with it being an endogenous antagonist of ERNI, regulating not only Sox2 expression, but also the onset of neural plate development. To confirm this, we examined the effects of coexpressing BERT with Geminin+ERNI (which does not induce Sox2; see above and Figure 5A and 5B) in the area opaca. Indeed, when all three constructs are cointroduced, induction of Sox2 is seen (12/12; Figure 9E and 9F).
Is BERT required to control the onset of Sox2 expression in the neural plate? To address this, we designed a fluorescein-labeled Morpholino oligonucleotide (MO) to the 5′ end of the coding sequence (see Materials and Methods) and introduced this (together with GFP) by electroporation into the prospective neural plate at stage 3–3+ and examined Sox2 expression at stages 4+–5. BERT-MO caused down-regulation of Sox2 expression in this domain (Figure 9G and 9H; 5/6), unlike control MO (Figure 9I and 9J; 0/5). Staining of BERT-MO–electroporated embryos with an antibody against BERT/SCOCO (see Materials and Methods) confirmed that the MO does indeed inhibit translation of BERT protein (Figure 9K) in the electroporated domain (Figure 9L). Together, these findings implicate BERT as an endogenous antagonist of ERNI, required to regulate the onset of Sox2 expression in the neural plate.
From the results presented above, the evidence that BERT binds to ERNI directly is based entirely on the two-hybrid screen used to isolate BERT. To confirm that the two proteins can interact physically, we used BiFCo assays, which further revealed that BERT, Geminin, and ERNI all bind to each other through their coiled-coil domains (Figure 11 and Table 1). This finding raises the possibility that BERT disrupts Geminin-ERNI dimers by binding to both proteins, thus removing ERNI-HP1γ from the complex to activate Sox2 (Figure 9M). To test this further, we used BiFCo competition assays (Figure 12). When BERT is added to Geminin-Venus(N)+ERNI-Venus(C), fluorescence is lost (Figure 12A and 12B). When Dlx5 is used as a control instead of BERT in this assay, there is no effect (Figure 12C). Conversely, when BERT-Venus(C) is added to Geminin-Venus(N)+ERNI, fluorescence is generated (Figure 12D and 12E), and this is not mimicked by addition of Dlx5-Venus(C) (Figure 12F). Likewise, when BERT-Venus(C) is added to Geminin+ERNI-Venus(N), fluorescence is produced (Figure 12G and 12H), which is not mimicked by the use of Dlx5-Venus(C) instead of BERT-Venus(C) (Figure 12I). Together, these findings support the idea that BERT can disrupt Geminin-ERNI heterodimers by binding to both proteins.
The experiments described above tested the protein–protein interactions and their effects on Sox2 expression, but their physical association with the N2 enhancer was extrapolated from published results in a cultured cell line, unrelated to the early neural plate [22]. To test whether these interactions can regulate Sox2 expression directly at the N2 enhancer, we coelectroporated a reporter construct consisting of the N2 enhancer and a minimal TK promoter [20] driving expression of LacZ together with either Geminin alone, with Geminin+ERNI, or with Geminin+ERNI+BERT, into the extraembryonic epiblast. No expression of the reporter was seen when it was coelectroporated with the control construct pCAβ-GFP (Figure 13A and 13B; 0/16) or with Geminin+ERNI (unpublished data; 0/6). However, expression was induced by both Geminin (unpublished data; 6/6) and by Geminin+ERNI+BERT (Figure 13C and 13D; 5/5). This shows that ERNI can block the activity of the N2 enhancer of Sox2 and that BERT inhibits this.
Finally, to confirm that these proteins do indeed interact physically with the N2 enhancer, we conducted chromatin immunoprecipitation (ChIP) assays using chromatin extracted from embryonic day (E)7.5 mouse embryos and an antibody against mouse Geminin. The antibody specifically precipitates the N2 enhancer of Sox2 (Figure 13E, lane 3), unlike control experiments performed either without chromatin (Figure 13E, lane 1) or without anti-Geminin antibody (Figure 13E, lane 2). These findings demonstrate that Geminin does indeed associate physically with the N2 enhancer of Sox2 in vivo at an appropriate stage in development.
Sox2 is an important gene that plays multiple roles in development especially in controlling cell fate and proliferation. Its expression pattern is complex and regulated by multiple noncoding elements [20,21]. In the normal embryo, one of the earliest conserved sites of expression is the nascent neural plate, where Sox2 constitutes the earliest definitive marker for this tissue. It is therefore of particular interest to understand the mechanisms that regulate the location and timing of expression of this gene in the neural plate, as this process is critical for normal nervous system development. Here, we propose that interactions between several coiled-coil proteins, heterochromatin proteins, and chromatin-remodeling molecules regulate the time of onset of Sox2 expression in the chick neural plate.
The most parsimonious model to explain our findings in terms of how Sox2 expression is regulated in the early neural plate comprises the four steps shown in Figures 1S, 4G, 7E, and 9M. Since Brm and HP1α are expressed ubiquitously ([24] and results from the present study), we propose that there is a basal state in which Brm is bound to the N2 enhancer of Sox2 [22], but the latter is not expressed because the repressor HP1α occupies the chromoshadow-binding domain of Brm (Figure 1S). Early in development, FGF activity induces both ERNI [18,36] and Geminin (this study) in the epiblast. Geminin binds to the chromoshadow-binding domain of Brm, displacing HP1α (Figure 4G). However, the interaction of ERNI with Geminin recruits the transcriptional repressor HP1γ, thus continuing to prevent premature expression of Sox2 in the epiblast (Figure 7E). Later in development (stage 4–4+), BERT is up-regulated within the neural plate, where it binds to both ERNI and Geminin and displaces ERNI-HP1γ complexes away from Brm, freeing the latter to activate N2 and thus Sox2 expression (Figure 9M). At around the same time, ERNI transcription starts to be down-regulated in the neural plate. This model accommodates all of our results and those in the literature; its significance is explored further in the following sections.
The N2 enhancer of Sox2 is about 550 bp long and is predicted to contain multiple binding sites for transcription factors [20], many of which are expressed in the epiblast prior to the stage at which Sox2 expression is initiated. In principle, binding of the appropriate activators to the N2 enhancer should turn on Sox2. However, the spatial and temporal patterns of expression of these factors do not account for the timing or spatial distribution of Sox2 transcription at this stage in development, as many of them are expressed ubiquitously (unpublished data). We therefore propose that, irrespective of the binding of putative activators to the N2 enhancer, the conformation of chromatin, maintained in a closed configuration by HP1 proteins, prevents activation at early stages. It is only when HP1 proteins are removed and the chromatin-remodeling activity of Brm is released that N2 is activated.
Chromatin-remodeling complexes may turn out to have a widespread role in the transcriptional activation of specific genes, as exemplified by Smad-activated genes whose transcriptional regulation also requires the activity of such complexes [46]. Likewise during skeletal muscle differentiation, chromatin-binding proteins “mark the spot” for activation of genes by other transcription factors together with chromatin remodeling by SWI/SNF proteins: MyoD binding to chromatin is regulated by the homeodomain protein Pbx1 in cooperation with the Brahma-related enzyme Brg1 [47–50]. To our knowledge, however, this is the first report suggesting that a SWI/SNF chromatin-remodeling complex can recruit HP1 proteins to a specific enhancer to repress transcription of a target gene.
Our model proposes mutually inhibitory interactions between several proteins. Why does Sox2 need to be regulated by such a complex mechanism, rather than by merely recruiting a single or a few activators to a simple enhancer? We suggest that this is one in a series of steps that act to separate different functions for signals that are common to different developmental processes.
Previously, we showed that 3–5 h of exposure to signals from the organizer (Hensen's node) is sufficient to induce transient expression of the pre-neural marker Sox3, but not sufficient to induce later neural plate markers (such as Sox2), and that the BMP antagonist Chordin can stabilize the expression of Sox3 induced by such a graft (but again not induce Sox2) [16]. Based on these findings, we conducted a screen to identify genes induced within 5 h of exposure to the organizer [36]. We identified several genes induced within this time, among them ERNI, which is induced very rapidly, within 1–2 h. FGF8 is sufficient to mimic this effect, and during normal development, ERNI is expressed even before gastrulation, in a domain identical to that covered by the underlying hypoblast (which expresses FGF8).
FGF is required for both mesodermal [51–54] and neural induction [36,55,56]. How do cells that have received FGF signals decide between these two incompatible fates? A likely scenario is that cooperation with other factors, present at different times and in different locations, contributes to refine this choice. To allow this to happen, it may be necessary for cells to retain a “memory” that they have received FGF signals yet be prevented from being allocated prematurely to inappropriate fates. ERNI appears to fulfill such a role: while it is expressed, cells are multipotent, as its early domain of expression encompasses the prospective neural and mesendodermal domains as well as some nonneural ectoderm. At the end of gastrulation, ERNI transcription starts to be down-regulated from the future neural plate, remaining only at the border between neural and epidermal domains [36,57]. At the same time, BERT is up-regulated in the domain that is losing ERNI expression while Sox2 starts to be expressed in the same domain (stage 4–4+). This sequence of events could help to explain why it takes such a long time (about 9 h) following a graft of a node for Sox2 expression to begin and for a neural plate to be induced [14–17]. Consistent with the proposal that ERNI is part of a mechanism to prevent premature expression of Sox2, we have observed that transfection of BERT into the prospective neural plate region of stage 2–3 embryos can induce premature expression of Sox2 (unpublished data).
The present and previous studies [36] reveal that FGF signaling activates ERNI as well as Sox3 and Geminin expression in the epiblast. However, FGF does not induce BERT, whose expression is also not regulated by BMP antagonists or any combination of known factors implicated in neural induction to date (unpublished data). In future, it will be interesting to determine whether BERT is induced by some other combination of factors or whether its expression is regulated simply by a cell-autonomous timer in cells that are still in the epiblast at the end of gastrulation, but does not require input from other cells.
In all likelihood, the mechanisms responsible for regulating Sox2 expression and the acquisition of neural fate will turn out to be considerably more complex, and our model does not rule out additional mechanisms. It will be interesting in future to investigate whether other developmentally expressed genes are regulated by similar processes.
Our findings provide a mechanism for how Sox2 expression is initiated as part of the events that define the early neural plate. We propose that ERNI functions as an inhibitor of premature Sox2 expression during early gastrulation: cells expressing ERNI are multipotent and can generate any cell type. Cells that remain in the epiblast at the end of gastrulation and acquire expression of BERT to activate Sox2, which, most likely together with other genes involved in neural specification, assigns a neural plate fate.
Fertile hens' eggs (Brown Bovan Gold; Henry Stewart & Co.) were incubated at 38 °C to the desired stages. Electroporations were performed as described [35]. The coding region of full-length ERNI, ERNI coiled-coil domain (aa 1–164), chick BERT, chick Geminin, human BrmK755R (kind gift from Dr A Imbalzano), mouse HP1α, mouse HP1α chromoshadow domain (aa 106–180), and mouse HP1γ chromoshadow domain (aa 118–176) were cloned into pCAβ and electroporated at 0.2 μg/μl (except ERNI and BrmK755R and HP1α, which were used at 0.4 μg/μl) together with 1 μg/μl of pCAβ-GFP, which was used to mark the electroporated cells. The N2-TK-LacZ reporter plasmid was constructed from N2-TK-GFP, kindly provided by Dr H. Kondoh, and was electroporated at 1 μg/μl. FGF8b (Sigma) was delivered bound to heparin beads (prepared as described [19]) at 50 μg/ml. In situ hybridization and immunostaining for GFP were performed as described [35].
To establish the role of different components in regulating the expression of Sox2, three different types of assays were used for gain- and loss-of-function experiments. First, to assess the effects on endogenous expression of Sox2 in the normal neural plate, constructs were introduced into the prospective neural plate at mid-primitive streak stage (stage 3–3+) and the embryos incubated about 6–9 h so that the embryo had reached stages 4+–7, just beyond the stages at which Sox2 expression begins (4+) and also because at these stages the neural plate is still open, allowing easier visualization of expanded expression. Please note that stages 4+–7 are particularly short, this entire period lasting only about 3 ± 1.5 h at 38 °C. To determine whether a construct can induce ectopic expression of Sox2, two different locations were chosen. In one set of assays, the construct is introduced as a continuous line between the prospective neural plate of the embryo and the inner aspect of the extraembryonic epiblast, covering most of the prospective epidermis. In the other assay, the construct is introduced as a discrete domain within the inner third of the extraembryonic (area opaca) epiblast and the embryos incubated 12–15 h (by which time they have reached stages 6–9).
The reasons for choosing both of the latter two assays for induction is that extensive embryological studies have revealed differences in their reactivity to neural inducing stimuli. For example, inhibition of BMP signaling is sufficient to expand the endogenous neural plate laterally (and BMP misexpression to narrow it), but only when the territory is continuous with the embryo's own neural plate [13,16,17], suggesting that induction of neural markers by certain stimuli in this region requires cellular continuity with the neural plate and/or its border. On the other hand, a graft of the organizer (Hensen's node) is able to induce a complete, patterned ectopic nervous system from the extraembryonic epiblast of the inner area opaca [13,29–32]. A period of 9–13-h contact is required to induce Sox2 after a graft of the organizer, which is why 12–15 h was chosen in this assay. To date, no single factor or any combination thereof has been found to mimic this activity of the organizer. It is therefore particularly important, to assess the full inducing properties of a treatment, to test its ability to induce Sox2 in the area opaca. We therefore used all three assays to compile a more comprehensive understanding of the inducing or inhibiting activities of each of the constructs in this study.
A translation-blocking MO against BERT with the sequence CAGCGTCCATGTCAGCGTTCATCAT, targeting the 5′ end of the ORF of the gene or a standard control MO (Gene Tools LLC), both labeled with fluorescein, were electroporated by injecting a small volume (about 0.1 μl) of a stock of the MO at 1 mM exactly as described for electroporation of constructs (see above). Antibody against human SCOCO was kindly provided by Dr. Richard Kahn. This was used in whole mounts by indirect immunoperoxidase with anti-rabbit-HRP using the same method as described for GFP (see above).
For two-hybrid screens with embryonic cDNA, poly-A RNA was isolated from 600 chick embryos (stage 3–6) using the Ambion Poly(A)Pure Kit. The mRNA was used to synthesize a cDNA library which was cloned into the pMyr vector using the CytoTrap XR Library Construction Kit (Stratagene). The library was transformed into XL10-Gold Ultracompetent Cells (Stratagene). Full-length ERNI was cloned into the pSOS vector and used as bait in the CytoTrap two-hybrid screen, which was performed according to the manufacturer's instructions (Stratagene).
For two-hybrid screens on chick ES cells, poly-A RNA was isolated from ES cells [58]. cDNA was synthesized using Stratagene's cDNA synthesis kit and introduced into pGAD424 vector (Clontech), and this was transformed into XL1-blue MRF' bacteria by electroporation. All plasmids, yeast strains, and media used were purchased from Clontech. The bait ENS-1/ERNI coding sequence was cloned in NdeI/SalI sites of pGBKT7, introduced into AH109 yeast, and checked for lack of self-activation of the reporter. Screening was performed according to the Yeast Protocols Handbook (Clontech). pGAD424 recombinant plasmids from 18 candidates were purified, of which seven encoded the CHCB2 protein [41] and all included the chromoshadow domain. The smallest one, encoding the 87 carboxy-terminal amino acids, was used in further experiments. The full ENS-1/ERNI coding sequence was cloned into pGADT7 and various truncated forms (Figure 5) subcloned into pGBKT7. Point mutations were introduced into pGBKT7:ENS-1 using the QuikChange site-directed mutagenesis kit from Stratagene and checked by sequencing. Yeast two-hybrid assays were performed by rapid cotransformation of strain AH109.
The Xenopus Geminin amino acid sequence was used to BLAST the GenBank EST database. The full-length chick homolog sequence was recovered and cloned by PCR from the CytoTrap cDNA library described above.
The N- and C-terminal halves of Venus (aa 1–154 and 155–229) were PCR-amplified from pCS2 vectors and cloned into pcDNA3.1A. Geminin, ERNI, BERT, and human E2F3 were cloned in frame into the 5′ end of each of the two Venus halves, giving rise to six plasmids expressing each of the three genes fused to either of the two Venus halves. Dlx5 control vectors were a kind gift of Andrew Bailey. COS cells and cES cells were transfected as described [16], and the cells were observed the next day by epifluorescence in a compound microscope.
The method used closely followed one previously described [59]. Briefly, 20 E7.5 mouse embryos were fixed in 4% formaldehyde, homogenized in lysis buffer, and sonicated. Cell extracts were harvested by centrifugation, incubated overnight with an antibody against mouse Geminin (Santa Cruz Biotechnology FL-209, 5 μg), and then immunoprecipitated with Protein-A-Sepharose. Precipitates were heated to reverse the formaldehyde cross-linking. The DNA fragments in the precipitates were purified by phenol/chloroform extraction and EtOH precipitation and used as a template for a PCR, using the following mouse N2-specific primers: forward: AACTCTCATAGCCCTAACTGTC, reverse: CCCTCCTCTCCTAATCTCCTTATGG. After 20 cycles of amplification, one-tenth of the reaction product was used as a template for a second round of a further 20 cycles. The final PCR products were run on a 1% agarose gel.
The GenBank (http://www.ncbi.nlm.nih.gov/Genbank) accession number for the chick homolog of Geminin is EU118174, and for BERT, it is EU118175.
|
10.1371/journal.pntd.0005652 | Neighborhood-targeted and case-triggered use of a single dose of oral cholera vaccine in an urban setting: Feasibility and vaccine coverage | In June 2015, a cholera outbreak was declared in Juba, South Sudan. In addition to standard outbreak control measures, oral cholera vaccine (OCV) was proposed. As sufficient doses to cover the at-risk population were unavailable, a campaign using half the standard dosing regimen (one-dose) targeted high-risk neighborhoods and groups including neighbors of suspected cases. Here we report the operational details of this first public health use of a single-dose regimen of OCV and illustrate the feasibility of conducting highly targeted vaccination campaigns in an urban area.
Neighborhoods of the city were prioritized for vaccination based on cumulative attack rates, active transmission and local knowledge of known cholera risk factors. OCV was offered to all persons older than 12 months at 20 fixed sites and to select groups, including neighbors of cholera cases after the main campaign (‘case-triggered’ interventions), through mobile teams. Vaccination coverage was estimated by multi-stage surveys using spatial sampling techniques. 162,377 individuals received a single-dose of OCV in the targeted neighborhoods. In these neighborhoods vaccine coverage was 68.8% (95% Confidence Interval (CI), 64.0–73.7) and was highest among children ages 5–14 years (90.0%, 95% CI 85.7–94.3), with adult men being less likely to be vaccinated than adult women (Relative Risk 0.81, 95% CI: 0.68–0.96). In the case-triggered interventions, each lasting 1–2 days, coverage varied (range: 30–87%) with an average of 51.0% (95% CI 41.7–60.3).
Vaccine supply constraints and the complex realities where cholera outbreaks occur may warrant the use of flexible alternative vaccination strategies, including highly-targeted vaccination campaigns and single-dose regimens. We showed that such campaigns are feasible. Additional work is needed to understand how and when to use different strategies to best protect populations against epidemic cholera.
| Oral cholera vaccine (OCV) is becoming part of the standard cholera-control toolkit, although experience in deploying OCV is limited. Adapting vaccination strategies to the global availability of vaccines and the local context (i.e., population movement, security constraints, etc.) is key to maximize the impact of OCV as a cholera-control tool. Here we describe the operational details of the first field use of a single-dose of OCV, which was deployed in a targeted manner, both at high-risk neighborhoods and then to neighbors of suspected cases after the main OCV campaign when sporadic cholera case reports continued. We show that it is feasible to conduct micro- and macro-targeted vaccination campaigns in urban areas like Juba with moderate to high coverage and without social unrest due to vaccinating some groups and not others. Flexible and context-adapted OCV dosing regimens and strategies should be considered in future deployments of the vaccine.
| Oral cholera vaccine (OCV) is an effective tool to prevent and control cholera both in endemic settings and in response to outbreaks [1,2]. On 23-June-2015, the Republic of South Sudan Ministry of Health (MoH) declared a cholera outbreak in Juba, the nation’s capital. Initial cases were traced back to 18-May in the United Nations Protection of Civilians Camp, where approximately 28,000 internally displaced people (IDP) resided. By the time the epidemic was declared, cases had been confirmed throughout the city and public health officials believed that Juba was at risk for a large cholera outbreak, with the threat of spread to other areas of the country. The MoH convened the National Cholera Taskforce to guide a comprehensive outbreak response involving case management, water and sanitation interventions, health education and hygiene promotion. Following a situation assessment and in light of the 2014 cholera outbreak with 6,269 reported cases and 156 deaths in multiple areas of the country [3], the MoH, supported by Médecins sans Frontières (MSF), decided to integrate OCV into the cholera response in Juba.
Only 270,000 doses were released from the global emergency OCV stockpile due to severe supply limitations despite the much larger at-risk population (500,000–1,000,000 people). The MoH decided to use an off-label, single-dose, regimen in a targeted vaccination campaign. The rationale was based on preliminary results from a large randomized clinical trial [4] demonstrating significant 1-dose protection in Bangladesh, immunogenicity studies [5] and modelling analyses [6] showing that even with a significantly less effective one-dose regimen, one-dose campaigns may save more lives than their two-dose counterparts when supply is limited. The goal was to quickly provide protection to the maximum number of people at highest risk with a, perhaps less effective, single-dose regimen, rather than covering half the number of people with the standard two-doses. The possibility of providing a second dose, to potentially increase the effectiveness and extend the duration protection, when supplies were available was not ruled out.
Two targeted vaccination approaches were used with an aim to halt transmission and shorten the duration of the epidemic. First, OCV was targeted to neighborhoods with evidence of significant transmission just prior to the campaign and vulnerable groups at higher risk of cholera including IDPs, prisoners and health care workers. After the main campaign, sporadic case reports continued, mostly in unvaccinated neighborhoods. Given that the risk of cholera has been shown to be highly elevated among those living around a cholera case in the days after the case presents for care at a clinic [7,8], the remaining vaccine was delivered to neighbors living around suspected cholera cases together with water sanitation and hygiene measures (case-triggered interventions). Further details related to the decision-making process, timeline and vaccine effectiveness are described in detail elsewhere [9,10]
Here, we describe the operational details and vaccine coverage of these spatially targeted OCV delivery approaches, from campaigns that represent, to our knowledge, the first field-use of a single-dose of OCV in response to an epidemic, and the first use of case-triggered cholera interventions including OCV. We explore vaccine uptake in the different areas targeted, including neighborhoods and smaller areas around the households of suspected cases, identify difficult-to-reach population groups and discuss alternative campaign strategies to improve vaccine coverage.
The campaign setting was particularly challenging amid a humanitarian crisis with significant population displacement. Accurate population estimates were not available, so we used two approaches to define the target populations when planning the campaigns. First, we extrapolated population estimates for different areas of the city using data from the latest population census in 2009 [11]. This census was conducted prior to independence and civil war and extrapolated population size estimates are believed to vastly underestimate the true population size. Next, we used estimates of the number of built structures in each area from recent digitized satellite images (http://wiki.openstreetmap.org/wiki/WikiProject_South_Sudan). We initially assumed 70% of the structures with a roof footprint from 5 to 250 m2 were residential and that each had an average of 6 people [11].
In this rapidly evolving epidemic, the decisions of where to target vaccine were based on the most up to date cumulative attack rates, recent incidence and local knowledge of known cholera risk factors. Three main areas were identified, using unofficial boundaries and referred to here as Kator, Northern Juba and Gumbo, with combined population estimates ranging from 53,543, from census data, to 368,136 from satellite imagery (Fig 1). The three areas differed greatly by socio-demographics. Kator is a densely populated area including a semi-commercial part of the city that had experienced a spike of suspected cholera cases just prior to the campaign and a large slum-like area bordering the Nile river. Northern Juba is a fairly-isolated settlement next to a large military base, predominantly inhabited by military members and their families. Gumbo is an area with moderate population density on the south-eastern side of the river Nile with predominantly poor-housing and persistent notification of cholera cases preceding the campaign. In addition to the three targeted areas of the city, IDPs living in informal camps, inmates in Juba’s prison, health care workers and residents living close to suspected cholera patients presenting after the main campaign were targeted by mobile teams.
A single dose of OCV (Shanchol®, Shantha Biotechnics Ltd, Hyderabad, India) was offered to all persons older than 12 months presenting at vaccination sites, regardless of her/his area of residence. Twenty fixed vaccination sites operated from 08:00–17:00 from 31-July to 5-August each with a team of approximately 20 people per site (3–4 vaccinators, 3–4 individuals preparing the vaccine, 8–10 registrars filling out vaccine cards and tally sheets, 1 security guard, 2 health promoters and 1 team supervisor). As the number of individuals coming to the sites slowed (3-August), vaccination teams split into semi-mobile units and set up mini-vaccination sites to reach those not yet covered. Vaccines were stored under cold chain (2–8°C) using a refrigerated truck, transported to the vaccine site in their original Styrofoam box with icepacks and then used at ambient temperature the day of vaccination. We distributed a vaccination card to each vaccinee indicating her/his name, age, vaccination location, date of vaccination and vaccine lot number.
We avoided the widespread use of radio and other media to publicize the campaign due to concerns that offering limited vaccine only in selected parts of the city could spark civil unrest. Trained health promoters disseminated information regarding the campaign in the targeted communities, and members of the target populations were recruited to spread key messages using megaphones.
To assess vaccine coverage in the neighborhood-targeted (main) campaign, a random sample of the population living in each of the target areas was selected using a stratified spatial sampling approach, with households serving as the primary sampling unit. A total of 128 households, including all household members, were required in each of the three target areas to estimate each area-specific coverage with a precision of ±5%. Detailed methods for the coverage survey, conducted 9–14 August, are provided in S1 Text.
From 13–26 August, after the main campaign but before cholera case reports had stopped within the city, OCV was included as part of a case-triggered comprehensive targeted intervention (CTI) approach. Vaccine was offered together with soap, water purification tablets, a leaflet on cholera prevention and health promotion by mobile teams made up of staff from multiple organizations, including the MoH, South Sudanese Red Cross, Oxfam and MSF. This activity required close coordination with multiple governmental and non-governmental actors to (1) detect and test suspected cases, (2) rapidly communicate results from cholera diagnostic tests, (3) locate the case’s household and decide on the location for intervention in conjunction with local leaders and (4) deploy the intervention in a timely manner.
Patients from Juba reporting to any of the cholera treatment centers or oral rehydration posts with a stool sample positive for cholera using the Crystal VC rapid diagnostic test (RDT), either directly or after a 4–6 hour enrichment in alkaline peptone water [12], were put on a list for CTI eligibility. Due to limited human resources, cases coming from areas that had not been covered in the neighborhood-targeted OCV campaign and those testing positive to the, more specific [12], enriched RDT were prioritized. When possible, teams also conducted CTI in previously vaccinated areas where it appeared that cholera transmission may have continued. These CTIs were also targeted to the homes of individuals who died of acute watery diarrhea, either in the community or in a health-facility, even if no stool sample had been tested.
The Juba County Health Department’s rapid response team and MSF staff travelled to the home of CTI-eligible suspected cases. Together with a community leader, they identified a suitable intervention site as close as possible to the home of the patient and recruited four community members: two to assist with security/crowd control and two for going door-to-door informing the neighbors about the intervention and encouraging residents to come to the sites. Volunteers were not informed of the specific rationale behind the location of intervention site (i.e. details of the suspect cholera case triggering the CTI) to respect the patients’ privacy, but they were given a specific geographic focal area. The site was typically set-up the following day by a 5-7-person team (1 site supervisor, 1–2 vaccinators, 1–2 registrars for completing vaccination cards and tally sheets and 2–3 people delivering the water/sanitation/hygiene intervention). All individuals coming to the site were eligible for OCV regardless of whether they had been vaccinated in the main vaccination campaign.
To assess vaccine coverage in each case-centered targeted intervention cluster, we selected 30 spatially random points within 350-meters (assumed as the catchment area of the interventions) of suspected case households (Fig 1, S1 Text). As with the population-based survey for the main campaign, the closest household to each GPS point was included in the survey, but instead of ascertaining the vaccination status of all individuals living in the household as done in the main coverage survey, one person was selected at random from those residing in (but not necessarily present at the time of the first survey visit) the household.
We collected data on age, sex, vaccination status (both verbal and confirmed with card) and reasons for non-vaccination (when applicable) for everyone included in the coverage surveys. We also collected household-level variables including the number of household members at the time of the campaign, the number of built structures included in each household and their spatial coordinates.
We estimated mean vaccination coverage and 95% confidence intervals for individual vaccination target areas (both neighborhoods and areas around case-triggered interventions) and for the entire target population. In secondary analyses, we estimated the coverage by age group, sex and distance to the closest vaccination site. Individuals with missing information on vaccination status were excluded from the analysis. Relative risks and 95% confidence intervals were estimated using a generalized linear model with a log link. All confidence interval estimates for vaccine coverage and relative risks took into account the survey design (clustering by household and vaccination area) using the svy commands in Stata 12.0 (College Station, TX, USA).
This was a public health intervention designed to prevent the spread of cholera, informed consent for participation was not required. The activities presented in this study were conducted as standard monitoring and evaluation exercises, thus approval from ethical review committees was not obtained. Although written informed consent was not solicited for the coverage surveys, all interviewees provided verbal consent and no identifiable information was collected other than household coordinates.
From 31-July to 26-August-2015, 162,377 people were vaccinated through targeted campaigns in Juba. 127,191 vaccines were distributed at fixed vaccination posts as part of a neighborhood-targeted approach from 31-July to 5-August-2015, 8,592 (6.1%) were distributed in informal IDP settlements. Mobile vaccination teams provided 1,011 doses (0.7%) in a local prison and 3,455 doses (2.5%) to healthcare workers. A further 22,128 people received the vaccine as part of 17 case-triggered CTI deployments from 13–26 August.
Most of the vaccines (91,953 doses, 65.6%) were distributed in the targeted area of Kator. In the neighborhoods targeted in Northern Juba, 21,039 individuals received OCV, which was greater than the population estimated by both census and using satellite imagery (Table 1). Just over half of those receiving vaccine during the neighborhood-targeted campaign were male (71,945, 51.3%) and 75,638 (53.9%) were at least 15 years old (Table 2, S1 Table).
A total of 371 households were included in the coverage survey of targeted neighborhoods (Fig 2). No households refused to participate in the survey. Household size varied from 1 to 20 with a median of 6 (S2 Table). The mean number of built structures per household was 2 (S2 Table). All but two households with available coordinates were within 1 kilometer from the closest vaccination site (Fig 3), with a median distance to the closest vaccination site of 156 meters. We ascertained the vaccination status for 96.8% (2578/2662) of individuals, with 94% of those who reported to have been vaccinated providing a vaccination card. We estimated the vaccine coverage (self-report) to range from 60–70% (S1 Table) with an overall population-weighted coverage across the three targeted areas of 68.8% (95% CI: 64.0–73.7, Table 2).
Vaccination status between individuals in the same household was more correlated than expected, with a survey design effect of 7.3. In Northern Juba, nearly 1 in 3 households (30%) reported that no household members were vaccinated (S2 Table). The proportion of household members vaccinated decreased with distance to the closest vaccination site (Fig 3). Coverage was highest among children 5–14 years (90.0%; 95% CI: 85.7–94.3). While overall vaccine coverage was similar between women (68.9%; 95% CI: 63.7–74.0) and men (64.7%; 95% CI: 57.7–71.7) on average; adult women tended to have higher coverage than adult men (RR 0.81, 95% CI: 0.68–0.96), with less than half the men 15 years or older reporting to have been vaccinated (Fig 4, S3 Table). The main reasons for non-vaccination in the main campaign were; (1) not being aware of the campaign (256, 30% of unvaccinated individuals), (2) being absent during the time of the campaign (202, 23%), and (3) not having time (129, 15%, S4 Table).
Of the 54 suspected cholera cases from Juba screened by direct and enriched RDT during the CTI period, 17 were positive by enriched RDT. We carried out CTI at the homes of 14 (82%) of these enriched RDT positive cases. One additional direct RDT positive household (out of 9 direct-RDT positive only) was included in CTI as it occurred on a day with no other priority activities for the teams. The remaining two CTIs occurred around the residence of individuals who had reportedly died due to acute watery diarrhea for whom no sample was available (1 community- and 1 facility-death). Two additional deaths were reported during this period although the team was unaware of these deaths at the time of the activities.
All but two of the CTIs took place in areas that had not been covered in the main campaign. The CTIs occurred 1–6 days after the suspected cholera case had presented at the health facility (mean delay 3.4 days). Ten CTIs were single-day events and the remaining 7 took place over a two-day period. Based on tally sheets collected at the CTI sites, 11,491 (54.4%) of those who received the intervention were female, 4,091(19.4%) were children 1–4 years old and 7,886 (37.3%) were children 4–14 years old.
Coverage surveys were carried out in 13 of 17 CTIs (exact location of the patient’s home was unavailable for 3 and one was outside of Juba town), with a total of 390 individuals sampled. Vaccine coverage per CTI site ranged from 30% (95%CI 12.6–47.4) to 86.7% (95%CI 73.7–99.6). Overall, the coverage was 51.0% (95%CI 41.7–60.3), with no significant difference between those sites targeted inside and outside the main campaign target area. Coverage patterns were like those observed in the main neighborhood-targeted campaign. Adult men were less likely to have received the vaccine compared to adult women (RR 0.73; 95%CI 0.60–0.89). Overall coverage was 45.7% (95%CI 35.5–55.8) among men and 55.8% (95%CI 45.5–66.1) among women. Coverage was also highest among school-aged children (Table 2).
We provided a single-dose of OCV through spatially-targeted campaigns to over 160,000 individuals in Juba, South Sudan. We achieved nearly 70% vaccine coverage within the main, neighborhood-targeted, campaign and had no significant challenges in using a targeted strategy within this large urban setting. Our experience should ease concerns about targeting specific populations with OCV in urban settings, even during an outbreak. Similarly, targeted OCV campaigns have been successfully implemented in urban slums of Haiti [13]. These findings support the possibility of targeting particular neighborhoods that may be responsible for driving urban cholera epidemics, which may provide an efficient way to minimize cost and maximize public health impact [14].
We also demonstrated that it is feasible for multiple actors (e.g., MoH and humanitarian organizations) to work together to rapidly provide a suite of cholera control interventions to the high-risk group living near cholera cases with moderate coverage. While this type of approach is intuitively appropriate for cholera control given the evidence of elevated risk around cases [7,8], evaluations of the effectiveness of similar interventions in the future are needed. Here, the case-triggered CTI approach was used at the end of the outbreak, when cases were sporadic, with the hopes of quelling the outbreak. Consequently, numbers were low and given the approach was hastily devised during the outbreak, there was limited time for detailed planning to optimize impact and to incorporate any detailed evaluation of effectiveness. More work is needed to best define the best mix of components to include in CTI, including the possibility of prophylactic antibiotics, to halt cholera transmission. This approach is not likely to be a silver bullet for cholera control, but may prove to be an efficient strategy in periods of low transmission, perhaps seasonally as has been proposed in Haiti [15], or to accelerate the end of an outbreak after mass campaigns.
Although some OCV campaigns have achieved higher coverage, our estimates are consistent with others in urban areas [16–19]. Despite initial concerns, vaccine sites were not over-run with population from elsewhere in the city, and even in the targeted areas, coverage was less than expected. One potential reason that public interest in vaccine was lower than expected may be due to ‘cholera fatigue,’ where after the much larger 2014 outbreak [3], individuals and the media paid much less attention to cholera in 2015.
While vaccine coverage was lower in adult men, use of the administrative data alone masked this difference and suggested roughly equal OCV coverage by sex. This was especially apparent in Kator where a substantial proportion of those that received the vaccine during the campaign were adult men (53.6% per tally sheets), but little over 40% of the men who lived in the area received the vaccine (S1 and S2 Tables). Being a commercial part of town, it is possible that these were male businessman working in the area during the day but who lived elsewhere in the city.
A door-to-door strategy may have been more appropriate for this highly targeted campaign, however other campaigns using a mixture of fixed sites and door-to-door vaccine delivery report a similar coverage among urban populations and similar challenges reaching adult men [19]. Keeping vaccination sites open later (security situation dependent), and perhaps moving them near places where people congregate in the early evening could prove a useful strategy to improve coverage among men, given that lack of time was the most common reason for non-vaccination. On the other hand, if the reason for not having time was due to work commitments, campaigns could target workplaces during the day. The high coverage among school-aged children likely reflects the success of using schools as vaccination sites, as has been observed in other settings [20].
We observed heavy clustering of vaccination status within households of the main coverage survey. Over 15% of the households sampled had no vaccinated individuals, despite most households visited being well within walking distance from a vaccination site (e.g., more than half being within 160 meters). On the other hand, a third of the households had 100% coverage among eligible members. The most common reason for non-vaccination in the neighborhood campaign was not being aware of the campaign, perhaps reflecting the limited use of radio and other measures to publicize the campaign. Ensuring at least one person in every house is aware of the OCV campaign could be an important approach to increasing overall coverage. A door-to-door strategy for social mobilization rather than for vaccine delivery could help increase household-level knowledge of the campaign.
This experience in Juba highlights the key challenge of designing public health interventions in low-resource, volatile settings such as South Sudan, where accurate and up-to-date demographic information is not always available. The use of satellite imagery has been used to estimate population size in unstable settings [21] and innovative initiatives like MissingMaps (www.missingmaps.org) make the task more feasible, even in the world’s most vulnerable populations. Nevertheless, local information regarding the different observable characteristics of residential and non-residential built structures and the number of persons per built structure are needed to obtain accurate estimates. Developing standardized methods for gathering and sharing information to aid population estimation in low-resource, data-poor settings is a key priority for efficient public health programming.
Our findings come with several limitations. We based our spatial sampling on building density from recent satellite images rather than true population density. It is possible that some areas of the city, especially the most vulnerable, overcrowded areas may have more people but less built structures and could therefore be underrepresented. Furthermore, we interviewed the senior household member for information regarding vaccination status of the other household members in the main coverage survey. This may have led to information bias and contribute to our findings of high intra-household clustering. This also may have led to less precise and accurate estimates of the reasons for non-vaccination within the household.
In conclusion, we showed that targeting OCV in response to an outbreak within a large urban population both to neighborhoods and neighbors of cholera cases is feasible and well accepted by the population. Developing and testing new ways to reach traditionally hard-to-reach groups, including adult men, remains a priority. While cholera continues to strike in complex settings with mobile populations and dynamic security constraints, flexible targeted approaches and alternative dosing schedules, like the one described here, are needed to maximize the potential impact of the vaccine.
|
10.1371/journal.pntd.0007256 | Aurora kinase protein family in Trypanosoma cruzi: Novel role of an AUK-B homologue in kinetoplast replication | Aurora kinases constitute a family of enzymes that play a key role during metazoan cells division, being involved in events like centrosome maturation and division, chromatin condensation, mitotic spindle assembly, control of kinetochore-microtubule attachments, and cytokinesis initiation. In this work, three Aurora kinase homologues were identified in Trypanosoma cruzi (TcAUK1, -2 and -3), a protozoan parasite of the Kinetoplastida Class. The genomic organization of these enzymes was fully analyzed, demonstrating that TcAUK1 is a single-copy gene, TcAUK2 coding sequence is present in two different forms (short and long) and TcAUK3 is a multi-copy gene. The three TcAUK genes are actively expressed in the different life cycle forms of T. cruzi (amastigotes, trypomastigotes and epimastigotes). TcAUK1 showed a changing localization along the cell cycle of the proliferating epimastigote form: at interphase it is located at the extremes of the kinetoplast while in mitosis it is detected at the cell nucleus, in close association with the mitotic spindle. Overexpression of TcAUK1 in epimastigotes leaded to a delay in the G2/M phases of the cell cycle due a retarded beginning of kinetoplast duplication. By immunofluorescence, we found that when it was overexpressed TcAUK1 lost its localization at the extremes of the kinetoplast during interphase, being observed inside the cell nucleus throughout the entire cell cycle. In summary, TcAUK1 appears to be a functional homologue of human Aurora B kinase, as it is related to mitotic spindle assembling and chromosome segregation. Moreover, TcAUK1 also seems to play a role during the initiation of kinetoplast duplication, a novel role described for this protein.
| The cell cycle is a complex and highly regulated cellular process in which different checkpoints are coordinated and a unique pattern of protein activities is present. In trypanosomatids, this process is even more complex because of the presence of the kinetoplast, a network of circular DNA inside a large mitochondrion that coordinates its division and segregation with the nuclear mitosis. Aurora kinases comprise a family of enzymes that regulate key steps during the cell cycle by varying their subcellular localization. In this work, we identified three Aurora kinase genes (TcAUK1, -2 and -3) in T. cruzi, the agent of Chagas disease, and established that TcAUK1 has the typical behavior of a mammalian Aurora B kinase, localizing inside the nucleus and associating with the mitotic spindle during mitosis. On the other hand, during interphase TcAUK1, localizes at each side of the kinetoplast, showing a novel localization that has not been described before. Finally, here we present evidence for the relationship between the novel localization of TcAUK1 and its participation in kinetoplast division.
| Cell cycle in eukaryotic cells involves the sequential transition between G1, S, G2 and M phases and this progression is tightly regulated by protein kinases and phosphatases [1]. During the M phase, the cell divides its nucleus to originate two daughter cells with the same genetic content. In this event, Aurora kinase proteins play a crucial role. The Aurora kinase family of proteins presents a variable number of members among different organisms. While yeasts present a single Aurora kinase gene [2,3], organisms like C. elegans and D. melanogaster have two genes [4–7], whereas in vertebrates three Aurora kinase proteins are present [8]. In this last group, the three members of the Aurora kinase family are named Aurora-A, -B and -C and each protein plays specific functions during the cell cycle. In organisms with a single Aurora kinase gene, the encoded protein combines the function of both Aurora-A and -B, whereas in organisms with two Aurora proteins, one behaves as an Aurora-A while the other has functions similar to Aurora-B. In humans, Aurora-C is expressed in germinal cell lines and its function has not been elucidated yet. A distinctive feature of all Aurora proteins is that they change their cellular location during mitosis progression, according to their different roles. Aurora-A, the so-called Polar Aurora, is involved in centrosome maturation/migration and bipolar spindle formation/stabilization [9] and therefore, it is found in the neighborhood of the dividing centrosome in early mitosis, after which it moves with each of duplicated centrosome to opposite extremes of the cell, where the spindle poles locate during G2 phase [10]. Aurora-B is named Equatorial Aurora because it locates in the mid-plane of the cell during mitosis (metaphase) and, as nuclear division proceeds, it is tightly associated with segregating chromosomes. This Aurora forms the Chromosomal Passenger Complex (CPC) with other three proteins (INCENP, Survivin and Borealin). As part of this complex, Aurora-B promotes chromatin condensation in prophase through phosphorylation of histone H3 at Ser10 [11]. In metaphase, it participates at the Spindle Checkpoint to ensure correct chromosome segregation during anaphase [12]. Finally, in cytokinesis the CPC settles in the cell midzone and participates in the formation of the contractile ring and the cleavage furrow [13].
Trypanosoma cruzi is a protozoan of the Kinetoplastida order and is the etiological agent of the American Trypanosomiasis, also known as Chagas disease. This disease is endemic in Latin America but in the last decades, due to the increasing migratory flux, a growing number of infections have been detected in non-endemic countries like the United States and Spain [14]. The complex life cycle of T. cruzi involves different forms: the epimastigotes and metacyclic trypomastigotes are present in the insect vector, and the amastigotes and bloodstream trypomastigotes are found in the vertebrate host. All these forms contain a single flagellum emerging from the basal body, a nucleus and a mitochondrion carrying the DNA complex known as kinetoplast. During cell division, all these organelles are replicated and segregated into two daughter cells in a synchronized manner. Moreover, both the existing and newly synthesized basal bodies physically interact with the duplicating kinetoplast and drive its division [15], meaning that the synchronization is due in part to a physical contact between these organelles. During T. cruzi mitosis, contrary to what is observed in other organisms, the nuclear membrane remains intact and the chromosomes do not condensate. Despite this, the formation of a mitotic spindle inside the nucleus has been described [16] and it is well known that chromosome segregation is mediated by a microtubules-dependent mechanism [17]. While most eukaryotic cells contain many mitochondria with separate copies of circular DNA molecules, kinetoplastids have a single mitochondrion with a genome in which DNA molecules are physically interlocked forming a big network, the so-called kinetoplastid DNA (kDNA). These DNA molecules consist of two type of circles: the larger maxicircles that are the equivalent of the mitochondrial genome in other organisms and are present as several dozen identical copies per cell, and the smaller minicircles that codified for guide RNAs (gRNAs) and are in thousands copies per cell. During cell division, this DNA does not only need to be duplicated but a well-orchestrated de-concatenation and segregation, driven by the recently duplicated flagella basal bodies, needs to take place [18]. The complexity of the mitochondrion genome adds for a particular kDNA S-phase besides from the classical G1, S and G2/M cell cycle stages. Although contributions have been made by different authors to the knowledge about the kDNA replication mechanism [19], many of the molecular players involved in the later steps of this process remain obscure, mainly the ones involved in kDNA division and segregation.
Aurora kinase genes have been identified and described in different protozoan organisms. In Leishmania major, a single Aurora gene (Lmairk) has been reported but its function has not been studied yet [20]. In the apicomplexa Plasmodium falciparum, three Aurora genes have been described (Pfark -1, -2 and -3) Pfark-1 was defined based on its subcellular localization as the classic Aurora gene present in other organism, whereas the remaining two Pfark proteins seem to be involved in cellular processes exclusive of this organism [21]. Davids and coworkers found a single Aurora gene in the biflagellate Giardia lamblia [22], showing that this protein adopts a cellular localization similar to mammalian Aurora-A and Aurora-B and is involved in microtubules dynamics reorganization at mitosis and interphase. In Trypanosoma brucei three Aurora genes were identified by Wang and coworkers but gene silencing experiments demonstrated that only TbAUK1 is a functional gene, at least in the procyclic form [23]. This protein is involved in mitotic spindle assembling and it was defined by the authors as a human Aurora-B ortholog. Later, these same authors showed that TbAUK1 silencing in the bloodstream form leads to failure to conclude cytokinesis and cell shape alteration, both effects associated to a microtubule filaments disruption [24]. TbAUK1 seems to be associated to others proteins conforming a complex like the CPC of mammals and, as with Aurora B, the proteins of this complex affect TbAUK1 localization and function [25].
A comparative analysis of the kinomes of three pathogenic kinetoplastids—including T. cruzi–by Parsons and co-workers reported the presence of several kinases normally associated to roles in cell division, the Aurora kinases being part of this group [26]. In this work we report the initial characterization of three Aurora kinase proteins in T. cruzi (TcAUK1, TcAUK2 and TcAUK3). In addition, by a detailed analysis of TcAUK1 localization, we have found that this protein shows the canonical behavior of a chromosome passenger protein, being associated with the mitotic spindle during nuclear division. Furthermore, we detected that during interphase, TcAUK1 is located at both sides of the kinetoplast. Finally, we report that TcAUK1 overexpression in epimastigote forms causes a delayed G2-M transition, presumably by affecting the onset of kinetoplast duplication.
Radio chemicals were purchased from PerkinElmer Life Sciences, and restriction endonucleases were from New England Biolabs, Beverly, MA. Bacto-tryptose, yeast nitrogen base, and liver infusion were from Difco. All other reagents were purchased from Sigma.
The gene sequences corresponding to TbAUK1 (Tb927.11.8220), TbAUK2 (Tb927.3.3920) and TbAUK3 (Tb927.9.1670) were used to screen T. cruzi sequences in TryTrip database using BlastN algorithm. Pairwise alignment and motif search were performed on high-scored targets by EMBL-EBI tools [27] and Pfam [28], as well as manual inspection. Multiple sequence alignment was performed in MEGA 5 software [29] with ClustalW algorithm and visualized with BioEdit software [30].
T. cruzi epimastigote of CL Brener strain were grown at 28°C with 5% CO2 in LIT medium [5 g.l-1 liver infusion, 5 g.l-1 Bacto-tryptose, 68 mM NaCl, 5.3 mM KCl, 22 mM Na2PO4, 0.2% (w/v) glucose, 0.002% (w/v) hemin] containing 10% v/v fetal bovine serum (NATOCOR, Argentina), 100 units.ml-1 penicillin and 100 μg.l-1 streptomycin. Cell density was maintained between 1x106 and 1x108 cells.ml-1 sub-culturing parasites every 6–7 days. For growth curve determinations, a sample of culture supernatant was taken and swimming epimastigotes were fixed by incubation in 4% formaldehyde in PBS for 5 min at room temperature. Cell density was determined by counting at least three independent cultures in an hemocytometer. Specific growth rate (μ, expressed as h-1) was estimated by the slope of the graphic “Ln of culture cell density” vs “culture time” (h). Cells duplication time (DT) was calculated according to the formula:
DT=ln2μ
Cercopithecus aethiops (green monkey) Vero cells (ATCC CCL-81) were cultured at 37°C and 5% CO2 supplied in Minimum Eagle Medium (MEM, Gibco) supplemented with 10% fetal bovine serum (HyClone), 2 mM L-Glutamine (Sigma), 100 units.ml-1 penicillin and 100 μg.l-1 streptomycin. For the obtention of T. cruzi trypomastigotes and amastigotes Vero cells were infected with trypomastigotes (1:50 ratio) 24 h after being plated and maintained in MEM supplemented with 3% FBS. Trypomastigotes in culture supernatant were harvested by centrifugation and processed as needed. Amastigotes were collected from 10–11 day old cultures from the supernatant (90% or higher amastigotes/trypomastigotes ratio), centrifuged, and processed as needed.
Based on the sequences of Aurora kinases homologs found in T. cruzi database, specific primers were designed: TcAUK1 (TcAUK1-NcoI-fwd 5´-CCATGGTGAGTGCGGCGGAGGGCGGCCAA-3´ and TcAUK1-XhoI-rev 5´-CTCGAGGTTCTCCTTTCCGCCCGAGAAGT-3´), TcAUK2 (TcAUK2-BamHI-fwd 5´-GGATCCGCAGCACCACAACTTGAGTTCC-3´ and AUK2-XhoI-rev 5´- CTCGAGCTTCTTCTTCTTCTTCTCCCCATTT-3´), TcAUK3 (AUK3-NcoI-fwd 5´- CCATGGTGTGGTCGCTGGATGACTTTGAT-3´ and AUK3-XhoI-rev 5´-CTCGAGTAAATTCTCTGCCGCATCAACCGT-3´). Polymerase Chain Reaction (PCR) was performed in a PTC-150 MiniCycler (MJ Research). For this, genomic DNA of T. cruzi was isolated as described previously [31] and gene amplification was performed by using a high fidelity DNA polymerase (Herculase II, Stratagene). The thermal cycling conditions were specific for each TcAUK gene. Amplification products were gel-purified, subcloned into pGEM-T Easy vector (Promega), transformed into E. coli DH5α competent cells and both strands were sequenced (Macrogen, Korea).
Genomic DNA of epimastigote forms was digested (approx. 30 μg) overnight with 30 units of the indicated restriction enzyme (New England Biolabs). After digestion, DNA was SpeedVac concentrated (Jouan RCT 60 Refrigerated Cold Trap) and electrophoresed for 8–12 h in 0.8% agarose gel (1 V/cm) and was then denatured, neutralized and transferred onto nylon membrane (GeneScreen, Perkin Elmer) for Southern blot analysis [32]. For this, specific radiolabeled probes were generated by primer extension using full-length TcAUK genes and [α-32P]-dCTP (NEBlot kit, New England Biolabs). dCTP radiolabeled probes were then purified with MicroSpin G-50 columns (GE Healthcare) and heat denaturalized before proceeding with hybridization. Probes were hybridized at 65°C (overnight) and washed at 65°C using 2x SSC, 1xSSC and 0.5x SSC with 0.1% SDS sequentially to remove excess of probe. Blot was developed by exposing membrane to Phosphoimager Storm system (Pharmacia-Biotech).
To perform Pulsed Field Gel Electrophoresis, CL Brener exponentially growing epimastigotes were washed with PBS-Glucose 2%, suspended in PBS and mixed with one volume of low melting point agarose 1.4% in PBS. After polymerization, the agarose plug was incubated with LIDS buffer (1% Lauryl Sulfate Lithium Salt, 10 mM Tris-HCl pH8.0, 0.1 M EDTA) for 48 h at 37°C. Afterwards the blocks were washed 6 times with NDS 0.2% buffer (0.2% N-Lauroylsarcosine-Sodium Salt, 0.1 M EDTA, 2 mM Tris base) and, before running the agarose gel, they were equilibrated in TE buffer pH 8,0. Gel electrophoresis was performed at 16°C in three steps: 1) 3 V/cm changing periods every 90–200 sec during 30 h; 2) 3 V/cm changing periods every 200–400 sec during 30 h; 3) 2.7 V/cm changing periods every 400–700 sec during 24 h. After electrophoresis, DNA was transferred to a nylon membrane and hybridized with TcAUKs probes as described above for Southern blot analysis.
Total RNA was isolated from epimastigotes, trypomastigotes and amastigotes forms using Trizol according to the manufacturer’s protocol (Invitrogen). cDNA was obtained from mRNA by Transcriptor First Strand cDNA Synthesis Kit (Roche) using oligo(dT)18 primer, following the supplier´s instructions. These cDNA samples were used to amplify a fragment of TcAUKs genes and the housekeeping Actin gene (TcCLB.510945.30). For all TcAUKs targets the same forward Tc-SL primer was used (5´-AACGCTATTATTGATACAGTTTC-3´) whereas a specific reverse primer for each one was designed: TcAUK1-RT-rev (5´-CCACCCAAAGTCTGCCAACTTA-3´), TcAUK2-RT-rev (5´-AGCGTGCGGTGAACGTTGATCT-3´) and TcAUK3-RT-rev (5´-AATCCATCGTGCCGCAAAGCGT-3´). In the case of the housekeeping actin gene the primers used were: TcActin-fwd (5´-ATGATCATCGTGGACTTTGGGT-3´) and TcActin-rev (5´-TTCCGCTTGGGTGTGAACAGC-3´). The PCR reactions included an initial denaturalization step at 95°C for 2 min, followed by 35 cycles at 95°C for 1 min, annealing at 60°C for 45 sec, extension at 72°C for 45 sec and a final extension at 72°C for 5 min. Amplification products were visualized by electrophoresis on 1.5% agarose gel. Then, they were gel extracted, subcloned into pGEM-T Easy vector (Promega) and sequenced (Macrogen, Korea).
For western blot analysis cells were suspended in lysis buffer (50 mM Tris-HCl pH 8.0, 1 mM EDTA, 1 mM DTT, 0.1% Triton X-100, 1% NP-40, 1 mM PMSF and 1 μg. ml-1 E-64) and lysed by freeze/thaw cycles in liquid nitrogen. The obtained lysate was centrifuged at 10,000 x g for 30 min and the pellet was discarded. Total protein concentration in the extracts (supernatant) was estimated by the Bradford quantification, and an aliquot containing 20–80 μg of proteins was loaded onto 12% (w/v) SDS-polyacrylamide gel, solved by electrophoresis as described by Laemmli [33] and electro-transferred to nitrocellulose membranes (Hybond-C, Amersham Pharmacia Biotech). The membranes were then blocked with 5% (w/v) non-fat milk suspension in TBS-Tween 0.05% for 2 h and TcAUK1 was detected with a rabbit antiserum to TcAUK1 (custom produced by GenScript Corporation against the peptide PRGKRMRGAADFSG, amino acids 292 to 305 of TcAUK1) and a goat antiserum to rabbit IgG HRP-labeled secondary antibody (PerkinElmer). Specific TcAUK1 signal was developed with the ECL Plus Western blotting detection system (PerkinElmer Life Sciences).
The full-length coding sequence of TcAUK1 gene sub-cloned into pGEM-T Easy vector (see above) was isolated by digestion with EcoRI and XhoI nucleases and cloned into pTREX plasmid [34] digested with the same restriction enzymes. T. cruzi epimastigotes of CL Brener strain were electro-transfected with empty pTREX plasmid or the pTREX-TcAUK1 construct as described previously [34]. Stable transfectant pools were achieved after 60 days of treatment with 500 μg.ml-1 G418 (Gibco BRL, Carlsbad, CA). Once selection has finished, single clone cell cultures were obtained for pTREX-TcAUK1 transfectants by the limit-dilution cloning method. Transgenic condition of several clones was confirmed by Southern and Western blot analyses.
For the expression of the fusion protein TcAUK1-GFP, the coding sequence of TcAUK1 was amplified by PCR with primers TcAUK1-NcoI-fwd (5´-CCATGGTGAGTGCGGCGGAGGGCGGCCAA-3´) and TcAUK1-Rev-STOPLess-BamHI (5´-GGATCCGTTCTCCTTTCCGCCCGAGAAGTCC-3´). The amplification product was sublconed into pGEM-T Easy vector, isolated by digestion with EcoRI and BamHI endonucelases and cloned into pTEX-eGFP-TEV-HA-EEF plasmid (kindly donated by Dr. Leon A. Bouvier, Instituto de Investigaciones Biotecnológicas, IIB-INTECH) digested with the same restriction enzymes. T. cruzi epimastigotes of CL Brener strain were electro-transfected with this construct and localization of fusion protein TcAUK1-GFP was evaluated after 48 h by fluorescence microscopy.
T. cruzi epimastigote and trypomastigote forms were harvested by centrifugation, washed once with PBS and allowed to adhere to poly-L-lysine coated coverslips. Vero cells cultured in 24 wells plate with a sterile coverslip, were infected and at different days after infection, culture medium was removed and cells were washed once with PBS previous to further processing. Parasites and infected Vero cells were fixed in 4% paraformaldehyde and washed twice with PBS. After been permeabilized with 0.2% Triton-X100 in PBS (PBT solution), cells were treated with blocking solution (1% BSA in PBS) for 1 h at room temperature. For TcAUK1 and mitotic spindle double staining, epimastigote forms were first incubated with a mix of the rabbit antiserum to TcAUK1 (1:200 dilution, GenScript Corporation) and the monoclonal mouse anti-β-tubulin KMX-1 antibody (1:400 dilution, Chemicon International) and then incubated with a mix of goat anti-rabbit IgG Alexa Fluor 488-labeled (1:500 dilution, Invitrogen) and goat anti-mouse IgG Alexa Fluor 594-labeled (1:500 dilution, Invitrogen). For TcAUK1 labeling in trypomastigote and amastigote, cells were first incubated with the rabbit antiserum to TcAUK1 and then with the secondary antibody goat anti-rabbit IgG Alexa Fluor 488-labeled. In all cases cells were incubated with the first antibody for 1 h at room temperature, then washed three times with PBT solution and finally incubated with the secondary antibody another hour at room temperature. After being washed three times with PBT solution, the slides were mounted in VectaShield mounting medium (Vector Labs) containing DAPI and examined with a fluorescence microscope (model BX41, Olympus). For actin filament staining, Vero cells were incubated for 10 min at room temperature with Rhodamine Phalloidin (1:1000 dilution, Invitrogen), washed and mounted as described previously for the secondary antibody. Images were processed with the ImageJ software [35].
Synchronization of epimastigote forms of T. cruzi in G1/S of the cell cycle was achieved using hydroxyurea (HU). Cells in exponential growth phase were arrested by incubation with 15 mM of HU for 20–24 h and then released by washing twice with PBS and suspending the cells in culture medium. Cells continued to be cultured for 20 h; samples were taken at the indicated time points and processed as indicated. For flow cytometry analysis, 3x105 cells were harvested by centrifugation, washed with PBS-EDTA 2 mM and fixed in 70% ethanol at -20°C for 30 min. Then, they were washed once with PBS and suspended in staining solution (69 μM propidium iodide, 38 mM citrate buffer pH 7.40, 0.2 mg.ml-1 RNase). The DNA content of propidium iodide-stained cells was analyzed with a fluorescence-activated cell sorting (FACSAria II) analytical flow cytometer (BD Biosciences). Percentages of cells at different phases of the cell cycle were evaluated by Cyflogic software. For microscopic observation of cell cycle progression, synchronized cells were processed as described for the immunolocalization assay, using KMX-1 antibody.
The protein sequences of T. brucei Aurora kinase genes (TbAUK1, TbAUK2 and TbAUK3) were used as baits to search for orthologue sequences in the T. cruzi genome database (http://tritrypdb.org/tritrypdb/). As this database has been made from sequencing the genome of CL Brener strain, a hybrid [36,37] that arose from two different lineages (Esmeraldo-like and Non-Esmeraldo-like haplotypes), we found several putative genes for each TbAUK. When TbAUK1 and TbAUK2 were used as query, two coding sequences (CDS) for each one were found, representing in both cases the same allele for the different haplotypes. In the case of TbAUK3, three CDS were detected, two of them corresponding to the Esmeraldo-like and the other one from the non-Esmeraldo-like haplotype. After an exhaustive sequence analysis, we established that both CDS related to TbAUK1 codify for a single amino acid sequence, and the same result was found with the three CDS related to TbAUK3. Nevertheless, the two CDS related to TbAUK2 showed conspicuous differences, including an insertion of 21 nucleotides. These identified sequences were named TcAUK1, TcAUK2 and TcAUK3, in accordance with the TbAUKs given names (for details see Table 1). In the case of TcAUK2, the two protein sequences were named TcAUK2S (short isoform) and TcAUK2L (long isoform with the 21 nucleotides insertion). Fig 1A shows the sequence alignment of TcAUK2S and TcAUK2L where the dissimilarities and the insertion of seven residues are highlighted. After establishing the CDS for each TcAUK, specific oligonucleotides were designed and used to amplify these genes, using genomic DNA of T. cruzi CL Brener strain as template. The obtained amplification products were then subcloned, sequenced and compared to nucleotide sequences found in the database. While sequenced TcAUK1 was identical to the gene found in the T. cruzi Genome Project database, TcAUK2 as well as TcAUK3 sequences showed some minor discrepancies. Particularly in the case of TcAUK2, the sequencing reactions confirms the presence of the variants TcAUK2S and TcAUK2L in the genome of the parasite. Once the final sequence of each gene was determined, they were annotated in GenBank under the following accession numbers: TcAUK1 EU494590.1, TcAUK2S EU494591.1, TcAUK2L EU494592.1, and TcAUK3 EU494593.1.
A series of multiple sequence alignments (MSA) were performed to analyze the detected Aurora kinase genes of T. cruzi. A MSA of TcAUK proteins with the catalytic domain of human Protein kinase A (PKA) showed that most amino acids with key roles in the kinase activity of PKA are conserved in TcAUKs (S1 Fig). The most relevant is the presence of catalytic glutamic acid of PKA preceded by an Arginine, which is conserved in TcAUKs, allowing to classify them as members of the RD kinases group (S1 Fig, indicated as (1)). In a second MSA, TcAUKs deduced protein sequences were aligned with Aurora kinase proteins from model metazoans. The four TcAUKs present the two most characteristic domains of Aurora proteins: the Activation loop (DFGWSxxxxxxRxTxCGTxDYLPPE) and the Destruction-box (LLxxxPxxRxxLxxxxxHPW). The Threonine residue found in the highly conserved RxT motif from the Activation loop is phosphorylated in the active form of Aurora kinases enzymes (Fig 1B). Described in detail in human Aurora A protein, the 3D active site of this family of proteins involves the Activation loop and the Glycine Rich Loop [38]. In between them is the Hinge region, which along with other residues, forms a hydrophobic pocket where the purine ring of the ATP’s adenosine substrate is located. TcAUKs proteins also show a strong conservation of the main residues implicated in the folding of this hydrophobic pocket. The 3D structure of human Aurora kinase A in complex with ATPγS reported by Nowakowsky and co-workers [39] reveals that in its active state, the αC helix in the protein adopts a position that allows a salt bridge formation between residues Glu181 and Lys162. This salt bridge is essential for the catalytic activity of the enzyme, and it is very close to the β phosphate from de ATPγS ligand. The key residues are conserved in all the Aurora kinase proteins, including the TcAUKs. Finally, a phylogenetic analysis was carried out based on a third MSA of TcAUKs with Aurora kinases proteins of model metazoan and Aurora proteins described in other protozoans. The full-length amino acid sequences were aligned and a maximum parsimony tree was constructed (Fig 1C). As expected, the three TcAUKs group with its orthologues in T. brucei. On the other hand, the longer C-terminal sequences for TcAUK3 and a presence of a large inserted region in the Activation Loop of TcAUK2S and TcAUK2L make these proteins to group separately from the Aurora kinases of other species. Similarly, when a MSA was performed considering only the catalytic domain (S2 Fig) of these proteins, again TcAUK3 and both TcAUK2 grouped apart from the proteins of others organisms. Thereby, is TcAUK1 the one closest related to Aurora proteins of other protozoan and metazoans.
The complete analysis of the TcAUKs sequences show strong evidences that allow us to conclude that these genes most likely code for the T. cruzi aurora kinase functional orthologues.
As described above, Aurora kinases from T. cruzi are represented in the database as more than one CDS. A detailed analysis of the information found in the database suggests that the CDS retrieved for TcAUK1 and TcAUK2 correspond to allelic variants of single copy genes. For TAUK3 the information in the database it is not conclusive about the number of CDSs and their corresponding genomic localization. A Southern blot analysis with specific probes for each TcAUK was performed on genomic DNA of T. cruzi CL Brener strain to experimentally determine the number of copies (Fig 2A). When TcAUK1 or TcAUK2 probes were used, the results confirm that both are single copy genes. Notably, the pattern obtained with the TcAUKs probe on genomic DNA digested using PstI demonstrated the existence of the TcAUK2S and TcAUK2L variants. Endonuclease PstI has two recognition sites within the sequence of the TcAUK2S variant but none inside the TcAUK2L. Consistent with the presence of both variants of TcAUK2, the total number of bands observed in the Southern blot were four: three corresponding to TcAUK2S and one for TcAUK2L (Fig 2A, TcAUK2 panel, line Pst1, asterisks). The Southern blot results obtained using a probe against TcAUK3 showed that when restriction enzymes without recognition sites within the protein coding sequence were used two bands were observed, indicating the presence of more than one copy of this gene. However, when restriction enzymes with one digestion site within the sequence were employed, there was a difference with the expected number of bands. The analysis of the CDS of TcAUK3 and their surrounding sequence obtained from TriTrypDB showed that TcAUK3 is present in three copies per haploid genome with two CDS in a tandem arrange on chromosome TcChr-33 and the third CDS in another contig without chromosome assignation. The tandem CDSs show high sequence identity on its flanking regions (98% identity), indicating the occurrence of a duplication event involving a large chromosomic region in which TcAUK3 is included. Given the difficult interpretation of this Southern blot results, Pulsed Field Gel Electrophoresis (PFGE) to separate intact chromosomes followed by hybridization with a TcAUK3 probe (Fig 2B) was carried out. Probes for TcAUK1 and TcAUK2 were included in this experiment to confirm the Southern blot results. While only one band was observed for TcAUK1 and TcAUK2, supporting results that indicated these are single copy genes, two bands were detected for TcAUK3, confirming the presence of more than one CDS located in different chromosomes. This result, together with the restriction profile observed in the Southern blot analysis, allows us to conclude that TcAUK3 gene is present in three copies per haploid genome of T. cruzi CL Brener strain. Additionally, in a recent publication [40] Dr. Robello and collaborators present the analysis of the genome sequences of two T. cruzi clones TCC (TcVI) and Dm28c (TcI), determined by PacBio Single Molecular Real-Time technology. The assemblies obtained with this technology permitted accurately estimate gene copy numbers. Analyzing this improved genome sequence, it was possible to confirm the presence of three copies of TcAUK3 gene in Dm28c: two at scaffold 196 and one at scaffold 24 (Dr. Robello, personal communication).
After TcAUKs genes were identified, their expression through parasite life cycle was evaluated. The presence of the different TcAUK transcripts was evaluated by RT-PCR in epimastigote, trypomastigote and amastigote forms. Specific reverse primers for each TcAUK were used, while a primer corresponding to the Splice Leader (SL) region was used in all three cases. Amplification products were detected for each TcAUKs in all the three parasite forms (Fig 3A). In order to confirm their identity all amplification products were subjected to sequencing reactions. Sequence data obtained from TcAUK2 RT-PCR products confirms that both variants of this gene–TcAUK2S and TcAUK2L–are transcribed. Furthermore, 5´UTR region shows 100% of identity between variants, supporting the hypothesis that these two forms correspond to alleles inherited from two different parental lineages. In the case of TcAUK3, two specific amplification products of similar length (Fig 3A, I and II) were obtained. Sequencing showed that there a 100 bp stretch in the 5´UTR present in only one of the transcripts (Fig 3B, TcAUK3, II, underline), while the rest of the sequence had 100% identity with the shorter product (Fig 3B, TcAUK3). Thereby, the existence of two amplification products for TcAUK3 with differences in the 5´UTR agrees with what was observed in the Southern blot and the PFGE analysis, indicating the presence of more than one gene copy per haploid genome.
The phylogenetic tree in Fig 1C suggests that TcAUK1 shows greater similarity to metazoan Aurora genes than the rest of the TcAUKs. Since TbAUK1 from T. brucei is regarded as the protozoan counterpart of human Aurora B [23], we focused on unravelling the role of TcAUK1 in T. cruzi biology. Taking into account that in T. cruzi most of the regulation of protein expression occurs at the post-transcriptional level, the existence of TcAUK1 protein was evaluated by immunoblotting using a specific polyclonal antiserum against this protein. TcAUK1 protein was detected in whole cell extracts from epimastigotes, trypomastigotes and amastigotes (Fig 3C), hence consistent with mRNA expression data.
A main characteristic of Aurora kinase proteins is their dynamic redistribution during cell cycle. Human Aurora-B localization during mitosis has been well documented, at the beginning of the mitotic process it is dispersed in the nucleus, after which becomes concentrated in the nuclear midzone and migrates with chromatids to the cell poles, to finally locate in the constriction ring during cytokinesis. To address if TcAUK1 shares this dynamic localization during cell division, this protein was followed along cell cycle in epimastigote forms by immunostaining. DNA of the nucleus and kinetoplast were detected by DAPI staining while mitotic spindle microtubules were labeled using KMX-1 monoclonal antibody. This allowed to identify the different cell cycle stages in an asynchronous population based on the determination of the number of flagella (F), nuclei (N) and kinetoplasts (K), three organelles that duplicate in a tightly coordinate manner. In cells at interphase (1F1K1N), TcAUK1 was observed as a punctuated pattern localized in both extremes of the kinetoplast (Fig 4A, upper panel). The same localization was observed at late G2 when cells have already duplicated their flagellum and kinetoplast (2F2K1N) (Fig 4A middle row panels). In these cells, the two kinetoplasts are arranged in an anteroposterior alignment indicating the beginning of kinetoplasts segregation. During mitosis (2F2K2N), when both kinetoplasts are moving in opposite directions, TcAUK1 was detected inside the nucleus. In cells where the nuclei have not segregated yet, TcAUK1 adopted a circular conformation (Fig 4A lower row panels, yellow arrowhead), whereas in cells where both nuclei were segregating TcAUK1 was observed as an elongated structure (Fig 4A lower row panels, white arrowheads). Furthermore, TcAUK1 adopted the same configuration as the mitotic spindle, as shown by microtubule staining, revealing a close association of TcAUK1 with this structure.
To study TcAUK1 localization in the other forms of T. cruzi, trypomastigote and amastigote forms were collected from culture supernatant of infected Vero cells and subjected to immunostaining with TcAUK1 antiserum (Fig 4B). In amastigotes, TcAUK1 appeared as a single focus close to the kinetoplast, whereas in trypomastigotes no TcAUK1 signal was detected. Given that amastigotes represent the intracellular replicative form of T. cruzi we prompted to study TcAUK1 localization in Trypanosoma cruzi during the progression of cell infections, using Vero as a host cell. In Fig 4C it can be observed that at day 2 after infection, both amastigotes and trypomastigotes were present inside the cytoplasm of Vero cells. Surprisingly, here TcAUK1 was detected not only in amastigotes but also in trypomastigotes and in both cases located inside the parasite nucleus (Fig 4C, 2 days). At days 7 and 11 post-infection, only amastigotes could be observed inside Vero cells, and TcAUK1 was detected as two foci located one at each side of the kinetoplast, similar to what was observed in epimastigote forms during interphase (Fig 4C, 7 days and 11 days).
The above observations indicate that during the cell cycle TcAUK1 shows a changing localization. To accurate define each cell cycle phase, we synchronized epimastigote forms by hydroxyurea (HU) treatment. After release, samples were taken at different time points for flow cytometry and for determine TcAUK1 localization by immunofluorescence. Fig 5 shows that in G1 and S phases, when the cell has 1 nucleus and 1 kinetoplast, TcAUK1 was located at both extremes of the kinetoplast. At the moment that the kinetoplast has duplicated in G2 phase, TcAUK1 was still close to the kinetoplast but it moved out from the extremes and appeared as a single point. Finally, in M phase TcAUK1 was detected inside the dividing nucleus. According to what was observed in Fig 4A, and considering the role that Aurora proteins from other organisms play, during M phase TcAUK1 could be involved in mitotic spindle dynamics and chromatin segregation. Nevertheless, the localization of TcAUK1 in the proximity of the kinetoplast during interphase leads to ask about the role that it is playing in that location.
The cell cycle stage dependent localization in epimastigotes of TcAUK1 is in accordance to the typical behavior of an Aurora B kinase, localizing inside the nucleus and appearing in association with the mitotic spindle during mitosis. Therefore, TcAUK1 could be fulfilling a possible role in spindle organization and chromosome segregation during cell division. The localization at each side of the kinetoplast during interphase is a novel observation that it has not been described in any other kinetoplastid. For this reason, to confirm this result with a different experimental approach, we generated epimastigote forms expressing a TcAUK1-GFP fusion protein by using the episomal, low-expression level vector pTEX. Fig 5B shows that TcAUK1-GFP presents a similar localization pattern to that observed by immunostaining. This novel localization could be reflecting roles of TcAUK1 in process other than nuclear division. In order to address the study of TcAUK1 possible roles, epimastigotes of T. cruzi overexpressing TcAUK1 were generated. Cells were transfected with the construction pTREX-TcAUK1 and stable overexpressing cells were selected under presence of G418 in the culture medium. After the selection period pTREX-TcAUK1 vector integration into the genomic DNA was confirmed by Southern blot analysis of the transfectant pool (Fig 6A). Parasites transfected with an empty pTREX vector were used as a control. Individual clones were isolated by the limiting dilution method, and a single clone in which TcAUK1 overexpression was confirmed by western blot was used for subsequent studies (Fig 6B). Considering that TcAUK1 seems to be involved in mitosis, we hypothesized its overexpression could alter cell cycle progression, modifying the culture growth rate. Three independent cultures of pTREX (control) and pTREX-TcAUK1 epimastigotes were initiated at 1x106 cells.ml-1 and counted for cell density every 24 h for six consecutive days. Fig 6C shows the growth curve for pTREX-TcAUK1 and pTREX epimastigotes cultures, where it can be observed that pTREX-TcAUK1 epimastigotes had slightly lower growth rate. When cell duplication time (DT) was calculated based on the specific growth rate (μ) between days two to five in each culture (Fig 6C, inset), the mean of pTREX parasites was 20.6 (SEM 0.2) hours whereas for TcAUK1 overexpressing epimastigotes was 25.4 (SEM 0.895) hours (statistical significance was determined by Student's Paired t-test, P<0.05). In order to elucidate how cell cycle progression could be affected by TcAUK1 overexpression, hydroxyurea (HU) synchronized cultures were subjected to propidium iodide DNA staining followed by flow cytometry. For this, cultures of pTREX and pTREX-TcAUK1 cells were arrested at G1/S transition by incubating with hydroxyurea (HU). After this treatment cells were released and samples were taken every two hours for cell cycle analysis. Results depicted as histograms in Fig 6D show that both cultures progressed similarly through G1 and S phases (0 to 8 h after HU removal), but when cells entered to G2/M (11 h after HU removal) a delay was detected in TcAUK1 overexpressing cells when compared to pTREX epimastigotes. While at 13 h after-release from HU most of pTREX parasites completed mitosis (56% cells in G1 and 38% in G2/M), at this same time point pTREX-TcAUK1 parasites were still retained in G2/M phase (41% cells in G1 and 50% in G2/M). It was not still up to 14 hs post-release that most of pTREX-TcAUK1 cells complete mitosis (55% cells in G1 and 36% in G2/M), meaning that mitosis concretion in these cells was at least one hour retarded. Despite this observation, at 15 h after HU release, both cultures showed similar profiles in the cell cycle histogram. Thereby, the delay in the normal progression of the cell cycle in TcAUK1 overexpressing parasites supports the idea that TcAUK1 is involved in events occurring during cell division.
A more detailed analysis of the progression through G2/M (11 h to 14 h after HU removal) of synchronized control and TcAUK1 overexpressing cultures was performed (Fig 7A).
Considering the limited information available on T. cruzi mitosis and the wide differences between lab strains, we initially set to analyze the dynamic of organelle duplication in WT epimastigotes. Progression through G2/M is characterized by the organized and timed duplication of 1- the flagellum, 2- the kinetoplast and 3- the nucleus. According to our observations, once the new flagellum is synthesized and protrudes from the flagellar pocket, the kinetoplast starts to duplicate, first increasing its length (Fig 7B, panel I DAPI) and then acquiring a “V” shape with the concave side facing the base of the flagella (Fig 7B, panel II DAPI). After that the kinetoplast “breaks” giving rise to the two daughter kinetoplasts (Fig 7B, panel III DAPI), that localize one behind each other (anteroposterior arrangement). Concurrently, the mitotic spindle starts to form in the nucleus, observed as a circular accumulation of β-tubulin (Fig 7B, panel III KMX-1). Next, while the new flagellum continues increasing its length, both kinetoplasts adopt a side by side configuration (lateral arrangement) and the mitotic spindle assembling begins (polygonal structure) in the nucleus (Fig 7B, panel IV KMX-1 and DAPI). This last event marks the end of the G2 phase and the beginning of mitosis (M). During this phase, the spindle elongates while the nucleus is divided in two (Fig 7B, panel V KMX-1 and DAPI). At the end of the nuclear division, sister nuclei and kinetoplasts continue to migrate in opposite direction while cytokinesis begins with the cleavage furrow formation in the anterior edge of the cell (Fig 7B, panel VI, VII and VIII KMX-1 and DAPI). As a result of this in-depth analysis of G2/M progression, two specific structural characteristics can be considered as landmarks: one is the length of the new flagellum in 2F1N1K cells and the other is the relative localization of the two kinetoplasts in 2F1N2K cells.
The microscopic analysis of the above described events in pTREX and pTREX-TcAUK1 synchronized epimastigotes between 11 h and 14 h after HU removal showed that entry into mitosis was delayed in TcAUK1 overexpressing parasites when compared to control cultures (Fig 7C). This is concluded from the observation that at 11.5 h after HU removal, control cultures showed a significant rise in the 1F1K1N population, indicating that most epimastigotes have undergone cell division. However, at this time point, most parasites in the overexpressing culture corresponded to the 2F1K1N subpopulation, with an augmentation of the 1F1K1N population only after the 12.5 h time point. The number of epimastigotes corresponding to the mitotic configurations 2F2K1N and 2F2K2N remains low throughout the analyzed period both in pTREX and pTREX-TcAUK1 cultures, indicating that this process occurs within a short time span. Moreover, this observation indicates that TcAUK1 overexpression does not alter mitotic events, in which case accumulation of the previously mentioned mitotic configurations would have been observed.
With the aim to confirm the effect of TcAUK1 overexpression on kinetoplast replication, we defined a method that allowed us to time this event with considerable accuracy. As described above, the new flagellum protrusion from the flagellar pocket and the beginning of mitotic spindle assembling (polygonal structure) constitute two independent events that indicate the start and the end of kinetoplast replication, respectively. Considering this, 2F1N1K cells were analyzed for kinetoplast division and new flagellum length, while 2F1N2K cells were examined for mitotic spindle formation and the lateral arrangement of duplicated kinetoplasts. In pTREX epimastigotes, at every time point all cells that displayed the new flagellum protruding from the flagellar pocket also presented kinetoplasts in the process of duplication. However, in pTREX-TcAUK1 cultures, cells bearing a long new flagellum outside the flagellar pocket but with a kinetoplast that had not entered the duplication process could be observed (Fig 7D, panel IV). Indeed at 12.5 h post-release, out of 53 2F1N1K cells, 47 had a protruding flagellum but only 22 showed a dividing kinetoplast. This structural configuration was detected in parasites from TcAUK1 overexpressing cultures obtained at every time point (from 11 to 14 h post release). Cells with 2F1N2K configuration were evaluated for the kinetoplasts arrangement (posterior or lateral) as well as presence or absence of mitotic spindle polygonal structure. The total population of pTREX epimastigotes examined showed the presence of a polygonal structure of the spindle as well as a lateral arrange of the kinetoplasts, as expected according to what was described for WT cells. However, when 2F1N2K TcAUK1 overexpressing epimastigotes were analyzed, 50% of the cells that showed the polygonal structure of the mitotic spindle displayed kinetoplasts still placed in an anteroposterior arrangement at 12.5 h after HU was removed (Fig 7D, panels III and V). This phenotype of 2F1N2K could observed at every time-point in pTREX-TcAUK1 epimastigotes. Moreover, an extreme phenotype of 2F1N1K cells with the polygonal structure of the mitotic spindle but a V-shaped kinetoplast was detected (Fig 7D, panels I and II) in epimastigotes overexpressing TcAUK1.
Taking into account that the function of Aurora kinases is closely related to their localization through the cell cycle, we next studied TcAUK1 dynamics in overexpressing epimastigotes during the different phases of cell cycle, considering kinetoplast and nucleus duplication and mitotic spindle assembly as hallmarks of this process. Interestingly, TcAUK1 in overexpressing cells was present almost exclusively in the nucleus during the entire cell cycle (Fig 8). This result differs from what was observed in WT epimastigotes, in which TcAUK1 localizes in both extremes of the kinetoplast during the interphase (G1 and S phases) and only is detected in the nucleus just before mitosis begins (Fig 5). This observation led us to hypothesize that, when overexpressed TcAUK1 loses its localization at the kinetoplast and this could be affecting replication process of this organelle.
In this work we identified three Aurora kinase genes present in T. cruzi genome: TcAUK1, -2 and -3. By analysis of genomic sequences, we found that TcAUK2 has two different forms differing only in its amino acid sequences by 3%, including a 7-residue insertion in the longer form of the gene (TcAUK2L). By Southern blot and PFGE, we determined that TcAUK3 gene is represented by more than one copy per haploid genome. Moreover, cDNA sequencing shows that transcripts of these copies differ in their 5´-UTR. Considering all this data we postulate that TcAUKs family is not conformed only by 3 members, but 5 members form part of this group of genes. The presence of two different forms of TcAUK2 could indicate that these proteins have different and specific functions. In the case of TcAUK3, considering that gene expression is mainly regulated at the post-transcriptional level in trypanosomatids the differences in the 5´ UTR of the different transcripts identified for these genes could denote differential regulation in their mRNA stability or translation efficiencies, being therefore subjected to selective protein expression regulation. These findings propose TcAUK2 and TcAUK3 as interesting proteins to be studied on the near future. TcAUK1, which is present as a single copy gene, appears as the closest T. cruzi Aurora kinase to metazoans Aurora kinases, based on the phylogenetic analysis. Moreover, its orthologue in T. brucei (TbAUK1) has been reported as the protozoan counterpart of mammalian Aurora B kinase [23]. These led us to focus our efforts on the in-depth characterization of TcAUK1.
After confirming the expression of TcAUK1 in the amastigote, trypomastigote and epimastigote stages (Fig 3), we next aimed to study its localization during epimastigotes cell cycle. By immunofluorescence, we detected TcAUK1 at two discrete cell cycle stage-dependent subcellular localizations: during interphase, it is localized at the extremes of the kinetoplast while in mitosis it resides inside the nucleus, associated with the mitotic spindle (Fig 4A). Furthermore, we extended the localization analysis to trypomastigotes and amastigotes isolated from culture supernatant as well as in intracellular amastigotes. Although the result of the immunoblotting shows that this protein is expressed in trypomastigotes obtained from culture supernatants, TcAUK1 could not be detected in this form of the parasite by immunofluorescence. Nevertheless, when TcAUK1 was labeled in intracellular parasites we found trypomastigote forms that express this protein located in the nucleus (Fig 4C, 2 days after infection). Therefore, it seems that TcAUK1 could be expressed in trypomastigotes at specific times during the infection cycle, probably indicating its participation in specific cellular processes. It is important to highlight that despite the trypomastigote, amastigote and epimastigote are the three forms that have been more deeply studied, during the differentiation that mediates the passage between these forms, the parasite adopts different intermediate stages. This involves a plethora of morphological events occurring in a dynamic way until the cell reaches its final form. According to what we have observed in the trypomastigote form, it is possible to postulate that TcAUK1 could be expressed for a short period of time and participates in events taking place during cell differentiation. Furthermore, the hypothesis of a short-lived TcAUK1 could explain why it can be detected when techniques that evaluate a wide population are used (e.g. western blot) but its detection remains elusive when observing individual events (e.g. immunofluorescence). In amastigotes, similarly to what was observed for epimastigotes, TcAUK1 adopts two different locations. It is found inside the nucleus of intracellular amastigotes at two days after infection (Fig 4C) but it locates at the extremes of the kinetoplast in extracellular amastigotes and intracellular amastigotes at advanced days of infection (Fig 4B and 4C). The fact that TcAUK1 adopts the same nuclear location in trypomastigotes than in amastigotes at early days of infections is another observation that suggests that this protein could be involved in the differentiation process. On the other hand, in actively replicating intracellular amastigotes (advanced cellular infections) and the epimastigote replicative form, the location of TcAUK1 at the extremes of the kinetoplast could indicate that this protein is involved in cell division processes.
The role of TcAUK1 during cell division was studied in detail in the epimastigote form of T. cruzi. TcAUK1 localization alternates between kinetoplast extremes at interphase and the nucleus during mitosis. The arrangement that TcAUK1 adopts in the nucleus (Fig 4A, 2F2K2N cells) in close association with the mitotic spindle and migrating with segregating chromosomes strongly suggests that this protein is the orthologue of human Aurora B. The overexpression of TcAUK1 leads a delay in the duplication time of the epimastigote form, as observed in the growth curves of control and pTREX-TcAUK1 parasites (Fig 6C). The results of cell cycle analysis by flow cytometry on synchronized pTREX and pTRX-TcAUK1 cultures showed that in the later, the time employed to conclude mitosis is longer than in control epimastigotes (Figs 6D and 7A). This result confirms that TcAUK1 plays an important role in cell division. To further investigate this, we analyze organelle division dynamics during G2/M. The prevalence of 2F1K1N population at longer times after HU removal in TcAUK1 overexpressing cultures (Fig 7C) suggests that in these cells kinetoplast duplication is altered. By timing the sequential duplication of flagellum, kinetoplast and nucleus in WT cells (Fig 7B), we noticed that overexpression of TcAUK1 leads to a delay in the start of kinetoplast duplication (Fig 7D). Considering that the role of Aurora proteins is closely related to its localization, a possible explanation for the altered kinetoplast duplication is the fact that overexpressed TcAUK1 lost its location at the extremes of the kinetoplast during cell cycle interphase, as is detected in Fig 8. The subcellular location of TcAUK1 in WT epimastigotes during interphase detected by immunofluorescence is very similar to the localization of the so-called Antipodal sites–a nucleation of proteins that are involved in different kinetoplast processes–in Kinetoplastids [18,41]. Hence, it is very likely that TcAUK1 interacts with different targets that participate in kinetoplast division at these sites. Further experiments should be performed to confirm this hypothesis. It has been widely documented that Aurora kinase B in metazoans is associated with other proteins to conform the so-called CPC. More recently it was found that T. brucei AUK1 (TbAUK1) it is also associated with other proteins conforming a complex and that, like the CPC, this is crucial to guide TbAUK1 across different targets during cell division [25]. Nevertheless, the possible role of Aurora kinases in other cellular processes has not been explored yet. Here, we report for the first time that TcAUK1 is not only involved in mitosis but also in the duplication of the kinetoplast. It is well known that Aurora proteins cellular function is regulated by different post-translational modifications, such as protein phosphorylation and SUMOylation [42–45]. It is possible that TcAUK1 localization in the extremes of the kinetoplast is determined by its interaction with other proteins or by post-translational modifications, or a combination of these. The loss of its localization could be associated with the failure of one or both of these possibilities due to non-physiological expression levels. This hypothesis is supported by the observation that in western blot experiments, in whole cell extracts of pTREX-TcAUK1 epimastigotes only the lower molecular weight band increased its intensity, indicating that there are at least two different states for TcAUK1 (probably corresponding to different post-translationally modified proteins) but only one of them is augmented in overexpressing parasites (Fig 6B). If TcAUK1 needs to be post-translationally modified to interact with specific proteins responsible in order to be recruited to the kinetoplast extremes, then it could be possible that unmodified overexpressed TcAUK1 interacts and sequesters specific proteins, therefore impeding the modified TcAUK1 to reach the kinetoplast.
In summary, in this work we have identified the members of the Aurora kinase proteins in the protozoan T. cruzi. We have demonstrated that TcAUK1 protein behaves as the human Aurora kinase B regarding its role in nuclear division. Also, we have found that this protein is present in the three forms of the parasites life cycle and evidence shows that this protein could play a role in the differentiation between trypomastigote to amastigote forms. Interestingly, we have described a novel function for this protein in direct involvement in the initiation of kinetoplast duplication. This represents the first description of a role for Aurora kinases other than its widely studied function in mitosis. This stands as a valuable contribution in the attempt to understand in a molecular level the complex processes that take place during the life cycle of protozoan organisms.
|
10.1371/journal.pgen.1007113 | PRM1 and KAR5 function in cell-cell fusion and karyogamy to drive distinct bisexual and unisexual cycles in the Cryptococcus pathogenic species complex | Sexual reproduction is critical for successful evolution of eukaryotic organisms in adaptation to changing environments. In the opportunistic human fungal pathogens, the Cryptococcus pathogenic species complex, C. neoformans primarily undergoes bisexual reproduction, while C. deneoformans undergoes both unisexual and bisexual reproduction. During both unisexual and bisexual cycles, a common set of genetic circuits regulates a yeast-to-hyphal morphological transition, that produces either monokaryotic or dikaryotic hyphae. As such, both the unisexual and bisexual cycles can generate genotypic and phenotypic diversity de novo. Despite the similarities between these two cycles, genetic and morphological differences exist, such as the absence of an opposite mating-type partner and monokaryotic instead of dikaryotic hyphae during C. deneoformans unisexual cycle. To better understand the similarities and differences between these modes of sexual reproduction, we focused on two cellular processes involved in sexual reproduction: cell-cell fusion and karyogamy. We identified orthologs of the plasma membrane fusion protein Prm1 and the nuclear membrane fusion protein Kar5 in both Cryptococcus species, and demonstrated their conserved roles in cell fusion and karyogamy during C. deneoformans α-α unisexual reproduction and C. deneoformans and C. neoformans a-α bisexual reproduction. Notably, karyogamy occurs inside the basidum during bisexual reproduction in C. neoformans, but often occurs earlier following cell fusion during bisexual reproduction in C. deneoformans. Characterization of these two genes also showed that cell fusion is dispensable for solo unisexual reproduction in C. deneoformans. The blastospores produced along hyphae during C. deneoformans unisexual reproduction are diploid, suggesting that diploidization occurs early during hyphal development, possibly through either an endoreplication pathway or cell fusion-independent karyogamy events. Taken together, our findings suggest distinct mating mechanisms for unisexual and bisexual reproduction in Cryptococcus, exemplifying distinct evolutionary trajectories within this pathogenic species complex.
| Sexuality is ubiquitous in eukaryotic systems, but it is present in diverse forms, ranging from distinct sexual individuals to parthenogenic organisms in both animals and plants. Consequently, different organisms have evolved different reproduction strategies in which cell-cell fusion and nuclear fusion (karyogamy) play fundamental roles. The opportunistic human fungal pathogen Cryptococcus neoformans can undergo both bisexual reproduction between a and α cells and selfing unisexual reproduction, which offsets the cost of finding a mating partner, coinciding with the observation that 99% of clinical and environmental isolates are mating type α. It has been a central interest to elucidate the similarities and differences between these two sexual cycles. Here, we identified and characterized two genes in the Cryptococcus species complex, PRM1 and KAR5, which play conserved roles in plasma membrane fusion and karyogamy in fungi. We showed that unisexual reproduction is largely independent from cell-cell fusion and is mechanistically different from bisexual reproduction. We also demonstrated that karyogamy takes place at different stages during bisexual reproduction between two sister species, exemplifying distinct evolutionary trajectories within the pathogenic species complex.
| Sexual reproduction is ubiquitous in eukaryotic systems and promotes genetic diversity important for successful evolutionary adaptation to ever-changing environments [1]. In addition to bisexual reproduction between mating partners of opposite sexes, many eukaryotic systems, including fish, amphibians, and reptiles, can undergo unisexual reproduction, termed parthenogenesis, often in the absence of the opposite sex [2]. During bisexual reproduction, parental gametes undergo cell fusion and nuclear fusion to produce recombinant progeny, whereas during parthenogenesis, the maternal genome undergoes reduplication through either cell-cell fusion or endoreplication to produce clonal offspring of the mother [2]. Analogous to parthenogenesis, several human fungal pathogens have been reported to undergo both unisexual and bisexual reproduction [3, 4]. In Candida albicans bisexual reproduction, a/a and α/α cells first undergo white-opaque switching to become mating competent and then form tetraploid cells via cell fusion and nuclear fusion. These cells then undergo a parasexual cycle to return to the diploid state. During C. albicans unisexual reproduction, loss of the Bar1 protease in a/a cells enables auto-response to MFα pheromone and promotes cell and nuclear fusion producing tetraploid cells [5]. During bisexual reproduction in the Cryptococcus species complex, cell fusion triggers a dramatic yeast-hyphal morphological transition, producing dikaryotic hyphae. The growing tips of these hyphae differentiate into basidia, in which two nuclei undergo nuclear fusion to produce basidiospores through meiosis [6]. During the unisexual cycle, α or a cells initiate hyphal growth and form monokaryotic hyphae, during which the haploid nucleus undergoes a ploidy increase through either cell-cell fusion followed by nuclear fusion, nuclear fusion between mother and daughter cells, or an endoreplication pathway, and the diploid nucleus inside the basidium then undergoes meiosis and produces haploid spore progeny [7, 8].
Sexual reproduction has only been observed under laboratory conditions In the Cryptococcus species complex. However, spore-like cells have been harvested from the environment, suggesting the sexual cycle may occur in natural environments [9, 10]. Unisexual reproduction has been documented for C. neoformans, C. deneoformans, and C. gattii [7, 11, 12]. Based on evidence from population genetics studies, natural isolates also recombine through unisexual reproduction, which may be of ecological significance because more than 99% of environmental and clinical isolates are the α mating type [13–16]. Of note, the unisexual cycle generates genotypic and phenotypic diversity de novo, similar to the bisexual cycle [17]. A common set of genetic circuits govern both unisexual and bisexual reproduction, [8, 18–20] and both sexual cycles involve similar meiotic recombination mechanisms [21]. The recombining nature of the unisexual cycle can enable a clonal population to reverse Muller’s ratchet and avoid an evolutionary dead end [22].
Despite similar regulatory genetic circuits, fundamental differences are obvious between the two modes of sexual reproduction [23–25]. Genetically, the unisexual cycle is initiated in the absence of an opposite-mating type partner, whereas the bisexual cycle is initiated upon a-α cell-cell fusion. Morphologically, the unisexual cycle produces monokaryotic hyphae with unfused clamp cells, while the bisexual cycle produces dikaryotic hyphae with fused clamp cells, which allow a nucleus to migrate between adjacent hyphal compartments to maintain dikaryotic hyphae [24, 25]. While diploidization is achieved through karyogamy in the bisexual cycle, it is not yet clear how diploidization is achieved during the unisexual cycle. Three hypotheses have been proposed, including 1) cell fusion followed by karyogamy; 2) karyogamy between mitotically dividing mother-daughter cells followed by either mis-segregation of the nucleus or cytokinesis arrest; and 3) endoreplication during hyphal growth [25, 26].
In all bisexually reproducing organisms, gamete fusion is a fundamental process requiring a set of dedicated fusion proteins [27]. In the fungal kingdom, Prm1 (Pheromone regulated multi-spanning membrane protein 1) is a conserved plasma membrane protein required for plasma membrane fusion during cell-cell fusion [28–30]. In Saccharomyces cerevisiae and Neurospora crassa, deletion of PRM1 reduces fusion frequency by approximately half and leads to cell lysis. The mutant phenotype is alleviated in the presence of a high calcium and exacerbated upon calcium depletion [31, 32]. Prm1 is also required for asexual hyphal fusion in N. crassa [29]. In Schizosaccharomyces pombe, deletion of PRM1 causes a 95% reduction in cell fusion frequency independent of extracellular calcium concentration, but does not lead to a cell lysis phenotype [30].
Cell fusion has been well studied in Cryptococcus sexual cycles. During bisexual reproduction, a-α cell-cell fusion is required for hyphae induction and clamp cell-hyphal fusion is required for proper nuclear migration between adjacent hyphal compartments to maintain dikaryotic hyphal growth [6, 33]. During unisexual reproduction, α-α cell-cell fusion occurs at a low frequency whereas the presence of a cells can enhance α-α cell fusion ~1000 fold in a ménage à trois fashion [7]. G proteins in the pheromone response pathway are required for cell-cell fusion [34], and the master transcription factor Mat2 governs the yeast-hyphal morphological transition [18]. An evolutionarily conserved Ire1 kinase/endoribonuclease in the unfolded protein response pathway has been shown to negatively regulate the pheromone response pathway and is required for cell-cell fusion [35]. However, genes that are directly involved in plasma membrane fusion during cell-cell fusion have not been identified. A transcriptomic study showed that expression of the S. cerevisiae PRM1 homolog in C. deneoformans is highly upregulated during hyphal growth, suggesting it may function in the sexual cycle, but its involvement in cell-cell fusion had yet to be determined [18].
Karyogamy is an essential step for intermixing of parental genetic information during sexual reproduction. Two sets of genes regulate karyogamy in S. cerevisiae. The class I genes, including KAR1, KAR3, KAR4, and KAR9, regulate nuclear congression, while the class II genes, including KAR2, KAR5, KAR7, KAR8, and PRM3, mediate inner and outer nuclear membrane fusion [36, 37]. Lee and Heitman identified the Cryptococcus karyogamy genes KAR2, KAR3, KAR4, KAR7, and KAR8 based on homology to S. cerevisiae [38]. While homologs of KAR2 and KAR7 were identified in Cryptococcus with roles in filamentation and meiosis, respectively, homologs of KAR3, KAR4, and KAR8 did not show karyogamy defects during unisexual or bisexual reproduction. This suggests that these genes are either rewired in Cryptococcus compared with S. cerevisiae or are functionally redundant in regulating nuclear fusion. KAR2, an ER-resident chaperone protein, is essential in Cryptococcus, and its overexpression partially rescues the filamentation defect of the ire1 mutant [35, 39]. KAR7 maintains a conserved role in mediating nuclear membrane fusion during both Cryptococcus unisexual and bisexual reproduction. However, a diploid strain without KAR7 produced hyphae and basidia but failed to undergo sporulation, suggesting KAR7 may play additional roles in meiotic processes. In S. cerevisiae, Kar5 localizes to both inner and outer nuclear membranes at the spindle pole body, and coordinates the outer and inner nuclear membrane, facilitating the inner nuclear membrane fusion step during karyogamy [40–42]. However, a KAR5 homolog was not identified in Cryptococcus. A study on the Chlamydomonas nuclear fusion gene GEX1 by Ning and colleagues [43] showed that protist and plant GEX1 genes and fungal KAR5 genes belong to an ancient cysteine rich domain (CRD) containing protein family that is conserved throughout eukaryotes, suggesting that they may share a conserved role in nuclear membrane fusion. In that same study, a KAR5 ortholog was identified for a basidiomycetous fungus, Puccinia graminis [43].
In this study, we identified PRM1 and KAR5 orthologs in both C. neoformans and C. deneoformans and investigated their conserved functions in mediating plasma membrane and nuclear membrane fusion. Utilizing these two genes, we studied cell fusion and nuclear fusion in the C. neoformans bisexual cycle and the C. deneoformans unisexual and bisexual cycles. C. neoformans and C. deneoformans bisexual cycles were dependent on cell and nuclear fusion at different stages during sexual development, whereas, cell fusion was largely dispensable in the solo unisexual cycle of C. deneoformans and the ploidy duplication during unisexual reproduction is dependent on either endoreplication or cell fusion-independent karyogamy events. Our results provide mechanistic insights relevant to studies of mating mechanisms of unisexual reproduction and parthenogenesis in other eukaryotic systems.
To study cell-cell fusion during the Cryptococcus sexual cycles, we performed BLASTP searches to identify plasma membrane fusion protein, Prm1, known to orchestrate cell-cell fusion during mating in other fungi. BLASTP searches using S. cerevisiae, C. albicans, Aspergillus fumigatus, S. pombe, and N. crassa Prm1 protein sequences [28–30] identified CNAG_05866 (Cn Prm1) and CNF01070 (CdPrm1) as candidate PRM1 genes in C. neoformans and C. deneoformans, respectively (S1A and S1B Fig). The CnPrm1 and CdPrm1 proteins share 91% sequence identity and are the only candidate proteins that shared significant sequence similarity with Prm1 proteins from other fungal organisms. Reciprocal BLASTP searches confirmed the orthologous nature of these fungal PRM1 genes. Both CnPrm1 and CdPrm1 are predicted to share a similar protein topology with ScPrm1 and SpPrm1, and contain four transmembrane domains based on Phobius prediction [44]. However, the Cryptococcus Prm1 proteins have a long C-terminal tail following the last transmembrane domain (Fig 1A).
Another crucial cellular process during sexual reproduction is karyogamy, the fusion of nuclei. One of the karyogamy proteins in S. cerevisiae, Kar5, facilitates nuclear membrane fusion during mating [40, 42]. We identified CNAG_04850 as the KAR5 gene in C. neoformans using the Kar5 protein sequence of Puccinia graminis [43], which belongs to the same phylum (Basidiomycota) as Cryptococcus. The same BLASTP search failed to identify the CdKAR5 gene, but using the CnKAR5 genomic sequence we identified an unannotated region on chromosome 10 from bp 790071 to 792560 that encodes the KAR5 ortholog in C. deneoformans. BLASTP searches and phylogenetic analyses of Kar5 proteins from several fungal organisms suggested that Kar5 protein sequences are divergent across different fungal species (S1C and S1D Fig). Multiple sequence alignment and topology predictions by Phobius prediction and COILS/PCOILS confirmed that CnKar5 and CdKar5 share a similar protein topology with ScKar5 and SpKar5, with an N-terminal signal peptide and a CRD domain, followed by coiled-coiled domains and a C-terminal transmembrane domain, except that SpKar5 does not have the N-terminal signal peptide (Fig 1B and S1E Fig) [44, 45].
Deletion of PRM1 caused a significant filamentation delay during C. neoformans bisexual reproduction (Fig 2A). However, abundant hyphal production and sporulation were still observed after 10 days (S2A Fig). To evaluate the overall impact of PRM1 deletion on overall mating progress, we quantified the relative spore production of prm1 mutants compared to the wild type at 7 days by Percoll gradient centrifugation. Deletion of PRM1 caused a mild reduction in spore production (87.3 ± 9% of wild type, p = 0.207) (S3A Fig).
We conducted a wild type mating between CF757 (JEC20a URA5-NAT) and CF762 (JEC21α ADE2-NEO) as a control. A total of 47 spore derived colonies were randomly chosen and analyzed (S4A Fig). Among the 47 progeny, all eight genotypes of Mendelian inheritance were recovered at a distribution of frequency ranging from 2.1% to 23.4% (17% parental genotype MATa URA5-NAT, 2.1% for parental genotype MATα ADE2-NEO, 12.8% for MATα URA5-NAT, 23.4% for MATa ADE2-NEO, 6.4% for MATa URA5-NAT ADE2-NEO, 6.4% for MATα URA5-NAT ADE2-NEO, 17% for MATa, and 14.7% for MATα) (S4B and S4C Fig). This provides evidence that the cells isolated by Percoll gradient centrifugation are indeed spores.
To address the involvement of Prm1 in cell-cell fusion, we performed cell fusion assays using two genetically marked mating partners. prm1 mutants showed a bilateral (prm1Δ X prm1Δ) cell fusion defect with a fusion frequency of 12% ± 4% relative to the wild type level (Fig 2B), but no defect in unilateral (prm1Δ X WT) cell fusion. The basal level of cell fusion activity may allow prm1 mutants to produce abundant hyphae after a 10-day incubation on mating inducing medium (S2A Fig).
During C. neoformans bisexual reproduction, the dikaryotic hyphae generate clamp cells, which fuse with adjacent hyphal compartments to allow a nucleus to translocate between hyphal compartments and maintain the dikaryon status [6]. To test whether Prm1 plays a role in clamp cell-hyphal fusion, we examined hyphae by scanning electron microscopy (SEM). The clamp cell and a peg from the adjacent hyphal compartment both exhibited elongated tubular morphology in prm1 mutants compared to clamp cell connections in the wild type (Fig 2C), suggesting that these clamp cells and peg protrusions failed to undergo cell fusion. Transmission election microscopy (TEM) showed that the plasma membranes failed to undergo fusion in the clamp cells (S5 Fig). DAPI staining of hyphal nuclei showed that a single nucleus was trapped in the prm1 mutant clamp cells, resulting in an abnormal number of nuclei in a single hyphal compartment (Fig 2D).
Clamp cell fusion is regulated by the pheromone signaling pathway, both PRM1 and MFα expression were maintained at a significantly high level after mating for seven days on mating inducing V8 medium compared to non-mating inducing YPD medium (3.4-fold increase for PRM1, p <0.005; and 50.8-fold increase for MFα, p <0.005) (Fig 2E and 2F). prm1 mutants exhibited a significant increase in MFα expression compared to wild type (1.9-fold increase, p <0.005), suggesting that the cell fusion defect dampens MFα repression that occurs in response to SXI1α-SXIa repression following nuclear pairing (Fig 2F). These results indicate that Prm1 plays a role in both cell-cell fusion and clamp cell-hyphal fusion during C. neoformans bisexual reproduction (Fig 3G).
Like prm1 mutants, kar5 mutants showed a significant delay in filamentation during C. neoformans bisexual reproduction (Fig 3A); the mutants produced abundant hyphae after 10 days (S2B Fig). In contrast to other prm1 mutant phenotypes, kar5 mutants were not defective in cell fusion but exhibited sporulation defects (Fig 3A and 3B). SEM studies showed that the abnormal basidia were either bald or had more than four budding sites compared to the four sites in the wild type (Fig 3C). However, the wild type phenotype (four spore chains) was observed in kar5 mutants after longer mating incubation periods. Similar to prm1 mutants, deletion of KAR5 caused a mild reduction in spore production (77.2% ± 8.8% p <0.05) (S3B Fig), suggesting that deletion of KAR5 did not completely block sporulation. We stained the nuclei within the abnormal basidia generated by kar5 mutants with DAPI and found two nuclei in close contact within the kar5 mutant bald basidia in contrast to either one nucleus or four meiotic nuclei present in wild type basidia (Fig 3D and 3E). Quantification of 129 wild type basidia and 131 kar5 mutant basidia stained with DAPI showed that 5.7% wild type basidia versus 48.9% kar5 mutant basidia contained two nuclei, suggesting that deletion of KAR5 inhibited, but did not completely block karyogamy inside the basidia (Fig 3E). The nuclear morphology of the C. neoformans kar5 mutant was similar to the kar5 mutant karyogamy phenotype in S. cerevisiae [42], supporting the hypothesis that KAR5 plays a conserved role in mediating karyogamy during C. neoformans bisexual reproduction. KAR5 expression was upregulated upon mating induction and maintained at a significantly high level after mating for a week compared to non-mating inducing conditions (1.6-fold increase, p <0.05). Deletion of PRM1 significantly reduced KAR5 expression (1.6-fold decrease, p <0.05), suggesting control of gene expression following cell-cell fusion during C. neoformans bisexual reproduction (Fig 3F and 3G).
In contrast to C. neoformans, C. deneoformans prm1 mutants showed a mild delay in hyphal production (Fig 4A), and exhibited a significant reduction in spore production compared to wild type C. deneoformans (27% ± 2.2%) bisexual reproduction (S3A Fig). PRM1 deletion caused both bilateral and unilateral cell fusion defects with fusion frequencies of 6.9% ± 2.6% and 8.2% ± 1.8% of the wild type levels, respectively (Fig 4B). To understand the mechanistic requirement for Prm1 in cell-cell fusion during C. deneoformans bisexual reproduction, we monitored cell-cell fusion of prm1 mutants with confocal microscopy. In the same prm1 mutant cell fusion sample, both fused and unfused cells were detected by the presence and absence of inter-cellular mixing of fluorescent signals between the Nop1-GFP and mCherry labeled fusion pairs (Fig 4C, S1 and S2 Movies). Based on quantification of fluorescent signal intermixing, the wild type cell fusion frequency was 90.6%, while the prm1 mutant unilateral cell fusion and bilateral cell fusion frequencies were 51.7% and 12.8%, respectively (Fig 4D and S6 Fig). The unilateral cell fusion defect suggests that the Prm1 plays a more important role during bisexual reproduction in C. deneoformans compared to C. neoformans; in agreement with this observation, PRM1 expression level was maintained at a similar high level after mating for seven days compared with 36 hours (Fig 4E).
To visualize the structures of the plasma membrane at the conjugation sites between the fusion pairs, we stained both wild type and prm1 mutant fusion pairs with the lipophilic dye FM4-64. The prm1 mutant fusion pairs exhibited robust staining of the plasma membrane boundaries at the conjugation site (Fig 4F). Compared to the fused wild type cells, the plasma membrane of the prm1 mutant fusion pairs failed to undergo membrane fusion at the conjugation sites, and a layer of cell wall material was present between the plasma membranes (Fig 4G and S7A Fig). In 2 out of 20 observed fusion pairs by TEM, the plasma membranes formed extensive invaginations into the opposite cytosolic compartments without membrane fusion (Fig 4G). Similar to C. neoformans bisexual reproduction, prm1 mutants also exhibited a clamp cell-hyphal fusion defect during C. deneoformans bisexual reproduction (S7B Fig). However, both wild type and prm1 mutant crosses produced hyphae with unfused clamp cells, which are characteristics of monokaryotic hyphae (S7B and S8A Figs). Overall, these results suggest that PRM1 plays a more significant role in C. deneoformans bisexual reproduction in comparison to C. neoformans.
Like prm1 mutants, kar5 mutants showed a mild delay in hyphal production, and produced significantly fewer spores compared to wild type (9.5% ± 2.7%) (Fig 5A and S3B Fig). SEM studies demonstrated that C. deneoformans kar5 mutants produced basidia with abnormal sporulation patterns during bisexual reproduction, similar to C. neoformans kar5 mutants (Fig 5B). Deletion of KAR5 caused both bilateral and unilateral cell fusion defects with fusion frequencies of 32.3% ± 6% and 24.5% ± 2.7% compared to the wild type level, respectively (Fig 5C). To test whether KAR5 is directly involved in cell-cell fusion, we quantified cell-cell fusion events for kar5 mutants based on fluorescent signal intermixing, and found that kar5 mutant cell-cell fusion frequency was 86.4%, similar to wild type (Fig 4D and S6D Fig), suggesting that Kar5 plays a role in post-fusion survival mechanisms for cell fusion products, and that the function of CdKar5 in bisexual reproduction is diverged from CnKar5 (Fig 3B, and 5C). Furthermore, CdKAR5 and CdMFα expression were upregulated upon mating induction, but returned to basal level after mating for seven days, which is distinct from CnKAR5 and CnMFα expression patterns (Figs 2F, 3F, 5D and 5E). However, CdKAR5 expression was significantly reduced for Cdprm1 mutants compared to wild type (2.7-fold decrease, p <0.05) (Fig 5D), similar to C. neoformans (Fig 3F).
To elucidate the phenotypic differences of prm1 and kar5 mutants during bisexual reproduction between C. neoformans and C. deneoformans, we stained wild type and mutant hyphal nuclei with DAPI. In contrast to the dikaryotic hyphae produced by C. neoformans (Fig 2D), C. deneoformans bisexual reproduction produced monokaryotic hyphae (S8A Fig), similar to those produced during C. deneoformans unisexual reproduction (S8B Fig). To dissect the involvement of Prm1 and Kar5 in monokaryotic hyphae formation during C. deneoformans bisexual mating, we tracked nuclear dynamics using the nucleolar marker Nop1-GFP [38]. During early bisexual mating at 48 hours, wild type produced both monokaryotic and dikaryotic hyphae (Fig 5F and S9A Fig), whereas prm1 mutants mainly produced monokaryotic hyphae (S9A Fig). In both wild type and kar5 mutant hyphae, pairs of congressed nuclei were observed, resembled the C. neoformans kar5 mutant karyogamy phenotype inside basidia during bisexual reproduction (Figs 3D and 5F and S9A Fig). After 10 days, monokaryotic and dikaryotic hyphae were present in the wild type cross, while prm1 mutants mainly produced monokaryotic hyphae and kar5 mutants mainly produced dikaryotic hyphae (S9B Fig). After six weeks, wild type and prm1 mutants mainly produced monokaryotic hyphae, whereas, kar5 mutants produced both monokaryotic and dikaryotic hyphae (S9C Fig). Live cell imaging of hyphal nuclear morphology suggests the following: 1) karyogamy may take place early in bisexual reproduction in C. deneoformans; 2) deletion of PRM1 leads to monokaryotic hyphae formation; and 3) deletion of KAR5 blocks early karyogamy in fused cells and could explain the observed post-fusion survival defect for the fused cells, which in turn promoted dikaryon hyphae formation.
To confirm that karyogamy occurs early during C. deneoformans bisexual reproduction, we followed the hyphal nuclear dynamics between mating partners labeled with fluorescent markers (nucleolar marker Nop1-GFP and nuclear marker H3-mCherry). We observed nuclear congression in fused a-α cells (Fig 6A); and a single nucleus labeled with both fluorescent protein markers was observed, confirming that karyogamy can occur immediately after cell fusion (Fig 6B). We also observed both dikaryotic hyphae with fused clamp cells and monokaryotic hyphae with unfused clamp cells during the early mating process (Fig 6C and 6D). These hyphae expressed both parental fluorescent markers, indicating that karyogamy can occur at different stages during C. deneoformans bisexual reproduction. To test whether Kar5 functions in karyogamy immediately after cell fusion and deletion of KAR5 leads to dikaryon formation, we quantified monokaryon and dikaryon fusion products and mature hyphae labeled with both nuclear fluorescent markers in wild type and kar5 mutant crosses (Fig 6E and 6F). Among 125 wild type and 126 kar5 mutant cell fusion products, 60.8% wild type versus 22.2% kar5 mutant fused cells were monokaryotic (Fig 6E). Among 133 wild type and 132 kar5 mutant mature hyphae, 68.4% wild type versus 36.4% kar5 mutant mature hyphae were monokaryotic (Fig 6F). These results confirmed that deletion of KAR5 inhibited, but did not completely block karyogamy in early cell fusion products, and promoted dikaryon formation.
In contrast to bisexual reproduction, deletion of PRM1 and KAR5 in C. deneoformans did not impact filamentation during solo unisexual reproduction (Fig 7A), and caused less reduction in spore production relative to the wild type level (57.3% ± 7.2% and 52.8% ± 4.6%, respectively) (Fig 7A and S3 Fig). PRM1 and MFα expression were upregulated upon mating induction, but, KAR5 expression was maintained at a low level and was not affected by pheromone induction (Fig 7B), suggesting KAR5 may play a less important role in solo unisexual reproduction. Similar to what is seen during bisexual reproduction in C. deneoformans, PRM1 expression was maintained at a significantly high level after mating for seven days compared to non-mating inducing conditions (6.2-fold increase, p <0.005), whereas pheromone signaling subsided to basal level, indicating that PRM1 expression is not tightly coordinated with the pheromone signaling pathway in C. deneoformans (Fig 7B). Although PRM1 expression was significantly upregulated, cell fusion occurred at a 1000-fold lower frequency during unisexual reproduction in C. deneoformans compared to both C. neoformans and C. deneoformans bisexual reproduction (Fig 7C). Among cells that underwent cell-cell fusion during unisexual reproduction, deletion of PRM1 and KAR5 caused both bilateral and unilateral cell fusion defects (Fig 7D), and deletion of KAR5 produced basidia with abnormal sporulation patterns (Fig 7E), which were also observed during bisexual reproduction. These results suggest that during unisexual reproduction, a minority of cells undergo α-α cell fusion followed by karyogamy, similar to C. deneoformans bisexual reproduction.
Although cell fusion is largely dispensable for solo unisexual reproduction, karyogamy may function independently of cell fusion between mother and daughter cells or inside basidium. Deletion of KAR7 has been indicated to block nuclear congression inside the basidium during unisexual reproduction [38]. To test whether KAR5 has similar functions, we stained wild type, kar5 mutant, and kar7 mutant basidia with DAPI. Interestingly, all strains produced basidia with one, two, or more than two nuclei, which may represent three different stages of meiosis inside basidia (one nucleus as pre-meiosis, two nuclei as post meiosis I, and more than two nuclei as post meiosis II) (S10A Fig). Among 114 wild type, 116 kar5 mutant, and 115 kar7 mutant basidia, only 1.8% wild type, 4.3% kar5 mutant, and 1.7% kar7 mutant basidia contained two nuclei (S10B Fig), which is different from the cnkar5 mutant with 48.9% basidia containing two pre-karyogamy nuclei during bisexual reproduction (Fig 3E), suggesting that nuclear fusion occurs differently during unisexual reproduction of strain XL280α. If KAR5 and KAR7 were required for karyogamy in the basidia, we would have expected to see a higher population of basidia with 2 nuclei trapped at a pre-karyogamy stage compared to wild type. However, wild type and kar5 mutants exhibited similar basidia nuclear morphology with few two nuclei basidia (S10B Fig), indicating that KAR5 is not required for a later stage of unisexual reproduction, and nuclear fusion is not occurring inside the basidium. The kar7 mutant produced 24.3% basidia versus 60.5% basidia in wild type with more than two nuclei (S10B Fig), suggesting that KAR7 plays a role in meiosis during unisexual reproduction, supporting the previous observation that a diploid kar7/kar7 mutant has a defect in sporulation [38]. To validate these results, we examined basidia nuclear morphology based on nuclear fluorescent signals of wild type (CF836), kar5 mutant (CF718), and kar7 mutant (CF1442) cells labeled with Nop1-GFP, and observed similar results (S11 Fig), supporting the hypothesis that karyogamy occurs at a low frequency and karyogamy defects do not impact basidia nuclear morphology during solo unisexual reproduction.
Given that cell fusion is dispensable in solo unisexual reproduction, and kar5 is not required for meiotic basidia formation, we aimed to confirm that meiosis was involved during spore production. We generated prm1 spo11 and kar5 spo11 double mutants and observed two short spore chains compared to the four long spore chains produced by prm1 and kar5 single mutants (S12 Fig). The lack of normal spore chains confirms that spore production in unisexual reproduction is indeed dependent on the key meiotic gene SPO11 as shown previously [8].
It is unclear how diploidization occurs during solo unisexual reproduction. By following mating partners of the same mating type labeled with different fluorescent markers (nucleolar marker Nop1-GFP and nuclear marker H3-mCherry), we showed that hyphae frequently originated from single cells rather than as of α-α cell fusion products (S3 Movie), which further confirmed that cell fusion is dispensable for the solo yeast-hyphal morphological transition.
To understand when and where diploidization takes place during solo unisexual reproduction, we dissected nascent blastospores from the growing hyphae (Fig 8) and analyzed their ploidy by FACS (Table 1 and S3 Table). In the wild type, 66 out of 71 blastospores dissected from eight budding sites germinated with a survival rate of 93%. FACS analysis of 16 blastospore derived colonies, including two from each budding site, showed that all were diploid (Fig 8). prm1 and kar5 mutants exhibited blastospore germination defects with survival rates of 19.2% and 62.1% respectively. prm1 spo11 and kar5 spo11 double mutants exhibited blastospore survival rates of 63.3% and 92.5% respectively. FACS analysis revealed all blastospores of the prm1 mutant, 12 blastospores from 6 budding sites (four blastospores from two budding sites failed to germinate) of the prm1 spo11 double mutant, and 14 blastospores from seven budding sites (two blastospores from one budding site failed to germinate) of the kar5 mutant were diploid. In analysis of 19 blastospores from 10 budding sites in the kar5 spo11 double mutant, 6 blastospores were haploid, and 13 were diploid (Fig 8). The six haploid blastospores were dissected from three budding sites, suggesting that blastospores originating from the same budding site may have the same ploidy composition. To infer whether the observed single nucleus in the kar7 mutant basidia might be a product of karyogamy (S10 and S11 Figs), we dissected 248 blastospores from 46 budding sites for the kar7 mutant, and only 16 blastospores from 10 budding sites germinated with a survival rate of 6.45%, suggesting Kar7 is required for wild type blastospore survival (Table 1 and S3 Table). Among 15 blastospores analyzed, 9 were diploid, 3 were haploid, and 3 were aneuploid (Table 1 and S13 Fig), suggesting that the nuclei inside kar7 mutant basidia are likely largely diploid and that diploidization occurs earlier and outside of the basidium. The limited sample size of dissected blastospores presented here may explain why a few haploid blastospores were only recovered from the kar5 spo11 double mutant and kar7 mutant but not from wild type or the other mutant strains. That 74 out of a total of 86 (86%) tested blastospores were diploid suggests that diploidization occurs early in the hyphae during unisexual reproduction, and this process may be dependent on an endoreplication pathway or early karyogamy events between mother and daughter cells or inside the growing hyphae. Prm1 and Kar5 were dispensable for diploidization, but may contribute to blastospore survival, implying that Prm1 and Kar5 could have additional cellular functions.
Without an obligate requirement for a mating partner, unisexual reproduction mitigates the two-fold cost of bisexual reproduction in finding an opposite mate. However, lacking genome diversity, clonal unisexual reproduction could be considered an evolutionary dead-end. In Cryptococcus, this assumption is challenged, as unisexual reproduction can generate genotypic and phenotypic diversity de novo by forming aneuploid progeny through meiosis [17]. Given that more than 99% of the natural isolates are α mating type, the presence of a unisexual cycle allows a clonal population to adapt to changing environments, which provides ecological significance to the Cryptococcus pathogenic species complex [46]. In this study, we demonstrated that a small population of cells undergo cell-cell fusion and nuclear fusion during unisexual reproduction, which enables recombination between cells of the same mating type. In response to selection pressures in the environment, the cell fusion dependent unisexual reproduction could facilitate selection of beneficial alleles in a large same sex population and reverse Muller’s ratchet [22]. Same sex cell-cell fusion can be further stimulated by the presence of small population of the opposite mating type [7]. Besides the similar ecological benefits conferred by unisexual and bisexual reproduction, many studies have shown that both modes of sexual cycles share a common signaling network that regulates the yeast-to-hyphal morphological transition and meiotic recombination [25, 26, 47]. Despite the similarities, there are key mechanistic differences between the two. In this study, we focused on two key cellular processes involved in sexual reproduction, cell-cell fusion and nuclear fusion, and studied their involvement in unisexual and bisexual reproduction in two sister Cryptococcus species harboring different sexual cycles.
Cryptococcus orthologs of the S. cerevisiae cell fusion gene PRM1 perform conserved roles during Cryptococcus sexual reproduction. Prm1 facilitates cell fusion between a and α mating partner cells, cell fusion between α-α cells, and clamp cell-hyphal fusion during dikaryotic hyphal growth. During C. neoformans bisexual reproduction, deletion of PRM1 caused a bilateral (prm1Δ X prm1Δ) cell fusion defect, which is similar to what has been observed in S. cerevisiae and N. crassa [29, 31]. However, during C. deneoformans bisexual reproduction, deletion of PRM1 caused both unilateral (prm1Δ X WT) and bilateral (prm1Δ X prm1Δ) cell fusion defects, suggesting that Prm1 plays a more significant role in C. deneoformans.
Cell-cell fusion and clamp cell-hyphal fusion in Cryptococcus is analogous to cell fusion between conidial anastomosis tubes and hyphal fusion in filamentous fungi [48, 49]. Like in S. cerevisiae, N. crassa, and S. pombe, deletion of PRM1 resulted in plasma membrane curvature at the membrane merger site (Fig 4D), but these membranes were separated by a layer of cell wall (Fig 4G and S6 Fig), similar to the prm1 mutant phenotype in S. pombe. Although deletion of PRM1 caused a cell fusion defect, it did not completely block cell fusion in Cryptococcus, suggesting that Prm1 is not the sole membrane fusion protein. Additional candidate cell fusion genes have been identified in S. cerevisiae and N. crassa, including FIG1, LFD1, and LFD2; but BLASTP searches failed to identify homologs of these genes in Cryptococcus [31, 32]. Prm1 may be the evolutionary conserved core component for cell fusion in the fungal kingdom, and species-specific plasma membrane fusion machinery may have evolved independently.
Similarly, the Cryptococcus karyogamy machinery has been previously shown to function differently than that of S. cerevisiae [38]. Deletion of KAR5 did not completely block either unisexual or bisexual reproduction, suggesting that additional karyogamy genes may have redundant functions with KAR5. The nuclear morphology inside cnkar5 mutant basidia and cdkar5 mutant early fusion products is similar to the kar5 mutant karyogamy defect phenotype in S. cerevisiae, indicating KAR5 plays a conserved role in nuclear fusion between Saccharomyces and Cryptococcus [41]. During C. neoformans bisexual reproduction, deletion of KAR5 blocked nuclear fusion inside basidia, whereas, during C. deneoformans bisexual reproduction, deletion of KAR5 blocked nuclear fusion at an early developmental step and caused growth arrest for the cell fusion products leading to an apparent cell fusion defect. Early karyogamy in C. deneoformans wild type relieved the requirement for pheromone signaling for directing clamp cell-hyphal fusion during dikaryotic hyphal growth, and the pheromone expression level was rapidly reduced to a basal level in the wild type. However, deletion of KAR5 promoted dikaryotic hyphal growth, and as a consequence the pheromone signaling pathway in kar5 mutants was significantly upregulated compared to the wild type. The pheromone expression patterns validated KAR5’s function in karyogamy.
During C. neoformans and C. deneoformans bisexual reproduction, the involvement of KAR5 in nuclear fusion revealed that karyogamy machinery takes place at different sexual development stages between these two closely related sister species. As reported by Ning and colleagues, the Kar5 protein belongs to a divergent nuclear fusion protein family [43]. Neither CnKar5 nor CdKar5 share sequence similarities outside of the conserved CRD domain with Kar5 proteins from other ascomycetous fungi. Interestingly, the CnKar5 and CdKar5 protein sequences share 85% identity, compared to the average of 93% identity for the 5569 orthologs shared by these two sister species [50]. This suggests that the KAR5 gene has undergone more rapid divergent evolution. The divergence of these proteins may contribute to the mechanistic differences in the karyogamy machinery and may represent a barrier for inter-species nuclear fusion (Fig 9). Several diploid or aneuploid environmental and clinical hybrid isolates of the two Cryptococcus species have been reported, but the few that produce spores have a <10% germination rate [51]. Incompatibility in components of the karyogamy machinery may help to generate a physical barrier for mating and drive speciation events within the Cryptococcus species complex.
Although we validated the conserved roles for PRM1 and KAR5, neither is the sole fusion protein for plasma membrane fusion or nuclear membrane fusion; and deletion of these two factors caused different impacts on bisexual cycles in Cryptococcus (Fig 9). In C. neoformans, Prm1 participates in cell-cell fusion during the initial mating process and mediates clamp cell-hyphal fusion, which is required for maintaining dikaryotic hyphal growth, and Kar5 functions in karyogamy inside the basidia during bisexual reproduction. whereas, in C. deneoformans, Prm1 plays a more significant role in cell-cell fusion, and Kar5 can function in karyogamy immediately following cell fusion, which produces monokaryotic diploid hyphae (Fig 9). However, the observed monokaryotic hyphae could be derived from unisexual reproduction, as pheromone produced by cells of the opposite mating type can promote unisexual reproduction [7]. To address this, we used GFP- and mCherry-labeled nuclear markers to show that the nuclei inside of the monokaryotic hyphae are indeed karyogamy products labeled with both fluorescent markers and thus the products of bisexual reproduction (Fig 6). Collectively, these results demonstrated that there are major differences in both the cell fusion machinery and the karyogamy program during bisexual reproduction between these two closely related sister species (Fig 9).
In contrast to bisexual reproduction, deletion of PRM1 did not cause a significant phenotypic defect during solo unisexual reproduction in C. deneoformans. Although PRM1 was highly upregulated during the unisexual cycle, α-α cell fusion occurred at a 1000-fold lower frequency compared to a-α cell fusion. Furthermore, live cell imaging of yeast cell germination during unisexual reproduction provided compelling evidence that the yeast-hyphal morphological transition is largely independent of cell-cell fusion. It is likely PRM1 may be a fortuitous transcriptional target during unisexual reproduction. However, it is worth noting that those cells that undergo cell-cell fusion do complete the unisexual cycle follow a pathway similar to the bisexual mating mechanism in C. deneoformans, and both PRM1 and KAR5 mediate cell-cell and nuclear fusion during modes of unisexual reproduction that results from α-α cell fusion as detected with genetically marked strains.
In bisexual reproduction, pheromone expression is dampened by the formation of the transcription factor complex Sxi1α-Sxi2a after a-α cell fusion [52]. Interestingly, pheromone expression was also dampened quickly during unisexual reproduction, but the transcriptional downregulation trigger must differ from bisexual reproduction because the opposite mating type was absent. During bisexual reproduction in both C. neoformans and C. deneoformans, KAR5 expression was upregulated and dampened by PRM1 deletion. KAR5 expression was maintained at a basal level and was not affected by the deletion of PRM1 during unisexual reproduction. Furthermore, deletion of KAR5 did not change basidia nuclear morphology compared to wild type, demonstrating that KAR5 is not required for unisexual reproduction. The fact that wild type, the kar5 mutant, and the kar7 mutant produced very few basidia with the two nuclei, indicating either that karyogamy does not occur inside the basidia during unisexual reproduction or that karyogamy occurs transiently and it is hard to capture by DAPI staining or nucleolar fluorescent marker Nop1-GFP. Interestingly, FACS analyses showed that the majority of blastospores produced along the hyphae from unisexual reproduction were diploid, supporting the hypothesis that nuclear fusion does not occur inside the basidium. Despite the fact that deletion of KAR5 does not impact unisexual reproduction and nuclear fusion does not occur inside basidium, we can not entirely rule out that karyogamy could occur during unisexual reproduction, as deletion of KAR5 did not completely block karyogamy during bisexual reproduction, and karyogamy genes in Cryptococcus share redundant functions [38]. Karyogamy occurs early during C. deneoformans bisexual reproduction, and it could also occur early in mother and daughter cells or growing hyphae, which leads to ploidy duplication. However, we favor the interpretation that karyogamy is dispensable for solo unisexual reproduction and an endoreplication pathway, which has been implicated in the formation of polyploid titan cells during Cryptococcus animal infection, contributes to ploidy duplication [53, 54] (Fig 9), which must be differentially controlled compared to titan cell formation, as titan cells reach a much higher ploidy [53].
With the ability to undergo both unisexual and bisexual reproduction, Cryptococcus serves as a model system to study the mating mechanisms for different sexual cycles. Our findings reveal the evolutionary differences in bisexual reproduction within the Cryptococcus species complex and suggest that the unisexual mating mechanism is plastic and complex, providing mechanistic insights to studies of mating mechanisms of unisexual reproduction and parthenogenesis in other eukaryotic systems.
Strains and plasmids used in this study are listed in S1 Table. All strains used to study bisexual reproduction in C. neoformans were generated in the congenic MATα H99 and MATa KN99 strain backgrounds [33]. All strains used to study bisexual reproduction in C. deneoformans were generated in the congenic MATα JEC21 and MATa JEC20 strain backgrounds [55]. All strains used to study unisexual reproduction in C. deneoformans were generated in the MATα XL280 strain background [7]. Yeast cells were grown at 30°C on Yeast extract Peptone Dextrose (YPD) medium. Strains harboring dominant selectable markers were grown on YPD medium supplemented with nourseothricin (NAT) or G418 (NEO). Mating assays were performed on either 5% V8 juice agar medium (pH = 5.0 for C. neoformans and pH = 7.0 for C. deneoformans) or Murashige and Skoog (MS) medium minus sucrose (Sigma-Aldrich) in the dark at room temperature for the designated time period.
To identify the PRM1 orthologs in C. neoformans and C. deneoformans, BLASTP searches using the S. cerevisiae, S. pombe, C. albicans, N. crassa, and A. fumigatus Prm1 protein sequences were conducted against C. neoformans H99 and C. deneoformans JEC21 genomes on FungiDB (www.fungidb.org) [56]. This approach identified CNAG_05866 in C. neoformans and CNF01070 for C. deneoformans as candidate PRM1 genes. Reciprocal BLAST searches confirmed that these two genes are PRM1 orthologs in Cryptococcus spp. Phobius prediction suggested that both CdPrm1 and CnPrm1 have four transmembrane domains at the same amino acid positions (67–87, 352–371, 433–455, and 647–688) [44].
To identify the KAR5 othologs in C. neoformans and C. deneoformans, a BLASTP search using the P. graminis Kar5 protein sequence against the C. neoformans H99 genome identified CNAG_04850 as a candidate KAR5 gene for C. neoformans. However, the same BLASTP search failed to identify a candidate KAR5 gene for C. deneoformans. A subsequent BLASTP search using the C. neoformans KAR5 gene sequence against the C. deneoformans JEC21 genome identified a region from bp 790071 to 792560 on chromosome 10 encoding a protein that shares 85% identity with the C. neoformans candidate Kar5 protein sequence. Multiple sequence alignment of candidate Cryptococcus Kar5 protein sequences with predicted Kar5 protein sequences from other fungal species using the MUSCLE program confirmed they contain Cysteine Rich Domain (CRD) [43, 57]. Phylogenetic analyses for Prm1 and Kar5 were tested with 1000 bootstrap replicas by using the maximum likelihood method in MEGA7 [58, 59]. Phobius prediction predicted that both CdKar5 and CnKar5 have an N-terminal signal peptide and a C-terminal transmembrane domain at amino acid positions 1–16 and 476–501 for CdKar5, and 1–21 and 477–502 for CnKar5 [44]. The COILS/PCOILS program predicted that CdKar5 has four coiled-coil domains at amino acid positions 179–199, 216–236, 318–339, and 368–389, and that CnKar5 has two coiled-coil domains at amino acid positions 180–200 and 370–390 [45].
S1 Table and S2 Table lists the plasmids and primers, respectively, used in this study. To generate deletion mutants for genes of interest, deletion constructs consisting of the 5’ and 3’ regions of the targeted genes flanking an appropriate selection marker (NAT or NEO cassette) were generated by overlap PCR as previously described [60]. The deletion constructs were introduced into the respective strains via biolistic transformation as previously described [61]. Stable transformants were selected on YPD medium supplemented with NAT (100 mg/L) or G418 (200 mg/L). Gene replacements by homologous recombination were confirmed by PCR and Southern hybridization. To generate C. deneoformans wild type strains with dominant selectable markers for cell fusion assays, an analogous method was used to insert a dominant selectable marker (NAT cassette) into the intergenic region immediately downstream of the URA5 gene (CNG03730) and a dominant selectable marker (NEO cassette) into the intergenic region between CNE02520 and CNE02530, which is downstream of the ADE2 gene (CNE02500).
To visualize the cytosol in Cryptococcus, a plasmid encoding the cytosolic mCherry gene and containing a dominant selectable marker (NEO cassette) was generated. The mCherry coding sequence was amplified from pLKB25 [62] and inserted into pXL1 after the GPD1 promoter using the Gibson assembly method, which assembles multiple DNA fragments with 20 to 40 bp overlap sequences in a single reaction containing exonuclease, DNA polymerase, and ligase [63], resulting in pCF1. To monitor nuclear morphology and dynamics during Cryptococcus sexual reproduction, plasmid pSL04 encoding a GFP-tagged nucleolar protein Nop1 from a previous study [38] and a plasmid encoding an mCherry-tagged histone H3 were used. To express the H3-mCherry chimera, the 1075-bp 5’UTR and the 683-bp 3’UTR of the H3 gene were used as promoter (P) and terminator (T), respectively. The H3 promoter and coding sequences before the stop codon and the H3 terminator sequence were amplified from JEC21α genomic DNA, and the mCherry coding sequence was inserted between the H3 coding sequence and H3 terminator by overlap PCR. The chimera expression cassette H3P-H3-mCherry-H3T was then inserted into pAI3 using the Gibson assembly method [63], resulting in pCF9. C. deneoformans strains were biolistically transformed with the pCF1, pSL04, and pCF9 plasmids, and the fluorescent protein expression cassettes were randomly inserted into the genomes. Stable transformants were screened based on fluorescent signals and the selectable markers.
In C. neoformans bisexual reproduction, YSB119 (H99α aca1Δ::NAT ura5 ACA1-URA5) and YSB121 (KN99a aca1Δ::NEO ura5 ACA1-URA5) were used as genetically marked wild type strains to study the fusion competency of prm1 (CF56 and CF562) and kar5 (CF57 and CF549) mutants. In C. deneoformans bisexual reproduction, CF757 (JEC20a URA5-NAT) and CF762 (JEC21α ADE2-NEO) were used as wild type strains to study the fusion competency of prm1 (CF1 and CF313) and kar5 (CF487 and CF364) mutants. InC. deneoformans unisexual reproduction, CF750 (XL280α URA5-NAT) and CF752 (XL280α ADE2-NEO) were used as wild type strains to study the fusion competency of prm1 (CF317 and CF659) and kar5 (CF150 and CF260) mutants. Strains for each fusion pair were grown overnight in YPD liquid medium at 30°C. Cells were washed twice with ddH2O and diluted to a final density of OD600 = 2. Then, 50 μl of equal-volume mixed cells were spotted on V8 medium and incubated for 48 hours (for bisexual reproduction) or 72 hours (for unisexual reproduction) in the dark at room temperature. The cells were then removed, washed with ddH2O, and plated in serial dilution on both YPD medium and YPD medium supplemented with both NAT and G418. The cells were incubated for five days at room temperature. Cell-cell fusion frequency was measured by counting the average number of double drug resistant cfu/total cfu.
To quantify the cell-cell fusion frequency during C. deneoformans bisexual reproduction based on fluorescent signal mixing, CF830 (JEC21α NOP1-GFP-NAT) was mated with JEC20a for wild type fusion frequency, CF768 (JEC20a prm1Δ::NEO NOP1-GFP-NAT) was mated with either JEC21α for prm1 mutant unilateral cell fusion frequency or with CF1 (JEC21α prm1Δ::NEO) for prm1 mutant bilateral cell fusion frequency, and CF723 (JEC20a kar5Δ::NEO NOP1-GFP-NAT) was mated with CF487 (JEC21α kar5Δ::NEO) for kar5 mutant bilateral cell fusion frequency. Cells were prepared as described above and collected for direct fluorescence microscopic observation after 24 hours of incubation. Approximately 100 fusion events were recorded for each mating and were identified by the presence of conjugation tubes connecting the fusion pairs. Fusion frequency was determined by the number of fusion pairs with Nop1-GFP labeled nuclei in both cellular compartments/total fusion events.
To determine whether prm1 and kar5 mutants were defective in spore production, spores were isolated by Percoll gradient centrifugation as previously described [64]. For C. neoformans bisexual reproduction, CF56 (H99α prm1Δ::NAT) crossed with CF562 (KN99a prm1Δ::NEO) and CF57 (H99α kar5Δ::NAT) crossed with CF549 (KN99a kar5Δ::NEO) were compared to the wild type cross between H99α and KN99a. For C. deneoformans bisexual reproduction, CF1 (JEC21α prm1Δ::NEO) crossed with CF313 (JEC20a prm1Δ::NAT) and CF487 (JEC21α kar5Δ::NEO) crossed with CF364 (JEC20a kar5Δ::NAT) were compared to wild type cross between JEC21α and JEC20a. For C. deneoformans unisexual mating, CF317 (XL280α prm1Δ::NEO) and CF260 (XL280α kar5Δ::NEO) were compared to the wild type XL280α. For each mating, triplicates were performed for statistical analysis. Strains were grown overnight in YPD liquid medium. Cells were washed twice with ddH2O and diluted to a final cell density of OD600 = 0.5. Then, 10 μl of equal-volume mixed cells were spotted on V8 medium and incubated for seven days in the dark at room temperature. The entire mating patch was suspended in 60% Percoll (GE Health) in PBS with 0.1% Triton X100. After centrifugation at 10,000 X g for 30 mins in an SW41Ti ultracentrifuge rotor (Beckman-Coulter), a band of spores near the bottom of the Percoll gradient was recovered with a 1-ml tuberculin syringe and transferred into an Eppendorf tube. The total spore production was determined by multiplying the spore density, measured by hemocytometer, with the final volume.
Wild type matings between CF757 (JEC20a URA5-NAT) and CF762 (JEC21α ADE2-NEO) were conducted as controls. The isolated cells were serially diluted and plated on YPD medium and allowed to recover for five days at 30°C. A total of 47 colonies were randomly chosen and grown on YPD medium supplemented with either NAT or G418 to assess growth phenotypes (S2 Fig). Mating type specific primer pairs were used to determine the MAT locus for the progeny.
For all three modes of sexual reproduction studied, prm1 and kar5 mutant strains and wild type strains were grown overnight in YPD liquid medium. Cells were washed twice with ddH2O and diluted to OD600 = 2. Then 250 Δl of an equal-volume mixture of cells were spotted on V8 medium or YPD medium and incubated for 36 hours (YPD and V8) or one week (V8), as the pheromone pathway has been shown to be upregulated upon mating induction on V8 medium and the expression levels are maintained at relatively high levels between 24 and 48 hours [8]. Mating patches were harvested and flash frozen in liquid nitrogen. RNA was extracted using TRIzol reagent (Thermo) following the manufacturer’s instructions. RNA was treated with Turbo DNAse (Ambion), and single-stranded cDNA was synthesized by AffinityScript RT-RNAse (Stratagene). For each sample, cDNA synthesized without the RT/RNAse block enzyme mixture was used as a “no RT control” to control for genomic DNA contamination. The relative expression level of target genes was measured by quantitative real-time PCR using Brilliant III ultra-fast SYBR green QPCR mix (Stratagene) in an Applied Biosystems 7500 Real-Time PCR System. For each target, a “no template control” was performed to analyze melting curves to exclude primer artifacts. Technical triplicates and biological triplicates were performed for each sample. Gene expression levels were normalized using the endogenous reference gene GPD1 and determined by using the comparative ΔΔCt method. The primers used for RT-PCR are listed in S2 Table. The Student’s t-test was used to determine if the relative gene expression levels between different strains exhibited statistically significant differences (P <0.05).
To visualize the nuclei during sexual reproduction, cells were stained with DAPI as previously described [62]. In brief, a 1-mm3 MS agar block containing hyphae on the edge of mating patches was excised and transferred to a small petri dish. The agar block was fixed in 3.7% formaldehyde and permeabilized in 1% Triton X100. The agar block was stained with 2 Δg/ml DAPI (Sigma) and transferred to a glass slide and covered with a cover slip for fluorescent microscopic observation.
To visualize the plasma membrane of the conjugation tubes during C. deneoformans prm1 mutant bisexual reproduction, strain CF1 (JEC21α prm1Δ::NEO) was crossed with CF768 (JEC20a prm1Δ::NEO NOP1-GFP-NAT). After incubation on V8 medium for 24 hours, cells were harvested and resuspended in cold YPD liquid medium on ice. FM4-64 (Thermo) was added at a final concentration of 10 μM and the cells were stained on ice for 15 mins. The cells were then washed with cold YPD medium and fixed in 3.7% formaldehyde in PBS for 10 mins. After a final wash with PBS, the stained cells were examined immediately by confocal microscopy.
Hyphal growth on the edge of mating patches, basidia, and spore chains were captured using a Nikon Eclipse E400 microscope equipped with a Nikon DXM1200F camera.
For fluorescence imaging of hyphae, an agar block supporting hyphal growth was excised and transferred onto a glass slide and covered with a coverslip. For fluorescence imaging of short early hyphae and fusion pairs, early mating patches were harvested and suspended in ddH2O and cells were placed on a glass slide containing a 2% agar patch and covered with a coverslip. Fluorescent images were obtained using a Deltavision system (Olympus IX-71 base) equipped with a Coolsnap HQ2 high resolution SSD camera. Images were processed using the software FIJI.
Confocal fluorescent images were captured by confocal laser scanning microscopy using a Zeiss LSM 710 Confocal Microscope at the Duke Light Microscopy Core Facility. Plan-Apochromat 63X/1.40 Oil DIC M27 objective lenses were used for imaging, and a smart setup was used for image acquisition configuration. Confocal fluorescent images and movies were processed using the ZEN software.
SEM and TEM were performed at the North Carolina State University Center for Electron Microscopy, Raleigh, NC, USA. Samples were prepared for SEM as previously described [8]. In brief, 1-mm3 MS agar blocks containing hyphae on the edge of mating patches were excised and fixed in 0.1 M sodium cacodylate buffer, pH = 6.8, containing 3% glutaraldehyde at 4°C for several weeks. Before viewing, the agar block was rinsed with cold 0.1 M sodium cacodylate buffer, pH = 6.8 three times and post-fixed in 2% osmium tetroxide in cold 0.1 M cacodylate buffer, pH = 6.8 for 2.5 hours at 4°C. Then the block was critical-point dried with liquid CO2 and sputter coated with 50 Å of gold/palladium using a Hummer 6.2 sputter coater (Anatech). The samples were viewed at 15KV with a JSM 5900LV scanning electron microscope (JEOL) and captured with a Digital Scan Generator (JEOL) image acquisition system. For TEM, conjugation tubes were prepared by crossing strain CF712 (JEC21α prm1Δ::NAT mCherry-NEO) with CF768 (JEC20a prm1Δ::NEO NOP1-GFP-NAT). After incubation on V8 medium for 24 hours, cells were harvested and analyzed with a B-C Astrios Sorter to enrich fusion pairs that were positive for both GFP and mCherry fluorescence at the Duke Cancer Institute Flow Cytometry Shared Resource. Hyphae were prepared by crossing strain CF56 (H99α prm1Δ::NAT) with CF562 (KN99a prm1Δ::NEO). After incubation on V8 medium for four weeks, hyphae on the edge of the mating patches were harvested for observation of clamp cell morphology. Upon harvest, cells or hyphae were immediately fixed in 3% glutaraldehyde in 0.1 M sodium cacodylate buffer, pH = 6.8, at 4°C for several weeks. The sample preparation was performed as previously described [62]. In brief, cells were post-fixed with 4% KMnO4 and pre-embedded in 2% agarose. After dehydration with an increasing gradient of ethanol solutions and filtration with Spurr’s resin, the agarose block was embedded in 100% Spurr’s in BEEF capsules. Thin sections were cut and collected on 200-mesh grids, followed by staining with 4% aqueous uranyl acetate and Reynold’s lead citrate. Grids were viewed using a Philips 400T transmission electron microscope. TEM images were processed with Photoshop (Adobe).
Ploidy of blastospores was determined by Fluorescence Activated Cell Sorting (FACS) analysis as previously described [65]. XL280α and MN142.6 (XL280α/α ura5Δ::NAT/ura5Δ::NEO) were used as haploid and diploid controls respectively. Dissected blastospores were grown on YPD medium between three and five days at 30°C to yield colonies. Cells were harvested and washed with PBS buffer. After fixation in 70% ethanol at 4°C overnight, cells were washed once with 1 ml of NS buffer (10 mM Tris-HCl, pH = 7.2, 250 mM sucrose, 1 mM EDTA, pH = 8.0, 1 mM MgCl2, 0.1 mM CaCl2, 0.1 mM ZnCl2, 0.4 mM phenylmethylsulfonyl fluoride, and 7 mM β-mercaptoethanol), and stained in 180 μl NS buffer with 20 μl 10 mg/ml RNase and 5 μl 0.5 mg/ml propidium iodide at 4°C overnight. Then, 50 μl stained cells were diluted in 2 ml of 50 mM Tris-HCl, pH = 8.0 and sonicated for 1 min. For each sample, 10,000 cells were analyzed on the FL1 channel on the Becton-Dickinson FACScan at Duke Cancer Institute Flow Cytometry Shared Resource. Data analysis was performed using the software FlowJo.
|
10.1371/journal.pntd.0002934 | Comparative Pathogenesis of Alkhumra Hemorrhagic Fever and Kyasanur Forest Disease Viruses in a Mouse Model | Kyasanur Forest disease virus (KFDV) and Alkhumra hemorrhagic fever virus (AHFV) are genetically closely-related, tick-borne flaviviruses that cause severe, often fatal disease in humans. Flaviviruses in the tick-borne encephalitis (TBE) complex typically cause neurological disease in humans whereas patients infected with KFDV and AHFV predominately present with hemorrhagic fever. A small animal model for KFDV and AHFV to study the pathogenesis and evaluate countermeasures has been lacking mostly due to the need of a high biocontainment laboratory to work with the viruses. To evaluate the utility of an existing mouse model for tick-borne flavivirus pathogenesis, we performed serial sacrifice studies in BALB/c mice infected with either KFDV strain P9605 or AHFV strain Zaki-1. Strikingly, infection with KFDV was completely lethal in mice, while AHFV caused no clinical signs of disease and no animals succumbed to infection. KFDV and high levels of pro-inflammatory cytokines were detected in the brain at later time points, but no virus was found in visceral organs; conversely, AHFV Zaki-1 and elevated levels of cytokines were found in the visceral organs at earlier time points, but were not detected in the brain. While infection with either virus caused a generalized leukopenia, only AHFV Zaki-1 induced hematologic abnormalities in infected animals. Our data suggest that KFDV P9605 may have lost its ability to cause hemorrhagic disease as the result of multiple passages in suckling mouse brains. However, likely by virtue of fewer mouse passages, AHFV Zaki-1 has retained the ability to replicate in visceral organs, cause hematologic abnormalities, and induce pro-inflammatory cytokines without causing overt disease. Given these striking differences, the use of inbred mice and the virus passage history need to be carefully considered in the interpretation of animal studies using these viruses.
| Kyasanur Forest disease virus (KFDV) and Alkhumra hemorrhagic fever virus (AHFV) are tick-borne flaviviruses that cause severe hemorrhagic disease in humans. The pathogenesis of the disease is still not very well understood mostly due to the lack of suitable animal models. Despite sharing a high degree of genetic sequence similarity, KFDV replicates primarily in the brain and is uniformly lethal for BALB/c mice. In contrast, AHFV does not cause clinically overt signs in mice, replicates in the visceral organs, and induces pro-inflammatory cytokines and hematological changes. Given the striking differences in pathogenesis and tissue tropism, the use of inbred mice as well as the passage history of the virus needs to be carefully considered in the interpretation of animal studies using these viruses.
| Kyasanur Forest disease virus (KFDV) and Alkhumra hemorrhagic fever virus (AHFV) are tick-borne flaviviruses that cause severe hemorrhagic disease in humans. KFDV and AHFV share a high degree of genetic sequence similarity (>90% amino acid identity for the E glycoprotein) despite occupying very different ecological niches [1]–[4]. KFDV was first isolated in 1957 [5]–[7] and causes outbreaks with 400–500 cases annually in Karnataka State, western India. AHFV was first isolated in 1995 [8]–[10] and has caused recent outbreaks of febrile disease in Saudi Arabia [11], [12], as well as illness in travellers returning to Europe from southern Egypt [13]. KFDV infection is primarily associated with bites from Haemaphysalis ticks [14], [15], while AHFV is thought to be transmitted by Ornithodoros savignyi and Hyalomma dromedarii ticks [16], [17]. Both viruses have case fatality rates of up to 20% [reviewed in 18], and work with infectious material is restricted to biosafety level 4 (BSL4) laboratories in North America.
Flaviviruses in the tick-borne encephalitis (TBE) complex typically cause neurological disease in humans. The TBE complex includes human pathogens such as tick-borne encephalitis virus (TBEV), Powassan virus (POWV), and Louping ill virus (LIV), as well as a number of other viruses that are apathogenic in humans [19]. TBEV, POWV, and LIV cause encephalitis of varying severity in humans, but some members of the TBE complex, such as Omsk hemorrhagic fever virus (OHFV), cause predominantly hemorrhagic disease with very little neurological involvement [20], [21]. KFDV and AHFV typically cause hemorrhagic disease as well [6], [8], [19], [21], but there is some evidence of central nervous involvement in infections by KFDV [22], [23] and AHFV [10], [11], [24]. Tick-borne flaviviruses therefore cause a spectrum of disease ranging from neurological to hemorrhagic manifestations, and KFDV and AHFV may occupy an intermediate disease phenotype between TBEV and OHFV. This question needs to be addressed in a small animal model. Furthermore, regulatory bodies such as the World Health Organization (WHO) require that flavivirus countermeasures have to be initially evaluated in a mouse model [25]. In recent laboratory studies, mice have been successfully used as a model for OHFV disease that generally follows the clinical presentation seen in humans [26]–[28]. However, no studies of KFDV pathogenesis in small animal models have been published in the past decade. Older publications from the 1960s and 70s sought to evaluate lethality of KFDV in wild-caught mammals from southwestern India and surrounding regions, but the amount of information contained in these studies is rather limited [29]–[35]. The body of scientific literature on the related AHFV is even smaller, and no animal studies have been published.
To evaluate the utility of existing mouse models for tick-borne flavivirus pathogenesis studies, we performed serial sacrifice studies of immunocompetent mice infected with either KFDV or AHFV to determine the tissue distribution of these viruses and to assess the physiological effects of infection. Although genetically closely related, KFDV and AHFV have very different clinical and virological presentations in BALB/c mice. Our data show that KFDV replicates primarily in the brains of infected animals, while AHFV replicates in visceral organs (kidneys, spleen, and liver) that are commonly associated with hemorrhagic diseases. Given the striking differences in pathogenesis and tissue tropism, the use of inbred mice as well as the passage history needs to be carefully considered in the interpretation of animal studies using these viruses.
Vero E6 were maintained in Dulbecco's modified Eagle's medium (DMEM) containing 5% fetal bovine serum (FBS) and BHK21 clone 13 cells were maintained in minimal essential medium (MEM) containing 10% FBS at 37°C with 5% CO2. All virus isolates were obtained from the World Reference Center for Emerging Viruses and Arboviruses (WRCEVA), which is housed at the University of Texas Medical Branch (UTMB), Galveston, Texas. KFDV P9605 (Genbank accession number JF416958) had 9 suckling mouse brain (SMB) passages and 2 Vero E6 passages; AHFV 200300001 (AHFV 2003; accession number JF416954) had 1 SMB passage and one Vero E6 passage; AHFV Zaki-1 (accession number JF416956) had one SMB passage, 2 mouse brain passages, 4 Vero passages, and 3 Vero E6 passages; and OHFV Guriev (accession number AB507800) had at least 27 SMB passages and one Vero E6 passage. All infections were performed under biosafety level 4 (BSL4) conditions at the Galveston National Laboratory (GNL), UTMB. Virus stocks were grown in Vero E6 cells. Titrations were performed using BHK21 cells by limiting dilution in MEM containing 5% (v/v) FBS and are expressed as the 50% tissue culture infectious dose (TCID50).
For growth kinetics, 5×105 cells per well were seeded into 12-well dishes and were allowed to attach. The respective viruses were added at a multiplicity of infection (MOI) of 0.01 (5×103 TCID50 per well) in a total volume of 500 µl per well. After incubation at 37°C for 1 h, virus inoculum was removed and replaced with 500 µl per well of fresh MEM containing 5% FBS. Beginning at 1 day post-infection, culture supernatants were harvested every 24 h and centrifuged for 5 min at 500×g. The clarified supernatants (∼500 µl) were transferred to fresh tubes and stored at −80°C. Cell-associated virus was collected by scraping the cell monolayers into 500 µl of fresh MEM containing 5% FBS. The resuspended cell fractions were stored at −80°C.
All animal procedures were reviewed and approved by UTMB's Institutional Animal Care and Use Committee (IACUC) in strict compliance with the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health.
Four- to six-week old female BALB/c mice (strain code 028) were purchased from Charles River (Wilmington, Massachusetts). Subdermal transponders IPTT-300 measuring body temperature (Biomedic Data Solutions, Seaford, Delaware) were implanted. Subsequently, animals were moved into the ABSL-4 in the GNL and allowed to acclimate for 5 days before challenge. Mice were kept under barrier conditions in individually ventilated micro-isolator cages at five mice per cage (Tecniplast, Buguggiate, Italy) with corn-cob beading. Food and water was provided ad libitum and environmental enrichment material such as nestles were provided. All animal procedures were reviewed and approved by UTMB's Institutional Animal Care and Use Committee and in strict compliance with the guide for the care and use of laboratory animals. Mice were infected via the footpad with 2×103 TCID50 per animal (20 µl total volume) or by intraperitoneal infection with 103–104 TCID50 per animal of virus diluted in serum-free MEM. Uninfected control animals were injected with serum-free MEM. Blood was collected by intracardial terminal bleeds in EDTA tubes. The whole blood was centrifuged for 5 min at 9,300×g and the resulting plasma was removed to fresh cryovials.
For quantification of virus in post-mortem tissue specimens, a maximum of 0.3 g of tissue was homogenized in 0.6–1.0 ml of MEM containing 5% FBS for 2 rounds of 2.5 min each at 30 cycles s−1 using a Tissuelyser II homogenizer (Qiagen), followed by a 30 s spin at 9,300×g to pellet debris. The virus titers in tissue lysate supernatant and plasma samples were determined by limiting dilution and are expressed as TCID50 per gram of tissue or milliliter of plasma. Additional samples of each tissue were collected in PBS-buffered 4% paraformaldehyde. The fixed organs were kept at 4°C in the BSL4 laboratory, after which the fixative was changed completely and the tissue samples were brought out of the BSL4 according to institutional UTMB operating procedures.
For complete blood counts, 25 µl of EDTA blood was used. All cell counts were quantified using HemaVet 950FS hematology analyzer equipped with software to measure white blood cell count, red blood cell count, platelet count, hemoglobin concentration, hematocrit, mean corpuscular volume, mean corpuscular hemoglobin, and mean corpuscular hemoglobin concentration. For clinical chemistry, 100 µl of lithium heparin plasma was used. Analysis was performed using a VetScan2 Chemistry Analyzer (Abaxis Inc., Sunnyvale, CA, USA), which provides a complete diagnostic panel that includes albumin, alkaline phosphatase, alanine aminotransferase, amylase, total bilirubin, blood urea nitrogen, calcium, creatinine, glucose, and potassium. These hematological and clinical chemistry analyses were performed using whole blood in the BSL4 laboratory.
Brain, spleen, liver, kidney, and lung were harvested during necropsy, fixed in 10% neutral buffered formalin, and removed from the BSL4 based on Galveston National Laboratory (GNL) inactivation procedures. Tissues were then routinely processed by UTMB's Research Histopathology Core, cut, and stained with hematoxylin and eosin (H&E) for histopathologic examination. For immunohistochemistry, 5 µm sections were cut, air dried overnight, and placed into a 60°C oven for 1 h. The deparaffinized and rehydrated sections were quenched for 10 min in aqueous 3% hydrogen peroxide and rinsed in deionized water. Epitopes were retrieved using Biocare's rodent decloaking solution at 98°C for 40 min. Once slides were cooled, they were placed into Tris buffered saline containing 0.05% Tween-20 (TBS-T) for 5 min. Slides were blocked for 30 min with Rodent Block M from the MM HRP-polymer kit (Biocare Medical, USA), and were then rinsed with TBS-T. The slides were incubated with mouse ascites fluid antibodies against KFDV strain P9605 (R157 HMAF) and AHFV strain Zaki-1 (R204 HMAF), both kindly provided by Dr. T. Ksiazek (WRCEVA, UTMB) at a dilution of 1∶1000 overnight at 4°C. Slides were rinsed with TBS-T, incubated for 20 min with MM polymer-HRP, and rinsed again in TBS-T. Betazoid DAB Chromogen Kit (Biocare Medical, USA) was used as the substrate chromogen and the slides were counterstained with Gill's hematoxylin.
To determine cytokine levels in the plasma and tissues of mice, 25 µL of plasma or organ homogenate was run in duplicate with a Bio-Plex Pro Mouse Cytokine Assay kit (Bio-Rad) in the BSL4 laboratory based on the manufacturer's instructions. The kit simultaneously quantifies interleukin (IL)-1β, IL-5, IL-6, IL-10, IL-13, interferon (IFN)-γ, monocyte chemotactic protein (MCP)-1 (also known as CCL-2), and tumor necrosis factor (TNF)-α. The samples were run on the Bio-Plex 200 (Bio-Rad) and analyzed using the Bio-Plex Manager software version 6.0. Sample groups were compared by 1-way analysis of variance (ANOVA) with Tukey post-test using GraphPad Prism version 5.03. Levels of statistical significance are given as either p<0.05 or 0.01.
We performed growth kinetics experiments with OHFV Guriev, KFDV P9605, AHFV 200300001 (AHFV 2003), and AHFV Zaki-1 by infecting BHK21 cells at an MOI of 0.01 and followed the virus titers daily. All viruses reached peak replication in the supernatant (Figure 1A) and cell-associated (Figure 1B) fractions at day 2 post-infection, and titers decreased thereafter. There was approximately 5- to 10-fold more virus found in the supernatant than in cell-associated samples. The growth patterns of AHFV Zaki-1 and KFDV P9605 were very similar in both the supernatant (Figure 1A) and cell-associated fractions (Figure 1B), whereas AHFV 2003 was more comparable to OHFV Guriev (Figures 1A and 1B). The cytopathic effects were similar for all viruses (data not shown). Taken together, these data indicate that none of the viruses have obvious growth defects.
We next determined the lethality of KFDV P9605, AHFV Zaki-1, and AHFV 2003 by infecting groups of 5 BALB/c mice with 103 TCID50 per animal via footpad inoculation. BALB/c mice have been used in prior studies of tick-borne flavivirus infection and disease [26], [27]. KFDV P9605 was uniformly lethal, and animals died between days 6 and 12 post-infection (Figure 2A). Animals typically developed ruffled fur and hunched posture the day before death (data not shown) and had elevated temperatures and weight loss beginning at day 6 (Figures 2B and 2C). In contrast, mice infected with AHFV 2003 (data not shown) or AHFV Zaki-1 did not succumb to infection (Figure 2A), nor did they exhibit any signs of disease (data not shown) or significant weight loss (Figure 2C). AHFV Zaki-1 animals had slightly elevated temperatures on day 3 post-infection, but then returned to normal levels for the duration of the study (Figure 2B). We also infected mice with AHFV 2003 at doses of 104 and 103 TCID50 via the footpad and intraperitoneal route, respectively, but neither route of infection was lethal (data not shown).
In order to demonstrate that animals were infected and the evaluate virus dissemination in infected mice, we performed serial sacrifice studies of BALB/c mice infected by footpad inoculation with 2×103 TCID50 per animal of either KFDV P9605 or AHFV Zaki-1. We chose to use KFDV P9605 and AHFV Zaki-1 for detailed pathogenesis studies because they had similar growth kinetic profiles in cell culture (see Figure 1). Groups of five mice were sacrificed on days 2, 4, 6, and 8 post-infection and organ homogenates were titrated to determine the presence of infectious virus. Mice infected with KFDV P9605 had small amounts of virus in the lung on day 6, which increased on day 8 (Figure 3A). The same pattern was observed in the brain, where several animals were positive on day 6, but all animals had virus in the brain on day 8 (Figure 3B). Virus was not detected in the homogenates of the kidney, spleen, or liver from animals infected with KFDV P9605 (Figure 3C, 3D, and 3E). AHFV Zaki-1 was only detected in the lung and brain of 1 of 5 infected mice on day 2 post-infection, but others were negative throughout the study (Figures 3A and 3B). However, relatively high amounts of AHFV Zaki-1 were found in kidney, spleen, and liver (Figures 3D, 3D, and 3E). Peak titers in the kidney were detected on day 2 post-infection, but declined during the study and virus was cleared by day 8 (Figure 3C). Few animals had virus in the spleen and liver, with the highest titers detected on day 4 post-infection, and were then cleared by day 8 (Figures 3D and 3E); we did not detect virus in the plasma (Figure 3F).
In the brain, lesions were first observable at day 6 in some of the animals and were characterized by mild meningitis with infiltrates primarily composed of lymphocytes (data not shown). KFDV P9605 immunoreactivity was observed in the brains of two of the five animals euthanized at this time (data not shown). Positive immunostaining was observed in scattered neurons of the superficial cerebral cortex. At day 8 brain lesions were more widespread and severe (Figure 4B). Meninges were expanded by inflammatory infiltrates including lymphocytes, plasma cells, and neutrophils (Figure 4B). There were few migrating inflammatory cells directly underlying the affected pia mater within the molecular layer of the cerebral cortex. There was occasional perivascular cuffing and endothelial hypertrophy. KFDV P9605 immunoreactivity was observed in the brains of all animals at day 8 within scattered foci of neurons and occasionally in astrocytes (Figure 5B). Interestingly, meninges and affected vessels were commonly devoid of antigen. Two animals showed mildly increased cellularity of alveolar walls in the lung at day 8 due to infiltration of macrophages and lymphocytes in localized foci (data not shown). Antigen could be detected sporadically by immunohistochemistry in these foci. No lesions or antigen were detected in the liver, spleen, or kidney of mice infected with KFDV P9605.
Small, scattered foci of inflammatory cells were detected throughout the liver of most of the animals infected with AHFV Zaki-1 at days 6 and 8 post-infection. Foci were composed of macrophages and lymphocytes (Figure 4D). Interestingly, viral antigen could not be detected in these foci but was rather found in Kupffer cells lining the sinusoids (Figure 5F). Kidneys appeared congested and an increased cellularity of the glomeruli was noted at day 4 in all animals (Figure 4F). AHFV Zaki-1 immunoreactivity was observed in the kidneys of three out of the five animals from days 2 through 6 post-infection (data not shown). AHFV Zaki-1 immunostaining was observed in cells of the distal cortex, outer and inner medulla (Figure 5D), and sporadically in the mesangial cells of the glomeruli (Figure 5D). At day 4 there was dropout of lymphocytes in the germinal centers of the spleen and increased numbers of macrophages. By day 6 there was pronounced reactive hyperplasia. Immunoreactivity for AHFV Zaki-1 was observed in scattered mononuclear cells in three animals at day 2, one at day 4, and two at day 6 post-infection (data not shown). Antigen was detected within mononuclear cells scattered throughout the parenchyma. No lesions or antigen was detected in the brains of mice infected with AHFV Zaki-1. There were also no lesions or antigen detected in any of the tissues harvested from mock-infected control mice.
We further characterized the responses of mice to KFDV P9605 and AHFV Zaki-1 by evaluating panels of hematological and clinical chemistry parameters. Both KFDV P9605 and AHFV Zaki-1 induced leukopenia on day 2 post-infection, although these levels recovered partially in mice infected with KFDV P9605 after this point (Figures 6A and 6C). The recovery of white blood cell levels was delayed until day 6 in mice infected with AHFV Zaki-1 (Figures 6A and 6C). There was a specific decrease in lymphocyte levels in mice infected with both viruses and these levels recovered over the course of the study (Figures 6B and 6C). The proportion of white blood cells (lymphocytes, monocytes, neutrophils, basophils, and eosinophils) in mice infected with AHFV Zaki-1 remained stable, whereas lymphocyte proportions decreased and neutrophil proportions increased in mice infected with KFDV P9605 (Figure 6D). AHFV Zaki-1-infected mice also had decreased red blood cells counts (Figure 7A), hemoglobin (Hb) levels (Figure 7B), hematocrit (Figure 7C), and platelet counts (Figure 7D), whereas these parameters remained stable in KFDV P9605-infected mice. Thus, despite a lack of clinical disease, AHFV Zaki-1 induces several of the hallmark signs of hemorrhagic fever observed in other animal models [36]–[39].
We have already shown that the highest KFDV P9605 virus titers were present in the brain at days 6 and 8 post-infection, while no AHFV Zaki-1 was detected in the same tissue (Figure 3B). Likewise, AHFV Zaki-1 virus titers are the highest at day 2 and 4 in the kidney and spleen, but KFDV P9605 was not detected in this organ (Figures 3C and 3D). We therefore analyzed the homogenates of these organs from the indicated time points by Bio-Plex Pro Mouse Cytokine Assay for a panel of cytokines. In the brain, KFDV P9605 induced large increases in the amount of IL-10 and IFN-γ, and MCP-1 at day 8, while IL-6 levels were elevated at both days 6 and 8 (Figure 8A). Mice infected with AHFV Zaki-1 had levels of IL-6, IL-10, IFN-γ, and MCP-1 in the brain that were lower than those found in the control or KFDV P9605 animals (Figure 8A). The levels of TNF-α in the brain were generally lower than what was found in the control animals for both viruses (Figure 8A). AHFV Zaki-1 infection induced the production of IL-6, IL-10, IFN-γ, MCP-1, and TNF-α in the kidney at days 2 and 4 post-infection, while the response to KFDV P9605 infection was similar to control animals for these cytokines, with the exception of significantly elevated levels of MCP-1 on day 4 (Figure 8B). Finally, infection with KFDV P9605 also resulted in a significant induction of IL-6, IFN-γ, and MCP-1 in the spleen on day 2, while both KFDV P9605 and AHFV Zaki-1 induced significantly higher levels of IL-10 (Figure 8C). However, the levels of these cytokines diminished rapidly by day 4 for both viruses (Figure 8C). The levels of TNF-α were lower than in control animals on day 2 after AHFV Zaki-1 infection but increased by day 4 (Figure 8C), while the levels of IL-6 and IFN-γ were unchanged on both days (Figure 8C). Both KFDV P9605 and AHFV Zaki-1 are therefore able to induce pro-inflammatory cytokines in organs where virus replication is detected, although the inflammatory response in the spleen is brief.
To the best our knowledge, we have performed the first study that systematically examines the pathogenesis of the tick-borne flaviviruses KFDV and AHFV in an immunocompetent mouse model. KFDV P9605 was lethal in BALB/c mice, but did not cause disease consistent with a hemorrhagic fever. The AHFV strains 2003 and Zaki-1, in contrast, were not lethal in mice and did not cause any overt clinical disease, but the pattern of virus detected in organs (kidney, liver, and spleen), and the observed hematologic abnormalities (leukopenia, lymphopenia, and decreased hematocrit) were consistent with some aspects of hemorrhagic fever.
Previous studies of tick-borne flavivirus pathogenesis have used either of the immunocompetent mouse strains BALB/c or C57BL/6 mice [26]–[28] and the Bogoluvovska strain of OHFV, which has approximately two suckling mouse brain (SMB) passages, and extensive tissue culture passages [26]. This virus caused enlarged spleen and some evidence of hemorrhages in the liver, as well as rapid onset of clinical signs with death occurring around day 9 post-infection [26]. Of particular note is the lack of neurological disease and detection of virus only within the cerebellum of infected animals [26]. Further studies with OHFV strain Guriev resulted in high virus titers in the brain and relatively low titers in other peripheral organs [27]. Although KFDV P9605 does not have as many SMB passages as OHFV Guriev (9 compared to 24 for OHFV Guriev), we found that its distribution in tissues was similar to the results of Tigabu et al. [27], namely that the highest virus titers were found in the brain, but only at low levels elsewhere. This suggests that this neuroinvasive phenotype is perhaps due to neuroadaptation over the course of many SMB passages. In prior studies, mice infected with OHFV Guriev showed some signs of mild neurological disease [27], but we did not observe comparable signs in mice infected with KFDV P9605. KFDV with higher 9 SMB passages (strain 1639, which is similar to KFDV P9605 used in our experiments) has also been found in the CSF of infected macaques at later stages of disease [29], supporting the conclusion that KFDV may have acquired a neuroinvasive phenotype during SMB passage. The practice of passaging virus isolates in mouse brains was common in the 1950s before the widespread availability of tissue culture, so it is consequently quite challenging to obtain virus isolates that have not been extensively passaged in mice.
We found virus distributed in tissues commonly associated with hemorrhagic fever (kidneys, liver, and spleen) in mice infected with AHFV Zaki-1, which appeared at early time points, but was then cleared. We did not find significant amounts of virus in the brain or lungs, which also corresponds with what has been observed for OHFV [26]. Similar to what has been observed by others upon OHFV infection, AHFV induces a pro-inflammatory cytokine response early in the kidney and spleen, and differences in some hematologic parameters, namely leukopenia combined with decreases in lymphocytes, RBC count, hemoglobin levels, and hematocrit [28], indicating that AHFV Zaki-1 causes a sub-clinical hemorrhagic-like syndrome. Lethal infection of mice with TBEV is not associated with elevated body temperatures [40], while animals infected with KFDV in our study had transient fever during the later phase of disease. This suggests that tick-borne flaviviruses that are primarily associated with neurological disease do not induce fever. The presence of KFDV in the brain may therefore reflect a spillover event into the CNS due to factors such as increased vascular permeability rather than a true neuroinvasive phenotype as is seen with TBEV infection. Initial studies with AHFV showed that it was lethal in suckling mice when injected intracerebrally (IC) or intraperitoneally (IP), and in adult mice when administered via the IP route [8]. This discrepancy likely reflects the fact that injection by the IC and IP is a more efficient means of infecting mice compared to footpad injection, which we used for our experiments. We infected mice IP with AHFV Zaki-1, but the dose was not sufficient to cause disease or death. Footpad injection also requires a much smaller volume of inoculum than other routes, so higher doses of virus are not always practical or even possible.
We were unable to detect virus in the plasma of infected animals for both KFDV P9605 and AHFV Zaki-1. This confirms the findings of previous studies, which also failed to detect free virus in animals infected with other tick-borne flaviviruses [27], [28]. It is unclear why there is no detectable viremia in infected mice, especially since we found that there was generally more virus in the in vitro culture supernatant than in the cellular fraction (see Figure 1). We suspect that the virus is associated with circulating leukocytes (macrophages or dendritic cells), as has been described for other viruses such as morbilliviruses [41], [42] and henipaviruses [43]. Interestingly, another study found that KFDV was lethal in bonnet macaques (Macaca radiata) and high virus titers were detected in the serum [29], which suggests that the lack of free virus in the serum may be restricted to rodent species.
In this study we evaluated the pathogenesis of KFDV and AHFV in the BALB/c mouse model. Here we found that AHFV Zaki-1 infection did not cause a clinically evident disease but showed viral tropism, hematological changes, and pro-inflammatory cytokine response suggestive of a viral hemorrhagic fever. In contrast, infection with KFDV was uniformly lethal with evidence of a neuro-inflammatory disease. Histopathology and immunohistochemistry findings support the tissue distribution phenotype of KFDV P9605 and AHFV Zaki-1 observed by titration of virus in organs, and demonstrates the invasiveness of these viruses in the respective tissues. KFDV P9605 seems to invade the brain on or before 6 days post-infection and causes moderate meningitis, with virus replicating in neurons in all areas of the brain. Surprisingly, KFDV P9605 replication could not be detected in any other visceral organs, with the exception of the lung. AHFV Zaki-1 was found in a range of visceral tissues where it only causes mild histopathologic changes. It should be noted that although there is no detectable viremia, AHFV Zaki-1 immunoreactivity was found in multiple tissues with filtering function such as glomeruli in the kidney, Kupffer cells in the liver, and macrophages in the spleen. Despite finding AHFV Zaki-1 antigen and infectious virus in multiple organs, mice appear to be able to control and eliminate the virus effectively. These data also suggest that KFDV is more neurotropic than AHFV and support the hypothesis that KFDV infection can cause neurological disease despite its infrequent occurrence in humans. It is possible that serial passage of KFDV P9605 in the brains of suckling mice contributed to the enhanced neurovirulence. Clearly, the use of authentic low-passage KFDV isolates would be desirable for this study to empirically evaluate the neurovirulence of KFDV. Unfortunately, low-passage KFDV strains were not available to us. Future studies will examine the influence of progressive virus passage of AHFV and KFDV in different organs on disease presentation.
|
10.1371/journal.pgen.1006751 | Semaphorin-1a prevents Drosophila olfactory projection neuron dendrites from mis-targeting into select antennal lobe regions | Elucidating how appropriate neurite patterns are generated in neurons of the olfactory system is crucial for comprehending the construction of the olfactory map. In the Drosophila olfactory system, projection neurons (PNs), primarily derived from four neural stem cells (called neuroblasts), populate their cell bodies surrounding to and distribute their dendrites in distinct but overlapping patterns within the primary olfactory center of the brain, the antennal lobe (AL). However, it remains unclear whether the same molecular mechanisms are employed to generate the appropriate dendritic patterns in discrete AL glomeruli among PNs produced from different neuroblasts. Here, by examining a previously explored transmembrane protein Semaphorin-1a (Sema-1a) which was proposed to globally control initial PN dendritic targeting along the dorsolateral-to-ventromedial axis of the AL, we discover a new role for Sema-1a in preventing dendrites of both uni-glomerular and poly-glomerular PNs from aberrant invasion into select AL regions and, intriguingly, this Sema-1a-deficient dendritic mis-targeting phenotype seems to associate with the origins of PNs from which they are derived. Further, ectopic expression of Sema-1a resulted in PN dendritic mis-projection from a select AL region into adjacent glomeruli, strengthening the idea that Sema-1a plays an essential role in preventing abnormal dendritic accumulation in select AL regions. Taken together, these results demonstrate that Sema-1a repulsion keeps dendrites of different types of PNs away from each other, enabling the same types of PN dendrites to be sorted into destined AL glomeruli and permitting for functional assembly of olfactory circuitry.
| In the Drosophila olfactory system, olfactory projection neurons (PNs) are derived from four neural stem cells (called neuroblasts) during the development. Intriguingly, these PNs generate complex dendritic patterns within the primary olfactory center of the brain, the antennal lobe (AL), to relay odorant information from olfactory sensory neurons in the periphery to neurons in higher olfactory centers. In this study, we investigate how various types of PNs use a repulsive transmembrane protein Semaphorin-1a (Sema-1a) to establish appropriate dendritic patterns within the AL. Previously, Sema-1a was proposed to globally control initial PN dendritic targeting along the dorsolateral-to-ventromedial axis of the AL. In contrast, we disclose an unknown role of Sema-1a, in which this neuronal protein acts to keep dendrites of various types of PNs produced from different neuroblasts away from select AL regions, thereby enabling the dendrites of the same types of PNs to sort correctly into destined glomeruli within the developing AL for assembly of the functional olfactory neural circuitry.
| In the olfactory system, odorant inputs are detected by olfactory sensory neurons (OSNs) in the periphery and converged into individual glomeruli of the primary olfactory center, termed the antennal lobe (AL) in Drosophila and the olfactory bulb in mice, where projection neurons (PNs in Drosophila and mitral/tufted cells in mice) relay these inputs to other brain regions for decoding [1]. In Drosophila, most PNs are generated from four neural stem cells (called neuroblasts) and therefore can be assigned into four neural lineages: anterodorsal PNs (adPNs) in the ALad1 lineage, ventral PNs (vPNs) in the ALv1 lineage, lateroventral PNs (lvPNs) in the ALlv1 lineage, and lateral PNs (lPNs) from a lateral group of mixed PNs and local interneurons (LNs) in the ALl1 lineage [2, 3] (also see S1 Fig). Among these PNs, dendrites of most types of adPNs and lPNs innervate individual glomeruli (as uni-glomerular PNs) within the AL [4–9], whereas many types of vPNs and lvPNs establish poly-glomerular dendritic arborization patterns in the AL [10, 11]. Intriguingly, various types of adPNs, lPNs, vPNs and lvPNs distribute their dendrites in distinct but overlapping patterns within the AL [11]. Elucidating the molecular mechanisms underlying how types of PNs within different neural lineages generate the complex patterns of dendritic arborizations in discrete AL glomeruli is crucial for comprehending the formation of the functional olfactory circuitry.
Since it is unclear whether the same molecular mechanisms are utilized to generate the complex dendritic patterns of adPNs, lPNs, vPNs and lvPNs, it is important to examine the roles of the same organizing cues in the formation of appropriate dendritic patterns for different types of PNs. For example, it has been previously reported that initial PN dendritic targeting in the developing AL is mediated through opposing gradients of repulsive semaphorin cues, Sema-2a/-2b, and a receptor for these cues, the transmembrane protein semaphorin-1a (Sema-1a). The ventromedial (VM) expression of secreted Sema-2a/-2b from degenerating larval OSN axons is proposed to influence PN dendritic elaboration that is dependent upon dorsolateral (DL) expression of membrane-tethered Sema-1a in PNs [12, 13]. Dendrites of DL1 adPNs and DA1 lPNs underwent a DL-to-VM shift when the whole animal was deficient for Sema-2a/-2b, or when Sema-1a was selectively removed from PNs, suggesting a crucial role for a repulsive Sema-2a/-2b gradient that is read by the receptor Sema-1a in setting up appropriate dorsal dendritic patterns for adPNs and lPNs [12, 13]. In contrast, RNAi knock-down of Sema-1a caused the dendrites of DA1 vPNs to no longer be constrained within the DA1 glomerulus, with a substantial fraction of these dendrites invading the DA3 glomerulus [14], implicating Sema-1a as a regulator of vPN dendritic morphogenesis. However, it is rather puzzling why the shifted DA1 vPN dendrites in the absence of Sema-1a, which are perpendicular to those of Sema-1a-deficient DL1 adPNs, do not exhibit a DL-to-VM shift, a prediction of the current model [12, 13]. Therefore, it is possible that the Sema-1a signal is transmitted differently in adPNs and lPNs compared to vPNs, or even that an alternative model accounts for these Sema-1a loss-of-function (LOF) dendritic phenotypes.
Here, using genetic LOF and rescue studies we identify a previously unknown role for Sema-1a in preventing aberrant dendritic invasion of both uni-glomerular and poly-glomerular PNs into select AL regions, including the DA3 glomerulus and the region close around the VC1 glomerulus; this role is distinct from previously explored functions of Sema-1a in global control of initial PN dendritic targeting along the DL-to-VM axis of the AL [12]. Intriguingly, the prevention of dendritic mis-targeting to the DA3 glomerulus mediated by Sema-1a seems to be PN-origin dependent, i.e., the occurrence of the Sema-1a-deficient dendritic mis-targeting phenotype only in the types of PNs derived from adPN and vPN neuroblasts but not from the lPN neuroblast. Further, ectopic expression of Sema-1a caused DA3 adPNs that normally send their dendrites to the DA3 glomerulus to mis-project their dendrites into adjacent glomeruli. Taken together, our results suggest that repulsive Sema-1a signals in adPNs, lPNs and vPNs keep different types of PN dendrites away from each other, ensuring that they instead navigate to their destined glomeruli to establish appropriate dendritic patterns for assembling the functional olfactory circuitry to decode odorant information from the external world.
We sought to label adPNs (or lPNs) and vPNs in distinct colors that permits the simultaneous visualization of how the dendrites of different PN populations distribute within the AL during development. Applying the twin-spot MARCM (mosaic analysis with repressible cell markers) system [15], we induced independent fluorescent labeling of adPN (or lPN) and vPN neuroblasts of newly hatched larvae (NHL) to visualize larval-born-adPNs (or -lPNs) and -vPNs in two different colors (labeled by GAL4-GH146 and GAL4-MZ699) [6, 10]. We found that dendrites of adPNs (or lPNs) and vPNs were initially segregated at the early pupal stage, became apparently mixed at 48 hours APF and turned into fully intermingled in the adult AL (S2 Fig). The observation of dendritic mixing among adPNs (or lPNs) and vPNs in the developing AL raises an interesting question as to whether or not vPNs employ similar or different molecular mechanisms from those used by adPNs and lPNs to generate appropriate dendritic patterns during their morphogenesis.
Previous work demonstrates that dendrites of DL1 adPNs and DA1 lPNs, as opposed to those of DA1 vPNs, have qualitatively distinct phenotypes in LOF studies of Sema-1a (see S3A Fig for the illustrative drawing of Sema-1a-deficient dendritic phenotypes observed in DL1 adPNs and DA1 vPNs) [12, 14]. To verify that mis-targeting of dendrites to the DA3 glomerulus we observed in the Sema-1a RNAi knock-down DA1 vPN (also see S3B–S3G Fig) actually resulted from the absence of Sema-1a rather than from an off-target effect of Sema-1a RNAi [16], we conducted MARCM experiments on a severe Sema-1a LOF mutation (Sema-1aP1) using GAL4-GH146, which labels four types of vPNs: DA1, diffuse, VA1lm and VL1 (Fig 1) [4, 5, 17]. In contrast to the wild-type DA1 vPN dendrites, which predominantly innervated the DA1 glomerulus (Fig 1A; Table 1), the Sema-1aP1 mutant DA1 vPN dendrites robustly mis-target into the DA3 glomerulus (Fig 1B; 95%, n = 21; Table 1). Notably, this DA3-glomerular dendritic mis-targeting defect was completely rescued by restoring the expression of Sema-1a in DA1 vPNs such that dendritic innervation was almost exclusively within the DA1 glomerulus (Fig 1C; Table 1). These results show that DA1 vPN dendrites aberrantly invade into the DA3 glomerulus when the expression of Sema-1a is disrupted.
It is unclear whether this dendritic mis-targeting phenotype is unique to DA1 vPNs or occurs in other types of vPNs (e.g., diffuse-, VA1lm- and VL1-vPNs) as well in the absence of Sema-1a. Interestingly, diffuse vPNs also exhibited a similar dendritic mis-targeting to the DA3 glomerulus in the Sema-1aP1 mutant: they accumulated in the DA3 glomerulus and aberrantly projected to the subesophageal zone (SEZ), in contrast to wild type, where the dendrites of diffuse vPNs were loosely distributed to nearly all of the AL glomeruli (Fig 1D–1E and their insets; 100%; Table 1). Both the DA3-glomerular dendritic accumulation and SEZ mis-projection phenotypes disappeared when Sema-1a was over-expressed in Sema-1aP1 diffuse vPNs (Fig 1F and its inset; Table 1). On the other hand, we did not observe the DA3-glomerular dendritic mis-targeting phenotype in the other two types of GAL4-GH146-positive vPNs, VA1lm- and VL1-vPNs, when the expression of Sema-1a was altered (S4 Fig; S1 Table). Taken together, the DA3-glomerular dendritic mis-targeting defect observed in the Sema-1a-deficient DA1- and diffuse-vPNs demonstrates that Sema-1a plays a crucial role in establishing appropriate dendritic patterns of both uni-glomerular and poly-glomerular vPNs, supporting a model that Sema-1a counteracts putative attractive force of the DA3 glomerulus, and this specific dendritic mis-targeting defect seems deviated from the prediction of the current model in which PN dendrites shift along the DL-to-VM axis of the AL in the absence of Sema-1a [12].
Since dendrites of wild-type DA1- and diffuse-vPNs are distributed close to the DA3 glomerulus, and since dendrites of Sema-1a-deficient DA1- and diffuse-vPNs mis-target into the DA3 glomerulus (Fig 1 and S3 Fig), we wondered whether Sema-1a signaling serves to prevent aberrant dendritic invasion into the DA3 glomerulus by surrounding PNs. To test this hypothesis, we examined Sema-1a LOF effects on the adPNs and lPNs which normally project their dendrites to surround the DA3 glomerulus (Fig 2): these adPNs and lPNs include DL3- and DA1-lPNs and VA1d-, DA4l-, DA4m- and D-adPNs, which project their dendrites clockwise to surround the DA3 glomerulus; they also include DL4-, DC3- and DC1-adPNs, which send their dendrites posteriorly covering the DA3 glomerulus (Fig 2A). Notably, all the PNs we examined except DA1- and DL3-lPNs mis-targeted their dendrites into the DA3 glomerulus in the absence of Sema-1a (Fig 2B–2S; S5 and S6 Figs; Table 1), similar to the dendritic mis-targeting phenotype we observed in the Sema-1aP1 DA1- and diffuse-vPNs (Fig 1). Among the adPNs we examined, DC3 adPNs displayed the most severe defect, with full dendritic invasion into the DA3 glomerulus in all Sema-1a-deficient animals (Fig 2E, 2F and 2S; Table 1). The rest of the adPN types exhibited differing degrees of penetrance and expressivity of the DA3-glomerular dendritic mis-targeting phenotype in the Sema-1aP1 mutant and Sema-1a RNAi knock-down samples (Fig 2B, 2C, 2H, 2I and 2K–2S; S5 Fig; Table 1). The DA3-glomerular dendritic mis-targeting phenotype in the Sema-1aP1 D-, DC3- and VA1d-adPNs was no longer observed when wild-type Sema-1a was over-expressed in these same PNs (Fig 2D, 2G and 2J; Table 1). Notably, in the rescue experiments the dendrites of D- and VA1d-adPNs remained situated on the edge and outside of the D glomerulus (insets of Fig 2D; 100%, n = 13; Table 1) and at the ventral portion of the VA1d glomerulus (Fig 2J; 100%, n = 14; Table 1), implicating that dendrites of the rescued and remaining wild-type adPNs may repel with each other. Taken together, the dendritic mis-targeting defect we observed in Sema-1a-deficient adPNs and vPNs suggests that the DA3 glomerulus serves as a select AL region for extending dendrites in the absence of Sema-1a. Intriguingly, in this DA3-glomerular dendritic mis-targeting phenotype, the PNs with dendrites that surround the DA3 glomerulus, including seven types of adPNs (D, DA4l, DA4m, DC1, DC3, DL4 and VA1d) and two types of vPNs (DA1 and diffuse), but not DA1- and DL3-lPNs, tend to aberrantly extend their dendrites into the DA3 glomerulus when Sema-1a is absent.
The disappearance of dendrites from the DA3 glomerulus in diffuse vPNs observed in the Sema-1aP1 mutant with Sema-1a over-expression (Fig 1F and its inset) prompted us to ask how PN dendrites that normally project into the DA3 glomerulus (e.g., DA3 adPNs) would behave when Sema-1a expression is altered (Fig 3). Since dendrites of wild-type DA3 adPNs already distribute themselves into the DA3 glomerulus that attracts Sema-1a-deficient dendrites (Fig 2), we predicted that DA3 adPN dendritic projections to the DA3 glomerulus should remain unaffected when Sema-1a is mutated. Indeed, we found that DA3 adPNs rarely displayed abnormal dendritic phenotypes in the Sema-1aP1 mutant (Figs 2S, 3A and 3B; Table 1; see S2 Table for information on the birth-order of Sema-1aP1 DL1-, DA3- and DC2-adPNs in our synchronized MARCM experiments), implicating that the endogenous Sema-1a expression may be low (if there is any expression) and may not play a crucial role in the dendritic targeting of the DA3 adPNs. Furthermore, if Sema-1a counteracts the attraction in the DA3 glomerulus, DA3 adPN dendrites should be sensitive to an excessive level and ectopic time window of the Sema-1a gain-of-function paradigm. Since we did not systematically conduct the synchronized MARCM experiments to over-express Sema-1a in the wild-type PNs, we, instead, altered Sema-1a expression in DA3 adPNs by over-expressing Sema-1a in the Sema-1aP1 mutant PNs. Using this approach, we did not find any adPNs with dendritic projections into the DA3 glomerulus. Instead, we found many adPNs whose dendrites projected into the DL3 glomerulus in the majority of the cases (63%, n = 16; Table 1; Fig 3C) and in a few cases into the DA4l glomerulus (12%, n = 16; Table 1; Fig 3D) or both the DA4l and DL3 glomeruli (25%, n = 16; Table 1; Fig 3E and 3F) during the developmental time window for the generation of DA3 adPNs. We noted that our determination of the identity of these DL3/DA4l PNs as DA3 adPNs was based on their anterodorsal soma position, ruling out their being lPNs (Fig 3C–3F; the only wild-type DL3 PNs labeled by GAL4-GH146 are DL3 lPNs [6]). Further, the birth of these DL3/DA4l PNs occurred prior to the birth of DC2 adPNs but after the birth of DL1 adPNs in our synchronized MARCM experiments, establishing their identity as DA3 adPNs (S3 Table; the birth order of the embryonic-born DA4l adPN and larval-born DL1-, DA3-and DC2-adPNs has been reported previously [6]). Taken together, these results using manipulation of Sema-1a expression in DA3 adPNs reinforces our hypothesis that the DA3 glomerulus acts as a select AL region to attract nearby PN dendrites when counteracting Sema-1a signaling is absent.
When we analyzed the phenotypes of those PNs that mis-targeted their dendrites into the DA3 glomerulus, we observed that Sema-1aP1 DC1 adPN dendrites were also mis-projected to a region close to the VC1 glomerulus (100%, n = 3; Fig 2P and its inset; Table 2). This observation led us to look for additional select AL regions (besides the DA3 glomerulus) that could attract dendrites from different sets of PNs when Sema-1a is absent. In our MARCM experiments additional seven types of lPNs and nine types of adPNs were also labeled using GAL4-GH146, allowing us to examine their dendritic patterns (Fig 4). Similar to the DC1 adPN, DC2-, DP1m- and VL2p-adPNs and VA5- and VA7m-lPNs also displayed the phenotype of dendritic mis-targeting to the region around the VC1 glomerulus, with variable mis-projection positions, penetrance and expressivity (31%~87%; Fig 4A–4J; Table 2). Again, over-expressing wild type Sema-1a in Sema-1aP1 DC2 adPNs and VA5- and VA7m-lPNs rescued this mis-targeting phenotype: dendrites remained in relatively wild type locations, with dendritic occupancy at the edge and outside of the DC2, VA5 and VA7m glomeruli (S7 Fig; Table 2; similar findings were seen in the DL1 adPNs and DA1- and DL3-lPNs, see S6F, S6I and S8C Figs). The possibility of there being multiple select AL regions that attract dendrites in the absence of Sema-1a was further strengthened by our observation of another striking dendritic mis-targeting phenotype, in which dendrites mis-projected into the DA3 glomerulus and into regions in close proximity to the VC1 glomerulus when Sema-1a was mutated in two embryonic-born adPNs (Fig 4K and 4L; Tables 1 and 2; the only wild-type DA3 adPNs labeled by GAL4-GH146 are larval-born [6]). Moreover, the VM7v and VA4 glomeruli and the region ventral to the DP1m glomerulus were also prone to aberrant dendritic invasion by DM1-, DM2- and VA7m-lPNs and DL1-, DL5- and DM3-adPNs (S8B and S9 Figs; S4 Table; VM7v adPNs are the only PNs labeled by GAL4-GH146 that innervate the VM7v glomerulus [6]). However, whether the VM7v and VA4 glomeruli and the region ventral to the DP1m glomerulus behave as the select AL regions to attract other PN dendrites remains unclear and awaits the characterization of dendritic patterns for the rest of the PNs that we did not examine here. In contrast, VA2-, VA3- and VM3-adPNs and VA4-, VC1- and VC2-lPNs did not exhibit specific dendritic mis-targeting phenotypes when Sema-1a was absent (S1 Table). We also noted that many embryonic-born, and a few larval-born, Sema-1aP1 adPNs/lPNs mis-projected their dendrites into the SEZ without AL innervation, which we never observed in the wild-type samples (S10 Fig; S5 Table). Taken together, these results suggest that one additional select AL region (i.e., the region close to the VC1 glomerulus) may co-exist along with the DA3 glomerulus and that they serve to attract PN dendrites when Sema-1a is absent.
Secreted ligands and cell surface molecules act in concert to regulate the cell-morphogenetic and neurite-sorting processes that generate appropriate patterns of axonal branches and dendritic arbors in neurons of the olfactory system, resulting in constructing the complex, functional olfactory circuitry [18, 19]. In the present study, we discover a new role of Sema-1a to prevent dendrites of adPNs, lPNs and vPNs from aberrantly invading into the select AL regions, which is crucial for generating appropriate discrete PN dendritic patterns to construct the olfactory map within the AL.
Construction of the Drosophila adult AL is a complex process to integrate neurites of multiple populations of PNs, LNs and OSNs during the pupal stage [8, 20]. What are the roles of specific molecules in this complex process of neurite sorting and integration among PNs, LNs and OSNs? Sema-1a was proposed to control initial dendritic targeting of PNs along the DL-to-VM axis of the AL based on observations of dorsolateral-enriched expression of Sema-1a in the developing AL and mis-targeting of Sema-1aP1 DL1 adPN dendrites into the region ventromedial to the developing AL (Fig 4C of the Komiyama study) [12]. Interestingly, we observed similar phenotypes in Sema-1aP1 adPNs and lPNs, in which their dendrites mis-projected into the SEZ, a neuropile ventromedial to the adult AL, with or without entering the AL (Figs 1E and 4F; S5B–S5D Fig, S6E and S10 Figs). Therefore, the repulsive Sema-1a signal does play an essential role in the step of initial dendritic targeting by preventing PN dendrites from mistakenly invading the region ventromedial to the AL (e.g., the SEZ; also see the schematic drawing in Fig 5).
The above described step of initial dendritic targeting directed by Sema-1a was further proposed to link with the later refinement and sharpening of boundaries among glomeruli through intercellular interactions, for example dendrite–dendrite interactions among PNs mediated by N-cadherin [12, 21]. A correlated range of severity of the DL-to-VM dendritic shift phenotypes found in Sema-1aP1 PNs—in which DL1 adPNs (having the farthest dorsolateral dendrites) displayed the most severe DL-to-VM dendritic shift compared to the moderate and mild phenotypes of the dendrites of DA1 lPNs and DC3 adPNs—supported the idea that the distribution of PN dendrites in the AL is determined by the Sema-1a expression gradient [12]. Although we observed similar DL-to-VM dendritic shift defects in Sema-1aP1 DL1 adPNs and DA1 lPNs (S6D and S8B Figs), these phenotypes may be also interpreted as mis-targeting of dendrites of DL1 adPNs and DA1 lPNs into unidentified select AL regions (e.g., the region ventral to the DP1m glomerulus for DL1 adPNs shown in the inset of S8B Fig) in the absence of Sema-1a. In contrast, Sema-1aP1 DC3 adPNs exhibited the DA3-glomerular dendritic mis-targeting defect and did not show the DL-to-VM dendritic shift phenotype in our study (Fig 2G). Upon close examination of two Sema-1aP1 DC3 adPN images from Komiyama et al. (their Fig 3D) [12], we noted that the bottom right image displays a mild dendritic mis-targeting defect, occupying the ventral tip of the DA3 glomerulus, and the bottom left image was horizontally flipped, so it is not likely to be a DC3 adPN. Both DC3 adPN examples shown in the Komiyama study [12], taken together with the DA3-glomerular dendritic mis-targeting defect we observed in DC3 adPNs (Fig 2G) and also in other PN samples (D-, DA4l-, DA4m-, DC1-, DC3-, DL4- and VA1d-adPNs and DA1- and diffuse-vPNs in Figs 1 and 2), complicates the straightforward DL-to-VM dendritic shift model that was proposed in order to account for the function of Sema-1a in the establishment of appropriate dendritic patterns once PN dendrites have projected into the developing AL.
How then does Sema-1a regulate the formation of dendritic patterns once multiple populations of PNs (e.g., adPNs, lPNs and vPNs) have sent their dendrites into the developing AL? We found here that different sets of uni-glomerular- and poly-glomerular PNs among adPNs, lPNs and vPNs mis-targeted their dendrites into select AL regions, including the DA3 glomerulus and a region close around the VC1 glomerulus, when Sema-1a was mutated (Figs 1, 2 and 4). Intriguingly, the PNs that displayed the Sema-1a-deficient DA3-glomerular dendritic mis-targeting defect seem to associate with their deriving origins, i.e., the occurrence of the dendritic mis-targeting phenotype only in adPNs and vPNs but not lPNs (Figs 1 and 2 and S6 Fig). In the light of these results, it may be reasonable to speculate the presence of sorting centers in the developing AL for controlling dendrites of uni-glomerular- and poly-glomerular-PNs among adPNs, lPNs and vPNs toward their destined glomeruli. Interestingly, dendrites of adPNs and vPNs seem to accumulate in the anterior dorsolateral portion of the developing AL at 24h APF (double-arrowhead in S2C Fig), which maybe correlate with PN dendritic innervation in the DA3 glomerulus and anterodorsal glomeruli of the AL. When the Sema-1a repulsion (as a driving force) is gone, PN dendrites will be trapped in the sorting centers, which results in to dendritic mis-targeting into select AL regions. However, without figuring out the identity of those PN dendrites in the anterior dorsolateral portion of the developing AL (e.g., whether the green adPN dendrites are destined toward the DA3 glomerulus and its surrounding glomeruli, and whether the magenta PN dendrites are derived from vPNs but not from the DL1 adPN; double-arrowhead in S2C Fig), the presence of dendritic sorting centers in the developing AL remains elusive.
No matter whether the dendritic sorting centers exist in the developing AL or not, how does Sema-1a transmit the repulsion signal in PN dendrites to prevent inappropriately mixing together? Previously, the expression of Sema-2a/-2b in the ventromedial corner of the developing AL has been reported as ligands of Sema-1a to regulate the PN dendritic targeting [13]. However, it may take a slightly complicate mechanism to only involve the graded expression of Sema-2a/-2b to elicit the repulsive Sema-1a signal in PNs to avoid dendritic accumulation in select AL regions. Therefore, it is reasonable to speculate the presence of not-yet identified factors which may work in concert with the repulsive Sema-2a/-2b-Sema-1a signal in PNs. Dependent on the distribution of those not-yet identified factors in the developing AL, various types of uni-glomerular- and poly-glomerular-PNs among adPNs, lPNs and vPNs may interpret the repulsive Sema-1a signal instructively or permissively to generate distinct but overlapping dendritic patterns in discrete glomeruli of the AL. Based on the differing degrees of penetrance and expressivity of the Sema-1a-deficient DA3-glomerular dendritic mis-targeting phenotype, we speculate that the repulsive Sema-1a signal may be differentially transmitted in various types of PNs with high (e.g., DC3 adPNs and DA1- and diffuse-vPNs), moderate (e.g., D-, DA4l-, DA4m-, DC1-, DL4- and VA1d-adPNs) to low (e.g., DA3 adPNs) degrees (see Fig 5 for the schematic drawing). Removing Sema-1a from PNs (e.g., DC3- and DA4l-adPNs and DA1 vPNs) that are normally affected by the repulsive Sema-1a signal may turn these Sema-1a-deficient PN dendrites to behave more like dendrites of PNs that normally do not respond to the repulsive Sema-1a signal (e.g., DA3 adPNs), which results in dendritic mis-targeting in a wrong AL region (e.g., DC3 and DA4l Sema-1aP1 adPNs and DA1 Sema-1aP1 vPNs mis-target their dendrites into the DA3 glomerulus). On the other hand, excessive and ectopic expression of Sema-1a in the PNs (e.g., DA3 adPNs) that presumably expresses low or no endogenous level of Sema-1a may convert these Sema-1a-manipulated PNs into sensitive to the repulsive Sema-1a signal, which results in steering their dendrites away from their destined region (e.g., the DA3 glomerulus) into adjacent glomeruli (e.g., DA4l and DL3 glomeruli; see Fig 5 for the schematic drawing). While our proposed model is able to account for the dendritic mis-targeting phenotypes observed among Sema-1a-deficient adPNs, lPNs and vPNs, questions remain concerning how and what not-yet identified factors work together with Sema-1a in various types of PNs to set up correct dendritic patterns. Future identification and investigation of these factors may reveal how extracellular signals that elicit the Sema-1a repulsion in PNs and how Sema-1a works in concert with ensembles of molecules among various types of PNs to generate a precise and functional olfactory map that decodes external olfactory inputs into essential internal information for the survival of animals.
Standard molecular biological techniques were used to generate UAS-Sema-1asy, in which the Sema-1a open reading frame from a UAS-Sema-1a cDNA construct [17] was cut with XhoI and XbaI and re-cloned into pJFRC7-20XUAS-IVS-mCD8::GFP [22]. The UAS-Sema-1asy construct was injected into a fly stock carrying an attP docking site (VK00033; e.g., Bloomington stock number (BL) 9750) to generate the transgenic fly stock via the service provided by Rainbow Transgenic Flies, Inc.
The fly strains used in this study were as follows: (1) hs-FLP122 [8]; (2) FRT40A,UAS-rCD2::RFP,UAS-GFP-miRNA [15]; (3) FRT40A,UAS-mCD8::GFP,UAS-rCD2-miRNA, GAL4-GH146 [6]; (4) GAL4-MZ699 [23]; (5) GH146-FLP [24]; (6) UAS-FRT<stop<FRT-myrGFP [22]; (7) R95B09-GAL4 (BL47267); (8) UAS-Sema-1a RNAiTRiP (BL34320); (9) FRT40A,UAS-mCD8::GFP,GAL4-GH146 [25]; (10) FRT40A, UAS-mCD8::GFP,Sema-1ak13702,GAL4-GH146 (Sema-1ak13702 is Sema-1aP1 [17] and the original source of Sema-1ak13702 was from Kyoto Stock Center with the stock number 111328); (11) FRT40A,tubP-GAL80 [26]; (12) UAS-Sema-1asy (generated in this study); (13) R38B04-GAL4 (BL49984); (14) GAL4-MZ19 [27]; (15) actin-FRT<stop<FRT-GAL4 [2].
Flippase-out mediated intersection samples and mosaic clones for the MARCM, flip-out MARCM and twin-spot MARCM studies were generated as previously described [2, 14, 15, 26]. MARCM samples were obtained by collecting embryos in vials and inducing mosaic clones at various developmental periods with heat-shock for 10 to 15 minutes. To pin-point birth-order of PNs in Sema-1a LOF and Sema-1a rescued MARCM experiments (S2 and S3 Tables), larvae were picked up as NHL to synchronize samples, and mosaic clones were induced at four-hour intervals from 26h ALH to 56h ALH with heat-shock for 10 to 15 minutes. For the flip-out MARCM and twin-spot MARCM experiments, mosaic clones of larval-born adPNs, lPNs, vPNs and lvPNs were generated by heat-shock for 10 to 25 minutes in NHL.
Dissection, immunostaining and mounting of more than 5,000 fly brains were performed as described in a standard protocol [26]. Primary antibodies used in this study included rabbit antibody against GFP (1:800, Invitrogen), rabbit antibody against RFP (1:800, Clontech), rat antibody against DN-cadherin (DN-Ex #8, 1:50, DSHB), rat antibody against mCD8 (1:100, Invitrogen), and mouse antibody against Bruchpilot (nc82, 1:50, DSHB). Secondary antibodies conjugated to different fluorophores (Alexa 488, 546, and 647 (Invitrogen)) were used at a 1:800 dilution. Immunofluorescent images of single neurons from single-cell MARCM clones and intersection experiments (around 900 images) and groups of neurons from multi-cellular neuroblast clones of flip-out MARCM and twin-spot MARCM experiments (around 100 images) were collected by Zeiss LSM 700 or 780 confocal microscopy, processed using the Zeiss LSM image browser, and the image intensity adjusted using Photoshop. For the purpose of presentation, wild-type mCD8::GFP- and rCD2::RFP-positive multi-cellular neuroblast clones in the twin-spot MARCM experiments are shown in green for adPNs and lPNs and in magenta for vPNs in S2 Fig. The original LSM files used in the present study are available upon request.
Scoring of the dendritic patterns of DL1 adPNs was modified from a previous report [12]. Images of interest were projected from confocal stacks containing DL1 adPN dendrites using the Zeiss LSM image browser. Dendritic regions of DL1 adPNs were manually selected and the DL-to-VM axis of the AL was rotated to make it vertical by using Image J. Fluorescent signals were converted into binary numbers using Huang's fuzzy thresholding method provided by Image J [28]. The scoring region was selected and divided into ten bins along the DL-to-VM axis based on Brp-positive staining. The value of the dendritic pattern of DL1 adPNs within the scoring region was summed by Image J. Dendritic intensity and mean position within the scoring bins for dendrites of DL1 adPNs were then calculated. Student’s t-test was used for statistical analysis.
|
10.1371/journal.ppat.1007689 | Human metapneumovirus activates NOD-like receptor protein 3 inflammasome via its small hydrophobic protein which plays a detrimental role during infection in mice | NOD-like receptor protein 3 (NLRP3) inflammasome activation triggers caspase-1 activation-induced maturation of interleukin (IL)-1β and IL-18 and therefore is important for the development of the host defense against various RNA viral diseases. However, the implication of this protein complex in human metapneumovirus (HMPV) disease has not been fully studied. Herein, we report that NLRP3 inflammasome plays a detrimental role during HMPV infection because NLRP3 inflammasome inhibition protected mice from mortality and reduced weight loss and inflammation without impacting viral replication. We also demonstrate that NLRP3 inflammasome exerts its deleterious effect via IL-1β production since we observed reduced mortality, weight loss and inflammation in IL-1β-deficient (IL-1β-/-) mice, as compared to wild-type animals during HMPV infection. Moreover, the effect on these evaluated parameters was not different in IL-1β-/- and wild-type mice treated with an NLRP3 inflammasome inhibitor. The production of IL-1β was also abrogated in bone marrow derived macrophages deficient for NLRP3. Finally, we show that small hydrophobic protein-deleted recombinant HMPV (HMPV ΔSH) failed to activate caspase-1, which is responsible for IL-1β cleavage and maturation. Furthermore, HMPV ΔSH-infected mice had less weight loss, showed no mortality and reduced inflammation, as compared to wild-type HMPV-infected mice. Thus, NLRP3 inflammasome activation seems to be triggered by HMPV SH protein in HMPV disease. In summary, once activated by the HMPV SH protein, NLRP3 inflammasome promotes the maturation of IL-1β, which exacerbates HMPV-induced inflammation. Therefore, the blockade of IL-1β production by using NLRP3 inflammasome inhibitors might be a novel potential strategy for the therapy and prevention of HMPV infection.
| Human metapneumovirus (HMPV), a negative-stranded, enveloped RNA virus, is recognized as one of the leading causes of acute respiratory disease in children since its discovery in 2001. Nevertheless, there is currently no licensed vaccine for the prevention of HMPV infection and treatment modalities are limited to the use of ribavirin, a weak antiviral agent or immunoglobulins. NOD-like receptor protein 3 (NLRP3) inflammasome has been shown to be involved in the pathogenesis of several RNA viral diseases but its role during HMPV infection has not been fully studied. Here, we report for the first time that NLRP3 inflammasome is activated by the small hydrophobic protein of HMPV, leading to the release of IL-1β, which has the potential to exacerbate inflammation. However, NLRP3 inflammasome has no direct influence on viral replication. Thus, IL-1β-mediated inflammatory process plays an important role during HMPV infection and, therefore, anti-IL-1β strategies such as the use of NLRP3 inhibitors may be a novel potential approach for the prevention and therapy of HMPV disease.
| The inflammasomes are cytosolic multiprotein complexes responsible for caspase-1 activation [1]. Once activated, caspase-1 proteolytically cleaves interleukin (IL)-1β and IL-18 precursors (pro-IL-1β and pro-IL-18), leading to the release of mature forms [2, 3]. Among identified inflammasomes, the NOD-like receptor protein 3 (NLRP3) inflammasome containing NLRP3, adapter protein apoptosis-associated speck-like protein (ASC) and pro-caspase-1 is the most fully studied [4]. NLRP3 inflammasome activation is a two-step process. The first step involves a priming signal provided by microbial molecules or endogenous cytokines, which upregulates the transcription of inactive NLRP3, pro-IL-1β and pro-IL-18. The second step is characterized by the oligomerization of NLRP3 and subsequent assembly of NLRP3, ASC and pro-caspase-1 into a complex [5]. This signal is provided by numerous stimuli such as ATP, pore-forming toxins, viral RNA, etc. Most of them induce potassium efflux, calcium signaling, reactive oxygen species generation, mitochondrial dysfunction and lysosomal rupture [6].
The NLRP3 inflammasome has been demonstrated to be activated by many RNA viruses and could play distinct roles during viral infections [7, 8]. NLRP3 inflammasome activation has been reported to aggravate Newcastle virus, murine hepatitis virus, coxsackievirus B3, Dengue virus, and Zika virus infections [9–13] but exerts beneficial effects to the host response against enterovirus 71 and rabies virus diseases [14, 15]. Surprisingly, although NLRP3 inflammasome could be activated by vesicular stomatitis and encephalomyocarditis viruses, it seemed to have no influence on the pathogenesis of these two viruses [16]. In the case of influenza A virus infection, several studies investigating the role of NLRP3 inflammasome have yielded controversial results [17–20], leading to the conclusion that NLRP3 inflammasome might play a dual protective or detrimental role at different stages of influenza A virus infection [21].
Human metapneumovirus (HMPV) is a member of the Metapneumovirus genus within the new Pneumoviridae family of non-segmented, negative-stranded, enveloped RNA viruses [22]. This virus is one of the leading causes of respiratory tract disease in both children and adults. The adaptive immune response generated against HMPV is usually inefficient at protecting from reinfections, which are repeated throughout life [23]. There is currently no licensed vaccine to prevent HMPV infection and its treatment is still limited to the use of ribavirin, a weakly effective antiviral agent, and immunoglobulins [24]. Thus, highlighting the role of NLRP3 inflammasome during HMPV infection may provide a new perspective on the prevention and treatment of this viral disease. To date, only one study has reported increases in the production of IL-1β and IL-18, accompanied by an upregulation of NLRP3 mRNA expression in HMPV-infected children, as compared to control healthy individuals [25]. However, the authors did not clarify the role of NLRP3 inflammasome during HMPV infection.
In the current study, by using a pharmacological approach, small hydrophobic protein-deleted recombinant HMPV (HMPV ΔSH) as well as IL-1β-deficient (IL-1β-/-) mice, we show that NLRP3 inflammasome can be activated by HMPV SH protein. Once activated, this multiprotein complex exerts a deleterious effect during HMPV infection in mice by triggering IL-1β release. Therefore, targeting NLRP3 inflammasome as well as IL-1β may be of interest for the development of new therapeutics against HMPV infections.
MCC950 has been recently synthesized and recognized as a specific inhibitor of NLRP3, but not NLRP1, NLRC4 or AIM2 inflammasomes [26]. Since then, it has been preferentially used in various models of NLRP3-related diseases [27]. A recent study has reported that MCC950 did not impact the viability and proliferation of high-glucose-induced human retinal endothelial cells at a concentration of 100 μM [28]. In the current study, we also showed that this inhibitor was safe and usable for both human THP-1 (CC50 > 250 μM) and murine J774.2 (CC50 > 125 μM) cells (S1 Fig). To determine if NLRP3 inflammasome impacts HMPV replication, THP-1 or J774.2 cells were treated or not with 10 μM of MCC950 because this dose has been shown to be able to block NLRP3 activation in mouse bone marrow derived macrophages, human monocyte derived macrophages and human peripheral blood mononuclear cells [26], and then infected with HMPV. The viral loads evaluated on days 1, 2 and 3 post-infection did not differ between MCC950-treated and control DMSO-treated groups (Fig 1A). We also observed that viral titers were relatively low, but as expected since no trypsin was added during the cell culture assay. Thus, NLRP3 inflammasome had no influence on HMPV replication in vitro.
To investigate if NLRP3 inflammasome is responsible for IL-1β and IL-18 production, THP-1 or J774.2 cells were treated or not with MCC950 and then infected or not with HMPV. We found that NLRP3 inflammasome inhibition suppressed IL-1β and IL-18 secretion in THP-1 cells and significantly decreased their concentrations in J774.2 cells (Fig 1B). We also confirm that NLRP3 is responsible for the maturation of only IL-1β and IL-18 [29], as evidenced by no difference in IL-6 and TNF-α levels between MCC950-treated and control groups during HMPV infection (Fig 1C). Of note, we observed no IL-6 production in THP-1 cells upon HMPV inoculation. In order to confirm those results, we used wild-type (WT) bone marrow derived macrophages (BMDM) and NLRP3 KO BMDM cell lines and observed that HMPV-infected BMDM cells induced the production of IL-1β at 48 h, but not HMPVΔSH-infected BMDM (S2 Fig). Conversely, no IL-1β was detected in NLRP3 KO BMDM cells following HMPV infection. Notably, TNF-α was detected in both WT and NLRP3 KO BMDM cells.
To investigate if NLRP3 inflammasome is involved in the pathogenesis of HMPV, BALB/c mice were treated with MCC950 and infected with HMPV at a LD50 dose. No mortality and a less important weight loss were observed in MCC950-treated groups, as compared to controls (Fig 2A). Nevertheless, the protective effect of MCC950 treatment was slightly decreased when treatment was delayed 24 h post-infection (Fig 2B), in comparison with immediate treatment (Fig 2A). Thus, NLRP3 inflammasome plays a detrimental role during HMPV infection and the blockade of its activation may be useful not only for the prevention but also for the therapy against HMPV disease, at least in mice.
We then investigated the effect of NLRP3 inflammasome on HMPV replication by determining viral loads in the lungs. In agreement with in vitro results, mice inoculated with HMPV at sublethal or LD50 doses both showed no difference in lung viral titers between MCC950-treated and control groups (Fig 2C). Therefore, it is likely that the involvement of NLRP3 inflammasome in the pathogenesis of HMPV does not occur via a direct viral replication-related pathway.
Because NLRP3 inflammasome did not impact HMPV replication, we hypothesized that it possibly exerts a deleterious effect via IL-1β and/or IL-18-dependent pathways. To verify this hypothesis, we measured IL-1β, IL-18 and other cytokine levels in BAL at different time points from BALB/c mice infected with HMPV at sublethal or LD50 doses and treated or not with MCC950. IL-18 levels were not different between MCC950-treated and DMSO-treated mice. By contrast, IL-1β levels in HMPV-infected mice were significantly decreased upon MCC950 treatment on day 1 post-infection (Fig 3). As presented above, NLRP3 inflammasome did not impact HMPV-induced IL-6 and TNF-α secretion in vitro. By contrast, interferon (IFN)-γ, IL-6, and TNF-α levels in BAL were significantly decreased upon MCC950 treatment (Fig 3). In parallel, NLRP3 inflammasome inhibition also reduced the alteration of pulmonary capillary permeability and leukocyte recruitment, as evidenced by significant decreases in total protein levels and cell number in BAL from MCC950-treated mice compared to controls during HMPV infection (Fig 4A and 4B).
Thus, NLRP3 inflammasome inhibition protects mice against HMPV disease by exerting an anti-inflammatory effect. Moreover, this anti-inflammatory effect seems to be virus dose-dependent because NLRP3 inflammasome inhibition decreased more efficiently inflammatory parameters in the case of sublethal dose than LD50 dose of virus. Indeed, MCC950 treatment induced significant decreases in IL-6 and TNF-α levels on day 1 post-infection and total protein levels on days 3 post-infection (sublethal dose) and 5 post-infection (sublethal and LD50 doses) (Figs 3 and 4A).
We then investigated if NLRP3 inflammasome impacts the recruitment of particular cell type(s). Only lymphocyte percentage was decreased on day 3 post-infection upon MCC950 treatment accompanied by an increase in macrophage percentage in mice inoculated with HMPV at a LD50 dose (Fig 4C). Lymphocytes decrease on day 3 was characterized by a reduction in % of B and CD8 T cells (S3 Fig). A time-dependent different contribution of each cell type during HMPV infection was detected, as evidenced by the predominance of polymorphonuclear neutrophils on day 1, and then lymphocytes on days 3 and 5 post-infection.
Since IL-18 levels were unaltered during in vivo infection, we further determined if this cytokine exerts some effects during HMPV disease. IL-18 was therefore inhibited in HMPV-infected BALB/c mice by using IL-18 binding protein (IL-18BP), which functions as an IL-18 antagonist by binding to IL-18 and blocking its biological activities [30–32]. No difference in survival and weight loss between IL-18BP-treated and non-treated mice was observed during HMPV infection (Fig 5A). Thus, the involvement of NLRP3 inflammasome in the pathogenesis of HMPV is IL-18-independent. Interestingly, we noticed that NLRP3 inflammasome inhibition always protected mice against HMPV disease even if virus was administered at a lethal dose, as evidenced by no mortality and slight weight loss (<10% initial weight) in MCC950-treated compared to DMSO-treated mice. This finding enables us to suggest that NLRP3 inflammasome is essential for the pathogenesis of HMPV in BALB/c mice.
Because IL-18 was not required for the infectivity of HMPV, we determined if the implication of NLRP3 inflammasome in the pathogenesis of this virus could be associated with the release of IL-1β. Therefore, C57BL/6 (IL-1β+/+) and IL-1β-/- mice were infected with HMPV at a LD50 dose. MCC950-treated IL-1β+/+ and untreated IL-1β-/- mice showed less weight loss and mortality, reduced IFN-γ and total protein levels, as well as leukocyte number in BAL, as compared to IL-1β+/+ mice without MCC950 treatment on day 5 post-infection (Fig 5B–5E). Other parameters including IL-1β, IL-6, TNF-α, IL-18, leukocyte differentiation and viral replication did not differ between IL-1β+/+ mice, IL-1β-/- and IL-1β+/+ treated with MCC950 mice (S4 Fig). Thus, the involvement of NLRP3 inflammasome in the pathogenesis of HMPV seems to predominantly occur via IL-1β secretion.
To investigate whether HMPV SH protein is responsible for NLRP3 inflammasome activation, we first designed and generated HMPV ΔSH virus from strain C85473. We then evaluated caspase-1 cleavage as a marker of NLRP3 inflammasome activation because caspase-1 cleavage depends on the assembly of NLRP3, ASC and procaspase-1 to form inflammasome [5]. Western Blot analysis showed that HMPV inoculation induced caspase-1 cleavage in THP-1 cells (Fig 6). This cleavage, however, was prevented by MCC950 treatment. These results confirm the capacity of HMPV to induce NLRP3 inflammasome. In parallel, we found that HMPV ΔSH could not induce caspase-1 cleavage in THP-1 cells (Fig 6). Thus, SH protein seems to be the viral component triggering NLRP3 inflammasome activation, but other experiments are needed to determine if the mutant virus simply becomes inaccessible to pro-inflammatory danger sensors via compartmentalization or if it is physically unable to prime or activate the inflammasome. Cleaved caspase-1 was detected in both cell lysates and supernatants at 1 h but was only present in the cell supernatants at 24 h post-infection. This indicates that caspase-1 was rapidly cleaved and released into the supernatants [33] upon HMPV inoculation.
We also investigated if the absence of SH protein could attenuate the infectivity of HMPV. No difference in viral replication between HMPV and HMPV ΔSH was observed in THP-1 cells (Fig 7A) but IL-1β production was abrogated in HMPV ΔSH-infected THP-1 cells (Fig 7B). The same tendency was observed during in vivo studies. Indeed, at the image of MCC950-treated mice, no mortality was seen in HMPV ΔSH-infected mice whereas slight weight loss as well as reduced IFN-γ, IL-6, total protein levels, leukocyte numbers in BAL and lung histopathological scores were observed, compared to HMPV-infected mice on day 5 post-infection (Fig 7C–7G & S5A Fig). No difference in IL-1β and TNF-α levels as well as leukocyte differentiation was observed on day 5 between HMPV ΔSH- and HMPV-infected mice (S5B and S5C Fig). However, a significant decrease of IL-1β was seen on day 1 in the HMPV ΔSH group (Fig 3). Thus, both NLRP3 inflammasome inhibition and SH protein deletion attenuated inflammation and lung injury. Moreover, SH protein did not impact on viral replication [34, 35], as demonstrated by no difference in the viral loads of lungs between and HMPV ΔSH- and HMPV-infected groups (Fig 7H). Altogether, we conclude that in the case of HMPV infection, NLRP3 inflammasome activation is triggered by the viral SH protein.
Our study clearly shows the role of the inflammasome and in particular IL-1β in the pathogenesis of HMPV using a pharmacological approach, BMDM NLRP3 KO cells and IL-1β-/- mice. As a crucial component of the innate immune system, NLRP3 inflammasome serves an important role in host defense by recognizing RNA viral pathogens and triggering immune responses [36]. Although NLRP3 inflammasome has been reported to be implicated in many RNA viral diseases with distinct functions [7, 8], little is known about the involvement of this protein complex in the pathogenesis of HMPV. In such a context, this present study shows for the first time that NLRP3 inflammasome plays a detrimental role during HMPV infection and that such effect is mediated by the viral SH protein.
The contribution of NLRP3 inflammasome in the pathogenesis of RNA viruses occurs through its role as a trigger of only inflammation [12, 15, 18, 19] or both inflammation and viral replication [9, 10, 16, 17, 37]. Herein, we demonstrate that the involvement of NLRP3 inflammasome in the pathogenesis of HMPV only proceeds via its pro-inflammatory effect. Indeed, both in vitro and in vivo studies showed that viral replication was almost unaltered whereas inflammation was attenuated upon NLRP3 inflammasome inhibition during HMPV infection. This finding is consistent with two other studies which have also shown that NLRP3 inflammasome did not impact on viral replication during influenza and chikungunya diseases [21, 38]. Nevertheless, the suppression of NLRP3 inflammasome has been demonstrated to decrease fulminant hepatitis and Zika virus replication [10, 13] but increase Newcastle virus replication [9]. In other words, NLRP3 inflammasome plays distinct roles in the replication of RNA viruses.
NLRP3 inflammasome, once activated, will promote caspase-1-induced IL-1β and IL-18 maturation [29], but not other cytokines. In this study, NLRP3 inflammasome-independent secretion of IL-6 was observed in HMPV-infected J774.2 cells and TNF-α in THP-1 and J774.2 cells. This finding is consolidated by a recent in vitro study investigating RSV [39], the closest virus related to HMPV [40] also belonging to the Pneumoviridae family [22]. The authors reported that the secretion of IL-1β, not IL-6 was triggered by RSV-induced NLRP3 activation in primary human lung epithelial cells.
Although NLRP3 inflammasome is responsible for the secretion of only IL-1β and IL-18 in infected cells, the inhibition of this protein complex decreased the levels of not only IL-1β but also IL-6, TNF-α and IFN-γ in HMPV-infected mice. Furthermore, three other inflammatory parameters including the alteration of pulmonary capillary permeability, leukocyte recruitment and lung histopathological scores were also decreased upon NLRP3 inflammasome inhibition. These data are not unique and they are consistent with several previous reports [17–19, 21, 26, 38]. Thus, NLRP3 inflammasome may impact not only IL-1β and IL-18 secretion but also exert proinflammatory effects via unknown pathways [18] during RNA viral diseases in general and HMPV infection in particular.
To explain the proinflammatory function of NLRP3 inflammasome, it has been suggested that its activation may occur in concert with other proinflammatory pathways such as lipotoxicity-, oxidative stress- and TLR4-related pathways [41]. We suggest that NLRP3 inflammasome exerts pro-inflammatory effect during HMPV infection through biological activities of IL-1β. Indeed, this cytokine has been identified as an important regulator of inflammation, as evidenced by its capacity to stimulate neutrophil and macrophage recruitment and infiltration in some conditions [42, 43] and induce lung vascular permeability damage [44]. IL-1β has also been identified as an activator of IL-6 and IL-8 production [45, 46]. Most importantly, we reported that both NLRP3 inflammasome inhibition (BALB/c and C57BL/6 mice) and deletion of the gene encoding IL-1β (C57BL/6 mice) induced less weight loss with decreased mortality and inflammation in HMPV-infected mice. Moreover, the protective effect against HMPV disease did not differ between NLRP3 inflammasome inhibition and IL-1β deletion (C57BL/6 mice). Briefly, NLRP3 inflammasome-induced IL-1β release plays a crucial role during HMPV infection, at least in mice.
In the case of BALB/c mice, no mortality was found in animals infected with HMPV at a LD50 dose and treated with MCC950 (Fig 2A). By contrast, some mortality (25%) was detected in infected C57BL/6 mice given MCC950 treatment (Fig 5B). Furthermore, NLRP3 inflammasome inhibition was found to decrease IL-6 and IFN-γ levels in BAL from BALB/c mice (Fig 3) but only IFN-γ levels in the case of C57BL/6 mice on day 5 post-infection (Fig 5C). In parallel, IL-6, IFN-γ and TNF-α levels in BAL from BALB/c mice were strongly higher than those from C57BL/6 mice (Figs 3 and 5C & S4A Fig). All these findings indicate that HMPV-induced inflammation is more severe in BALB/c than C57BL/6 mice and that the role of NLRP3 inflammasome and IL-1β is more important in the former mice during HMPV infection. The different susceptibility of these two murine strains to HMPV [47] is a possible explanation.
Although NLRP3 inflammasome activation triggers the maturation of IL-1β and IL-18, we show that only IL-1β is involved in the pathogenesis of HMPV. Moreover, this involvement occurs at an early stage of infection process because IL-1β levels were only decreased upon NLRP3 inhibitor treatment on day 1 post-infection. This finding is not surprising since IL-1β as well as other IL-1 family cytokines are widely considered as early-response cytokines as they are released in the earliest stage of an immune response [48]. By contrast, IL-18 secretion was unaltered during infection and had no influence on the pathogenicity of HMPV in mice. This might be explained by the limited presence of macrophages during HMPV infections, which were shown to be an abundant source of IL-18 during in vitro studies [49] (Fig 1B). By contrast, polymorphonuclear neutrophils were abundant on day 1 and then lymphocytes were dominant on days 3 and 5 post-infection (Fig 4C).
NLRP3 inflammasome activation generally employs a two-step mechanism. In general, the first signal permitting the generation of pro-IL-1β and pro-IL-18 is triggered by the recognition of viral pathogens by Toll-like receptors (TLRs) or retinoic acid-inducible gene-I-like receptors [5]. Herein, we did not investigate the mechanisms by which the first signal occurs. However, we think that the TLR4 receptor may be responsible for this process because a recent study has demonstrated that among various TLRs including TLR2, TLR3, TLR4, TLR7 and TLR8, only TLR4 provides the first signal of NLRP3 inflammasome activation in RSV-infected lung epithelial cells [39]. Furthermore, TLR4-/- mice induced less weight loss, decreased inflammation and no difference in viral replication, as compared to wild-type mice during HMPV infection [50]. These findings are consistent with our results when using either IL-1β-/- mice or pharmacological approach for blocking NLRP3 inflammasome activation.
Recently, it has been shown that RNA viruses trigger NLRP3 inflammasome activation through a receptor interacting protein (RIP) 1/RIP3/dynamin-related protein 1 signaling pathway [51, 52]. Briefly, RNA virus infection initiates the assembly of RIP1/RIP3 complex, promoting activation of dynamin-related protein 1 and its translocation to mitochondria. This results in mitochondria damage, excessive reactive oxygen species generation and subsequent NLRP3 inflammasome activation [6]. RNA viral components responsible for triggering this pathway have been identified in several viral infections such as influenza virus M2 and PB1-F2 proteins [53–55], Measles virus V protein [56], RSV SH protein [39], encephalomyocarditis virus and rhinovirus 2B proteins [57, 58], coronavirus E protein [59] and enterovirus 71 3D protein [60]. Among these proteins, encephalomyocarditis virus and rhinovirus 2B proteins as well as RSV SH protein were recognized as viroporins. Viroporins from RNA viruses have been reported to be responsible for mitochondrial alteration [61]. Furthermore, once inserted on host cell membrane, viroporin will enable virus to tune ion permeability to stimulate a variety of viral cycle stages [62]. Taken together, we hypothesized that HMPV SH protein may be an activator of NLRP3 inflammasome during HMPV disease because it has been suggested to act as a viroporin [63] and the genomic structure of HMPV is closely related to that of RSV [40].
HMPV ΔSH viruses have been previously reported to be generated from the CAN97-83 (group A) or NL/1/99 (group B) strains [34, 35]. Here, we generated HMPV ΔSH using C85473 strain (group A) to verify our hypothesis [64]. In addition to the blockade of caspase-1 cleavage, a reliable marker of NLRP3 inflammasome activation, resulting from the lack of SH protein or MCC950 treatment, we found that identically to MCC950 treatment, SH deletion had no effect on the viral replication both in vitro and in vivo. This finding is consolidated by previous studies using other HMPV ΔSH viruses [34, 35]. We also detected that HMPV ΔSH-infected mice were protected against severe infection, as evidenced by no mortality, less weight loss and reduced inflammation. The evolution of HMPV infections was not different between HMPV ΔSH- and wild-type HMPV-infected mice receiving MCC950 treatment, but we acknowledge that no measures were taken to detect defective interfering particles in both viral preparations. In parallel, NLRP3 inflammasome inhibition had no influence on the pathogenesis of HMPV ΔSH. Taken together, we conclude that HMPV SH protein might be an activator of NLRP3 inflammasome in addition to its identified other roles to modulate type I IFN signaling pathway [65, 66], deteriorate cell host membrane permeability, regulate viral fusogenic function [63] and reduce CD4+ T cell activation [67].
In summary, we report for the first time a detrimental role of NLRP3 inflammasome during HMPV infection in murine models. Mechanistically, HMPV SH protein triggers NLRP3 inflammasome activation, leading to the cleavage of pro-IL-1β to form mature IL-1β. Although this cytokine is not crucial for controlling viral replication, it plays a major role in inflammatory process which is identified as an important feature for the pathogenicity of HMPV. Thus, the involvement of NLRP3 inflammasome in HMPV disease occurs via IL-1β-related inflammatory process rather than virus replication. In such a context, we believe that anti-inflammatory treatments in general and anti-IL-1β drugs in particular (i.e. the use of NLRP3 inhibitors) may be considered as novel potential strategies for the prevention and treatment of HMPV disease.
Six-week old female BALB/c mice with a body weight of 16.5–18 g were purchased from Charles River Laboratories (Senneville, QC, Canada). IL-1β-/- mice were kindly provided by Dr Steve Lacroix (Infectious Disease Research Centre, Quebec City, QC, Canada). Age-matched wild-type C57BL/6 mice were purchased from Charles River Laboratories. Mice were housed under pathogen-free conditions in the animal research facility of the Quebec University Health Centre (Quebec City, QC, Canada) and allowed to acclimatize for one week prior to the start of experiments.
All experimental procedures with mice were approved by the Animal Protection Committee of the Quebec University Health Centre in accordance with guidelines of the Canadian Council on Animal Care (Protocol number: CPAC 2017-139-1). Before inoculation of substances or euthanasia, mice were anaesthetized by inhalation of isoflurane vaporized at concentrations of 3–4% and oxygen flow rate adjusted to 1.5 l/min. Individual body weight and clinical signs were used to monitor animal health and response to infection and were recorded daily. Mice were euthanized by CO2 inhalation upon loss of 20% of initial body weight.
LLC-MK2 cells (ATCC, Manassas, VA, USA) were maintained in minimal essential medium (Thermo Fisher Scientific, Burlington, ON, Canada) supplemented with 10% fetal bovine serum (FBS) (Wisent, Saint-Jean-Baptiste, QC, Canada) and HEPES buffer (2.5 g/l). Murine macrophage J774.2 cells were kindly provided by Dr Sachiko Sato (Infectious Disease Research Centre, Quebec City, QC, Canada) and maintained in Dulbecco’s modified eagle medium (Thermo Fisher Scientific) supplemented with 10% FBS and 1% penicillin-streptomycin (Thermo Fisher Scientific). Human monocyte-like THP-1 cells were generously provided by Dr Francesca Cicchetti (Quebec University Health Centre, Quebec City, QC, Canada) and maintained in RPMI 1640 medium (Thermo Fisher Scientific) supplemented with 10% FBS, 1% penicillin-streptomycin, 1% non-essential amino acids (Thermo Fisher Scientific) and 0.05 mM 2-mercaptoethanol (Sigma Aldrich, Oakville, ON, Canada). Differentiation of THP-1 cells into macrophages by the addition of phorbol 12-myristate 13-acetate (100 ng/ml) [68] (Sigma Aldrich) was carried out in all in vitro experiments. Immortalized murine bone-marrow derived macrophages WT (BMDM iWT) or NLRP3 -/- (BMDM iNLRP3KO) were kindly provided by Dr Bénédicte Py (Centre International de Recherche en Infectiologie CIRI, Lyon, France) and maintained in Dulbecco’s modified eagle medium (Thermo Fisher Scientific) high glucose, supplemented with 10% FBS and 1% penicillin-streptomycin (Thermo Fisher Scientific).
The HMPV strain C85473, a clinical isolate, and the recombinant HMPV strain C85473 ΔSH were grown in LLC-MK2 cells and concentrated as previously described [69]. Viruses were concentrated by ultracentrifugation and pellets resuspended in PBS. Viral stocks were sequenced and titers were determined by immunostaining [70] and expressed as plaque-forming units (PFU) per milliliter.
The strategy for the construction of the plasmid encoding the full-length genomic cDNA of HMPV A1/C-85473 strain (GenBank accession number KM408076.1) and the subsequent production of recombinant viruses is described in details in [64]. Briefly, the full-length genomic cDNA of HMPV A1/C-85473 strain was generated by RT-PCR using the Superscript II reverse transcriptase (Thermo Fisher Scientific) and amplified by Phusion DNA polymerase (New England Biolabs, Whitby, ON, Canada). A Gibson Assembly (Cloning Kit, New England Biolabs) was performed to integrate the genomic viral cDNA into a pSP72 plasmid (Promega, Madison, WI, USA) containing a T7 terminator, the hepatitis delta virus (HDV) ribozyme and a T7 promoter. To generate HMPV ΔSH virus, the mentioned plasmid was amplified using specific primers, designed to match before the SH gene start sequence (5’-GGGACAAGTAGTTATGGA-3’) and after the intergenic SH-G region (5’-ACTCTGATGTGTTTTTACTAAC-3’), in order to extract completely the SH gene sequence. After amplification, linear DNA was phosphorylated and ligated with the T4 Ligase to re-circularize the shortened HMPV genome. The newly generated HMPV ΔSH genomic plasmid was validated by complete sequencing prior to transfection. BSR-T7 cells were then co-transfected (Lipofectamin 2000, Thermo Fisher Scientific) with the HMPV ΔSH genomic plasmid and 4 supporting plasmids expressing the N, P, L, and M2.1 viral ORFs. Seven hours after transfection, the medium was replaced by Opti-MEM supplemented with 1% Non-Essential Amino Acids (Thermo Fisher Scientific). Transfected cells were incubated at 37°C and 5% CO2 for four days. At this point, LLC-MK2 cells were added to the transfected BSR T7 cells and co-cultured at 37°C and 5% CO2 with the addition of fresh trypsin (0.0002%) after two days. Two or three days after co-culture, cells were harvested, sonicated and centrifuged at 2000 x g for 5 min at room temperature. The supernatant was collected, diluted into infection medium (Opti-MEM supplemented with 0.0002% trypsin), and inoculated onto confluent LLC-MK2 monolayers. Infected monolayers were monitored for the appearance of characteristic cytopathic effect, and the harvested virus was further amplified through serial passages in LLC-MK2 cells.
The CC50 concentration of NLRP3 inhibitor MCC950 (Tocris Bioscience, Bristol, UK) was determined in J774.2 and THP-1 cells using the CellTiter 96 Aqueous One Solution Cell Proliferation Assay (Promega) according to the manufacturer’s instructions.
J774.2 or THP-1 cells were treated with 10 μM of MCC950 [26] and incubated at 37°C for 1.5 h. Equivalent dilutions of dimethyl sulfoxide (DMSO) (Sigma Aldrich) served as control. The cells were then inoculated with wild-type HMPV or HMPV ΔSH at a MOI of 0.001 (THP-1) or 0.01 (J774.2) per well and incubated at 37°C. Cell lysates and supernatants were harvested at 1; 24; 48 and 72 h post-infection for ELISA or Western Blot analyses. In addition, the viral titers were determined by immunostaining three days post-infection and expressed as PFU per milliliter. BMDM iWT and iNLRP3KO cells in 24-well plates were washed in PBS and inoculated with Opti-MEM (mock), wild-type HMPV or ΔSH HMPV at a MOI of 0.1 in Opti-MEM + 0.0002% trypsin. After 3 h adsorption at 37°C, inocula were removed and changed by fresh cell culture medium DMEM high glucose + 10% FBS and 1% penicillin-streptomycin (Thermo Fisher Scientific). Cell supernatants were harvested at 1, 24, 48 and 72 h post-infection to perform IL-1β or TNF-α quantification by ELISA assays (DuoSet ELISA, R&D Systems), according to manufacturer’s instructions, and viral titrations, as previously described [70].
A preliminary study has been carried out to determine the sublethal, LD50 and lethal doses of virus in mice. BALB/c mice were inoculated intranasally with HMPV strain C85473 (sublethal dose = 3 x 105; LD50 = 5 x 105 or lethal dose = 106 PFU per mouse) or HMPV ΔSH whereas IL-1β-/- and wild-type C57BL/6 mice were inoculated with 2 x 106 PFU of HMPV strain C85473. The LD50 dose in C57BL/6 mice was four-fold higher than that of BALB/c mice because C57BL/6 mice are less susceptible to HMPV infection, as compared to BALB/c mice [47]. Equal volumes of Opti-MEM medium served as mock infection.
To block NLRP3 inflammasome activation, MCC950 (5 mg/kg) [21, 71] was mixed and administered intranasally at the same time with the virus. However, MCC950 was also given 24 h post-infection in a single experiment. Equivalent dilutions of DMSO (Sigma Aldrich) served as control. This treatment was repeated once a day for two consecutive days. For IL-18 inhibition, immediately following inoculation of virus, mice underwent intraperitoneal injections of IL-18BP at a dose of 75 μg/kg (R&D Systems, Minneapolis, MN, USA) [72]. This treatment was repeated once a day for two consecutive days during infections. Control mice were given sterile saline in a similar manner.
To evaluate viral titers on days 1, 3 and 5 post-infection, mice were euthanized and whole lungs were harvested and then homogenized in PBS (1 ml/sample) using TH Tissue Homogenizer (Omni International, Kennesaw, GA, USA). Supernatants were collected after centrifugation at 350 x g for 10 minutes at 4°C and methylcellulose was used to determine viral titers by immunostaining and expressed as PFU per gram of lung.
On days 1, 3 and 5 post-infection, mice were euthanized and broncho-alveolar lavage (BAL) was performed with sterile cold phosphate-buffered saline (PBS). The cells in the lavage fluid were pelleted by centrifugation at 300 x g for 5 min at 4°C, and then suspended in PBS whereas BAL supernatants were collected for evaluating other inflammatory parameters. Viable cell number was determined using a hemocytometer and expressed as number per milliliter of BAL.
For differential cell counts, 100 μl of suspended cells were spun onto a slide by using a Shandon Cytospin 3 cytocentrifuge (Thermo Fisher Scientific) at 100 x g for 5 min at room temperature. Slides were then air-dried and stained with May-Grunewald Giemsa solutions (Sigma Aldrich) according to the manufacturer’s instructions. Differential cell counts were made with standard morphological criteria by counting at least 300 cells per sample. The results were expressed as differential percentage.
The concentrations of IL-1β, IL-6, TNF-α, IFN-γ and IL-18 in the cell supernatants or BAL fluids were determined using the Mouse or Human IL-1β, IL-6, TNF-α, IFN-γ, IL-18 DuoSet ELISA (R&D Systems) or the Mouse IL18/IL-18 ELISA Pair Set (Sino Biological, Beijing, China) according to the manufacturer’s instructions. The results were expressed as picogram per milliliter of BAL.
Total protein levels in the BAL supernatants were determined using Quick Start Bradford Protein Assay (Bio-Rad Laboratories, Mississauga, ON, Canada) according to the manufacturer’s instructions. The results were expressed as milligram per milliliter of BAL.
In order to analyze lung-infiltrating immune cells, mice were deeply anesthetized and perfused intracardially with D-PBS without Ca2+ and Mg2+ prior to (day 0) and on days 1, 3 and 5 post-infection. Whole lungs were collected and digested with Liberase TL (Roche Diagnostics, Mannheim, Germany). Lung homogenates were incubated for 1 h at 37°C then filtered through a 70-μm cell strainer (BD Biosciences, Mississaugo, ON, Canada). The cell suspension was centrifuged at 300 x g for 10 min at room temperature. The supernatant was aspirated and cells were washed twice with D-PBS plus 2% FBS. Cells were first incubated on ice for 30 min with fixable viability stain 510 (BD Biosciences, CA, USA), then washed and incubated again on ice for 30 min with purified rat anti–mouse CD16/CD32 (Mouse Fc Block; BD Biosciences, CA, USA). Red blood cells were lysed with BD Pharm Lyse (RBC Lysis Buffer 10X –BioLegend, San Diego, CA, USA) and the recovered leukocytes were washed and resuspended in D-PBS. After the washing step, cells were incubated on ice for 40 min with a pool of antibodies (anti-CD45, anti-CD11b, anti-CD170 (Siglec-F), anti-Ly6C, anti-Ly6G, anti-CD11c, anti-CD115, anti-B220, anti-CD3ε, anti-CD4 and anti-CD8a /BD Bioscience, CA, USA). Number of cells was determined with Precision Count Beads (BioLegend, San Diego, CA, USA). Labeled cells were then washed and resuspended in DPBS. Flow cytometry analyses and data acquisition were performed by using a BD SORP LSR II and the BD FACSDiva software, respectively.
The total proteins in cell supernatants were concentrated using Amicon Ultra-15 Centrifugal Filters (Millipore Canada, Etobicoke, ON, Canada) according to the manufacturer’s instructions. The concentrations of protein in cell lysates and supernatants were determined using Quick Start Bradford Protein Assay. Equal protein amounts were separated on 10% SDS-PAGE gels and then transferred to nitrocellulose membranes (GE HealthCare Life Sciences, Mississauga, ON, Canada) and blocked using 5% BSA (Sigma Aldrich). Primary antibodies were used at a dilution of 1:1,000 goat anti-caspase-1 (R&D Systems); rabbit anti-cleaved caspase-1 (p20) or mouse anti-α-tubulin (Cell Signaling Technology, Boston, MA, USA). Secondary antibodies were used at a dilution of 1:1,000 HRP-conjugated rabbit anti-goat (R&D systems) or 1:5,000 HRP-conjugated rabbit anti-mouse or mouse anti-rabbit (Cell Signaling Technology). Signal detection was carried out using the West Pico Plus Chemiluminescent Substrate (Thermo Fisher Scientific).
On day 5 post-infection, mice were euthanized and their lungs were removed. Tissue was fixed in 4% paraformaldehyde, embedded in paraffin, sectioned in slices of 5 μm, and stained with hematoxylin and eosin. Slides were digitalized at 40X magnification using a Nanozoomer slide scanner (Hamamatsu, Japan) and scored using NDP viewer 2.0 software (Hamamatsu, Japan). The histopathological scores were determined by a pathologist and a medical biologist who were blinded to the experimental data. A semi-quantitative scale was used to score bronchial/endobronchial, peribronchial, perivascular, interstitial, pleural and intra-alveolar inflammation [73]. Scores represent consensus between the two observers. The results were expressed as lung total inflammatory scores.
All statistical tests were conducted using the GraphPad Prism version 6.0 (GraphPad Software, La Jolla, CA, USA). The results were expressed as the mean ± S.E.M for each group and 'n' referred to the sample size. Survival data were analyzed by comparing Kaplan-Meier curves using the log-rank test. Viral titers, cytokines and total protein levels, immune cell recruitment, cell differentiation as well as lung histopathological scores were analyzed using unpaired Student t-test, Mann-Whitney U-test, one-way analysis of variance (ANOVA) followed by Tukey post hoc or Kruskal-Wallis test followed by Dunn’s post hoc for multiple comparisons. Differences were considered statistically significant when P < 0.05.
|
10.1371/journal.pcbi.1006256 | Proteome-scale relationships between local amino acid composition and protein fates and functions | Proteins with low-complexity domains continue to emerge as key players in both normal and pathological cellular processes. Although low-complexity domains are often grouped into a single class, individual low-complexity domains can differ substantially with respect to amino acid composition. These differences may strongly influence the physical properties, cellular regulation, and molecular functions of low-complexity domains. Therefore, we developed a bioinformatic approach to explore relationships between amino acid composition, protein metabolism, and protein function. We find that local compositional enrichment within protein sequences is associated with differences in translation efficiency, abundance, half-life, protein-protein interaction promiscuity, subcellular localization, and molecular functions of proteins on a proteome-wide scale. However, local enrichment of related amino acids is sometimes associated with opposite effects on protein regulation and function, highlighting the importance of distinguishing between different types of low-complexity domains. Furthermore, many of these effects are discernible at amino acid compositions below those required for classification as low-complexity or statistically-biased by traditional methods and in the absence of homopolymeric amino acid repeats, indicating that thresholds employed by classical methods may not reflect biologically relevant criteria. Application of our analyses to composition-driven processes, such as the formation of membraneless organelles, reveals distinct composition profiles even for closely related organelles. Collectively, these results provide a unique perspective and detailed insights into relationships between amino acid composition, protein metabolism, and protein functions.
| Low-complexity domains in protein sequences are regions that are composed of only a few amino acids in the protein “alphabet”. These domains often have unique chemical properties and play important biological roles in both normal and disease-related processes. While a number of approaches have been developed to define low-complexity domains, these methods each possess conceptual limitations. Therefore, we developed a complementary approach that focuses on local amino acid composition (i.e. the amino acid composition within small regions of proteins). We find that high local composition of individual amino acids is associated with pervasive effects on protein metabolism, subcellular localization, and molecular function on a proteome-wide scale. Importantly, the nature of the effects depend on the type of amino acid enriched within the examined domains, and are observable in the absence of classically-defined low-complexity (and related) domains. Furthermore, we define the compositions of proteins involved in the formation of membraneless, protein-rich organelles such as stress granules and P-bodies. Our results provide a coherent view and unprecedented resolution of the effects of local amino acid enrichment on protein biology.
| Low-complexity domains (LCDs) in proteins are regions enriched in only a subset of possible amino acids. LCDs can be composed of homopolymeric repeats of a single amino acid, short tandem repeats consisting of only a few different amino acids, or aperiodic stretches with little amino acid diversity [1]. Proteins containing LCDs are relatively common among organisms from all domains of life, and are particularly common among eukaryotes [2–4]. For example, approximately 70% of genes in the Saccharomyces cerevisiae genome possess at least one classically-defined LCD [3]. Furthermore, the total number of LCDs far exceeds the total number of yeast genes (~2-fold more LCDs than genes), indicating that many genes contain multiple distinct LCDs.
Various methods have been developed to assess biopolymer sequence complexity [1,5–9]. One of the most commonly employed methods to define LCDs is the SEG algorithm [1], which scans protein (or nucleic acid) sequences using a short sliding window, and calculates the local Shannon entropy for each window (see [10] for a detailed description). Subsequences with a Shannon entropy value below a pre-determined “trigger” threshold are classified as LCDs. LCD boundaries are later extended and refined by merging overlapping LCDs and calculating combinatorial sequence probabilities. Another metric commonly used to assess relative sequence complexity is compositional bias, which involves determining the statistical probability of a sequence given whole-proteome frequencies of the individual amino acids [11,12]. These approaches (or closely-related approaches) have been used extensively to examine LCDs on a proteome-wide scale [1,3,12–17].
LCD-containing proteins have been implicated in a variety of normal and pathological cellular processes. For example, Q/N-rich yeast proteins often play a role in transcription regulation, endocytosis, and cell cycle regulation, among other functions [11,18]. Many proteins containing Q/N-rich LCDs, or LCDs of related types (Q/N/G/S/Y-rich LCDs) have been linked to prion or prion-related processes [11,18–21]. Additionally, many prion-like LCDs, which are often composed of short tandem repeats of low-complexity [22], have been linked to stress granules and processing bodies (P-bodies) in eukaryotes (see [23] for recent review). The amino acid composition of these LCDs confers unusual biophysical properties to these domains [24], which likely relates to their unique behavior in vitro and in vivo [25–30]. However, these unusual characteristics appear to be inseparably linked to pathological processes as well. For example, genetic expansion of regions encoding homopolymeric glutamine repeats (the simplest type of LCD) in various proteins can lead to a multitude of neurodegenerative disorders, including Huntington’s Disease and spinocerebellar ataxias (for review, see [31]). Furthermore, mutations in the LCDs of stress granule proteins can alter stress granule dynamics and lead to degenerative diseases [26,28,30,32,33]. The importance of LCDs extends well beyond Q/N-rich LCDs, as LCDs of other compositions have also been linked to normal and pathological cellular processes [12,14,17,34,35].
Although LCDs can clearly impact protein regulation and function, a number of challenges have thus far limited a proteome-scale understanding of these relationships. One major challenge lies in defining LCDs. Current approaches use statistically-defined thresholds for sequence complexity or compositional bias [1,11], or arbitrarily-chosen repeat lengths for proteins with homopolymeric repeats [34–41]. Although these definitions of LCDs, compositionally biased sequences (herein referred to as “statistically-biased domains” to avoid later confusion), or homopolymeric repeats have facilitated important discoveries, the biological relevance of these thresholds has not been rigorously examined. Furthermore, these proteins are often grouped into a single class even though their compositions, and therefore physical properties, can differ dramatically (a limitation that was appreciated in a recent review [42]).
To address these limitations, we have developed an alternative approach to infer relationships between amino acid composition and protein metabolism and function. By focusing on amino acid composition, which is the fundamental feature underlying both sequence complexity and statistical amino acid bias, we examined links between local compositional enrichment and various aspects of protein regulation and function without appealing to pre-defined sequence complexity or statistical bias thresholds. We find that local compositional enrichment correlates with differences in nearly all core aspects of a protein’s tenure in the cell, including translation efficiency, abundance, half-life, protein-protein interaction promiscuity, subcellular localization, and function. However, enrichment for different amino acids is associated with different effects, even for residues often grouped based on physicochemical similarities, highlighting the importance of distinguishing LCDs of different types. These relationships are discernible at compositions below those required for classification as low-complexity or statistically-biased, suggesting that the thresholds in traditional methods may not be biologically optimized. Finally, analysis of experimentally-defined protein components of stress granules and P-bodies reveals both shared and distinct compositional features associated with these organelles.
Fundamentally, both sequence complexity and statistical amino acid bias are indirect measures of local amino acid composition. Since composition is a more direct indication of overall protein domain properties, we sought to examine whether composition alone could be used to infer residue-specific relationships between local amino acid composition and protein regulation and function. We first developed an algorithm to partition the yeast proteome on the basis of maximum local composition for each amino acid using a series of scanning window sizes (Fig 1; see Methods). For all amino acids, the majority of proteins are partitioned into composition bins of ≤ 25% (Fig 2 and S1 Table). However, the number of proteins achieving higher local compositions, indicated by a right-hand shoulder or tail in the distribution, were strongly residue-dependent. For example, proteins containing local enrichment of highly hydrophobic residues (I, L, M, and V), aromatic residues (F, W, and Y), or cysteine are almost exclusively limited to composition bins of ≤ 45% for the smallest window size, whereas alanine and proline distributions extend to slightly higher composition ranges (up to 60–65%). Proteins containing local enrichment of polar (G, N, Q, S, and T) or charged (D, E, and K) residues in composition bins of ≥ 40% are relatively common even among larger window sizes (albeit to differing degrees), whereas histidine and arginine rich regions are relatively rare. These data indicate that relatively high local enrichment is tolerated for some amino acids, while compositional enrichment for other amino acids appears to be restricted in yeast.
While the origins and evolution of LCDs have been extensively explored [3,4,14,38,43,44], the regulation and metabolism of LCD-containing proteins remain poorly-understood. Proteins with intrinsically disordered segments, which often qualify as LCDs [45,46], have been associated with lower protein half-lives [47]. However, not all intrinsically disordered regions lead to short protein half-lives, and not all LCDs are intrinsically disordered [15]. Additionally, proteins with homopolymeric repeats, when considered as a single class, are associated with lower translation efficiency, lower protein abundance, and lower protein half-life compared to proteins lacking homopolymeric repeats [37]. However, the regulation and structural properties of proteins with LCDs or homopolymeric repeats is likely strongly dependent on the predominant amino acids within the domain of interest [42].
To explore relationships between local compositional enrichment and protein metabolism, we first examined possible links between local compositional enrichment and protein abundance. Recent advances in proteomic methods have facilitated remarkable proteome coverage for both protein abundance [48] and protein half-life [49] measurements in yeast. At each window size/percent composition bin, the distribution of protein abundance values for all proteins partitioned into that bin was compared to the protein abundance distribution for all other yeast proteins (Mann-Whitney U test). Transitions from significantly lower median abundance to significantly higher median abundance or vice versa are observed upon enrichment for many amino acids individually (Fig 3). However, the direction of the trends upon progressive compositional enrichment are dependent on amino acid type. For the majority of amino acids (C, D, F, H, I, L, M, N, P, Q, R, S, T, W, or Y) compositional enrichment is associated with lower median protein abundance. However, compositional enrichment of A, G, or V is associated with higher median protein abundance. Two very similar transitions are observed for both E-rich and K-rich sequences: as compositional enrichment increases, the relative median protein abundance transitions from high to low, then back to high. Collectively, these trends are consistent with, yet much stronger than, previously observed correlations between protein abundance and whole-protein composition [50,51]. This suggests that the trends observed previously may actually reflect the effects of local compositional enrichment, which would increase apparent whole-protein composition for the enriched amino acid yet be dampened by confounding effects from the remainder of the protein sequence.
Similar trends are observed when compositional enrichment is compared to protein half-lives (Fig 4). Compositional enrichment for the majority of amino acids (C, H, K, M, N, P, S, or T) is associated with lower protein half-life, whereas enrichment for A, G, I, or V is associated with higher protein half-life. Enrichment for F leads to an initial transition from lower to higher half-lives, while further enrichment leads to a transition back to lower half-lives. It is worth noting that similar trends were observed in an independent protein half-life dataset when the proteins were analyzed based on whole-protein amino acid composition [52], suggesting that maximum local composition is sufficient to detect associations between amino acid composition and half-life. Although for many amino acids the trends are readily apparent, the strength of the association between compositional enrichment and protein half-life appears to be slightly weaker than the association between compositional enrichment and protein abundance. This is likely due, at least in part, to limited proteome coverage (relative to the protein abundance dataset). However, a recent study also suggested that protein half-life is strongly affected by factors other than sequence characteristics [53], which would likely further dampen relationships between compositional enrichment and protein half-life. Finally, protein half-life is generally less-conserved than protein abundance [54], perhaps suggesting that specific relationships between conserved sequence features and protein half-life may not be particularly strong. Therefore, it is rather surprising that we observe the indicated trends in spite of these limitations, and could suggest that half-life is more strongly influenced by local composition than particular primary sequence motifs.
Direct measurement of protein synthesis rates is more experimentally challenging. Consequently, proteome-wide coverage for experimentally-derived translation efficiency remains substantially lower than coverage for protein abundance and half-life. The normalized translation efficiency (nTE), a reported metric of translation elongation efficiency [55], is based on codon usage frequencies and tRNA gene copy numbers, allowing for calculation of translation efficiency for the entire proteome. Therefore, we first examined relationships between local compositional enrichment and calculated translation elongation efficiency. nTEs were calculated for whole-protein sequences using the corresponding coding region on mRNA transcripts (see Methods). Translation efficiency is strongly dependent on the locally-enriched amino acid (Fig 5). For the majority of amino acids (C, D, E, F, H, I, K, L, M, N, P, Q, R, or Y), local enrichment is associated with significantly lower median nTEs suggesting that, as a single class, proteins with local compositional enrichment tend to be translated relatively inefficiently. Proteins with domains enriched in S, T, or W are generally associated with significantly lower median nTEs, although proteins with very high S, T, or W enrichment are associated with significantly higher median nTEs. However, proteins with domains enriched in A, G, or V residues are consistently associated with significantly higher median nTEs, suggesting that these proteins may be translated relatively efficiently. Remarkably, nearly identical trends are observed between local compositional enrichment and the experimentally-derived protein synthesis rates reported for a limited proteome (S1 Fig) despite a substantial reduction in sample size (n = 1115; [56]), suggesting that nTE can serve as a good surrogate for overall protein synthesis efficiency. Collectively, these results indicate that local amino acid enrichment is associated with differences in protein production rates in a composition-dependent manner.
For most amino acids, we noticed a remarkable correspondence in the trends for translation efficiency, protein abundance, and protein half-life, despite the fact that these values are derived from entirely different methods and experiments. For example, local enrichment for many amino acid types is associated with low nTE values, low protein abundance, and low protein half-life (Table 1). While translation efficiency and protein degradation rate are largely functionally independent in cells, protein abundance depends, at least in part, on both translation efficiency and protein half-life [49]. This may suggest that protein abundance for these proteins is limited in cells by a combination of poor translation efficiency and rapid degradation rate. In contrast, local enrichment for some amino acids is associated with high protein abundance also tended to have higher nTE values and higher half-lives, perhaps suggesting that high protein abundance for these proteins is achieved by a combination of efficient translation and poor degradation.
As a model eukaryotic organism, S. cerevisiae provides a number of important advantages in proteome-scale studies relating protein sequence to protein metabolism and function. In addition to the unmatched proteome coverage in protein abundance and protein half-life datasets, and the availability of yeast-specific tools such as nTE, sequence-function analyses in yeast are further simplified by the absence of tissue-specific effects and limited alternative splicing (only ~4% of yeast genes contain introns and, of those genes, only a small fraction is capable of producing alternative protein products [57,58]).
With these caveats in mind, we sought to examine whether similar relationships between local amino acid composition and protein abundance could be detected in a model multicellular eukaryotic organism. We decided to focus on whole-organism protein abundance measurements in C. elegans [59] for four main reasons: 1) due to technical experimental challenges, protein abundance measurements in C. elegans are substantially more robust than protein half-life measurements; 2) on a proteome-wide scale, protein abundance is more strongly conserved across yeast species than protein half-life [49], suggesting that final protein levels tend to be constrained across organisms, while regulation of the metabolic pathways that contribute to protein abundance may vary; 3) protein abundance is, at least partially, a function of translation efficiency and protein half-life; and 4) the parameters underlying the translation efficiency method (namely the “s-vector”, or the efficiency of wobble base pairing between tRNA isoacceptors) were optimized for yeast [60]. Therefore, the nTE method may not be amenable to application in other organisms.
In order to examine relationships between maximum local composition and protein abundance, we first determined the proteome distribution of C. elegans proteins as a function of maximum local composition for each amino acid. The C. elegans-specific proteome distributions (S2 Fig and S2 Table) were overall quite similar to the yeast proteome distributions (Fig 2). However, the maximum local composition for S and N appear to be slightly more constrained in C. elegans (indicated by contraction of the shoulder to lower maximum compositions), while G, P, and T achieve slightly higher maximum local compositions, indicating relaxed constraints on local enrichment of these residues. These results are consistent with previous observations noting both shared and organism-specific homopolymeric repeat signatures or bulk proteome compositions across proteomes from different organisms [4,38,40,41,44–46,61].
As observed in yeast, progressive compositional enrichment results in a transition from higher to lower median abundance for the majority of amino acids with a clear trend (C, F, I, M, N, P, S, W, and Y; Fig 6). Furthermore, all three amino acids (A, G, and V) that exhibit a transition from lower to higher median abundance upon progressive enrichment in yeast exhibit the same trend in C. elegans as well. Indeed only one amino acid with a clear transition in protein abundance upon local enrichment differs between C. elegans and S. cerevisiae: in yeast, local K enrichment is associated with mixed effects on protein abundance (depending on the degree of K enrichment), whereas in C. elegans local K enrichment is weakly (yet consistently) associated with higher protein abundance. Therefore, nearly identical residue-specific relationships are observed between local amino acid enrichment and protein abundance in a more complex eukaryote.
An important advantage of approaching LCDs from a composition-centric perspective is the ability to examine relationships between amino acid composition and protein outcomes without appealing to pre-defined thresholds of statistical amino acid bias [11] or sequence complexity [1,10], which may not reflect biologically-relevant thresholds. Indeed, the transitions observed in the median translation efficiencies, protein abundances, and protein half-lives often occur at surprisingly mild levels of compositional enrichment, suggesting that these trends may be observed even in the absence of classically-defined statistically-biased or low-complexity domains.
Statistical amino acid bias conceptually parallels our investigation of compositional enrichment, and has been used to investigate the functions of proteins with statistically-biased domains [11,12]. To examine whether compositional enrichment may be linked to biologically-relevant effects on protein metabolism independently of statistically-biased domains, a conservative bias threshold was employed to define statistically-biased domains using previously developed methodology [12] (also, see Methods). Proteins with statistically-biased domains were then filtered from the yeast proteome (n = 866 statistically-biased proteins for the yeast translated proteome of sequences ≥ 30 residues in length). However, even in the absence of statistically-biased domains, compositional enrichment resulted in robust trends in translational efficiency, protein abundance, and protein half-life that re-capitulated those originally observed (S3–S5 Figs). This suggests that compositional enrichment affects protein metabolism at thresholds preceding those required for classification as statistically-biased by alternative methods.
The SEG algorithm, by default, employs substantially more relaxed criteria when classifying protein domains as low-complexity [1]. Indeed, of the 5,901 proteins of length ≥30 amino acids in the translated ORF proteome, 4,147 proteins contain at least one LCD, which is consistent with previous estimates [3]. Nevertheless, despite a large reduction in proteome size, many of the trends in protein metabolism are discernible even when all proteins with a SEG-positive sequence are filtered from the proteome (S6–S8 Figs). This suggests that compositional enrichment exerts biologically relevant effects even among non-LCD-containing proteins.
Proteins containing homopolymeric amino acid repeats (often defined as five or more identical amino acids in succession), were recently reported to have lower translation efficiency, lower protein abundance, and lower protein half-life when compared to proteins without homopolymeric repeats [37]. Homopolymeric repeats are effectively short sequences of maximum possible single-amino acid density. Therefore, proteins with homopolymeric repeats are expected to be disproportionately common among compositionally enriched domains, raising the possibility that the trends observed in the present study have been mis-attributed to compositional enrichment alone. To examine this possibility directly, the relationship between compositional enrichment and nTE, abundance, and half-life was re-evaluated for a filtered proteome that excludes all proteins containing at least one homopolymeric repeat (n = 755 proteins excluded). While exclusion of these proteins preferentially reduces the sample sizes at higher compositional enrichment percentages, the absence of homopolymeric repeat proteins has little effect on the trends in nTE, abundance, and half-life as a function of compositional enrichment (S9–S11 Figs). This does not definitively rule out the possibility that homopolymeric repeats may, in some way, specifically affect translation efficiency, abundance, and half-life. However, since homopolymeric repeats per se are not absolutely required, the effects of homopolymeric repeats may instead be explained simply by local compositional enrichment.
Collectively, these results suggest that compositional enrichment affects translation efficiency, protein abundance, and protein half-life at thresholds preceding those required for classification as low-complexity or statistically-biased by traditional methods. It is worth noting that in the course of eliminating proteins with classically-defined low-complexity, statistically-biased, or homopolymeric domains, proteins with multiple distinct domains strongly enriched in different amino acid types, or with single domains strongly enriched in more than one amino acid, are eliminated from the proteome before re-evaluation. Therefore, the trends in protein metabolism observed upon enrichment of a given amino acid are not due to confounding effects of domains strongly enriched in other amino acids occurring within the same protein sequences.
Local enrichment of a single amino acid can dramatically influence the physicochemical properties of a given protein domain [24]. In a cellular context, these physicochemical properties likely influence interactions between proteins and surrounding molecules, including other proteins.
To examine whether local compositional enrichment affects protein-protein interactions, we explored relationships between enrichment for each of the amino acids and protein-protein interaction promiscuity (defined as the number of unique interacting partners per protein). Proteins found in a range of high-percent composition-bins for most amino acids (A, D, E, G, K, N, P, Q, R, and V) are associated with significantly more interacting partners relative to all other proteins (Fig 7), suggesting that these domains are relatively promiscuous. Additionally, proteins with mild enrichment for select hydrophobic residues (I, L, and M) are generally associated with more interacting partners, although fewer comparisons reach statistical significance (blue or red dots). These results are consistent with previous reports that, as a single class, proteins with LCDs or homopolymeric repeats tend to have more protein-protein interaction partners [16,37]. However, proteins in a range of high-percent composition-bins for each of the aromatic residues (F, W, and Y) are associated with significantly fewer interacting partners relative to other proteins, suggesting that aromatic residues tend to lack the interaction promiscuity observed at higher percent compositions for other amino acids. Furthermore, proteins with moderate to high local C content and proteins with extremely high maximum local S or T content are also associated with significantly fewer interacting partners relative to other proteins, suggesting that these domains are relatively non-promiscuous as well. This is particularly interesting, given that these trends were not observed upon enrichment for other polar residues. Again, this highlights the potential pitfall of grouping amino acids with related physicochemical properties into a single category. Collectively, these results indicate that protein-protein interaction promiscuity varies for proteins with high compositional enrichment in a residue-specific manner.
Previous studies have attempted to associate proteins containing LCDs, statistically-biased domains, and homopolymeric repeats with particular cellular functions [12,16–18,34,37]. However, one important consideration when inferring relationships between proteins with LCDs and cellular functions, for example, is the prevalence of proteins with multiple LCDs [3], and of LCDs strongly enriched in more than one amino acid type [11,14,18,36]. Therefore, attempts to associate cellular functions to specific LCD types, without controlling for other LCDs within the same protein sequences, risk mis-attributing functions to unrelated protein features [12,14,34,36]. While multiple LCDs within the same protein (or multiple amino acid types enriched within the same LCD) may cooperate to generate novel structures or functions, this complicates interpretation of the role of each individual amino acid type within LCDs. Furthermore, because some types of LCDs are more common than others, general attempts to associate cellular functions with LCDs, statistically-biased domains, or homopolymeric repeats likely reflect the functions associated with only the most common types when considered as a single, unified class [16,37]. Therefore, definitive assignment of cellular functions to each individual class of LCD necessitates exclusion of proteins with other types of LCDs.
In order to minimize possible confounding effects introduced by proteins with multiple regions enriched in different amino acid types, a modified version of the initial calculation performed by the SEG algorithm (namely, the Shannon entropy; see Methods) was employed to define proteins with only a single type of compositionally-enriched domain (CED). In an effort to incorporate our results (which indicate that compositional enrichment may exert biologically-relevant effects at compositions preceding the SEG algorithm threshold) into our definition of single-CED proteins, percent composition bins for which at least 75% of the residing proteins contained a SEG-positive sequence (as defined above) were pooled to generate a single list of CED-containing proteins for each amino acid. Proteins that contain multiple types of CEDs were then removed from the dataset, resulting in a non-redundant set of proteins with only one type of CED. Importantly, this method captures the exclusion of proteins containing more than one type of CED, as well as proteins with CEDs strongly enriched in more than one amino acid type.
Gene Ontology (GO) term analysis was performed separately for each window size within each single-CED category. For each type of CED, there is strong overlap in the enriched GO terms across the range of window sizes, suggesting that the associations between functions and residue-specific CEDs are not strongly length-dependent at this scale. Therefore, for simplicity of interpretation, significantly enriched GO terms for each window size were pooled to generate a single non-redundant list of enriched GO terms for each CED type.
Removal of proteins with multiple types of CEDs reveals a remarkable degree of specialization for CEDs of different types (Fig 8, and S3 Table), which is often not observed for CEDs when considered as a single category or when multi-CED proteins are not excluded. For example, L-rich proteins are predominantly associated with functions at the ER and vacuole membranes, whereas I-rich proteins are more strongly associated with carbohydrate transport at the plasma membrane. A-rich proteins are associated with a variety of processes or cellular components, including translation, protein kinase activity, the cell wall, and carbohydrate/alcohol catabolism. N-rich proteins are strongly (and perhaps exclusively) associated with functions related to transcription, whereas Q-rich proteins appear to be more weakly associated with transcription and, instead, are associated with a larger variety of functions including endocytosis, mating projection of the membrane, and response to glucose. Finally, although yeast cell wall proteins are often radically S/T-rich, after controlling for co-enrichment of S and T in the same proteins, S-rich proteins are more strongly associated with membrane-related processes (cell wall, cellular bud tip, cellular bud neck, mating tip projection, etc.), protein kinase activity, and transcription, whereas T-rich proteins tend to be associated with nucleic acid binding and helicase activity, with fewer associations with membrane-related processes. Therefore, after controlling for the presence of multiple CEDs within the same proteins, specialized functions emerge even among commonly grouped amino acids.
Furthermore, CEDs enriched in some amino acids share functions despite the removal of multi-CED proteins, suggesting some degree of co-specialization. For example, D-, E-, and K-rich CEDs were each associated with functions in the nucleus/nucleolus, including ribosomal RNA processing, nucleic acid binding, transcription, and histone/chromatin binding. Intriguingly, intrinsically disordered domains with opposite net charges (along with other charged macromolecules such as nucleic acids and polyADP-ribose) can drive phase separation or complex coacervation in the nucleus [62–64]. It is possible that these domains, along with nucleic acids and other polyionic molecules, may participate in nuclear processes via dynamic electrostatic association with these or other membraneless assemblies. By contrast, H-rich CEDs are associated with processes related to zinc ion transport and regulation. There were no GO terms significantly associated with R-rich CEDs. However, compositional enrichment for R appears to be constrained, as evidenced by the sharp decline in the number of proteins with R-rich domains toward higher maximum local percent compositions (see Fig 2), which may be further impacted by the removal of proteins with other types of CEDs.
In summary, when examined as separate classes, different types of CEDs can have overlapping or specialized roles in the cell.
The molecular specialization observed for CEDs indicates that proteins with enrichment of particular residues may localize to particular subcellular compartments in order to execute their specialized functions. Furthermore, protein quality control factors can differ between subcellular compartments (for review, see [65]), which may contribute to composition-dependent differences in protein metabolism. Therefore, we applied a bottom-up approach to infer the composition profiles associated with the major subcellular compartments (see Methods).
Largely aqueous subcellular compartments are almost exclusively associated with proteins containing domains enriched in charged residues, polar residues, and proline (Fig 9; see also S12 Fig). However, differences in compositional enrichment profiles are apparent even among related aqueous compartments. For example, significant associations with charged, Q, or N residues reach more extreme percent compositions in the nucleus, whereas as significant associations with P enrichment reach higher percent compositions in the cytoplasm. By contrast, the highly membraneous internal organelles (e.g. the endoplasmic reticulum and Golgi apparatus) are predominantly associated with enrichment of hydrophobic or aromatic residues (Figs 9 and S13). The yeast vacuole is also associated with composition profiles resembling those of membraneous compartments, with additional weaker associations with S and C enrichment. Few weak associations are observed for mitochondria. The yeast cell wall is strongly associated with S enrichment (likely related to its ability to be glycosylated), with additional moderate associations with T and A enrichment, and a weak association with mild V enrichment (Figs 9 and S14). As expected, the plasma membrane is associated with enrichment for a variety of hydrophobic and aromatic residues. However, the plasma membrane is also significantly associated with enrichment of a select subset of polar residues (namely C, G, S, and T), further corroborating the specialized roles observed for these CEDs at the outer membrane. Indeed, G-rich CEDs are significantly associated with amino acid transport (see S3 Table), and S- or T-rich CEDs of the plasma membrane could have overlapping functions or interactions with S- and T-rich CEDs of the cell wall. Together, these observations indicate that subcellular compartments may tolerate or prefer proteins with specific types of CEDs.
Recent observations indicate that a variety of Q/N-rich and G-rich domains can form highly dynamic protein-rich droplets in aqueous environments [25–30], a process referred to as liquid-liquid phase separation. These types of LCDs are prevalent among components of membraneless organelles such as stress granules and P-bodies [23]. Furthermore, stress granules and P-bodies share many properties with protein-rich liquid droplets formed in vitro, suggesting that the fundamental biophysical properties of these domains are related to the formation of membraneless organelles in vivo. However, while amino acid composition is acknowledged as a critical determinant of this behavior, the precise compositional requirements associated with membraneless organelles remain largely undefined.
Therefore, we also applied our bottom-up approach to infer the compositional enrichment profiles associated with protein components of stress granules and P-bodies (as defined in [66]). Stress granules and P-bodies have overlapping protein constituents and can exchange protein components [67,68], suggesting that they are closely related yet distinct organelles. Accordingly, we observe both shared and unique features in the composition profiles associated with stress granule and P-body proteins (Fig 10). As expected, both stress granules and P-bodies are strongly associated with proteins containing Q-rich or N-rich domains. For example, minimum Q or N compositions significantly associated with stress granules range from ~15–100% at small window sizes (≤30 amino acids) and ~10–30% at large window sizes (≥80 amino acids), although these values vary slightly depending on window size and residue. Similarly, minimum Q or N compositions significantly associated with P-bodies range from ~15–100% at small window sizes and ~10–40% at larger window sizes.
In addition to the commonly appreciated link between stress granule/P-body components and Q/N-rich domains, we identify and define a variety of currently underappreciated compositional features common to stress granule and P-body components. Components of both stress granules and P-bodies are strongly associated with P-rich domains, weakly associated with K-rich domains, and very weakly (yet significantly) associated with Y-rich domains. Furthermore, while both stress granules and P-bodies are associated with proteins containing G-rich domains, stress granule components are associated with a much broader range of G enrichment, suggesting that G enrichment may be a more characteristic feature of stress granules than P-bodies. This is particularly striking in light of recent observations indicating that high glycine content helps maintain the liquid-like characteristics of phase-separated droplets and prevents droplet hardening in vitro [69].
Additionally, some compositional features are unique to either stress granules or P-bodies. For example, stress granule constituents are significantly associated with A-rich, M-rich, E-rich, and R-rich domains, whereas P-body constituents exhibit little or no preference for these compositional features (a key role for arginine in the phase separation of stress granule-associated proteins was also recently reported [69]). By contrast, P-body components are weakly associated with H-rich domains, whereas stress granule components are not enriched among proteins containing H-rich domains.
To our knowledge, this represents the first attempt to systematically define the range of amino acid compositions associated with membraneless organelles such as stress granules and P-bodies. These observations suggest that components of related, membraneless organelles have overlapping yet distinct compositional preferences. It is possible that shared compositional features facilitate the physical interactions between stress granules and P-bodies and allow for the exchange of components, while differences in compositional features facilitate their ability to function as independent organelles.
Protein domains categorized as low-complexity, statistically-biased, or homopolymeric encompass broad, heterogeneous classes of sequences with diverse physical properties and cellular functions. These domains can play important roles in normal and pathological processes. However, challenges in categorizing proteins on the basis of sequence complexity or statistical bias have thus far precluded a complete, proteome-wide view of the effects of these domains on protein regulation and function. Here, we adopt an alternative, unbiased approach to examine proteome-wide relationships between local amino acid enrichment and the birth, abundance, functions, subcellular localization, and death of proteins. For nearly all amino acids, progressive local enrichment corresponds to clear transition thresholds with regard to translation efficiency, protein abundance, and protein half-life. Transition thresholds ubiquitously occurred at compositions preceding those required for classification as low-complexity or statistically-biased by traditional methods, indicating that our observed transition thresholds more closely reflect biologically-relevant composition criteria.
Protein sequences can range from perfectly diverse (i.e. a completely homogeneous mixture of amino acids with maximal spacing between identical amino acids) to lacking any diversity (i.e. homopolymeric sequences). While homopolymeric regions represent an extreme on this spectrum and can influence protein metabolism [37], classically defined homopolymeric regions are not absolutely required for these effects (see S9–S11 Figs). This suggests that compositional enrichment may affect protein metabolism even upon some degree of primary sequence dispersion (i.e. greater linear spacing between identical amino acids). Defining the limits of this dispersion may shed additional light on the relationship between amino acid composition and protein metabolism.
An advantage of assessing compositional enrichment (as opposed to sequence complexity) is the ability to distinguish the effects of compositional enrichment for each amino acid type. The nature of the trends in translation efficiency, protein abundance, and protein half-life depend on the amino acid enriched in the protein sequences, indicating that local enrichment of different amino acids can have opposite effects. This highlights a key limitation when considering low-complexity, statistically-biased, or homopolymeric domains as a single class–grouping domains composed of radically different amino acids effectively skews any trends observed toward those of the most common type and, in some cases, can completely mask the effects of less common low-complexity, statistically-biased, or homopolymeric domains. Furthermore, even grouping these domains on the basis of common physicochemical properties can introduce the same complication. This is exemplified by the non-aromatic hydrophobic amino acids; while I-rich, L-rich, and M-rich domains are associated with poor translation efficiency, low abundance, and rapid degradation rate, A-rich and V-rich domains are associated with high translation efficiency, high abundance, and slow degradation rate. Additionally, the cellular functions associated with domains enriched in hydrophobic residues tend to differ; L-rich domains are predominantly associated with the ER or vacuole membrane, whereas I-rich domains are predominantly associated with carbohydrate transport at the plasma membrane. Similarly, N-rich domains are strongly associated with transcription-related processes, whereas Q-rich domains are more strongly associated with endocytosis and other processes in the cytoplasm. While there is some overlap between these two groups, this suggests that domains enriched in remarkably similar amino acids may yet be favored for specialized roles in the cell.
Finally, a bottom-up application of our composition-centric algorithm to membraneless organelles provides the first step in defining the distinct compositional profiles associated with each type of organelle. We find that even closely related and physically interacting organelles are associated with discernible differences in compositional enrichment, which may relate to differences in their properties, regulation, and function in vivo. It is important to note that, while the observed trends in compositional enrichment are significantly associated with stress granule proteins or P-body proteins as respective groups, these features may not be absolutely required for individual proteins to be incorporated into stress granules and/or P-bodies. It is possible, for example, that two proteins possessing non-overlapping subsets of the associated compositional features may still be recruited to stress granules, and that some stress granule proteins may be recruited for reasons entirely distinct from compositional enrichment (e.g. via RNA-binding domains). One might even imagine that differences in compositional features, while still allowing recruitment to stress granules and/or P-bodies, could favor differences in the dynamics of individual protein components (e.g. the kinetics of entry/exit, dwell time, the strength of the interactions, or the depth of penetration within the stress granule/P-body). Therefore, while the associated composition ranges observed here are collectively enriched among proteins associated with these membraneless organelles, each individual protein need not possess all of the compositional features simultaneously in order to function as a stress granule or P-body protein.
While a great deal of attention is rightfully devoted to understanding the effects of primary amino acid sequence on protein fates (including folding, regulation, and functions), amino acid composition is increasingly believed to drive a variety of cellular and molecular processes. Here, we have developed an approach to examine relationships between local compositional enrichment and protein fates for each of the canonical amino acids, in the absence of a priori assumptions or pre-defined thresholds. Our results provide a coherent, proteome-wide view of the relationships between compositional enrichment and the fundamental aspects of protein life cycle, subcellular localization, and function in model eukaryotic organisms.
Protein sequences were parsed using FASTA sequence parsing module from the Biopython package [70]. For each amino acid in the set of 20 canonical amino acids, each protein in the translated ORF proteome (latest release from the Saccharomyces Genome Database website, last modified 13-Jan-2015) or the ORF coding sequences (organismID:UP000001940_6239, release date 23-May-2018 downloaded from the UniProt website) was scanned using a sliding window of defined size (ranging from 10 to 100 amino acids, in increments of 10). The percent composition of the amino acid of interest (AAoI) is calculated for each window, and the protein is sorted into bins based on the maximum percent composition achieved for the AAoI (ranging from 0 to 100 percent composition in 5 percent increments). Analyses were performed for all possible AAoI, window size, and percent composition combinations.
Translation efficiency for each gene was estimated using the normalized translation efficiency (nTE) scale [55], which is based on tRNA gene copy number, codon-anticodon wobble base-pairing efficiency, and transcriptome-wide codon usage. However, the original nTE algorithm plots all nTE values for each codon to generate a separate translation efficiency profile for each gene. In order to condense translation efficiency information to a single value for each gene (in a manner analogous to the tRNA adaptation index; [60]), the geometric mean of nTE values across the transcript was calculated as
nTEgene=(∏k=1lsnTEiks)1ls
(1)
where nTEiks represents the translation efficiency value of the ith codon defined by the kth triplet in nucleotide sequence s, and ls represents the length of the nucleotide sequence excluding stop codons. Therefore, nTE values reported in the current study represent whole-gene nTE values. nTE analyses were performed using an in-house Python script.
The Shannon entropy of each sequence was calculated as
SE=−∑i=1N=20niL(log2niL)
(2)
where N represents the size of the residue alphabet (N = 20, for the canonical amino acids), ni represents the number of occurrences of the ith residue within the given sequence window of length L. For comparison with established measures of sequence complexity, we defined low-complexity domains by using the default window size (12 amino acids) and Shannon entropy threshold (SE ≤ 2.2bits) used in the first pass of the SEG algorithm to initially identify LCDs [1,10].
In the SEG algorithm, the complexity state vector used to calculate the Shannon entropy is blind to the amino acid composition (i.e. the ni values in Eq 2 are not attributed their respective amino acids). Therefore, when indicated, in order to distinguish LCDs on the basis of the predominant amino acid, sequences for which the SE ≤ 2.2bits and nAAoI ≥ nmax within the complexity state (indicating that the AAoI is a major contributor to the sequence’s classification as an LCD) were assigned to the corresponding amino acid category (e.g. A-rich LCDs, C-rich LCDs, etc.). Single-LCD/CED proteins are proteins classified as LCDs or CEDs that do not appear on multiple amino acid-specific LCD/CED lists.
Statistical amino acid bias was calculated as described in [12]. Briefly, the lowest probability subsequence for each protein was determined by exhaustively scanning proteins with window sizes ranging from 25 to 2500 amino acids. For each window, the subsequence bias probability (Pbias) was defined as
Pbias=[w!n!(w−n)!]×(fx)n×(1−fx)w−n
(3)
where w denotes the window size, n denotes the number of occurrences of the amino acid of interest within the subsequence, and fx denotes the fraction of the amino acid of interest in the yeast translated proteome. The lowest probability subsequence for each protein is the subsequence with the lowest Pbias.
A suitable threshold to define statistically-biased proteins within the yeast protein was determined as previously described [12], except that more relaxed criteria were used in order to include additional proteins with less extreme biases. Briefly, the Pbias corresponding to the lowest probability subsequence (Pmin) for each protein was plotted on a log-log plot against whole-protein sequence length. A line was fitted, then the y-intercept was decreased until only 15% of the proteome had Pmin values below the line (previous analyses used a more stringent cutoff of 10% to define statistically-biased proteins [12]). Additionally, a length-independent threshold was designated as the Pmin value at which 15% of the proteome had smaller absolute Pmin values. This threshold was used when it was less than the Pmin threshold given by the length-dependent method to avoid unreasonably relaxed bias criteria for small protein sequences. Amino acid bias was calculated using values from the translated orf proteome only, and implemented via an in-house Python script with pre-computed look-up tables for computational efficiency.
Proteins containing homopolymeric sequences were defined simply as any protein with a subsequence of five or more contiguous residues of the same amino acid, as previously described [37].
Yeast protein abundance values (in average number of molecules per cell per protein) were obtained from [48] (n = 5,391). Protein abundance values for C. elegans were obtained from [59]. Yeast protein half-life data were obtained from [49]. For simplicity of interpretation, only proteins with unambiguous, non-zero half-life or abundance values were included in the datasets. Proteins listed on separate lines with identical half-life or abundance values were retained, whereas protein half-life or abundance measurements assigned to more than one protein on the same line were excluded (these were often highly homologous genes, suggesting that the measurement could not be unambiguously assigned to one of the proteins). Furthermore, all proteins corresponding to “low-confidence” measurements in the half-life dataset were excluded (see [49] for criteria). n = 3,525 for the filtered yeast half-life dataset, and n = 5,952 for the filtered C. elegans protein abundance dataset.
For all AAoI/window size/percent composition bins, the distribution of nTE, abundance, or half-life values for proteins included in the given bin was compared to the distribution of the respective values of all proteins excluded from the given bin. Statistical significance was estimated using a two-sided Mann-Whitney U test (also referred to as the Wilcoxon rank-sum test; refer to Supplemental Experimental Procedures from [47] for a detailed description and rationale). Where indicated, p-values were adjusted within each window using the Bonferroni correction method for multiple hypothesis testing. All statistical tests were performed using modules available in the SciPy package with default settings, unless otherwise specified. All plots were generated using Matplotlib or Seaborn modules.
GO term enrichment tests were performed using the GOATOOLS package (version 0.7.9) [71] for each set of proteins contained in a given amino acid/window size/percent composition bin. For each test, the set of background proteins was defined as all proteins from the translated ORF proteome of sequence length greater than or equal to the given window size. All reported p-values were adjusted using the Bonferroni correction during GO term association. To evaluate the compositional enrichment profiles associated with GO terms related to subcellular compartments, we applied a minimum-threshold-scanning approach to all partitioned proteomes. For each AAoI, window size, and percent composition bin, all proteins with maximum local compositions greater than or equal to the current percent composition under consideration are pooled and evaluated for possible enriched GO terms. This effectively evaluates possible GO term enrichment iteratively with increasing maximum local composition criteria. GO term results were subsequently evaluated for significant enrichment of a single GO term describing each subcellular compartment (or two related GO terms, “outer membrane” and “plasma membrane”, in the case of the plasma membrane). p-values were further adjusted within each window size using the Bonferroni correction method.
Similar analyses were performed for the sets of experimentally-defined stress granule (n = 83) and P-body (n = 52) proteins [66]. Specifically, a minimum-threshold-scanning approach was applied to all partitioned proteomes. For each AAoI, window size, and percent composition bin, all proteins with maximum local compositions greater than or equal to the current percent composition under consideration are pooled. Significant enrichment of experimentally-defined stress granule or P-body proteins within each pool of proteins was evaluated using Fisher’s exact test (p < 0.05).
|
10.1371/journal.pntd.0006962 | Backpack PCR: A point-of-collection diagnostic platform for the rapid detection of Brugia parasites in mosquitoes | Currently, molecular xenomonitoring efforts for lymphatic filariasis rely on PCR or real-time PCR-based detection of Brugia malayi, Brugia timori and Wuchereria bancrofti in mosquito vectors. Most commonly, extraction of DNA from mosquitoes is performed using silica column-based technologies. However, such extractions are both time consuming and costly, and the diagnostic testing which follows typically requires expensive thermal cyclers or real-time PCR instruments. These expenses present significant challenges for laboratories in many endemic areas. Accordingly, in such locations, there exists a need for inexpensive, equipment-minimizing diagnostic options that can be transported to the field and implemented in minimal resource settings. Here we present a novel diagnostic approach for molecular xenomonitoring of filarial parasites in mosquitoes that uses a rapid, NaOH-based DNA extraction methodology coupled with a portable, battery powered PCR platform and a test strip-based DNA detection assay. While the research reported here serves as a proof-of-concept for the backpack PCR methodology for the detection of filarial parasites in mosquitoes, the platform should be easily adaptable to the detection of W. bancrofti and other mosquito-transmitted pathogens.
Through comparisons with standard silica column-based DNA extraction techniques, we evaluated the performance of a rapid, NaOH-based methodology for the extraction of total DNA from pools of parasite-spiked vector mosquitoes. We also compared our novel test strip-based detection assay to real-time PCR and conventional PCR coupled with gel electrophoresis, and demonstrated that this method provides sensitive and genus-specific detection of parasite DNA from extracted mosquito pools. Finally, by comparing laboratory-based thermal cycling with a field-friendly miniaturized PCR approach, we have demonstrated the potential for the point-of-collection-based use of this entire diagnostic platform that is compact enough to fit into a small backpack.
Because this point-of-collection diagnostic platform eliminates reliance on expensive and bulky instrumentation without compromising sensitivity or specificity of detection, it provides an alternative to cost-prohibitive column-dependent DNA extractions that are typically coupled to detection methodologies requiring advanced laboratory infrastructure. In doing so, this field-ready system should increase the feasibility of molecular xenomonitoring within B. malayi-endemic locations. Of greater importance, this backpack PCR system also provides the proof-of-concept framework for the development of a parallel assay for the detection of W. bancrofti.
| Molecular xenomonitoring has demonstrated significant potential as a non-invasive means of providing reliable surveillance for the presence of lymphatic filariasis (LF)-causing parasites. Given the continuing successes of global mass drug administration efforts, the need for such non-invasive surveillance techniques is expanding. However, considering the significant infrastructural demands which such surveillance requires, the development of simplified surveillance methodologies will be fundamental to future programmatic implementation efforts. Accordingly, we have developed a novel, simplified diagnostic platform for point-of-collection-based detection of the LF-causing parasite, Brugia malayi in pools of mosquitoes. By coupling a rapid and inexpensive DNA extraction methodology with a field-friendly amplification platform and test strip-based detection assay, this backpack PCR system eliminates the need for expensive instrumentation and laboratory-based infrastructure. Furthermore, adaptation of the platform described here will allow for the straightforward and rapid development of a parallel assay for the detection of Wuchereria bancrofti, facilitating the increased use of xenomonitoring and enabling mosquito surveillance efforts in regions lacking sophisticated laboratory infrastructure.
| Lymphatic filariasis (LF) is a disfiguring and disabling tropical disease caused by parasitic filarial nematodes. It is estimated that more than 120 million people currently suffer from this mosquito-borne infection, with approximately 90% of the global LF burden caused by Wuchereria bancrofti and the remaining 10% caused by the parasites Brugia malayi and Brugia timori [1–3]. Despite this significant burden of disease, the World Health Organization has targeted LF for elimination by the year 2020 [4–7], and accordingly, the Global Programme to Eliminate Lymphatic Filariasis (GPELF) has implemented mass drug administration (MDA) programs in most endemic countries in order to interrupt disease transmission and reduce infection rates.
An important component of LF elimination efforts is monitoring changing infection and exposure rates over the course of yearly MDAs in order to evaluate programmatic success and determine when treatment can be stopped [8–11]. Following the cessation of MDA, similar monitoring is required to ensure that infection recrudescence has not occurred and is unlikely to occur going forward [12]. While human blood sampling is currently the standard procedure for monitoring these rates [13–16], PCR-based detection of parasite DNA in insect vectors, termed molecular xenomonitoring (MX), is an effective and non-invasive alternative, capable of indirectly measuring parasite burden within endemic locations [17–27]. Given this potential, the World Health Organization has championed the development of novel methodologies capable of increasing the practicality of this approach to disease surveillance [28–29], since implementation of currently available MX techniques are expensive in terms of both reagents and infrastructure requirements.
Currently, field-adaptable loop-mediated isothermal amplification (LAMP) assays exist for the detection of both W. bancrofti [30] and Brugia parasites [31], and a helicase-dependent isothermal amplification (HDA) assay exists for the detection of B. malayi DNA in human blood samples [32]. However, widespread programmatic implementation of such assays has not occurred. Limited use likely stems primarily from insufficient large-scale comparative evaluation efforts. Concerns regarding detection ambiguities and uncertainties may also contribute. When running LAMP assays detection of positive samples relies on the visualization of either sample turbidity or sample fluorescence [31]. As such, results are not “presence” or “absence”-based, but rather occur on a spectrum. This potentially raises concerns regarding susceptibility to technician bias, fatigue, or interpretation. Equally problematic, while HDA assays typically rely on agarose gel electrophoresis, a proven means of effective target detection, such techniques are prone to sample contamination [33], and are difficult to perform in a field setting. In contrast, the field-friendly, “Backpack PCR” platform we describe here utilizes test strip-based DNA detection methods. Such technology partners the advantages of standardizing assay readouts with the capacity to minimize laboratory equipment needs. Furthermore, given the ongoing work towards the development of automated card reader technologies [34], the consistency of results will likely continue to improve. For these reasons, a number of assays conducted in conjunction with the Global Programme to Eliminate Lymphatic Filariasis and other tropical disease control and elimination programs [35] make use of strip or card-based detection techniques [36–39].
Here we describe the development of a novel diagnostic platform, which we have termed “Backpack PCR”, coupling a rapid mosquito DNA extraction procedure with the test strip-based detection of B. malayi DNA. This assay amplifies a 132 base pair region of the non-coding HhaI repetitive DNA sequence element, selected as the target due to its high copy number and proven reliability in previously described molecular diagnostic assays [32, 40–45]. Following a resource-minimizing NaOH-based DNA extraction procedure, amplification occurs using a field-friendly, miniaturized PCR technology. Parasite DNA-derived amplicons are then detected using a test-strip-based methodology, enabling point-of-collection-based sample processing. While acknowledging the potential utility of backpack PCR as a tool in Brugia-endemic locations, the primary contribution of this work is likely its service as a proof-of-concept study for the parallel development of a similar assay for the detection of W. bancrofti. Such development is currently underway.
While B. malayi-infected mosquitoes can be successfully reared in the laboratory, due to the possible failure of individual mosquitoes to ingest microfilariae when taking a blood meal, the definitive assumption cannot be made that all exposed mosquitoes actually harbor parasites. For this reason, positive mosquito pools were prepared by spiking groups of uninfected, laboratory reared Aedes aegypti mosquitoes (Oxitec Ltd., Abingdon, UK) with a single B. malayi infective-staged larva (L3) (Filarial Research Reagent Resource Center [FR3], Athens, GA). Uninfected A. aegypti mosquito pools were also prepared to serve as experimental controls. Positive pools containing 5, 10, 15, 20, and 25 uninfected mosquitoes were prepared by adding a single L3 larva to each pool. Twenty positive pools of each pool size were prepared, while two negative control pools, containing only uninfected mosquitoes with no added L3 larva, were also prepared for each pool size. Thus, 22 pools total were prepared for each of the 5 pool sizes, resulting in 110 total pools.
Ten positive mosquito pools and 1 negative mosquito pool of each mosquito pool size (5, 10, 15, 20 and 25) were extracted using a modified version of a previously published, cost-effective, NaOH-based DNA extraction procedure [46]. Briefly, in a 1.7 ml microfuge tube, a sterile plastic micro-pestle (Axygen Scientific, Union City, CA) was used to grind each mosquito pool with 180 μl of 0.2 N NaOH for 3 min. The micro-pestle was then rinsed with an additional 180 μl of 0.2 N NaOH into the same 1.7 ml tube containing the ground mosquitoes to be sure all mosquito debris was removed from the pestle. A new, clean, sterile micro-pestle was used for each pool. Each tube was incubated at 75 °C for 10 min. 115.2 μl of 1 M Tris (pH = 8.0) and 364.8 μl of nuclease-free water were added to each sample and the tubes were thoroughly mixed using a vortex mixer for 10 sec. Samples were centrifuged at 10,000 x g for 3 min, and the supernatant, containing extracted DNA, was collected and transferred into a clean 1.7 ml microcentrifuge tube. Utilizing an additional volume of nuclease-free water, ten-fold dilutions of each sample were prepared to minimize the effects of PCR inhibitors on downstream amplification reactions. In parallel, ten positive and 1 negative mosquito pool of each mosquito pool size (5, 10, 15, 20 and 25) were also extracted utilizing the Qiagen DNeasy Blood & Tissue Kit (Qiagen, Hilden, Germany) in accordance with the previously described extraction protocol [43, 47–48].
Real-time PCR was used to determine whether or not each pool of mosquito DNA, extracted by either the NaOH-based or Qiagen-based method, contained B. malayi DNA. For testing purposes, the real-time PCR assay developed by Rao, et al. was considered to be the gold standard for parasite detection [43], and it was assumed that every pool containing B. malayi DNA would result in positive detection using this assay. All real-time PCR reactions were carried out in triplicate using the previously described primers, probe, reaction mixture, volume, and cycling conditions [43]. Thermal cycling was performed using the StepOnePlus Real-Time PCR System (Applied Biosystems, Foster City, CA). Results were considered to be positive when amplification occurred with a mean Ct value of less than 38.
A comparison of amplification efficiency between a standard thermal cycler (Veriti Thermal Cycler, Applied Biosystems, Foster City, CA) and the field-friendly miniPCR platform (Amplyus LLC, Cambridge, MA), was conducted using conventional PCR-based amplification of B. malayi DNA. The reactions were carried out using the Phire Hot Start II enzyme (Thermo Fisher Scientific, Waltham, MA) with the previously described real-time PCR primer set [43], modified to contain a biotin tag on the reverse primer (Fwd: 5’—GCAATATACCGACCAGCAC—3’ / Rev: 5’—Biotin-ACATTAGACAAGGAAATTGGTT—3’). Reactions were performed in 20 μl total volumes containing 100 nM concentrations of each primer, 0.5 μl of enzyme, 5 μl of 5X reaction buffer, 0.5 μl of 10 mM dNTPs, and 1 μl of a 10-fold dilution of the DNA extraction product. Parallel amplification reactions were performed using both a standard thermal cycler, (Veriti Thermal Cycler, Applied Biosystems), and the miniPCR instrument (Amplyus LLC, Cambridge, MA). Cycling conditions consisted of an initial 30 sec hold at 95 °C, followed by 35 cycles of 30 sec at 95 °C, 40 sec at 55 °C, and 1 min at 72 °C. Following cycling, a final 5 min extension step at 72 °C was performed.
Fifty pools of mosquitoes were prepared which contained either zero or one B. malayi-L3-staged larva in pools of 5, 10, 15, 20, or 25 uninfected mosquitoes. Pools were then coded blind to the technician performing the assay. All mosquito pools underwent DNA extraction using the NaOH-based method described above, and extracts underwent amplification using the miniPCR platform, a standard conventional PCR platform, and the StepOnePlus Real-Time PCR System. All conventional PCR products were visualized using both test strip-based detection and agarose gel electrophoresis. Real-time PCR results were used to confirm the presence or absence of parasite DNA within each sample. After recording all results, the samples were un-blinded, and results were assessed.
The efficiency of the rapid NaOH-based DNA extraction methodology was examined through comparative real-time PCR testing of NaOH and Qiagen column-extracted sample panels. Qiagen extractions resulted in positive parasite detection in 50/50 samples, while NaOH-based extractions resulted in positive parasite detection in 49/50 samples. The NaOH-based extraction method was sufficiently efficient to give consistent detection of a single B. malayi L3-staged worm in all pools of up to 20 uninfected mosquitoes and in 9/10 pools containing 25 mosquitoes (Table 1). All mosquito pools not containing an L3-staged parasite were negative by both methods.
Following re-blinding of the panel of NaOH-extracted samples described above, conventional PCR-based amplification of all samples was conducted using both the miniPCR and the standard Veriti PCR instruments. Results were compared with those previously obtained by real-time PCR. For all samples tested, real-time PCR results agreed with the miniPCR results using both test strip-based detection and electrophoresis-based detection (Table 2).
In order to verify the specificity of the “Backpack PCR”-based assay for the detection of B. malayi DNA, gDNA samples from D. immitis, M. perstans, W. bancrofti, B. pahangi, A. viteae, and L. loa were assayed along with B. malayi DNA as a positive control. To verify that these samples contained amplifiable filarial parasite DNA, extracts from all species were subjected to amplification utilizing the previously described pan-filarial primer set (DIDR) [49], and all samples amplified successfully (S1 Fig). When tested using the B. malayi field-friendly PCR platform described above, all pools containing non-Brugian filarial DNA tested negative, regardless of whether results were visualized by gel electrophoresis or test strip-based detection. However, B. pahangi genomic DNA produced positive results, indicating cross-reactivity, when examined by both visualization techniques (S1 Fig). Although the primers and probe selected for use with this assay were designed using the HhaI repeat DNA sequence from B. malayi (GenBank Accession No. AF499129.1), these results indicate that this assay also detects the closely related animal parasite B. pahangi. This is not surprising since the HhaI repeat found in B. malayi has a sequence that is 89% identical to the HhaI repeat found in B. pahangi [40]. Thus, the assay we describe here is Brugia genus-specific, but not B. malayi species-specific.
To further validate the use of our field-friendly platform and to further demonstrate its comparable sensitivity to real-time PCR, a series of 50 mosquito samples were prepared and coded blind to the processing technician. Following NaOH-based DNA extraction, each sample underwent cPCR-based amplification using both the standard and miniPCR instruments, followed by analysis using both gel electrophoresis and test strip-based detection methodologies. Quantitative real-time PCR was also performed for comparison. For 48 out of 50 samples, results for all assays were in agreement (Table 3). Examination of the two discordant samples revealed that in one instance (Sample #13) all assays yielded negative results with the exception of standard amplification coupled with test strip-based detection. In the second instance of discordance (Sample #46), real-time PCR and test strip-based detection assays returned positive results, while the gel electrophoresis results were negative. Thus, real-time PCR and the test strip-based detection agreed on 49 of 50 samples.
As a platform for filarial parasite detection, the coupling of a simple DNA extraction method with a field-friendly amplification platform and test strip-based detection technology has the potential to greatly expand the reach of MX efforts. Due to the reduced need for infrastructure and expensive, highly technical equipment, “Backpack PCR” is ideally suited for use in endemic locations currently lacking the capacity to perform real-time PCR reactions for MX purposes. We have demonstrated that this “Backpack PCR” platform has the capacity to reliably detect a single B. malayi L3 infective larva in pools of up to 25 uninfected A. aegypti mosquitoes. "Backpack PCR” minimizes equipment needs since it requires only a dry bath, a low-speed microcentrifuge, and a portable, battery-powered and smartphone-controlled thermal cycler (miniPCR) which weighs less than 1 lb. All of the equipment and materials needed for the assay can be easily transported in a backpack (Fig 3) and sample analysis can be carried out at, or near the point-of-collection in remote locations with limited resources. Building upon the proof-of-concept work described here, the development of a parallel assay for the detection of W. bancrofti will further empower local scientists by enabling their independent use of MX as a tool for the mapping of filarial parasite prevalence and for post-MDA surveillance.
Despite the many advantages of this point-of-collection platform, the current design does present some challenges. Foremost, at a cost of approximately $7.50 per pool of mosquitoes, extraction and detection using this assay requires a significantly greater reagent investment than does extraction coupled with real-time PCR testing, estimated to cost $5.45 per equally-sized pool. Much of the cost of this system is due to the expensive test-strips, so we are actively researching less expensive alternatives that could reduce the cost to be competitive with real-time PCR. Utilization of this field-friendly platform will eliminate the need to maintain and service the sophisticated thermal cyclers required to perform real-time PCR diagnostics in a laboratory setting. In addition, the large capital commitments required for the initial purchase of such real-time PCR instruments would be eliminated, substantially lowering both initial overhead costs, and the recurring costs that arise from maintenance contracts and service fees. Furthermore, in many settings, increased reagent costs could be offset by eliminating the expenses associated with transporting and/or shipping samples to a reference laboratory for analysis. Eliminating the need for shipment, oftentimes out of country, has added benefits, as government regulations in many endemic countries require samples to be tested in their country of origin.
Although not unique to MX assays [48], the capacity of this assay to amplify DNA from both the human parasite B. malayi and the animal parasite B. pahangi presents an additional limitation. While sufficient data does not exist to reliably estimate the prevalence of B. pahangi within most locations co-endemic for both Brugian parasites, this shortcoming certainly merits additional investigation in such co-endemic areas.
Effectively trapping large numbers of Anopheles and Mansonia mosquitoes, primary vectors of B. malayi [50], presents the most substantial obstacle for all Brugia-based xenomonitoring efforts. Historically, such trapping difficulties have greatly restricted xenomonitoring efforts in Brugia-endemic locations, resulting in very few published examples of implementation [47, 51]. However, novel trap designs and improved trapping techniques continue to emerge [52–56] and it is imperative that the appropriate molecular tools be available to most effectively capitalize upon trap improvements as they occur. More importantly, the successful development of the “Backpack PCR” platform described here provides proof-of-principle for the development of future MX assays for the detection of other mosquito-borne infections. Accordingly, the development of a similar “Backpack PCR”-based assay for the detection of W. bancrofti, the parasite responsible for approximately 90% of the global LF burden, would be of significant use to the research community, and efforts to create such a platform in our laboratory are currently underway. Similarly, this method could also be extended to the detection of pathogens causing other mosquito-borne infections such as malaria, Zika, dengue fever, Chikungunya and others.
|
10.1371/journal.pcbi.1005118 | Error Correction and the Structure of Inter-Trial Fluctuations in a Redundant Movement Task | We study inter-trial movement fluctuations exhibited by human participants during the repeated execution of a virtual shuffleboard task. Focusing on skilled performance, theoretical analysis of a previously-developed general model of inter-trial error correction is used to predict the temporal and geometric structure of variability near a goal equivalent manifold (GEM). The theory also predicts that the goal-level error scales linearly with intrinsic body-level noise via the total body-goal sensitivity, a new derived quantity that illustrates how task performance arises from the interaction of active error correction and passive sensitivity properties along the GEM. Linear models estimated from observed fluctuations, together with a novel application of bootstrapping to the estimation of dynamical and correlation properties of the inter-trial dynamics, are used to experimentally confirm all predictions, thus validating our model. In addition, we show that, unlike “static” variability analyses, our dynamical approach yields results that are independent of the coordinates used to measure task execution and, in so doing, provides a new set of task coordinates that are intrinsic to the error-regulation process itself.
| During the repeated execution of precision movement tasks, humans face two formidable challenges from the motor system itself: dimensionality and noise. Human motor performance involves biomechanical, neuromotor, and perceptual degrees of freedom far in excess of those theoretically needed to prescribe typical goal-directed tasks. At the same time, noise is present in the human body across multiple scales of observation. This high-dimensional and stochastic character of biological movement is the fundamental source of variability ubiquitously observed during task execution. However, it is becoming clear that these two challenges are not merely impediments to be overcome, but rather hold a key to understanding how humans maintain motor performance under changing circumstances, such as those caused by fatigue, injury, or aging. In this work, by studying skilled human participants as they play a virtual shuffleboard game, we demonstrate the fundamental importance of adopting a dynamical perspective when analyzing the motor variability observed over many trials. Using this dynamical approach, we can not only study the geometry of observed inter-trial variability, but can also theoretically describe and experimentally characterize how it is temporally generated and regulated. Furthermore, our theoretical framework and model-based data analysis approach helps to unify previous variability analysis approaches based on stability, correlation, control theory, or task manifolds alone. This conceptual unification supports the idea that such seemingly disparate features of motor variability arise from a single, relatively simple underlying neurophysiological process of motor regulation.
| During the repeated execution of goal-directed movements, statistical variability is always observed from one trial to the next, and this motor variability has long been a major focus of movement neuroscience [1–3]. It is generally believed that these inter-trial fluctuations contain crucial information about how the neuromotor system organizes itself to meet task requirements in the face of physical constraints, external perturbations, and motor noise [4–9]. Indeed, there is increasing evidence that inherent biological noise, which is present at multiple scales from the level of motor units down to the level of genes, may play a crucial physiological function in the nervous system [7, 10, 11]. However, the process by which this multiscale noise comes to be expressed as variability at the organismic level is still far from completely understood.
There is an excess of body-level degrees of freedom over those needed to specify the outcome of a typical goal-directed movement, and it is natural to expect this redundancy to affect the structure of observed variability. A number of data analysis approaches [12–14] have been developed to examine the effect of this redundancy using task manifolds, which are surfaces in a suitably-defined space of biomechanical observables, or “body states” (e.g., joint kinematic variables), that contains all possible task solutions. By definition, every point in a task manifold corresponds to a body state that results in perfect task execution, and so, as a consequence, only body-level deviations away from the manifold result in error at the goal level.
Originally inspired by ideas from research in redundant robotics, uncontrolled manifold (UCM) analysis [12, 15–17] assumes that the task manifold is defined at each instant along a given movement trajectory, and in typical applications takes the task’s goal to be represented by the average movement in a time-normalized set of trials. The ratios of normalized variances orthogonal and tangent to a candidate manifold are then used to identify possible “control variables”, with the expectation that there should be a larger variance along the manifold than normal to it. In a similar vein, motor learning has been studied by statistically decomposing observed body-level variability into tolerance, noise, and covariation (TNC) empirical “costs”, [13, 18–20], all three of which are defined with respect to a task manifold. In contrast with UCM analysis, the TNC approach conceives of the task manifold as existing in a minimal space of variables needed to specify task execution (e.g., the position and velocity of a ball at release when throwing at a target). In addition to using its covariation cost to characterize the alignment of body-level variability with the task manifold, TNC analysis crucially relates the goal-level variability to error at the body level via its tolerance cost.
This relationship between body and goal-level variability was the initial focus of a sensitivity analysis method based on the goal equivalent manifold (GEM) concept [14]. Like TNC, the GEM analysis defines its task manifold using only a minimal set of variables needed for task specification, however it makes direct use of an explicit goal function that serves as a hypothesis on the task strategy being used. The zeros of the goal function give body states yielding perfect task execution, and the set of all such solutions then gives the GEM. In addition to defining the GEM, the goal function provides a theoretical definition of the “passive” sensitivity (i.e., sensitivity independent of any applied control) to body-level disturbances, via the singular values of the goal function’s Jacobian matrix [14, 21].
While the initial GEM-based sensitivity analysis was useful for describing the geometrical structure of observed variability and quantifying motor performance, like the UCM and TNC approaches it did not provide an analysis of the temporal structure of observed inter-trial fluctuations. This limitation was addressed by subsequent developments that incorporated optimal control ideas with the GEM to create a dynamical, model-based data analysis framework. Optimal control in the presence of redundancy has been proposed as a theoretical basis for models of the neuromotor system [22, 23], and the minimum intervention principle (MIP) [23, 24] posits that little or no control will be exerted along the task manifold, since to do so would entail a waste of control effort. The expanded GEM data analysis framework allows one to create theoretical models of inter-trial fluctuations that can be used for hypothesis testing against movement data from human participants [25–27].
This initial work has demonstrated the central importance of taking a dynamical approach when analyzing motor variability. A fundamental feature of variability highlighted by these studies is that inter-trial fluctuations are found to be dynamically anisotropic with respect to the GEM [25–29]: that is, it is found that the local stability and correlation properties are congruent with the local GEM geometry, with greater stability and lower temporal correlation being associated with the components of time series transverse to the GEM, and lower stability and greater correlation for times series components along the GEM. A similar directionality in correlation properties has been found in a study of skill acquisition [30]. However, such studies have tended to examine these dynamical properties in isolation, and it is not completely clear how the various temporal properties (e.g., local stability multipliers, lag-1 correlations, etc.) relate, if at all, to the purely geometrical features of inter-trial variability arising from the task manifold itself (e.g. variance ratios, passive sensitivity). In particular, it remains an open question whether these various features of inter-trial variability should be considered as manifestations of unique neurophysiological phenomena each in their own right, or if, conversely, they are epiphenomena that naturally arise from a single, underlying regulatory process. In this paper we present evidence that supports the latter, more parsimonious interpretation.
To this end, we examine the performance of human participants as they play a virtual shuffleboard game. We chose shuffleboard for this study because it is among the simplest tasks exhibiting task-level redundancy, and is thus both mathematically and experimentally tractable. As such, it serves as a “model problem” for a much broader class of goal-directed tasks which can be expected to exhibit similar variability characteristics. Observed inter-trial fluctuations are modeled as the output of the perception-action system as participants attempt to hit the target in each trial by correcting error in the previous trial. We focus on skilled performance, and, starting with a previously-developed general model for inter-trial error correction [21, 26, 28], we present a theoretical analysis using the shuffleboard task as an illustrative example. The analysis yields theoretical predictions about the geometrical and temporal structure of inter-trial variability, culminating in a prediction of how GEM geometry, passive sensitivity, and active error correction combine to yield task performance. Specifically, we show that the scaling of the root mean square (RMS) error at the target is determined by the total body-goal sensitivity, which is, in effect, a total “gain” mapping body-level fluctuations to the goal level.
We also address a critical technical issue that arises when experimentally testing our theoretical predictions. For skilled performance, the local geometric stability properties of the fluctuations play a fundamental role, with such properties being determined theoretically by an eigenanalysis of a linearized model. Unfortunately, numerical estimates of eigenvalues and eigenvectors are known to be highly sensitive to errors in the matrix estimate [31], which are themselves unavoidable when the matrix is found using regression on experimental data. This problem is compounded by the relatively small data sets available in typical human subjects experiments. In this paper we present a new method for estimating all of our dynamical quantities based on bootstrapping [32–34], which allows us to estimate the complete underlying probability distribution for each quantity considered, resulting in the most robust demonstration to date of the degree to which dynamical anisotropy is present in inter-trial movement data. Furthermore, this data analysis allows us to confirm the theoretical performance scaling prediction to high precision, not only showing how the individual participants performed in this particular task, but also validating the many assumptions underlying our theoretical derivation.
Studies of variability using task solution manifolds typically assume that they are embedded in a space of variables with identical physical dimension, such as, for example, joint angles [14, 15, 35], muscle activation [36, 37], or finger forces [16, 38, 39]. Such situations have tended to obscure a fundamental difficulty if one intends to make inferences based on the relative magnitude of fluctuations normal and tangent to any hypothesized manifold: namely, that multivariate statistics are not invariant under coordinate transformations. This issue was recently recognized in the context of movement variability analysis [30, 40], but is a well-known problem in multivariate statistics. Indeed, the widespread utility of principal component analysis [41, 42] is based in part on the fact that correlations between variables can be completely removed with properly selected linear coordinate transformations.
It is clearly highly desirable that the inferences we make about the motor system be invariant under coordinate transformations. While it is possible to normalize the variables and make the data dimensionless, such an approach does not completely resolve the scaling issue because the choice of the normalizing constant is, in most cases, arbitrary. This problem becomes even more acute when the task manifold resides in a space composed of different physical quantities, for example positions and velocities. Given the central role played by local geometric stability in our approach, we are able to exploit the well-known fact that such dynamical properties are invariants that do not depend on the coordinates used [43, 44]. We therefore show that our approach provides a coordinate-independent characterization of the variability observed in our experiments, suggesting that the local geometric stability analysis of inter-trial fluctuations provides a new set of task coordinates that are intrinsic to the error regulation process itself.
This section begins with a discussion of the key concepts and models that theoretically ground our approach, and that culminate in a set of four experimental hypotheses. With this theoretical background as foundation, we then describe our experimental virtual shuffleboard game, the experimental protocol, and our data analysis methods.
All participants provided informed consent, as approved by the Institutional Review Board at The Pennsylvania State University.
Fig 1 shows a schematic of a theoretical shuffleboard game. The entire game takes place along a straight line. Starting the puck at x = 0, the shuffleboard cue is accelerated from rest while in contact with the puck. Thereafter, the cue decelerates and, when the contact force between it and the puck reaches zero, the puck is released with position and velocity x and v, respectively. Once released, the puck slides on the board and is decelerated by the force of Coulomb friction, with kinetic coefficient μ, between the board and the puck. The puck eventually comes to rest at x = xf. The goal-level error, e = xf − L, is the distance between the final puck position and the target.
Elementary Newtonian mechanics gives the equation of motion for the puck after release as x ¨ = - μ g, where g is the gravitational acceleration constant. For arbitrary initial conditions x and v just after release, and final velocity vf = 0, the equation of motion is easily integrated to give −v2 = −2μg(xf − x). Since perfect execution (hitting the target) requires e = xf − L = 0, we then obtain a goal function for the task as
e = f ( x , v ) = v 2 + 2 μ g ( x - L ) . (1)
Any values of x and v for which e = f(x, v) = 0 result in perfect task execution (zero error at the goal level).
Dimensionless quantities x ˜ = x / R, v ˜ = v / 2 g R, and L ˜ = L / R can be defined for some length scale R. Note that the exact value of R used in this rescaling has no significant bearing on our results: it was chosen for convenience so that when plotting experimental data the rescaled release position x ˜ = x / R ≈ 1. For the experiments described in what follows, we took L = 200cm and R = 20cm, so that the target was located at a distance of L ˜ = 10 dimensionless units. Using these rescalings in Eq (1) gives, after rearranging and dropping tildes, the goal function in dimensionless form as
f ( x , v ) = v 2 μ + x - 10 . (2)
Henceforth we use the dimensionless goal function of Eq (2).
There are an infinite number of states (x, v) that are zeros to Eq (2), corresponding to trials that hit the target perfectly. In this simple case, we can solve for this set analytically, and find, as shown in Fig 2, that it forms a 1D goal equivalent manifold (GEM)
G = ( x , v ) | v 2 = μ ( 10 - x ) , (3)
which has the shape of a parabola in the (x, v) plane. Since the performance is completely determined by the values of x and v at release, we take as our body state x = (x, v)T (where the superscript T denotes the transpose). Note that the goal function f(x) ≠ 0 for “strategies” x that are not exactly on the GEM: for this task, this value is identical to the goal-level error, e.
The GEM represented in Fig 2 exists independently of who or what performs the task. Actuating the shuffleboard cue with a single degree of freedom pneumatic actuator, a robot with tens of degrees of freedom, or a biological organism with thousands of degrees of freedom does not affect the requirements in the (x, v) body state space needed to hit the target. Furthermore, the GEM has been defined without any consideration of the control that might be applied to correct errors from one trial to the next: even a completely uncontrolled system that randomly assigned values of x and v for each trial would have the same GEM.
For a skilled participant whose performance is perfect on average, we assume that the state will be near the GEM and write x = x* + u, where the operating point x * = ( x * , v * ) T ∈ G represents the average perfect trial on the GEM, and u = (p, q)T is a small fluctuation. Substitution into the goal function Eq (2) and linearizing about u = (0, 0)T then gives
e = ( v * + q ) 2 μ + ( x * + p ) - 10 ≈ 1 2 v * μ p q ≜ A u , (4)
where A = ( ∂ f ∂ x ∂ f ∂ v ), with derivatives evaluated at (x*, v*), is the 1 × 2 body-goal variability matrix [14] that maps body-level perturbations u into goal-level error e.
The null space N of A, defined by N = { u | A u = 0 }, contains fluctuations that are goal equivalent, i.e., that to leading order have no effect on the goal level error. Using this definition, the unit tangent vector to the GEM is found to be
e ^ t = 1 1 + 2 v * μ 2 - 2 v * μ 1 , (5)
giving the 1D goal-equivalent subspace as N = span { e ^ t }, which is also the subspace tangent to the GEM at x* (again, see Fig 2). In contrast, the row space R of A contains fluctuations that result in error at the goal and, hence, are goal relevant. This 1D space is orthogonal to the GEM, so that R = span { e ^ n }, where e ^ n is the unit normal to the GEM given by
e ^ n = 1 1 + 2 v * μ 2 1 2 v * μ . (6)
Given a fluctuation u from the operating point x*, its goal-relevant and goal-equivalent components are found using the inner products
u R = u · e ^ n and u N = u · e ^ t , (7)
respectively. Using these, one can readily compute from observations the sample standard deviations of goal-relevant and goal-equivalent fluctuations, σ R and σ N, respectively.
The singular values of the body-goal matrix A determine how fluctuations u get amplified onto the target [14], and so determine the sensitivity of the performance to body-level errors. Since the sensitivity depends only on the goal function, it is independent of any specific inter-trial control mechanism, and so is considered to be a passive property of the task. For the shuffleboard game, A has one singular value s, which is given by [31]
s = 1 + 2 v * μ 2 . (8)
Thus, the passive sensitivity is a function of the friction coefficient, μ, and the speed at the operating point, v*, with the latter indicating that s is not constant along the GEM. Given s, Eq (4) can then be used to obtain the RMS goal-level error as
σ e = s σ R , (9)
which is a special case of the general expression obtained in [14]. Thus, the passive sensitivity “explains” the goal level error, but only when the goal-relevant fluctuations are taken as given. However, the scale of those fluctuations, σ R, is itself determined by the active process of inter-trial error correction.
As discussed previously, the GEM and body-goal sensitivity are passive properties of the task that exist prior to the imposition of any error-correcting control. Here, we “close the loop” on the problem by discussing simple perception-action models of inter-trial error correction. For clarity, we present our modeling framework with a bit more generality than will ultimately be needed. Additional background and details can be found in [26, 28].
A typical experiment for a goal-directed task with N trials results in a time series of the body state variable, { x k } k = 1 N, and a corresponding time series of goal-level errors, { e k } k = 1 N. We consider these time series to result from the process of error-correction used by participants as they make adjustments after each trial, and model the fluctuation dynamics with update equations of the form [21, 26, 28]:
x k + 1 = x k + G I + N k c ( x k ) + ν k , (10)
in which: c(xk) is an inter-trial, error-correcting controller depending on the current state; Nk is a matrix representing signal-dependent noise in the motor outputs [45]; and νk is an additive noise vector representing unmodeled effects from perceptual and neuromotor sources. The diagonal matrix of gains, G, is included as a convenient way to detune the model away from optimality when c is an optimal controller designed initially with G = I [26].
Error-correcting models with mathematical form similar to Eq (10) have been used to study motor learning [46–48] and to understand the effect of motor noise. These previous efforts have not focused on the role of task level redundancy, or attempted to relate body-level fluctuations to those at some external goal, as we do here. However, in contrast to these previous studies, we do not make reference to hidden internal state variables related, for example, to motor planning, but instead construct our models at the level of experimentally-observable task-relevant kinematic variables. As a consequence, our models cannot be used to disambiguate the effect of noise due to motor planning from that due to motor execution [46]. Our focus here is not on how internal “neuronal” state variables are dynamically mapped to kinematic output variables, but rather how the body-level task variables are mapped onto the goal-level task error in the presence of redundancy. Hence, our study takes place at a different level of description than studies aimed at understanding the physiological origin of motor noise and its role in motor learning. Models with the general form of Eq (10) can be viewed as the between-trial component of a hierarchical motor regulation scheme that makes error-correcting adjustments to an approximately “feed forward,” within-trial component.
Focusing once again on skilled movements, we write xk = x* + uk as was done leading up to Eq (4), where uk are small perturbations from the operating point x*. Assuming, in addition, small noise terms Nk and νk, we can linearize the controller Eq (10) [21, 28] about uk = 0 to obtain:
u k + 1 = B u k + ν k , (11)
where the matrix B = I+GJ, and J = ∂c/∂x is the Jacobian of the controller evaluated at x*. Note that, to leading order, signal dependent noise does not affect the inter-trial dynamics near the GEM [28]. Thus, small fluctuations are governed by the linear map of Eq (11), and the eigenvalues and eigenvectors of B determine the local dynamic stability properties of the system [44, 49, 50]. Specifically, eigenvalues λ with magnitude near zero (|λ|≈0) indicate that deviations from the GEM are rapidly corrected, whereas positive eigenvalues strictly less than but closer to one (0 ≪ λ < 1) indicate that deviations are only weakly corrected (that is, they are allowed to “persist”). Note that values of λ > 1 indicate instability, indicating that deviations would continue to grow in successive trials, something that is not expected in experiments. For the shuffleboard task, the body states are 2-dimensional, so that B is a 2 × 2 matrix possessing two eigenvalues, {λw, λs}, and two eigenvectors, { e ^ w , e ^ s }, where the subscripts w and s indicate weakly and strongly stable directions, as described below. We limit our discussion to the case of real, distinct eigenvalues, which has been found to be sufficient in experimental applications to date.
In [26], c was found analytically as an optimal controller using different specified cost functions. Because goal-level error was minimized as a cost, the goal function (which, for the current paper, is given by Eq 2) was built into the model, and so the effect of the GEM was explicitly included. In studies of this type, the model is used to generate simulated data, which is then statistically compared to experimental data to “reverse engineer” the controller used by human participants. Furthermore, if one wishes to study local stability properties via Eq (11), the matrix B can, in principle, be obtained analytically by differentiation.
In contrast, in this work we take a simpler, empirical approach: instead of formulating an explicit optimal controller, linear regression is used to estimate the matrix B of Eq (11) directly from the experimental fluctuation data. The eigenstructure of the estimated B is then obtained and compared to the geometry of the shuffleboard GEM (Fig 2). Thus, other than the assumption of closeness to an operating point x * ∈ G (i.e., of linearity), the controller is not assumed to to be optimal, nor is the GEM encoded into it in any way. Thus, any structure in the data related to the presence of the GEM is a property of the observed fluctuation dynamics: it has not been imposed by the model.
Task manifold methods applied to a variety of motor tasks have shown that the body-level variability observed during skilled task execution will tend to have greater variance along the task manifold than normal to it. Indeed, anisotropy in the variability is typically taken to demonstrate that a hypothesized task manifold is being used to organize motor control [12, 16]. Such results are consistent with a generalized interpretation of the UCM hypothesis and the MIP: namely, that while disturbances along the task manifold are not truly “uncontrolled”, they are, at least, more weakly controlled than those normal to it. However, movement variability may be “structured” (i.e., may exhibit anisotropy) for biomechanical and/or neurophysiological reasons that are unrelated to control [36]. In addition, variance-based analyses are vulnerable to ambiguities related to the coordinate dependence of variability statistics [28, 40], and by themselves do not provide any insight into how observed fluctuations are dynamically generated and regulated [28, 51].
A number of researchers have addressed this last limitation by combining task manifold ideas with time series analysis of statistical persistence [25–27, 30, 51–54], as measured either via detrended fluctuation analysis (DFA) [55, 56] or autocorrelations. Generally speaking, a time series exhibits statistical persistence if, given fluctuations in one direction, subsequent fluctuations are likely to be in the same direction. If subsequent fluctuations are likely to be in the opposite direction, the time series is said to be antipersistent, and if subsequent fluctuations are equally likely to be in either direction the time series is non-persistent or, alternatively, uncorrelated. As was shown in [25], the coherent interpretation of persistence results requires the consideration of error-correcting control near the task manifold: there is greater statistical persistence along the manifold, where the control is weak, than perpendicular to it, where the control is strong. These types of results are, again, consistent with a generalized interpretation of the MIP [28].
All of the above-cited studies lead us to expect dynamical anisotropy in inter-trial fluctuations. That is, the temporal structure of fluctuations should reflect the operation of a controller that strongly acts against goal-relevant deviations by pushing subsequent body-states toward the GEM, while only weakly acting to correct goal-equivalent deviations along the GEM.
Since in this paper we focus on skilled movements, we make direct use of the linearized model Eq (11). For an ideal MIP controller, the complete absence of control along the GEM would result in neutral stability along it, as well, meaning that one eigenvector of the matrix B (Eq (11)) would be identical to the unit tangent e ^ t, and its associated eigenvalue would be λ = 1. However, such a scenario in the presence of motor noise would result in an unbounded random walk along the GEM, something which has yet to be observed in experiments. Thus, we expect the inter-trial dynamics to be slightly perturbed from what one would expect for a perfect MIP controller, giving one weakly stable eigenvalue less than, but somewhat close to, 1 (i.e., 0 ≪ λw < 1) with an associated unit eigenvector ew that is close to e ^ t, but slightly rotated. In contrast, the strongly stable eigenvalue, λs, indicates vigorous correction of deviations off of the GEM, so that |λs|≈0 and es is transverse (but not necessarily perpendicular) to the GEM. The general geometry of the situation, in which local stability properties are overlaid on the GEM near an operating point x*, is show schematically in Fig 3.
The fluctuations uk in the original, laboratory coordinates (e.g., representing speed and position for the shuffleboard game) can be transformed into new fluctuations expressed in eigencoordinates via the linear coordinate transformation
u k = E z k , (12)
where E is the matrix containing e ^ w and e ^ s as its columns. Note that E is not typically an orthogonal matrix because the eigenvectors of B are not usually perpendicular. Using this transformation, Eq (11) becomes
z k + 1 = E - 1 B E z k + E - 1 ν k ≜ Q z k + n k . (13)
where z = (zw, zs)T are the fluctuations expressed in weak-strong eigencoordinates, the diagonal matrix Q = E−1 BE has λw and λs along its diagonal, and n = (nw, ns)T is the transformed additive noise term. That is, the transformation Eq (12) decouples the dynamics in the weak and strong directions so that Eq (13) can be written as
z w , k + 1 = λ w z w , k + n w , k (14) z s , k + 1 = λ s z s , k + n s , k , (15)
in which zw, k and zs, k are simply the components of zk in the weak and strong directions, respectively. This “diagonalized” form of the system illustrates the action of each eigenvalue on fluctuations in their respective directions: in the absence of noise an eigenvalue close to zero will eliminate a given fluctuation on the very next trial, whereas a positive eigenvalue a bit less than 1 will allow fluctuations to persist over many trials. The decomposition of Eqs (14) and (15) is intrinsic to the fluctuation dynamics created by inter-trial error correction, and so differs significantly from “static” decompositions using, for example, the normal and tangent to the GEM, or principal component analysis [42].
From Eq (7) and the transformation Eq (12) we can relate the standard deviations of fluctuations in the goal-relevant and strongly-stable directions as
u R = e ^ n · u = e ^ n · z w e ^ w + z s e ^ s ≈ β z s ⟹ σ R ≈ β σ s , (16)
where β ≜ e ^ n · e ^ s = sin ( θ s ) (see Fig 3) and we have assumed, consistent with a generalized MIP, that the weakly stable direction is nearly tangent to the GEM, so that e ^ w ≈ e ^ t ⇒ e ^ n · e ^ w ≈ 0. Squaring both sides of Eq (14), taking the ensemble average (as indicated by angle brackets), and assuming that the noise and fluctuations at trial k are uncorrelated, yields
z w , k + 1 2 = λ w 2 2 m u z w , k 2 + n w , k 2 ⟹ σ w = σ n w 1 - λ w 2 , (17)
where σ n w 2 ≡ 〈 n w , k 2 〉, and in which we have used the fact that at steady state 〈 z w , k + 1 2 〉 = 〈 z w , k 2 〉 ≡ σ w 2. A similar calculation with Eq (15) gives
σ s = σ n s 1 - λ s 2 . (18)
Eqs (17) and (18) show that as the eigenvalues approach 0, the “output” variance of the fluctuations approaches a minimum value equal to the variance of the “input” noise. Conversely, as the eigenvalues approach the stability boundary of 1, the output variance becomes unbounded (i.e., the fluctuations approach the behavior of a random walk).
Finally, substituting from Eq (16) into Eq (9), using Eq (18), and rearranging we find
σ e σ n s ≈ β s 1 - λ s 2 ≜ s TOT , (19)
where sTOT is the total body-goal sensitivity, which quantifies how much intrinsic body-level fluctuations are amplified at the goal level. Note that sTOT results from the interaction of the passive sensitivity (via s), the local GEM geometry (via β = sinθs) and active control “strength” (via λs).
Given zw and zs time series from the diagonalized controller of Eqs (14) and (15), we can compute the normalized lag-1 autocorrelations of the fluctuations in the weak and strong directions as
R w ( 1 ) = ( z w , k + 1 ) ( z w , k ) σ w 2 and R s ( 1 ) = ( z s , k + 1 ) ( z s , k ) σ s 2 , (20)
respectively. This provides a simple quantification for the statistical persistence in both directions. However, multiplying Eq (14) by zw, k, taking the ensemble average, and assuming the additive noise is uncorrelated with the fluctuations so that 〈(zw, k)(nw, k)〉 = 0 gives
( z w , k + 1 ) ( z w , k ) = λ w ( z w , k ) ( z w , k ) ≡ λ w σ w 2 . (21)
Solving for λw in the above and comparing it to the definition Rw(1) in Eq (20), we see that Rw(1) ≡ λw. Likewise, a similar calculation with Eq (15) shows Rs(1) ≡ λs. Thus, as a persistence measure the normalized lag-1 autocorrelation does not, theoretically speaking, provide information distinct from the eigenvalues λw and λs. We include it here to demonstrate the connection between stability and this simple persistence measure. We use it later, as well, to serve as a consistency check on our experimental eigenvalue estimates.
To test for statistical persistence with a method independent from the eigenanalysis, one can apply detrended fluctuation analysis (DFA) [55, 56] with linear detrending to the zw and zs time series. The DFA algorithm yields a positive exponent, α, where α < 0.5 indicates antipersistence in a time series, α > 0.5 indicates persistence and α = 0.5 indicates non-persistence. Contrary to its most common use in the literature, in this work we are not using DFA to claim that observed fluctuations exhibit long-range persistence, but instead employ α merely as a convenient overall measure of persistence that, unlike the autocorrelation, does not require consideration of specific lags. Additional discussion regarding the application of DFA to movement variability data can be found in [28], including a review of its vulnerability to false positives when testing for long-range persistence [57–59].
In this subsection we show how the dynamical analysis of inter-trial fluctuations allows us to characterize observed variability in a way that is insensitive to the choice of coordinates. Starting with some original body state variable x, consider a new variable y of the same dimension as x, with each being related by a general differentiable, invertible coordinate transformation x = g(y). Thus, the operating point expressed for each choice of coordinates is related by x* = g(y*), and we find that small fluctuations are related to lowest order by a linear transformation from:
x * + u k = g ( y * + v k ) ≈ g ( y * ) + T v k ⟹ u k = T v k , (22)
where uk and vk are the fluctuations expressed in terms of the old and new coordinates, respectively, and T is the square Jacobian matrix of the transformation g evaluated at y*.
Using Eq (22) to substitute for uk into the linearized controller Eq (11) then gives, in a manner analogous to that used to obtain Eq (13):
v k + 1 = T - 1 B T v k + T - 1 ν k . (23)
Clearly, the matrix T−1BT on the right-hand side of the above equation is congruent to the original B, and so will have the same eigenvalues, and, hence, the same stability properties.
As discussed in [28], the GEM itself is transformed when using the new coordinates. Recall from the discussion prior to Eq (5) that the tangent to the GEM is determined from the null space of the Jacobian to the goal function, A. That is, to leading order the fluctuation uk is on the GEM whenever Auk = 0. However, again using the transformation Eq (22), we see that Auk = ATvk, showing that whenever uk is on the GEM expressed in terms of the original coordinates, vk is on the GEM expressed using the new coordinates. Thus, not only are the stability properties unaffected by coordinate transformations, the eigenvectors and GEM are transformed in a predictable way that preserves the topology near the operating point: that is, while changing coordinates will typically rotate and shear the picture somewhat, the overall arrangement illustrated in Fig 3 is preserved.
Following the above discussion, we are led to the following four theoretical predictions, presented here as experimental hypotheses, which we here simply state directly. Additional computational details, as required to test the hypotheses, are presented in the Data Analysis section below. As a convenience to the reader, Table 1 contains a glossary of the key symbols used in stating the hypotheses.
Hypotheses H1–H3 can be tested directly by examining the eigenstructure of the matrix B in Eq (11). They are dynamical consequences of the more general hypothesis that Eq (11) is derived from a “GEM aware” controller, and hence strives to eliminate goal-relevant deviations quickly, after only one trial, while allowing goal-equivalent deviations to persist for multiple trials. In contrast, hypothesis H4 emphasizes how the overall goal-level performance (as measured by σe) will result from the interaction between the strongly-stable component of the intrinsic “input” noise (measured by σns), inter-trial error correction, and passive sensitivity.
The total body-goal sensitivity, sTOT, is an overall “gain” between body-level noise and goal-level error. We expect λs ≈ 0, and β = sin(θs)<1 (Fig 3). Thus, β / 1 - λ s 2, which is the “active factor” of sTOT will have a value on the order of unity. In contrast, the “passive factor” of sTOT, which is simply the passive sensitivity s (Eq (8)), may be substantially greater than unity. Thus, a somewhat counterintuitive effect of error-correcting control is that the passive sensitivity, which is determined by task properties independent from control, may play a dominant role in determining motor performance at the goal level.
Fig 4 shows a schematic representation of the experimental set-up for the shuffleboard game in a virtual environment. The participant was seated in an upright position, and in each trial moved a custom-built input device consisting of a manipulandum affixed to a low friction, single degree of freedom, linear bearing. Participants held the manipulandum with their dominant hand and pushed it in a direction parallel to the ground plane. The apparatus was configured for each participant so that at rest the upper arm was aligned with the midaxillary line and the angle between the upper arm and the forearm was approximately 90°.
Each trial started with the puck at x = 0 (recall Fig 1). The participant accelerated the manipulandum from rest. Position data was acquired from the manipulandum’s motion and used to generate the motion of a virtual shuffleboard cue in real time, via custom software, which pushed the puck on the virtual court. The release of the puck happened as the cue decelerated and the virtual contact force between the cue and the puck decreased to zero. At the point of release, the position and velocity, x and v, of the puck were acquired, defining the body state for a given trial. Thereafter, the acquired values of x and v were used to compute the motion of the puck as it slid on the virtual court and was decelerated by Coulomb friction before coming to rest. The movement of the shuffleboard cue and puck during the entire trial was generated in real time by the control software and projected onto a screen. Participants could see an animated 3D scene showing the movement of the puck on the court as it moved toward a visible target line before coming to a stop. The projector (InFocus LP70+) was located to the right and just behind the participants, approximately 3m from a 1.7m × 1.3m screen, with the settings adjusted for flicker-free images that filled the screen.
The position and velocity data were obtained from two transducers placed on the manipulandum and collected through two 12-bit channels: an accelerometer (ADXL320, Analog Devices, Inc., Norwood, MA) was used to collect acceleration data, which was integrated to provide the velocity; the other channel collected position data from a linear variable displacement transducer (LVDT) (Daytronic Corporation, Dayton, OH). The LVDT was also used to calibrate the accelerometer by scaling the doubly integrated acceleration signal to match the position signal. A National Instruments NIDAQCard-6024E data acquisition card was used to acquire the data to a laptop computer. A virtual instrument written in LabVIEW (National Instruments, Austin, TX) passed the velocity and position information in real time to a C++ program which used the Visualization Toolkit (VTK, http://www.vtk.org), an open-source graphics library, to render the 3D virtual environment. Both signals were sampled at 5kHz to provide smooth animation in the virtual environment. Even though the virtual environment has no physical units per se, we designed the system so that all VTK representations of lengths matched centimeters in the physical world: the accelerometer and LVDT were calibrated and data was recorded in cm/s2 and cm, respectively.
We expected the dynamical anisotropy predictions (H1–H3) to depend primarily on the local geometry of the GEM, and to not, therefore, depend on the friction coefficient μ. On the other hand, the scaling prediction, H4, depends on μ via the passive sensitivity, since s = s(μ) from Eq (8). Therefore, we had each participant perform the task with two different friction levels in the virtual world, giving a total of eight different participants/conditions. For a given velocity and position at release, the time of motion before the puck stops is inversely proportional to the coefficient of friction. We therefore selected values of μ so that the time for a hypothetical ideal trial varied uniformly between 3s and 5s. This ideal trial was defined by a release position of x = 0 and release velocity v determined from the goal function Eq (2) so that the puck would stop exactly at the target. The resulting set of 8 μ values were split into two sets: the lowest 4 gave “low friction” (LF) conditions, and the highest 4 “high friction” (HF) conditions. These different friction conditions gave us inter-trial data sets generated with different passive sensitivity properties, via Eq (8).
Four healthy, right-handed male participants aged 25, 28, 29 and 33 years (labeled P1–P4) participated in this study. Each participant was randomly assigned one HF and one LF friction condition to perform the shuffleboard task. The participants were instructed to launch the puck so that its center stopped on the target in every trial. Participants had the visual feedback from the 3D scene showing the error from a given trial. The goal-level error was also displayed momentarily on the screen providing a second, more precise, feedback on their performance. All participants were allowed to familiarize themselves with the task and the equipment, and practiced hitting the target until their average error e (Fig 1) over 50 trials was less than 10% of the target distance. That is, participants practiced until the average state x ¯ = ( x ¯ , v ¯ ) T acquired over 50 trials lay within the error contours of Fig 2. All participants achieved this level of performance within four blocks of 50 trials.
Once the participants achieved the required level of performance, the data collection phase began. The body state x = (x, v)T and goal-level error e were recorded for each trial. For each of the two friction conditions (LF and HF) the participant was required to perform 500 trials. All of the data was collected over three days: two days each of four 50-trial blocks, with two blocks before noon and two in the afternoon, followed by a day of two 50 trial blocks. Each block took no more than seven minutes and the participant was given up to five minutes of rest between blocks. The last block of P1-HF was incomplete due to an experiment malfunction, so only data from the first 9 blocks (450 trials) were subsequently analyzed; P3-HF had only 350 usable trials due to the entry of an erroneous friction coefficient. Typical inter-trial time series of states x = (x, v)T obtained from one participant over 500 trials are shown in Fig 5(a)–5(c).
The complete data set for each of the 8 friction conditions (4 participants × 2 conditions each) consisted of time series of release position and velocity, { x k } k = 1 N and { v k } k = 1 N, respectively, and the corresponding error, { e k } k = 1 N, for each of N = 500 trials. The data was rescaled into dimensionless form, as for the goal function of Eq (2). Note, however, that the stability and persistence properties studied here depend only on the temporal relations between consecutive trials, so the rescaling does not affect the results presented in this paper. Except as noted, all data analyses were performed using Matlab (Mathworks, Natick, MA). All data and software used for this study is contained in Supporting Information S1 Data and Code.
The sample mean body state x ¯ = ( x ¯ , v ¯ ) T over all trials was used to define the operating point used in Eq (4): that is, we took x * ≡ x ¯. Fluctuation time series were then obtained from u k = x k - x ¯, and Eq (11) was used to estimate B via linear regression. That is, we used ordinary least squares to minimize the single-step mean-square prediction error 〈(uk+1 − Buk)T(uk+1 − Buk)〉, where, again, the angle brackets denote the ensemble average. A requirement for the use of this straightforward approach to estimation [60–62] is that the state measurement error or “noise” (as distinct from the process noise νk in Eq (11)) not be too large. While there is no firm cutoff for how much measurement noise becomes problematic, Kantz and Screiber suggest (see [62], p. 251 ff.) that ordinary least squares works well as long as the measurement errors are under about 10%. In our case the measurement precision after calibration was approximately 2%, well under the suggested cutoff. Furthermore, we cross validate the estimate of B by comparing its eigenvalues against the lag-1 autocorrelation, which is computed independently, as discussed previously following Eq (21).
The eigenvectors of B, { e ^ w , e ^ s }, and their corresponding eigenvalues, {λw, λs}, were then obtained as solutions to B e ^ = λ e ^. A typical result of this eigenanalysis is shown in Fig 5(d). The alignment of the eigenvectors to the GEM was computed using the theoretical tangent vector from Eq (5) (recall the schematic of Fig 3). Because the empirically-determined operating point x ¯ was always close to, but never exactly on the GEM, as a check we also computed the eigenvector orientation using the tangent to the error contour passing through the operating point (determined from by f ( x ¯ ) = e ¯, where f is the goal function Eq (2)). This was found to give identical results, confirming the closeness of x ¯ to the GEM. Together with the alignment information so obtained, the estimated eigenvalues of B, which quantify the stability of the inter-trial dynamics, were used to test H1 and H2.
Next, the fluctuation time series { u k } k = 1 N in the original position-speed coordinates were transformed into time series { z k } k = 1 N expressed in eigencoordinates, via the linear coordinate transformation Eq (12). Following the discussion surrounding Eqs (20) and (21), statistical persistence in both directions was quantified using the lag-1 autorcorrelations Rw(1) and Rs(1), as well as the DFA exponents αw and αs. These results allowed us to test H3.
To test the scaling relationship of H4, the RMS goal-level error σe was computed directly from the time series, { e k } k = 1 N. Using Eq (8), the value of μ for a given set of trials, and the velocity component of the average operating point, v ¯ ≡ v *, we obtained an estimate of s. The values of β and λs were available from the eigenanalysis. For σns, we used the estimated B and Eq (12) to compute the residual of the regression expressed in eigencoordinates, via rk = E−1(uk+1 − Buk). We then took 〈 | r s , k 2 | 〉 as an estimate of σns, where rs,k is the strongly stable component of rk. Using these estimates to evaluate Eq (19) allowed us to test H4.
All of the above analyses depend critically on the eigenvalues and eigenvectors of the matrix B. To estimate B via regression we require only data from a set of trials, which need not themselves be consecutive, together with the subsequent states that are presumed to follow under the action of B via Eq (11). To eliminate the spurious “state update” between the last trial in each block and the first trial in the next block, we only consider the first 49 trials within each 50 trial block. In addition, to avoid possible transient “retraining” effects at the beginning of each block, we removed the first 4 trials, leaving 45 trials within each block, for a total of 450 trials per friction condition. Finally, to overcome known problems associated with the sensitivity of eigenvalue and eigenvector estimates to matrix errors [31], such as are unavoidable with matrices estimated via regression, we used bootstrapping [32–34] to estimate the various quantities needed to test our hypotheses.
For each iterate of the bootstrap, we selected a uniformly-distributed random sample of 450 states (with replacement) from the 450 available for each friction condition, together with the state from the next trial. In this way, we obtained an ensemble of “current states” (xk) and an ensemble of the corresponding “next states” (xk+1) that were used to obtain one estimate of B via linear regression. This estimate of B was then used to compute one set of eigenvalues and eigenvectors. The eigenvectors were then used to obtain the fluctuation components in the weakly and strongly stable directions, zw and zs, via the transformation Eq (12). These allowed us to estimate the lag-1 autocorrelations using Eq (20). By choosing many such random samples, each resulting in its own estimate of B, we were able to generate an empirical probability distribution for all quantities needed to test H1 and H2, and to partially test H3 using R(1). The bootstrapping gave us reliable estimates of mean values together with 95% confidence intervals. For the above results, we used 10000 bootstrap iterates.
Since DFA relies on the proper temporal sequence of an entire data set (not just over a single lag as for the autocorrelation), the sampling procedure outlined above could not be used. In addition, because DFA does not give reliable estimates for small data sets, we concatenated all 10 trial blocks, again with the first four trials removed, and analysed the resulting data set of 460 trials at once. Such a concatenation procedure was shown in an analysis of Parkinsonian gait [63], using data sets of 25 strides each, to give results with sufficient accuracy to distinguish Parkinsonian and healthy participants. While perhaps not accurate enough to characterize subtle differences in long-range correlated data sets, as stated earlier this is emphatically not our aim here: we merely use DFA to provide a convenient, lag-independent measure of statistical persistence, which we checked against the lag-1 autocorrelation for consistency. For this paper, once the eigenvectors were found within each iterate of the bootstrap, the entire time series of fluctuations was transformed into eigencoordinates, again via Eq (12). The DFA exponents, αw and αs, for the two eigencoordinate fluctuations were then obtained, allowing us to complete the test of H3. To reduce the computation time required to carry out 10000 DFA calculations for each friction condition, we used a version of the algorithm written in C [64], that was then called from Matlab.
Finally, to test H4, another variant of the bootstrap was used. In each bootstrap iteration, 450 samples with replacement were drawn and used to estimate σe, σns, s, β and λs, as needed for Eq (19); this was done for all 8 friction conditions. Within this bootstrap iteration, regression was then used to estimate the parameters a and b of a fit σe/σns = asTOT + b: following Eq (19), we expected a ≈ 1 and b ≈ 0. Thus, after repeating this process 10000 times, we obtained estimates and confidence intervals for the slope a and y-intercept b, as required to test H4.
Fig 6 shows empirical probability density functions (EPDFs), obtained using bootstrapping, for the eigenvalues {λw, λs} of the matrix B (Eq (11)). We see that in all cases they satisfy 0 ≈ |λs| ≪ λw < 1. In aggregate, across all participants (P1–P4) and friction conditions, we found λs = −0.03 [−0.24, 0.14] and λw = 0.76 [0.62, 0.90], where here and throughout the stated estimate is the aggregate mean, and the closed interval represents the aggregate 95% confidence interval (CI). The orientation of the eigenvectors is shown in Fig 7, which plots the EPDFs for the angles θ w = cos - 1 ( e ^ w · e ^ t ), and θ s = cos - 1 ( e ^ s · e ^ t ). We see that, for all participants/conditions, the weakly stable eigenvector was very close to the tangent, and the strongly stable eigenvector made a larger transverse angle with it, so that 0 ≈ |θw| ≪ θs. Specifically, we found θw = 0.90° [−2.36°, 3.99°] and θs = 79.75° [20.66°, 144.75°]. We note that the orientation of the weakly stable subspace is tightly regulated to be near the GEM’s tangent (i.e., its CI is small, spanning less than 7°), whereas the orientation of the strongly stable subspace is not tightly regulated (its CI spans over 124°). The aggregate values of the matrix components of B were found as B(1, 1) = 0.76 [0.62, 0.90], B(1, 2) = −0.26 [−2.03, 1.19], B(2, 1) = −0.01 [−0.04, 0.03], and B(2, 2) = −0.03 [−0.25, 0.14]. Using the mean matrix components as a simple consistency check, we found values of λw and λs equal to the means obtained via bootstrapping, above.
The results shown in Figs 6 and 7 strongly support hypotheses H1 and H2. We found that the component of the inter-trial dynamics directed along the strongly stable subspace acted to quickly correct deviations off of the GEM that caused goal-level errors. For example, for the estimated mean value λs = −0.03, Eq (15) shows that a deviation transverse to the GEM would be, in the absence of noise, reduced to 3% of its initial magnitude after only one trial. In contrast, the dynamics in the weakly stable subspace did not rapidly correct deviations that were approximately tangent the GEM, and which therefore had little effect on error at the target. For the mean value of λw = 0.76, Eq (14) shows that, in the absence of noise, 9 iterates would be required to reduce an initial deviation to less than 10% of its initial value.
In Fig 8 we show the EPDFs obtained for the normalized lag-1 autocorrelations of fluctuations in the two eigendirections, for all friction participants/conditions. We find in all cases that 0 ≈ |Rs(1)| ≪ Rw(1). Specifically, we estimate Rs(1) = −0.03 [−0.24, 0.14] and Rw(1) = 0.76 [0.64, 0.88]. These results indicate that the trial-to-trial fluctuations in the weakly stable direction show greater persistence than those in the strongly stable direction. Furthermore, the strong control results in fluctuations that are close to uncorrelated white noise (since Rs(1) ≈ 0). As anticipated in the discussion following Eq (21), these results are nearly identical to the local stability results in Fig 6. The EPDFs obtained for the DFA exponents αw and αs for fluctuations in the weakly and strongly stable subspaces, respectively, are shown in Fig 9. We found αs = 0.52 [0.44, 0.59] and αw = 0.99 [0.89, 1.16]. Thus, in all cases 0.5 ≈ αs ≪ αw, showing substantial persistence between successive fluctuations in the weakly stable direction, and nearly uncorrelated fluctuations in the strongly stable direction. Thus, the persistence results of Figs 8 and 9 are consistent with each other and, taken together, strongly confirm H3.
Finally, Fig 10 illustrates how the variability ratio σe/σns, which represents an empirical “gain” between intrinsic body-level noise and goal-level variability, was found to linearly scale with the total body-goal sensitivity sTOT (Eq (19)). The light gray dots in the plot represent values obtained by bootstrapping: one such point was generated for all 8 friction conditions and linear regression was applied within each of 10000 iterations. This process yielded estimates for the slope, a = 0.99 [0.93, 1.03], and y-intercept, b = 0.21 [−0.98, 1.52]. The resulting aggregate fit had an R2 of 0.996. As a check, we used all 8 × 10000 points at once for a single linear fit; this did not change the fit parameters or the R2 value. The figure also includes the average values obtained for each participant/condition, computed independently by bootstrapping, together with error bars representing 95% CIs. The uneven size of the error bars, especially in the horizontal direction, reflects the nonlinearity of sTOT, particularly the factor of β = sin(θs). We see that in each case the mean points fall very near the linear fit, indicating that the scaling relationship held not only in aggregate, but for each participant/condition individually. Indeed, similar fits done for each participant independently yielded R2 estimates of 0.962, 0.991, 0.979 and 0.992, values not meaningfully different from the overall value. Thus, we concluded that for all participants/conditions Eq (19) holds, confirming hypothesis H4.
We conclude this section with an illustration of how our approach overcomes the potential interpretive ambiguity stemming from the coordinate dependence of variance [28, 40]. As discussed when presenting Eqs (22) and (23), the dynamical analysis carried out here yields quantities that are intrinsic to the observed temporal fluctuations, and hence are coordinate invariant. As a demonstration of this invariance, and its advantage in analyzing motor variability, we constructed a “worst case” coordinate transformation similar in form to Eq (12). However, in this case we defined new fluctuation coordinates q = (q1, q2)T via u = Pq, where the matrix P was obtained from principal component analysis [42], as follows: let P = SC, in which C is a matrix with columns composed of the eigenvectors (i.e., the principal components) of the fluctuation covariance 〈uuT〉, and S is a diagonal matrix with the square root of the inverse principal values, 1/σ1 and 1/σ2, along its diagonal. The result of applying this transformation is that both of the new coordinates q1 and q2 have identical variance, and hence the variance “cloud” in the (q1, q2) plane is isotropic by construction (i.e., the variance ellipse is a circle).
Fig 11 shows what happens when we apply this transformation to typical data from a single participant and friction condition. In Fig 11(a) we see the original data and the local stability results estimated from it, whereas in Fig 11(b) we see the equivalent analysis carried out on the transformed data. The eigenvalues obtained are identical in both cases, since the original matrix, B (Eq (11)), and the transformed matrix, P−1BP, are congruent. Furthermore, as discussed following Eq (23), the transformed eigenvectors maintain their qualitative relationship with the transformed GEM: that is, the weakly stable subspace is nearly tangent to the GEM, whereas the strongly stable subspace is transverse to the GEM at a much greater angle. Thus, in both cases 0 ≈ θw ≪ θs so that the local stability picture is qualitatively unchanged by the coordinate transformation, and can be used to test a candidate GEM in either case. In sharp contrast, using the shape of the variance ellipse to identify the GEM location works reasonably well for Fig 11(a), but clearly fails for the case shown in Fig 11(b). Indeed, using an approach similar to that used to create Fig 11(b), one can change the shape of the variance ellipse at will, while in all cases maintaining the proper qualitative relationship between the GEM and the weakly and strongly stable subspaces.
Understanding how humans are able to perform accurate and repeatable goal-directed movements in the presence of inherent biological noise [7–11] and neuromotor redundancy [22–24] has been a critical goal of neuroscience research (e.g., [45, 46, 48]) since the pioneering work of Bernstein [1]. In recent years, studies addressing this question have focused on using either task manifold ideas to address redundancy (e.g., [12–14]), or time series analysis methods to study temporal correlation structure (e.g., [25, 51, 54, 55]).
However, these often divergent perspectives have not yet been fully unified into a comprehensive theoretical framework, and it remains an open question whether these various aspects of inter-trial variability represent distinct neurophysiological phenomena, or can be traced back to a single underlying motor regulation process. The work in this paper expands on previous efforts [25, 28] suggesting that such a unification can be achieved by considering the inter-trial dynamics of fluctuations near a task’s goal equivalent manifold (GEM). These studies have shown that a fundamental feature of such inter-trial fluctuations is that they are dynamically anisotropic in a manner that respects the local geometry of the GEM [25–29], an observation supported by work carried out from different task manifold perspectives [30, 54, 65].
Using a custom-built interactive virtual environment, we studied the variability exhibited by skilled participants as they carried out repeated trials of a simple shuffleboard game. The experiments were used to test theoretical predictions obtained from a new analysis, presented in this paper, of a previously-developed general model for inter-trial error correction [25, 28]. The assumption of skilled performance, for which body states will remain close to the GEM, yields a simple linear inter-trial control model. The further empirically-supported assumption that inter-trial error correction satisfies a generalized interpretation of the minimum intervention principle (MIP), together with an analysis of geometric stability, yielded theoretical predictions about the geometrical and temporal structure of inter-trial variability, showing analytically how body-level variability generates variability at the goal level. In particular, we showed that the assumptions underlying our analysis give rise to a new scaling relationship (Eq (19)), which introduces the total body-goal sensitivity, sTOT, a quantity showing how intrinsic goal-relevant fluctuations at the body level are mapped into fluctuations at the goal level. This relationship provides a unification of task manifold, control theoretic, and dynamical (time series) perspectives by showing specifically how the GEM geometry, passive sensitivity, and active error correction combine to yield task performance.
The predictions resulting from our analysis were summarized in the form of four experimental hypothesis, which were tested using data from four participants playing the shuffleboard game. To demonstrate the generality of the dynamical anisotropy predictions (H1–H3), and, more importantly, to allow us to tease apart active and passive effects in task performance as specified by the scaling prediction H4, we had each participant perform the task with two different friction levels, giving a total of eight different participants/conditions. All of our hypotheses were very strongly confirmed: in all cases, the difference between local stability and correlation properties in the weakly and strongly stable directions was just as predicted by theory (Figs 6–9), confirming H1–H3; and the goal-level performance scaled as predicted across all participants and friction conditions (Fig 10), confirming H4.
Given the nature of H4, which concerns the scaling relationship Eq (19) and therefore depends on all assumptions used in its derivation, these experimental results do more than characterize the behavior for these particular participants executing this particular task. Rather, they serve to validate our general model for inter-trial error-correcting control near the GEM. Thus, while this work does not make any direct ties to underlying physiological mechanisms, our results indicate that the combined geometrical and temporal structure of observed fluctuations can be explained by a single, relatively simple process. This supports the idea that one need not posit separate neurophysiological mechanisms for controlling such disparate features as the geometric distribution of trials about the GEM, the stability of inter-trial fluctuations, and the goal-level performance, but, rather, that all such behaviors arise from a single, unified process of error regulation in the presence of task-level redundancy.
Another contribution of this paper is the introduction of statistical bootstrapping [32–34] to the analysis of movement variability data. Using this approach, we were able to estimate the underlying probability distribution for quantities required by each hypothesis (e.g., eigenvalues, correlations, etc.), thus demonstrating that the predicted dynamical anisotropy is very highly significant in each case individually (Figs 6–9), without the need for conventional significance testing. Furthermore, this data analysis allowed us to confirm the theoretical performance scaling prediction (Fig 10) to high precision, thus demonstrating that task performance was largely determined by passive sensitivity, which in this case was a function of the friction condition (Eq (8)). This theoretical prediction is perhaps counterintuitive, because the passive sensitivity is determined entirely by the task’s goal function (Eq (2)), independent from any consideration of control. However, this behavior occurs precisely because error-correcting control strongly compresses variability onto the GEM. Thus, as shown theoretically by using Eq (18) in Eq (16) (with the understanding that λs ≈ 0, as shown in Fig 6), the scale of goal-relevant fluctuations is minimized, taking a value proportional to the scale of the strongly-stable component of the intrinsic noise. Therefore, for skilled participants, the resulting performance (as measured by the RMS error at the goal) is largely determined by the passive sensitivity, which is a property of the task as defined by the goal function.
Finally, as shown in our theoretical discussion and demonstrated with our experimental data, the dynamical approach used for this study yields results that are invariant under quite general (differentiable and invertible) coordinate transformations, something that is not true for variability analyses based only on the spatial distribution of body states near a given task manifold. Even in the “worst case” scenario for which coordinates are chosen that render the variability cloud isotropic, so that it contains no information about the location of the GEM, as shown in Fig 11, the dynamical approach yields correct information about the structure of inter-trial fluctuations. Thus, our data analysis methods resolve the persistent problem of coordinate dependence of variability measures [30, 40]. This suggests that the dynamical coordinates, as obtained via the transformation Eq (12), provide a set of objective, canonical coordinates for the study of inter-trial variability: that is, they represent coordinates that are intrinsic to the regulatory process responsible for inter-trial error correction.
These findings again highlight the critical importance of considering fluctuation dynamics [25–27, 30, 51–54] in both theoretical and experimental studies aimed at understanding the neuromuscular control of complex movements. While time series analyses alone can yield important descriptive information, in the absence of any underlying model they often have limited explanatory power. Recent efforts have seen the use of time series analyses to interpret model outputs and/or predictions [46, 48, 54, 66]. These efforts have yielded findings qualitatively similar to ours, and consistent with our interpretations of inter-trial variabilty presented both here and elsewhere [25, 26, 28, 29]. Even though these efforts have focused on motor learning, which we do not, conceptually there is a strong affinity between these papers and the work presented here. In [46, 54, 66], van Beers and colleagues used simple linear models with direct error feedback to analyze task performance when reaching to a point [46, 66] or a line [54]. Their lag-1 autocorrelation analyses for the redundant task of reaching to a line showed strong statistical persistence along the target line and uncorrelated fluctuations perpendicular to it, precisely as we would theoretically predict and very similar to our own findings (our Figs 8 and 9). In parallel work, Abe & Sternad [30] also obtained similar results applying both lag-1 autocorrelation and DFA analyses to van Beers’ model of the same task. Both studies thus independently support the experimental results presented here.
The analytical formalisms presented in the present paper, however, add several important extensions to these experimental observations. First, here we tie these time series analysis approaches directly to the stability properties of the dynamical system that generates the observed fluctuations, as determined by its eigenvalues and eigenvectors (Figs 6 and 7). Second, by formally defining the task in terms of a goal function (Eq (2)), we are able to show analytically (Eq (19)) how active and passive properties of the task interact to affect goal level fluctuations, a theoretical prediction that we test and confirm experimentally (Fig 10). Finally, van Beers’ model accounts only for the correction of goal-relevant errors, that is, of body-level fluctuations perpendicular to the GEM, and thus implements an ideal MIP-based controller with no control acting along the task manifold. However, as we have shown in previous work using models derived using a stochastic optimal control framework [25], and as discussed here and demonstrated experimentally by us [28] and others [36], such “pure” MIP controllers are not observed experimentally: that is, we find that the fluctuations along the GEM do not exhibit an unbounded random walk. Furthermore, our approach allows us to demonstrate this deviation from ideal MIP behavior geometrically, as well as in terms of stability and correlation properties. A conclusion of our work is that, while the control observed experimentally is congruent with the task manifold, it is not perfectly aligned with it: instead, the direction of “minimum intervention” (i.e., of weakest control) is close to, but not exactly tangent to the GEM. Nor is the direction of strongest control necessarily perpendicular to the GEM. One possible interpretation of these observations is that there are other competing costs, beyond simple error correction, that are at play during repeated task execution.
Other recent attempts to connect temporal analyses to task manifold geometry [27, 51] have similarly supported our experimental findings, but have not directly shown how such results can be predicted from a general model-based analysis, as the current work does. Dingwell et al. [27] applied lag-1 correlation analyses to a redundant reaching task, but did not directly connect those experimental analyses back to any underlying computational model. Rácz & Valero-Cuevas [51] used DFA analyses on data from a redundant, 3-finger grasping task to provide an experimental demonstration of the need to consider control as acting across both spatial and temporal domains. However, their work again did not provide mathematical theory able to explain and predict the observed behaviors. Nevertheless, in spite of these differences in experimental and/or computational approaches, each of the studies described above obtained findings consistent with our conclusion that the diverse geometrical and temporal aspects of inter-trial variability likely derive from a single underlying motor regulation process.
Our approach fully integrates task manifold geometry with ideas from control and dynamical systems theory, and thereby can be used to explain the structure of observed motor variability from a model-based, theoretical perspective. The theory and methods presented in this paper are quite general, and should be applicable to the study of skilled motor performance for a wide range of discrete, or discretizable, tasks. That said, general application can be expected to encounter difficulties, especially for tasks in which the relevant body and/or goal variables are high-dimensional (so that visualizing the GEM is difficult, if not impossible), as well as for tasks in which the goal function and GEM are not readily available in analytical form. In such cases, the basic theory will have to be used to formulate suitable, purely abstract, computational methods.
The assumption of skilled motor behavior, which implies that all fluctuations are near the GEM, permitted us to employ linear mathematics in our study. Without this linearity, it would have been much more difficult to make such precise, analytically-derived predictions. However, we did not impose linearity as a mere analytical convenience. On the contrary, our results show that a linear model of “GEM-aware” error correction captures key facets of the observed variability structure with substantial accuracy. The main aims of this paper were to robustly demonstrate the nature of dynamic anisotropy, to show how task performance is generated by the interaction of the GEM geometry and inter-trial error correction, and to demonstrate that such an approach yields results that are not sensitive to the coordinates chosen. As such, our focus on the steady state (i.e., learned) behavior of the inter-trial regulation system was appropriate. But this does not mean that the models and methods presented here would not have value for studies related to motor learning. Indeed, as discussed at some length above, models with a very similar mathematical structure have been used to precisely that end. From a dynamical systems perspective, our approach treats skilled movements as a “stochastic attractor” of the more general perception-action system engaged in motor learning. A logical point of departure for future work aimed at extending our methods to motor learning would be to study how the the “transient” portion of the a learning data set approaches the “steady-state” local geometrical structure uncovered using the methods of this paper. While such explorations would no doubt pose multiple challenges, in principle the theoretical concepts presented here could be extended to address questions of learning and/or adaptation, topics that we see as interesting aims of future work.
|
10.1371/journal.pcbi.1002465 | Optimization of Muscle Activity for Task-Level Goals Predicts Complex Changes in Limb Forces across Biomechanical Contexts | Optimality principles have been proposed as a general framework for understanding motor control in animals and humans largely based on their ability to predict general features movement in idealized motor tasks. However, generalizing these concepts past proof-of-principle to understand the neuromechanical transformation from task-level control to detailed execution-level muscle activity and forces during behaviorally-relevant motor tasks has proved difficult. In an unrestrained balance task in cats, we demonstrate that achieving task-level constraints center of mass forces and moments while minimizing control effort predicts detailed patterns of muscle activity and ground reaction forces in an anatomically-realistic musculoskeletal model. Whereas optimization is typically used to resolve redundancy at a single level of the motor hierarchy, we simultaneously resolved redundancy across both muscles and limbs and directly compared predictions to experimental measures across multiple perturbation directions that elicit different intra- and interlimb coordination patterns. Further, although some candidate task-level variables and cost functions generated indistinguishable predictions in a single biomechanical context, we identified a common optimization framework that could predict up to 48 experimental conditions per animal (n = 3) across both perturbation directions and different biomechanical contexts created by altering animals' postural configuration. Predictions were further improved by imposing experimentally-derived muscle synergy constraints, suggesting additional task variables or costs that may be relevant to the neural control of balance. These results suggested that reduced-dimension neural control mechanisms such as muscle synergies can achieve similar kinetics to the optimal solution, but with increased control effort (≈2×) compared to individual muscle control. Our results are consistent with the idea that hierarchical, task-level neural control mechanisms previously associated with voluntary tasks may also be used in automatic brainstem-mediated pathways for balance.
| The nervous system has the ability to rapidly and flexibly coordinate many muscles and limbs to produce movements. This neuromechanical transformation must robustly achieve motor goals under the changing mechanics of the body and environment, and select one solution amongst many alternatives. What computational principles govern such decisions? Although optimality principles have predicted features of biological movement in simple models, here we show that this computational principle can robustly predict detailed experimental measures in an unrestrained, whole-body balance task. Detailed patterns of muscle activity and forces across multiple movement directions and body configurations were predicted based on interactions between musculoskeletal mechanics of the limbs, and task-level neural strategy of controlling the CoM mechanics while minimizing control effort. Moreover, similar muscle activity and forces were generated when muscles were coupled together in groups called muscle synergies, reducing the number of independent variables that are controlled. Our work is consistent with the idea that the nervous system may learn to coordinate muscles and limbs by minimizing effort in producing natural movements, and may use approximate solutions based on muscle synergies. Understanding such neural mechanisms may allow us to predict the effects of neural injury and disease on motor function.
| Although optimality principles have been presented as a general framework for understanding motor control in animals and in humans [1], the ability of optimization to explain experimental data using high-dimensional musculoskeletal models remains largely unknown. Studies using optimization approaches have demonstrated an impressive ability to predict qualitative features of motor behaviors, such as the presence of low-dimensional muscle patterns [2], [3], and the presence of high levels of noise in some redundant degrees of freedom and low levels of noise in others [4]. Further, studies using approaches based on optimal feedback control have even predicted features such as countermovements [1], [5]. However, much of this evidence relies on biomechanical models that are abstract [1], that lack muscles [6], [7] or that have reduced degrees of freedom for computational efficiency [3], [8], [9], [10]. When complex musculoskeletal models are used to predict experimental data, the greatly increased complexity often precludes investigation of more than a single experimental condition [2], [11], which may be insufficient to discriminate different candidate control strategies or cost functions [12], [13]. Here, our goal was to test optimization as a predictive tool for understanding motor control by predicting detailed changes in experimentally-measured quantities across multiple biomechanical conditions.
The postural response to perturbations during standing balance is a motor paradigm in which consistent patterns of motor outputs are elicited across different biomechanical contexts, but the degree to which these patterns reflect neural control or biomechanical mechanisms is unknown. To maintain balance, the center of mass (CoM), a task-level variable, must be maintained above the base of support of the feet. Robust patterns of muscle activity referred to as the automatic postural response (APR) occur about 40 ms after horizontal translations of the support surface [14] and are consistently tuned to the direction of CoM motion across different perturbation types [15], suggesting that these long-latency responses reflect task-level control of the CoM by the nervous system. This robustness is surprising given that in a quadruped, the net force acting on the CoM can be produced by many combinations of individual limb forces. Further, each limb force can be produced by many patterns of muscle activity due to muscular redundancy [16], [17]. Active ground reaction forces during the postural response (∼100 ms latency) tend to be directed along a diagonal axis either towards or away from the CoM across perturbation directions [18]. The distribution of individual limb force direction and magnitude in the horizontal plane is consistently altered by varying the distance between the fore- and hind-feet, yet surprisingly, the directional tuning of muscle activity remains intact (Figure 1) [19], suggesting that the limb force variation may be due largely to differences in biomechanical context across postural configurations.
However, we demonstrated that biomechanical constraints alone are insufficient to determine the active production of limb forces during perturbations to standing balance. Measured postural forces are nearly ten times smaller than the absolute force production capability of a detailed musculoskeletal model of the isolated cat hindlimb in all directions [20]. The diagonal orientation of the forces is not predicted from anisotropies in the force-generating capability of the limb, which is greatest in the anterior-posterior direction. Further, changes in postural force directions across biomechanical contexts cannot be attributed to alterations in the force-generating capability of the limb, as peak force directions do not change appreciably across postural configurations [21]. Therefore, here we sought to improve our predictions of experimental measures through the addition of a model of a neural control mechanism that could achieve appropriate task-level forces and moments at the CoM while coordinating redundancy across both multiple muscles and across multiple limbs.
The optimal feedback control of CoM dynamics predicts the timecourse of activity in single muscles during balance control in both quadrupeds and bipeds [22], [23], [24]; however, it remains unknown whether task-level constraints at the CoM are sufficient to predict execution-level motor patterns across multiple muscles and limbs in a complex and redundant musculoskeletal system. Such redundancy has previously been resolved by minimizing neural control effort, assumed to be equivalent to the sum squared muscle activation or sum squared motor commands [2], [13], [25], [26]. Such optimizations have been applied to predict muscle tuning curves across conditions in relatively simple or quasi-static motor tasks [2], [26], or to deduce complex muscle activation patterns from detailed kinetic and kinematic measures [27]. Moreover, effort minimization is also sometimes treated as equivalent to energy minimization, which can predict aspects of gait in simple models of locomotion in humans and other animals [28], [29], [30]. However, predicting muscle coordination in detailed musculoskeletal models by minimizing quantities like effort remains challenging [11].
Further, it has been argued that low-dimensional muscle patterns emerge from optimization of the activation of individual muscles, without explicit neural constraints on muscle activation [1], [2]. While low-dimensional patterns in the form of muscle synergy groupings have been observed experimentally [19], [31], [32], [33], [34], [35], [36], studies using planar musculoskeletal models have noted similarities in motor behaviors predicted by optimally controlling individual muscles or muscle synergies [3], [37]. Such predictions have been based on relatively simple or abstract musculoskeletal models, and thus it is not clear whether such emergent low-dimensional patterns are competent to predict forces and muscle activation patterns in more behaviorally-relevant motor tasks. It has also been argued that muscle synergies may allow for near-optimal performance with simplified computations based on a reduced number of controlled variables [37], [38], but may increase control effort due to additional coactivation [39]. However, direct comparisons of the energetic cost associated with controlling individual muscles or muscle synergies in a 3D model of a natural behavior have not been performed.
Here, we sought to identify a task-level optimization framework that could predict execution-level limb forces and muscle tuning measured in an unrestrained balance task across different biomechanical contexts. We hypothesized that features of execution-level patterns of limb forces and muscle activity reflect the minimum-effort solution for achieving appropriate forces and moments at the CoM. We compared predictions using a static quadrupedal musculoskeletal model of the cat to data from experiments. Specifically, we predicted that limb forces would be directed along the diagonal for long stance distances, and more evenly distributed in direction at short stance distance. Further, we predicted that muscle activity would be low-dimensional, and that muscle tuning to perturbation direction would scale, but not shift as postural configuration varied. By varying cost functions and task-level variables we demonstrated that the predicted outputs depended on the optimization formulation, and not simply the biomechanical constraints. Finally we compared results from optimal control of individual muscles to those based on controlling experimentally-derived muscle synergies. Our work suggests that the neural control of this natural behavior can be well described by a cost function that minimizes effort expended in the muscles in order to achieve appropriate forces and moments to stabilize the CoM. Further, our results are consistent with the idea that the computation may be implemented in a hierarchical control framework that allows for approximately-optimal motor patterns with a reduced number of controlled variables.
To test the hypothesis that execution-level variables reflect optimal control of task-level variables, we predicted patterns of limb forces and muscle activity in response to multidirectional postural perturbations in cats based on achieving task-level mechanics while minimizing different formulations of control effort (Table 1). Using a detailed static quadrupedal musculoskeletal model of standing balance, we first identified patterns of muscle activity that produced forces and moments at the CoM necessary to maintain balance in response to postural perturbations in twelve different perturbation directions while minimizing neural control effort (model MMe). We considered multiple postural configurations with altered stance distance between the fore- and hind-feet. We compared identified muscle activation patterns and the resulting ground reaction forces to mean values measured experimentally during the initial response. In order to demonstrate that biomechanical constraints alone could not account for the identified solutions, we demonstrated that alternate cost functions and task goals produced qualitatively different results. We compared predictions from minimum effort control of CoM force and moment to predictions from minimizing an alternative cost function designed to be a better representation of the metabolic energy used in the muscles (model MMm). Additionally, we compared predictions of controlling an alternate task-level variable, the position of the center of pressure (CoP; model MPe). Finally, to investigate whether task-level control of the CoM could be accomplished with a small number of muscle synergies, rather than with individual muscles, we constrained the muscles in the model to activate in muscle synergies adapted from previously-observed experimental data (models SMe and SMc). We estimated and compared the energetic cost, the computational cost, match to experimental data, and the dimensionality of the muscle activation patterns predicted by controlling individual muscles or postural muscle synergies.
We parameterized the musculoskeletal model and assessed predicted limb forces and muscle activation patterns using previously-collected data of three cats during quiet standing and postural perturbations in multiple postural configurations [19]. The cats (bi, 2.7 kg; ru, 4.2 kg; ni, 3.5 kg) were trained to stand unrestrained with weight evenly distributed on four 8 cm-square force plates mounted on a moveable perturbation platform that could translate in any of 12 directions in the horizontal plane (Figure 2). Translations were 15 cm/s velocity and 5 cm amplitude. Data were collected in a self-selected postural configuration (preferred configuration), and in postural configurations in which the stance distance between the fore- and hind- force plates was altered. The following stance distances were examined in each of the animals: bi, 30 cm, 27 cm (preferred), 20 cm, and 13 cm; ru, 40 cm, 29 cm (preferred), 24 cm, and 18 cm; ni, 29 cm (preferred), 24 cm, and 18 cm. Stance width between the left and right force plates was 8 cm in all conditions.
We modeled muscle activity and limb forces associated with the initial period of the automatic postural response (APR) to perturbation, which can be studied as a quasi-static process. Multiple experimental and modeling studies have demonstrated that the forces during the initial portion of the APR can be attributed primarily to muscular forces [19], [40], [41]. During this period the acceleration- and velocity-dependent terms in the equations of motion are negligible so that the influence of dynamic terms on ground reaction forces is minimal [42] and the task can be approximated as quasi-static. This feature is due to the fact that there are distinct delays between the perturbation onset, the evoked muscular activity, and the subsequent active force. EMG activity due to the initial perturbation acceleration occur approximately 60 ms after the onset of the perturbation and only produces active forces at the ground after an additional 60 ms delay. Thus, there is no interaction between the perturbation acceleration and the active forces which occur during the constant-velocity, e.g. quasi-static phase of the perturbation [15]. Similarly, the acceleration of the body segments is largest while the acceleration of the platform is transmitted across all body segments [43], whereas after this period, the CoM has approximately constant horizontal-plane velocity (note the approximately constant slope of the CoM displacement during the active period indicated by the gray bar, Figure 2). Therefore, inertial forces associated with segment accelerations are not appreciable during the active response. Second, due to the relatively short latency of the active response compared to the overall motion, the posture of the animal has not changed appreciably from quiet standing at the onset of the active response. The posture of the animal affects gravitational forces, as well as torque generation via the muscle moment arm matrix. However, at the onset of the active force, the total displacement of the CoM is typically less than 1 cm and the effective tilt angle of the CoM is 1–2° [15]. Therefore the posture can be considered to be static, with no appreciable changes in gravitational forces or muscle moment arms. Therefore, our model assumes that all of the ground-reaction forces during the initial period of the APR are due to muscular activation, rather than dynamic terms.
We created the quadrupedal musculoskeletal model by modifying and assembling four instances of an existing static, 3-D musculoskeletal model of the cat right hindlimb [20], [21]. The hindlimb model relates 31-element muscle excitation vectors to the six-element force and moment system produced at the hindlimb endpoint:(1)where the vector is comprised of the model's seven kinematic degrees of freedom: three at the hip, and two each at the knee and ankle, designates the Moore-Penrose pseudoinverse of the transpose of the geometric system Jacobian (pinv.m), designates the moment-arm matrix, and and are diagonal matrices of maximum isometric forces and scaling factors based on muscle force-length properties [44]. Hindlimb model parameters are provided for each animal and experimental condition in Dataset S1. The muscles included in the hindlimb model and recorded in experimental data are summarized in Table 2. A closed-form expression for the Jacobian was identified with AutoLev software (Online Dynamics, Inc., Sunnyvale, CA, USA; currently being developed as MotionGenesis Kane) and implemented in Matlab (Mathworks, Natick, MA). The model of the left hindlimb was created by duplicating the right hindlimb model and reversing the sign of the lateral force component.
Prior analyses demonstrated that the hindlimb model is insensitive to the pseudoinverse operation, although the choice of pseudoinverse can be particularly important in robotics applications [45], [46]. Two previous studies demonstrated that the overall hindlimb force production capability is unchanged whether one degree of freedom (hip rotation) is locked, making the Jacobian 6×6 and exactly invertible [20], or whether the pseudoinverse is used [21], because the majority of the muscles in the model have hip rotation moment arms that are small in comparison to other degrees of freedom at the hip. Further, very similar endpoint force directions are produced by the muscles in the model in these two conditions. Across muscles, animals, and experimental conditions, the average difference in predicted endpoint force direction between the hip-locked and pseudoinverse conditions was only a few degrees (2.8±5.0°, dorsal plane; 4.4±11.3°, sagittal plane). These results are consistent with recent experimental results in which similar mappings between muscle forces and endpoint forces and torques were identified when mechanical degrees of freedom were locked or freed [46].
Because a detailed musculoskeletal model of the forelimb was unavailable, we approximated the forelimb by modifying the hindlimb model into a vertical strut that transformed muscle activation to vertical force. Although the forelimbs do not always contribute to horizontal-plane forces during the postural response [18], they contribute non-negligible vertical forces, of magnitudes several times larger than their horizontal force magnitudes. There also may be less potential for horizontal-plane forces to be produced by the extensor muscles in the cat forelimb because the morphology is more columnar than that of the hindlimb. Therefore, we approximated the forelimb as a transformation from muscle activity to vertical force by eliminating all rows of Equation 1 except for the row corresponding to vertical force.
The transformation from muscle activation to CoM force and moment in the quadrupedal musculoskeletal model was found using the forces from each limb and the approximate location of the CoM. Resultant CoM force was calculated as the sum of the individual limb forces. Resultant CoM moment was calculated as the sum of the vector cross products between the vectors from the CoM to the limb endpoints and the limb forces. Limb endpoint moments were assumed to make negligible contributions to the net moment at the CoM. The net force and moment at the CoM due to individual limb forces is thus:(2)Where designates the vector from the CoM to the endpoint of limb . The transformation from muscle activation to force and moment at the CoM was formulated as a 6×124 matrix equation for each postural configuration and animal relating muscle activation levels (31 muscles in each limb, for 124 total) to the 6D CoM force and moment.
We identified joint angles in the musculoskeletal model that best approximated the recorded kinematics of each cat during quiet standing in each postural configuration (Figure 3A). Positions of kinematic markers located on the platform and the left sides of the body were collected at 100 Hz during each trial for each cat. Locations of joint centers were estimated from marker positions by subtracting off joint radii, skin widths, and marker widths. The joint angles that minimized the squared error between the sagittal- and frontal-plane angles of the femur, shank, and foot in the model and in the background-period kinematics of each trial of each cat were identified using numerical optimization (fmincon.m) [20]. All residual segment angle errors were ≤10−4°. Joint angles were averaged across like trials. Muscle moment arm values and fiber lengths were determined with SIMM software (Musculographics, Inc., Santa Rosa, CA) and averaged across like trials.
We approximated the location of the CoM with respect to the feet in the musculoskeletal model separately for each cat in each postural configuration based on kinematic data and morphological parameters. For all conditions, the CoM was assumed to be located midway between the limb endpoints in both the anterior-posterior and medial-lateral directions. The height of the CoM above the plane of the feet was estimated from kinematic data and morphological parameters separately for each cat in each postural configuration. Across postural configurations, average CoM heights for each cat were (mean ± SD): bi, 12.6±0.4 cm; ru, 15.2±0.4 cm; ni, 12.7±0.8 cm.
To determine whether minimum-effort task-level control of the CoM could predict execution-level limb forces and muscle activity, we first identified patterns of muscle activity in the musculoskeletal model that produced forces and moments at the CoM similar to observed values while minimizing squared muscle activation (model MMe). Task-level constraints on CoM force and moment were based on average values from experimental data (Figure 3B). Average limb forces and CoM positions during the active period 120–200 ms after perturbation onset [19] were combined to estimate the average forces and moments at the CoM for each perturbation direction and postural configuration of each animal. Moments generated at the limb endpoints were assumed to make negligible contributions to the net CoM moment. Because values were similar across animals and postural configurations, a single set of average CoM forces and moments that was considered representative for all animals was then created and used as the optimization constraint: net horizontal-plane forces directed in the perturbation direction of 2.5 N magnitude, net vertical forces of 30 N, and net pitch-roll moments of 0.75 N-m magnitude directed perpendicular to the perturbation direction. CoM yaw moment was left unconstrained. Muscle activation patterns that satisfied task-level constraints could not be identified analytically without violating physiological bounds on muscle activation [44]. Therefore, optimizations were formulated as quadratic programming problems (quadprog.m) to identify muscle activation patterns that satisfied task-level constraints while minimizing total squared muscle activation:(3)Where designates a vector containing the activity levels of all muscles in the model (124). Additional constraints ensured that the activation levels of each muscle were in the interval (0,1) and that vertical ground reaction forces were ≥0. Separate optimizations were performed for each animal, postural configuration, and perturbation direction.
To investigate whether similar force predictions could arise from optimization criteria other than the minimum effort criterion used in model MMe, we next altered the cost function to better approximate metabolic energy consumption in the muscles, in terms of Joules/second, than minimizing Equation 3, but without the added complexity of Hill-type muscle models [47]. In single muscle fibers, metabolic energy usage (Joules/sec) is proportional to stress [48], equivalent to muscle activation in the model used here [44]. We assumed that the number of fibers in a muscle, and therefore its energy consumption, is proportional to its mass. Therefore, we performed additional optimizations with constraints and methods identical to the first model formulation, but minimizing total squared muscle activation weighted by muscle mass:(4)Where is a diagonal matrix of muscle masses. Masses for each muscle are included in Dataset S1. The majority of muscle masses (23/31 hindlimb muscles) were taken from the literature [49]. Because muscles for which no data were available were typically small, these masses were all set to a common low-midrange value.
To investigate whether the minimum-effort control of an alternate task-variable could predict similar limb forces, we tested a formulation similar to model MMe, except constrained to match displacements of the CoP in each perturbation direction, leaving the net force at the CoM unconstrained. Some studies of sagittal-plane balance in humans have suggested that the location of the CoP is the task-level variable controlled during balance [50]. Task-level constraints on CoP displacement were based on average values from experimental data (Figure 3C). The average displacement of the CoP at the midpoint of the active period in each perturbation direction for each postural configuration of each animal was calculated from the four vertical forces [15]. Similar to the first model formulation, a single set of corrections in CoP location (3.3 cm in magnitude and directed opposite the direction of the perturbation) was created and used as task-level constraints in the optimizations.
Next, to determine whether task-level control of the CoM could be accomplished with a small number of muscle synergies, rather than individual muscles, we constrained the muscles in each limb of the musculoskeletal model to activate in 5 muscle synergies based on muscle synergy force vectors previously observed in the same animals during the balance task [19]. The model [21], assumes that the activation of each muscle results from the additive combination of a few muscle synergies , recruited by scaling coefficients . The activation level of the muscles in the model is therefore:(5)where each element of represents the activation of the muscle by the muscle synergy, restricted to be within the interval (0,1), and the elements of scaling coefficients are restricted to be greater than zero. Five muscle synergies and related ground reaction force vectors were previously extracted from experimental data of each animal using nonnegative matrix factorization [19]. The muscle synergy patterns used in the model were subsequently derived by identifying patterns of muscle activation in the hindlimb model that could produce each ground reaction force vector while minimizing squared muscle activation (Equation 3) [21]. Identical muscle synergies were used in each limb and in all postural configurations. The constraints and solution method in this formulation were very similar to model MMe, with the exception that muscle synergy activation levels were identified rather than muscle activation levels. Synergy activation levels were constrained to be positive with respect to a level that created a background net vertical force. We considered two different cost functions in optimizations of muscle synergy control. Optimizations were performed that minimized muscle effort, (model SMe), as in Equation 3, but with the addition of muscle synergy constraints:(6)Further, to determine whether optimal solutions could be identified entirely in reduced-dimension space, optimizations were also performed (model SMc) that satisfied task-level constraints on CoM force and moment while minimizing sum squared muscle synergy activation:(7)
We calculated goodness-of-fit between predicted left hindlimb and right forelimb forces and experimental data from each animal across experimental conditions. Because vertical force (VF) magnitudes are several times larger than horizontal force (HF) magnitudes, they were analyzed separately. We compared predicted left hindlimb (LH) HF direction, LH HF magnitude, LH VF magnitude, and right forelimb (RF) VF magnitude with experimental data. R2 values for each force component were calculated across perturbation directions for each postural configuration for each animal and subjected to two-way ANOVAs (postural configuration×animal) evaluated with a significance level of α = 0.05 adjusted with a Bonferroni correction for multiple comparisons (α = 0.0125) to determine whether the predictive ability of each formulation depended on the experimental condition. Left hindlimb HF magnitudes in perturbation directions that loaded the hindlimb (0° through 90°) were also subjected to two-way ANOVA (postural configuration×animal) evaluated at α = 0.05 to determine whether magnitudes decreased as postural configuration was varied. R2 values predicted by the different model formulations were subjected to three-way ANOVAs (postural configuration×animal×formulation) and evaluated at the Bonferroni-corrected level of α = 0.0125.
We compared predicted muscle tuning curves to mean values from each animal and experimental condition. Mean values of EMG were calculated during the initial burst of muscle activity 60–140 ms after perturbation onset, and averaged across like trials. We compared the scaling and shifting in predicted tuning curve peak values across postural configurations to changes observed in data. Muscle tuning curves were normalized to maximum values observed in the preferred postural configuration of each cat. The peak magnitude and perturbation direction of each muscle tuning curve in each postural configuration of each animal was identified and expressed as a change from the preferred configuration value, either as a magnitude change, or as a direction change in degrees. In tuning curves with more than one peak, we tracked the peak value that was dominant in the preferred postural configuration. Tuning curve scaling was assessed by regressing peak values onto postural configuration (L,P,S,SS) and comparing the resulting regression coefficients for each cat and model. Tuning curve shifting was assessed by calculating the maximum change in peak direction across postural configurations. These values were then subjected to one-way ANOVA evaluated at α = 0.05 to determine whether shifts predicted by each model were comparable to observed values.
We assessed the dimensionality of muscle activation patterns predicted by models MMe, SMe, and SMc using a simple criterion based on principal components analysis (PCA). As we were primarily interested in comparing muscle activity pattern dimension predicted by controlling individual muscles (MMe) versus that predicted by controlling postural muscle synergies (SMe, SMc), we used a simple criterion that excludes components that contribute less variance than any individual variable in the original dataset [51], [52]. Vectors of predicted left hindlimb muscle activation were assembled into matrices arranged with perturbation directions along the rows and muscles along the columns. Separate matrices were assembled for each postural configuration and animal. The dimensionality of each matrix was then estimated as the number of eigenvalues of the data correlation matrix ≥1.0. Dimensionality estimates were pooled across animals and postural configurations and subjected to one-way ANOVA at a significance level of α = 0.05 to determine whether the formulations predicted similar muscle activity dimensionality. Dimensionality estimates from each model were compared to 5, the previously reported value [19]. Comparisons were performed with t-tests at a significance level of α = 0.05, adjusted with a Bonferroni correction for multiple comparisons to α = 0.0167.
We compared the total control effort required for controlling individual muscles (MMe) versus that required for controlling postural muscle synergies (SMe, SMc). The control effort required for the muscle activity predicted by each model formulation was calculated with Equation 3. Values were normalized to 100% of the value predicted by optimal muscle control in the preferred postural configuration of each cat. We then performed one-way ANOVA on the resulting values, at a significance level of α = 0.05, to determine whether the three formulations predicted similar sum-squared muscle activity. We estimated the computational cost predicted by the three formulations by measuring and comparing the time required for each formulation to identify muscle activity patterns in all perturbation directions in each experimental condition. Resulting values were subjected to one-way ANOVA at a significance level of α = 0.05, to determine whether the three formulations required similar computation time.
Task-level constraints on CoM force and moment or CoP location were satisfied by all of the models considered, but each predicted different patterns of muscle activity and limb forces, demonstrating the high level of redundancy of the quadrupedal musculoskeletal system. Experimentally-measured horizontal plane limb forces at preferred stance distance were predicted by task-level control of CoM forces and moments using either the minimum-effort or the minimum-energy cost functions (models MMe and MMm), whereas solutions predicted by control of CoP control (MPe) differed substantially. However, differences between forces and moments predicted by models MMe and MMm were revealed when limb forces were examined across postural configurations; although MMe solutions varied in magnitude across stance distances in a similar fashion to experimental measures, MMm solutions did not predict any qualitative differences in limb forces across stance distances. Limb forces similar to MMe predictions were found when a muscle synergy constraint was enforced (models SMe and SMc). In all three models that matched experimental limb forces across postural configurations (MMe, SMe, SMc), muscle tuning directions were found to be invariant across postural configurations, similar to experimental data, resulting in low-dimensional overall muscle activity patterns. However, using muscle synergies derived from experimental data (SMe, SMc) allowed better predictions of activity in flexors, some of which were not activated in the independent muscle coordination conditions (MMe). Finally, control effort increased by several times, but the time required for the quadratic programming search was decreased, when muscle synergies were controlled (SMe, SMc) rather than individual muscles (MMe).
Although we did not explicitly try to match experimentally-measured limb forces with the model, task-level control of CoM force and moment using either the minimum effort (model MMe) or minimum energy (MMm) cost functions nonetheless predicted horizontal plane forces directed towards and away from the CoM characteristic of the force constraint strategy described previously [18] in the preferred postural configuration (Figure 4A,B, 27 cm; Figure 5A, 27 cm). Across all perturbation directions, predicted left hindlimb HF directions were similar to data (MMe: mean R2 = 0.89±0.08, P<1e-3; MMm: 0.84±0.08, P<1e-3). In perturbation directions that loaded the left hindlimb (0° to 90°), predicted HF forces were directed towards the CoM, similar to data (data: mean direction 56±28°; MMe: 67±19°; MMm: 76±8°). In perturbation directions that unloaded the left hindlimb (180° to 270°), horizontal-plane forces were directed away from the CoM, again similar to data (data: mean direction 263±9°; MMe: 254±13°; MMm: 258±6°).
Left hindlimb HF magnitudes predicted by both cost functions varied as bimodal functions of perturbation direction similar to experimental data, particularly in loaded perturbation directions (MMe: mean R2 = 0.94±0.09; MMm: 0.91±0.05). Fits of left hindlimb HF magnitudes across all perturbation directions were reduced somewhat because of the small recorded force magnitudes in the unloaded perturbation directions (MMe: mean R2 = 0.77±0.29; MMm: 0.48±0.09). Maximal left hindlimb HF magnitudes were observed near 30° perturbations that loaded the hindlimb and minimal values for perturbations towards 120°, near the opposite diagonal axis. Average hindlimb HF magnitudes in perturbation directions where the left hindlimb was loaded (0° to 90°) were 1.2±0.4 N in data vs. 2.4±0.9 N and 3.3±0.3 N, in the MMe and MMm models, respectively. Absolute predicted HF magnitudes were larger than recorded values, which was necessary in order to account for the absent contributions of the forelimbs.
VF magnitudes predicted by both cost functions exhibited a realistic exchange between the forelimbs and hindlimbs as a function of the perturbation direction (R2>0.98). For perturbations diagonally to the right (near 30°), left hindlimb vertical forces were maximal (data: 11.4±3.4 N; MMe: 12.4±1.9 N; MMm: 12.3±1.8 N), whereas recorded right forelimb vertical forces were near minimal (data: 3.8±2.5 N; MMe: 2.1±2.3 N; MMm: 2.0±2.4 N). Both cost functions predicted complete unloading (0 N) of the left hindlimb and right forelimb in some cases, whereas the minimum vertical reaction forces observed in data were 1.0 N in the hindlimb and 0.6 N in the forelimb.
Differences between the predictions of models MMe and MMm became apparent when other postural configurations were considered. Variations in left hindlimb HF direction and magnitude were observed across stance distances similar to data [53] in model MMe, but not in model MMm. As stance distance was decreased, a wider range of HF directions was observed in MMe but not MMm (e.g., compare changes between 27 cm and 13 cm in Figure 4A versus Figure 5A). Similarly, HF magnitude from longest to shortest stance had a greater decreasing trend in MMe (−9±22%, P<0.25) than MMm (−4±8%; P<0.55) in unloaded directions (180° to 270°). However, neither reached the degree of HF magnitude change observed experimentally (−59±31%; P≪0.001) in unloaded directions. Over all directions, HF magnitude fits to data were significantly higher (P<0.001) in MMe compared to MMm (Table 3). Moreover, HF magnitude fits were similar across postural configurations in MMe, but were significantly decreased at shorter stance distances in MMm (P<0.0002).
Differences in forces across postural configurations were due to the fact that MMe favored recruitment of large muscles whereas MMm favored recruitment of small muscles. Large muscles that produce downward and backward endpoint forces relative to the limb axis were preferentially activated in MMe. When stance distance is shortened, the force rotates to have a more vertical orientation, thus reducing the component of force in the horizontal plane [19], [21]. In contrast, smaller muscles produce forces with relatively small elevations in the horizontal plane, so that horizontal plane force components are relatively constant as stance distance is shortened. Compared to MMe, model MMm reduced the activation of large antigravity muscles by several times (LG, mass 12.4 g, 1/3×; VL, 19.6 g, 1/4×) and increased the activation of small muscles by 5–1000 times (PSOAS, 4.0 g, 4×; SOL, 4.03 g, 20×; VI, 4.39 g, 5×; PT, 1.06 g, 1000×).
Unlike experimental data, model MPe predicted HF directions near the strongest axis of force production in the isolated hindlimb [20] in all perturbation directions and postural configurations (Figure 5B) to achieve task-level constraints on CoP location. Because CoP location is measured about the projection of the CoM on the ground, predicted CoM forces and moments deviated significantly from experimental measures (peak deviations: anterior force, 18.7±2.0 N; rightwards force, 1.9±0.4 N, roll-right moment, 0.2±0.1 N-m; pitch-up moment, 2.5±0.3 N-m). Although VF magnitudes predicted by model MPe were similar to data (R2>0.86), HF direction fits were poor (R2 = 0.36±0.15), and CoM-directed horizontal-plane forces were never observed. Instead, average left hindlimb HF directions were 90±3° and 98±3° for perturbation directions that loaded, and unloaded the left hindlimb respectively, near the direction of maximum force production of the hindlimb [20]. CoP control requires only modulation of VF magnitude across all four legs; large horizontal forces result from the fact that the minimum-effort muscle activation pattern to produce a vertical force component also has a very large horizontal component. These predictions were similar when the minimum energy cost function (Equation 3) was used (not shown).
Adding muscle synergy constraints (models SMe and SMc) resulted in limb forces that were similar overall to predictions of model MMe (Table 1); however, SMc additionally predicted a reduction in HF magnitude at shorter postural configurations that was comparable to the data (see arrows in Figure 6). As in MMe predictions, muscle synergy control models predicted characteristic HF directions towards (SMe, 83±100°; SMc, 68±80°) and away (SMe, 254±45°; SMc, 255±42°) from the CoM; however, visual inspection suggested that HF directions were more dispersed compared to MMe. Superior to MMe predictions, both muscle synergy control models predicted statistically-significant decreases in HF magnitudes in unloaded perturbation directions as stance distance decreased from preferred to shortest (SMc, −31±39%, P≪0.0001; SMe, −4±44%, P<0.04), although decreases were still less than those observed experimentally (−59±31%). VF magnitudes were predicted well in both the left hindlimb and right forelimb in SMe and SMc (R2 = 0.93±0.05), although MMe predictions remained superior (P<0.001). As in MMe predictions, both the left hindlimb and right forelimb completely unloaded in some cases for SMe and SMc (Figure 7). In some perturbation directions of the shortest postural configuration of cat bi (SMe, 5/132 total; SMc, 6/132) VF magnitude constraints were relaxed to allow CoM constraints to be achieved; these were excluded from further analysis.
All models that predicted realistic limb forces across postural configurations (MMe, SMe, SMc) predicted smooth cosine muscle tuning to perturbation direction similar to experimental data, particularly in morphologically simple extensors (Figure 8). Experimentally-observed tuning curves from left hindlimb extensors were typically cosine-shaped and centered around rightwards perturbations (0°) with approximate widths of 90°–120° at half-maximum (e.g., VM, GMED). Models MMe, SMe, and SMc made similar predictions for several extensors, including GMAX, GMED, VI, VM, VL, and SOL. Recruitment was not identical across models; for example, hip extensor BFA was recruited with similar tuning in SMe and SMc, but only in 1/3 cats in MMe. Some multifunctional extensors were more difficult to predict; for example, hip flexor/knee extensor RF was recruited in posterior/rightwards perturbations towards 330° experimentally, but predicted tuning curves (MMe, SMe, SMc) were centered about 0°. Ankle extensor/knee flexor MG was recruited with tuning curves centered near 180° by all models, unlike experimental results [15]; this tuning was similar to that observed in flexors, suggesting that the function at the knee might be dominating, with ankle extension being provided by extensor-tuned SOL. Ankle extensor/knee flexor LG was also recruited with tuning near 180° (1/3 cats, MMe) or with bimodal tuning to leftwards and rightwards perturbations (3/3 cats, SMe, SMc).
In some cases, the activation of flexor muscles was predicted by models SMe and SMc, but not by model MMe. Although some flexors were recruited with realistic cosine tuning about 180° in MMe, including PSOAS and SART (Figure 8), others were recruited in SMe and SMc but were never recruited in MMe. Ankle flexor TA was recruited with realistic cosine tuning to leftwards perturbations only in SMe, and only in cat bi. Some bifunctional muscles with flexor contributions were recruited in SMe and SMc but not in MMe. For example, hip extensor/knee flexor BFP was never recruited in MMe, but was recruited in 3/3 cats in SMe and SMc. Hip extensor/knee flexor GRAC was similar (2/3 cats, SMe, SMc; 0/3 cats, MMe), although predicted tuning curves were phase shifted somewhat from the anterior/leftwards tuning observed experimentally. Although hip extensor/knee flexor STEN was never recruited in MMe, it was recruited in SMe and SMc, but with either a bimodal (2/3) or extensor pattern (1/3).
Models MMe, SMe, and SMc all predicted muscle tuning curves that scaled in magnitude and shifted as stance distance was decreased comparable to experimental data (Figure 9). EMG peak magnitude increased as stance distance was shortened both in experimental data (regression slopes of 0.25, P<0.0001, bi; 0.10, ni; 0.10, ru) and in model predictions (MMe, 0.19±0.01; SMe, 0.26±0.27; SMc, 0.22±0.29; all P<0.022). Tuning curves predicted by all three models exhibited shifting with postural configuration that was not significantly different (P>0.05) from recorded values (average variation in peak tuning direction in data, 24±24°; MMe, 18±24°; SMe, 23±21°; SMc, 29±28°), although model SMc predicted increased tuning curve shifting compared to predictions of model MMe.
Models MMe, SMe, and SMc all predicted low dimensional muscle activity patterns, with muscle synergy control predicting lower dimensional EMG than individual muscle control. Patterns of left hindlimb muscle activity predicted in MMe were characterized by 4.3±0.5 principal components across cats and postural configurations, significantly higher (P<0.0001) muscle synergy control predictions (SMe: 3.2±0.6; SMc: 3.1±0.7). Dimensionality estimates from models MMe, SMe, and SMc were all significantly lower (P<0.0001) than 5, the number of muscle synergies previously identified in the balance task [19].
Models of muscle synergy control required more control effort, but less computation time during the quadratic programming search, than model MMe (Figure 10B). Using muscle synergy control reduced the computation time by a factor of 8 compared to MMe (P≪0.001) whereas control effort increased 2–4 times (P<0.0005). Post hoc analyses revealed a significant contrast between the control effort required for the MMe and SMc models (P<0.05).
To test whether model MMe might be predicting unrealistic endpoint moments, MMe optimizations in the preferred postural configuration of each animal were repeated with additional constraints such that the moments at each limb endpoint were limited to zero. This formulation predicted fits to experimentally-observed left hindlimb HF directions that were similar to those of the MMe model (P<0.83, paired t-test). Due to the additional constraints, 5/12 optimizations of cat Ni failed to converge and were excluded. Convergence failures occurred in the same conditions in ten repetitions of these optimizations.
Our results demonstrate how optimality principles can be used to understand how the nervous system may distribute effort across redundant muscles and limbs to achieve task-level goals during the automatic postural response, a natural motor behavior. Importantly, this work demonstrates that optimality principles can predict experimental data in the context of a detailed musculoskeletal model. We demonstrate that achieving task-level constraints on the forces and moments at the CoM while minimizing the control effort to the muscles can simultaneously resolve redundancy at the level of both muscles and limb forces during the initial portion of the automatic postural response. Moreover, by examining a rich repertoire of experimental conditions, we were able to distinguish amongst candidate task-level variables and effort cost functions, which often generated indistinguishable predictions in a single biomechanical context. Predictions were further improved by imposing constraints based on experimentally-derived muscle synergies and muscle synergy force vectors, demonstrating the feasibility of muscle synergies as physiological mechanisms for the implementation of near-optimal motor solutions, as well as suggesting additional costs and constraints that were not included in our original optimization framework. These results are consistent with the idea that the hierarchical, task-level neural control mechanisms previously identified in cortically-mediated tasks may also be relevant in understanding brainstem-mediated motor tasks.
Although prior studies demonstrated that temporal patterns of activation of individual muscles during balance could be predicted from task-level optimal control of CoM dynamics and control effort, they did not address the partitioning of control effort across redundant muscles or limbs. Temporal patterns of individual muscle activity during balance can be predicted from an optimal tradeoff between minimizing CoM excursion and control effort in both humans and cats [23], [24]. However, previous models of CoM control during balance have eliminated redundancy by examining single-plane movements, as well as by controlling the joints with torques [7], [54], [55], [56], [57], [58], [59] or single muscles [23], [24]. In contrast, we focused on predicting spatial patterns of activity at the initial timepoint of the CoM feedback response in order to understand the coordination of multiple muscles and limbs across multiple perturbation directions spanning the horizontal plane.
Here, we found that detailed patterns of muscle activity and limb forces across biomechanical contexts were predicted from interactions between a common optimization framework – achieving task-level constraints while minimizing effort – and the changing properties of the musculoskeletal system. Prior studies demonstrated that the properties of single-limb biomechanics [20], [21] were insufficient to predict the force directions observed across multiple postural configurations [14], [18], [53], leaving the role of biomechanics in determining this behavior unclear. These results suggest that control effort costs influence the way that the nervous system distributes effort across the redundant musculature when different combinations of muscles can realize the constraints of the task, and that the characteristic changes in forces observed during the balance task emerge as optimal patterns of distribution are applied in different biomechanical configurations. Moreover, constraints on net CoM mechanics allowed both muscle and limb force redundancy to be simultaneously resolved by minimizing control effort [2], [13], [26], eliminating the need to explicitly minimize limb force [25], [60].
Our results also demonstrate the feasibility of muscle synergies to produce approximately optimal motor patterns in the context of a detailed model in a realistic motor task. Multiple studies have demonstrated that muscle synergies might be a feasible and effective way for the nervous system to produce movement [61], [62], [63], and that the control of muscle synergies can closely approximate the optimal control of individual muscles, particularly in planar or idealized tasks [37], [64], [65]. We found that muscle synergy control was sufficient to achieve the task constraints, in some cases recreating the activation of flexors that was not well-predicted by minimizing the activation of individual muscles. However, in general, solutions from optimal muscle control and muscle synergy control were broadly similar, consistent with the results of other studies [2], [66]. For example, extensor muscle activity and the limb forces in perturbations for which the hindlimb was loaded were well-predicted whether the activity of individual muscles or of muscle synergies was optimized. Although our study does not resolve the debate over whether low-dimensional muscle activation patterns reflect optimal patterns of individual muscle control or explicit muscle synergy constraints, these results demonstrate the feasibility of muscle synergies for the implementation of near-optimal motor solutions in a realistic motor task.
Taken together, the results of this and previous studies are consistent with the idea that the temporal and spatial patterning of muscle activity during the automatic postural response can be well-described by a hierarchical optimal control framework. Hierarchical optimal control is based on the idea that higher levels of the nervous system operate on increasingly abstract variables, such as CoM kinematics, while relying on lower-level controllers to locally control high-dimensional musculoskeletal dynamics [67], [68]. We hypothesize that the high-level representation is critical because multiple studies have demonstrated that lower-level kinematic variables such as joint angles are insufficient to predict the activation of individual muscles during balance control, whereas CoM kinematics robustly predicts which muscles will be activated [15], [54], [69], [70], [71], [72]. Such a hierarchical structure may be required in neural control structures due to neural conduction and computation delays. One idea proposed for the low-level control architectures is that they might implement local feedback control to linearize the nonlinear, fast dynamics of the musculoskeletal system, or implement other regulatory functions [68], [73]. Our concept of a muscle synergy is proposed as a transformation between high-level task goals and low-level dynamics, that may be parameterized to optimally actuate musculoskeletal mechanics [64] or to provide stability [74], but not necessarily to function as a controller per se. We speculate that CoM feedback may be used to recruit muscle synergies, and in support of this, a recent study in human balance control demonstrated that CoM kinematics are sufficient to describe the temporal recruitment of postural muscle synergies throughout complex perturbations [75]. Despite the various differences, the similarity between solutions arising from optimization of the activity of individual muscles and optimization of the activity of muscle synergies are consistent with the idea that muscle synergies may reflect mid- or low-level control structures within a general hierarchical optimal control scheme for movement.
While control of the CoP was sufficient to explain the results of previous studies that considered a limited range of biomechanical conditions, we were able to compare CoP and CoM as task-level variables by examining their ability to predict individual limb forces across multiple directions of perturbation. Both the CoM and CoP have been proposed as controlled variables for balance control [50], [70], but control of CoP involves fewer constraints and is based on the control of vertical and not horizontal limb forces. These candidate control variables have typically been investigated in models of only a single plane of movement [7], [23], [24], [54], [55], [56], [57], [58], [59], where they may make indistinguishable predictions. Here, forces predicted by the two candidate task-variables were similar for the direction of primary limb loading in which lateral forces were small. Predictions of sagittal-plane limb forces were also similar across both models (MMe vs. MPe) in the directions across all directions in which the limb was loaded. Given the anisotropic force generation characteristics of the hindlimb [20], it seemed plausible that the control of vertical forces could be sufficient to determine shear forces as well. However, the models produced qualitatively different horizontal plane forces, suggesting that additional constraints on CoM moment and force were necessary to predict the observed force patterns in a quadruped. It is possible that CoM and CoP control are indistinguishable in sagittal plane balance control in humans where force generation is primarily in the vertical direction [76], [77]. However, the predictions of CoP control are likely to break down when significant horizontal place forces are required such as in our quadrupedal model, or in medial-lateral human balance control. Further, CoP control in human and robot walking has been limited to quasi-static conditions [57], [78], [79], whereas more dynamic conditions suggest that angular momentum about the CoM due to CoM moments is an important control variable [80], [81], [82], [83], [84], [85]. Importantly, these results demonstrate that the observed muscle activity patterns and forces could result from an optimization framework in which task-level goals are specified, independent of individual limb forces.
We noted that different cost functions produced qualitatively different patterns of limb forces, demonstrating that the experimentally measured patterns are not simply due to musculoskeletal constraints, but indeed depend upon the nature of the optimization framework. Prior studies have found that multiple cost functions could produce similar results [13], [86], suggesting that solutions may be qualitatively determined by biomechanical constraints, independent of any optimization framework or control policy. In contrast, our study and other recent studies demonstrate that some cost functions can be eliminated based on their robustness across a wider range of experimental conditions [12], [25]. Here, minimization of muscle effort (MMe) versus energy (MMm) predicted similar horizontal plane forces in the preferred postural configuration, but not in short or long stance configurations. In order to more precisely determine a physiological cost function inverse optimization approaches could be used [25], [87], [88]. However, it is unlikely that composite cost functions based on weightings between MMe and MMm [25], [89] would improve fits to recorded muscle activity (e.g. absent flexors, SOL recruited rather than MG), as both cost functions strongly penalize muscle coactivation. Neither are these differences likely to be resolved using alternative cost functions such as minimization of signal dependent noise, which predicts muscle activity patterns similar to minimization of control effort [90].
To further investigate either the task-level variable or the cost function would require implementation of task-level control within a dynamic musculoskeletal model. Although balance control is a dynamic task, we were able to use a static musculoskeletal model to examine the force-sharing problem at a specific instant in time during the postural response that is most amenable to description by a quasi-static model (see Methods). Here we sought only to reproduce the net CoM forces and moments observed in the initial postural response, which in turn can be predicted by an optimal feedback control model in a low-dimensional biomechanical model [22], [23], [24]. Integrating an optimal controller with a realistic musculoskeletal model would allow us to test various optimal control models for dynamic balance control, which might implicate criteria relevant to the balance task beyond the control cost formulations presented here. Specifically, considering the longer time constants required to deactivate versus activate muscle [91] would likely improve model predictions by encouraging activation of the flexors. Similarly, rewarding recruitment of muscles with fast fiber types would likely encourage the ankle extensor function of MG (primarily fast muscle fibers), over that of SOL (primarily slow muscle fibers; [92]). Other criteria such as those related to mechanical stability might also be used to explain the absent coactivation [16]. For example, arm impedance is increased in unstable environments, likely requiring additional coactivation [93]. It is possible that these costs could be incorporated within an optimal control formulation penalizing response time in a tradeoff with costs such as control effort, as optimal control models without fixed terminal time have recently been developed for motor tasks [94], [95], [96]. A dynamic model would also allow for further refinement of the task variable. Although we were able to differentiate between CoM and CoP, the current model cannot differentiate between CoM and some other candidate task-level variables – for example, translations of the CoM along the anterior-posterior axis – since a static model ignores inertial contributions such that an equivalent moment can be computed about any point.
We consider it unlikely that adding additional detail to either the hindlimb or the forelimb models would appreciably influence the forces predicted here. Based on the high level of similarity in the force production capability between the static hindlimb model used here and previous dynamic models, it is unlikely that including a linearized dynamic model with the mass matrix would appreciably influence the results. Previous linearized and fully dynamic versions of the hindlimb model that include the mass matrix have demonstrated nearly identical force production capability to the static model used here [20], with force production capability biased along the anterior-posterior axis [74], [97]. Based on earlier versions of the present model and experimental results, it is also unlikely that including a detailed forelimb model would appreciably influence the predicted forces. A previous model that included forelimbs as hindlimbs with reflected anterior-posterior force production capability did not fundamentally change the forces predicted by model MMe [98] in the preferred and long postures. However, as the stance distance shortens, the geometry of the forelimbs in a real animal becomes increasingly like that of a vertical strut, whereas the hindlimbs remain flexed, breaking the symmetry of the forces between the fore- and hind-limbs Although the fore-hind force asymmetry in the shorter postures was not very pronounced in these particular animals modeled here, in some cases the forelimb forces are not elongated at all [14], [18], suggesting that the forelimbs can be very well approximated as vertical struts in these conditions.
Significant electrophysiological evidence exists for the neuroanatomical substrates required for the hierarchical, task-level neural control mechanisms investigated by this and other studies. While we and others have demonstrated that muscle activity and movements can be described by mathematical tools like optimization, these techniques do not explain how such relationships and computations are achieved within the nervous system [99]. Importantly, electrophysiological evidence from both cortically-mediated as well as brainstem-mediated motor tasks exists to support the idea that the hierarchical, task-level control frameworks suggested here may describe aspects of the organization of the neural substrates for motor control. For example, electrophysiological evidence demonstrates that task-level variables such as the direction of the limb endpoint are represented in motor cortex during reaching [100], [101]. Although lesion studies demonstrate that the balance task considered here does not require the cortices [102], [103], similar task-level representations are found in brainstem, where neurons in the pontomedullary reticular formation respond equivalently to perturbations of different limbs [104], [105]. Electrophysiological evidence also demonstrates that increasingly abstract representations of the motor periphery are assembled in increasingly higher levels of the nervous system. For example, higher-level representations of limb length and orientation, rather than individual joint angles, are encoded in the dorsal root ganglia and dorsal spinocerebellar tract [106], [107]. Muscle synergies may describe how task-level representations are mapped to execution-level activity of motoneurons, via the divergent projections to multiple muscles that have been identified at various levels of the nervous system [108], [109], [110], [111]. For example, both cortical and brainstem neurons project to multiple motoneurons, or to spinal interneurons [112] whose activity has been shown to reflect the patterns of muscle synergies rather than individual muscles [113].
These results support the hypothesis that muscle synergies may be important physiological mechanisms for the implementation of near-optimal motor solutions with a reduced number of controlled variables. The original concept of the muscle synergy hypothesis was that it would offer computational “simplification” due to the large numbers of independent variables that must be simultaneously controlled by the nervous system [114]. In our study, using muscle synergies significantly decreased the search time the optimization algorithm required to identify a motor solution, similar to a previous report [64]. This search time decrease illustrates the possible benefits of a reduced dimension solution space during gradient-based searches, although the computational mechanisms in the nervous system are certainly different than a computer. Stochastic search approaches, for example, might realize less benefit from reducing the dimension of the solution space. Moreover, the results do not imply that the nervous system is re-optimizing the cost function de novo every time the motor task is presented [25], but instead are consistent with the idea that optimal motor solutions could be refined over the course of motor learning and adaptation. Such refined solutions could be encoded within the nervous system in sparse representations that use small number of neurons at any given time. Sparse representations have been hypothesized to increased storage capacity in associative memories and increased energy efficiency [115] as well as accelerate motor learning. For example, a neural-network model demonstrated accelerated motor learning with decreases in the number of independent neural commands [38]. However, this interpretation may be somewhat controversial, as other evidence demonstrates that sparse motor representations based on muscle synergies may slow the learning of motor tasks for which the library of available muscle synergies is inappropriate [116]. We speculate that muscle synergies implement a transformation from task-level goals to muscle activation patterns that is computationally similar to a lookup table that is assembled over motor learning, the structure of which likely reflects the statistics of the behavioral repertoire as well as the motor system [117]. Similar to the arguments advanced for sparse coding of sensory inputs, we speculate that muscle synergies are reinforced over the course of motor learning through biologically-plausible local learning rules (e.g. “cells that fire together wire together”). Through such learning rules, simple model neurons can learn the principal components of their synaptic input weightings [118]. We speculate that groups of muscles would be reinforced, rather than individual muscles, because the function of individual muscles (in this case, the output force) may vary depending on the activity of the other muscles in the limb [119].
We speculate that the increased control effort required when using experimentally-derived muscle synergies versus individual muscles may be physiologically reasonable, particularly if considerations beyond energy efficiency are important in balance control. Whereas prior work demonstrated that similar efficiency could be found by controlling individual muscles or muscle synergies developed from optimality criteria [64], [65], we show that controlling experimentally-derived muscle synergies requires additional control effort. Although minimizing energetic cost may be critical in some contexts, particularly in ongoing movement tasks like locomotion over evolutionary timescales [28], [29], [30], we speculate that in discrete tasks like the balance responses presented here strictly effort-minimal solutions may not be necessary. For example, in discrete arm posture tasks, subjects can be cued to maintain high levels of coactivation out of habit even at levels of muscle activation that are considerable proportions of maximal voluntary contraction [120]. The forces observed during balance are well within the boundaries of the absolute musculoskeletal capabilities [21], and the magnitudes of the individual muscle activations predicted by model MMe were moderate, as proportions of MVC (notice that the scale maxima in Figure 8 vary between 0.002 and 0.4). Thus the additional effort cost predicted by muscle synergy control may be physiologically plausible. The fact that experimentally measured co-activation is absent in the MMe model predictions further suggests that the physiological state does not necessarily correspond to the minimum effort solution. We speculate that muscle synergies may be organized to implicitly account for criteria related to the dynamic response described above (e.g. fiber type, etc.). Particularly in balance control, using more than the absolute minimum amount of muscle activation required to achieve stability may be advantageous.
|
10.1371/journal.pntd.0003559 | Toxocariasis Diagnosed in International Travelers at the Institute of Tropical Medicine, Antwerp, Belgium, from 2000 to 2013 | Although infection with Toxocara canis or T. catis (commonly referred as toxocariasis) appears to be highly prevalent in (sub)tropical countries, information on its frequency and presentation in returning travelers and migrants is scarce. In this study, we reviewed all cases of asymptomatic and symptomatic toxocariasis diagnosed during post-travel consultations at the reference travel clinic of the Institute of Tropical Medicine, Antwerp, Belgium. Toxocariasis was considered as highly probable if serum Toxocara-antibodies were detected in combination with symptoms of visceral larva migrans if present, elevated eosinophil count in blood or other relevant fluid and reasonable exclusion of alternative diagnosis, or definitive in case of documented seroconversion. From 2000 to 2013, 190 travelers showed Toxocara-antibodies, of a total of 3436 for whom the test was requested (5.5%). Toxocariasis was diagnosed in 28 cases (23 symptomatic and 5 asymptomatic) including 21 highly probable and 7 definitive. All but one patients were adults. Africa and Asia were the place of acquisition for 10 and 9 cases, respectively. Twelve patients (43%) were short-term travelers (< 1 month). Symptoms, when present, developed during travel or within 8 weeks maximum after return, and included abdominal complaints (11/23 symptomatic patients, 48%), respiratory symptoms and skin abnormalities (10 each, 43%) and fever (9, 39%), often in combination. Two patients were diagnosed with transverse myelitis. At presentation, the median blood eosinophil count was 1720/μL [range: 510–14160] in the 21 symptomatic cases without neurological complication and 2080/μL [range: 1100–2970] in the 5 asymptomatic individuals. All patients recovered either spontaneously or with an anti-helminthic treatment (mostly a 5-day course of albendazole), except both neurological cases who kept sequelae despite repeated treatments and prolonged corticotherapy. Toxocariasis has to be considered in travelers returning from a (sub)tropical stay with varying clinical manifestations or eosinophilia. Prognosis appears favorable with adequate treatment except in case of neurological involvement.
| Toxocariasis is a zoonosis of worldwide distribution caused by dog (Toxocara canis) or cat (T. catis) roundworm that can be fully asymptomatic or may cause significant disease such as a the systemic syndrome called visceral larva migrans as well as neurological or eye manifestations. Toxocariasis prevails in tropical areas, but information about this disease in travelers and migrants is scarce. In this study, we describe in detail a case series of 28 international travelers, mostly adults, diagnosed with toxocariasis from 2000 to 2013 at the reference travel clinic of the Institute of Tropical Medicine of Antwerp, Belgium. We found this infection in all types of travelers returning from any part of the world. Clinical symptoms, when present, varied widely and an increase of the blood eosinophil count was almost always present. Morbidity was substantial and 2 patients had severe neurological complications. Diagnosis was difficult in travelers because the illness often resembled other tropical infections. Recovery was, however, complete, either spontaneously or with anti-parasitic drugs, except in both cases with neurological involvement. Toxocariasis is one of the numerous parasitic infections to consider in travelers returning from the tropics with any type of symptoms or with an increased blood eosinophil count.
| Toxocariasis, caused by intestinal roundworm of dogs (Toxocara canis) or cats (Toxocara catis), is a zoonotic infection with a worldwide distribution [1,2]. Humans can get infected by ingestion of embryonated eggs present on the soil, plants or soil-dwelling invertebrates contaminated by dog or cat feces, and less frequently by ingestion of encapsulated larvae from undercooked paratenic hosts such as chickens, cattle and sheep. Infection is followed by the migration of third-stage larvae through the tissues which is usually asymptomatic but may be associated with a variety of non-specific clinical features [2,3]. Presentation may be acute or sub-acute, with systemic, abdominal or respiratory manifestations, classically described as the syndrome of visceral larva migrans (VLM) sometimes associated with dermatological symptoms as well. Occasionally, the course is complicated by the involvement of the central nervous system (neurological toxocariasis) or the eyes (ocular toxocariasis). Like in other helminthic infections with larval migration, blood eosinophilia is common and/or numbers of eosinophils may be increased in other relevant tissues or fluids; other laboratory abnormalities such as elevated inflammatory parameters or liver enzymes disturbances are often mild and inconstant. Of note, a distinct clinical presentation called “covert” (in children) or “common” (in adults) toxocariasis has been described more recently, with more subtle and chronic symptoms such as cough, abdominal pain or pruritus, associated with mild eosinophilia and Toxocara spp. seropositivity [4–6]. It is however assumed that most cases of human toxocariasis go unrecognized.
A definitive diagnosis of human toxocariasis can only be made by histological examination of infected tissue, demonstrating Toxocara spp. larvae within eosinophilic granulomatous lesions. Since sampling of appropriate tissue is rarely justified on clinical grounds, diagnosis of human toxocariasis relies in the daily practice on a constellation of suggestive symptoms (if present), combined with laboratory abnormalities (blood eosinophilia) and the detection of circulating immunoglobulin G (IgG) antibodies to Toxocara excretory–secretory (TES) antigens [7]. In Western countries with low prevalence of helminthic infections, the association of clinical symptoms (although frequently poorly specific) with eosinophilia usually triggers clinicians aware of the disease to perform the specific serological investigations [1]. In the tropics however, where polyparasitism is highly prevalent, the low specificity of eosinophilia, the lack of diagnostic facilities and the cross-reactivity with other helminthic infections often preclude the etiologic diagnosis [2].
While larvae may remain viable for years in the tissues, human toxocariasis is usually a self-limiting disease, although its course can sometimes be invalidating and prolonged (up to several weeks). Treatment firstly aims at controlling the inflammatory reaction when needed, and despite the limited knowledge about its clinical benefit, anti-helminthic therapy is often associated. In such cases, preference usually goes to albendazole for its parasitological efficacy and good clinical tolerance [8].
The global burden of human toxocariasis is poorly quantified [9]. Serological surveys demonstrate that the infection is more frequent in tropical settings and in rural areas, with population-based seroprevalence ranging from 2.5% in urban Europe up to 85% in rural tropics [9–14]. Visceral larva migrans has been reported in individuals born in tropical countries having migrated to Europe [15]. However, although many susceptible European travelers are assumed to be at increased risk of exposure to Toxocara spp. during a stay in highly endemic developing countries, frequency and presentation of toxocariasis are largely unknown in this population. A recent large multicenter GeoSentinel study has reported 16 cases diagnosed as visceral larva migrans among 42,173 travelers evaluated from 2007 to 2011 (38 cases per 100,000 travelers) but there was no standard diagnostic approach in the 53 participating travel clinics and no clinical description [16]. In the present study, we aimed to assess the frequency, presentation and outcome of toxocariasis diagnosed among travelers and migrants presenting at the travel clinic of the Institute of Tropical Medicine of Antwerp, Belgium.
For this retrospective study, a query was first undertaken in the database of the Central Laboratory of Clinical Biology of the Institute of Tropical Medicine, Antwerp (ITMA), to retrieve all results of Toxocara serology requested in patients having attended the travel clinic of the ITMA from 2000 to 2013. The ITMA is the national reference center for tropical medicine in Belgium, with on average about 6500 consultations a year for post-travel care. The medical records of all travelers and migrants found with a positive Toxocara serology during this period were then reviewed. Relevant clinical and laboratory data were extracted, de-identified and entered in a Microsoft Access 2010 database. Variables included: demographic data, month and year of first Toxocara positive serology, most recent travel destination, time of symptom onset after travel, duration of symptoms before consultation, clinical presentation, result of the chest X-rays if requested, absolute blood eosinophil count (and percentage of white blood cell count), Toxocara antibody optical density result, results of parasitological and serological tests prescribed by the physician targeting other helminths according to epidemiological relevance (Ascaris spp., Echinococcus granulosus, Fasciola spp., Filaria spp., Schistosoma spp., Taenia solium, Strongyloides stercoralis, Trichinella spp. and Anisakis simplex), administered treatment(s), clinical and laboratory evolution and outcome.
Next, we categorized all reviewed Toxocara-seropositive cases in four groups according to the clinical reasons for requesting Toxocara serology: 1) presence of symptoms compatible with VLM combined with eosinophilia; 2) asymptomatic eosinophilia; 3) symptoms compatible with VLM without eosinophilia; and 4) possible exposure/no clear reason (Fig. 1).
For this clinical study, strict case definitions of symptomatic (Toxocara-associated visceral larva migrans) and asymptomatic toxocariasis were used. Diagnosis of Toxocara-associated VLM was considered as highly probable when the following criteria were all fulfilled: (1) positive Toxocara serology (entry criteria) AND (2) presence of any systemic symptom compatible with toxocariasis (including fever, respiratory signs such as wheezing, dry cough, dyspnea or an infiltrate on the chest X-ray, abdominal symptoms as abdominal pain, vomiting, diarrhea or hepatomegaly, neurological signs such as focal deficit or encephalopathy, and ocular signs such visual disturbances), with or without dermatological symptoms such as pruritus, urticarial rash or angio-edema AND (3) blood eosinophilia (defined as an absolute blood eosinophil count above 500/μL, or > 7% of the white blood cell count at first presentation [17,18]), or presence of eosinophils in another relevant fluid/tissue) AND (4) reasonable exclusion of alternative diagnosis. Diagnosis of VLM was considered definitive when all 4 criteria were present together with unequivocal Toxocara seroconversion documented on paired serum samples. Diagnosis of asymptomatic toxocariasis relied on the presence of blood eosinophilia at presentation in asymptomatic Toxocara-seropositive individuals and no evidence of other infection likely to explain the eosinophilia.
To comply with the case definition, we therefore did not further analyze groups 3 and 4 (since there was no eosinophilia), and then carefully assessed for all cases of groups 1 and 2 whether an alternative diagnosis or a co-infection were possible. In accordance with the case definitions, we finally excluded from this study all patients with parasitological or serological evidence of another infection or clinically suspect of alternative diagnosis (such as allergy, scabies,…).
From 2000 to 2009, the serological diagnosis of toxocariasis was performed with a commercial anti-TES IgG enzyme-linked immunosorbent assay (ELISA) (Toxocara canis, Bordier Affinity Products SA, Crissier, Switzerland) according to the instructions of the manufacturer. The assay’s threshold for positivity was set as “weak positive” (from 2000 to 2009) and at 1.0 (measured by optical density, since 2010 onwards) for use in clinical settings [13]. In this context, the assay has been reported to provide a sensitivity and a specificity of 91% and 86% respectively, the latter being evaluated in a population of patients with protozoan and helminthic infections [19].
When toxocariasis is strongly suspected or confirmed, the standard treatment regimen at ITMA is albendazole 2 x 400 mg (in adults) for 5 days, according to international guidelines. Corticosteroids may be added at the physician’s discretion according to severity of symptoms. Diethylcarbamazine (DEC) is restricted as second line therapy for refractory cases [20,21].
Analysis was performed with the SPSS program version 20.0 (SPSS Inc., Chicago, IL, USA). Differences were compared using Student’s t-test or Mann-Whitney U-test, when appropriate, for continuous outcome and chi-square and Fisher’s exact tests for categorical outcomes. P values of less than 0.05 were considered to indicate statistical significance.
This was a retrospective analysis of data collected during routine clinical care over a 13-year period. Ethical clearance was obtained from the institutional review board at ITMA. Laboratory queries were obtained in an anonymous way. Clinical data were then retrieved through an encoded link and de-identified for analysis according to the Belgian legislation. No written informed consent was obtained from individual participants because an opt-out strategy is in place at ITMA covering surveillance use of clinical and laboratory data.
From January 2000 to August 2013, 3436 Toxocara serological tests were ordered for diagnostic purpose in post-travel care by the ITMA physicians. Of these tests, 190 (5.5%) had positive anti-TES IgG (Fig. 1). Of 187 patients with complete clinical data, 44 had VLM symptoms and eosinophilia (group 1), 35 were found with asymptomatic eosinophilia (group 2) and 54 had symptoms compatible with VLM but no eosinophilia (group 3, excluded from the study); for the remaining 54 individuals with no symptoms and no eosinophilia, the reason for Toxocara testing was considered as unclear (group 4, also excluded from further analysis).
In the groups 1 and 2, after exclusion of the cases with alternative diagnosis or possible co-infection, a diagnosis of human toxocariasis was retained in 28 patients, including Toxocara-associated VLM in 23 and asymptomatic toxocariasis in 5 (Fig. 1). Diagnosis of VLM was considered as definitive in 7 cases for whom seroconversion was observed. Histological examination was not performed in any case. Of note alternative diagnoses in the excluded cases of groups 1 and 2 (n = 51) were mainly strongyloidiasis (n = 15), allergic reaction (n = 7), filarial infection (n = 7), Ancylostoma/Ascaris infection (n = 7) and schistosomiasis (n = 6). Suspicion of co-infection was also frequent.
The clinical presentation of the 28 cases is detailed in Table 1. Cases were evenly distributed throughout the study period with no cluster phenomenon. All patients were adult travelers born or residing in Europe, except one Ethiopian child evaluated after adoption and one Lebanese adult living in the Democratic Republic of the Congo. Mean age was 46 years (range: 4–68 years) with a male/female ratio of 0.87. Regions of most recent travels (and presumed exposure) were North and sub-Saharan Africa in 10 patients, Southern and Southeast Asia in 9 and Southern Europe (including Turkey) in 6; for 3 patients (numbers 18, 19 and 26), the continent of acquisition could not be traced with certainty because of multiple travel destinations within a short timeframe. Duration of travel was less than 1 month in 12 (43%) patients and more than 3 months in 10 (36%).
For the 23 cases presenting with VLM, symptoms started during the stay abroad in 11 (48%) or developed within 3 weeks in average (range: 0–8 weeks) after return. For one patient (number 11), the dates of recent travel were not clearly reported. Clinical manifestations included abdominal complaints in 11 (48%) patients, respiratory symptoms and skin abnormalities in 10 (43%) each and fever in 9 (39%); in most cases, symptoms were combined or developed sequentially. Radiological pneumonia was found in 5 patients and one of them (number 22) had to be admitted elsewhere because of the severity of respiratory symptoms. Two patients (numbers 11 and 16) presented with progressive neurological features of transverse myelitis but no other symptoms. In both cases, the diagnosis was made by the demonstration of an increased eosinophil count and of anti-TES IgG in the cerebrospinal fluid, in the absence of another etiology (Table 1).
For the 21 patients without neurological complications, median duration of symptoms before first documented medical evaluation at ITMA or elsewhere was 3 weeks (range 5 days- 4 months), while both patients with neurological toxocariasis were diagnosed several months (2,5 and 7 respectively) after symptom onset.
At first presentation, the median blood eosinophil count was 1720/μL (range: 510–14160) in the 21 VLM cases without neurological complication and 2080/μL (range: 1100–2970) in the 5 asymptomatic cases. Of note, the median eosinophil count was significantly lower in the VLM patients presenting more than 4 weeks after symptom onset than in those consulting earlier (700/μL versus 2340/μL, p<0.001). The blood eosinophil count was within the reference ranges for both patients with neurological toxocariasis.
Of the 5 patients with asymptomatic toxocariasis, 3 received the 5-day course of albendazole upon diagnosis, and the eosinophil count normalized uneventfully; the other 2 patients were empirically treated before the complete results were available: one with ivermectin (who required albendazole secondarily because eosinophilia persisted) and the second with ivermectin and praziquantel (with resolution of the eosinophilia).
Of the 23 symptomatic cases, 15 received a standard course of albendazole within one or two weeks after presentation (when laboratory results were available): 11 clinically recovered and eosinophil counts normalized (of note 4 had already clinically improved before the start of therapy); one was lost to follow-up; another had to switch to DEC because of immediate anosmia attributed to albendazole; both neurological cases were given steroids concomitantly with albendazole, but since the improvement was slow DEC was also administered (with resolution of the radiological abnormalities and moderate clinical recovery). Four patients were initially treated empirically with praziquantel (n = 1), or ivermectine (n = 3) but 3 of them required albendazole secondarily in the absence of substantial clinical/laboratory improvement. Finally the remaining 4 patients preferred at first not to be treated; 3 of them recovered spontaneously but the 4th patient needed a course of albendazole a few weeks later and persisting symptoms eventually subsided. Of note only one non-neurological patient (number 22) required corticosteroids during hospitalization to control the severe respiratory symptoms.
Although toxocariasis is highly prevalent in most tropical areas, this condition has been hardly studied in returning travelers and migrants. We describe here a case series of 28 patients diagnosed with toxocariasis acquired from all over the world. Clinical presentation was extremely varied and resembled that of many other endemic helminth infections. Morbidity was important and complications sometimes serious.
This study has many limitations. It was indeed a retrospective single-center study conducted in a reference travel clinic, meaning that collection of data was not systematic and that findings may not be generalizable to all clinical settings. For instance, some cases with longer incubation period may have been directly attended in the primary care setting given the fact that the link between symptoms and travel had become less obvious. Also our observations are not transposable as such to the features observed in autochthonous toxocariasis sporadically seen in Belgium or elsewhere, mainly in children [22]. Other limitations could be related to the restrictive case definition that may have missed several true cases without eosinophilia (as sometimes observed in milder cases of “covert/common” toxocariasis) or with negative Toxocara serology (during the serological window period or because sensitivity does not reach 100%, in particular in low burden infection such as ocular toxocariasis [2]). In the same line, some true toxocariasis cases may have been disregarded just because serological tests against other helminths were also positive, either by cross-reactivity or as part of infection with multiple parasites. Conversely, false positive result was also possible if anti-Toxocara seropositivity was just reflecting remote exposure or cross-reaction while another helminthic infection was missed during the workup. The commercial test we used is widely considered as an adequate screening tool for clinical practice but indeed detects anti-TES IgG that may persist for years; IgM or IgE-based serological tests that could better discriminate recent infection are not routinely available. On the other side, for European travelers with little previous exposure to parasites, cross-reaction is probably not a major issue, since test threshold has been set at value providing good specificity. In addition, the consistent clinical and laboratory diagnostic approach by a stable group of expert physicians throughout the study probably reflected the best accuracy that can be obtained in routine care. Finally, in this series, infection was most likely acquired abroad since symptoms developed during or shortly after travel, but infection in Belgium before travel or after return cannot be fully excluded.
The observation of toxocariasis in travelers is not surprising, although poorly studied so far. In a 10-year retrospective study in Spain, Toxocara antibodies were detected in 31 (4.9%) of 634 Latin American migrants and VLM was diagnosed in 4 of them [15], but this was in a particular segment of the travel population. With 28 highly probable/definitive toxocariasis cases diagnosed in about 85,000 travelers during the 13-year study period (33/100,000 travelers), our findings are in line with the recent multicenter GeoSentinel study (38/100,000 travelers) although the actual frequency was probably somewhat underestimated given the very restrictive case definition. Because of the high prevalence of toxocariasis in tropical countries and the inherent risks related to visiting regions with substandard hygiene (exposure to locally prepared food, incidental contacts with animals [23], it is reasonable to include toxocariasis in the differential diagnosis of most travel-related illnesses. However, surprisingly, the rate of Toxocara seropositivity in the (suspected) travelers for whom the test was requested (5.5%) was quite similar to that found in suspected autochthonous cases in Denmark (5.5%) [13] or the Netherlands (5–10%) [12] at similar ages. This observation does not support the idea that exotic travel by itself represents a major risk factor for Toxocara seropositivity, but since the frequency of clinical toxocariasis was not reported in those studies, comparisons with our findings remain inconclusive. Finally, only one study conducted in the Netherlands has investigated the incidence rate of Toxocara infection in travelers by comparing pre- and post-travel serology and found 1.1 seroconversion per 1000 person-months [24]. We confirm here that, even if infrequent, toxocariasis does occur in travelers and has to be considered after any type or duration of travel and from any destination.
Several factors contribute to underdiagnosis of toxocariasis, even in settings with higher resources [25]. Observed symptoms were often little specific and mimicked many other parasitic infections occasionally seen in travelers [26,27]. They were sometimes mild and self-limiting, not always triggering a complete etiological workup. Eosinophilia was rather high within the first weeks after symptom onset but tended to normalize quite rapidly. Laboratory investigations often detected evidence of other helminthic infection, with almost no possibility to discriminate between cross reaction, dual infection or remote exposure. Finally, we observed almost always excellent clinical and laboratory responses to albendazole, which is liberally used as empiric anti-helminthic treatment in travel medicine [28]. Clinicians often tend therefore to consider the specific diagnosis of toxocariasis as somehow difficult and of secondary importance.
Morbidity of toxocariasis should however not be underestimated. In the present series, most patients experienced a protracted illness, also because diagnosis was often delayed. Admission was necessary for three patients (10%), including both neurological cases for whom the diagnosis was particularly challenging. The clinical features of transverse myelitis had indeed developed insidiously, with no concomitant systemic symptoms and no blood eosinophilia at diagnosis, but well laboratory findings indicating an eosinophilic meningitis and non-specific alterations at the Magnetic Resonance Imaging of the spine [29–32]. Both cases did not fully recover despite maximal anti-helminthic therapy and prolonged corticosteroids. For all other cases however, administration of a 5-day course of albendazole was, when tolerated, immediately beneficial, suggesting that the current recommended practice is adequate [8]. This was not the case for ivermectin treatment with which clinical failures were observed [33]. Of note, no clinical exacerbation was observed during anti-parasitic treatment in contrast with what often occurs in other helminthic infections with larval invasion such as acute schistosomiasis [26]. Finally, we did not observe any case of ocular toxocariasis, but such cases are often serologically negative (due to low parasite load) and usually diagnosed in specialized ophthalmologic clinics [34].
In conclusion, symptomatic and asymptomatic toxocariasis was sporadically diagnosed in international travelers attending our center and had clinical and laboratory features overlapping those of many other tropical infections. In the present series, morbidity was non negligible and occasionally severe. A standard 5-day course of albendazole provided substantial clinical benefit without evidence of clinical exacerbation. Research is needed to develop antigen-based tests that would better reflect the disease activity both for diagnostic and monitoring purposes in clinical care.
|
10.1371/journal.pgen.1003189 | Tbx2 Controls Lung Growth by Direct Repression of the Cell Cycle Inhibitor Genes Cdkn1a and Cdkn1b | Vertebrate organ development relies on the precise spatiotemporal orchestration of proliferation rates and differentiation patterns in adjacent tissue compartments. The underlying integration of patterning and cell cycle control during organogenesis is insufficiently understood. Here, we have investigated the function of the patterning T-box transcription factor gene Tbx2 in lung development. We show that lungs of Tbx2-deficient mice are markedly hypoplastic and exhibit reduced branching morphogenesis. Mesenchymal proliferation was severely decreased, while mesenchymal differentiation into fibrocytes was prematurely induced. In the epithelial compartment, proliferation was reduced and differentiation of alveolar epithelial cells type 1 was compromised. Prior to the observed cellular changes, canonical Wnt signaling was downregulated, and Cdkn1a (p21) and Cdkn1b (p27) (two members of the Cip/Kip family of cell cycle inhibitors) were strongly induced in the Tbx2-deficient lung mesenchyme. Deletion of both Cdkn1a and Cdkn1b rescued, to a large degree, the growth deficits of Tbx2-deficient lungs. Prolongation of Tbx2 expression into adulthood led to hyperproliferation and maintenance of mesenchymal progenitor cells, with branching morphogenesis remaining unaffected. Expression of Cdkn1a and Cdkn1b was ablated from the lung mesenchyme in this gain-of-function setting. We further show by ChIP experiments that Tbx2 directly binds to Cdkn1a and Cdkn1b loci in vivo, defining these two genes as direct targets of Tbx2 repressive activity in the lung mesenchyme. We conclude that Tbx2-mediated regulation of Cdkn1a and Cdkn1b represents a crucial node in the network integrating patterning information and cell cycle regulation that underlies growth, differentiation, and branching morphogenesis of this organ.
| During organ formation, proliferation rates and differentiation patterns vary widely between different stages and tissue compartments. It is poorly understood how cell cycle progression is locally controlled and integrated with patterning processes in these developmental programs. Here, we used the mouse lung as a model to study how growth and differentiation are controlled on a transcriptional level. Combining genetic loss- and gain-of-function approaches, we show that the T-box transcription factor gene Tbx2 is required and sufficient to direct appropriate lung growth by maintaining proliferation and inhibiting differentiation in the mesenchymal compartment of the lung. We found that expression of the cell cycle inhibitor genes Cdkn1a (p21) and Cdkn1b (p27) inversely correlates with expression of Tbx2 and that deletion of both genes rescues, to a large degree, the growth deficits of Tbx2-mutant lungs. We further show by biochemical assays that Tbx2 directly binds to Cdkn1a and Cdkn1b loci in vivo, defining these two genes as direct targets of Tbx2 repressive activity in the lung mesenchyme. We conclude that Tbx2-mediated regulation of Cdkn1a and Cdkn1b represents a crucial module for the tissue-specific control of cell cycle progression that underlies growth, differentiation, and branching morphogenesis of this organ.
| The development of organs and organisms depends on the precise control of the progression through and the exit from the cell cycle to achieve appropriate patterns of proliferation and differentiation in time and space. Progression through the cell cycle is regulated predominantly by a series of serine/threonine kinases, the cyclin-dependent kinases (CDKs) that link proliferative signals with mechanical aspects of cell duplication. CDK function is controlled by a variety of mechanisms, including a group of molecules that inhibits CDK activity by complex formation. These CDK inhibitors (CKIs) have been categorized into two families, the Cip/Kip (Cdkn1) family with three members in mammals (Cdkn1a, Cdkn1b, Cdkn1c (also known as p21, p27 and p57)), that inhibit all kinases involved in G1/S transition, and the Ink4 (Cdkn2) family with four mammalian members (Cdkn2a, Cdkn2b, Cdkn2c, Cdkn2d (also known as p16/p19ARF, p15, p18, p19)) that specifically inhibit Cdk4 and Cdk6. Biochemical and cell culture experiments have identified CKIs as primary effectors of signaling pathways that control cell cycle exit, an event critical for differentiation. Expression or stability of CKIs is reduced in tumors, and deletion of six of the seven family members leads to organ hyperplasia and increased tumor susceptibility. In contrast to the obvious relevance of CKIs in tissue homeostasis, their role in development of tissues and organs, and the transcriptional mechanisms that mediate their precise temporal and spatial expression in the embryo have been much less well defined. This may relate to functional redundancy between family members as well as to the complexity of their regulatory modules (for reviews on CKIs see [1]–[3]).
T-box (Tbx) genes encode an evolutionary conserved family of transcription factors that regulate patterning and differentiation processes during vertebrate development [4]. Tbx2 and Tbx3 are two closely related members of the Tbx2-subfamily that are required in the development of numerous organs during mammalian embryogenesis including the heart, the palate, the limbs, and the liver [5]–[10]. In these contexts, these two transcriptional repressors mainly seem to regulate cell fate decisions and differentiation. However, in vitro studies indicated a role for Tbx2 and Tbx3 in the progression of the cell cycle [11]–[13]. Expression of Tbx2 and Tbx3 is upregulated in a number of tumors including those of the breast, pancreas, liver and bladder, and in melanomas, and both genes can function as immortalizing agents to bypass senescence, i.e. escape irreversible growth arrest (for reviews see [14], [15]). In cell culture assays, this phenomenon is mediated by transcriptional repression of Cdkn1a and Cdkn2a [12], [13], [16], [17]. Although often speculated (e.g. [6]), the relevance for this molecular function in a developmental context has remained unclear. Intriguingly, mice analyzed in our lab that were mutant for Tbx2, showed severely hypoplastic lungs, pointing to a possible role of this T-box factor in the regulation of proliferation and/or differentiation during development of this organ.
The architecture of the mammalian lung arises from a complex developmental program in which the tight orchestration of proliferation and differentiation processes assures the formation of an appropriately sized organ with a correct distribution of differentiated cell types for air-conduction and gas-exchange. In the mouse, the conducting airways develop from two primary buds that emerge from the ventral wall of the foregut endoderm at embryonic day (E) 9.5 by iterative rounds of stereotyped outgrowth and branching until E16.5. The gas-exchange units, the alveoli, only arise subsequently to this pseudoglandular stage from the terminal buds until late in postnatal life. Normal morphogenesis and patterning of the bronchial tree critically depends on the underlying mesenchyme that is derived from the splanchnic mesoderm. This mesenchyme is a source of signals that mediate proliferation of epithelial precursors, and direct their correct spatial differentiation. It also gives rise to a number of different cell types, including parabronchial and vascular smooth muscle cells, lipocytes, fibrocytes and endothelial cells. In turn, epithelial signals from the endoderm but also from the mesothelium maintain proliferation of mesenchymal precursors, closing a reciprocal signaling loop that directs outgrowth of the distal epithelial buds (for a recent review see [18]).
Previous work reported the expression of Tbx2 and Tbx3 in the mesenchymal compartment of the developing lung, but a functional significance has not been assigned to this expression [19], [20]. Here, we show by loss- and gain-of-function experiments in the mouse that Tbx2 is required and sufficient to maintain proliferation and inhibit differentiation in the mesenchymal compartment of the developing lung. Expression, organ culture and biochemical assays identify the cell cycle inhibitors encoded by the Cdkn1a and Cdkn1b genes as direct targets of Tbx2 repressive activity in this developmental program in vivo. Lung growth was substantially rescued by genetically limiting Cdkn1a and Cdkn1b expression in Tbx2-deficient mice, indicating that suppression of these genes is a critical function of Tbx2 in the control of organ growth during development.
Mice homozygous for a null allele of Tbx2 (Tbx2cre) that is maintained on an NMRI outbred background survive embryogenesis but die shortly after birth due to a cleft palate [9], [21]. Morphological and histological examination of mutant embryos at E18.5 revealed hypoplastic lungs that frequently manifested with alveolar haemorrhages. Air was present in the bronchial network but the lung was poorly inflated. Lobulation was normal but all four right lung lobes and the left lung lobe were reduced in size; the tissue appeared thickened. The weight of the mutant lung was reduced to approx. 50% of that of the littermate control whereas the liver and the spleen were unaffected excluding a general growth retardation problem (Figure 1A–1C). At E16.5, the mutant lung was visibly smaller and haemorrhagic. Its weight was reduced to 33% of the wildtype level (Figure 1D–1F). No obvious difference in morphology, histology and weight of the lung between wildtype and Tbx2-deficient embryos was observed at E14.5 (Figure 1G–1I).
To evaluate whether the decreased size of Tbx2-deficient lungs after E14.5 relates to a reduction in branching morphogenesis, we explanted E11.5 lung rudiments and analyzed their (2-dimensional) outgrowth after 6 days of culture. Whole-mount in situ hybridization analysis for expression of the epithelial tip marker gene Id2 showed an almost 3-fold reduction of branching endpoints in the Tbx2-mutant lung explants suggesting that epithelial branching morphogenesis is indeed severely hampered by loss of Tbx2 (Figure 1J–1L). However, reduction of branching morphogenesis was restricted to the late phase of lung outgrowth as revealed by non-significant changes of the number of branching endpoints in Tbx2-deficient cultures at 2 and 4 days (Figure S1). We conclude that Tbx2 is required to maintain normal branching morphogenesis and growth of the developing lung after E14.5.
Lung growth during the pseudoglandular stage is driven by branching morphogenesis of the distal lung buds. This, in turn, relies on rapid proliferation of the precursor cells in the bud epithelium and its underlying mesenchyme. Reduced size of Tbx2-deficient lungs could therefore relate to increased apoptosis and/or to decreased proliferation of distal epithelial and mesenchymal tissue compartments as shown for other models of lung hypoplasia [22]. Terminal deoxynucleotidyl transferase-mediated nick-end labeling (TUNEL) staining revealed that apoptosis was absent both in wildtype and mutant lungs at E14.5 and E16.5 but was increased in Tbx2-deficient lungs at E18.5 indicating a late contribution to the hypoplasia of this organ (Figure 2A and 2A′, 2D and 2D′, 2G and 2G′).
Analysis of 5-bromo-2′-deoxyuridine (BrdU) incorporation showed that the epithelial and mesenchymal tissue compartments of the lung were highly proliferative irrespective of the genotype at E14.5 (Figure 2B–2C). However, at E16.5 the BrdU labeling index showed a highly significant reduction from 29.6+/−1.5% in the wildtype to 16.0+/−3.4% in the mutant mesenchyme, and a significant reduction from 19.7+/−1.6% in the wildtype to 14.2+/−4.7% in the mutant distal lung epithelium (marked by expression of SRY-box containing protein (Sox)9 [23], [24]) while the proximal lung epithelium (marked by expression of Sox2 [25], [26]) or a control tissue (the diaphragm) were unaffected (Figure 2E–2F, Figure S2). At E18.5, proliferation as indicated by the BrdU labeling index was dramatically decreased in the whole lung to levels similar in wildtype and mutant embryos (Figure 2H–2I). Thus, Tbx2 is required to maintain normal proliferation of the mesenchyme and distal epithelium of the lung in a narrow temporal window.
As proliferation and differentiation are often inversely correlated, we next investigated the occurrence of changes in the differentiation patterns of both mesenchyme and epithelium in Tbx2-deficient lungs at E14.5, E16.5 and E18.5 to cover the period before, around and after the histological and cellular defects were apparent in the mutant (Figure 3). At all analyzed stages a normal distribution of networks of endomucin (Emcn)-positive endothelial cells [27] was present throughout the mutant lung. In the mutant mesenchyme, transgelin (Tagln)-positive smooth muscle cells were restricted to the proximal airways as in the wildtype. It has recently been shown that expression of S100 calcium binding protein A4 (S100a4) marks fibroblasts that are highly proliferative and express low levels of extracellular matrix proteins indicating the precursor character of these cells [28]–[30]. We found that in E14.5 wildtype lungs all mesenchymal cells that were positive for S100a4 also incorporated BrdU confirming the proliferative character of this cell type (Figure S3). In the Tbx2-deficient lung, expression of S100a4 was completely abolished. Fibronectin (Fn) and periostin (Postn), extracellular matrix proteins that are secreted by mature fibrocytes at low levels in proximal airways in the wildtype [30]–[32], were expressed throughout the mesenchyme starting from E14.5 (Fn) and E16.5 (Postn) in the mutant lung (Figure 3). This suggests that Tbx2 is required to maintain the precursor state of a subpopulation of future fibrocytes in the lung mesenchyme.
We next investigated whether these mesenchymal changes are accompanied by alterations in proximal-distal patterning of the respiratory tree and cell differentiation in the epithelium during development in Tbx2-deficient lungs (Figure S4). In the wildtype lung, Sox2 was expressed in the trachea and proximal airways and was excluded from distal endoderm at all analyzed stages. Sox9 was expressed in the distal tip endoderm and excluded proximally at E14.5 and E16.5. At E18.5, Sox9 was downregulated distally and reactivated in the mesenchyme of the proximal airways possibly indicating onset of cartilage formation in this region. Expression of keratin 14 (Krt14, also known as cytokeratin 14) in basal cells in the trachea, of tubulin, beta 4A class IVA (Tubb4a) in ciliated cells, and of secretoglobin, family 1A, member 1 (Scgb1a1, also known as CC10 and uteroglobin) in secretory or Clara cells was activated at E16.5 and maintained at E18.5. Krt14 was also found in myofibroblasts surrounding the proximal airways at E16.5. Surfactant associated protein C (Sftpc1, also known as SP-C) was expressed in alveolar epithelial cells type II (AEC2) from E16.5. Expression of podoplanin (Pdpn) and aquaporin 5 (Aqp5) was activated in AEC1 at E18.5. All of these markers (described in [18]) were appropriately activated and maintained in Tbx2-deficient lungs with the exception of Pdpn and Aqp5 that showed reduced expression levels at E18.5. We conclude that mesenchymal loss of Tbx2 does not affect proximal-distal patterning of the lung epithelium. Reduced or delayed differentiation of AEC1 from AEC2 may relate to the loss of appropriate signaling from the prematurely differentiated mesenchyme.
Phenotypic changes of Tbx2-deficient lungs were confined to the late phase of branching morphogenesis suggesting a narrow temporal window of expression and/or activity of this gene. Alternatively, Tbx2 may act redundantly with other T-box genes during early lung development. In fact, previous work reported expression of Tbx2 as well as of Tbx3, Tbx4 and Tbx5 in the pulmonary mesenchyme [20], [33]. To assess the comparative temporal expression patterns of these genes during lung development, we performed in situ hybridization analysis on sagittal sections of the lung (Figure 4). We observed coexpression of Tbx2 and Tbx3 in the mesenchymal compartment from E10.5 to E14.5. Expression of Tbx3 declined sharply after this stage, whereas Tbx2 was maintained at high levels at subsequent embryonic stages. Coexpression of Tbx4 and Tbx5 was found between E10.5 to E16.5 in the lung mesenchyme. Hence, late onset of phenotypic changes in Tbx2-deficient lungs may relate to functional redundancy with the closely related Tbx3 gene during the initial phase of branching morphogenesis. This notion is supported by the finding that mice homozygous for a null allele of Tbx3 exhibit lungs morphologically and histologically indistinguishable from the wildtype at E14.5, shortly before these mice die ([8] and data not shown). Since mice with more than two mutant alleles of Tbx2 and Tbx3 die around E9.5 due to cardiac defects [10], analysis of the functional redundancy of the two genes in early lung development was not possible with the mouse lines available to us.
An antisense oligonucleotide approach with cultured lung rudiments and more recently conditional gene targeting demonstrated a requirement for mesenchymal Tbx4 and Tbx5 in the regulation of pulmonary branching morphogenesis [19], [33]. Tbx4 and Tbx5 genetically interact with Fgf10 during lung growth and branching, and may direct transcriptional activation of Fgf10 that encodes a potent growth factor in the lung but also in other developmental contexts [33], [34]. Given the molecular nature of Tbx2 and Tbx3 as transcriptional repressors, Tbx2 and Tbx3 may compete with Tbx4 and Tbx5 for binding to conserved DNA-binding sites in the promoter of Fgf10, similar to the antagonistic control of Nppa expression in the heart by Tbx5 and Tbx2/Tbx3 [35].
To test this hypothesis and determine the molecular changes underlying the lung phenotype, we analyzed components as well as targets of bone morphogenetic protein (Bmp)-, fibroblast growth factor (Fgf), sonic hedgehog (Shh) and wingless-related MMTV integration site (Wnt) pathways that collectively confer outgrowth and branching morphogenesis of the respiratory tree [18]. To accurately identify expression changes we used quantitative RT-PCR of whole lung extracts at different developmental stages. We started our analysis with lungs at E16.5, when morphological, histological and proliferation defects were fully apparent (Figure 5A, grey bars). At this stage, we observed a significant downregulation of components of the Bmp pathway such as Bmp4 and Bmp receptor (Bmpr)2 as well as the Bmp target gene homeobox, msh-like (Msx)1 [36]. Bmp2 and Bmpr1a expression, however, was not significantly altered. Expression of Shh was markedly reduced but not accompanied by decreased intracellular signaling as revealed by almost normal expression of the target gene patched (Ptch)1 [37]. Wnt ligands Wnt2 and Wnt5a were strongly reduced in their expression as was the target of the canonical (Ctnnb1-dependent) sub-branch of Wnt signaling, Axin2 [38]. Unexpectedly, no changes in Fgf pathway components were found. Fgf10 expression was at wildtype level as was the receptor Fgfr2 and the known Fgf target ets variant gene 4 (Etv4, also known as Pea3) [39]. At E14.5, i.e. prior to the observed phenotypic changes, components and targets of Shh-, Fgf- and Bmp-activity were unchanged in their expression. The canonical Wnt target gene Axin2 and the non-canonical ligand Wnt5a were strongly and Wnt2 expression was slightly reduced in mutant lungs (Figure 5A, black bars). In situ hybridization analysis showed that downregulation of Wnt2, Wnt5a and Axin2 was confined to the mesenchymal compartment of E14.5 Tbx2−/− lungs (Figure 5B). These data suggest, that Tbx2 does not counteract the transcriptional activation of Fgf10 transcription and Fgf signaling by Tbx4/Tbx5 but targets canonical Wnt signaling in the lung mesenchyme, what, in turn, may secondarily affect Bmp signaling.
Next, we analyzed expression of cell cycle regulators potentially involved in proliferation control of lung mesenchyme (Figure 5C, grey bars). Among the tested cell cycle activators cyclin-dependent kinase (Cdk)1 and cyclin D (Ccnd)1 showed significant reduction whereas Ccnd2 and Ccnd3 expression was unchanged at E16.5. As Ccnd1 has been described as target of canonical Wnt signaling [40], its reduced expression may relate to the observed downregulation of this pathway. The cell cycle inhibitors Cdkn1a, Cdkn1c, Cdkn2a and Cdkn2d were unchanged whereas Cdkn1b was upregulated more than 7 times in the mutant at this stage. At E14.5, all cell cycle regulators were unaffected except Cdkn1b and Cdkn1a that were upregulated 4 and 3.5 times, respectively, in the Tbx2-deficient lung (Figure 5C, black bars). Expression of Tbx3 was unaltered at both analyzed stages excluding a compensatory upregulation of this gene in the Tbx2-mutant background (Figure 5C). In situ hybridization and immunofluorescence analyses confirmed strong upregulation of Cdkn1a/Cdkn1b mRNA and Cdkn1a/Cdkn1b protein both in the mesenchymal and in the distal epithelial compartment of E14.5 Tbx2−/− lungs (Figure 5D). These results argue that reduced proliferation in the mesenchyme and distal epithelium (that are probably secondary to altered mesenchymal signals) of E16.5 Tbx2-deficient lungs may be caused by de-repression of cell cycle inhibitors Cdkn1a and Cdkn1b. Decreased (canonical) Wnt signaling in Tbx2-deficient lungs may reflect an independent branch of Tbx2 activity, or may merely present a secondary consequence of de-repression of cell cycle inhibitors.
To unravel the contribution of increased expression of Cdkn1a and Cdkn1b to the growth deficit of Tbx2-deficient lungs, we ablated the two genes in the mutant background. Compound Tbx2;Cdkn1a and Tbx2;Cdkn1b mutants, respectively, exhibited lungs that were morphologically indistinguishable from the Tbx2-single mutant organ. In contrast, triple Tbx2;Cdkn1a;Cdkn1b mutants exhibited visibly larger lungs at E18.5 (Figure 6A). To quantify the observed changes, we determined the relative lung weight (lung weight to body weight ratios, normalized to that of Tbx2+/− control embryos) of the different compound mutants. Statistical analysis did not detect significant weight changes between Tbx2−/−;Cdkn1a−/− and Tbx2−/−;Cdkn1b−/− lungs, whereas the increase in weight in Tbx2−/−;Cdkn1a−/−;Cdkn1b−/− lungs was highly significant (Figure 6B). Although Tbx2−/−;Cdkn1a−/−;Cdkn1b−/− lungs reached 80% of the control weight, the difference remained significant indicating an incomplete rescue. This suggests that the combined de-repression of Cdkn1a and Cdkn1b accounts predominantly but not completely for hypoplasia of Tbx2-deficient lungs.
Since the individual deletion of Cdkn1a and Cdkn1b in the Tbx2-mutant background did not lead to even a partial rescue of growth, we tested for the presence of a compensatory mechanism by analyzing expression of Cdkn1a and Cdkn1b, respectively, by quantitative RT-PCR analysis on mRNA of E16.5 (compound) mutant lungs (Figure 6C). Cdkn1a expression was increased 2-fold in Cdkn1b−/− lungs and 9-fold in Tbx2−/;−Cdkn1b−/− lungs whereas Cdkn1b was increased 2-fold in Cdkn1a−/− lungs and 3-fold in Tbx2−/−;Cdkn1a−/− lungs at this stage. Thus, either gene shows a compensatory upregulation upon loss of the other gene.
Binding sites for TBX2 within the Cdkn1a promoter have recently been described in cell culture experiments [12] whereas Cdkn1b has not been recognized as a direct target of Tbx2 repressive activity before. In silico analysis of the mouse Cdkn1a and Cdkn1b genes identified a consensus DNA-binding site for T-box proteins (T-box binding element (TBE): AGGTGTGA) [41] in the Cdkn1a promoter and two putative TBEs in the Cdkn1b locus. The first element (AGGTGTGTG) was detected 3 kbp upstream of the start codon, the second element with the reverse complementary sequence CACACCT was localized within an intron of that gene (Figure S5). ChIP experiments with E15.5 lung tissue revealed in vivo binding of Tbx2 to the known TBE in the Cdkn1a locus and to the 5′ located but not the intronic TBE in the Cdkn1b gene (Figure 6D) compatible with the notion that Cdkn1a and Cdkn1b represent direct targets of Tbx2 repressive activity in the lung mesenchyme.
It has previously been shown that Cdkn1a expression is elevated on inactivation of endogenous Tbx2 in the murine B16 melanoma and the human MCF-7 breast cancer cell line [12]. Using the previously published conditions [12], we downregulated Tbx2 in both cell lines using a Tbx2-specific siRNA approach. Immunofluorescence analysis showed that in the non-silencing control nuclear Tbx2 protein was present in all cells whereas Cdkn1b was not detected. In contrast, in cells treated with the Tbx2-specific siRNA Tbx2 expression was extinguished in almost all cells examined whereas Cdkn1b expression was strongly upregulated in the cytoplasm (in MCF-7 cells) and in the nucleus (in B16 melanoma cells) (Figure S6). This further supports that Cdkn1b similar to Cdkn1a is a true target of Tbx2.
To further evaluate the mechanistic role of Tbx2 in the lung mesenchyme, we additionally employed an in vivo gain-of-function approach. For this, we crossed the Tbx2cre line and an HprtTBX2-allele, that was generated by integration of a bicistronic transgene-cassette containing the human TBX2 ORF followed by IRES-GFP in the ubiquitously expressed X-chromosomal Hypoxanthine guanine phosphoribosyl transferase (Hprt) locus [10], [42] to maintain TBX2 expression in its endogenous domains including the lung mesenchyme. Male (Tbx2cre/+;HprtTBX2/y) embryos were not recovered after E12.5 most likely due to cardiac defects. In contrast, female (Tbx2cre/+;HprtTBX2/+) embryos, which exhibit a mosaic expression due to random X-chromosome inactivation, survived embryogenesis and puberty. Lungs of E18.5 Tbx2cre/+;HprtTBX2/+ embryos were slightly bigger than those of control littermates and showed a looser tissue organization (Figure 7A). Apoptosis was not detected in either genotype, but Tbx2cre/+;HprtTBX2/+ lungs exhibited a strong increase of proliferation in the mesenchyme as shown by the BrdU assay (Figure 7B). Notably, Western blot analysis of lungs of E18.5 Tbx2cre/+;HprtTBX2/+ embryos showed that transgenic TBX2 expression did not reach unphysiologically high levels (Figure S7).
Branching morphogenesis is downregulated after E16.5 concomitant with the shut-down of signaling pathways involved in epithelial-mesenchymal tissue interactions at the distal lung buds. Therefore, increased proliferation in Tbx2cre/+;HprtTBX2/+ lungs may relate to continued branching by maintained activity of these signaling pathways. Morphological inspection did not detect changes of branching between E12.0 wildtype and Tbx2cre/+;HprtTBX2/+ lung rudiments cultured for 6 days (Figure S8). Furthermore, RT-PCR analysis found unchanged expression of targets of Shh (Ptch1), Fgf (Etv4), Bmp (Msx1) and canonical Wnt (Axin2) pathways in Tbx2cre/+;HprtTBX2/+ lungs at E18.5 showing that Tbx2 is not sufficient to induce these pathways, thus, branching morphogenesis (Figure 7C). However, when testing cell cycle regulators in this assay, we detected a selective downregulation of Cdkn1a and Cdkn1b in Tbx2cre/+;HprtTBX2/+ lungs showing that Tbx2 is not only required but also sufficient to repress expression of Cdkn1a and Cdkn1b (Figure 7D).
To evaluate long-term consequences of prolonged TBX2 expression in the lung mesenchyme, we analyzed Tbx2cre/+;HprtTBX2/+ mice at postnatal day (P) 40, a stage when they were present in the expected numbers. Although Tbx2cre/+;HprtTBX2/+ mice were visibly smaller than their littermate controls at this stage, the relative lung mass was increased by a factor of 1.27 (Figure 8A and 8B). Immunofluorescence for GFP and TBX2 expression on lung sections confirmed the widespread expression of the transgene in the mesenchymal compartment of P40 Tbx2cre/+;HprtTBX2/+ mice. Histological analysis by haematoxylin and eosin staining uncovered clusters of tissue thickenings, and alveolar air spaces were surrounded by multiple cell layers in these transgenic lungs. Histological staining for keratin and collagen (Masson's trichrome) did not detect changes in the transgenic lung, excluding the possibility that tissue thickening is caused by excessive deposition of extracellular matrix (Figure 8C). Analysis of BrdU incorporation showed that the lung tissue was highly proliferative in the transgenic animals at P40 (Tbx2cre/+;HprtTBX2/+: 31.0%±5.2, control: 2.1%±0.8) (Figure 8D). Apoptosis as detected by TUNEL staining was similarly absent from control and transgenic lungs (Figure 8E).
Analysis of cell differentiation by immunofluorescence of marker proteins showed normal presence of lung epithelial cell types, of endothelial cells, and of mesenchymal smooth muscle cells around the proximal airways of Tbx2cre/+;HprtTBX2/+ lungs (Figure S9). Fn1 and Postn deposition in the extracellular matrix was augmented, and S100a4-positive cells were increased in number. Expression of the cell cycle inhibitors Cdkn1a and Cdkn1b was dramatically downregulated in the mesenchymal compartment (Figure 8F). Together these findings indicate that prolonged expression of TBX2 maintains mesenchymal proliferation at a high level. While a part of these mesenchymal cells differentiate into ECM-producing cell-types, a substantial fraction retains a S100a4-positive precursor character.
To determine the contribution of reduced expression of Cdkn1a and Cdkn1b to the observed histological, immunohistochemical and molecular changes in Tbx2cre/+;HprtTBX2/+ mice, we additionally analyzed Cdkn1a−/−;Cdkn1b−/− mice at P40 using a similar panel of assays. Double mutant lungs, normalized against the increased body weight, were significantly larger than lungs of their littermates (Figure S10A, S10B). Histological analysis did not find changes in the tissue organization (Figure S10C) but Cdkn1a−/−;Cdkn1b−/− lungs exhibited a 5-fold increase of proliferation as shown by the BrdU assay compared to the wildtype. Apoptosis was unaffected (Figure S10D, S10E). Fn1 and Postn deposition in the extracellular matrix was normal and immature fibroblasts (S100a4) were absent as in the wildtype. Immunofluorescence analysis of Cdkn1a and Cdkn1b confirmed that both proteins were completely absent in the Cdkn1a−/−;Cdkn1b−/− lung (Figure S10F). Hence, Cdkn1a−/−;Cdkn1b−/− mice do not feature the histological and cellular changes seen in Tbx2cre/+;HprtTBX2/+ mice, but exhibit increased lung mass due to increased proliferation. We conclude that downregulation of Cdkn1a and Cdkn1b mediates the pro-proliferative effects of Tbx2 overexpression to a large degree but may not account for changes in tissue architecture and cell differentiation.
Branching morphogenesis and growth of the lung requires the coordination of cellular behaviors of its epithelial and mesenchymal tissue compartments. Here, we have identified Tbx2 as a crucial mesenchymal factor that maintains the mesenchymal signaling center for epithelial branching morphogenesis. We suggest that Tbx2 promotes mesenchymal proliferation and inhibits terminal differentiation partly via direct transcriptional repression of cell cycle inhibitor genes. Irrespective of its precise mode, Tbx2 additionally maintains canonical Wnt signaling in the mesenchyme, which, in turn, may account for maintenance of epithelial growth and branching at the distal tips of the lung buds (Figure 8G).
Cdkn1a, Cdkn1b together with Cdkn1c constitute the Cip/Kip family of CKIs that inhibit cell cycle progression by binding to and inhibition of a broad range of cyclin-CDK complexes via a shared N-terminal cyclin-CDK binding domain. Cdkn1 activity correlates with cell cycle exit and differentiation, and is, thus, under tight control of anti-mitogenic signals. In tissue homeostasis, expression and activity of CKIs is regulated by a large number of molecular mechanisms including protein binding and posttranslational modification that affect cyclin/CDK binding as well as stability and degradation of CKIs (for a review see [43]). Gene targeting experiments have unambiguously shown that all three members are important players in tissue homeostasis and cancer (for a review see [2]) whereas Cdkn1c is the only CKI to be uniquely required for embryonic development [44], [45]. However, additional congenital defects have been described in mice lacking more than one member of this gene family pointing to redundant functions in some but not all developmental processes (see e.g. [46], [47].
To exert a precise timing of cell cycle exit and differentiation in development, expression of Cdkn1 genes must be tightly controlled on the transcriptional level. In fact, Cdkn1a and Cdkn1c have specific patterns of expression in development that correlate with terminal differentiation of multiple cell lineages including skeletal muscle, cartilage, skin, and nasal epithelium. In contrast, Cdkn1b expression appears more widespread (for a review see [2]. Cell culture experiments identified Cdkn1a as a transcriptional target of p53 [48], [49] whereas the transcriptional regulation of Cdkn1c is mediated by factors that play critical roles during embryogenesis such as Notch/Hes1, MyoD and p73 [50]–[52]. To our knowledge, the in vivo relevance of these regulatory modules has remained unclear. Interestingly, previous efforts were largely directed towards the identification of transcriptional activators of Cdkn1 genes, and the possibility that these genes are subject to negative regulation in vivo, i.e. that activation of expression in a certain cell type results from attenuation or abolition of a prior transcriptional repression, was neglected.
Here, we have shown that Cdkn1a and Cdkn1b are derepressed in the pulmonary mesenchyme in Tbx2-deficient mice prior to other molecular changes, that Cdkn1a and Cdkn1b are repressed upon ectopic expression of TBX2 in mature lung mesenchyme, and that deletion of Cdkn1a and Cdkn1b largely rescued the growth defects of Tbx2-deficient lungs. Furthermore, we identified by ChIP analysis Tbx2 binding to Cdkn1a and Cdkn1b loci in the developing lung. Together, our genetic and biochemical analyses provide evidence that Cdkn1a and Cdkn1b are subject to direct repression by Tbx2 and are crucial downstream mediators of this gene in the mesenchymal compartment of the developing lung. In turn, it is the first clear evidence, that Tbx2 directly regulates cell cycle control genes in a developmental context in vivo. Intriguingly, ChIP-seq analysis of genomic binding of Tbx3 in cardiomyocytes in vivo, identified a large number of loci with binding peaks containing a variant TBE [53]. Tbx3 and Tbx2 are closely related family members that recognize the same DNA binding site. Re-inspection of this data set identified binding peaks of Tbx3 in both the Cdkn1a and Cdkn1b loci. In fact, the DNA-element used in our ChIP analysis precisely mapped to a major Tbx3 peak in the promoter of the Cdkn1a locus which contained additional less conserved TBEs, whereas the DNA element used for our Cdkn1b ChIP located closely to a minor peak (Figure S11). This together with enhanced expression of Cdkn1a and Cdkn1b in melanoma and breast cancer cell lines depleted of endogenous Tbx2 [12], [13], [17] (and this study), indicates that Tbx3 and the closely related Tbx2 protein occupy DNA sites in the Cdkn1a and Cdkn1b loci in other cell types and may regulate these genes in other developmental contexts as well.
It should be noted that changes of Tbx2 did not only (inversely) affect proliferation in the lung mesenchyme but directly correlated with the precursor state of at least one mesenchymal sub-population, S100a4-positive fibroblasts. Although Cdkn1-mediated cell cycle arrest has been associated with cellular differentiation in different developmental contexts [44], [54], [55], we did not observe differentiation defects in lungs double mutant for Cdkn1a and Cdkn1b. This may indicate that in this developmental context negative control of cell differentiation by Tbx2 is not mediated by repression of Cdkn1 and Cdkn1b. Changes of Tbx2 expression did not affect differentiation of other mesenchyme-derived cell types including smooth muscle cells. This may indicate that Tbx2 does not control differentiation of these cell types, or it may simply reflect the fact that these cell types differentiate prior to E14.5 when Tbx3 expression is downregulated and Tbx2 is uniquely required. In the future, it will be interesting to study the relation between mesenchymal proliferation and differentiation in mice deficient for both Tbx2 and Tbx3, which are likely to act redundantly throughout the pseudoglandular stage until E14.5.
Combined deletion of Cdkn1a and Cdkn1b function in Tbx2-deficient embryos restored lung growth largely but not completely suggesting that additional factors or pathways may act downstream of Tbx2 to mediate mesenchymal proliferation. Our RT-PCR analysis of signaling pathways relevant for branching morphogenesis did not detect changes of Fgf and Shh signaling but uncovered reduced activity of canonical Wnt and Bmp signaling. Notably, we detected decreased expression of Wnt components and signaling as early as E14.5 in the pulmonary mesenchyme, whereas Bmp4 expression and Bmp signaling was unchanged at that stage suggesting a secondary mode of change of the latter. As Bmp4 was shown to act as an autocrine signal for distal endoderm proliferation [56], reduced expression may contribute to the reduced proliferation in the distal endoderm at E16.5 in Tbx2-deficient lungs.
A number of studies have implicated different Wnt genes in lung development. Mice deficient for the non-canonical Wnt ligand gene Wnt5a, which is expressed in the distal lung mesenchyme, exhibit increased cell proliferation in both epithelium and mesenchyme with a resulting expansion of the distal lung and increased lung size [57]. Wnt2 is a canonical Wnt ligand robustly expressed in the mesenchyme of the developing lung. Wnt2, in cooperation with Wnt2b, is essential for specification of the respiratory lineage in the anterior foregut endoderm [58]. Later, Wnt2 acts upstream of Fgf10 and the critical transcription factor myocardin to regulate early airway smooth muscle cell differentiation in the multipotent lung mesenchyme [59]. Finally, Wnt7b, a canonical ligand expressed in the pulmonary epithelium stimulates embryonic lung growth by increasing proliferation in both tissue compartments of the developing lung without affecting the differentiation patterns [22]. Furthermore, tissue-specific deletion of the unique signaling mediator of the canonical pathway, Ctnnb1, in the epithelium led to defects in proximal-distal differentiation of airway epithelium [60] whereas mesenchymal deletion of Ctnnb1 resulted in hypoplasia due to reduced epithelial and mesenchymal proliferation [61].
Maintained differentiation of airway smooth muscle cells but decreased proliferation in the epithelial and mesenchymal compartments during the late phase of branching morphogenesis is compatible with the idea that loss of Tbx2 affects the canonical Wnt pathway in the mesenchyme triggered by the epithelial Wnt7b signal. The growth-promoting effect of this pathway may at least partly be mediated by activation of the pro-proliferative gene Ccnd1 that was previously recognized as a target of Wnt signaling [62]. This is compatible with the finding that proliferation defects observed in Tbx2-mutant lungs at E16.5 coincide with a strong decline of expression of this gene at this stage. However, the significance of downregulation of Wnt2 and Wnt5a in the Tbx2-deficient lung remains unclear. We assume that it provides only a minor contribution to the observed changes.
At present, we cannot distinguish whether changes of Wnt signaling activity are secondary to cell cycle exit and/or upregulation of Cdkn1a and Cdkn1b or represent an independent branch of Tbx2 transcriptional activity in the lung mesenchyme. Unfortunately, the recovery of mice triple mutant for Tbx2, Cdkn1a and Cdkn1b for analysis of signaling pathways at E14.5 is extremely inefficient. The finding that constitutive expression of Tbx2 in the lung mesenchyme of adult mice did not increase canonical Wnt signaling, suggests that Tbx2 is not sufficient to activate this pathway. However, Tbx2 may be required for repression of an inhibitor of Wnt signaling to maintain this pathway during branching morphogenesis. The relevance of the control of canonical Wnt signaling by Tbx2 in the lung mesenchyme will be addressed in future experiments.
All animal work conducted for this study was approved by H. Hedrich, state head of the animal facility at Medizinische Hochschule Hannover and performed according to German legislation.
Mice carrying a null allele of Cdkn1a (Cdkn1atm1Tyj, synonym Cdkn1a−) [63], a null allele of Cdkn1b (Cdkn1btm1Mlf, synonym: Cdkn1b−) [64] or a null allele of Tbx2 (Tbx2tm1.1(cre)Vmc, synonyms: Tbx2−, Tbx2cre) [21], and mice with integration of the human TBX2 gene in the Hprt locus (Hprttm2(CAG-TBX2,-EGFP)Akis, synonym: HprtTBX2) [10] were maintained on an outbred (NMRI) background. For timed pregnancies, vaginal plugs were checked in the morning after mating; noon was taken as embryonic day (E) 0.5. Pregnant females were sacrificed by cervical dislocation; embryos were harvested in phosphate-buffered saline, decapitated, fixed in 4% paraformaldehyde overnight, and stored in 100% methanol at −20°C before further use. Genomic DNA prepared from yolk sacs or tail biopsies was used for genotyping by polymerase chain reaction (PCR). For primers and conditions see Table S2.
Embryos were embedded in paraffin and sectioned to 5 µm. For histological analyses, sections were stained with haematoxylin and eosin (HE), Masson's trichrome (Masson's) and picrosirius red (Sirius red) following standard protocols. For the detection of antigens, antigen retrieval was performed using citrate-based antigen unmasking solution (H-3300, Vector Laboratories Inc). Sections were pressure-cooked for 5 min and signal amplification was performed with the Tyramide Signal Amplification (TSA) system (NEL702001KT, Perkin Elmer LAS) or the DAB substrate kit (SK-4100, Vector Laboratories Inc). The following primary antibodies were used: rabbit anti-mouse E-cadherin (gift from Rolf Kemler, MPI for Immunobiology and Epigenetics, Freiburg/Germany) [65], rabbit polyclonal antibody against GFP (1∶200, sc-8334, Santa Cruz), mouse monoclonal antibody against GFP (1∶200, 11814460001, Roche), monoclonal antibody against alpha smooth muscle actin, Cy3-conjugate (1∶200, C 6198, Sigma), monoclonal antibody against alpha smooth muscle actin, FITC-conjugate (1∶200, F3777, Sigma), rabbit polyclonal against SM22a (transgelin, 1∶200, ab14106, Abcam), rat monoclonal antibody against endomucin (1∶2, gift from Dietmar Vestweber, MPI for Molecular Medicine, Münster/Germany) [66], rabbit polyclonal antibodies against Tbx2 (1∶100, ab33298, Abcam), Cdkn1a (1∶200, sc-397, SantaCruz), Cdkn1b (1∶200, 554069, BD Biosciences), uteroglobin (1∶200, ab40873, Abcam), cytokeratin14 (1∶200, ab7800, Abcam), Tubb4a (1∶100, ab11315, Abcam), prosurfactant protein C (1∶200, ab40879, Abcam), Sox2 (1∶100, ab97959, Abcam), Sox9 (1∶200, ab5535, Millipore), aquaporin5 (1∶100, ab92320, Abcam), hamster monoclonal against podoplanin (1∶50, ab11936, Abcam) and mouse monoclonal against BrdU (1∶100, 1170376, Roche). For immunofluorescent stainings on adult sections or double immunofluorescent stainings with two primary mouse antibodies the Biotinylated Mouse on Mouse (M.O.M.) Anti-Mouse Ig Reagent (Vector laboratories) was used.
For analysis of branching morphogenesis E11.5 or E12.0 lung rudiments were dissected and kept on Transwell permeable 0.4-µm pore size, PET 6-well plates (Corning) supplied with DMEM supplemented with 10% fetal calf serum (Biowest), 2 mM Glutamax, 100 units/ml Penicillin, 100 µg/ml Streptomycin (Gibco). Lungs were cultivated at 37°C and 5% CO2 for 2 to 6 days and the number of branching endpoints was counted.
Human MCF-7 breast adenocarcinoma cell line was cultured in RPMI 1640 with Glutamax (Gibco) supplemented with 10% FBS, MEM non-essential amino acids (Gibco), 1 mM sodium pyruvate (Gibco), 10 µg/ml human insulin (Roche) and 100 units/ml Penicillin, 100 µg/ml Streptomycin (Gibco). Mouse B16 melanoma cells were cultured in RPMI 1640 with glutamax, supplemented with 10% FBS and 100 units/ml Penicillin, 100 µg/ml Streptomycin.
Downregulation of TBX2 or Tbx2 was achieved by siRNA exactly as recently described [12].
Whole-mount in situ hybridization was performed following a standard procedure with digoxigenin-labeled antisense riboprobes [67]. Stained specimens were transferred in 80% glycerol prior to documentation. In situ hybridization on 10 µm paraffin sections was done essentially as described [68]. For each marker at least three independent specimens were analyzed.
Cell proliferation in embryonic and adult lungs was investigated by detection of incorporated 5-bromo-2′-deoxyuridine (BrdU) similar to published protocols [69]. At least nine sections from three individual embryos per genotype and stage were used for quantification. Statistical analysis was performed using the two-tailed Student's t-test. Data were expressed as mean ± standard deviation. Differences were considered significant when the P-value was below 0.05.
For detection of apoptotic cells in 5 µm paraffin sections of embryos, the terminal deoxynucleotidyl transferase-mediated nick-end labeling (TUNEL) assay was performed as recommended by the manufacturer (Serologicals Corp.) of the ApopTag kit used.
Total RNA was extracted from dissected lungs with RNAPure reagent (Peqlab). RNA (500 ng) was reverse transcribed with RevertAid H Minus reverse transcriptase (Fermentas). For semiquantitative PCR, the number of cycles was adjusted to the mid-logarithmic phase. Quantification was performed with Quantity One software (Bio-Rad). Assays were performed at least twice in duplicate, and statistical analysis was done as previously described [9]. For primers and PCR conditions see Table S3.
2ChIP was performed essentially as previously described [70]. Dissected E15.5 lung tissue was treated with 4% paraformaldehyde overnight. The DNA-containing supernatants were incubated overnight with anti-Tbx2 antibodies and collected on protein G beads. Cross-linked products were reversed by cooking for 15 min, treated with Proteinase K and RNAse H at 56°C for 30 min and the immunoprecipitated DNA was purified. Primers for PCR amplification were 5′-CCGAGAGGTGTGAGCCGC-3′ (Cdkn1a-f1) and 5′- GTCATCCACCTGCCGCGG-3′ (Cdkn1a-r1); 5′-GGCTTAGATTCCCAGAGGG-3′ (Cdkn1af2) and 5′-TTCTGGGGACACCCACTGG-3′ (Cdkn1a-r2) for the Cdkn1a promoter and 5′- CAAGTTCAGTAAACTAAGTAGG-3′ (Cdkn1b-f1) and 5′- GCACATATGTGGACAAACTCG-3′ (Cdkn1b-r1) for the 5′-T-site in the Cdkn1b promoter. For the intron located T-site 5′-ATATACCTTCTACAGACATAGC-3′ (Cdkn1b-f2) and 5′- GCTTTTGACTAGAGTCTTATGG-3′ (Cdkn1b-r2) primers were used. Primers for the negative control region were 5′-CTCTGAAACTCGAACAGGCC-3′ (ncr-f1) and 5′- ACTCTGAATTGGATTCCTAGC-3′ (ncr-r1).
Sections were photographed using a Leica DM5000 microscope with a Leica DFC300FX digital camera. Whole mount specimens were photographed on a Leica M420 microscope with a Fujix digital camera HC-300Z. Images were processed in Adobe Photoshop CS3.
|
10.1371/journal.ppat.1003449 | Location of the CD8 T Cell Epitope within the Antigenic Precursor Determines Immunogenicity and Protection against the Toxoplasma gondii Parasite | CD8 T cells protect the host from disease caused by intracellular pathogens, such as the Toxoplasma gondii (T. gondii) protozoan parasite. Despite the complexity of the T. gondii proteome, CD8 T cell responses are restricted to only a small number of peptide epitopes derived from a limited set of antigenic precursors. This phenomenon is known as immunodominance and is key to effective vaccine design. However, the mechanisms that determine the immunogenicity and immunodominance hierarchy of parasite antigens are not well understood.
Here, using genetically modified parasites, we show that parasite burden is controlled by the immunodominant GRA6-specific CD8 T cell response but not by responses to the subdominant GRA4- and ROP7-derived epitopes. Remarkably, optimal processing and immunodominance were determined by the location of the peptide epitope at the C-terminus of the GRA6 antigenic precursor. In contrast, immunodominance could not be explained by the peptide affinity for the MHC I molecule or the frequency of T cell precursors in the naive animals. Our results reveal the molecular requirements for optimal presentation of an intracellular parasite antigen and for eliciting protective CD8 T cells.
| Toxoplasma gondii is a widespread intracellular parasite that can cause severe disease in immunocompromised individuals and lead to fetal abnormalities if contracted during pregnancy. Establishment of protective immunity relies on CD8 T cells, which recognize antigenic peptides presented by MHC class I molecules on the surface of T. gondii-infected cells. Intriguingly, while the proteome of T. gondii is large, CD8 T cell responses target a very limited set of peptides. These peptides can be ranked according to the magnitude of the associated CD8 response (from immunodominant down to subdominant). Yet, little is known about the rules that define their immunogenicity and the hierarchy of the associated T cell responses. Using a panel of genetically modified T. gondii where the GRA6 dominant antigen was mutated, we show that the C-terminal location of the epitope within the source antigen is the critical parameter for immunodominance. Interestingly, when placed at the C-terminus of GRA6, the subdominant status of an epitope can be overturned. Our results unravel the mechanisms that make parasite antigens accessible for the MHC I presentation pathway. They may help to ameliorate natural immune responses and improve vaccine design against intravacuolar pathogens.
| CD8 T cells play a critical role in immune-mediated protection against intracellular apicomplexan parasites. Antigenic determinants recognized by CD8 T cells are short peptides of 8 to 10 amino acids presented by class I molecules of the major histocompatibility complex (MHC I). Antigenic peptides are typically degraded by cytosolic proteasomes, transported into the endoplasmic reticulum (ER), trimmed by ER-resident aminopeptidases and loaded on peptide-receptive MHC I molecules [1]. The spectrum of peptides that can theoretically be presented by a given MHC I is far larger than the peptides that actually elicit CD8 T cell responses. Furthermore, not all the peptide-MHC I complexes that can be recognized are equal: rather they elicit a hierarchy of specific CD8 T cells. This phenomenon of “selection and ranking” is termed immunodominance. Immunodominant peptide-MHC I elicit the most abundant cognate T cell populations, whereas subdominant peptide-MHC I induce less abundant T cells (reviewed in [2], [3]). Knowledge of the mechanisms that enhance immunogenicity and determine immunodominance hierarchy is central to design of optimal vaccines.
Mechanisms of immunodominance have been widely studied in the context of viral infections. The dominant position in the hierarchy has been positively correlated with 1) efficiency of peptide generation by the antigen processing pathway, e.g. due to proteasomal activity [4], ER aminopeptidase activity [5] or the nature of epitope-flanking sequences [6]), 2) antigen abundance [7], 3) ability of the antigen-presenting cells (APCs) to stimulate T cells, e.g. dendritic cells (DCs) versus non-professional APCs [8], 4) MHC binding affinity [4], [9] and 5) size of the naïve pool of specific T cells [9], [10], [11]. This latter parameter is increasingly being considered as a good predictor of immunodominance hierarchy, although, like the other parameters, it does not seem to be absolute [12].
During infection by intracellular parasites, the parameters that promote immunogenicity of a protein and that determine T cell immunodominance remain largely unknown. Unlike viruses, parasite-derived antigens are not synthesized by the host cell translation machinery, thus bypassing a preferential linkage between protein synthesis and MHC I presentation [13]. Moreover, except for antigens that may be directly injected into the host cytoplasm (e.g. T. gondii rhoptry proteins), most antigens from parasites that live in vacuoles are segregated from the cytosol by one or more membranes. We hypothesize that, despite the greater genomic complexity of apicomplexan parasites relative to viruses [14], these key differences could determine the limited number of hitherto characterized antigenic peptides from Plasmodium yoelii [15] Theileria annulata [16] or T. gondii [17] parasites.
In the present study, we addressed the causes and consequences of immunodominance during T. gondii infection. T. gondii is a widespread intravacuolar parasite that can cause severe disease in humans [18]. T. gondii replicates in a specialized parasitophorous vacuole (PV) and CD8 T cells play a protective role, especially against toxoplasmic encephalitis which is caused by the persistence of cysts in the brain [19]. We previously identified a decamer peptide (HF10, derived from the GRA6 protein) presented by the Ld MHC I molecule and recognized by a large CD8 T cell population during toxoplasmosis [17]. Two other epitopes, also presented by Ld, have been reported: the ROP7-derived IF9 and the GRA4-derived SM9 peptides [20]. Although the source antigens for each of these epitopes are contained in T. gondii secretory organelles, the GRA6-specific response appeared immunodominant based on its magnitude [17]. The molecular mechanisms for the potent immunogenicity of GRA6-derived HF10 epitope are not known.
We generated transgenic parasites that do or do not express the GRA6-derived HF10 epitope and established that even in the absence of the immunodominant GRA6-specific CD8 T cell response, the subdominant GRA4 and ROP7 responses remain poorly immunogenic and fail to protect mice from toxoplasmosis. We show that the location of the epitope at the C-terminus of the GRA6 antigenic precursor is a critical parameter that allows efficient processing, determines immunodominance and provides protection during chronic stage.
In order to study the pathophysiological consequences of HF10 immunodominance, we generated parasites that do or do not express the GRA6-derived HF10 epitope. We took advantage of the genetic diversity among three common T. gondii strains (type I, II and III). While the GRA4-derived SM9 and ROP7-derived IF9 peptides are conserved (data not shown and ToxoDB.org), the GRA6-derived HF10 peptide is polymorphic between type II and type I/III strains. Within the last 10 residues of GRA6, four non-synonymous single-nucleotide polymorphisms differentiate the type II sequence (HF10: HPGSVNEFDF) from the type I/III sequence (HY10: HPERVNVFDY) (Fig. 1A and Fig. S1 in Text S1). We noted that instead of a phenylalanine (F), the C-terminal residue in GRA6III is a tyrosine (Y), a residue not expected to be an appropriate anchor residue for Ld binding [21]. To evaluate the ability of HF10 and HY10 to bind to Ld, we used an MHC I stabilization assay. TAP-deficient RMA-S cells display empty, unstable MHC I molecules and addition of exogenous peptides that can bind to MHC I can stabilize their expression on the cell surface, as read out by flow cytometry. Expression of Ld on the surface was stabilized by addition of HF10 at a 1000-fold lower concentration compared to HY10 (Fig. 1B), which confirmed its poor Ld binding capacity.
Given that a type III strain like CEP, expresses HY10 (and not HF10), we inferred that it would provide a useful “HF10-null” background to analyze immunodominance in vivo. Therefore we engineered CEP parasites to stably express type II or (as a control) type I versions of GRA6 (designated CEP+GRA6II and CEP+GRA6I respectively). To facilitate tracking of parasites and infected cells, we used a CEP strain previously modified to express the GFP and the luciferase reporter genes [22]. We assessed the amount of transgenic GRA6 protein expressed by CEP+GRA6II and CEP+GRA6I parasites by immunoblot, using an antibody that detects all forms of GRA6 (I, II and III). The slower migration of GRA6II allowed us to discriminate between endogenous GRA6III and transgenic GRA6II. We confirmed expression of GRA6II, at slightly higher levels as compared to endogenous GRA6III in the same parasites. Transgenic GRA6I and endogenous GRA6III were undistinguishable thus precluding a precise analysis of the GRA6I transgene level. (Fig. 1C). Next, we infected B10.D2 mice (H2d MHC haplotype) with CEP+GRA6II and CEP+GRA6I parasites and 3 weeks post-infection, we measured the CD8 T cell response induced by HF10 and HY10 peptides. As observed with the type II Pru strain [17], nearly 25% of CD8 T cells from CEP+GRA6II-infected spleens produced IFN-γ in response to HF10 peptide. In contrast, no response was detected above background in CD8 T cells from CEP+GRA6I-infected mice after restimulation with the HF10 or HY10 peptides (Fig. 1D). We conclude that CEP strains are “HF10-null” and suitable for assessing the immunogenicity of various transgenes.
We used these transgenic parasites to confirm the immunodominance hierarchy among the 3 known natural peptides presented by Ld and to analyze the consequences of HF10 absence on T cell responses to the other antigens. Four weeks post-infection, we examined the T. gondii-specific CD8 T cell response in the spleen using peptide-loaded Ld dimers (Fig. 2A). As expected, a large fraction of HF10-specific CD8 T cells were detected only in CEP+GRA6II-infected mice (9.5%+/−3.7%, mean +/− s.d., p = 10−4). The IF9-specific CD8 T cells were found at a much lower frequency (0.9+/−0.8%, p = 0.11) and SM9-specific CD8 T cells were hardly detectable. Interestingly, even in the absence of HF10 (such as in CEP+GRA6I-infected mice), the frequencies of IF9- and SM9-specific CD8 T cells did not increase, suggesting that the subdominant status of IF9 and SM9 was not the result of competition (also called immunodomination) exerted by HF10-specific T cells. Similar results were obtained when we assessed IFN-γ production by effector CD8 T cells following in vitro restimulation (Fig. 2B). Likewise, the epitope-specificity of CD8 T cells in brain infiltrates showed the same HF10>IF9>SM9 hierarchy (Fig. 2C).
While the above experiments define the immunodominance hierarchy among already known epitopes, unknown epitopes could also play a role in the parasite-specific response. To analyze the entire repertoire of T. gondii-specific CD8 T cells, we used parasite-infected, rather than peptide-pulsed, APCs to restimulate T cells ex vivo. The magnitude of IFN-γ response elicited by parasite-infected macrophages (Fig. 2D) was no higher than that observed after peptide restimulation (see Fig. 2B). Thus, CD8 T cells of other specificities do not play a major role in our experimental model system.
Taken together, our data demonstrate that the presence of GRA6II in the parasites triggers a strong and dominant HF10-specific CD8 response in the spleen and brain of chronically infected animals but does not negatively affect (“immunodominate”) the subdominant SM9 and IF9 responses.
We have previously reported that immunization of H2d mice with HF10-loaded bone marrow-derived dendritic cells protects against lethal type II parasite challenge [17]. We predicted that presence of HF10 may decrease parasite burden. To test this hypothesis, we took advantage of the luciferase expression to analyze parasite dissemination by bioluminescence imaging in BALB/c mice (H2d). Regardless of the presence of HF10, all strains were cleared by day 13 (Fig. 3A,B). CEP+GRA6II parasites appeared to be cleared slightly earlier than control CEP (HXGPRT+) and CEP+GRA6I parasites, although this difference did not reach statistical significance (Fig. 3A,B). In addition, parasite signal in the brain was detected only in mice infected with control CEP or CEP+GRA6I and was never observed with CEP+GRA6II (Fig. S2 in Text S1). This suggested that the control of parasitemia by HF10-specific T cells may be more effective at the chronic stage, when parasites are found mostly as brain cysts. When we measured parasitemia in chronically infected B10.D2 mice at 4 weeks post-infection, we found a significantly higher proportion of splenocytes harboring parasites in CEP+GRA6I-infected mice (Fig. 3C). Accordingly, the number of brain cysts in CEP+GRA6I-infected mice was nearly 5 times higher than in mice infected with the HF10-expressing parasites (Fig. 3D). These results could not be attributed to an intrinsic growth difference between clones since they behaved comparably in a plaque assay in vitro (Fig. S3A in Text S1). Furthermore, the influence of HF10 on cyst number was visible only in B10.D2 mice and not in C57BL/6 mice, which have a different MHC haplotype (H2b) and therefore do not elicit HF10-specific T cells (Fig. S3B in Text S1). Combined, the data show that the HF10-specific response has a modest protective effect on parasite control during acute toxoplasmosis but is essential for controlling parasite load during chronic infection.
To uncover possible causes of HF10 immunodominance, we used the MHC I stabilization assay described above (see Fig. 1B) and compared the affinity for Ld of HF10 to other Ld-restricted peptides. These other peptides were derived either from T. gondii (SM9, IF9), from a mouse minor antigen (QL9) or from a mouse cytomegalovirus protein (YL9) (Fig. 4A). HF10 affinity appeared ∼10-fold higher than that of IF9, QL9 and YL9 but fell in the same range as the T. gondii subdominant peptide SM9. Therefore, the dominance hierarchy did not correlate with peptide affinity for MHC I.
We next assessed whether abundance of peptide-specific T cells in the repertoire of naïve mice may control immunodominance. We employed a tetramer-based enrichment procedure [23], [24] to enumerate naïve T cells isolated from spleen and lymph nodes of uninfected mice and specific for each of the 3 epitopes. Numbers of T cell precursors were analyzed by flow cytometry after gating on a population of live dump− (dump = B220, F4/80, MHC II) CD3+ CD8α+ cells (Fig. S4 in Text S1). Surprisingly, we observed an inverse correlation between the number of naïve T cells per mouse and the immunodominance (Fig. 4B), with the frequency of HF10-specific CD8 T cells around 10- and 3-fold lower than SM9- and IF9-specific CD8 T cells respectively (Fig. 4C). Thus the immunodominance of HF10 cannot be explained by a high starting precursor frequency of specific T cells.
Having ruled out two plausible hypotheses, we wondered whether HF10 immunodominance might be related to processing efficiency. We noted that HF10 is located at the very C-terminus of GRA6II, a position that may facilitate processing since no C-terminal cut is required. To test the importance of epitope position, we changed the C-terminus of HF10 by extending GRA6II with one or more amino acids.
We first transfected C-terminally extended versions of GRA6II in mouse fibroblasts. Extensions were either single amino acids (lysine, K ; leucine, L ; proline, P) or several residues such as the GRA6I/III-derived HY10 peptide or the entire GFP. We used CTgEZ.4 T cell hybridomas, a β-galactosidase-inducible reporter cell line [17], to read-out HF10 presentation. The response of CTgEZ.4 T cells was mildly decreased (GRA6II-K), severely disrupted (GRA6II-L) or totally abrogated (GRA6II-P, GRA6II-HY10, GRA6II-GFP) (Fig. 5A). These data suggest that the C-terminal location determines optimal processing and presentation of the HF10 peptide when the precursor protein is expressed ectopically by the antigen-presenting cell.
To assess the impact of HF10 position in T. gondii, we used CEP parasites expressing longer versions of GRA6II, extended either by a leucine (GRA6II-L) or by the HA tag (GRA6II-HA). First, we verified that transgene levels were comparable by Western blot (Fig. 5B). To investigate whether these additional C-terminal residues might perturb GRA6 transport, we took advantage of the HA tag and evaluated the subcellular distribution of transgenic GRA6-HA, as compared to total GRA6. Analysis of the overlap between HA and the GRA2 and GRA5 dense granule proteins in extracellular tachyzoites (Fig. S5A,B in Text S1) and in infected fibroblasts (Fig. S5C,D in Text S1), indicated that GRA6II-HA is packaged in the dense granules and secreted in the vacuolar space, as known for wild-type GRA6 [25]. Although the distribution of GRA6II-L could not be directly assessed, we inferred from the above data that the extra leucine did not alter protein transport either. When used to infect bone marrow-derived macrophages (BMDMs), the CEP+GRA6II-L and CEP+GRA6II-HA transgenic parasites led to similar infection rates (data not shown) but HF10 presentation was abrogated (Fig. 5C).
Finally, we examined the importance of HF10 C-terminal location in vivo. We infected mice and analyzed the induction of HF10-specific CD8 T cells in the spleen (Fig. 5D) and the brain (Fig. 5E) at chronic stage. In accordance with our in vitro findings, only the CEP+GRA6II parasites elicited a detectable HF10-specific response. The absence of HF10-specific response in mice infected by CEP+GRA6II-L and CEP+GRA6II-HA was consistent with a dramatically higher cyst burden in their brains (Fig. 5F). We conclude that the precise C-terminal location of HF10 is required for optimal processing and presentation by T. gondii-infected APCs and for eliciting parasite-specific T cells that could provide in vivo protection.
We further assessed whether the location of an epitope at the GRA6II C-terminus may be sufficient for enhancing presentation and immunogenicity. We generated CEP parasites expressing the subdominant SM9 peptide either at the C-terminus of GRA6II (CEP+GRA6II-SM9Cter) or, as a control, internally within GRA6II (CEP+GRA6II-SM9internal) (Fig. 6A). The selected clones expressed comparable levels of transgenes (Fig. 6B). We measured SM9 presentation at the surface of parasite-infected cells using a new β-galactosidase-inducible T cell hybridoma specific for Ld-SM9 complex (BDSM9Z) (Fig. 6C). Interestingly, although the natural SM9 precursor, GRA4, was expressed in type III parasites (data not shown), the presentation of Ld-SM9 complexes remained below detection in CEP-infected BMDMs. SM9 presentation was also undetectable when the peptide was placed at an internal position within GRA6. In contrast, BDSM9Z T cells were strongly stimulated when SM9 was located at GRA6 C-terminus (Fig. 6C). HF10 presentation by infected BMDMs was abrogated by the presence of SM9 at the C-terminus but not by the presence of SM9 at the internal position (Fig. 6D), consistent the C-terminal extension studies described above (Fig. 5C),. In conclusion, placing the SM9 peptide at the C-terminus of GRA6II was sufficient to enhance its presentation by parasite-infected cells in vitro.
We next measured the SM9 and HF10-specific CD8 T cell responses in the spleen (Fig. 6E) and brain (Fig. 6F) of mice infected for 3 weeks with the transgenic parasites. A SM9-specific response was hardly detectable in mice infected with the control CEP. Infection with CEP+GRA6II-SM9internal elicited SM9-specific CD8 T cells but these T cells were between 3-fold (spleen, Fig. 6E) and 5-fold (brain, Fig. 6F) more abundant when SM9 was grafted at GRA6 C-terminus. This difference was not due to reduced infectivity of the CEP+GRA6II-SM9internal parasites since HF10-specific CD8 T cells were abundant in those mice (Fig. 6E,F). Similar results were obtained in mice immunized with irradiated tachyzoites (Fig. S6 in Text S1). To ask whether the enhanced SM9-specific response participates in parasite control, we enumerated the brain cysts in the 3 groups. As compared to mice infected with control CEP, the parasite load was lower when either a strong HF10- or a strong SM9-specific response was elicited (Fig. 6G). These data indicate that the nature of the antigenic peptide itself does not seem to determine the protective effect. We conclude that location of a subdominant peptide at GRA6 C-terminus dramatically enhanced its immunogenicity, changed the epitope hierarchy and had beneficial repercussions for parasite control.
In this study, we have identified the molecular bases underlying the marked immunodominance of a CD8 T cell response that controls the intracellular T. gondii parasite. Rather than peptide affinity for MHC I and naïve T cell frequency, we find that immunodominance is determined by the location of the epitope within the antigenic precursor.
The endeavor to characterize natural T cell antigens from T. gondii has started only recently [17], [20], [26], [27] but it has provided much needed tools to better understand T cell immunity to this widespread opportunistic pathogen. We report here that the 3 known Ld-restricted responses follow an immunodominance hierarchy. At chronic stage, GRA6II-specific CD8 T cells were between 10-fold (in the spleen) and 30-fold (in the brain) more abundant than CD8 T cells specific for the 2nd dominant epitope: IF9 derived from ROP7. Response to the 3rd dominant epitope, SM9 derived from GRA4, was hardly detectable. Remarkably, we did not observe immunodomination by the GRA6II dominant epitope. Immunodomination refers to situations in which the T cell response to a given epitope is inhibited by T cells specific for another, more dominant, epitope [2]. This phenomenon has been reported during infection by simian immunodeficiency virus [28] and by Trypanosoma cruzi, another protozoan parasite phylogenetically related to T. gondii [29]. A mechanism commonly proposed to explain immunodomination is elimination of APCs by the dominant cytotoxic T cells. Perhaps immunodomination does not occur here because, as compared to IFN-γ production, perforin-mediated cytolysis by CD8 T cells plays only a limited role during T. gondii infection [30]. The absence of immunodomination also suggests that accessibility of peptide-loaded APCs for T cells is not limiting. This may be because T. gondii is able to invade and be presented on MHC I by many cell types, even non-professional APCs [31].
Another major conclusion is that during chronic stage, subdominant responses could not compensate and provide efficient parasite control in the absence of the GRA6II dominant peptide. These data designate GRA6 as a strain-specific component which determines chronic parasitemia and is targeted by adaptive immunity. This is in contrast to already known T. gondii virulence factors which mostly interfere with innate processes, such as ROP16 which interferes with STAT transcription factors [32], ROP18 which disarms immunity-related GTPases involved in host defense [33], [34]) or GRA15 which promotes NF-κB activation [35]. Of note, GRA6 is among the 20 most polymorphic genes in the T. gondii genome and many polymorphisms are located in its C-terminal region (see ToxoDB.org and Fig. S1 in Text S1). Beyond the 3 prototypic strains, sequence polymorphisms in GRA6 have been characterized in more exotic strains (or haplogroups) [36]. These atypical strains either express HF10, HY10 or alternative versions of the decamer peptide with distinct polymorphisms. Interestingly, in chronically infected humans, some of these variations are specifically recognized by natural antibodies that are used as a tool to serotype the parasite [37], [38]. Given that GRA6 C-terminus is targeted both by humoral and cellular responses, we speculate that selective pressure by adaptive immunity may have contributed in shaping GRA6 polymorphisms.
By exploring the possible causes of HF10 immunodominance, we were able to rule out two possible explanations. We found no positive correlation of immunodominance hierarchy with peptide affinity for Ld or naïve T cell frequency. The numbers of naïve HF10- and IF9-specific T cells fall within the previously reported range of 15 to 1500 naïve CD8 T cells per mouse [39] but the number of SM9-specific cells (4300) may look unusually high. Although the exact reason remains unclear, this may be related to a large amount of positive selecting ligands available for this population. In agreement with this idea, an F1 H2bxd mouse strain gives half the value measured with the B10.D2 H-2d strain (data not shown). Remarkably, we found an inverse correlation between size of the naïve population and magnitude of the parasite-specific response. This latter result may seem paradoxical in the light of other situations, such as peptide immunization [23] or viral infection [7], [9], [10], [24], where size of the naïve pool was a good predictor of immunodominance. However, there are known exceptions to this rule [12]. Here, the low abundance of HF10-specific precursors may facilitate their expansion by limiting interclonal competition. Alternatively, the TCRs used by HF10-specific CD8 T cells may have a high affinity for the HF10-Ld complex, which could promote stronger signaling and proliferation. These hypotheses remain to be investigated.
Our central finding is that C-terminal position of the epitope within GRA6 plays a crucial role for immunodominance. It is illustrated by the fact that the weakly immunogenic GRA4 epitope elicited strong SM9-specific responses when grafted at GRA6 C-terminus. Intriguingly, the internal position of SM9 gave rise to a lower, but substantial, CD8 response whereas we did not detect any HF10-specific T cells when HF10 was placed at the same internal position (Fig. S7 in Text S1). The bases for such a different outcome are unknown but they may lie in different epitope-dependent processing efficiencies or compensatory effects of the high frequency of SM9-specific T cell precursors counteracting the less favorable position for processing. Most notably, addition of residues to GRA6II C-terminus greatly impaired presentation. It is theoretically possible that the C-terminal flanking sequences abrogated HF10 presentation by altering the vacuolar trafficking of GRA6 and/or its membrane insertion. We think it is unlikely because 1) we found that transport of GRA6II-HA was undistinguishable from transport of total GRA6 and 2) presentation of the C-terminal SM9 peptide could occur efficiently in vitro and in vivo. Rather than regulating global GRA6 trafficking, we favor the idea that additional C-terminal sequences impair processing. Epitope-flanking sequences are indeed known to positively or negatively affect protease cleavage capacity and generation of the final peptide, with clear consequences on immunodominance [4], [6], [40], [41], [42]. Using minigenes, it was shown that the nature of C-terminal flanking residues profoundly impacts excision of the processed peptide [42]. In the context of a full-length viral protein, single changes to the epitope-flanking residues dramatically reduced presentation [41] and it was later proposed that the subdominant nature of certain peptides bearing appropriate consensus motifs might result from suboptimal C-terminal sequences [40]. Finally, it is interesting to note that the influence of the flanking motifs may differ whether the antigen is presented by the direct MHC I pathway or by cross-presentation [43]. In our case, we observed a dramatic impact of the absence or presence of C-terminal residues on presentation, suggesting a key role for antigen processing in modulating immunodominance. A systematic screening of C-terminal extensions may be useful to precisely define the rules that govern processing of GRA6 C-terminus.
Since GRA6 behaves as an integral transmembrane protein in the vacuole [44], the importance of the C-terminus could be related to the topology of GRA6 membrane insertion. One possibility is that GRA6 C-terminal domain is displayed in the cytosol and thus potentially accessible to host proteases. This mechanism was proposed for antigens from the intravacuolar bacteria Chlamydia trachomatis that are inserted in the surrounding membrane of the bacteria-containing vacuole [45], [46]. This hypothesis remains to be tested. Alternatively, unfolded GRA6 may access the cytosol thanks to the recruitment of host endoplasmic reticulum components on the parasitophorous vacuole, as proposed for the soluble OVA model antigen [47], [48]. In any case, we consider it likely that access of GRA6 to the MHC I pathway is less efficient than in the situation of a viral antigen directly synthesized by the host cell translation machinery. Consequently, any parameter that would facilitate processing (e.g. being at C-terminus) may become the determining factor for the presentation outcome.
To our knowledge, this is the first evidence that C-terminal position can be positively correlated with immunodominance. Given the variety of parameters that can influence immunodominance, a remaining question is the degree of peculiarity of our current findings with respect to other antigens. A recent study interrogating the Immune Epitope Database (www.iedb.org) for a positional bias of viral epitopes reported that epitopes from both ends of a protein tended to be underrepresented [49]. An indirect way to assess the general relevance of the C-terminal position would be to transfer subdominant epitopes to the C-terminus of their respective antigens (e.g. GRA4, ROP7) and evaluate the impact on CD8 responses. Future studies, not only with T. gondii but also with other intracellular parasites, should shed light on the general relevance of this position.
During T. gondii infection in vivo, two scenarios of MHC I presentation could co-exist. On the one hand, phagocytosed parasite material may be processed by bystander cells present in the vicinity of infected cells [50]. On the other hand, parasite proteins may be directly presented by actively infected cells [51]. Our work shows that antigen access to the MHC I pathway and efficient processing are the limiting factors that control immunodominance. Beyond amino acid mutations within the peptide sequence, modifying the epitope position may provide the parasite with a strategy to manipulate how it is detected by CD8 T cells. Understanding the features that make certain peptides immunogenic will shed light on the strategies used by parasites to interact with their host immune system.
In the US, animal studies were carried out in accordance with the recommendations of the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health and in compliance with the guidelines of the Institutional Animal Care and Use Committee (IACUC) of the University of California. The animal protocols were approved by the IACUC of the University of California, Berkeley (Animal Use Protocol # R057-0913BR) and of the University of Pittsburgh, PA (Protocol # 1210130). In France, animal studies were carried out under the control of the National Veterinary Services and in accordance with European regulations (EEC directive 86/609 dated 24 November 1986). The protocol was approved by the Regional Ethics Committee from the Midi-Pyrénées Region (Approval # MP/01/29/09/10).
C57BL/6J (B6), BALB/c and B10.D2-Hc1 H2d H2-T18c/nSnJ (B10.D2) mice were purchased from The Jackson Laboratory (Bar Harbor, ME, USA). B6xDBA/2 F1 (B6D2) mice were purchased from Charles River (France). For all experiments, sex and age-matched mice were used. Mice were handled with the approval of local ethics committees. Except for the CEP+GRA6II-HA clone which was a gift from J. Saeij (Cambridge, MA, USA), all transgenic parasites were generated from the parental CEP.ΔHXGPRT.GFP.Luc strain [22]. Tachyzoites were maintained by passage on confluent monolayers of human foreskin fibroblasts (HFF). For infections, parasites were harvested, filtered through 3 µm and 105 tachyzoites were injected intraperitoneally in 100 µl PBS.
For in vitro expression of antigenic sequences, all C-terminally extended GRA6II sequences were cloned into the pcDNA1 vector containing the pcDNA1-embedded 3′UTR. Plasmids used for T. gondii transfection were derived from the pGRA.HA.HPT vector, a gift from J.D. Dunn and J. Boothroyd (Palo Alto, CA, USA). More details on the construction of the plasmids are given in the Supporting Protocol S1 in Text S1.
L cells were triple transfected using a standard diethylaminoethyl dextran method with vectors coding for Ld, B7-2 and mutant GRA6, as previously described [17]. For parasite transfections, 1.5×107 tachyzoites were electroporated with 50 µg of HindIII-linearized plasmid DNA and inoculated in 4 confluent HFF flasks in order to obtain up to 4 independent clones. The next day, 25 µg/ml mycophenolic acid and 50 µg/ml xanthine were added for selection. After 2 passages, resistant tachyzoites were cloned by limiting dilution and presence of the transgene was verified by PCR. For each construct, one clone that acquired resistance but no transgene was kept as HXGPRT+ control.
RMA-S.Ld cells were a gift from T. Hansen (St-Louis, MO, USA). RMA-S.Ld cells were incubated at 37°C, 5% CO2 for 8 h to saturate the culture medium with CO2. The flask was sealed with parafilm and incubated at RT overnight. The next day, cells were washed with PBS and plated at 3×105 cells/well in a 96-W plate. Peptide was added to the cells in serial dilutions. The plate was incubated for 1 h at RT and 3 h at 37°C. Cells were stained with the 30-5-7 antibody (specific for conformed, peptide-bound Ld) and a phycoerythrine (PE)-coupled goat anti-mouse secondary antibody and analyzed by flow cytometry.
HFFs were disrupted with a 23-G needle and tachyzoites were lysed in a lysis buffer containing 1% NP-40, 10 mM Tris pH 7.4, 150 mM NaCl, and protease inhibitors (cOmplete EDTA-free, Roche) for 30 min on ice. Lysates were centrifuged for 15 min at 15,000 g. Solubilized proteins were boiled and reduced for 5 min in SDS sample buffer, separated by electrophoresis on 12% polyacrylamide gels and transferred to nitrocellulose membranes. Immunologic detection was achieved using rabbit anti-GRA6 serum (gift from L. D. Sibley, St-Louis, MO, USA), mouse anti-HA (gift from D. Raulet, UC Berkeley, CA, USA) or mouse anti-SAG1 (clone TP3, Santa Cruz) followed by secondary horseradish-peroxidase-conjugated antibodies. Peroxidase activity was visualized by chemiluminescence.
Mice were sacrificed 3 to 4 weeks after infection. Spleens were dissociated into single-cell suspensions in complete RPMI (Invitrogen) supplemented with 10% (vol/vol) FCS (Hyclone). Samples were depleted of erythrocytes with ACK lysis buffer (100 µM EDTA, 160 mM NH4Cl and 10 mM NaHCO3). Leukocytes from the brain were prepared as in [17]. In brief, brains were minced and digested for 1 h at 37°C with 1 mg/ml collagenase (Sigma) and 100 µg/ml DNAseI (Roche) in complete RPMI. Brain suspensions were filtered through 70-µm cell strainers and centrifuged for 10 min at 200 g. Cells were resuspended in 60% (vol/vol) Percoll (GE Healthcare), overlaid on 30% (vol/vol) Percoll and centrifuged 20 min at 1,000 g. Infiltrating mononuclear cells were collected from the gradient interface and the remaining erythrocytes were lyzed with ACK lysis buffer.
The number of infected splenocytes was determined by measuring the percentage of GFP+ cells by flow cytometry. Results from two samples with over 2×105 events collected per tube were averaged for each mouse. For cyst enumeration, the brain was homogenized over a 100 µm strainer and 5% of the entire brain was stained with fluorescein-conjugated Dolichos biflorus agglutinin (Vector Laboratories). Cysts were counted using an inverted fluorescence microscope. For bioluminescence imaging, BALB/c mice were infected intraperitoneally with 105 tachyzoites of CEP expressing GRA6I or GRA6II or control CEP HXGPRT+. Parasite burden was assessed using in vivo bioluminescence imaging as described previously [52]. Briefly, daily readings were performed using an IVIS Lumina II imaging system (Caliper). Ten minutes prior to imaging, mice were injected intraperitoneally with 200 µL of 15.4 mg/mL D-Luciferin in PBS and anesthetized using 2% isoflurane. Dorsal and ventral images were acquired for 5 minutes and luminescence (photons/s/cm2/sr, total flux expressed as photons/s) was quantified using IgorPro Image Analysis Software (Caliper).
Bone marrow cells were obtained from mouse femurs and tibias. Primary BMDMs were differentiated for 7 days in Petri dishes with RPMI supplemented with 20% (vol/vol) FCS and 10% (vol/vol) colony-stimulating factor–containing culture supernatant (purity, about 95% CD11b+). Colony-stimulating factor-producing 3T3 cells were a gift from R. Vance (UC Berkeley, CA, USA). BMDMs were infected for 24 h with γ-irradiated tachyzoites (120 Gy) at various multiplicities of infection and used in antigen presentation assays. In all experiments, the proportion of infected (GFP+) BMDMs was controlled by flow cytometry.
B6D2 F1 mice were immunized subcutaneously with 100 µg synthetic SM9 peptide in complete Freund's adjuvant and boosted after 2 weeks. One week later, spleens were harvested and restimulated with 10 nM SM9. Recombinant human IL-2 (50 U/ml) and 5% T-stim (both from BD Pharmingen) were added after day 2 to support CD8 T cell proliferation. Four days after restimulation, responding cells were fused to the TCRαβ-negative lacZ-inducible BWZ.36/CD8α fusion partner as described in [17]. Specificity of the resulting BDSM9Z hybridomas was tested by overnight incubation with peptide-pulsed or unpulsed Ld-transfected L cells. TCR-mediated stimulation of the BDSM9Z and the CTgEZ.4 hybridomas [17] was quantified using a chromogenic substrate: chlorophenol red-β-D-galactopyranoside (CPRG, Roche). Cleavage of the CPRG by β-galactosidase releases a purple product, which absorbance was read at 595 nm with a reference at 655 nm.
Spleen and major lymph nodes from individual naïve B10.D2 mice were harvested. Single cell suspension was stained with PE-labeled (Prozyme) SM9:Ld, IF9:Ld or HF10:Ld tetramers (NIH tetramer facility). Tetramer enrichment was performed on each sample with anti-PE magnetic beads (Miltenyi Biotec) and each sample was stained with antibodies (BD Biosciences) for flow cytometry analysis. Total numbers of CD8α+tetramer+ T cells per mouse were determined as before [23], [24].
For other stainings, surfaces were labeled according to standard procedures with flow cytometry buffer (3% (vol/vol) FCS and 1 mM EDTA in PBS). Intracellular IFN-γ was detected with the Cytofix/Cytoperm kit (BD Pharmingen). DimerX H-2Ld:Ig (fusion protein of H-2Ld and immunoglobulin; BD Biosciences) was used according to the manufacturer's instructions and as described in [17]. All flow cytometry data were acquired on an XL Analyzer (Coulter) or a LSRII (Becton Dickinson) and were analyzed with FlowJo software (Tree Star).
Prism software (GraphPad) was used for statistical analyses. All P values were calculated with the two-tailed Mann-Whitney test (nonparametric).
|
10.1371/journal.ppat.1007896 | Recognition of specific sialoglycan structures by oral streptococci impacts the severity of endocardial infection | Streptococcus gordonii and Streptococcus sanguinis are primary colonizers of the tooth surface. Although generally non-pathogenic in the oral environment, they are a frequent cause of infective endocarditis. Both streptococcal species express a serine-rich repeat surface adhesin that mediates attachment to sialylated glycans on mucin-like glycoproteins, but the specific sialoglycan structures recognized can vary from strain to strain. Previous studies have shown that sialoglycan binding is clearly important for aortic valve infections caused by some S. gordonii, but this process did not contribute to the virulence of a strain of S. sanguinis. However, these streptococci can bind to different subsets of sialoglycan structures. Here we generated isogenic strains of S. gordonii that differ only in the type and range of sialoglycan structures to which they adhere and examined whether this rendered them more or less virulent in a rat model of endocarditis. The findings indicate that the recognition of specific sialoglycans can either enhance or diminish pathogenicity. Binding to sialyllactosamine reduces the initial colonization of mechanically-damaged aortic valves, whereas binding to the closely-related trisaccharide sialyl T-antigen promotes higher bacterial densities in valve tissue 72 hours later. A surprising finding was that the initial attachment of streptococci to aortic valves was inversely proportional to the affinity of each strain for platelets, suggesting that binding to platelets circulating in the blood may divert bacteria away from the endocardial surface. Importantly, we found that human and rat platelet GPIbα (the major receptor for S. gordonii and S. sanguinis on platelets) display similar O-glycan structures, comprised mainly of a di-sialylated core 2 hexasaccharide, although the rat GPIbα has a more heterogenous composition of modified sialic acids. The combined results suggest that streptococcal interaction with a minor O-glycan on GPIbα may be more important than the over-all affinity for GPIbα for pathogenic effects.
| Infective endocarditis (IE) is a life-threatening infection of heart valves, and streptococci that normally reside in the mouth are a leading cause of this disease. Some oral streptococcal species express a protein on their surface that enables attachment to glycan (sugar) modifications on saliva proteins, an interaction that may be important for colonization of the tooth and other oral surfaces. These "Siglec-like adhesins" are hypervariable in the type and number of glycan structures they bind, ranging from just one to more than six of the structures displayed on the saliva proteins. If streptococci enter into the bloodstream, the Siglec-like adhesin can mediate attachment to similar glycans that decorate platelet or plasma proteins, which can impact the overall virulence of the organism. This study highlights how recognition of a specific type of glycan structure can cause a generally beneficial or neutral microbe to create damage to specific tissues—in this case the heart valves, illustrating one means by which commensal bacteria can become opportunistic or accidental pathogens. The findings further indicate that certain glycan-binding streptococci among the oral microbiota may be predisposed to produce infective endocarditis.
| Infective endocarditis (IE) is a life-threatening cardiovascular disease in which microbes colonize and persist in platelet-fibrin thrombi on cardiac valve surfaces. The interaction of bacteria with platelets is thought to play a central role in the pathogenesis of IE [1, 2]. Most bacterial species are unable to colonize an intact cardiac valve endothelium, but instead attach to platelet-fibrin thrombi or "sterile vegetations" that have deposited on damaged valve surfaces [3–5]. The subsequent deposition of platelets onto the infected endocardium, along with bacterial proliferation, contributes to the progression of disease, and results in the formation of macroscopic endocardial lesions [6–8].
Streptococcus gordonii and Streptococcus sanguinis are oral commensal bacterial species that are primary colonizers of tooth surfaces [9]. Although generally associated with oral health, these closely-related species are frequently found as the causative agent of infective endocarditis, especially infections of the aortic valve [10–14]. Only a small number of virulence factors of S. sanguinis or S. gordonii that contribute to IE have been verified using animal models of this disease [15–22]. Among the best characterized for S. gordonii are the platelet-binding proteins GspB and Hsa, expressed by strains M99 and DL1, respectively. These cell wall anchored adhesins are two members of the highly-conserved family of serine-rich repeat (SRR) glycoproteins expressed by Gram-positive bacteria (Fig 1). The ligand-binding regions (BRs) of the SRR glycoproteins are modular and often species-specific [23, 24]. SRR glycoprotein sequences have been found in the genomes of all S. sanguinis and S. gordonii strains sequenced to date [25], and invariably contain "Siglec-like" BRs that confer high-affinity binding to α2–3 linked sialic acid [23, 26]. This sialoglycan modification is displayed at the termini of O-glycans that decorate the salivary mucin MUC7 [27, 28], and binding of S. gordonii and S. sanguinis to MUC7 is thought to be important for oral colonization. In addition, previous studies indicate that when oral streptococci enter the bloodstream, binding to similar O-glycans on platelet GPIbα (the receptor for von Willebrand factor, or vWF) can contribute to the pathogenesis of IE [29, 30].
The Siglec-like BRs are an intriguing group of hypervariable adhesive domains, displaying both conserved and divergent features (Fig 1). They all contain Siglec and Unique domains that are important for sialoglycan binding [23, 26, 30–32]. The BRs of some S. sanguinis and most S. gordonii SRR adhesins, such as GspB, also include a CnaA domain, but this region appears not to have a role in sialoglycan binding [23]. The Siglec domain has a V-set Ig fold resembling that of mammalian Siglecs, and includes a conserved "YTRY motif" that makes important contacts with Neu5Acα2-3Gal at the termini of larger glycans [31, 32]. The Unique domain does not appear to make direct contacts with sialoglycans, but may modulate the conformation and thus influence the binding properties of the Siglec domain. Despite a conserved structural fold, the Siglec domain sequences can vary by more than 50%, and both small and large sequence variations can impact the number and type of sialoglycan structures recognized. Specific glycan targets have been identified for nearly a dozen of the Siglec-like BRs, and the ligand repertoires range from a single type of sialylated trisaccharide, to a broad set of related sialoglycans [23, 26]. For example, GspB is highly selective for sialyl T-antigen (sTa) [23, 26], whereas the 10712BR (from the SRR adhesin of S. gordonii UB10712) preferentially binds 3'sialyllactosamine (sLn; Fig 2) [23]. Hsa has a broader ligand range and can bind both sTa and sLn [23, 26]. The differences in binding to defined, synthetic glycans are also reflected in the interaction with O-glycosylated plasma proteins [33]. GspB most readily binds proteins bearing sTa (a core 1 O-glycan; Fig 2), while 10712BR prefers proteins with sLn at the termini of larger, branched, and often extended core 2 glycans. The ligand repertoire also impacts the strength of binding of the recombinant BRs to platelet GPIbα, with binding to sLn generally conferring a higher affinity for platelets and GPIbα compared with binding to sTa [23, 33]. As measured by surface plasmon resonance, the affinity of recombinant GspBBR for GPIbα is 2.38 × 10−8 M, whereas HsaBR has approximately 5-fold higher affinity (KD values of 3.05 × 10−8 M and 5.05 × 10−9 M when fit to a heterogenous ligand model, which is consistent with the ability to bind two glycan moieties) [29].
The role of the Siglec-like adhesins and sialoglycan binding in streptococcal endocarditis is not entirely clear. Deletion of gspB or hsa from S. gordonii strains, or even just a single amino acid substitution in the YTRY motif of GspB (GspBR484E), resulted in two-log lower levels of bacteria in aortic valve vegetations using a rat model of IE [15, 16, 30]. In contrast, deletion of srpA did not adversely impact the virulence of S. sanguinis SK36 in a rabbit model of IE [21]. Although the precise sialoglycan ligand for SrpA has not been determined, it does not readily bind sTa, but instead may recognize a core 2 hexasaccharide or larger di-sialylated O-glycan [32, 33]. Supporting the possibility that the type of sialoglycan recognized might influence disease progression, S. gordonii strain SK12 was found to be significantly less virulent than S. gordonii DL1 in a rat model of IE [34]. SK12 encodes an SRR glycoprotein with a BR identical to that of the 10712BR (see legend to Fig 1), and thus is likely to bind sLn rather than sTa. Additional analysis of Siglec-like BRs from a small number of streptococcal strains suggested that IE and commensal strains might bind different glycan structures, in that IE isolates were more often GspB-like, whereas oral isolates were more SrpA-like [23, 25]. However, the question of whether binding to a particular sialoglycan structure, versus sialic acid binding in general, affects the propensity of bacteria to establish endovascular infections has never been formally assessed. In this study, we generated a set of isogenic strains that display distinctly different sialoglycan binding properties and different levels of binding to platelets. We then compare the relative virulence of these strains in two rat models of IE. The results indicate that the sialoglycan binding spectrum can impact the overall virulence of streptococci, displaying different effects on the initial colonization of aortic valves, as well as the post-colonization progression of endocardial infection.
Our first goal was to generate isogenic variants of S. gordonii strain M99 that differ in their sialoglycan binding phenotypes. We selected three BRs that were previously determined to have distinctly different binding properties (Fig 2): 1) GspBBR demonstrates sTa selectivity (core 1 O-glycans), 2) the 10712BR has high affinity for sLn and related structures (core 2 O-glycans), and 3) HsaBR shows high-affinity binding to both sTa and sLn (core 1 and core 2 glycans). The design of these isogenic strains was not trivial, since SRR glycoprotein expression relies on a complex and highly specialized system that coordinates post-translational modification and transport to the bacterial cell surface. For example, in S. gordonii M99 and Streptococcus parasanguinis FW12, elements in the preprotein mature region, as well as the N-terminal signal peptide, must be matched to the dedicated SecA2/Y2 transporter [35–38]. It was also important to avoid any alterations in the flanking SRR regions, since the post-translational modification of these domains can impact binding [39–43]. In view of these issues, we chose to replace the entire BR of GspB with that of Hsa, or with the 10712BR, using the conserved SRR1-BR and BR-SRR2 junctions (Fig 3A), while retaining the native GspB signal peptide, AST, SRR1 and SRR2 domains. To ensure native expression levels in vivo we opted to replace a portion of the gspB gene in the native chromosomal locus, using a "knock in" strategy previously used to generate single amino acid substitutions in the YTRY motif of the Siglec domain (Fig 3B). This resulted in strains PS3515 (GspB::HsaBR) and PS3516 (GspB::10712BR). Importantly, the variant strains showed growth rates and cell-surface SRR glycoprotein expression levels (i.e. SDS migration patterns and western blot intensity) that were indistinguishable from the parental M99 strain (Figs 4A and 4B and S1).
We next examined bacterial binding to synthetic sialoglycans or to immobilized human platelets. The binding of these strains to sialoglycans resembled that of the respective recombinant BRs: M99 readily bound to sTa but not sLn, PS3515 bound both sialoglycan structures, and PS3516 bound sLn rather than sTa (Fig 4C). Likewise, strain PS3516 showed higher levels of binding to platelets as compared with the parental strain M99 or with PS3515 (p = 0.0001 or 0.0397, respectively; Fig 4C), paralleling what was previously reported for the recombinant BRs [23, 31, 33]. Thus, the isogenic strains display the anticipated sialoglycan binding specificities.
To assess the impact of binding to sTa versus sLn on endocarditis, we used two versions of a well-established animal model for this disease. The first was a competition assay, in which rats were catheterized to induce aortic valve damage and platelet-fibrin deposition, and then infected intravenously with an inoculum containing 2 x 105 CFU of M99 and an isogenic variant at a 1:1 ratio. At 72 h post-infection, animals were sacrificed and the relative number of each strain in aortic valve vegetations, kidneys and spleens were determined. Using this model, trends were apparent, with PS3515 showing higher average numbers in vegetations, kidney and spleen, and PS3516 showing lower densities compared with M99 (Fig 5A and 5B and Table 1). However, despite these trends (5 of 6 animals in the latter case) the differences were found not to reach statistical significance (p>0.05).
We then used a second established model of IE, in which catheterized animals were infected intravenously with 105 CFU of a single strain. At 72 h post-infection, animals infected with strain M99 or PS3515 had comparable levels (CFU/g) of bacteria within aortic valve vegetations (Fig 5C and Table 2). In contrast, rats infected with strain PS3516 had significantly lower densities of bacteria within aortic valve vegetations, when compared with either M99 or PS3515 (p = 0.011 and p = 0.002, respectively). Levels of bacteria within the kidneys of animals infected with strain M99 were significantly higher than in animals infected with either PS3515 or PS3516 (p = 0.049 and p = 0.001, respectively; Table 2). Importantly, no differences were seen in the number of bacteria in the blood or spleen 72 h post-infection (Table 2), indicating that the differences seen in the heart and kidney were not likely due to differences in the bacterial susceptibility to innate host defenses. These results indicate that the ability to bind sTa (M99 and PS3515) contributes to increased virulence, as measured by bacterial levels within aortic valve vegetations. In addition, selective binding to sTa (M99 versus PS3515 or PS3516) results in higher densities within kidneys, suggesting a greater tendency to disseminate from the heart to other organs.
We next examined whether the differences in bacterial densities within aortic valve vegetations at 72 h post-infection were likely due to differences in the initial attachment of circulating bacteria to valve surfaces. Catheterized rats were infected intravenously with 108 or 107 CFU of M99, PS3515 or PS3516. At one hour after infection with 108 CFU, rats given M99 had higher levels of bacteria on aortic valves, compared with either PS3515 or PS3516 (Fig 6A; p = 0.020 or 0.009, respectively). After infection with 107 CFU, levels of PS3516 on valves were again significantly lower than those of M99 (p = 0.001). Levels of PS3515 were intermediate between those of M99 and PS3516, but not significantly different from either (Fig 6B). No significant differences were seen in the numbers of bacteria in the peripheral blood at either inoculum level (Fig 6C and 6D). These results indicate that binding of bacteria to sLn rather than sTa (PS3516 versus M99) results in diminished initial colonization. The lower extent of initial colonization does not fully account for the reduced numbers of PS3516 seen at 72 h, since the initial attachment of this strain was similar to that of PS3515 (Fig 6A and 6B), yet the latter showed two-log higher density in vegetations after 72 h (Fig 5C). Thus, the combined in vivo animal studies indicate that streptococcal binding to sTa contributes to higher bacterial densities subsequent to colonization of the damaged endocardium.
A number of studies have linked bacterial binding to platelets with increased virulence in animal models of IE [15, 16, 30, 44–47]. It was therefore surprising that the isogenic variant strain that had the highest level of binding to human platelets in vitro (Fig 4C) showed lower binding to rat valves in vivo. One possibility was that the isogenic variants might be impaired for binding to rat platelets. In addition, since the SRR adhesins exhibit mechanically activated shear-enhanced adhesion [48, 49], it was conceivable that the isogenic variants could not bind to platelets on valve surfaces, due to the high shear conditions present in vivo. To assess these possibilities, we compared binding of the strains to immobilized human and rat platelets under various shear levels. At low shear (0.1 dyne/cm2), the strains bound to human platelets similarly to what was seen earlier (Fig 4C), although M99 displayed significantly lower adherence than both PS3515 and PS3516 (Fig 7A). The same relative binding of strains was observed with rat platelets (Fig 7B), with M99 less adherent than the isogenic variants. Binding to human or rat platelets at high shear (1.0 dyne/cm2) increased 2- to 4-fold for all strains as compared with binding under low shear. Thus, the lower attachment of PS3515 and PS3516 seen in vivo is not likely due to lower binding of these strains to rat platelets, or to differences in shear-enhanced binding.
We also measured the ability of bacteria to remain bound to platelets in variable flow conditions. After allowing the strains to attach under flow at 1 dyne/cm2, the shear stress was either decreased to low levels, or increased to the high levels found near the heart valve surfaces (20–80 dyne/cm2). In both cases, M99 detached from rat and human platelets at similar or greater levels than did PS3515 and PS3516 (Fig 7C and 7D). Therefore, the lower levels of PS3515 and PS3516 relative to M99 found on aortic valves at 1 h in vivo do not reflect a lesser ability of the variant strains to maintain attachment in the very high shear stress of the intracardiac environment. However, the results are consistent with an increased ability for M99 to detach from cardiac valves and disseminate to other organs via the bloodstream.
We previously determined that GspB and the 10712BR bind less readily to sialoglycans terminating in the Neu5Gc versus Neu5Ac form of sialic acid, whereas Hsa binds readily to both [23, 26]. We postulated therefore that the slightly lower binding of M99 and PS3516 to rat versus human platelets, at least at low shear, might be due to the presence of Neu5Gc on the former (unlike rats and many other mammals, humans do not produce Neu5Gc [50]). To examine this directly, we assessed the sialic acid composition of GPIbα from rat versus human platelets. We chose to examine a minimally-processed sample, to avoid the loss of labile groups (e.g. O-acetyl) or the unintentional selective enrichment of glycoform sub-populations that can occur during purification. GPIbα was the major sialylated glycoprotein in the crude extracts of both rat and human platelets, as determined by western blotting and by probing the samples with the sialic acid-binding lectin Mal-II (Fig 8A). HPLC of chemically-released sialic acids from both the human and rat GPIbα had minor amounts of O-acetylated sialic acids (contributing to 5% or 13% of the total sialic acids, respectively; Table 3). However, more than half of the sialic acid content of the rat platelet GPIbα extract was Neu5Gc, rather than Neu5Ac. This finding largely explains why M99 and PS3516 showed somewhat lower binding to rat versus human platelets.
We also examined the O-glycan structures, in order to look for differences in core 1 glycans such as sTa, versus core 2 glycans which typically have sLn branches. We found that a core 2 hexasaccharide constitutes 87% of the total O-glycans in the human GPIbα extract (Fig 8B and Table 4), consistent with earlier reports showing this as the major O-glycan on purified human GPIbα [51, 52]. Also consistent with previous reports, a relatively minor amount of sTa was detected. However, a larger core 2 octasaccharide. rather than di-sialylated core 1 glycan, was evident as an additional minor glycan. In comparison, the rat GPIbα sample had a more heterogenous distribution of O-glycans, largely due to the variety of modified sialic acid forms (Fig 8C and Table 5). In agreement with the total sialic acid analysis, slightly more Neu5Gc than Neu5Ac was evident. An unexpected finding was the presence of neuraminic acid (Neu), in addition to Neu5Ac and Neu5GC, and thus adding to the heterogeneity of O-glycans on rat GPIbα. Di-sialylated core 2 hexasaccharides were still the most abundant structures (although as a mixed population), and di-sialylated core 2 octasaccharides were also evident. Although lectin blotting with Mal-II indicated that sTa was present on rat platelet GPIbα (Fig 8A), the amount was apparently below the level of detection by mass spectrometry. The higher abundance of sLn-bearing core 2 glycans versus sTa (epitopes recognized by Hsa and 10712BR but not GspB) may explain why M99 shows relatively low binding to human and rat platelets, compared with the levels of binding by PS3515 and PS3516. The combined results suggest that streptococcal interaction with a minor O-glycan on GPIbα may be more important than the over-all affinity for GPIbα for pathogenic effects.
These studies aimed to determine whether the binding of streptococci to different sialoglycan structures has an impact on the pathogenesis of IE. Our results indicate that there are at least two means by which sialoglycan binding can impact virulence. First, binding to sLn is correlated with a reduced initial colonization of the aortic valves, as compared with sTa binding. Since binding to sLn versus sTa does not appear to promote clearance from the blood, it is likely that binding to one or more sLn-modified targets could divert bacteria away from the damaged endocardium. The potential "off-target" glycans could be displayed on plasma proteins or blood cells. The most likely off-targets for the GspB::10712BR expressed by PS3516 are the core 2 sialoglycans of GPIbα on circulating platelets. Support for core 2 sialoglycan off-targets in blood was seen in a previous study of S. sanguinis and S. gordonii, where one or more components of whole blood diverted a SrpA+ S. sanguinis strain (which has high affinity for platelet GPIbα and core 2 sialoglycans), but not S. gordonii M99, away from surfaces under conditions of flow [49]. The finding that binding to sLn and core 2 sialoglycans in vitro is associated with a negative impact on aortic valve colonization is also consistent with IE studies using S. sanguinis SK36, in which a ΔsrpA variant was slightly more virulent than the parental strain [21]. However, whether the ΔsrpA variant showed increased initial attachment was not determined. It is also likely that initial colonization of aortic valves by S. gordonii is strongly influenced by other surface adhesins, such as PadA (binding the IIbIIIa fibrinogen receptor on activated platelets), CshA (adherence to fibronectin), and SspA/B [53–57]. The relative contribution of these adhesins has not been assessed in vivo, and such studies would benefit from better in vitro models of damaged cardiac valve endothelium.
In addition to the negative impact of sLn binding on the initial colonization of aortic valves, sTa binding appears to enhance disease progression. That is, strains that can bind sTa (M99 and PS3515) had two-log higher levels of bacteria (CFU/g) within aortic valve vegetations at 72 h post-infection, compared with the strain that does not bind sTa (PS3516; Fig 5). These differences in bacterial densities are on par with previous assessments of ΔgspB or Δhsa strains [15, 16]. The findings confirm that binding to sTa, rather than to sialic acid per se, is a virulence property. In addition, the growth of PS3515 to high densities, despite lower initial colonization, indicates that sTa binding contributes to later stages in the formation of macroscopic vegetations. If platelet GPIbα is the key sTa-modified target, sTa-binding adhesins such as GspB and Hsa may play a critical role in the subsequent capture of circulating platelets, or in modulating the aggregation or activation of the captured platelets. Since sTa was confirmed to be a minor glycan on platelet GPIbα, the results suggest that binding to a unique glycosite on GPIbα is important for these events. For example, binding to sT-modified glycosites near the N-terminal leucine-rich repeat domain of GPIbα, which encompasses the binding sites for vWF and thrombin and contributes to dozens of indirect interactions with other clotting factors [58], could have localized effects on properties of the platelet-fibrin thrombus. In turn, this could impact the ability of streptococci to persist within aortic thrombi, thus contributing to the severity of disease. The impact of GspB and Hsa on platelet function and thrombus properties likely occurs in concert with other interactions, especially PadA with platelet IIbIIIa (the fibrinogen receptor) and secreted factors such as Challisin, which has been reported to cleave fibrinogen [44, 54, 59–61].
Aside from the role of sialoglycan binding in pathogenesis, a second question addressed in these studies is whether the rat and human GPIbα O-glycans are similar or different. We found that GPIbα from both species has a disialylated core 2 hexasaccharide as the major O-glycan, but the rat O-glycans display a greater variety of modified sialic acid forms. An unexpected finding was the presence of Neu, in addition to Neu5Ac and Neu5Gc, thus adding to the heterogeneity of O-glycans on rat GPIbα. Possibly due to a mix of Neu, Neu5Ac and Neu5Gc forms, sTa was not detected by mass spectrometry of the rat platelet GPIbα O-glycans. However, the binding of M99 versus the ΔgspB strain PS846 to rat platelets (Fig 7), and the strong reactivity seen with the Mal-II (Fig 8A), are strongly indicative of the presence of sTa. Since Hsa can readily bind the Neu5Gc form of sTa [23, 26], this may explain why PS3515 produced high bacterial densities in the aortic valve vegetations 72 hours after infection (Fig 5), despite the lower initial attachment (Fig 6). Based on our aggregate findings, we would predict that a variant of S. gordonii that is selective for sTa, but that can readily bind both the Neu5Ac and Neu5Gc forms, would be the most virulent in animal models of IE. Future studies will address this question.
An ongoing challenge in determining the precise mechanisms by which sialoglycan binding can drive or attenuate virulence, and whether interactions with sialylated glycoproteins beyond platelet GPIbα contribute to pathogenesis, is the limited knowledge regarding where and when specific O-glycan structures are expressed within the endovascular space. Regarding the role of sTa binding, it is possible that interactions with O-glycosylated proteins other than platelet GPIbα could contribute to streptococcal survival in the infected endocardium. However, the other sTa-modified glycoprotein ligands for Siglec-like adhesins identified thus far (red blood cells and several plasma proteins) are not known components of the aortic valve vegetations. Similarly, for sLn and core 2 glycans as off-targets, it is unknown whether other blood cells display a higher density of sialylated O-glycans than do platelets. Other potentially important off-target glycan ligands not yet specifically addressed, but recognized by the 10712BR and several other Siglec-like BRs, include sulfated or fucosylated derivates of sLn, such as sialyl Lewis X (Fig 2). Although there is little, if any, of these other structures on GPIbα or plasma proteins recognized by the Siglec-like BRs [33], in samples obtained from healthy individuals, it is unknown whether they may be more abundant in conditions of vascular damage or chronic valve disease that occur in susceptible human patient populations. As we continue to hone our understanding of the ligand specificities of the Siglec-like BRs, we can use the recombinant adhesins as probes to monitor spatial and temporal changes in specific sialoglycan epitopes in different human tissues and in the animal models of disease.
Human blood was collected from volunteers under a protocol approved by the Committee on Human Research at UCSF (IRB number 11–06207) or at the University of Washington (IRB number 29332). All donors provided written informed consent. Animal studies were approved by the Los Angeles Biomedical Research Institute animal use and care committee (IACUC number 31311–01, reference number 044163), and followed the United States Public Health Service Guide for the Use and Care of Laboratory Animals.
Todd-Hewitt broth (THB; Difco Laboratories), or Todd-Hewitt agar (THA) containing 8% (v/v) defibrinated sheep blood (Hardy Diagnostics) were used as bacterial culture media. Spectinomycin (100 μg/ml) or chloramphenicol (15 μg/ml) was added to solid media as indicated. Antibiotics and Dulbecco's phosphate buffered saline (DPBS) were from Sigma.
S. gordonii M99 is a previously-described strain that was recovered from the blood of an endocarditis patient [62]. Strains PS846 (M99ΔgspB) and PS2114 (M99Δ5'gspB::cat) lack expression of GspB and were described elsewhere [30, 36]. Replacement of the BR region of gspB in strain PS2114 was accomplished using a "knock in" strategy (Fig 3) similar to that used for generating point mutations in gspB [30]. We initially sought to replace just the Siglec domain of the BR, since this is the only region that contacts sialoglycans. However, when examining recombinant BRs, we found that fusing the Siglec domain to a heterologous Unique domain rendered the chimeric BR prone to degradation when expressed in E.coli. We learned subsequently that this was likely due to a mis-match at the domain interface, i.e. the inter-domain angle is quite different in HsaBR or the 10712BR versus GspBBR (manuscript submitted, and see Fig 1). We therefore chose to replace the entire BR as follows. Chimeric sequences spanning codons 222 to 703 of gspB, but with the BR coding sequence altered as detailed in Fig 3 and including a 3' NotI restriction site, were synthesized (Life Technologies GeneArt Strings) and used to replace the SalI-NotI fragment spanning codons 231 to 602 of gspB in plasmid pS326B602 (pS326 carrying 3'pdxU::spec::gspB1-602; the SalI restriction site is at gspB codon 231). The resulting plasmid was introduced to strain PS2114 by natural transformation. Note that this strategy was designed to force downstream recombination within the ~300 nucleotide stretch of the SRR2 coding region spanning codons 605 to 703, which is substantially different from the remainder of the SRR2 coding region, in order to avoid indiscriminate recombination further downstream and potential alteration of the length of SRR2. Colonies were selected on spectinomycin and scored for loss of chloramphenicol resistance, indicative of double crossover and gene replacement rather than plasmid insertion. To confirm the expected replacement, chromosomal DNA was extracted using the Wizard Genomic DNA Purification Kit (Promega). A 4 kb region spanning 5'pdxU to gspB codon 1060 was amplified by PCR, and then subjected to DNA sequence analysis (Sequetech). The insertion of spec upstream of gspB (PS2161) was previously determined not to affect virulence [30].
To determine growth rates, strains were grown in THB for 17 h at 37°C, diluted 1:10 into fresh medium and split to 9 × 1 ml in 5 ml snap-cap tubes. The cultures were incubated in a 37°C water bath without shaking, and tubes were removed after 1 to 23 h, vortexed, and the contents transferred to a cuvette to determine the optical density at 600 nm. The experiment was repeated twice, and a representative experiment is shown. To assess surface expression of the GspB variant adhesins, cell wall proteins were extracted with mutanolysin, and proteins were monitored by western blotting with a polyclonal antibody that recognizes the glycan moieties on GspB, as described previously [63].
Human platelets were prepared from citrate-anticoagulated blood donated by healthy volunteers as described [62]. Rat platelets were prepared from sodium citrate-treated pooled Sprague-Dawley rat blood (Innovative Research, Novi MI). Prostaglandin I2 (Cayman Chemical Company) was added to 1 μg/ml final concentration. Platelet-rich plasma was obtained by centrifugation of whole blood for 15 min at 250 × g, followed by 10 min at 500 × g. Platelets were separated by centrifugation at 1000 × g for 10 min. Platelets were washed twice with 140 mM NaCl, 6 mM dextrose, 1 mM EDTA, 20 mM Hepes pH 6.6, and then either fixed [64] or used for GPIbα extraction as indicated.
Washed, formaldehyde-fixed platelets were immobilized in microtiter wells, and the binding of S. gordonii was determined as described [63]. To assess binding to immobilized synthetic glycans, biotinylated glycans (Glycotech) were immobilized in NeutrAvidin-coated microtiter wells (Thermo Scientific). After incubating 1 h at RT, wells were rinsed twice with DPBS to remove any unbound glycans. Wells were blocked with 50 μl of a Blocking Reagent (Roche) 1X in DPBS. Excess block was removed, and 50 μl of S. gordonii strains that had been grown 17 h in THB, washed twice in DPBS, sonicated briefly to disrupt any chains, and then suspend at ~5 x 108 per ml in DPBS were added. Plates were incubated for 1.5 h at RT with moderately vigorous mixing on a rotational platform, and any unbound bacteria were removed by aspiration and washing the wells twice with 100 μl DPBS. Bound bacteria were released by adding 50 μl of a trypsin solution (1 mg/ml DPBS), incubating 30 min at 37°C followed by 30 min at RT, and then plating dilutions on sheep blood agar to enumerate the percent of the inoculum bound. Differences between means were compared for statistical significance using a one-way ANOVA, followed by the Sidak's multiple comparisons test, and using p≤0.05 as the threshold for significance.
S. gordonii binding to immobilized human or rat platelets under shear, using microfluidic flow chambers (GlycoTech), was performed as described previously for human platelets [49] and using rat platelets prepared as described above. Differences in the binding of M99 versus each of the variant strains were assessed by comparing the means for statistical significance using a one-way ANOVA, followed by the Dunnett's multiple comparisons test, using p≤0.05 as the threshold for significance. Differences in detachment were assessed only at the lowest and highest shear levels.
Infective endocarditis was produced in Sprague-Dawley female rats (250–300 g; Envigo) as described previously [15], with the following modifications. Animals were anesthetized with ketamine (35 mg/kg) and xylazine (10 mg/kg). A sterile polyethylene catheter was surgically placed across the aortic valve of each animal, such that the tip was positioned in the left ventricle, and left in place throughout the experiment. Three days post-catheterization rats were infected IV with an inoculum of either 1 x 105 CFU of single S. gordonii strains, or with 2 x 105 CFU of a pair of strains at a 1:1 ratio, as indicated. At 72 h post-infection, animals were sacrificed with pentobarbital (200 mg/kg, intraperitoneally). All cardiac vegetations, as well as samples of the kidneys and spleens, were harvested, weighed, homogenized in saline, serially diluted, and plated onto THA to determine the number of bacteria in the homogenized tissues. For the competition studies, bacterial colonies were plated onto THA and THA containing spectinomycin, in order to determine the CFU/g of M99 and the isogenic variant strain. The number of bacteria within tissues was expressed as the log10 CFU per gram of tissue. Differences between means were compared for statistical significance using a paired t-test (for competition studies), or by one-way ANOVA, followed by the Tukey correction for multiple comparisons (for single strain infections), using p≤0.05 as the threshold for significance.
Differences in the initial in vivo adherence of these strains to the endocardium were assessed using the single strain infection model described above, except that rats were infected with either 108 or 107 CFU (levels determined to be at or above the level of detection, and below the level of saturated binding). At 1 h post-infection, blood samples were obtained, animals were sacrificed and the cardiac vegetations harvested for quantitative culture.
A crude extract containing platelet GPIbα was prepared using the method of Korrel et al [52], with the following modifications. Washed platelets, obtained from 25 ml of healthy human donor blood or pooled rat blood as described above, were suspended in 1.5 ml DPBS supplemented with 2 mM CaCl2. The platelet suspension was sonicated for 15 sec, and then incubated at 37°C for 1 h. Cellular debris was removed by centrifugation at 16,000 × g, and the GPIbα-containing supernatant was filtered through a 0.45 μm membrane, and then stored at -20°C.
Human or rat platelet extracts were combined with LDS sample buffer (Invitrogen) and dithiothreitol (50 mM final concentration), boiled for 5 min, and then loaded to wells of a 3–8% polyarylamide gradient gel (Invitrogen). Following separation by electrophoresis, proteins were transferred to BioTraceNT (Pall Corporation) and then probed via western blotting with anti-GPIbα (Abcam anti-CD42b) or via lectin blotting with Mal-II (Vector Laboratories) as described [33].
Sialic acids were released from platelet GPIbα by treating the extract with acetic acid (2N final concentration) at 80°C for 3 h, filtered through a 10kD centrifugal filter (Microcon), and dried using a vacuum concentrator (SpeedVac). The released sialic acids were labeled with 1,2-diamino-4, 5-methylenedioxybenzene (DMB, Sigma Aldrich) for 2.5 h at 50°C [65]. HPLC analysis was performed using a Dionex UltiMate 3000 system with an Acclaim C18 column (ThermoFisher) under isocratic elution in 7% methanol, 7% acetonitrile, and 86% water. Sialic acid standards were derived from commercially available bovine submaxillary mucin, Neu5Gc and Neu5Ac (Sigma Aldrich) as well as from normal horse serum.
The analysis of O-glycans was performed on the same GPIbα extract used for sialic acid analysis. The glycoprotein sample was suspended in 5 mM dithiothreitol in 100 mM ammonium bicarbonate buffer (pH = 7.5) and denatured by heating in boiling water for 1 min. The N-glycans were released from the protein by digestion with peptide:N-glycosidase F (PNGase F, New England Biolabs), and the de-N-glycosylated proteins were precipitated with chilled ethanol. The O-glycans were chemically released via beta elimination by resuspending the precipitated proteins in 1 M sodium borohydride and 0.1 M sodium hydroxide. After 18 h at 45°C, the reaction was quenched with acetic acid. The released O-glycans were purified using solid phase extraction with porous graphitic carbon and hydrophilic interaction liquid chromatography. Glycan samples were analyzed on an Agilent 6520 Accurate Mass Q-TOF LC/MS equipped with a porous graphitic carbon microfluidic chip. A binary gradient consisting of (A) 0.1% formic acid in 3% acetonitrile, and (B) 1% formic acid in 89% acetonitrile was used to separate the glycans at a flow rate of 0.3 μl/min. Data were processed with Agilent MassHunter B.07 software, using the Find by Molecular Feature algorithm with an in-house library of O-glycan masses and chemical formulae to identify and quantitate the O-glycan signals.
|
10.1371/journal.pcbi.1005585 | Independent regulation of gene expression level and noise by histone modifications | The inherent stochasticity generates substantial gene expression variation among isogenic cells under identical conditions, which is frequently referred to as gene expression noise or cell-to-cell expression variability. Similar to (average) expression level, expression noise is also subject to natural selection. Yet it has been observed that noise is negatively correlated with expression level, which manifests as a potential constraint for simultaneous optimization of both. Here, we studied expression noise in human embryonic cells with computational analysis on single-cell RNA-seq data and in yeast with flow cytometry experiments. We showed that this coupling is overcome, to a certain degree, by a histone modification strategy in multiple embryonic developmental stages in human, as well as in yeast. Importantly, this epigenetic strategy could fit into a burst-like gene expression model: promoter-localized histone modifications (such as H3K4 methylation) are associated with both burst size and burst frequency, which together influence expression level, while gene-body-localized ones (such as H3K79 methylation) are more associated with burst frequency, which influences both expression level and noise. We further knocked out the only “writer” of H3K79 methylation in yeast, and observed that expression noise is indeed increased. Consistently, dosage sensitive genes, such as genes in the Wnt signaling pathway, tend to be marked with gene-body-localized histone modifications, while stress responding genes, such as genes regulating autophagy, tend to be marked with promoter-localized ones. Our findings elucidate that the “division of labor” among histone modifications facilitates the independent regulation of expression level and noise, extend the “histone code” hypothesis to include expression noise, and shed light on the optimization of transcriptome in evolution.
| Gene expression noise, or cell-to-cell expression variability, has been a topic of intense interest for more than a decade. The prevailing model of “burst-like transcription” mediated by the promoter transitions between on and off states explains the formation of noise in eukaryotes. Albeit widely accepted, the cis- elements that determine the burst frequency and burst size remain largely unknown. Here we systematically examined the relationship between transcriptional burst frequency/size and all major histone modifications in various cell types, including human embryonic cells, mouse embryonic stem cells, and yeast, and found that histone markers can be divided into two groups based on their associations with burst frequency/ size. Coincidently, promoter-localized histone markers are associated with both burst size and burst frequency whereas gene-body-localized ones are more associated with burst frequency. We further knocked out a gene that is responsible for “writing” a gene-body histone mark in yeast, and found that burst frequency is indeed reduced. Our findings reveal a new mechanism of transcriptional burst regulation and shed light on the simultaneous optimization of gene expression level and noise in evolution.
| Gene expression fluctuates among isogenic cells under identical conditions, which is frequently referred to as gene expression noise or cell-to-cell expression variability [1–3]. Noise can be decomposed into intrinsic noise and extrinsic noise, according to its origins. The inherent stochasticity of biochemical reactions (for example, collision between transcription factors and promoters) and the chromatin transition between on and off states together result in intrinsic noise [4]. Intrinsic noise can be amplified in the gene regulatory network, leading to varying concentrations of gene expression machinery (for example, polymerase II and ribosomes) among isogenic cells, which reinforces the cell-to-cell variability in gene expression and forms the basis of extrinsic noise.
Similar to expression level, expression noise is of central importance in growth, development, and responding to environmental fluctuations [1–3,5–12]. Thus, it is subject to natural selection and fine-tuned according to the gene function [7,8,10,13]. For instance, essential genes and genes encoding protein complex subunits, the expression variation of which is predicted to reduce fitness of the organism and thus collectively termed as dosage sensitive genes, often exhibit low expression noise [7,11,13]. By contrast, genes responding to environmental fluctuations are often associated with high expression noise, which could be a product of positive selection due to immediate or long-term benefits on fitness [6,10]. To summarize, expression level and noise are two facets of gene expression, both of which are subject to natural selection.
However, a strong negative correlation between expression level and noise has been observed in bacteria, yeast, and mammals [1,5,11,14–16]. For example, in Saccharomyces cerevisiae, Newman et al took advantage of the green fluorescent protein (GFP) collection in which each strain has a GFP fused to the C terminus of an endogenous protein, measured single cell protein levels in the cell population of each strain by fluorescence-activated cell scanning (FACS), and calculated expression level and noise for ~2000 genes. A strong negative correlation between expression level and noise was reported [11]. This negative correlation manifests as a potential constraint for simultaneous optimization of expression level and noise [17].
In yeast, a number of cis- acting elements have been suggested to regulate noise independent of expression level. For example, TATA-box containing genes have higher noise than TATA-less genes after controlling for expression level, suggesting its potential role in regulating gene expression noise [3,6,11,15,18]. Further, mutations in TATA-box result in marked decrease of both expression level and expression noise, suggesting the co-occurrence of high expression level and high noise enabled by TATA-box [14]. Another element is the sequence that determines the nucleosome occupancy around the transcriptional starting site (TSS). Occupied proximal-nucleosome (OPN) genes tend to have higher noise, while depleted proximal-nucleosome (DPN) genes tend to have lower noise, even though they display no significant difference in expression level [11,17,19]. In fact, adding nucleosome-disfavoring sequences and strengthening transcription factor binding sites exhibit different impacts on expression noise, even though they both elevate expression level [18]. Besides, a recent study shows that adaptive changes in the expression mean and noise of a gene with autoregulation occurred during the long-term evolution of yeast, suggesting that feedback is an alternative mechanism to decouple gene expression mean and noise [20].
In addition, epigenetic modification, which is closely associated with chromatin remodeling, has been suggested to play a role in regulating gene expression noise [21–26]. For example, Weinberger et al discovered that two histone deacetylation complexes, Set3 and Rpd3(L)C, play different roles in regulating gene expression noise although both repress expression level [27], implicating the potentially sophisticated regulation of transcription by histone modifications. However, the genomic landscape of all major histone modifications’ impacts on expression noise is yet to be reported. Here, we systematically investigated gene expression noise and discovered that histone modifications are associated with independent regulation of expression level and noise.
Although gene expression noise has been extensively studied in single-celled organisms [1,5,11,14,15,28], noise in mammals, especially in human, has only been rarely addressed, either genome-wide [16,29] or on a small scale [30–32]. Single-cell mRNA-seq was performed in human preimplantation embryos [33], which provides us with the opportunity to investigate gene expression noise in human. Here, the expression noise of a gene is defined as the coefficient of variation (CV, σ/μ) of transcript concentrations among isogenic cells. We chose to use the data from 8-cell stage embryos for two reasons. First, the total number of single cells at this stage (20 cells from 3 embryos) is larger than that at 2- and 4-cell stages (6 and 12 cells, respectively). Second, it’s suggested that embryonic differentiation has not started yet at 8-cell stage [34]. To minimize the impact of sampling error during library preparation, which is larger for lowly expressed genes [35], we focused on the 7741 genes with at least one sequencing read detected in all 20 cells. With these data, we first performed unsupervised hierarchical clustering analysis, and found that cells from the same embryo do not always cluster together (Fig 1A), suggesting that gene expression pattern in early embryos may not be fully determined by embryo identities. Consistently, the relatively long external branches compared to internal branches (Fig 1A) indicate that these cells show unique gene expression patterns that are largely independent of their embryo identities. Nevertheless, to fully exclude the embryo effect in estimating expression noise, we subtracted the average expression of a gene in cells of the same embryo from its expression when calculating gene expression noise (Fig 1A). Consistent with previous findings in yeast [4,11,14,15] and in a mouse cell line[36], we observed a strong negative correlation between expression level and noise in human embryonic cells (Fig 1B, r = -0.66, P < 1×10−100, N = 7741, Pearson’s correlation). Note that for most genes, the noise estimated from single-cell mRNA-seq data is higher than technical noise (indicated by the grey line in Fig 1B), suggesting that technical noise is not the main source of the observed correlation above.
It is worth noting that the observed correlation between expression level and noise is potentially a constraint that prevents simultaneous optimization of both. Nevertheless, we observed that many genes, low-expression-low-noise and high-expression-high-noise genes of particular interest, deviated from the regression line in the expression level-noise correlation (Fig 1B). To examine whether this is related to biological functions, we divided genes into four groups based on their expression level and noise, and found that genes in these groups are enriched in various biological pathways (Fig 1B and S1 Table). More importantly, an inspection of these pathways reveals that the observations are consistent with the direction of natural selection on expression level and noise. For instance, genes in signaling transduction pathways, which generally function through the accurate expression of minute amounts of specific products, are enriched in the-low-expression-low-noise group (Group 4, Fig 1B). On the contrary, genes in the pathways of energy generation, amino acid metabolism, and autophagy, which benefit the organism by sensing and responding to environmental fluctuations, usually demand both high expression and high noise [6,10,37]. Consistently, they are enriched in the high-expression-high-noise group (Group 1, Fig 1B). We further defined genes with significant deviation as those deviated from the 95% confidence intervals of the major axis and minor axis, and reanalyzed the enriched pathways in 4 groups of genes, which generated similar results (S2 Table). More generally, we calculated the ratio between the numbers of essential and nonessential genes in each group, and found that essential genes are significantly enriched in the high-expression-low-noise group (Group 2, S1A Fig). Similar result was obtained for genes encoding protein complex subunits (Group 2, S1B Fig). We further verified that the observed pattern is not an artifact of varying mRNA decay rates among human genes, because mRNA decay rate [38]is only weakly correlated with expression level and noise (S2 Fig). Taken together, we found that regardless of the strong correlation between expression level and noise, the decoupling of them is prevalent in human embryonic cells.
We next turned to the transcriptional process for a molecular interpretation of the decoupling. It is generally accepted that eukaryotes, the DNA of which is wrapped around histones forming nucleosomes, mainly adopt a burst-like transcription process [2–4,9,14,17,29,39], due to the chromatin remodeling mediated promoter on-and-off transitions (Fig 2A). It was previously observed that a large fraction of genes in human adopt the burst-like expression mode [29,40]. In this prevailing “burst model”, when the promoter is “on”, a number of mRNA are transcribed, which is called a burst event. The frequency of such burst events is defined as burst frequency and the average number of mRNA molecules made per event is defined as burst size. Therefore, expression level changes with both burst frequency and burst size, while expression noise changes mainly with burst frequency (Fig 2B & S3 Fig) [14,17]. Thus, if expression level is predominantly regulated by burst size at the genomic scale, the slope of the regression line between expression level (μ) and noise (standard deviation divided by mean, CV, σ/μ) should be approximately 0 (Fig 2B). By contrast, if expression level is predominantly regulated by burst frequency, the slope should be approximately -0.5, because CV is inversely proportional to the square root of mean in a Poisson process (Fig 2B). In Fig 2B, the slope falls between 0 and -0.5, suggesting that both burst size and burst frequency are modulated across the human genome. Meanwhile, deviation from the regression line suggests that the relative contributions of burst frequency and burst size vary among genes. Specifically, genes above the regression line have relatively larger burst size while those under the line have relatively larger burst frequency, given similar expression levels. For instance, genes in Groups 1 and 3 have relatively larger burst size but smaller burst frequency than genes in Groups 2 and 4, respectively (Fig 1B). Based on this, we hypothesized that the decoupling of expression level and noise is enabled by the separate modulation of burst frequency and size.
Therefore, we estimated burst size and burst frequency of transcription for each gene from the single-cell mRNA-seq data [33] following previous studies [29,41]. mRNA level measured by mRNA-seq is proportional but not necessarily equal to the mRNA number of a gene in a cell, due to the amplification during mRNA-seq library preparation and high throughput sequencing. Because this amplification influences the estimation of the Fano factor and thus burst size, we first calibrated the mRNA number of each gene in a cell by the total number of mRNA in a typical mammalian cell (see Methods and Materials for details) [42]. We then estimated burst size with the equation burst size = σ2/μ -1, in which σ2/μ is the Fano factor. We further calculated burst frequency by calculating the ratio between the mRNA number of a gene in a cell and the estimated burst size (see Methods and Materials).
Epigenetic modification, especially histone modification, has been reported to play vital roles in regulating expression noise [27]. To systematically examine the roles histone modifications play in transcriptional regulation, we first retrieved the genomic distribution data of all the major euchromatic histone modifications, H3K4me1, H3K4me2, H3K4me3, H3K9ac, and H3K27ac, H3K36me3, H3K79me2, and H4K20me1, which were obtained from the chromatin immunoprecipitation sequencing experiments (ChIP-seq) conducted in human embryonic stem cells (hESC), from the ENCODE project [43]. Specifically, we defined 2000 base pairs upstream of TSS to transcription end site (TES) as the range of a gene and obtained the called ChIP-seq peaks for each histone mark in this range. In this study, the strength of a histone modification on a gene was defined as the average intensity of this histone mark on the gene, which was calculated as the ratio between the total intensity (total “height” of all peaks) of a histone mark on a gene and the range of the gene (see Methods and Materials for details).
We calculated the Pearson’s correlation between the intensity of each histone modification and transcriptional burst frequency/size. Surprisingly, we found that these histone modifications can be divided into two distinct groups based on the correlations with burst frequency and size. Specifically, three histone modifications (H3K36me3, H3K79me2, and H4K20me1) have significantly stronger correlations with burst frequency (rBF) than with burst size (rBS), while other histone modifications in general exhibit little differences between rBF and rBS (Fig 2C and 2D, significance determined by permutation test). Importantly, this distinction coincides with that according to the localization of histone modifications (Fig 2E). That is, gene-body-localized histone modifications exhibit larger rBF, while promoter-localized ones (including the enhancer-localized marker H3K4me1) often show similar rBF and rBS. It is important to note that mRNA decay rate was not included in the estimation of burst parameters here, due to the lack of such data in human embryonic cells. Therefore, we used the approximation that the mRNA degradation rate is identical among transcripts and set this rate to 1 min-1. Nevertheless, when we used the mRNA decay rate obtained from human lymphoblastoid cell lines as a substitute [38], the observation in Fig 2 persisted (S4 Fig).
Since expression level is jointly determined by burst frequency and burst size, while expression noise is mainly determined by burst frequency, we predict that promoter-localized histone modifications are more strongly associated with expression level (|rmean| > |rnoise|) while gene-body-localized ones are more strongly associated with expression noise (|rnoise| > |rmean|), which was indeed observed in 2-cell, 4-cell, and 8-cell embryos (Fig 3A). This suggests that the independent regulation of expression level and noise may be enabled through separate modulation of burst size and burst frequency by histone modifications. As an example, the low-noise gene UBE2E1 is intensively modified by the gene-body-localized histone mark H3K79me2 but only sparsely modified by the promoter-localized histone mark H3K4me3, while the high-noise gene FBXO8 displays the opposite pattern (Fig 3B and 3C).
It is worth noting that expression level and expression noise were estimated in embryonic cells whereas histone modification data are from hESC, due to the lack of such data in human embryonic cells. To test whether this can lead to artifacts in our analysis, we further examined histone modification conservation between two substantially different cell types, hESC and GM12878 (a lymphoblastoid cell line), in ENCODE. We found that the intensity of each histone mark is highly correlated between these two cell types (Spearman’s correlation coefficient ρ ranges from 0.46 to 0.83, S5 Fig), suggesting that histone modification state may be largely similar across cell types. More importantly, we calculated correlation between the intensities of histone modifications in GM12878 and the expression level (noise) in cells from the 8-cell stage embryos. The pattern observed in Fig 3 is largely unchanged (S6 Fig), suggesting that our observation is not sensitive to the cell type in which histone modification was quantified.
To summarize, histone modifications in human embryonic cells can be divided into two distinct groups based on their associations with burst frequency and size. Although the impact of individual histone modification seems modest, the combined effect of multiple markers on the same gene could be strong. It is worth noting that the association between gene-body-localized histone markers and burst frequency is not likely the consequence of transcriptional elongation mediated by these markers [44–46], because transcriptional elongation is unlikely to determine transcriptional burst frequency (but see [22]).
The budding yeast Saccharomyces cerevisiae has been frequently used to study gene expression noise [2,5,6,8,11,14,15]. With the yeast GFP collection in which the coding sequence of GFP is fused to the C-terminus of an endogenous gene in each strain, Newman et al (2006) measured single-cell fluorescence of ~2000 strains with FACS [11]. With these data, we next investigated the association of histone modifications [47] with expression level and noise in yeast. We calculated burst frequency and burst size of these genes and calculated correlation coefficients between them and the intensity of each histone modification. Consistent with the findings in human embryonic cells, we found that rBF is significantly larger than rBS for the gene-body-localized histone modifications, such as H3K79me3 (rBF = 0.26, rBS = 0.03, permutation test P value < 0.001, Fig 4A and 4B), while promoter-localized histone modifications tend to have similar rBS and rBF. Consistent with this, genes with high H3K79me3 intensities have a steeper slope than those with low H3K79me3 intensities (Fig 4C and 4D, linear regression P = 3.5×10−12, df = 2163), which indicates that burst frequency is modulated to a larger extent among genes with high H3K79me3 intensities, suggesting the role of H3K79me3 in preferential modulation of burst frequency. The consistent pattern in multiple human embryonic stages (Fig 3A), in yeast (Fig 4), and in mouse embryonic stem cells (see below) suggests that the role of histone modifications in decoupling expression level and noise is evolutionarily conserved.
Since TATA box and nucleosome occupancy are also suggested to play a role in the decoupling of expression level and noise in yeast [3,6,11,15,18,19], we calculated partial correlations between H3K79me3 intensity and gene expression noise controlling for these two factors. It turns out that the correlation coefficients stay virtually unchanged (Fig 4A and 4B), suggesting that H3K79me3 is associated with noise through a TATA or nucleosome occupancy independent mechanism.
Our observations so far are mainly from correlation analyses. We next designed experiments to examine the causality between histone modifications and expression level/noise. Here, we used H3K79 methylation as an example, because it has only one “writer” (DOT1, a non-essential gene) in yeast [46], which makes the removal of H3K79 methylation more feasible. Specifically, we constructed a homozygous DOT1 knockout strain and confirmed the absence of H3K79 methylation with western blot (Fig 5A). A pseudo gene (HO) homozygous knockout strain was similarly constructed as a negative control. To examine whether intrinsic noise level is elevated in the DOT1 knockout strain, a two-color system was constructed following previous studies [1,3], in which GFP and a red fluorescent protein gene (dTomato) were respectively fused to two alleles of the same endogenous gene (Fig 5B). Since two fluorescent proteins are expressed from the same promoter in the same cellular environment, the fluorescence difference between them is only attributable to intrinsic noise. We first examined intrinsic noise of TEF1, which is reported to be extensively modified by H3K79 methylation on its gene body [47], and observed that the fluorescence difference between GFP and dTomato is larger in DOT1 knockout strain (Fig 5B), suggesting that removal of H3K79 methylation elevates the intrinsic noise of TEF1.
To further quantify intrinsic noise, we randomly chose 6 additional genes with different H3K79 methylation intensities in wild-type cells, as reporter genes (S3 Table), and measured their fluorescence intensities of GFP and dTomato with FACS (Fig 5C). We calculated intrinsic noise in DOT1 and HO knockout strains following previous studies [1,3], and found that DOT1 deletion indeed elevates expression noise, especially for highly expressed genes (SSB1 and TEF1, Fig 5D). Importantly, the slope difference between regression lines in HO and DOT1 knockout strains (Fig 5D) is consistent with that among genes with high and low H3K79me3 intensities (Fig 4D). In further support of our model, both SSB1 and TEF1 exhibit lower burst frequencies in DOT1 deletion strains (Fig 5E). These observations are consistent with a recent genome-scale study about the position effects on gene expression noise [25]. In that study, Chen and Zhang reported that genes inserted into genomic regions with high H3K79me3 tended to have lower expression noise in yeast. To summarize, H3K79 methylation plays a causal role in repressing gene expression noise, which supports our hypothesis that independent regulation of expression level and noise is enabled by histone modifications.
Unexpectedly, we also observed that expression level virtually does not change upon DOT1 deletion, potentially due to some compensatory change of burst size accompanying the decrease of burst frequency. Alternatively, the compensation may occur at the post-transcriptional level. For example, protein degradation rate may change upon dosage imbalance among genes [48], so the altered protein degradation rate could lead to the compensation of protein concentration in DOT1 deletion strains. Importantly, the observation of compensation in our experiment is consistent with the evolutionary change of gene expression between human and mouse (see Discussion).
Genes with divergent functions have unique combinations of expression level and noise (Fig 1B). Are such unique combinations enabled by the histone coding strategy described above? To address this question, we first examined the enrichment of histone modifications in each group of genes defined in Fig 1B. Specifically, for each histone modification, we classified genes into high-intensity ones and low-intensity ones with the median, and calculated the ratio between the numbers of them in each group, which reflects the usage preference of this histone modification among genes in the group. We found that genes with both high expression level and high noise (Group 1) are preferentially modified by promoter-localized histone markers, while genes with high expression level and low noise (Group 2) are preferentially modified by gene-body-localized ones in human (Fig 6A). We further illustrated this point with two pathways, autophagy pathway in Group 1 and Wnt signaling pathway in Group 4. Genes in autophagy pathway respond to intracellular and extracellular stimuli, and thus, are predicted to have higher noise. Indeed, we found that they are preferentially modified by the promoter-localized histone marker H3K4me3 (Fig 6B). By contrast, genes in Wnt signaling pathway require low noise to ensure the accurate expression of minute amounts of their products, thus guaranteeing the fidelity in signaling transduction. Consistently, these genes are preferentially modified by the gene-body-localized histone marker H3K79me2 (Fig 6C). Additional examples (oxidative phosphorylation signaling pathway in Group1 and Jak-STAT signaling pathway in Group4) are shown in S7 Fig. We then performed a similar enrichment analysis in yeast and observed virtually the same pattern as in human embryonic cells (S8 Fig). Interestingly, H3K36me3 is enriched among Group 4 genes but not Group 2 genes, suggesting that H3K36me3 may play different roles in highly expressed and lowly expressed genes in yeast.
More broadly, dosage sensitive genes (e.g., essential genes and genes encoding protein complex subunits) tend to have lower noise both in yeast [7,11] and in human (S1 Fig). To test if these genes are preferentially modified by gene-body-localized histone markers, we calculated usage preference of each histone modification among them. To exclude the confounding effect of expression level, we first divided genes into 10 equal-sized bins based on expression level. Then in each bin, we divided genes into four categories based on the intensity of histone modifications and dosage sensitivity. With that, we calculated an odds ratio of the contingency table in each bin, and then calculated a common odds ratio and 95% confidence interval with Mantel-Haenszel procedure (Fig 7A). Consistently, we found that dosage sensitive genes indeed prefer to use gene-body-localized histone modifications, which is reflected by the larger-than-one common odds ratios, in both human embryonic cells and yeast (Fig 7B and 7C). All these observations indicate that histone modifications play a vital role in independent regulation of expression level and noise in human and yeast genomes.
In this study, we calculated expression noise among cells from three 8-cell stage embryos, considering that the fraction of maternal mRNA has declined to reach parity with paternal transcripts [49] and the number of single cells (20 cells) is relatively large. However, it is recently reported that a small proportion of genes display bimodal expression at 2- or 4- cell stage in mouse, implying potential cell differentiation in human 8-cell stage embryos [50], which may confound the calculation of expression noise. Nevertheless, we hold that the observations in this study are not artifacts, with the following supporting evidence. First, we calculated expression level and noise from human 2-cell stage (N = 6), 4-cell stage (N = 12), and embryonic stem cells (N = 30), and observed “decoupling” of expression level and noise in all these samples (Fig 3A & S9A Fig). Second, we further excluded human homologs of the bimodally expressed genes in mouse at 2- or 4-cell stage [50] and obtained essentially the same pattern (S9B Fig). Third, similar pattern was also obtained in mouse embryonic stem cells (N = 94, S9D Fig), as well as in yeast (Figs 4 and 5). The consistent patterns observed in all these samples imply the negligible contribution of potential differentiation at the 8-cell stage to noise calculation.
So far, we have observed that the intensities of gene-body-localized histone modifications are associated with expression noise among genes (Figs 2–4). Next, we investigated whether the same pattern could be observed among orthologous genes between human and mouse. Specifically, we asked whether the change of H3K79me2 intensity on a gene in evolution could predict the divergence of expression noise. To this end, we obtained the H3K79me2 intensities as well as single-cell transcriptomes in embryonic stem cells of human and mouse [33,51]. We observed that the change of H3K79me2 intensity could successfully predict the divergence of expression noise between human and mouse (S10 Fig, ρ = -0.12, P = 1.1×10−6), suggesting that the modulation of H3K79me2 intensity may play a role in the evolutionary optimization of gene expression noise. Interestingly, we did not detect a significant correlation between the change of H3K79me2 intensity and the divergence of expression level (S10 Fig, ρ = -0.03, P = 0.2). This could be due to the same compensatory mechanisms underlying the pattern observed in our manipulative experiment in yeast (Fig 5).
In yeast, TATA box and TSS-proximal nucleosome occupancy are associated with high noise independent of expression level [3,6,11,14,15,18,19]. In our study, however, we found that in human embryonic cells, neither the presence/absence of TATA-box [52,53] (ρ = 0.05, P = 0.10, N = 1393) nor the TSS-proximal nucleosome occupancy [54] (ρ = -0.003, P = 0.88, N = 3350, Fig 2C) is correlated with burst frequency.
Consequently, the difference in expression noise between TATA-box containing and TATA-less genes is not significant, and TSS-proximal nucleosome occupancy is only weakly correlated with noise (S11 Fig). This result persists when we classified genes according to the presence of a canonical TATA-box (TATAAA, S11A Fig) and calculated nucleosome occupancy in various ranges (S11B Fig), which implies that alternative mechanisms should exist to overcome the constraint between expression level and noise in human embryonic cells. We speculated that the losses of functions of TATA-box and nucleosome occupancy in expression noise regulation are compensated by the role of histone modifications in mammals.
Previous studies have provided some potential mechanisms by which chromatin structure and histone modifications regulate gene expression noise [3,6,11,15,18,19,21–26]. Weinberger et al discovered that histone deacetylase complex Rpd3(L)C could regulate expression noise by modulating transcription initiation [27]. Benayoun et al observed a correlation between H3K4me3 breadth and transcriptional consistency among single cells [22]. Benayoun et al further confirmed that the perturbation of H3K4me3 breadth led to reduced transcriptional consistency. They proposed that H3K4me3 breadth might regulate transcriptional consistency through a positive feedback loop between transcription initiation and elongation. However, in a genome-scale experiment that examined the position effects on gene expression noise in yeast [25], Chen and Zhang completely knocked out the open reading frame of the gene at the GFP knock-in site; they still observed that the intensity of the gene-body-localized histone modification H3K79me3 was associated with the expression noise of GFP. This observation suggests the presence of mechanisms in addition to the positive feedback loop between transcription initiation and elongation. In yeast, TSS-proximal nucleosome occupancy regulates the accessibility of the promoter, which influences burst frequency and expression noise [3,6,11,15,17–19]. Because H3K79 methylation regulates the switch between heterochromatin and euchromatin [46], gene-body-localized histone modifications may regulate the accessibility of chromatin in a broader region so that genes localized in this region exhibit reduced gene expression noise [25]. Consistent with this mechanism, Chen and Zhang observed that essential genes tend to be localized in the chromosome regions associated with low expression noise[25], which supported a previous hypothesis that essential genes form clusters in low noise regions to optimize the robustness of transcription [55].
It is also worth noting that although we provided evidence that histone modification can regulate expression noise (Fig 5), we did not exclude the opposite mechanism that transcriptional bursts may influence the intensities of gene-body-localized histone modifications. The relative contributions of these two mechanisms to the observed correlation between gene-body-localized histone modifications and expression noise deserve further investigations in the future.
Histone code hypothesis states that the genetic information encoded in DNA is partly regulated by chemical modifications to histone proteins [21]. Although histone modification is tightly associated with transcription, it remains elusive in what specific aspects of transcription do they play a role. For example, both of H3K4me3 and H3K79me2 are associated with active transcription, do they have redundant or separate functions? Here we discovered that burst frequency and burst size, the two independent parameters of transcription, were likely modulated by two distinct groups of histone modifications. Specifically, three gene-body-localized histone markers (H3K36me3, H3K79me2, and H4K20me1) exhibited stronger correlations with burst frequency than with burst size, while one promoter-localized histone marker (H3K4me2) exhibited a stronger correlation with burst size than with burst frequency, in human embryonic cells (Fig 2). In yeast, two gene-body-localized histone markers (H3K36me3 and H3K79me3) exhibited stronger correlations with burst frequency than with burst size, while three promoter-localized histone markers (H3K4me2, H3K9ac, and H3K14ac) exhibited stronger correlations with burst size than with burst frequency (Fig 4). Through these histone modifications, the independent regulation of expression level and noise is enabled. Note that additional mechanisms may also be involved in the independent regulation of expression level and noise.
Our finding broadens our understanding of transcription regulation by histone modifications and expands the histone code hypothesis to include the regulation of gene expression noise. Importantly, the evolutionarily conserved patterns in human, mouse, and yeast imply that this epigenetic strategy is probably general and is adopted by many other species.
Single-cell mRNA-seq data in human preimplantation embryos were downloaded from Yan et al (GSE36552)[33]. Specifically, the genome-wide expression profiles of 20 single cells from three 8-cell stage embryos were retrieved, which contain 4 cells, 8 cells and 8 cells, respectively. Considering the larger technical error among lowly expressed genes, only genes with expression detected in all 20 cells were used. For each gene, noise (coefficient of variation, CV) was calculated as follows:
Noise(CV)=1N−3∑j=13∑i=1nj(xij−x¯j)21N∑j=13∑i=1njxij
(1)
where xij is the expression level of the gene (in the unit of Reads Per Kilobase per Million mapped reads, RPKM) in the ith cell of the jth embryo. N (= 20) is the total number of cells, and nj is the number of cells in the jth embryo. Note that in order to exclude embryo effect in noise estimation, average expression level of each embryo (x¯j) is used in calculating CV. Gene expression noise at the 2/4-cell stage was calculated similarly with Eq (1).
Single-cell mRNA-seq data of the primary outgrowth during human embryonic stem cells (hESC) derivation (passage 0) and hESC of passage 10 were also downloaded from Yan et al (GSE36552)[33]. Expression noise is calculated as follows:
Noise(CV)=1N−2∑j=12∑i=1nj(xij−x¯j)21N∑j=12∑i=1njxij
(2)
Eq (2) is similar to Eq (1), except that 3 embryos are replaced with 2 passages.
Single-cell transcriptomes of wild-type mouse embryonic stem cells (mESC) cultured in ground state condition were generated by Kumar et al [51]. Expression level and noise (CV) of 94 cells were retrieved from their supplementary data.
Gene expression data at the protein level in single cells were measured with FACS in yeast by Newman et al [11]. Expression level and noise (CV) in YPD were retrieved. Newman et al also calculated DM (distance of each CV to a running median of CV) to control the negative correlation between expression level and noise[11]. In our study, we separately calculated correlations between histone modifications and expression level or CV. Similar to DM, the difference between the absolute values of those two correlations reflects the deviation from the negative correlation between expression level and CV.
Single cell mRNA-seq data generally have relatively high technical errors, mainly from the variation of single-molecule capture efficiency [36,56]. However, none of known errors generates bias towards certain histone modifications after controlling for gene expression level. In yeast, single-cell protein concentrations were measured by FACS [11], which avoids the high variation of single- molecule capture efficiency in single-cell mRNA-seq.
Total noise includes both extrinsic and intrinsic noise. It was previously reported in yeast that after controlling for cell size, extrinsic noise is significantly reduced. The cell sizes (and cell cycle status—another major source of extrinsic noise) are largely identical in 2-cell and 4-cell human embryos, and we still observed a similar pattern in these stages (Fig 3). These observations suggest that extrinsic noise is unlikely a major factor influencing our major conclusions of this study.
Unsupervised hierarchical clustering analysis was performed on the single-cell transcriptome data from 20 cells with R. To minimize the effect of experimental error in RNA-sequencing on noise calculation, hierarchical clustering analysis was performed on 7741 genes with at least one sequencing read detected in all 20 cells.
Based on the estimation that a typical mammalian cell contains 200,000 mRNA molecules [42], we estimated the “amplification factor”, A, which equals RPKMtotal/200,000, where RPKMtotal is the sum of RPKM values of all genes. The number of transcripts for gene i, Ni, was estimated by RPKMi/A, where RPKMi is the RPKM value of gene i.
Based on the burst-like model in eukaryotic transcription, burst size can be estimated from the Fano factor (σ2/μ), following previous studies [29],
Burstsize=σ2/μ-1
(3)
where σ2 and μ are the variance and mean of the estimated numbers of transcripts for each gene among cells, respectively. And
Burstfrequency=μ×γm/burstsize
(4)
where γm represents mRNA decay rate. Because mRNA decay rate is not available in human embryonic cells, we used two approaches to approximate it. First, we used the average mRNA decay rate estimated from 7 human B-lymphoblastoid cell lines [38]. Second, we assumed that the variance of mRNA decay rates among transcript species was smaller than the variance of mRNA production rates, and used Eq (5) to approximate burst frequency.
Burstfrequency=μ/burstsize
(5)
Two approaches led to similar observations (Fig 2C and S4 Fig). Genes (N = 3350) with estimated burst size > 0.75 were presented in Fig 2C. The observation kept unchanged when the cutoff of burst size was changed to > 0.5 (N = 3733, S12 Fig) or > 1.0 (N = 3057, S12 Fig).
Burst size and burst frequency in yeast were estimated with the following equations.
Burstsize=σ2/μ-1
(6)
Burstfrequency=μ×γp/burstsize
(7)
where variance (σ2) and average expression level (μ) were retrieved from a previous study [11] and were normalized by the total number of proteins in a yeast cell (~5×107) [57]. Protein degradation rate (γp) was retrieved from a previous study that measured protein degradation rates based on stable isotope labeling with amino acids in cell culture (SILAC) [58]. This degradation rate dataset was used because protein synthesis inhibitors, which perturb normal cell status, were not added to the culture media in this study. Note that the burst size in Eqs 6 and 7 refers to the number of proteins synthesized in each burst (Burst sizep) whereas the burst size in Eqs 3–5 refers to the number of mRNAs synthesized in each burst (Burst sizem). They are related by Burst sizep = Burst sizem × kp / γm, where kp is the protein translation rate and γm is the mRNA degradation rate. The underlying assumption in Eqs 6 and 7 is that the mRNA degradation rate is much higher than the protein degradation rate [41]. We retrieved the half-life data from previous studies [58,59], and found that this was indeed the case (the median mRNA half-life is ~ 7 minutes, while the median protein half-life is ~521 minutes).
Histone modification data in human and mouse were downloaded from the encyclopedia of DNA elements (ENCODE) project [43], in which chromatin immunoprecipitation coupled with high throughput sequencing (ChIP-seq) experiments were performed to measure intensities of histone modifications. All Eight ChIP-seq datasets of euchromatic histone modifications in H1-hESC (BROADPEAK files) were retrieved, under GEO accession numbers GSM733782 (H3K4me1), GSM733670 (H3K4me2), GSM733657 (H3K4me3), GSM733773 (H3K9ac), GSM733718 (H3K27ac), GSM733725 (H3K36me3), GSM1003547 (H3K79me2), and GSM733687 (H4K20me1).
Five ChIP-seq datasets of euchromatic histone modifications in mouse embryonic stem cell line ES-Bruce4 (BROADPEAK files) were retrieved, under GEO accession numbers GSM769009 (H3K4me1), GSM769008 (H3K4me3), GSM1000127 (H3K9ac), GSM1000099 (H3K27ac), and GSM1000109 (H3K36me3). Because H3K79me2 data are not available in any mouse embryonic stem cells, H3K79me2 in cell line CH12 was used instead.
The average intensity of histone modification on each gene was calculated as follows:
1LG∑i=1kLi×Ei
(8)
where LG is gene length, which is defined as the distance between 2 kilo base pairs (kb) upstream of TSS and TES. k is the total peak number of a specific histone modification on a gene. Li and Ei are the length and the average intensity of the ith histone modification peak on the gene, respectively. Since this equation could lead to inaccurate estimation of the intensities of histone modifications that only localize in part of a gene, we also calculated the intensity of histone markers in a modified equation:
1LP∑i=1kLi×Ei
(9)
where gene length (LG) was replaced by the length of genome that cover all peaks on a gene (LP). Two equations result in similar observations (Fig 3A and S9C Fig). Thus, Eq (8) was used in the rest of this study.
Histone modification data in yeast were retrieved from a previous study [47]. In the study, chromatin immunoprecipitation coupled with DNA microarrays (ChIP-chip experiments) were performed for H3K9ac, H3K14ac, H3K4me1, H3K4me2, H3K4me3, H3K36me3, and H3K79me3. Average histone modification intensity from multiple probes was calculated for each gene.
To determine if the difference between rBF and rBS is significant, we performed permutation by shuffling genes for 1000 times, obtained 1000 rBF and rBS, and then calculated two-tailed P values.
TATA-box classification information in human was downloaded from Jin et al [52] and Yang et al [53]. In Jin et al, promoters were further classified into two categories, based on the presence of the canonical TATA-box (TATATAA).
Nucleosome protected DNA sequences in human were retrieved from Gaffney et al [54], where high throughput sequencing data were generated from micrococcal nuclease-digested chromatin (MNase-seq). This study contains the highest-resolution map of nucleosome occupancy to date in human. Nucleosome occupancy was calculated in the region between 250/200/150/100 base pairs (bp) upstream of the transcriptional starting site (TSS):
Nucleosome occupancyregion length=∑i=1region lengthxi
(10)
where xi is the number of reads whose midpoints are i bases upstream of TSS.
TATA box-containing genes in yeast were identified by Basehoar et al [60]. Genes with occupied proximal-nucleosome (OPN) and depleted proximal-nucleosome (DPN) were identified from the promoter nucleosome occupancy data in yeast [61] by Tirosh and Barkai [19].
Human orthologs of mouse essential genes were identified by Georgi et al [62], and were defined as essential genes in human in our study. Genes encoding protein complex subunits were retrieved from CORUM (http://mips.helmholtz-muenchen.de/genre/proj/corum) and Havugimana et al [63].
Assuming that technical-error-derived mRNA molecule number follows a Poisson distribution, we can deduce that the line corresponding to technical noise has a slope of -0.5 and an intercept of 1/2×log2(A), or 0.73 (Fig 1B), where A is the amplification factor estimated above. Considering that human embryonic cells are larger than average human cells and thus have more mRNA molecules than 200,000, the actual technical noise should be lower than estimated here.
GOstats [64] was used to calculate the KEGG (Kyoto Encyclopedia of Genes and Genomes) pathways that are enriched in each of the 4 groups defined in Fig 1B. The background gene set is all genes in the other three groups.
For each gene in a pathway, we calculated the average H3K79me2 and H3K4me3 intensity as previously described, and then calculated the intensity ratio between them. As a control, for each gene in autophagy or oxidative phosphorylation signaling pathway (Group 1), we identified 100 genes with similar expression levels from Group 2, calculated the H3K79me2/H3K4me3 intensity ratio for each of them, and get the median ratio of these 100 genes. Finally, the intensity ratio of each gene in autophagy or oxidative phosphorylation signaling pathway was divided by the median ratio of these 100 control genes. Relative intensity ratios in Wnt signaling pathway or Jak-STAT signaling pathway (Group 4) were calculated with the same method, only that the control genes were identified from Group 3.
Yeast strains dot1Δ0 and hoΔ0 were constructed by PCR-mediated gene disruption with an auxotrophic marker gene URA3. Briefly, URA3 sequence was PCR amplified from plasmid pRS416, and the amplicon was transformed into haploid yeast strains BY4741 (MATa his3Δ1 leu2Δ0 met15Δ0 ura3Δ0) and BY4742 (MATα his3Δ1 leu2Δ0 lys2Δ0 ura3Δ0), respectively. Yeast cells were selected on synthetic complete medium with uracil dropped-off (SC-uracil); genomic DNA was extracted from colonies and PCR was performed to verify successful gene deletion. HO is a site-specific endonuclease required for homothallic switching, and is non-functional in strain s288c [65]. Thus, ho::URA3 was used here to control the potential noise effect of the auxotrophic marker URA3[66] in dot1::URA3. The DNA oligos used in this study are listed in S3 Table.
To estimate protein expression noise in dot1Δ0 and hoΔ0 strains, seven genes with high-intensity H3K79 methylation[47] and average protein level larger than 1000 (arbitrary unit) in Newman et al [11] were randomly chosen as reporters (S3 Table). Lowly expressed genes were excluded to ensure accuracy in quantifying florescence intensity. A series of GFP strains were generated on the background of BY4741 ho::URA3 and BY4741 dot1::URA3; in each of them, GFP was fused to the C-terminus of a reporter protein, following the protocol at Yeast Resource Center (http://depts.washington.edu/yeastrc/). In brief, GFP-KANMX6 cassette was PCR amplified from plasmid pFA6a-GFP(S65T)-KANMX6, and was transformed into BY4741 ho::URA3 and BY4741 dot1::URA3, respectively. A series of dTomato strains were generated similarly on the background of BY4742 ho::URA3 and BY4742 dot1::URA3, respectively. dTomato sequence and HYGMX4 were first PCR amplified from plasmids pRSET-B dTomato and pBS10, respectively, and were seamlessly ligated and cloned into PUC19 by GeneArt Seamless Cloning and Assembly Kit (Life Technology). This dTomato-HYGMX4 cassette was used to fuse dTomato to C-terminus of reporter proteins. Double fluorescence diploid strains with homozygous deletion of DOT1 (MATa/MATα dot1::URA3/dot1::URA3 GeneX-GFP-KANMX6/GeneX–dTomato-HYGMX4; GeneX is one of the reporter genes) were obtained by crossing BY4741 MATa dot1::URA3 GeneX-GFP-KANMX6 with BY4742 MATα dot1::URA3 GeneX–dTomato-HYGMX4 on YPD agar plate, followed by selection on YPD agar plate supplemented by G418 (AMRESCO, 200μg/ml) and hygromycin B (AMRESCO, 300μg/ml). Double fluorescence diploid strains with homozygous deletion of HO (MATa/MATα ho::URA3/ho::URA3 GeneX-GFP-KANMX6/GeneX–dTomato-HYGMX4) were generated similarly.
Yeast cells were cultured in YPD liquid media (1% Yeast extract, 2% Peptone, and 2% Dextrose, mass/volume) and were collected at the mid-log phase. Cells were washed by 1×PBS (137 mM NaCl, 2.7 mM KCl, 10 mM Na2HPO4, and 1.8 mM KH2PO4, pH 7.4) and were placed on ice.
The fluorescence intensities of EGFP and dTomato in single cells were measured with FACSAria III cell sorter (BD Biosciences). GFP was excited by 488nm laser and was detected through 530/30 nm emission filter; dTomato was excited by 561nm laser and was detected through 610/20 nm emission filter. Cells containing single fluorescence protein were also prepared to perform fluorescence compensation in FACS. Three biological replicates were performed for each sample.
Cells were gated by FSC-A, SSC-A, and FSC-W/FSC-H ratio to exclude cells with extraordinary size or complexity, as well as doublets and cell bulks. Then cells with both green and red fluorescence were identified, and were used for subsequent analysis. About 40,000 double-fluorescence events were recorded in each replicate.
Intrinsic noise was calculated following previous studies [1,3].
μ=g¯r¯
(11)
IntrinsicCV2=σ2μ2=(g−r)2¯2g¯r¯
(12)
where g and r are the normalized EGFP intensity and dTomato intensity, respectively, in single cells.
Burst frequency was estimated based on gamma distribution [67]. The intensity of EGFP fluorescence was normalized with the dTomato intensity of the same cell and single-cell fluorescence intensity of normalized EGFP was fit to a gamma distribution, in which process rate parameter (β) and shape parameter (α) were estimated. Burst frequency was calculated with the following equation.
Yeast cells were grown to mid-log phase and washed once by 1×PBS (pH 7.4). Fluorescent images were acquired using confocal microscope (DIGITAL ECLIPSE C1Si, Nikon, Japan) equipped with 488 nm and 543 nm lasers, under an oil-immersed objective at ×100 magnification.
Yeast cells were cultured in YPD liquid media (1% Yeast extract, 2% Peptone, and 2% Dextrose, mass/volume), which were collected at log phase (50 ml, OD660 around 0.7), washed, and immediately suspended in 400μl lysis buffer (50 mM HEPES-KOH, pH 7.5, 140 mM NaCl, 1 mM EDTA, 1% Triton X-100, 0.1% sodium deoxycholate, and Protease Inhibitor Cocktail). The suspension was mixed with glass beads, vortexed for 10 min at 4°C, and sonicated without beads for 50 times on ice (10-sec pulse at 195W followed by 20-sec rest). After 20-minute centrifugation at 10,000g, the supernatant was transferred to a new tube and boiled for 5 minutes in SDS sample buffer. After centrifugation (14,000g for 5 minutes), 40μl of the resulting lysates was subjected to western blot assay. Anti-H3K79 antibody, Anti-H3 antibody, and Anti-GAPDH antibody were purchased from Abcam (ab3594, ab1791 and ab9484, respectively). H3 and GAPDH were used as loading controls. The anti-H3K79 antibody (ab3594) binds to all three H3K79 methylations. Primary antibodies were used at 1:8000 dilution. Horse radish peroxidase (HRP)-conjugated secondary antibodies were purchased from Cell Signaling Technology (Cat. # 7074, 1:5,000 dilution; # 7076, 1:10,000 dilution).
|
10.1371/journal.pcbi.1004822 | Systems Pharmacology and Rational Polypharmacy: Nitric Oxide−Cyclic GMP Signaling Pathway as an Illustrative Example and Derivation of the General Case | Impaired nitric oxide (NO˙)-cyclic guanosine 3', 5'-monophosphate (cGMP) signaling has been observed in many cardiovascular disorders, including heart failure and pulmonary arterial hypertension. There are several enzymatic determinants of cGMP levels in this pathway, including soluble guanylyl cyclase (sGC) itself, the NO˙-activated form of sGC, and phosphodiesterase(s) (PDE). Therapies for some of these disorders with PDE inhibitors have been successful at increasing cGMP levels in both cardiac and vascular tissues. However, at the systems level, it is not clear whether perturbation of PDE alone, under oxidative stress, is the best approach for increasing cGMP levels as compared with perturbation of other potential pathway targets, either alone or in combination. Here, we develop a model-based approach to perturbing this pathway, focusing on single reactions, pairs of reactions, or trios of reactions as targets, then monitoring the theoretical effects of these interventions on cGMP levels. Single perturbations of all reaction steps within this pathway demonstrated that three reaction steps, including the oxidation of sGC, NO˙ dissociation from sGC, and cGMP degradation by PDE, exerted a dominant influence on cGMP accumulation relative to other reaction steps. Furthermore, among all possible single, paired, and triple perturbations of this pathway, the combined perturbations of these three reaction steps had the greatest impact on cGMP accumulation. These computational findings were confirmed in cell-based experiments. We conclude that a combined perturbation of the oxidatively-impaired NO˙-cGMP signaling pathway is a better approach to the restoration of cGMP levels as compared with corresponding individual perturbations. This approach may also yield improved therapeutic responses in other complex pharmacologically amenable pathways.
| Developing drugs for a well-defined biochemical or molecular pathway has conventionally been approached by optimizing the inhibition (or activation) of a single target by a single pharmacologic agent. On occasion, drug combinations have been used that generally target multiple pathways affecting a common phenotype, again by optimizing the extent of inhibition of individual targets, semi-empirically adjusting their doses to minimize toxicities as they are manifest. Here, we present a computational approach for identifying optimal combinations of agents that can affect (inhibit) a well-defined biochemical pathway, doing so at minimal combined concentrations, thereby potentially minimizing dose-dependent toxicities. This approach is illustrated computationally and experimentally with a well-known pathway, the nitric oxide-cyclic GMP pathway, but is readily generalizable to rational polypharmacy.
| Signal transduction via the nitric oxide (NO˙)-cyclic guanosine 3', 5'-monophosphate (cGMP) pathway is involved in multiple and diverse biological responses, including smooth muscle relaxation, inhibition of platelet aggregation, and neural communication [1–6]. This pathway is composed of several molecular species acting in two opposing limbs, the cGMP-synthetic limb and the cGMP-degradative limb (see Fig 1). The proper function of these two limbs is crucial in controlling these biological responses. Within the cGMP-synthetic limb, NO˙ binds to soluble guanylyl cyclase (sGC) to catalyze the production of cGMP from guanosine-5'-triphosphate (GTP), whereas in the cGMP-degradative limb, cyclic nucleotide phosphodiesterase (PDE) converts cGMP to GMP. Impaired function of either or both limbs of the NO˙-cGMP signaling pathway has been reported in many cardiovascular disorders, including heart failure and pulmonary arterial hypertension.
Importantly, increased oxidative stress associated with malfunction of the NO˙-cGMP signaling pathway has been implicated in the pathobiology of several diseases [7, 8]. During oxidative stress, the pathway’s unresponsiveness can be explained by several mechanisms, among which sGC insensitivity to NO˙ (tolerance) is decisive. Elevated reactive oxygen species (ROS) may promote sGC insensitivity through either non-heme (cysteine) oxidation of sGC [9–14], S-nitrosation of sGC [15], heme oxidation of sGC [16], or oxidation of NO˙, such as via enhanced peroxynitrite (ONOO-) formation [17]. Potentially, there are several determinants of this oxidatively-adapted pathway, including oxidatively inactivated sGC itself, oxidized NO˙, and PDE. The pharmacological challenge is how best to deploy potential therapeutic options that focus on these determinants under increased oxidative stress in a way that optimizes restoration of the function of this pathway.
Investigating the complexity of biological systems using combinatorial perturbations is a rational strategy for predicting function and phenotype [18], understanding network mechanisms [19–22], and identifying new and more promising therapeutic targets for human diseases [23, 24]. In theory, using a combination of drugs that can perturb different components of a system could be a more effective strategy than treating a disease with a single drug [25]. Indeed, the most complex diseases, such as cardiovascular diseases, cancer, diabetes mellitus, neurodegenerative diseases, and asthma, are multifactorial diseases. Systems-based interventions using multi-component drug combinations have been used increasingly to treat these complex diseases, although these approaches have largely been developed empirically in the clinical setting. The main purpose of model-based drug discovery is to revisit classical pharmacology logically in order to replace the one-gene, one-protein, and one-mechanism perspective with a systems-oriented paradigm to improve the therapeutic index of potential drugs targeting these complex diseases [26, 27].
Relevant principles have emerged from different studies of combination therapies that do not always yield predicted outcomes. For example, the combination of niacin (vitamin B3) with a statin [5-hydroxy-3-methylglutaryl-coenzyme A (HMG-CoA) reductase] leads to an incremental decrease in low-density lipoprotein (LDL) cholesterol concentration and an increase in high-density lipoprotein (HDL) cholesterol concentration [28]. Combinations of drugs that perturb five different targets in the HIV life cycle have turned AIDS from a lethal infection into a manageable chronic disease [29]. Another interesting combination is that of nitroglycerin and N-acetylcysteine (NAC), which can potentiate the effects of nitroglycerin in the treatment of acute myocardial ischemia [30]. The combination of β2-adrenergic receptor activators with muscarinic receptor blockers is useful for the treatment of chronic obstructive pulmonary disease [31].
Opposing and independent regulatory mechanisms within the NO˙-cGMP pathway determine the biological level of cGMP in the steady-state. Model-based approaches have facilitated our understanding of these regulatory mechanisms for cGMP formation [32–34]. Thus far, modeling has been used to study two distinct limbs of the NO˙-cGMP signaling pathway separately [33–39]; however, here, we will build this model as an integrated system that also includes oxidative inactivation. We then pose the question of whether a combination of two or three agents with orthogonal therapeutic actions (and toxicities), used at lower concentrations than when used alone, will enhance cGMP formation beyond that of single agents in the presence of oxidative stress.
In this study, we pursued this question using a dynamical model of the NO˙-cGMP signaling pathway in the presence of hydrogen peroxide. Impaired activation of NO˙-cGMP signaling has been observed in several cardiovascular disorders, including heart failure [40] and pulmonary arterial hypertension [41], due, in part, to excess oxidants. Current treatments for these disorders that involve this pathway includes nitrovasodilators and phosphodiesterase inhibitors. Thus, this theoretical approach, were it to demonstrate benefit, may offer initial strategies for optimal drug combinations for the treatment of these (and other) disorders in which NO˙-cGMP signaling is dysfunctional using approved agents.
Previous studies have used models of cell signaling networks to evaluate the action of drug pairs as compared with corresponding individual drugs [42]. Modeling drug action using ordinary differential equations could be challenging without sufficient information about the integrated network kinetics of drug action. To address this challenge, Araujo and colleagues investigated an interesting concept, perturbation simulation, on the epidermal growth factor receptor (EGFR) signaling network. They found that pairwise perturbations of reaction rates was more effective at restoring optimal function of the network than individual perturbation of corresponding reaction rates [43].
We expanded this approach to a practically remediable pathway and examined how hydrogen peroxide (H2O2)-induced oxidant stress affects the key reaction steps of the NO˙-cGMP signaling pathway to diminish cGMP levels, and then developed a combinatorial approach to perturb the oxidatively-impaired NO˙-cGMP signaling pathway and restore cGMP levels toward normal. In addition, in contrast to [43], we also examined the consequences of lesser degrees of inhibition in combination modeling to infer lower dose-dependent toxicity. Lastly, in contrast to [43], we performed cell-based experiments to validate the modeling strategy. To do so, all rate constants were perturbed individually, in pairs, or in trios by step-wise ten-fold changes in their values to their original values, which is analogous to concentration-dependent inhibition of a given reaction by a specific inhibitor. Our goal was to identify an optimal perturbation that augments the cGMP levels toward normal during oxidative stress.
Using a single perturbation, we found that the potential therapeutic targets, including the oxidation of sGC, NO˙ dissociation from sGC, and cGMP degradation by PDE, had a profound effect on enhancing cGMP accumulation as compared with other reaction steps. Using combined perturbations, we were able to identify an optimal triple perturbation that increases cGMP levels beyond that observed with the corresponding individual or paired perturbations that comprise it. Importantly, these theoretical findings were confirmed in cell-based experiments in which a combination of a nitric oxide donor (S-nitroso-N-penicillamine), an antioxidant (N-acetyl-L-cysteine), and a phosphodiesterase type 5-inhibitor (sildenafil) significantly improved the cyclic GMP output of the pathway in the setting of oxidant stress (hydrogen peroxide) in pulmonary artery vascular smooth muscle cells.
Elevated reactive oxygen species (ROS) can affect both the cGMP-synthetic limb [9, 10] and the cGMP-degradative limb [44] of the NO˙-cGMP signaling pathway. In order to assess the combinatorial effects of different pharmacological modulators on dynamical pathway behavior (as determined by cGMP output), we first modeled the dynamical behavior and then assessed the effects of optimal combinations of pharmacological inhibitors in cell-based assays using pulmonary artery vascular smooth muscle cells.
Once NO˙ is generated in source cells (endothelial cells) in the vasculature, it diffuses into vascular smooth muscle cells and binds to sGC, a ferrous iron hemoprotein receptor, to generate the NO˙-sGC complex. Either sGC alone or the NO˙-sGC complex, whose specific activity is ~200 times greater than sGC alone [45], can convert GTP to the second messenger molecule, cGMP, which is degraded by cyclic nucleotide PDE(s) to GMP. However, under oxidative stress conditions, sGC is also oxidized by H2O2 and thereby desensitized (Fig 1). The biological reactions comprising this system were modeled using ordinary differential equations and mass action kinetics involving 12 molecular species and 13 rate constants (S1 Table).
The simulation time intervals were selected to monitor cGMP dynamics between 0 and 200 seconds (based on the cGMP dynamics in S1A Fig for pulmonary artery vascular smooth muscle cells). Next, an oxidant (500 μM H2O2) was added to the system to alter the dynamics of all molecular species, including cGMP, as compared with control. One prior experimental study showed that both cGMP-degrading enzymes and sGC desensitization cooperatively accounted for the diverse patterns of cGMP responses to NO˙. Two different temporal dynamic signatures of cGMP were reported within platelets and astrocytes that have high and low levels of PDEs, respectively [32, 46, 47]. In our system under control conditions, the cGMP concentrations increased abruptly to a peak concentration within 40 seconds, and then decreased to baseline within 200 seconds (mirroring the experimental dynamics of S1A Fig). When the system was exposed to H2O2, the cGMP levels decreased by ~6-fold (Fig 2A). In this model, we proposed that H2O2 impairs activation of sGC and its generation of cGMP [48]; however, conflicting results have been reported by others [49].
To restore cGMP generation to the normal level, there are several potential therapeutic targets, including oxidatively inactivated sGC, the NO˙-activated form of sGC, and PDE. Therapies for some diseases with PDE inhibitors have been successful at increasing cGMP levels in both cardiac and vascular tissues. However, to predict which one of these potential targets would be most effective at increasing cGMP levels, we perturbed either the synthetic limb (k3) or the degradative limb (k10) of the pathway in the absence or the presence of H2O2, and then evaluated cGMP dynamics. We found that targeting these two reaction steps can significantly increase the cGMP levels as compared with control if there is sufficient unoxidized sGC available. However, under significant oxidative stress, targeting these two reaction steps cannot be an effective strategy for restoring cGMP levels to normal (Fig 2B and 2C).
To evaluate more systematically the role of any given reaction in cGMP formation, we compared cGMP dynamics by reducing each of the thirteen rate constants to 10% of its original value (simulating significant reaction inhibition) in the presence of H2O2. We found that the cGMP levels were not restored toward control levels by decreasing k1 (oxidizing sGC), k3 (desensitizing sGC), or k10 (degrading cGMP) to 10% of their original values (Fig 2D). This observation suggested that under oxidative stress, targeting either the synthetic limb or the degradative limb of the pathway alone is not an effective approach for restoring cGMP to normal levels.
The relative involvement of both synthetic and degradative components of the NO˙-cGMP signaling pathway led us to propose that these components could exert autonomous effects on cGMP accumulation. This concept raised the possibility that combined perturbations may have more profound effects on cGMP levels than single perturbations. Addressing this concept, the NO˙-cGMP pathway was perturbed using all possible single, paired, or triple perturbations in the presence of H2O2, and then the time-integrated cGMP (cGMPT) levels were calculated (Fig 3). We found that the simultaneous perturbation (ρ) of several rate processes along with the perturbation of k1 (ρk1) yielded the highest cGMPT levels relative to other perturbations. This finding suggested that targeting the primary driver of pathway dysfunction (ρk1) along with other potential therapeutic targets might be a better approach for increasing cGMP levels (even) beyond control levels under oxidative stress.
We next perturbed the proposed rate constants, including ρk1, ρk3, ρk10, or all possible combinations of these three rate constants, in the presence of H2O2. Modeling a fractional linear reduction of values for these rate constants, we created a vector of eleven different values for each wherein the maximum value for each rate constant was its original value (S1 Table) and the minimum value was 0.1, 0.3, or 0.5 of its original value for a rate constant in single, paired, or triple perturbations, respectively (lesser minimal values were used with greater combined perturbations to attempt to demonstrate efficacy at combined doses that might limit dose-dependent toxicities). Rate constants that were not varied under each set of modeling conditions were maintained at their full values. The cGMP dynamics was then calculated using the range of rate constants (Fig 4). These results suggested that under oxidative stress, decreasing dissociation of NO˙ from the NO˙-sGC complex (ρk3) is the most sensitive reaction step for increasing cGMP levels as compared with the use of an anti-oxidant (ρk1) and PDE inhibitor (ρk10). Furthermore, perturbation of k1 (ρk1) is the best strategy by which to increase cGMP levels beyond perturbation of either k3 (ρk3) or k10 (ρk10).
Subsequently, we used the Bliss model [50] (which, based on probability theory, assumes two inhibitors work through independent mechanisms of action, and assumes that the two inhibitors do not interfere or compete with each other) to evaluate the power of paired perturbations. Under oxidative stress, optimal parameters were perturbed either individually or in pairs in order to compare the effects of perturbations on cGMP levels. The effects of single perturbations on cGMP levels were used to calculate the Bliss model, as indicated by eq (15). As depicted in Fig 5, paired perturbations of optimized single parameters increased cGMP levels beyond the Bliss model predictions. To assess whether these differences are additive or non-additive, we used an isobologram analysis. We examined the combined effects on cGMPT when two or three rate constants were perturbed simultaneously. The isobologram (contour plot) [51–56] was used to quantify the combined effects (Fig 6), wherein we observed that combined perturbations act additively to increase cGMP levels in this system.
We studied cGMP dynamics using human embryonic kidney (HEK) 293 cells and human pulmonary artery vascular smooth muscle (PAVSM) cells. PAVSM contain abundant PDE5 compared with HEK293 cells (35), thereby ensuring that both the cGMP-synthetic and degradative limbs determine the cGMP levels (S1 Fig). Thus, in PAVSM, the rapid accumulation of cGMP is followed by its equally rapid reduction (S1A Fig). In HEK293 cells, which contain lower amounts of PDE5 (35) as compared with PAVSM cells, the cGMP-synthetic limb of the pathway primarily determines the cGMP levels (S1B Fig). The cells were pretreated with either H2O2 or buffer for 30 minutes. Time points were selected to capture cGMP dynamics. When the HEK293 cells were exposed to H2O2 at 500 μM, we found that NO˙-stimulated cGMP production was significantly reduced as a function of time (S1B Fig). This result suggested that H2O2 blocked the cGMP-synthetic limb of the pathway, which plays a predominant role in determining the cGMP levels in HEK293 cells (as confirmed by the absence of a biphasic response in cGMP dynamics compared with the PAVSM cells in S1 Fig).
In order to determine the validity of the combinatorial modeling described above, we measured cGMP in PAVSM cells treated with various combinations of agents that act on different steps in the pathway of Fig 1. Agents were chosen because they have been used in human studies, and because they affect each of the limbs of the pathway in Fig 1. As shown in Fig 7A, we first showed that the addition of a NO.-donor, S-acetyl-N-penicillamine (SNAP), increased the cGMP produced by 58% over vehicle-treated control cells; hydrogen peroxide treatment, however, abrogated this increase. When cells were treated with hydrogen peroxide and the reducing agent, N-acetyl-L-cysteine (NAC), cGMP levels increased to ~2-fold above control. The addition of sildenafil, a PDE5 inhibitor (the primary PDE isoform found in PAVSM responsible for cGMP degradation), to SNAP and NAC in the presence of hydrogen peroxide further increased cGMP levels to ~3.2-fold above vehicle-treated control cells. With these baseline measurements, we next explored key comparative combinations of agents that mimicked the optimal modeled combinations, as shown in Fig 7B. Here, cGMP responses are reported as the % of the maximal response (to sildenafil and SNAP in the absence of hydrogen peroxide) owing to variation from experiment to experiment. We observed that the combinations of NAC and SNAP, sildenafil and SNAP, and sildenafil, NAC, and SNAP each increased cGMP in the presence of hydrogen peroxide, and that the relative magnitude of the increases was consistent with the modeled data in Fig 3. The use of NAC inhibits reactions 1 (and possibly 13), the use of SNAP ‘inhibits’ reaction 3 indirectly by driving reaction 2, and the use of sildenafil inhibits reaction 10 as a competitive inhibitor of PDE5 and indirectly inhibits reaction 12 by limiting the formation of the catalytic complex and hence substrate turn-over. Thus, as in Fig 3, the magnitude of increase in cGMP was of the following order: inhibition of reactions 1 + 3 < inhibition of reactions 3 + 10 (or 12) < inhibition of reactions 1 + 3 + 10, which is similar to the experimental reaction order we observed in the data of Fig 7.
Impaired activation of NO˙-cGMP signaling pathway has been observed in cardiovascular disorders and other common disease states. There are multiple enzymatic determinants of cGMP production in this pathway, including sGC itself, the oxidatively inactivated form of sGC, the NO˙-activated form of sGC, enzymatic sources of NO˙, and PDE. Therapies for these disorders with PDE inhibitors have been successful at enhancing cGMP levels in cardiac and vascular tissue with attendant improvement in lusitropy and vasodilation, respectively. However, PDE is only one of the enzymatic determinants of cGMP formation. In this systems-level approach, we used all possible single, paired, or triple perturbations to propose a combined perturbation that was more effective in cGMP accumulation than any single perturbation. The optimal number of the perturbations was three owing to there being only three key processes that determine independently the cGMP levels, i.e., cGMP synthesis, cGMP degradation, and oxidative inactivation of sGC. By having this modeled information, one can improve experimental design, curb cost, and save time in performing the experiments necessary for gaining useful results. Alternatively, one could randomly target any given component of this pathway either individually or in combination with other components of the pathway. Yet another approach is the maximal damage targeting strategy [57], theoretically a better approach relative to the random targeting of a pathway. However, using either the random targeting or the maximal damage targeting approach, we might overlook the optimal perturbations among many combinations that may never have been tested.
PDEs are essential enzymes within normal cells that degrade the phosphodiester bond in the second messenger molecules cAMP and cGMP. PDEs are, therefore, important regulators of signal transduction mediated by these second messenger molecules. As with many drugs that affect molecular pathways involved in (many) different signaling pathways, the side effects of PDE inhibitors are dose-dependent [58]; thus, to reduce the dose of a PDE inhibitor and then combine it with other potential drugs that have non-overlapping mechanisms of action and toxicities may significantly improve the overall therapeutic index of the treatment strategy.
One of the rationales for using combination therapy is to block redundant pathways that exist extensively within the molecular networks whose functions are modified in human diseases. To overlook this network property may limit the potential for reformulating existing drugs that can be used in combination with higher efficacy and fewer toxicities. Our results show how the combined perturbations of the NO˙-cGMP signaling pathway represent a useful strategy for increasing cGMP levels. A model-based analysis suggests that the combinatorial perturbation of biological networks is a promising approach by which to identify drug combinations with higher efficacy and perhaps lower toxicity (rational polypharmacy) [42]. Further work on other specific pathways will be required to validate the general approach.
Enzyme immunoassay (EIA): human pulmonary artery vascular smooth muscle (PAVSM) cells and growth media were obtained from Lonza Inc. (Walkersville, MD., USA). Confluent cells were pre-treated with phenol-red free DMEM (supplemented with 10% fetal calf serum) in the presence of absence of 10 mM N-acetyl-L-cysteine (NAC), a thiol reducing agent and antioxidant, to reverse mildly oxidized critical sulfhydryl groups in sGC [14] and, possibly, to prevent the oxidation of NO to NOx; and/or 100 nM sildenafil, a PDE5 inhibitor, for 90 min followed by 500 μM H2O2 for 30 min. Cells were next treated with either phosphate-buffered saline or 100 μM S-nitroso-N-acetylpenicillamine (SNAP), a NO.-donor, for 10 min. PAVSM cells were rinsed in ice-cold phosphate buffered saline and then solubilized in ice-cold 6% trichloroacetic acid. Samples were stored at -80° until the day of the assay. Samples were processed and cGMP and protein were measured as previously described [14]. cGMP formation was measured by immunoassay according to the cGMP Assay (Cayman Chemical Co., Ann Arbor, MI). H2O2, trichloroacetic acid, NAC, and sildenafil were purchased from Sigma-Aldrich (St. Louis, MO). Phenol-red-free DMEM was obtained from Gibco, Life Technologies, Grand Island, NY and fetal calf serum was from Atlanta biologicals, Flowery Branch, GA.
Assuming mass-action kinetics, the reaction scheme (Fig 2, S1 Table) was deconstructed into 12 ordinary differential equations (ODEs):
d [H2O2]/dt = −k1 [ H2O2] * [sGC]
(1)
d [sGC]/dt = − k1 [H2O2] * [sGC]−k2 [NO.]* [sGC] +k3 [NO.−sGC]−k4 [ sGC] * [GTP] + (k8 +k5) [sGC−GTP]
(2)
d [sGC−H2O2]/dt = k1 [ H2O2] * [sGC]
(3)
d [NO.]/dt = −k2 [ NO.] * [sGC]+k3 [NO.−sGC]−k13[NO.]
(4)
d [NO.−sGC]/dt = k2 [ NO.] * [sGC]−k3 [NO.−sGC]−k6 [GTP]* [NO.−sGC] + (k9+k7) [NO.−sGC−GTP]
(5)
d [GTP]/dt = −k4 [ GTP] * [sGC]+k5 [sGC−GTP]−k6 [GTP]* [NO.−sGC] + k7 [NO.−sGC−GTP]
(6)
d [sGC−GTP]/dt = k4 [ GTP] * [sGC]−(k5+k8) [sGC−GTP]
(7)
d [NO.−sGC−GTP]/dt = k6 [ GTP] * [NO.−sGC]−(k7+k9) [NO.−sGC−GTP]
(8)
d [cGMP]/dt = k8 [ sGC−GTP] +k9 [NO.−sGC−GTP]−k10[cGMP]*[PDE] +k11[cGMP−PDE]
(9)
d [PDE]/dt = −k10 [ cGMP] *[PDE]+(k11 +k12) [cGMP−PDE]
(10)
d [cGMP−PDE]/dt = k10 [ cGMP] *[PDE]−(k11 +k12) [cGMP−PDE]
(11)
d [GMP]/dt = k12 [cGMP−PDE]
(12)
to simulate the dynamics of the molecular species within the NO˙-cGMP signaling pathway. The ODEs were solved using a numerical ODE solver (ode15s). All mathematical modeling and simulations were performed using the SimBiology toolbox in MatLab (Version 8, 2012b, MathWorks, Natick, MA). The parameter values for this model include 13 rate constants and 12 initial concentrations (S1 Table), which were chosen or estimated from the literature, as indicated in the Table. The system dynamics were assessed in the absence or presence of H2O2.
In the presence of H2O2 (500 μM), the NO˙-cGMP pathway was perturbed by assuming the presence of an effective inhibitor of a given reaction(s) sufficient to impair the reaction kinetics (ρk = 0.1k). We perturbed all possible individual (13), pairs (78), or trios (286) of reactions in the model. The total number of perturbations (up to triple perturbations) was computed by inserting the total number of rate constants (q = 13) and the maximum number of perturbations (p = 3) into following equation:
Cqp=∑i=1pq!/[(q−i)! × i!
(13)
Thus, the total possible number of perturbations (for triple perturbations) is 13C3 = 377. To assess the relative role of each perturbation, cGMP dynamics were illustrated (Fig 3).
The time-integrated cumulative cGMP (cGMPT) level is defined as:
cGMPT = ∫0T[cGMP] (t) dt , T= 200sec
(14)
Perturbation of k1, k3, or k10 alone can induce a dose-dependent cGMPT response in the presence of H2O2. We varied the rate constants by fractional linear decrements.
To depict the matrix response of cGMPT, two vectors of rate constant values were combined in 11×11 matrices where the value of each rate constant is depicted along each axis (Fig 6). The contour plots were used to evaluate additive and non-additive effects. Thus, the combined actions were either additive (Fig 6) or synergistic if the isobole is a straight line or a convex line, respectively. Likewise, 11×11×11 matrix of all model parameters was constructed for triple perturbations and three-dimensional contour plots analyzed accordingly.
|
10.1371/journal.ppat.1000731 | Persistent ER Stress Induces the Spliced Leader RNA Silencing Pathway (SLS), Leading to Programmed Cell Death in Trypanosoma brucei | Trypanosomes are parasites that cycle between the insect host (procyclic form) and mammalian host (bloodstream form). These parasites lack conventional transcription regulation, including factors that induce the unfolded protein response (UPR). However, they possess a stress response mechanism, the spliced leader RNA silencing (SLS) pathway. SLS elicits shut-off of spliced leader RNA (SL RNA) transcription by perturbing the binding of the transcription factor tSNAP42 to its cognate promoter, thus eliminating trans-splicing of all mRNAs. Induction of endoplasmic reticulum (ER) stress in procyclic trypanosomes elicits changes in the transcriptome similar to those induced by conventional UPR found in other eukaryotes. The mechanism of up-regulation under ER stress is dependent on differential stabilization of mRNAs. The transcriptome changes are accompanied by ER dilation and elevation in the ER chaperone, BiP. Prolonged ER stress induces SLS pathway. RNAi silencing of SEC63, a factor that participates in protein translocation across the ER membrane, or SEC61, the translocation channel, also induces SLS. Silencing of these genes or prolonged ER stress led to programmed cell death (PCD), evident by exposure of phosphatidyl serine, DNA laddering, increase in reactive oxygen species (ROS) production, increase in cytoplasmic Ca2+, and decrease in mitochondrial membrane potential, as well as typical morphological changes observed by transmission electron microscopy (TEM). ER stress response is also induced in the bloodstream form and if the stress persists it leads to SLS. We propose that prolonged ER stress induces SLS, which serves as a unique death pathway, replacing the conventional caspase-mediated PCD observed in higher eukaryotes.
| Trypanosomes are the causative agent of major parasitic diseases such as African sleeping sickness, leishmaniasis and Chagas' disease that affect millions of people mostly in developing countries. These organisms diverged very early from the eukaryotic linage and possess unique molecular mechanisms such as trans-splicing and RNA editing. Trypanosomes lack polymerase II promoters that govern the transcription of protein coding genes. Eukaryotes respond to unfolding of proteins in the endoplasmic reticulum (ER) by a distinct transcriptional programming known as the unfolded protein response (UPR). In this study, we demonstrate that despite the lack of transcriptional regulation, procyclic trypanosomes change their transcriptome as a response to ER stress by differential mRNA stabilization. Prolonged ER stress induces a unique process, the spliced leader RNA silencing (SLS), that shuts off the trans-splicing and the production of all mRNAs. SLS is induced both by prolonged ER stress and by knock-down of factors involved in ER translocation in both life stages of the parasite. SLS induces programmed cell death (PCD) evident by the hallmark of apoptosis in metazoa (DNA fragmentation, membrane flipping and ultrastructural changes). We propose that SLS serves as a unique death pathway replacing the conventional caspase-mediated PCD observed in higher eukaryotes.
| The endoplasmic reticulum (ER) is the site for folding of proteins that traverse the ER on the way to the outer membrane or internal organelles of the vesicular transport machinery. The ER is also the site of lipid synthesis and it maintains the cellular calcium homeostasis [1]. The ER performs these functions using resident chaperones, high levels of calcium and an oxidating environment. Proteins that translocate to the ER undergo proper folding and those can exit the ER. However, if the protein fails to fold properly, it is exported back to the cytoplasm and is degraded by the proteasome [2],[3]. ER function is sensitive to changes in calcium level, inhibition of glycosylation, oxidative stress, and exposure to reducing agents. These conditions induce the ER stress response, which triggers specific signaling pathways including the unfolded proteins response (UPR), which leads to reduction in the load of proteins to be translocated, enhanced degradation of misfolded proteins, and increased folding capacity of the ER [3]–.
In yeast, UPR is dependent on IRE1, which is a bi-functional kinase and endonuclease that cleaves a non-conventional intron from HAC1 mRNA, which encodes a transcription factor (bZIP) [6],[7]. Homologues proteins exist in mammals; IRE1 cleaves an intron in XBP1 encoding a potent bZIP transcription factor. This transcription factor activates genes essential for executing the UPR [8]. In mammals, two additional ER transmembrane proteins, PERK and ATF6, participate in UPR. PERK phosphorylates eIF2α to mediate translational attenuation. ATF6 is cleaved in the Golgi and translocates to the nucleus to activate the transcription of chaperones, such as BiP [9]–[12]. When the ER stress is persistent, the cells activate the apoptotic pathway via the activation of Caspase 12 [13],[14].
Trypanosomatids are known for their non-conventional gene expression mechanisms, including trans-splicing [15],[16] and RNA editing [17],[18]. Trans-splicing is required for the maturation of all mRNAs in these parasites. In this process, a small exon, the spliced leader (SL), encoded by a small RNA, the SL RNA, is donated to all pre-mRNA by trans-splicing [16]. No promoters upstream to protein coding genes transcribed by polymerase II were identified. Protein coding genes are arrayed in long polycistronic transcription units which are processed post-transcriptionally by the concerted action of trans-splicing and polyadenylation [15]. It is believed that gene expression in these parasites is regulated primarily post-transcriptionally at the level of mRNA degradation and translation; for most genes, the signals that dictate this regulation are situated at the 3′ UTR [19]. However, trypanosomes do not completely lack transcriptional regulation. The SL RNA has a defined promoter that binds specific transcription factors such as tSNAPc [20]–[22].
Our collective knowledge of the regulation of trypanosome gene expression is consistent with the novel stress-induced regulatory mechanism that controls the transcriptome under variety of stresses; this mechanism was named SLS, for SL RNA silencing [23]. SLS was discovered in cells silenced for the signal recognition particle receptor, SRα. Depletion of SRα results in a dramatic decrease in the level of SL RNA, leading to major reduction in mRNAs. Nascent transcription in permeable cells showed that SL RNA transcription is specifically extinguished under these conditions. ChIP assay further demonstrated that under SRα depletion, the SL RNA transcription factor, tSNAP42, fails to bind to the SL RNA promoter. The hallmark of SLS is therefore shut-off of SL RNA transcription, followed by massive accumulation of tSNAP42 in the nucleus [23].
Most recently, a very detailed study on the trypanosome transcriptome was performed and revealed significant changes between procyclic and bloodstream form trypanosomes. However, the same study failed to detect changes generally elicited by classical UPR inducers or changes that are generally elicited under serum starvation [24]. It is currently unknown whether trypanosomes lack the UPR mechanism, which is replaced by SLS in these organisms, or whether both mechanisms exist, and SLS is activated to induce PCD when UPR fails. Accumulated data suggest that PCD exists in trypanosomatids, and plays a major role in maintaining and regulating the parasite population [25],[26]. In metazoa, type I PCD is induced by caspases. Effector caspases activate proteases and nucleases, eventually leading to apoptosis. Trypanosomes lack homologues to metazoan caspases, but have five metacaspases [27]–[29]. However, up till now, it is believed that these metacaspases are not involved in apoptosis [27],[29]. In recent years, PCD pathways that are caspase-independent, such as autophagy were shown to exist. Autophagy functions in protein degradation and organelle turnover during stresses, such as nutrient and growth factor deprivation, to assure cell survival. Autophagy machinery can be also recruited to kill cells under certain conditions, generating a caspase-independent form of PCD [30]. Recent studies indicate that trypanosomatids possess such an autophagic pathway [31],[32].
In this study, we demonstrate that despite the lack of IRE1/XBP1 homologues that mediate the UPR in all eukaryotes, trypanosomes change their transcriptome in response to ER stress much like yeast and metazoa. The mechanism of ER stress response is unique, and is based on stress-induced mRNA stabilization of mRNAs, which are essential for executing the response. SLS is triggered under prolonged ER stress induced either by chemicals or by silencing of SEC63 and SEC61- two major components of ER translocation machinery. Typical hallmarks of PCD such as DNA fragmentation, exposure of phosphatidyl serine, increase in cytosolic calcium, reduction in mitochondria membrane potential and ROS production, were observed in SLS-induced cells. We propose that SLS is the apoptotic branch induced under persistent ER stress that leads to programmed cell death in both procyclic and bloodstream form trypanosomes.
To examine whether procyclic stage trypanosomes possess an ER stress response, the transcriptome was examined for up and down-regulation upon Dithiothreitol (DTT) treatment. RNA was prepared from cells treated with 4 mM DTT for 1 and 3 hours and the transcriptome was compared to untreated cells. The Lowess-normalized data were used to identify genes whose expression appeared to be significantly (P<0.05) up-regulated with an arbitrary cutoff of 1.5 fold. A list of these genes is provided in Table S2. In order to visualize the difference in gene expression between normal and under ER stress, we compared the changes of transcriptome after 1 and 3 hours of treatment. Hierarchical clustering was performed on the regulated genes, and a heat map was constructed (Fig. 1-A). The results demonstrate that the changes in the transcriptome between 1 and 3 hrs of treatment were similar in both up- and down-regulated genes (r = 0.86), indicating that the changes were treatment-specific (Fig. 1-B). Moreover, the amplitude of the up and down-regulation was increased by 1.24 fold when comparing the differential expression from 1 to 3 hours of treatment.
Inspection of the up-regulated genes (Table S2) suggests that many genes which are involved in the core processes of classic UPR, namely, protein folding, degradation, translocation, sorting, and lipid metabolism, were up-regulated. Additionally, genes involved in mitochondrial functions, redox balance, metabolism, cytoskeleton, and movement were also up-regulated. Interestingly, a large number of genes involved in signal transduction and gene expression were affected, as well.
To compare the response of trypanosomes under ER stress to that of other eukaryotes, microarray data from UPR induced cells of Saccharomyces cerevisiae [3], Caenorhabditis elegans [33], Drosophila melanogaster [34], and Homo sapiens [35] were downloaded and analyzed. Genes found to be up-regulated were classified into the same functional categories (Table S3) that were already defined for the trypanosome transcriptome (Table S2). Note that for S. cerevisiae, C. elegans, and D. melanogaster, the list of up-regulated genes was refined by filtering out genes whose expression under ER stress was not XBP-1 or IRE-1 dependent and therefore are not directly related to the UPR response. In trypanosomes and humans, no such filtering was possible.
The results presented as pie diagrams in Figure 1-C, demonstrate that in trypanosomes, as in all other organisms, the largest category of up-regulated genes in response to ER stress is the category of genes that function in protein secretion. Moreover, except movement, all the other functions that were shown to be up-regulated in trypanosomes are also up-regulated in some if not all other organisms. Changes in motility are specific to trypanosomes. Of note is also that a significantly high fraction of the down-regulated genes (35%, P<10−6) encode for proteins destined to traverse the ER (i.e., harboring either a signal-peptide or trans-membrane domains), similarly to D. melanogaster [34].
Trypanosomes lack conventional transcription regulation, and therefore it is expected that their genome lacks homologues of IRE1/XBP1. Since mRNA stability is a very dominant mechanism of regulation, we examined the possibility that up-regulation results from mRNA stabilization during ER stress. To this end, procyclic stage trypanosomes (untreated) and treated with DTT for 90 minutes, were treated with sinefungin and actinomycin D, inhibitors of trans-splicing and transcription, respectively [36]. Incubation was for increasing time points to measure possible changes in the half life of mRNAs during ER stress. RNA was extracted and subjected to Northern analysis with probes specific to genes whose level was significantly changed upon DTT treatment and that were shown to also be induced in other organisms. The results in Figure 2 demonstrate that the half-life of DNAJ, protein disulfide isomerase (PDI), thioredoxin and syntaxin were all increased (from 45 to 90 min, 60 to 90 min, 13 to 65 min and 10 to 20, respectively) under DTT treatment, suggesting that the up-regulation of these transcripts during ER stress can be explained by stress-induced mRNA stabilization. The level of mRNAs that were unchanged during DTT treatment (based on microarray analysis) was also examined, and no changes in their stability was observed upon the ER stress (results not shown), indicating the specificity of this regulation.
One of the hallmarks of UPR in yeast and metazoa is the ability to induce the production of ER chaperones to increase the folding capacity of the ER [1],[3],[4],[37]. Induction of chaperones is not only mediated by increase in the transcription but also by preferential translation and proteolysis under stress [38]. Although BiP mRNA is not among the mRNAs whose level is most highly increased under ER stress ([3],[39] and this study) its elevation at the protein level serves as a hallmark for UPR. To examine if in trypanosomes BiP is increased upon induction of ER stress, its level was examined after treating cells with DTT or 2-deoxy D- glucose (2-DG), which affects glycosylation [40],[41]. Although 2-DG inhibits glycosylation and is known to induce the UPR response in other eukaryotes, it may also have other effects on carbon source sensing and metabolism [41]. The results presented in Figure 3-Ai and Figure 3-Aii suggest that both treatments result in increased levels of BiP in a time-dependent manner. Next, we examined if ER stress response can be induced also in the bloodstream form. The results (Fig. 3-Aiii) suggest that under the conditions used, BiP level was increased. These results are contrast to those previously published [24]. However, note that we used 4mM DTT, whereas the previous study used only 1 mM DTT. The results suggest that ER stress response operates in both stages of the parasite.
To observe cellular changes that may accompany response to ER stress, ultrastructure analysis using electron microscopy was performed. After 1 hour of DTT treatment, the most prominent change observed by transmission electron microscopy (TEM) is the expansion of the ER (compare Fig. 3 Bi to Bii). ER expansion was previously observed in yeast and mammals, and it was suggested that the increase in the ER may help to accommodate newly synthesized ER proteins and inhibit aggregation of unfolded proteins by reducing their concentration [42],[43]. The ER lumen was not only dilated, but seemed to contain dense material that might correspond to aggregated proteins (Fig. 3-Bii).
To examine if trypanosomes can recover following removal of ER stress, procyclic stage parasites were exposed to DTT for different time periods ranging from 30 to 240 minutes. After incubation with DTT, the DTT was washed out, and growth was monitored. The results (Figure 4-A) suggest that after exposure to DTT for 120 minutes or more, cells were unable to resume growth upon withdrawal of DTT, suggesting that prolonged exposure to DTT leads to cell death.
SLS was shown to be induced in procyclic stage trypanosomes upon silencing of SRP receptor or low pH [23]. We previously proposed that SLS might replace the conventional ER stress response because of the lack of conventional transcriptional regulation that would be required to induce the production of mRNAs needed for this process. The results in Figures 1–3 suggest that trypanosomes possess an ER stress response, and it was therefore of great interest to examine if this response is associated with SLS. SLS is characterized by two hallmarks, reduction in SL RNA level and changes in the amount and localization of tSNAP42. The results in Figure 4-Bi and Figure 4-Bii suggest that SLS is induced 180 and 120 minutes after exposure to DTT or 2-DG, respectively. SLS is induced after BiP induction, since induction of BiP was detected 120 minutes after DTT addition (Fig. 3-Ai) and 60 minutes after 2-DG addition (Fig. 3-Aii), suggesting that the ER stress response is induced before SLS. To examine if prolonged ER stress induces SLS also in bloodstream trypanosomes, SL RNA level was examined upon treatment with DTT. The results in (Fig. 4-Biii) demonstrate the induction of SLS in bloodstream trypanosomes after 180 minutes exposure to DTT.
The induction of SLS is clearly observed when comparing the localization of the SL RNA transcription factor, tSNAP42. Under normal conditions, tSNAP42 is confined to the nuclear domain where SL RNA is transcribed and assembled with the Sm core proteins [44],[45]. Under prolonged ER stress, the level of tSNAP42 increases, filling up the entire nucleus (Fig. 4-C). These data suggest that SLS is activated after the response to ER stress was initiated; 180 minutes (in DTT-treated cells) or 120 minutes (in 2-DG-treated cells).
We previously demonstrated that SLS is induced following silencing of the SRP receptor in the procyclic stage [23]; it was therefore of interest to examine whether SLS is induced following other perturbations that are connected to ER translocation functions. For this analysis, we chose to investigate and compare three functions that control translocation across the ER membrane, SRα, SEC63 and SEC61. We have previously demonstrated that SEC63 is essential for translocation of proteins to the ER via both pathways - the co-translational pathway mediated by SRP and the post-translational pathway mediated by chaperones [46]. The SEC61 protein constitutes the translocon. SEC61 was silenced using a stem-loop construct as described in the Materials and Methods. The results in Figure 5-Ai demonstrate that upon silencing of SEC61 or SEC63, the level of SL RNA was reduced, with no change in the level of 7SL RNA. The reduction elicited a major decrease in the level of mRNAs due to inhibition of trans-splicing (Fig. 5-Aii). The reduction in SL RNA was accompanied by the classical nuclear accumulation of tSNAP42 (Fig. 5-B). Indeed the level of tSNAP42 was increased upon SEC63 silencing as observed in Figure 5-C. The results suggest the perturbation in ER translocation due to elimination of SEC61 or SEC63 elicits SLS, much like that observed in SRα silenced cells [23]. To examine if SEC63 silencing also induces SLS in bloodstream form, transgenic parasites expressing a T7 opposing construct to silence the SEC63 gene (Figure S2) were generated as described in Materials and Methods. The results (Figure 5-D) demonstrate that silencing of SEC63 induces SLS as evident by marked reduction in the level of SL RNA.
Is SLS a programmed cell death pathway that is induced under prolonged ER stress much like apoptosis in mammalian cells? To examine this intriguing possibility, the hallmarks of programmed cell death were investigated in SLS-induced procyclic stage trypanosomes. First, the DNA fragmentation during SLS was measured by TUNEL. TUNEL identifies nicks in DNA by measuring the incorporation of fluorescently labeled dUTP using terminal deoxynucleotidyl transferase (TdT). The results presented in Figure 6-A demonstrate an increase in free ends of DNA, marking DNA fragmentation in SRα silenced cells. Two controls were used to demonstrate the specificity of the assay; in the absence of TdT no signal was observed (Fig. 6Ai), and maximal signal was obtained upon treatment with micrococcal nuclease, which induces DNA breaks (Fig. 6Aii). Another assay to monitor DNA fragmentation is by propidium iodide (PI) staining and flow cytometry. Apoptotic cells display a broad hypodiploid (sub-G1) peak, which can be easily discriminated from the narrow peak of cells with normal (diploid) DNA content. SRα silenced cells were examined for the content of sub-G1 DNA upon silencing. The results (Fig. 6-B) demonstrate an increase in the number of sub-G1 cells, suggesting that many cells in the population undergo DNA fragmentation and lose DNA fragments as response to SRα silencing. The same experiment was performed for all silencing treatments that were shown to induce SLS, and the results in Figure 6-C indicate increase in sub-G1 population in cells silenced for SRα, SEC61 and SEC63. The TUNEL and sub-G1 assays can also detect DNA fragmentation in necrotic cells. The inter-nucleosomal DNA digestion by endogenous nucleases to yield a characteristic laddering pattern is one of the hallmarks of apoptotic cells that does not take place in necrotic cells [47]. Laddering of DNA is one of the final steps in the apoptotic process, together with chromatin condensation, nuclear picnosis and the formation of apoptotic bodies. The results in Figure 6-D demonstrate DNA laddering in SRα silenced cells, and that laddering is increased during SLS induction. The maximal laddering is observed 4 days after the induction of silencing. Identical DNA laddering was observed under SEC63 silencing (data not shown). In sum, the results demonstrate that SLS induction leads to DNA fragmentation and laddering, suggesting that SLS-induced procyclic stage trypanosomes undergo PCD.
Induction of apoptosis causes externalization of phosphatidyl serine (PS) on the surface of the apoptotic cells. AnnexinV binds to the exposed PS of apoptotic cells. The dual staining with propidium iodide (PI) enables the distinction between necrotic and apoptotic cells [48]. As indicated in Figure 7-A, necrotic cells lose their membrane integrity and become permeable to PI (Figure 7-A, upper left panel). Live cells are stained neither with AnnexinV nor with PI (bottom left panel). During early apoptosis, PS is exposed on the surface but because of the integrity of the plasma membrane the cells are not be stained with PI (bottom right panel). Late apoptotic or necrotic cells lose their membrane integrity and are stained with both PI and AnnexinV (upper right panel). SLS induced cells - elicited by silencing of SRα, SEC63 and SEC61 or after treatment with DTT - were stained with AnnexinV before the cells became permeable to PI, indicating that PS was externalized. These data suggest that apoptosis is induced 2 days after induction of silencing of each of the factors (Fig. 7-Bi-iii, middle column), or 12 hrs after DTT treatment (Fig. 7-Biv). After 3 days of silencing, the apoptotic cells also became permeable to PI (Fig. 7-Bi-iii, right column). Cells treated with DTT for 20 or 24 hrs also became permeable to PI (Fig. 7-Biv). To visualize cells externally stained by AnnexinV, cells were incubated with AnnexinV and PI, and visualized by fluorescence microscopy. The results in Figure 7-C demonstrate that although the cells were not stained by PI, AnnexinV staining was observed only in the silenced cells and not in control, uninduced cells. The results suggest that at early time points of silencing, PS became externally exposed on the cells, most probably as a result of membrane flipping, a signature of apoptosis. Note that, PS exposure was not observed in cells silenced either for SmD1 [49] or PTB2 [50] that are dying as a result of silencing (Figure S1-A).
Recent studies have shown that an increase in cytoplasmic [Ca2+] occurs both at early and late stages of the apoptotic pathway. Ca2+ release from the ER induced by milder insults promotes cell death through apoptosis [51]. Trypanosomes, like other eukaryotes, maintain a low intracellular concentration of free Ca2+, and changes in the concentration of this ion were shown to affect the virulence of these organisms [52]. Several organelles were demonstrated to transport Ca2+ in an energy dependent manner, including the plasma membrane [53], endoplasmic reticulum [54], mitochondrion [55] and the acidocalcisome [56],[57]. In trypanosomes, the mitochondrion maintains a low resting level of Ca2+, but transiently accumulates large quantities of Ca2+ from the cytosol following Ca2+ influx across the plasma membrane or after release from the acidocalcisome [58]. Indeed, death in T. brucei was shown to be associated with changes in the ability of the mitochondrion to modulate Ca2+ levels [58]. To examine changes in cytoplasmic [Ca2+] during SLS, the level of cytoplasmic Ca2+ was examined using fluo-4-AM [59]. The results, shown in Figure 8-A, demonstrate an increase of cytoplasmic [Ca2+] mostly 42 to 48 hours following induction of SEC63 silencing. An increase in cytoplasmic [Ca2+] was also observed following treatment with DTT (Fig. 8-B). The results suggest that SLS is associated with perturbations of Ca2+ homoeostasis. Note that changes in the cytoplasmic [Ca2+] originate from internal pools, since these changes were unaffected by the presence of external EGTA (Fig. 8-C).
A reduction in mitochondrial membrane potential (ΔΨm) has been observed during apoptosis. Tetramethyl rhodamine methyl-ester (TMRM) is a cationic lipophilic dye that enters cells and reversibly accumulates in the negatively charged mitochondrial matrix, depending on mitochondrial membrane potential [60]. Cells (uninduced) or after 3 days of silencing of SRα, SEC61 or SEC63 were incubated with TMRM, and the fluorescence of the dye was measured by FACS. The results (Fig. 9-Ai) indicate that up to 80% of the cells showed a decrease in membrane potential as a result of the silencing. To examine if membrane depolarization also takes place during DTT treatment, cells were treated with DTT for different time periods, and ΔΨm was measured. The results in Figure 9-Aii suggest that changes in ΔΨm were already observed after 1 hr, and after 3 hours, depolarization was observed in all cells.
During apoptosis, the inner mitochondrial membrane loses its integrity, and oxidative phosphorylation is uncoupled. When this occurs, oxidation of metabolites by O2 proceeds with electron flux not coupled to proton pumping, resulting in dissipation of the transmembrane proton gradient and ATP production, leading to the production of ROS. To measure ROS production that may result in part from the dissipation of the mitochondrial proton gradient, we used the sensitive probe 2′–7′ dichlorodihydrofluorescein (DCFH-DA). This non-fluorescent dye diffuses across cell membrane and is hydrolyzed to DCFH intracellularly. In the presence of ROS, DCFH is rapidly oxidized to highly fluorescent, dichlorofluorescein [61]. Indeed, ROS production was observed in SRα, SEC61 and SEC63 silenced cells (Fig. 9-Bi). ROS production was also monitored in DTT treated cells (Fig. 9-Bii). Interestingly, ROS production only appeared 5 hrs after DTT addition, although ΔΨm was already decreased after 3 hrs of DTT treatment (Fig. 9-Aii). This delay in observing ROS production after mitochondria permeabilization may reflect the different sensitivity of the assays used to measure ΔΨm and ROS. Note that DTT treatment could also affect ROS measurements because of its reducing capabilities.
One of the characteristic of changes involving the mitochondria during apoptosis is the leakage of proteins such as cytochrome C, which in metazoa, induces the activation of caspases [62]. Recently, endonuclease G (endoG) was described in trypanosomatid species [63]. This enzyme resides in the mitochondria and is released when cell death is triggered. The release of endoG from the mitochondria may represent a caspase-independent cell death mechanism. Since DNA fragmentation was observed in SLS-induced procyclic stage trypanosomes it was of great interest to examine if endoG is released in these cells. Indeed, endoG “leaked” to the cytosol in the SEC63 silenced cells (data not shown).
Transmission electron microscopy (Fig. 10-A) of procyclic stage trypanosomes after 2 days of SEC63 silencing (v-x) or after 5 and 12 hrs of DTT treatment (xi-xiv, xv-xvii, respectively) show characteristic features of apoptotic cells. Untreated cells show nuclei with prominent central nucleolus (a) and equally distributed chromatin, mitochondria with no dilations (ii, iii), and a narrow ER (iv). In contrast, the SLS induced cells show condensed chromatin (v, vi, xv, xvi), expanded ER (xii, xiii, xiv, xv), and expanded, dilated mitochondria (vii, ix, x, xii, xvi). To verify that the expended organelle is indeed the ER, the procyclic cells slices were reacted with anti-BiP and visualized using gold-labeled antibodies by TEM. The results (xiii, xiv) demonstrate that the expanded compartment contained BiP and is therefore the ER lumen.
Note that chromatin condensation typical to apoptotic cells was not observed in SmD1 silenced cells, although these cells show major morphological perturbations such as accumulation of large vesicles in the cytoplasm (Figure S1-B). Autophagosomes appearing as membrane-surrounded bodies were observed in the cytosol (vii, xi, xvi), as well as inside the mitochondria (viii, xvii). Autophagy is induced as a survival response under starvation conditions [64], but autophagy was shown to regulate the endoplasmic reticulum during UPR [42],[65]. Autophagy involves the massive degradation of organelles such as mitochondria (mitophagy). Autophagy was also shown to induce PCD [66]. To examine if autophagy is induced during SLS, the formation of autophagosome was monitored using tagged ATG8. ATG8 is conserved from yeast to mammals and is commonly used as an autophagosome marker. ATG8 was tagged by in situ insertion of yellow fluorescent protein (YFP) to the authentic site both in parental cells and in cells expressing the SEC63 silencing construct. The results in Figure 10-B demonstrate that upon silencing of SEC63 (Fig. 10-Bi), or prolonged treatment with DTT (Fig. 10-Bii), clear autophagosome formation was observed. The most obvious change was the increase in autophagosomes number and size.
In this study, two major questions were addressed. We wished to determine whether trypanosomes are able to change their transcriptome as a result of ER stress like other eukaryotes, and whether this process is mediated or governed by the stress-induced mechanism, SLS [23]. Our results suggest that ER stress induces a very distinct response; the transcriptome is changed in a manner similar to changes observed in other eukaryotes during UPR. The changes in the transcriptome do not result from transcriptional programming as in other eukaryotes, but by stabilizing mRNAs that are needed to execute the ER stress response, resembling the heat-shock response in trypanosomes [19]. However, when ER stress is persistent, SLS is induced, suggesting that SLS does not replace the UPR but triggers programmed cell death under severe ER stress in both stages of the parasite. We therefore suggest that the biological role of SLS is to serve as the apoptotic branch of UPR and induce a very rapid and aggressive PCD response to eliminate unfit parasites from the population.
In this study, the changes of the transcriptome upon DTT treatment were compared to similar changes in yeast and metazoa. The data show that procyclic stage trypanosomes are able to change their transcriptome in response to ER stress. The most obvious change is the up-regulation of genes that are involved in protein secretory functions and the down-regulation of genes harboring signal-peptide and transmembrane domains. In yeast, the majority of the up-regulated genes play a role in secretion, phospholipid metabolism, protein glycosylation, vesicular transport, cell wall biosynthesis, vacuolar targeting and ER-associated degradation [3]. All together the UPR increased the level of proteins that lead to enlargement of the ER lumen, enhance export of the misfolded proteins for degradation [endoplasmic reticulum associated degradation (ERAD)] or enhance their transport from the ER via the vesicular transport. Very similar data was observed during UPR in C. elegans. [33]. In human cells, genes involved in targeting of proteins to the ER are induced, but most of these transcripts are only slightly induced, suggesting that modest changes in the level of their transcripts are sufficient for an adaptive response to ER stress [35]. In organisms such as S. cerevisiae, D. melanogaster and C. elegans, the specificity of the transcriptome changes during UPR was refined by examining the effect in IRE/XBP-1 mutants and their homologues [3],[33],[34]. In trypanosomes the specificity of the response to ER stress was refined only by kinetic analysis of the induced genes and therefore not all the affected genes might be directly linked to the ER stress response. However, the results indicate that in trypanosomes, the largest group of up-regulated genes, function in protein secretion and ER related functions similar to the proportion of such genes in yeast, fly, worm and human under UPR. Note, that as opposed to yeast, in trypanosomes, a large number of genes involved in signaling (17% vs. 8%) were up-regulated during the ER stress response, suggesting that signaling pathways are activated to execute this response. Of interest are also genes involved in movement, which resembles the signaling of chemotaxis in bacteria, enabling the microorganism to escape from chemical stressors by moving to a different microenvironment. Another important response of the UPR described in D. melanogaster suggest that IRE1 mediates the rapid degradation of a specific subset of mRNAs, encoding for genes harboring signal-peptide or transmembrane domains. This response helps to reduce the burden on the translocation and folding machinery [34]. Indeed, among the trypanosome genes that are down-regulated under ER stress, significant numbers contain signal-peptide or transmembrane domains.
A recent study investigated the changes of T. brucei transcriptome during development as well as in response to tunicamycin or DTT treatments in bloodstream form trypanosomes [24]. This study reached the conclusion that while trypanosomes regulate mRNA abundance during the developmental cycle, they do not show significant changes in response to ER stress. The study failed to detect changes in BiP in response to ER stress and concluded that UPR does not exist in trypanosomes, primarily because these parasites live in homeostatic conditions, especially in the mammalian host. Our study reached an opposite conclusion. First, we observed major changes in the transcriptome, and, in addition, we observed changes in BiP expression during ER stress in both procyclic and bloodstream form trypanosomes. Moreover, we observed ER dilation at the same time at which we observed BiP up-regulation (1 and 5 hours) [(Fig. 3Bii, Fig. 10A-(xii, xiii, xiv)]. Note that the discrepancy between these studies can be easily reconciled given the fact that different DTT concentrations were used (1mM DTT vs. 4 mM DTT in our study). As noted in the previous study and as presented here, prolonged exposure to ER stress induces death. Here, we show that this death is associated with SLS during the entire life cycle of the parasite.
Another recent study on the enzyme UDP-glucose:glycoprotein glucosyltransferase (UGGT) demonstrated that null mutations of this gene in T. brucei bloodstream form do not induce BiP production. These results suggest that proteins involved in ER glycosylation are not tightly regulated in the trypanosome bloodstream form which might reflect the need for an extremely high flux of glycoprotein synthesis and export of VSG [67]. Interestingly, deletion of UGGT in T. cruzi has been reported to increase BiP expression, suggesting that in T. cruzi there is an ER stress response that senses the misfolded proteins and compensates for the lack of UGGT by up-regulating BiP [68]. All these data suggest that the response to ER stress may vary during the life cycle and among trypanosomatid species. However, the SLS induction in both procyclic and bloodstream from parasites as presented in this study, suggest that induction of ER stress by DTT or by major perturbation to the ER integrity and other cellular membranes, induce a systematic response. The response might be induced only under very severe stress that perturbs the membranes, since RNAi silencing of the post-translational pathway by depletion of SEC71 that does not harm membrane protein biogenesis did not induce SLS [46].
Conventional transcription regulation is absent in trypanosomes and therefore it was accepted that these organisms lack IRE1 and XBP1. In this study, we demonstrate a unique novel mechanism to induce UPR based on differential stabilization of mRNAs. The ER stress response resembles the heat shock regulation in trypanosomes, which is regulated mostly by mRNA stability and preferential translation [69]. mRNA stability is considered the major post-transcriptional mechanism that determines the trypanosome transcriptome [19]. The mechanism of the preferential stabilization under ER stress observed in this study is currently unknown. Stability of particular mRNAs can be regulated by two different mechanisms. RNA binding proteins can either stabilize the mRNAs under ER stress or a protein that normally destabilizes the mRNA can be inactivated during stress. Future challenges will be to identify the target site on the mRNAs that are the target of this mechanism, and further, to identify the cognate RNA binding protein(s) that mediates this regulation.
PERK is one of the key factors regulating UPR in metazoa. As opposed to yeast, which lack PERK, trypanosomes possess three potential PERK-homologues (TbITF2K1-K3) [70]. However, only one of these kinases, TbeIF2K2, carries a transmembrane domain and is able to phosphorylate eIF2α on Thr169, homologues to Ser51 of other eukaryotes [70]. This kinase level is reduced in high density culture but there is no evidence that this kinase participates in the ER stress response. Unexpectedly, the protein was localized to the flagellar pocket of the parasite [70]. No change in protein synthesis arrest as a result of DTT treatment was observed in cells silenced for this factor using RNAi (our unpublished results). Thus, TbeIF2K2 is most probably not involved in ER stress. Note that heat-shock in trypanosomes causes polysome collapse and translational shut-off independently of eIF2α phosphorylation, which takes place during the heat-shock response in other eukaryotes [69]. It remains possible that the other TbeIF2K kinases [70] function in ER stress. However, these kinases lack a transmembrane domain, and to function in the ER stress response these kinases must somehow associate with the ER membrane. Studies are in progress to decipher the possible role, if any, of these two kinases (TbeIF2K 1, 3) in ER stress and/or SLS.
UPR was shown to induce autophagy in yeast and mammals [65]. It was therefore interesting to observe autophagy in SLS-induced cells. The autophagy pathway functions as backup mechanism for ERAD, when substrates destined for degradation overwhelm the ERAD capacity [65]. This could also be the function of autophagy under ER stressors or silencing of SEC63 observed in this study.
The presence of PCD in unicellular organisms and in trypanosomatids was initially viewed with skepticism [26]. The major argument against the presence of PCD in these organisms is the absence in trypanosomes of caspases that function in PCD. However, PCD was shown to be induced by prostaglandin D2 (PGD2), a molecule secreted primarily by the stumpy form of the parasite [71]. Trypanosomes control their density within the mammalian host by secreting a quorum sensing molecule (SIF) [72]. Slender parasites secrete SIF, which induces differentiation to the stumpy form. Stumpy parasites are pre-adapted for transmission to the fly but respond to PGD2 by PCD in the blood. Thus, the altruistic removal of stumpy forms assures persistent infection [71],[73]. The PCD associated with the prostaglandins involves an increase in ROS. It was suggested that trypanosomes might exhibit a primitive form of apoptosis that does not depend on proteases but instead depends on ROS formation [26]. This ROS signal can lead to stabilization of mRNAs encoding for proteases and nucleases that are needed to execute PCD. The death pathway executed by ER stress via SLS might be another altruistic pathway present in these parasites that was developed to remove the unfit parasites from the population under stress.
One can envision a mechanism by which ER stress induces imbalance of Ca2+ homeostasis. The increase in cytoplasmic [Ca2+] is most probably due to leakage from the malfunctioning ER, resulting from the reduced capacity of the ER to store Ca2+. This decreased capacity might stem from reduced levels of calreticulin and the ER-resident SERCA calcium pump, as well as the acidocalcisome Ca2+ transporters [74],[75]. These proteins either carry a signal-peptide or are polytopic membrane proteins. Indeed, major defects in biogenesis of both polytopic membrane proteins and signal peptide-containing proteins were observed in SEC63 and SEC61 silenced cells [46]. The ER stress induced by DTT and 2-DG may also affect the activity of the ER chaperone and the Ca2+ pumps.
In eukaryotes, Ca2+ from the ER or cytosol moves to the mitochondrial outer membrane through voltage dependent ion channel (VDAC) [76]. This leads to induced opening of the mitochondrial permeability transition pore (PTP) resulting in matrix swelling [51],[77]. Such changes cause the rupture of the outer membrane of the mitochondria, and release of apoptotic factors such as cytochrome C [78]. The rise in the mitochondrial Ca2+ stimulates the generation of ROS, and the opening of PTP causes dissipation of the ΔΨm, as was also observed in SLS-induced cells.
The PCD observed by silencing of ER functions (SRα, SEC61, SEC63) or by ER stressors can be explained by perturbations induced by Ca2+, leading to ROS production [79]. If so, what is the role of SLS and why is it required? We propose that SLS function is to speed up the death process. SLS is induced when response to ER stress fails to restore homeostasis, and it resembles apoptosis that takes place in mammalian cells under persistent ER stress [13],[80].
Very little is known about how the switch between protective and destructive pathways is regulated in eukaryotes. In metazoa a defined regulatory circuit that controls the switch from protective UPR to the pro-apoptotic pathway exists [81]. It was suggested that transient PERK signaling protects cells by temporarily dampening cellular protein synthesis and thus reducing the misfolded protein level in the ER. However, persistent PERK signaling could ultimately impair cell viability if extended translational inhibition interrupted the generation of proteins vital for cellular homeostasis [81]. An analogous situation may exist in trypanosomes. The severe inhibition of protein synthesis emerging from the drastic decrease in all mRNAs due to elimination of trans-splicing by SLS may lead to drastic reduction in the synthesis of proteins that are also required for the response to ER stress, and may resemble the physiological effects elicited by prolonged PERK induction in metazoa. SLS may accelerate cell death, rapidly eliminating unfit organisms from the population. One of the most intriguing and yet unresolved question is the nature of the signaling pathway that senses ER stress and transmits it to the nucleus to initiate the SLS response.
Procyclic forms of T. brucei strain 29–13 which carries integrated genes for T7 polymerase and the tetracycline repressor [82], were grown in SDM-79 [83] supplemented with 10% fetal calf serum in the presence of 50 µg/ml hygromycin B and 15 µg/ml G418.
The bloodstream form (BSFs) of T. brucei strain 427, cell line 1313–514 (a gift from C. Clayton, ZMBH, Heidelberg, Germany) [84] were aerobically cultivated at 37°C under 5% CO2 in HMI-9 medium [85] supplemented with 10% fetal calf serum containing 2 µg ml−1 G418 and 2.5 µg ml−1 phleomycin.
The oligonucleotides used in this study are listed in Table S1. Generation of the stem-loop silencing construct for SEC61α (Tb11.02.4100) using the oligonucleotides 104–106 was performed as previously described [86]. The construction of SEC63 and SRα silencing constructs was previously described [23],[46]. The construct to silence SEC63 in BSF was prepared by cloning a polymerase chain reaction (PCR) product generated using oligonucleotides 1111–1112 to the p2T7TA-177 vector. The transfection was performed as previously described [87]. Transgenic cell lines were selected on hygromycin B. To generate the YFP–ATG8 (Tb927.7.5910) fused protein, a PCR fragment was amplified using the oligonucleotides 829–830, covering positions 1 (first ATG) to the end of the protein. The vector, p2829 YFP-Dhh1, was digested with HindIII and NotI, and the fragment coding for ATG8 was cloned using the same sites [88]. The cloning creates an N-terminal fusion, and the HindIII site was used for fusing the gene and YFP. For integration, the plasmid was linearized with XcmI at position 276 from the ATG. Cloned transgenic cell lines were selected on blasticidin.
Total RNA was isolated from untreated cells and cells treated with DTT for 1 or 3 hours. RNA quality was determined using the Eukaryote Total RNA Nano 6000 assay kit (Agilent Technologies) on the Agilent Technologies 2100 Bioanalyzer. To generate the fluorescently labeled cRNA, total RNA was labeled using the Ambion Amino Allyl MessageAmp II aRNA kit (Ambion). A classical balanced block design with dye swap was used for comparing between RNA extracted following 1 hr DTT treatment, or 3 hr DTT treatment, and RNA extracted from uninduced cells. Arrays were blocked with 1% BSA, 0.1% SDS, 53 SSC at 42°C for 1 h, rinsed with double-distilled water, and air-dried. Next, 1.5 µg of cyanine 3-labeled and 1.5 µg of cyanine 5-labeled aRNA were fragmented and hybridized to the T. brucei DNA microarrays obtained through NIAID's Pathogen Functional Genomics Resource Center (managed and funded by the Division of Microbiology and Infectious Diseases, NIAID, NIH, DHHS, and operated by the J. Craig Venter Institute) using the Gene Expression hybridization kit (Agilent Technologies) for 16 h at 60°C in a hybridization oven. After hybridization, each array was subjected to three washes: first wash with 2× SSC, 0.1% SDS for 1 min at 55°C, and then 0.1× SSC, 0.1% SDS for 1 min, and finally with 0.1× SSC for 1 min, at room temperature. Arrays were scanned using the dual laser scanner (Agilent G2505B Microarray scanner). The data were then extracted from images using Spotreader software (Niles Scientific). Flawed features identified by visual inspection of the array images were flagged, and the data from these features were removed prior to further analysis.
Total RNA was prepared with Trizol reagent and 20 µg/lane was separated on a 1.2% agarose gel, containing 2.2 M formaldehyde. Small RNAs were fractionated on a 6% (w/v) polyacrylamide gel containing 7M urea. The RNA blots were hybridized to oligonucleotides or [α-32P]-random labeled probes (Random Primer DNA Labeling Mix, Biological Industries, Co.). The oligonucleotides used to prepare the probes are detailed in Table S1.
Untreated cells and cells after 1.5 hour of DTT treatment (107 cells/ml) were divided into 50 ml aliquots for each time point, concentrated and resuspended in 5 ml of the remaining growth medium. The concentrated cells were incubated for 30 min at 27°C, treated with 2 µg/ml sinefungin (Sigma) for 10 min, and then 30 µg/ml of actinomycin D were added (Sigma). Aliquots were taken at different time points (0–90 min). For each time point, total RNA was prepared and the samples were subjected to Northern analysis. Quantization by densitometry was performed using ImageJ software, and the results were used for calculating the half-life of the mRNAs.
Whole cell lysates (106 cell equivalents per lane) were fractionated by SDS-PAGE, transferred to PROTRAN membranes (Whatman), and probed with anti BiP (diluted 1∶5000), kindly provided by Prof. James Bangs (University of Wisconsin-Madison, Madison, USA); anti tSNAP42 antisera (diluted 1∶7500) kindly provided by Prof. Vivian Bellofatto (UMDNJ- NJ, USA); and anti p58 antibodies raised by our group against a cytoplasmic protein Tb11.46.0009 (1∶4000). The bound antibodies were detected with goat anti-rabbit immunoglobulin G (IgG) or anti-mouse IgG coupled to horseradish peroxidase, and were visualized by ECL (Amersham Biosciences).
Cells were washed with PBS, mounted on poly-L-lysine-coated slides, fixed in 4% formaldehyde, and immunofluorescence was performed as described in [23] using anti tSNAP42 antisera. The cells were visualized with a Zeiss LSM 510 META inverted microscope.
T. brucei cells, silenced for various genes, were fixed in 70% Ethanol-30% PBS and stored at 4°C overnight. Cells were then washed once with PBS and incubated on ice for 30 minutes to enable rehydration. The samples were resuspended in PBS containing 50 µg/ml RNase A (Roche Diagnostics) for 30 minutes at 4°C and stained with 50 µg/ml Propidium Iodide (Sigma). Samples were analyzed by FACS using the FACStar plus (Becton Dickinson, San Jose, CA) and the CellQuest list mode analysis software.
T. brucei cells following the different treatments were harvested and loaded with 150 nM TMRM (Invitrogen) in serum-free medium. The samples were incubated in the dark for 15 min at 27°C and then analyzed by FACS.
Reactive oxygen species (ROS) were measured by flow cytometry using DCFH-DA (Sigma). Following the different treatments cells were harvested and incubated at 27°C for 30 min with 10 µM DCFH-DA. The cells were then washed once with PBS and analyzed by FACS using the CellQuest list mode analysis software.
For calcium measurements, 106 cells were loaded with 1 µM Fluo-4-AM (Invitrogen) for 1 hour at 27°C. The cells were washed three times with PBS and resuspended in PBS. Staining was evaluated by FACS. Results were analyzed using FlowJo software.
Cells were pelleted by centrifugation, washed with PBS, resuspended in extraction buffer (10 mMTris pH 8.0, 0.1 mM EDTA pH 8.0, 20 µg/ml RNase, 0.5% SDS), incubated at 37°C for 1 hour, and for an additional 3 hours at 50°C in the presence of 4 µg/ml proteinase K. After phenol:chloroform (1∶1 V/V) and chloroform extractions, and ethanol precipitation, the DNA was treated with 10 µg/ml RNaseA for 1 hour at 37°C and 10 µg was electroporated on 1.8% Agarose gel containing 0.5 µg/ml ethidium bromide and visualized under UV light.
Detection of cleavage of genomic DNA was performed by using the In-Situ Cell Death Detection Kit, FITC (Roche Molecular Biochemicals) according to manufacturer's instructions.
Trypanosomes were reacted with fluorescein isothiocyanate-labeled Annexin V antibodies (MBL©) and stained with propidium iodide according to the manufacturer's instructions. The cells were analyzed by FACS or visualized under a Zeiss LSM 510 META inverted microscope.
Trypanosomes were fixed with Karnovsky (4% paraformaldehyde, 2.5% glutaraldehyde) solution for 1 h at room temperature, and then at 4°C over night. Thereafter, the cells were incubated with 1% OsO4 for 1 h at 4°C. Samples were dehydrated in ethanol and embedded in Epon-812 according to standard procedures. Ultrathin sections were prepared using a LKB Ultratome III, stained with uranyl acetate and contrasted with lead citrate. The samples were visualized using a transmission electron microscope FEI Tecnai 120 kV.
Trypanosomes were fixed with Karnovsky (2% paraformaldehyde, 0.5% glutaraldehyde) solution for 1 h at room temperature, and then at 4°C over night. Thereafter, the cells were incubated with 1% OsO4 for 1 h at 4°C. Samples were dehydrated in ethanol and embedded in Epon-812 according to standard procedures. Ultrathin sections were prepared using a LKB Ultratome III. The sections were floated in blocking buffer (10mM Tris pH 8.5, 0.05% Tween-20, 5% non-fat dry milk) for 30 minutes and then incubated with rabbit anti BiP antibodies, kindly provided by Prof. James Bangs (University of Wisconsin-Madison, Madison, USA) diluted (1∶20) in the blocking buffer (4°C, O.N) washed with wash buffer (10 mM Tris pH 8.5, 0.05% Tween 20) and incubated with Donkey anti rabbit IgG (H+L) conjugated to 18 nm gold particles (Jackson immune-research cat.# 711-215-152, diluted 1∶100) and washed again with wash buffer. The grids were stained with uranyl acetate. The samples were visualized using a transmission electron microscope FEI Tecnai 120 kV.
|
10.1371/journal.pcbi.1005820 | Competitive tuning: Competition's role in setting the frequency-dependence of Ca2+-dependent proteins | A number of neurological disorders arise from perturbations in biochemical signaling and protein complex formation within neurons. Normally, proteins form networks that when activated produce persistent changes in a synapse’s molecular composition. In hippocampal neurons, calcium ion (Ca2+) flux through N-methyl-D-aspartate (NMDA) receptors activates Ca2+/calmodulin signal transduction networks that either increase or decrease the strength of the neuronal synapse, phenomena known as long-term potentiation (LTP) or long-term depression (LTD), respectively. The calcium-sensor calmodulin (CaM) acts as a common activator of the networks responsible for both LTP and LTD. This is possible, in part, because CaM binding proteins are “tuned” to different Ca2+ flux signals by their unique binding and activation dynamics. Computational modeling is used to describe the binding and activation dynamics of Ca2+/CaM signal transduction and can be used to guide focused experimental studies. Although CaM binds over 100 proteins, practical limitations cause many models to include only one or two CaM-activated proteins. In this work, we view Ca2+/CaM as a limiting resource in the signal transduction pathway owing to its low abundance relative to its binding partners. With this view, we investigate the effect of competitive binding on the dynamics of CaM binding partner activation. Using an explicit model of Ca2+, CaM, and seven highly-expressed hippocampal CaM binding proteins, we find that competition for CaM binding serves as a tuning mechanism: the presence of competitors shifts and sharpens the Ca2+ frequency-dependence of CaM binding proteins. Notably, we find that simulated competition may be sufficient to recreate the in vivo frequency dependence of the CaM-dependent phosphatase calcineurin. Additionally, competition alone (without feedback mechanisms or spatial parameters) could replicate counter-intuitive experimental observations of decreased activation of Ca2+/CaM-dependent protein kinase II in knockout models of neurogranin. We conclude that competitive tuning could be an important dynamic process underlying synaptic plasticity.
| Learning and memory formation are likely associated with dynamic fluctuations in the connective strength of neuronal synapses. These fluctuations, called synaptic plasticity, are regulated by calcium ion (Ca2+) influx through ion channels localized to the post-synaptic membrane. Within the post-synapse, the dominant Ca2+ sensor protein, calmodulin (CaM), may activate a variety of downstream binding partners, each contributing to synaptic plasticity outcomes. The conditions at which certain binding partners most strongly activate are increasingly studied using computational models. Nearly all computational studies describe these binding partners in combinations of only one or two CaM binding proteins. In contrast, we combine seven well-studied CaM binding partners into a single model wherein they simultaneously compete for access to CaM. Our dynamic model suggests that competition narrows the window of conditions for optimal activation of some binding partners, mimicking the Ca2+-frequency dependence of some proteins in vivo. Further characterization of CaM-dependent signaling dynamics in neuronal synapses may benefit our understanding of learning and memory formation. Furthermore, we propose that competitive binding may be another framework, alongside feedback and feed-forward loops, signaling motifs, and spatial localization, that can be applied to other signal transduction networks, particularly second messenger cascades, to explain the dynamical behavior of protein activation.
| Calcium (Ca2+) is well-recognized as an important second messenger in cellular signaling. One of the most widely expressed Ca2+ binding proteins, calmodulin (CaM), is a highly conserved protein in the EF-hand family [1] (Fig 1A). CaM has over 100 reported downstream binding proteins, including enzymes that regulate a variety of cellular functions, such as neurotransmitter release in presynaptic neuronal axons[2], insulin secretion in the pancreas [3], and contractility in muscle [4]. Ca2+-dependent signaling in postsynaptic dendrites of excitatory neurons has been the frequent subject of computational studies (see a recent review [5]). Indeed, it comprises an ideal system for mathematical modeling. Its parameters (molecular concentrations and kinetic rate constants) have been measured using controlled experiments, and experimental interest has produced an abundance of published values for model parameterization [4, 6–21]. Two highly-studied functions of synaptic Ca2+ signaling are the induction and maintenance of long-term potentiation (LTP) and long-term depression (LTD) [22], which are correlated to learning processes and memory storage in various brain regions [23–26]. Both LTP and LTD are accompanied by persistent changes in postsynaptic gene transcription [27], actin polymerization [28], and AMPA receptor trafficking [29] that adjust cellular excitability and, in turn, synaptic strength. Among the best-studied forms of LTP and LTD are those initiated by transient, localized increases in intracellular Ca2+ through postsynaptic N-methyl-D-aspartate receptors (NMDARs). CaM translates Ca2+ signals into either LTP or LTD by forming Ca2+/CaM complexes that bind and thereby activate downstream proteins (Fig 1C) [30]. Upon activation, these CaM-dependent proteins, which include a variety of enzymes—kinases, phosphatases, cyclases, and synthases—initiate protein signaling cascades that differentially modulate gene transcription, actin polymerization, and AMPA receptor trafficking.
The frequency [31], amplitude, duration, and location [32] of Ca2+ fluxes determine the pattern of activation of CaM-dependent enzymes and, in turn, the fate of the synapse. For example, 1 Hz stimulation for 10–15 minutes both increases activation of the CaM-dependent phosphatase calcineurin (CaN, or PP3) [33] and induces NMDAR-dependent LTD [34]. On the other hand, 100 Hz stimulation for 1 second increases Ca2+/CaM-dependent protein kinase II (CaMKII) activation and induces NMDAR-dependent LTP [35]. These and similar observations have led to the consensus that kinase cascades induce LTP, while phosphatase cascades induce LTD [36]. But more recent studies have found that CaN may also contribute to LTP induction [37], and that activated CaMKII can promote LTD [38]. These results suggest that normal initiation and maintenance of LTP and LTD do not simply depend on the Boolean activation of kinases or phosphatases in response to a given Ca2+ signal, but rather on the precise activation of a variety of often-counteracting proteins. Therefore, elucidation of the mechanisms that regulate NMDAR-dependent long-term plasticity depends on a complete understanding of the endogenous tuning mechanisms that pair precise patterns of enzyme activation to certain Ca2+ signals.
Computational studies have demonstrated the role of binding dynamics [39], feedback loops [40], and spatial effects [41] in regulating enzyme activation during synaptic Ca2+ signaling. In this work, we hypothesize that competition among CaM binding proteins for access to CaM may serve as an additional tuning mechanism. The concentration of CaM binding partners in the cell far exceeds that of CaM itself [42], and in vitro studies have demonstrated competitive inhibition among neuronal CaM binding partners [43–45]. But, despite the implicit presence of competition in many computational models of Ca2+/CaM signaling in neurons [41, 46–51] and cardiac myocytes [52–56], just one study [46] has had the explicit aim of investigating competition among CaM binding partners as a regulator of enzyme activation. Antunes et al. use such a model to investigate competitive binding as a potential facilitator of the frequency-dependence of CaM binding partners at low frequency Ca2+ fluxes (5 mHz to 5 Hz) for generalized sets of CaM binding partners. However, it is worth noting that both He et. al. and Slavov et. al. both mention competition for CaM as a part of their broader studies on the frequency dependent behavior of networks of generalized CaM targets [51] and relative activation of kinase versus phosphatase signaling [50].
In this work we develop models of Ca2+ binding to CaM that explicitly includes Ca2+-binding to each of the two termini (N- and C-termini, Fig 1). Previous experimental work has shown that CaM is able to activate downstream binding proteins at sub-saturating levels of Ca2+[57]. Moreover, a previous computational study explicitly including Ca2+-binding to each of the two binding sites (N- and C-termini) of CaM has shown that Ca2+ bound at the C-terminus likely significantly contributes to activation of downstream binding partners [39]. Our models also include seven experimentally-characterized postsynaptic CaM binding proteins expressed in CA1 hippocampal neurons. These mathematical models are used to investigate competition’s potential role as a regulator of Ca2+-dependent protein activation across a range of Ca2+ flux frequencies (0.1 Hz to 1000 Hz) that spans those found in vivo and oft employed experimentally in vitro. Specifically, we first develop a set of “isolated” models simulating CaM binding to Ca2+ and just one binding protein. We then combine the isolated models into a “competitive” model that simulates Ca2+ binding to CaM and CaM binding to its binding partners. The CaM binding proteins in this study have been chosen because they are known neuronal proteins with relatively well-characterized CaM-binding kinetics: adenylyl cyclase type I (AC1), the adenylyl cyclase type VIII N-terminus (AC8-Nt), the adenylyl cyclase type VIII C2b domain (AC8-Ct), calcineurin (CaN, also known as PP2B and PP3), CaMKII, myosin light chain kinase (MLCK), neurogranin (Ng), and nitric oxide synthase (NOS) (Fig 1C). Because our model is devoid of feedback loops and spatial localization, the differences in CaM-binding between the competitive and isolated models are solely due to competitive effects. We demonstrate the ability of competition to “tune” the binding and activation profiles of CaM-binding proteins at various Ca2+ flux frequencies and use the model to explain the counterintuitive role of neurogranin in CaMKII activation and LTP induction.
We use the total concentration of CaM-bound protein as a primary output parameter. This is contrary to most published computational models, which investigate the concentration of Ca2+-saturated CaM (CaM4) bound to each protein. This approach is preferred for three main reasons. First, although most CaM-dependent enzymes are maximally activated by binding CaM4, sub-saturated forms of CaM have also been found to activate these enzymes, albeit at a lower catalytic rate [57]. Therefore, the concentration of CaM4-bound enzyme does not represent the total concentration of active enzyme. Second, not all binding sites in our model increase in catalytic activity upon CaM binding. For these proteins (Ng and AC8-Nt) the CaM4-bound concentration is no more relevant than the concentration bound to apo-CaM or, for that matter, any other sub-saturated form. Third, CaM-binding to non-catalytic sites has been found to influence CaM availability to CaM-dependent enzymes [89, 117], suggesting an important physiological role for minimally-active, yet still CaM-bound, enzymes. Therefore, although the total concentration of CaM bound to each binding site is not a direct measure of its activation, it provides important information about patterns of enzyme activation that cannot be inferred from the concentration bound to CaM4 alone. To obtain a representative measure of total CaM-binding during Ca2+ spiking at a particular frequency, the average value (henceforth designated the average bound concentration, Cb) is calculated by Eq 1:
Cb=1tf−t0∫t=t0tf∑i=02∑j=02[TbCaMNiCj]dt
(1)
Tb={AC1…NOS}
Where the subscript b indexes the binding partners, so the average bound concentration for a given binding partner (Cb) is found by integrating the total concentration of that binding partner (Tb) bound to each CaM state (CaMNiCj, i and j = 0, 1, or 2) over the stimulation period (to until tf) and dividing by the stimulus duration (tf—to). To measure relative levels of CaM-binding across various proteins and experimental conditions, for each binding partner we normalize Cb by its peak value from among all the Ca2+ frequencies simulated.
We observe that for competitive models, the frequency range at which Cb peaks may shift or narrow relative to the isolated case. To quantify this tuning, we define a metric of frequency specificity (Sb), where the subscript b indexes the binding partners. A binding partner with high frequency specificity is one that most significantly binds CaM over a narrow range of frequencies; correspondingly, this binding partner’s frequency-dependence curve would have a tall, narrow peak. First, the frequency-dependence curve is integrated and then normalized by the maximum Cb (Eq 2). We also divide by the total simulated frequency range and subtract from 1 to report Sb as a metric that identifies the most strongly tuned binding partners. In Eq 2, f denotes Ca2+ frequency.
To investigate how competition alters the CaM-binding dynamics of each of the eight binding partners, we plotted the normalized concentrations of individual partners bound to different CaM states: apo-CaM (CaM0), CaM bound to two Ca2+ ions at its N-terminus (CaM2N), CaM bound to two Ca2+ ions at its C-terminus (CaM2C), and CaM4 (Fig 3). In each simulation, 10 Ca2+ fluxes (not plotted) were introduced at 10 Hz, corresponding to the logarithmic midpoint of our chosen frequency range. In Fig 3, the different colors of the plotted traces correspond to the concentration of binding partner bound to each of the four CaM states normalized to the total concentration of all CaM-bound binding partner (CaMtot). The time-course of CaM binding partners bound to various states of CaM in micromolar for 1 second of 10 Hz Ca2+ flux is plotted in Fig S3 in S1 Appendix.
As expected, the presence of competitors decreases the concentration of CaM bound to each binding partner. Because the relative contributions of the various CaM states to each binding partner’s CaMtot in the competitive model were similar to those in the isolated model, competition did not appear to have a disproportionately large effect on the binding of any one CaM state. This suggests that CaM, and not Ca2+, is the major limiting factor in the activation of CaM-dependent enzymes in hippocampal dendritic spines. Furthermore, competition appears to change not just the concentration of CaM bound to each partner, but also the CaM-binding dynamics. To paraphrase, concentrations in the competitive model are not simply scaled versions of their counterparts in the isolated model. Instead, competition seems to change how each binding partner responds to rapid Ca2+ transients, including how CaM-binding changes with each subsequent Ca2+ flux. For example, after just three Ca2+ fluxes, the concentration of CaM-bound MLCK no longer changes in the isolated model, while it continues to increase in the competitive model. Conversely, while the CaM-binding of Ng decreases with each subsequent Ca2+ spike in the competitive model, it does not change in the isolated model. Therefore, the dynamic behavior of CaM targets in cellular environments cannot necessarily be inferred from computational studies that model them in isolation.
Finally, although competition attenuates the CaM-binding of all binding partners, the magnitude of their attenuation varies considerably in our model. For example, while NOS experiences virtually no change in CaM-binding in the presence of competitors, CaN experiences a more than 20-fold reduction in CaM-binding in the competitive model. Therefore, the binding partners are unequally competitive under the simulated conditions. From these observations, we hypothesize that the competitiveness of each binding partner (i.e., the ability of a binding partner to bind CaM in the presence of other binding partners) might not be absolute and, instead, that the competitiveness of each protein may change across environmental conditions. In this case, competition for CaM is well-positioned to serve as a tuning mechanism, suppressing the CaM-binding of each binding partner for all but a small range of internal conditions and external stimuli and allowing for the tight control of enzyme activation needed for the precise regulation of LTP, LTD, and other neurological processes. Therefore, we investigate how competition may tune the CaM-binding of each neuronal protein to certain Ca2+ frequencies.
To investigate our hypothesis that competition affects the frequency-dependence of CaM-binding, we construct frequency-dependence curves for all eight CaM binding sites (distinguishing between each AC8 terminus) using both the isolated and competitive models (Fig S2 in S1 Appendix). The frequency dependence of Cb is then projected onto heat maps (Fig 4A and 4B). For all simulations, Ca2+ oscillations consisted of 100 concentration spikes ranging from 0.1 Hz to 1 kHz.
The introduction of competition shifts the frequency-dependence curves of almost all binding partners. For some, such as AC1, AC8-Ct, CaMKII, and MLCK, this shift is slight, but apparent. For other partners, such as AC8-Nt and CaN, this shift is dramatic. In the competitive model, maximal CaM-binding occurs at frequencies almost one order of magnitude lower for AC8-Nt (10 Hz in the competitive model, as compared to 60 Hz in the isolated model). For CaN, maximal CaM-binding occurs at frequencies over two orders of magnitude lower (0.3 Hz in the competitive model, as compared to 80 Hz in the isolated model). For NOS, a frequency shift is present but not visible in Fig 4B.
Although, as stated earlier, total CaM-binding and enzymatic activation are not the same (particularly for CaN, which is subject to dual regulation by Ca2+/CaM and CaNB), it is worth noting that CaN is activated by low, but not high, frequency stimulation in vivo [33]. Therefore, it would be expected that maximal CaM-binding of CaN occurs at a similarly low frequency. The fact that this held true in the competitive, but not in the isolated, model suggests that the in vivo frequency-dependence of CaN may be reliant upon the presence of cellular competitors. Because of both the established role of CaN in LTD induction [36, 64] and the demonstrated ability of low frequency stimulation to induce LTD [33], our results further suggest that competition for CaM may be essential to normal LTD induction. Furthermore, because activated CaN downregulates LTP induction [63], competitive suppression of CaM-binding to CaN at high frequencies may be equally essential to normal LTP induction.
To investigate the effects of competition on each CaM binding partner’s level of preference for a certain frequency range, we used the frequency-dependence curves to calculate the frequency specificity of each binding partner in both the isolated and competitive models as defined in Eq 2. If a binding partner were only active at one frequency, it would have a frequency specificity of 100 percent.
The introduction of competitors sharpens the frequency-dependence curves of almost all binding partners, as also indicated by increased frequency specificity values in the competitive models relative to the isolated models (Fig 4C); frequency specificities increased for AC1 (42.80%, as compared to 38.73%), AC8-Nt (58.55%, as compared to 24.79%), CaN (39.13%, as compared to 27.19%), CaMKII (54.17%, as compared to 37.89%), MLCK (27.45%, as compared to 5.70%), Ng (58.23%, as compared to 16.64%), and NOS (1.77%, as compared to 0.08%). The sole decrease, AC8-Ct, was small (37.86%, as compared to 39.50%). Therefore, competition for CaM not only regulates CaM-binding by changing the frequencies of maximal CaM binding, but also by narrowing the range over which appreciable CaM binding occurs.
Two studies have reported decreased CaMKII autophosphorylation and CaMKII activity in CA1 hippocampal slices harvested from Ng genetic knockout (Ng-/-) mice [126, 127]. Although both studies reported about a 30% decrease in CaMKII autophosphorylation and CaMKII activity, they were in disagreement concerning the effect of the genetic knockout (Ng-/-) on LTP induction. Pak et al. (2000) found that wild type (Ng+/+) mice required a single tetanus to achieve potentiation, while Ng-/- mice required multiple tetanic stimulations [126]. In direct contrast, Krucker et al. (2002) found that Ng-/- mice required only a single tetanus to induce LTP [127]. Despite these inconsistent results, both sets of authors suggested that this phenomenon may be caused by abnormal regulation of local Ca2+ and CaM concentrations, a proposal that has since been supported by several studies.
For example, Huang et al. (2004) attributed diminished LTP in Ng-/- mice to lower levels of free Ca2+ following high frequency stimulation [73]. And using two sets of Ng mutants which, respectively cannot bind, and constitutively bind, CaM, Zhong et al. (2009) provided evidence that abnormal regulation of local CaM concentrations may also be responsible. Using a model of the interactions of Ca2+, CaM, CaMKII, CaN, and AMPARs, Zhabotinsky et al. (2006) reproduced the effects of Ng knockout on LTP induction reported by Huang et al., but did not address the diminished CaMKII activity reported by both Pak et al. and Krucker et al. To date, no mathematical model has replicated the paradoxical effect of Ng genetic knockout on autonomous CaMKII activity.
We hypothesize that these phenomena could be explained by competitive tuning. We simulate autonomous CaMKII activation by extending our model according to a previously-published model of CaMKII autophosphorylation by Pepke et al. (see Fig 6 in [39]). In that work, two CaM-bound (active) CaMKII monomers form a complex that enzymatically catalyzes the phosphorylation of one of the monomers. We stimulate this extended model according to an LTP induction protocol followed by Krucker et al., in which hippocampal slices were subjected to two tetanic stimuli of 100 pulses at 100 Hz, 20 seconds apart. Using this protocol, we assess our isolated (Fig 5A) and competitive (Fig 5B) models’ responses to simulated Ng knockout at 600 seconds after the last stimulus. Normalized results from the same experimental stimulation protocol by Krucker et al. are shown in Fig 5C (see activity data in Fig 1F in [127]).
In the absence of other competitors, the isolated model elicits similar levels of CaMKII autophosphorylation (pCaMKII) whether in the presence or absence of Ng. That is, the complete removal of Ng, which competes with CaMKII for CaM, results in only a slight increase in pCaMKII (Fig 5A). In contrast, in the presence of competitors for CaM, simulated Ng knockout decreases pCaMKII levels by 44% compared to WT (Fig 5C). Notably, this decrease in pCaMKII is quantitatively similar to the roughly 33% loss of Ca2+-independent CaMKII activity indicated by Krucker et al. [127]. Further, our competitive model results are also consistent with Pak et al., who report a 40% decrease in pCaMKII in KO Ng-/- mice compared to WT (Ng+/+) mice [126].
Because our model does not allow for either spatial effects or variations in free Ca2+ concentration, these results suggest that competition for CaM alone could explain the paradoxical effect of Ng genetic knockout on CaMKII autophosphorylation and activity. pCaMKII levels seem to be regulated, at least in-part, by the competition for CaM established by Ng. With Ng, an abundance of the CaM not bound to Ng preferentially binds CaMKII (at moderate Ca2+ levels) because CaMKII can out-compete the other candidate binding partners. Without Ng, as in Ng-/- knockout mice, this competitive advantage of CaMKII to bind CaM is reduced, likely because the CaM that would normally bind Ng instead binds other proteins that do not dissociate as readily when high levels of Ca2+ are introduced.
This interpretation predicts that the decreased CaMKII autophosphorylation and activity seen in the Ng-/- knockouts occurs as a result of increased CaM-binding to other partners. To identify which other partners most preferentially bind CaM upon decreasing Ng, we first employ “semi-isolated” models containing only Ng and one of the seven other CaM binding partners (Fig 5D). Semi-isolated models are utilized in Fig 5D to help ensure that shifts in binding partner activation with decreasing Ng are in fact due to decreasing Ng. The partners that experience the greatest relative increase in CaM-binding as Ng concentration is decreased are AC8-Ct and AC8-Nt (calculated according the average CaM bound concentration, Cb, in Eq 1). A more pronounced increase in the average bound concentration of AC8-Ct and AC8-Nt is seen in full competitive model simulations at decreasing Ng concentrations (Fig 5E). This could indicate that the decrease in CaMKII autophosphorylation and activity in Ng-/- mice is due to the shift in availability of (that is, the competition for) CaM due to its increased binding to AC8 during high frequency stimulation. To investigate this, the average bound concentrations of AC8-Ct and AC8-Nt are summed together into AC8 in Fig 5E and plotted along with the average bound concentration of CaMKII as a function of initial Ng concentration for both isolated and competitive model simulations. CaM-binding to AC8 appears sufficient to explain these changes, with the amount of increase in the average bound AC8 concentration at decreasing Ng concentration closely mirroring the decrease in the average bound CaMKII concentration.
In the present study, we use a system of ordinary differential equations to model the dynamic interactions of Ca2+, CaM, and seven CaM target proteins implicated in LTP and LTD of hippocampal synapses. By developing both “isolated” and “competitive” models of this system, we observe competition among these target proteins for CaM-binding and investigate competition’s role in regulating the frequency-dependent activation of downstream CaM binding proteins. The dynamic behavior of our model is largely determined by kinetic rate constants that describe the binding of CaM to Ca2+ and CaM binding to downstream binding to CaM binding proteins. Our models are parameterized using published values where available, and are calculated by applying experimentally supported assumptions and the thermodynamic principle of microscopic reversibility. Global sensitivity analyses are performed to determine the impact of these assumptions on our conclusions, and we find that very few of the parameters that significantly impacted our results are derived from these assumptions.
One of the major results of this work is that competitive binding could be among the mechanisms by which protein activation is dynamically tuned and regulated. We find that the presence of competitors affects not only the concentration of all respective CaM-bound proteins, but also the CaM-binding dynamics of these targets. Based on the results of the present work, we recommend at least the inclusion of Ng into models simulating the activation of CaM-dependent proteins in response to low frequency Ca2+ transients and the inclusion of CaMKII into models simulating the activation of CaM-dependent proteins in response to high frequency Ca2+ transients. Based on the results of our global sensitivity analyses, these two proteins appear to have the most significant impact on the CaM-binding of other CaM targets at these frequency ranges.
Another major result of this work is that competitive tuning may be able to explain the counter-intuitive results from studies of Ng knockouts in mice (Ng-/-) in which CaMKII autophosphorylation and activity levels were seen to decrease in the Ng-/- compared to WT. Our results suggest that under tetanic stimulation and normal initial Ng concentration, Ng buffers CaM from AC8 but not CaMKII. At low concentrations or in the absence of Ng, AC and particularly AC-Ct, is able to bind more CaM, while CaMKII binds less CaM (Fig 5F). Although the KD value of CaM4 binding to CaMKII and AC-Ct are only within 2-fold of each other (1.7 μM and 0.8 μM, respectively), they exhibit very different binding dynamics based on their binding of sub-saturated CaM (CaM2C and CaM2N). This is best seen in Fig 3. For AC-Ct, the dominant species of CaM that binds is CaM2N, making up greater than 50% of the total CaM species bound to AC-Ct. In contrast, for CaMKII there is no dominant species of CaM that binds; CaM2N and CaM4 are major contributors to the total CaMKII-CaM bound species. The binding dynamics of CaM-CaMKII interactions that are seen in the competitive model suggest, as previous work has suggested [39], that CaMKII binds to CaM2C and this CaM2C is then converted to CaM4 while still bound to CaMKII, as noted by the coincident decline in CaMKII-CaM2 and increase in CaMKII-CaM4 in Fig 3 and S3 Fig in S1 Appendix. The binding dynamics of AC8-Ct seem to indicate that AC8-Ct binds CaM2N and stays bound until the next Ca2+ spike. Thus, we hypothesize that AC8-Nt is able to out compete CaMKII for CaM binding in absence of Ng because of its relatively high affinity for CaM2N. Since the dynamic behavior that we see in the competitive model is so dependent on the rate parameters it would be ideal if more of them could be experimentally determined in the future. To test the hypothesis that AC8 activity would be increased in a Ng-/- model, we suggest an experiment in which cAMP production is measured in CA1 hippocampal slices from Ng+/+ and Ng-/- mice while employing forskolin and specific AC1 blockers to control for cAMP production by AC1 and G protein activation, respectively. If our proposed model is accurate, then increased cAMP production will be observed in Ng-/- mice.
Protein networks for which the initiating ligand is a limiting resource, such as the Ca2+/CaM network studied here, are common in biology. As in vivo ligand concentrations often approach the dissociation constants of their binding partners, the concentration of bound ligand could exceed that of free ligand, resulting in the phenomenon of ligand depletion [128]. Ligand depletion, as described by Edelstein et al., reduces cooperative interactions and broadens the range of signals to which the ligand is most responsive. It may be that we observe ligand depletion phenomena in our isolated models (Fig 4A), given the broad range of Ca2+ frequencies at which many binding partners are activated, especially for AC8 and MLCK. However, if ligand depletion really were the predominant regulatory phenomenon, we would expect that by introducing more binding partners (Fig 4B), the broadening effect of ligand depletion would become more conspicuous. Instead, we see a shift and narrowing of the Ca2+ frequencies over which the binding partners are activated. Thus, we are confident that it is competition among the CaM-binding proteins that is the mechanism underlying this tuning behavior.
Because competition seems to be important in our neuron-based model, we sought to compare our results to a different biological system with Ca2+/CaM-dependent signaling. The 2008 publication by Saucerman and Bers examines activation of CaMKII and CaN in a compartmentalized model of cardiomyocytes, stimulated at Ca2+ frequencies ranging from 0–4 Hz [52, 53]. Although this frequency range is much narrower than that used in our competitive model, we can still compare trends of frequency-dependent protein activation. For example, CaMKII activation increases with frequency for both models. Additionally, our isolated model agrees with the Saucerman-Bers model without CaM buffers, in which CaN activation dramatically increases over 0-4Hz. In our competitive model CaN activation is attenuated, in agreement with the Saucerman-Bers model with CaM buffers. This agreement lends further confidence to our model, as the Saucerman-Bers results were subsequently verified experimentally [129]. It appears our model using explicitly-defined CaM buffers (binding proteins) is consistent with the Saucerman-Bers implementation of generalized, unidentified CaM buffers.
The 2008 model by Saucerman and Bers, though not explicitly spatial, highlights how protein localization may affect model output. In Saucerman’s model, Ca2+ frequency-dependent activation levels are different for the cytosolic and membrane-localized (dyadic) CaN sub-populations. Our current model excludes spatial effects in order to scrutinize competitive binding in the absence of confounding factors. However, we acknowledge that spatial effects likely alter competition for CaM, especially in the PSD. Future work would investigate the effect of spatial localization on competition for CaM binding; in particular instantiating membrane-localized proteins such as AC1, AC8, NOS and especially Ng at or near the membrane. Sub-populations of CaN may also be localized to the PSD through binding with scaffolding proteins such as AKAP79 [130–132]. Indeed, because we describe Ng as freely diffusing, it is possible our model exaggerates the ability of Ng to compete for CaM relative to other proteins in our model. Therefore, it would be interesting to assess whether a competitive model accounting for membrane localization can still explain the paradoxical effect of Ng-/- on CaMKII autophosphorylation.
Together, our results suggest that the frequency-dependence of CaM targets observed in vivo is not an inherent property of these proteins, but rather may be an emergent property of their competitive environment. This competitive tuning may provide a mechanism by which otherwise-independent protein pathways can engage in crosstalk through the limited availability of CaM. We propose that competitive tuning, alongside binding dynamics, feedback loops, and spatial localization, may serve as a major regulator of CaM target protein activation. Furthermore, we have attempted to explain the paradoxical decrease in CaMKII activity seen in Ng-/- mice as a result of the dysregulation of this competitive tuning mechanism. In the absence of spatial effects or aperiodic variations in free Ca2+ concentration, competitive tuning is able to offer an explanation for this phenomenon. It is important to note that other proteins, mechanisms, or pathways not included in this model likely lend robustness and further regulatory mechanisms of this phenomenon. Further, it is unlikely that seven CaM-target proteins studied here are the only CaM target proteins that engage in this type of crosstalk through limiting CaM. If competitive tuning facilitates crosstalk among CaM binding proteins, then genetic disorders, neurological diseases, normal aging processes, and therapeutics that disrupt any one CaM target protein may have non-intuitive effects that extend into other signaling pathways. Computational modeling and analysis will continue to play a large role deciphering these oft counter-intuitive regulatory mechanisms that when disrupted, give rise to complex neurological disorders and other important diseases.
All numerical integration and data manipulation were performed in Mathematica as described in Model Analysis. Reaction equations were implemented using Mathematica [133] with the XCellerator package [134]. XCellerator uses the Law of Mass Action to create ordinary differential equations describing the time rate of change in concentration for each binding partner and their respective CaM-bound states. In Eq 3 we monitor the concentration of a generalized Ca2+/CaM state complexed with an arbitrary binding partner, Tb:
d[ TbCaMNiCj ]dt=konTbCaMNiCj[ Tb ][ CaMNiCj ]−koffTbCaMNiCj[ TbCaMNiCj ]+konTbjC[ Ca2+ ][ TbCaMNiCj−1 ]+konTbiN[ Ca2+ ][ TbCaMNi−1Cj ]+koffTb(i+1)N[ TbCaMNi+1Cj ]+koffTb(i+1)C[ TbCaMNiCj+1 ]−[ TbCaMNiCj ](koffTbiN+koffTbjC+konTb(i+1)N[ Ca2+ ]+konTb(j+1)C[ Ca2+ ])
(3)
where i and j = 0, 1, or 2.
For simulations involving autophosphorylation of CaMKII, we extend the system of differential equations generalized in Eq 3 to describe formation of a complex between two active (CaM-bound) CaMKII monomers (Eq 4). Finally, complexes of CaMKII monomers react such that one monomer behaves as an enzyme and the other becomes the phosphorylated substrate (Eq 5). As stated, we refer directly to the previously-published model of CaMKII autophosphorylation by Pepke et al. (see Fig 6 in [39]).
d[DimerCaMKIIN1,iN2,mC1,jC2,n]dt=konDimer[CaMKIICaMN1,iC1,j][CaMKIICaMN2,mC2,n]−koffDimer[DimerCaMKIIN1,iN2,mC1,jC2,n]
(4)
d[pCaMKIICaMNiCj]dt=kpCaMNiCj[DimerCaMKIIN1,iN2,mC1,jC2,n]
(5)
Where i, j, m, and n = 0, 1, or 2. Phosphorylated CaMKII monomers may also be one of the two participating species in Eq 4.
All the equations for this model can be found in S3 Appendix. Mathematica files for the complete models can be found on the Purdue PURR database: Romano, D.; Pharris, M. C.; Patel, N.; Kinzer-Ursem, T. L. (2017), "Mathematica Files: Competitive tuning: competition’s role in setting the frequency-dependence of Ca2+-dependent proteins." (DOI: 10.4231/R7154F7Q). Our model code is also being uploaded to the BioModels Database [135–137].
Despite our best efforts to constrain our models’ parameter values to those that have been experimentally-measured or those which can be calculated by the principle of thermodynamic equilibrium, it was still a valuable exercise to investigate the effects of the previously-described calculations and assumptions on model conclusions. Therefore, a global sensitivity analysis was used to investigate how uncertainty in parameter values impacted model outputs. Latin hypercube sampling (LHS) was used to simultaneously sample input parameter spaces, and partial rank correlation coefficients (PRCC) were calculated to measure the correlation between variation in parameter values and variation in model outputs. These methods have been previously described (see [39, 75]). In short, for each CaM target, a uniform probability distribution of input parameter values was assumed to either span the experimental range specified in S1 Table or, if a range of experimental values is not present, 50–200% of experimental, calculated, or assumed values. A perfect positive correlation gave a PRCC of 1, whereas a perfect negative correlation gives a PRCC of -1. A threshold of 0.5 was used to select for only the parameters that significantly impacted (either positively or negatively) the average bound concentration of each binding partner, and parameters were then ranked by the absolute value of their PRCCs. For the sake of completeness, the sensitivity analysis was done for the nine-state model of Ca2+-CaM binding. Further discussion and enumeration of our sensitivity analysis is in S2 Appendix.
|
10.1371/journal.pntd.0001649 | Landscape Ecology of Sylvatic Chikungunya Virus and Mosquito Vectors in Southeastern Senegal | The risk of human infection with sylvatic chikungunya (CHIKV) virus was assessed in a focus of sylvatic arbovirus circulation in Senegal by investigating distribution and abundance of anthropophilic Aedes mosquitoes, as well as the abundance and distribution of CHIKV in these mosquitoes. A 1650 km2 area was classified into five land cover classes: forest, barren, savanna, agriculture and village. A total of 39,799 mosquitoes was sampled from all classes using human landing collections between June 2009 and January 2010. Mosquito diversity was extremely high, and overall vector abundance peaked at the start of the rainy season. CHIKV was detected in 42 mosquito pools. Our data suggest that Aedes furcifer, which occurred abundantly in all land cover classes and landed frequently on humans in villages outside of houses, is probably the major bridge vector responsible for the spillover of sylvatic CHIKV to humans.
| Chikungunya is a mosquito-borne virus that infects and sickens people in many tropical, urban regions of the world. This virus circulates in forest cycles of West Africa, where mosquitoes transmit it among non-human primates. It also infects humans via bridge vectors, mosquitoes that feed on both non-human primates and humans. To date, little is known about the environmental factors that influence the abundance and distribution of mosquito vectors that participate in the forest cycle of this virus or about specific mosquitoes that are likely to act as bridge vectors. We studied the distribution and abundance of mosquitoes potentially involved in the forest cycle in southeastern Senegal, as well as their infection by this virus. Satellite imagery was used to classify the region into the 5 most abundant land cover elements, and mosquitoes attracted to humans were collected in sites representing each land cover class. We found that Aedes furcifer, a mosquito that occurs in all land cover types and also enters villages to feed on humans, is probably the most important bridge vector between forest circulation and human populations.
| Chikungunya virus (CHIKV, genus Alphavirus, family Togaviridae) is maintained in a sylvatic cycle in West Africa, where it is transmitted by a suite of sylvatic Aedes mosquito species among a group of reservoir hosts, including African green monkeys (Chlorocebus sabaeus), patas monkeys (Erythrocebus patas) and Guinea baboons (Papio papio), and possibly reservoir hosts in other orders of mammals [1]–[3]. Moreover, CHIKV has a history of emergence into humans followed by sustained human-to-human transmission, with the peridomestic mosquito Aedes aegypti serving as the primary vector [1], [3]. Aedes albopictus also serves as a vector of CHIKV in the human cycle. Indeed, this species, which originated from Asia, is a rapidly expanding exotic species in the Americas, Europe and Africa [3], [4] and was responsible for explosive CHIKV outbreaks in the Indian Ocean, Asia, Europe and Central Africa [1]–[6].
CHIKV infection results in an acute febrile disease accompanied by debilitating arthralgia that begins soon after infection but can persist for years [6]–[8]. CHIKV is usually confined to Africa and Asia. However recent transmission following the arrival of infected travelers has been observed in Europe [3] and there is considerable concern that CHIKV will invade the Americas, where both of its major peridomestic vectors are abundant and infected travelers have arrived from Asia and the Indian Ocean [9].
Although past studies have documented the ability of CHIKV to spill over from sylvatic habitats into humans in West Africa, little is known about the environmental factors that influence the risk of human infection or the participation of specific vector species in transmission from zoonotic reservoir hosts to humans. In eastern Senegal, amplifications of sylvatic CHIKV have been detected in mosquito pools in 1975, 1979, 1983, and 1992 in the Kédougou region. During these amplifications, CHIKV was isolated there from humans (one strain in 1975 and two strains in 1983) and monkeys (Cercopithecus aethiops in 1972, Papio papio in 1975 and Erythrocebus patas in 1983) [2], [10]. Following the 2003 amplification, a human outbreak of CHIKV occurred in 2004 in Kedougou among Peace Corps volunteers. In Western Senegal, three epidemics of CHIK fever have also been reported in 1966, 1982, and 1996 [2].
All of these data indicate frequent infection of humans by sylvatic CHIKV in southeastern Senegal. This transmission to humans may occur due to the movement of people into foci of infection in the forest, or to the movement of infected sylvatic vectors into areas occupied by humans. There is a low probability that humans are infected in the forest itself, as humans frequent the forest during daytime while the vectors described above are active at night. However, humans could be infected by sylvatic vectors in other biotopes that they enter at dusk or at night for farming purposes, or while commuting between their place of work and their village. Nonetheless, vector movement seems the more likely explanation for human infection, as dispersal of sylvatic Aedes vectors, particularly Ae. furcifer, into villages is well documented in Senegal [11], [12] and elsewhere in Africa [13].
In the current study, we sought to better understanding the environmental factors that influence the risk of human infection by CHIKV by rigorously testing the association between specific land cover elements and the abundance of Aedes vectors and of CHIKV infection of those vectors. We measured both the distribution and infection of vectors in multiple sampling plots within 5 different land cover classes (forest, savanna, barren, agriculture and village) and also the distribution and infection of these vectors within and among individual villages.
Our study was undertaken in the Kédougou Region of southeastern Senegal (12°33 N, 12°11 W) close to the borders of Mali and Guinea (Figure 1). The area (1,650 km2; 30 km in north-south and 55 km in east-west direction; center coordinates ∼12°36′N, 12°18′W) is located in the shield region of Senegal, with natural vegetation comprised of a mosaic of open savanna, woody savanna, outcrops of laterite (bowé), and relictual gallery forest, the latter concentrated along valleys and rivers [14]. Deforestation for cultivation and human habitations, as well as desertification, has greatly reduced the forested area, as in many other sub-Saharan regions of Africa. Characterized by a tropical savanna climate [15], the Kédougou region receives an average of 1,300 mm of total annual rainfall, with one rainy season from approximately May through November, and mean temperatures varying between about 25–33°C during the year (Figure 2)(http://www.worldclimate.com/). The human population of the region is ca. 80,000, of whom 55% are under the age of 20. It is primarily rural (84%) with a low density of inhabitants (4/km2), mostly living in small, dispersed villages averaging 60 inhabitants. The economy depends on horticulture and cattle farming, along with hunting, gathering and harvesting wood for crafts, necessitating human contact with forests. The primate fauna of the region includes three species, Guinea baboons, patas monkeys, and African green monkeys, which are known reservoir hosts of CHIKV [1], [2].
A six-stage sampling scheme, summarized in Figure 3, was used to identify ten sampling sites in each of the five predominant land cover classes (village, agriculture, barren, savanna, forest) in the study area. Stage I aimed at minimizing spatial autocorrelation among data collected in any given land cover type and entailed the division of the study area into ten equally sized sampling blocks (i.e., 5 north and 5 south of the central east-west line), each of which would eventually contain one representative sampling site per land cover class. In Stage II, a land cover map was generated by means of a maximum likelihood supervised classification of Landsat 5 Thematic Mapper satellite imagery acquired on June 11, 2009 (WGS Path 201/Row 51). Stage III entailed the extraction of only those areas from the land cover map that would likely be accessible in the field, and was accomplished by reducing the land cover map to a one-kilometer buffer around major roads.
In Stage IV, three 2-hectare sites were randomly selected within each of the five land cover classes (i.e., strata), within each of the 10 blocks, and within the one-kilometer buffer zone around major roads. Of the 150 sites, only one site per land cover and block was retained for mosquito sampling. However, three potential sites were identified initially because accessibility and land cover map accuracy in those specific sites were unclear prior to actual inspections. Stage V involved field visits of the 150 sites: sites that were accessible and representative of the mapped land cover type were retained for the final sampling site selection process (Stage VI); sites that were either inaccessible or unrepresentative of the mapped land cover type were removed from the pool of potential final sampling sites. As a result of Stage V endeavors, Block A1 had to be removed entirely from subsequent analyses due to inaccessibility. To avoid losing 5 sampling sites, Block D2—the most complex and centrally located block—was subdivided into two sub-blocks and Stages IV and V repeated in each. Finally, in Stage VI, one sampling site per land cover class per block was selected randomly from the pool of potential final sampling sites identified in Stage V.
Mosquitoes were sampled via human landing collections, the only effective method for sampling sylvatic Aedes and the most appropriate method for determining human risk of infection. Teams of three collectors working simultaneously in forest, savanna, agriculture, village and barren sites in a particular block from 6–9 PM, based on previous data on biting periodicity [12], collected all mosquitoes that landed on their legs. In each of the ten forest sites, mosquitoes were collected at ground level by 3 collectors. Additionally, in eight of the blocks (A2, B1, B2, C1, C2, D1, E1 and E2), a 9 m high platform was erected to enable collection by an additional 3 persons in the forest canopy. In each village, mosquito sampling was conducted by 6 landing collectors per evening. Five houses were selected in the village, following a transect going from one periphery to the opposite periphery via the center (one house in the center, one in each of the periphery sites, and one between each periphery and the center). Each sampling evening, one indoor and one outdoor collector were positioned at each house. On a given night, collectors would be set up at three houses on one half of the transect: one on the periphery, one at the middle point between the periphery and center, and one at the center. On the next night they were positioned on the opposite side to avoid bias due to possible vector confinement within villages. Sampling was performed monthly for 1 to 4 consecutive nights in each block.
At the end of each collection evening, mosquitoes were frozen and then sorted on a chill-table using morphological identification keys established by Edwards [16], Ferrara et al. [17], Huang [18], and Jupp [19] for the culicines and by Diagne et al. [20] for the anophelines. Mosquitoes were sorted into monospecific pools of up to 40 individuals and frozen in liquid nitrogen for virus detection attempts.
The ovaries from a sample of the unengorged mosquitoes were dissected on a slide containing distilled water. The degree of coiling of ovarian tracheoles was then observed to determine whether the female was parous or nulliparous [21].
To attempt virus isolation, monospecific mosquito pools were homogenized in 2.5 ml of Leibovitz 15 cell culture medium containing 20% fetal bovine serum (FBS) and centrifuged for 20 min at 10,000× g at 4°C. For each homogenate, 1 ml of the supernatant was inoculated into AP61 (Ae. pseudoscutellaris) or Vero African Green kidney cells as described previously [22]. Cells were incubated at 28°C (AP61) or 37°C (Vero), and cytopathogenic effects recorded daily. Within 10 d, slides were prepared for immunofluorescence assay (IFA) against 7 pools of immune ascitic fluids specific for most of the African mosquito-borne arboviruses. Viruses were identified by complement fixation and seroneutralization tests by intracerebral inoculation into newborn mice, as approved by the UTMB Institutional Animal Care and Use Committee.
For the real-time PCR assay, 100 µl of supernatant were used for RNA extraction with the QiaAmp Viral RNA Extraction Kit (Qiagen, Heiden, Germany) according to the manufacturer's protocol. RNA was amplified using real-time RT-PCR assay and an ABI Prism 7000 SDS Real-Time apparatus (Applied Biosystems, Foster City, CA) using the Quantitect kit (Qiagen, Hilden, Germany). The 25 µl reaction volume contained 1 µl of extracted RNA, 2x QuantiTect Probe, RT-Master Mix, 10 µM of each primer and the probe. The primer and probe sequences used those of Weidmann et al. (manuscript in preparation) for CHIKV, including the primers RP-CHIK (CCA AAT TGT CCY GGT CTT CCT) and FP-CHIK (AAG CTY CGC GTC CTT TAC CAA G) and the probe P-CHIK (6FAM –CCA ATG TCY TCM GCC TGG ACA CCT TT- TMR). The following thermal profile was used: a single cycle of reverse transcription for 10 min at 50°C, 15 min at 95°C for reverse transcriptase inactivation and DNA polymerase activation followed by 40 amplification cycles of 15 sec at 95°C and 1 min 60°C (annealing-extension step). Fluorescence was analyzed at the end of the amplification.
For analysis of the distribution of vector species among land cover classes, the average per site of female mosquitoes/person/evening (F/P/E) was used as a measure of absolute abundance. Abundance data were log transformed (log10 (n+1)) and analyzed using ANOVA followed by a Tukey-Kramer post-hoc test. In the case of Ae. africanus, there were too many zero values to conduct a valid ANOVA, so abundance data were recoded as present or absent in a designated site and compared using a contingency table analysis. Comparison of vector abundance between villages was conducted similarly. To analyze the distribution of each vector species in the periphery, middle and center of villages, the average abundance of a given species in each of the three regions of each of the 10 villages, collected outside of houses, was compared using ANOVA. For comparison of the abundance of all species in the periphery versus the center of the village, a paired t-test was used to compare the mean abundance, averaged across the 10 villages, of each of the 6 species at the periphery and center.
Spatial patterns of vector abundance were assessed using both global and local measures of spatial autocorrelation. At the global level, we quantified spatial autocorrelation with standard and cumulative spatial correlograms of Moran's I [23], i.e., graphs of Moran's I coefficients on the ordinate plotted against distance classes on the abscissa. We used eleven distance classes (0 to 5,000 m, 5,000 to 10,000 m, 10,000 to 15,000 m, etc. for the standard correlogram and 0 to 5,000 m, 0 to 10,000 m, 0 to 15,000 m, etc. for the cumulative correlogram), a compromise between Sturge's rule [24] and a straightforward lag distance, and an inverse distance weighting scheme. To test the significance of individual Moran's I coefficients at the 0.05 level, we used 9,999 permutations and a progressive Bonferroni correction to account for multiple testing. A correlogram was considered globally significant at the 0.05 level if at least one of the autocorrelation coefficients was significant at the Bonferroni-corrected level [25]. All Moran's I coefficients were computed using PASSaGE [26]. Moran's I values range from −1 (indicating dispersion) to +1 (indicating correlation). Negative values indicate negative spatial autocorrelation; positive values indicate positive spatial autocorrelation; a zero value indicates a random spatial pattern. At the local level, we quantified spatial autocorrelation with Anselin's [27] Local Indicators of Spatial Association (LISA) statistic using weights based on the four nearest neighbors, 9,999 permutations, and a 0.05 pseudo significance level.Statistically significant LISA statistics include two types of positive spatial autocorrelation (HH = High values surrounded by High values; LL = Low values surrounded by Low values) and two types of negative spatial autocorrelation (LH = Low values surrounded by High values; HL = High values surrounded by Low values
Parous and infection rates were compared using a contingency table analysis. The index of parous and biting was also calculated. Both analyses were conducted in StatView 5.0 ® (SAS Institute, San Francisco, CA) or JMP ® (SAS Institute, Cary, N.C.). The pooled infection rate program (PooledInfRate, version 3.0, Center for Disease Control and Prevention, Fort Collins, CO: http://www.cdc.gov/ncidod/dvbid/westnile/software.htm) was used to calculate minimum field infection rates with a scale of 1,000 and the 95% confidence intervals for the species found positive for CHIKV.
Between June 2009 and January 2010, 39,799 mosquitoes were collected comprising 50 species within 6 genera (Table 1). Among host-seeking females of known or suspected CHIKV vectors, Ae. vittatus (22.98%), Ae. furcifer (18.66%), Ae. dalzieli (15.63%) and Ae. luteocephalus (13.05%) had the highest relative abundance and Ae. taylori (2.00%), Ae. africanus (1.71%) and Ae. aegypti (1.24%) had the lowest relative abundance. Absolute vector abundance showed considerable seasonal variation: Ae. vittatus, Ae. luteocephalus and Ae. aegypti reached their peak abundance in June at the beginning of the rainy season and declined drastically during the following months (Figure 2). Other species peaked twice between July and November 2009. Indeed, Ae. africanus exhibited 2 peaks of roughly equal level in August and October.
The patterns of precipitation and temperature over the mosquito sampling period are shown in Figure 2. With a total precipitation of 1087.3 mm (http://www.tutiempo.net/en/Climate/Kedougou/616990.htm), 2009 had a lower rainfall compared to the average of 1263 mm between 1967 and 1990 (www.worldclimate.com). Total vector abundance peaked at the start of the rains in 2009 in June and declined thereafter as rainfall increased and temperature decreased. However there was a second, albeit much smaller peak in November as rainfall dropped off abruptly and temperatures began to climb.
Potential sylvatic CHIKV vectors also showed significant variation in their distributions among land cover classes (Table 2). All species were collected in all land cover classes, with the notable exception of Ae. africanus, which was absent from barren, agricultural and indoor village sites. A contingency table analysis showed a significant difference in the distribution of Ae. africanus among land cover classes (χ2 = 25.9, df = 6, P = 0.0001); results of the remaining statistical comparisons of absolute abundance are listed in Table 2.
Importantly, all of the mosquito species showed significant differences in absolute abundance among land cover classes except for Ae. furcifer, which showed similar abundance in all five classes (F = 2.13, df = 6, 61, P = 0.062). This species had it highest abundances in the forest-canopy and the village-outdoor. However Ae. furcifer preferred the village outdoor environment, and was significantly more abundant outdoors than indoors in villages. Moreover, compared to the others vectors, it also had the highest abundance in village environment both outdoor and indoor. Indeed, the ratio of the abundance of Ae. furcifer to Ae. dalzieli and Ae. taylori in village-outdoor was 4.5:1 and 146.0:1, respectively.
Aedes africanus, Ae. luteocephalus and Ae. taylori were most abundant in the forest, particularly in the forest canopy. Aedes aegypti was most abundant in the forest at ground level. Aedes vittatus was most abundant in barren, agricultural and ground level forest sites while Ae. dalzieli was most abundant in savannah. The global abundance of CHIKV vectors was comparable across all land cover classes but was significantly lower inside of houses in villages than in any other sites.
As shown in Figure 4, the spatial correlograms of Ae. aegypti, Ae. africanus, Ae. furcifer, Ae luteocephalus, and Ae taylori were not significant (p>0.05), indicating that the abundance of these vectors exhibited no global spatial autocorrelation. Ae. dalzieli exhibits significant positive spatial autocorrelation only in the first distance class and Ae. vittatus significant negative spatial autocorrelation in distance classes 3 and 4. The standard correlogram for the abundance of all vectors suggests significant positive spatial autocorrelation in the first distance class; spatial autocorrelation in subsequent classes is not significant at the Bonferroni-corrected significance level. The cumulative correlograms suggest that most of the spatial autocorrelation in our vector abundance data occured in the first lag (0 to 5,000 m). The vectors with no global spatial autocorrelation generally exhibited the least amount of local spatial autocorrelation (Figure 5). Ae. aegypti exhibited some positive spatial autocorrelation (LL: A2 urban and A2 savannah), Ae. africanus some negative spatial autocorrelation (LH: A2 barren and B2 urban), Ae. furcifer mostly positive spatial autocorrelation (HH: A2 barren, B2 urban, and B2 forest), and Ae taylori mostly negative spatial autocorrelation (LH: in Block C1). Aedes luteocephalus showed very notable clusters of positive spatial autocorrelation (LL in Blocks D2 and D2′) and Ae. vittatus has mostly positive spatial autocorrelation (HH in Blocks C1 and D1 and LL in Block D2). The LISA map for abundance of all vectors (Figure 5) showed that, when combined, there was essentially no local negative spatial autocorrelation. Spatial autocorrelation in the western half of the study area was mostly non-significant. Positive spatial autocorrelation clusters were quite common, with hot spots (HH clusters) limited to the northern half (Blocks C1 and D1) and cold spots (LL clusters) to the east/southeast (Blocks D2, D2′, and E1).
The majority of mosquitoes dissected were parous for all species (Tables 3 and 4). However, Ae. africanus showed the highest parous rate (P<0.0001), while Ae. vittatus had the lowest. The monthly parous rates of each vector, except Ae. africanus (P = 0.06), were significantly different and the highest rates were observed in October, November and December, when almost all females were parous (Table 3). The index of parous rate/biting rate increased from August to December except for a drop in November for Ae. taylori. All the vectors except Ae. furcifer, Ae. vittatus and Ae. luteocephalus had high and statistically comparable parous rates in the different land cover classes (Table 4; P>0.1). The highest parous rates for both Ae. furcifer (P = 0.02) and Ae. vittatus (P = 0.02) were in the village sites and the highest rates for Ae. luteocephalus were in the savanna and village sites (P = 0.06).
Within villages, 5,573 mosquitoes were collected, representing 38 species within 6 genera; Table 2 shows absolute abundance of these species. Aedes furcifer (34.7% of the mosquitoes collected), Ae. vittatus (25.4%), Ae. minutus (13.1%), Ae. dalzieli (8.5%), Culex quinquefasciatus (5.9%), Ae. luteocephalus (2.4%) and Ae. aegypti (3.1%) had the highest relative abundance. Aedes taylori, representing only 0.3% of the mosquitoes collected, had the lowest relative abundance within the villages. None of the individual species differed significantly in their absolute abundance in the periphery, middle and center of villages (Table 2). However, when mean abundance of each of the six species of mosquitoes was compared at village periphery versus center, abundance was found to be significantly higher at the periphery (paired t-test, df = 5, t = 2.6, P = 0.048).
Large and statistically significant differences in absolute vector abundance were observed among villages (Figure 6). Aedes africanus and Ae. taylori had low abundance and were collected at one village (E1) and 5 of the 10 villages (B1, B2, C1, E1 and E2), respectively. Absolute abundance of Ae. vittatus and Ae. dalzieli were highest in the village in block D2′, while absolute abundance of Ae. aegypti was highest in the village in block D1 and that of Ae. luteocephalus was highest in the villages in blocks C1 and B1. Aedes furcifer was least abundant in villages in blocks C2 and D2′. In total, potential CHIKV vectors were present at all villages but were most abundant at the village in D1 (Ngari) and least abundant at villages C2 and D2′.
CHIKV was detected in 42 of the 4,211 mosquito pools collected from June, 2009 to January, 2010. Table 1 lists the number of pools and CHIKV infection rates of mosquito species. The 42 infected pools were distributed as follows: Ae. furcifer (15 pools of females and 1 of males), Ae. taylori (5 female pools), Ae. dalzieli (4 female pools), Ae. luteocephalus (5 female pools), Ae. africanus (2 female pools) and Ae. aegypti, Ae. metallicus, Ae. neoafricanus, Ae. centropunctatus, Ae. hirsutus, An. domicola, An. funestus, An. coustani, Mansonia uniformis and Cx. poicilipes (1 female pool each) captured in September, October, November and December. No CHIKV was detected in mosquitoes collected in the other months. These data represent the first detection of CHIKV in Ae. metallicus, Ae. centropunctatus, Ae. hirsutus, An. domicola, and Cx. poicilipes, and the first observation of CHIKV in a male Ae. furcifer from Senegal. Mean infection rates among species differed significantly (P<0.05). Higher and statistically comparable infection rates were observed in Ae. furcifer males, Ae. taylori, Ae. centropunctatus, Ae. metallicus, Ae. hirsutus, An. domicola and Cx. poicilipes females (P = 0.48). Taking into account the temporal dynamics of CHIKV, the highest infection rates were those of An. domicola in October, Ae. centropunctatus in November and Ae. furcifer males in December. Detailed characterization of the CHIKV isolates and sequences will be described separately.
CHIKV infection rates showed temporal and spatial variation. They were higher in December for Ae. furcifer, Ae. luteocephalus Ae. taylori and Ae. dalzieli. The differences were statistically significant except for Ae. taylori (P = 0.42) and Ae. luteocephalus (P = 0.2). CHIKV was detected from mosquitoes collected in 8 of 10 blocks (A2, B1, B2, C1, C2, D1, E1, E2) and in all land cover classes (Table 1), including 7 forest (24 pools), 3 savanna (5 pool), 3 barren (pools), 2 agricultural (4 pools) and 3 village (5 pools) sites. To assess variation among land cover classes, each site was coded as positive (at least one CHIKV-positive pool) or negative (no CHIKV-positive pools). Based on this coding, there was no significant association between land cover class and presence of CHIKV (χ2 = 8.0, df = 4, P = 0.09). However, there was a significant difference among blocks (χ2 = 17.7, df = 9, P = 0.04), with CHIKV being detected in all land cover sites in block D1, no land cover sites in blocks D2 and D2′, and some but not all sites in the remaining blocks. There was a significant, positive correlation between total vector abundance and the number of CHIKV-positive pools across sites (Spearman rank correlation, N = 50, P = 0.003).
The mosquito fauna of the Kédougou region is very diverse. Since the initiation of entomological studies in the area, over 102 species belonging to more than 7 genera have been collected [2], [11], [28], [29]. This high diversity is due to the availability of a wide variety of larval habitats (such as clean slow-running streams and ponds, temporary and semi-permanent pools, and small water collections on the ground or phytotelmata), vertebrate hosts, nectar sources, resting and mating places. However, the amount of diversity detected varies widely among studies, depending on specific sampling methods used (human landing collections alone or with animal baited trap, and larval sampling) and the time period and area covered.
The goal of this study was to determine when and where humans may be exposed to sylvatic CHIKV infection and to identify the bridge vectors responsible for such spillover. To accomplish this, we measured the relative abundance and parity of all putative vectors across different land cover classes at the onset of, during, and immediately after the rainy season. Additionally we conducted detailed sampling within villages to assess exposure to vectors inside versus outside of houses and at the center versus the periphery of villages. The study was specifically designed to avoid spatial autocorrelation by random selection of sampling sites within larger sampling blocks, and as expected we detected minimal levels of such autocorrelation. We collected few potential CHIKV vectors inside houses, indicating an exophagic feeding behavior of these mosquitoes. However, these vectors actively sought human hosts in all land cover classes investigated. In the evening, when the vectors peak in landing rates [11], [30], humans are generally within villages, suggesting that most exposures to sylvatic arboviruses occurs within villages in this region. Additionally, the majority of mosquitoes we collected were parous, indicating that they were in their second or a subsequent gonotrophic cycle and thus had high vectorial capacity. The season increase in the index of parous rate/biting rate suggests little or no recruitment of new mosquitoes to the biting population in October, November and December. Parous rates of vectors were higher in villages than other land cover classes, so humans are at risk of being infected by sylvatic CHIKV in every type of land cover we sampled, but are at greatest risk while outside of houses within villages. Across all species, vector abundance was higher at the periphery of villages than in the center, suggesting that vectors invade villages from surrounding land cover types and that risk of infection may therefore be highest at the edges of villages.
The unexpectedly high host seeking activity of mosquitoes in land cover classes where their known, preferred hosts (humans and monkeys) are not generally present, such as barren areas, suggests that they probably feed on other crepuscular or nocturnal vertebrates. These other species could also be involved in undocumented enzootic cycles of CHIKV in the Kédougou area, as has been suggested by associations of CHIKV with birds, bats and other mammals in Africa [2], [31], [32], [33].A more comprehensive understanding of the enzootic ecology of this virus in the region will require the identification of other potential vertebrate hosts and the description of their roles in the sylvatic cycle of CHIKV. Collection and identification of bloodmeals from feral, engorged vectors will be necessary to achieve this objective.
We associated five mosquito species with CHIKV for the first time. These new associations may reflect the wide spatial and seasonal scope of our study, since all the previous studies of CHIKV in the Kédougou area focused on only one forest-gallery site and a few villages. Detection of CHIKV from a male Ae. furcifer in the Kédougou region during our investigation, and in Ivory Coast [34], may suggest vertical transmission of this virus. Dengue and yellow fever viruses have also been detected in male Ae. furcifer and Ae. furcifer-taylori in Kédougou in previous studies [11], [35].The ecology of sylvatic Aedes mosquitoes in Africa has been well studied because of their role in the transmission of yellow fever virus [30], [35]. We demonstrated that the distribution of some vector species, such as Ae. luteocephalus, Ae. taylori and Ae. africanus, was largely restricted to the forest canopy. This observation is consistent with most similar studies in East and West Africa [36], [37], although Ae. africanus was collected within human settlements and inside houses in southeastern Nigeria [38], [39]. In combination with data suggesting that these mosquitoes feed only during the evening [12], [30], our data suggest that these exophilic species are primarily involved in the maintenance of the zoonotic, sylvatic cycle of CHIKV with little impact on spillover into humans.
Aedes furcifer, in contrast, had high and comparable abundance in the forest canopy and in villages outside houses. It was the only species that frequently contacted humans in villages, corroborating previous observations [11], [40]. Abundance of this species differed significantly among villages and occurred at lowest density in the two most developed of the ten villages we studied. This species is also the only one of the putative sylvatic vectors that is commonly infected with sylvatic arboviruses within villages in the area [11], [40]. Thus it is likely that Ae. furcifer is the principal vector for spillover of sylvatic arboviruses into humans in this area. However, the extreme generalism of Ae. furcifer for different land cover classes is unusual, and we caution that investigation of the population genetics of this species is warranted before firm conclusions can be made about its role as spillover vector.
The fact that the CHIKV was detected in 3 of the 10 villages, and that the distribution of CHIKV was significantly different among sampling blocks, suggests that the risk of transmission to humans may be localized or spatially or temporally heterogeneous. These findings also suggest the need to further characterize the different land cover classes in order to identify subclasses that could differ among blocks. Vector abundance showed a positive correlation with the number of CHIKV-positive pools detected at a site, but vector density may not be the only explanation for variation in the distribution of CHIKV, and therefore this phenomenon merits further study. For example, these three villages in which CHIKV was detected are the closest to gallery forests of the ten villages studied.
Although Ae. dalzieli and Ae. vittatus were widely distributed within the study area (in forest floor, savanna, barren and agricultural sites), and had high abundance in some villages, they have never been found infected with CHIKV within villages in the Kédougou area. Thus, these two species could be involved in virus dissemination from the forest to other land cover classes and could also play a role in potential secondary transmission cycles of the virus among as-yet unidentified species, but are unlikely to be important for spillover of sylvatic CHIKV. Aedes aegypti showed low human landing rates in all land cover classes. Previous studies have also found that Ae. aegypti did not land on humans in high numbers in the Kédougou area [11], [12]. The low abundance of human-seeking Ae. aegypti females despite high larval population density of this species in villages is probably due to its zoophilic tendency in West Africa [41], [42]. Indeed, only the sylvatic form, Ae. aegypti subspecies formosus, occurs in the Kédougou area [43], and this subspecies is thought to feed mainly on wild animals other than primates. Thus, although Ae. aegypti aegypti is the main CHIKV epidemic vector worldwide [1], [8], [44], Ae. aegypti formosus probably plays no major role in either maintenance of sylvatic cycle or spillover to humans in this area.
In summary, our data give new insight into the temporal and spatial dynamics of the extraordinarily diverse guild of sylvatic CHIKV mosquito vectors in an area where, at regular intervals, this virus undergo amplifications in their animal reservoirs that result in spillover infection of humans. While many vectors may participate in maintenance of sylvatic CHIKV, Ae. furcifer is most likely to be responsible for spillover into humans due to its broad land cover preferences and rates of human contact within village perimeters. This information can be used to inform the local population of the places and times of greatest risk for exposure so that mosquito avoidance or protective measures can be implemented. The detection of CHIKV-infected mosquito pools only during the rainy season was expected, but the aggregation of infected pools in specific sampling blocks, rather than in particular land cover classes, was not. We recognize that limited sampling for only a few hours per day and during only one year could have resulted in some anomalous findings or biased results. Additional surveillance and further analysis will be needed to reveal the ecological factors that shape the distribution of CHIKV; our surveillance efforts in Kédougou are ongoing to accomplish this goal.
|
10.1371/journal.pgen.1004872 | Mutations in Global Regulators Lead to Metabolic Selection during Adaptation to Complex Environments | Adaptation to ecologically complex environments can provide insights into the evolutionary dynamics and functional constraints encountered by organisms during natural selection. Adaptation to a new environment with abundant and varied resources can be difficult to achieve by small incremental changes if many mutations are required to achieve even modest gains in fitness. Since changing complex environments are quite common in nature, we investigated how such an epistatic bottleneck can be avoided to allow rapid adaptation. We show that adaptive mutations arise repeatedly in independently evolved populations in the context of greatly increased genetic and phenotypic diversity. We go on to show that weak selection requiring substantial metabolic reprogramming can be readily achieved by mutations in the global response regulator arcA and the stress response regulator rpoS. We identified 46 unique single-nucleotide variants of arcA and 18 mutations in rpoS, nine of which resulted in stop codons or large deletions, suggesting that subtle modulations of ArcA function and knockouts of rpoS are largely responsible for the metabolic shifts leading to adaptation. These mutations allow a higher order metabolic selection that eliminates epistatic bottlenecks, which could occur when many changes would be required. Proteomic and carbohydrate analysis of adapting E. coli populations revealed an up-regulation of enzymes associated with the TCA cycle and amino acid metabolism, and an increase in the secretion of putrescine. The overall effect of adaptation across populations is to redirect and efficiently utilize uptake and catabolism of abundant amino acids. Concomitantly, there is a pronounced spread of more ecologically limited strains that results from specialization through metabolic erosion. Remarkably, the global regulators arcA and rpoS can provide a “one-step” mechanism of adaptation to a novel environment, which highlights the importance of global resource management as a powerful strategy to adaptation.
| Changing environmental conditions are the norm in biology. However, understanding adaptation to complex environments presents many challenges. For example, adaptation to resource-rich environments can potentially have many successful evolutionary trajectories to increased fitness. Even in conditions of plenty, the utilization of numerous but novel resources can require multiple mutations before a benefit is accrued. We evolved two bacterial species isolated from the gut of healthy humans in two different, resource-rich media commonly used in the laboratory. We anticipated that under weak selection the population would evolve tremendous genetic diversity. Despite such a complex genetic background we were able to identify a strong degree of parallel evolution and using a combination of population proteomic and population genomic approaches we show that two global regulators, arcA and rpoS, are the principle targets of selection. Up-regulation of the different metabolic pathways that are controlled by these global regulators in combination with up-regulation of transporters that transport nutrients into the cell revealed increased use of the novel resources. Thus global regulators can provide a one-step model to shift metabolism efficiently and provide rapid a one-step reprogramming of the cell metabolic profile.
| Adaptation to novel environments can proceed either through many mutations with small effects or through few mutations with large effects [1]. Adaptation to complex environments is the norm in biology, but a clear understanding of the adaptive processes employed by organisms in ecologically diverse environments is challenging. Ecological complexity can arise from increased species diversity, spatial or temporal heterogeneity or different resources. The availability of countless resources in complex environments can lead to a rapid and substantial increase in genetic diversity that can obscure broader biochemical principles of adaptation. For example, if resources are varied and plentiful, the population can accumulate mutations in genes that are not essential for survival in the selective environment [2], [3]. Thus, in nutrient-rich environments, specialists with narrower niches can persist by using alternative resources without necessarily improving their fitness relative to the ancestor. Adaptation then occurs through specialization via fitness improvements or via metabolic erosion possibly without fitness improvements relative to the ancestor [3]–[5]. We hypothesized that, despite the tremendous opportunities for increased genetic diversity under conditions of plenty, consistent adaptive responses could be observed as parallel evolution across and within independently evolved populations. We further reasoned that such general and consistent adaptive responses could be driven by global metabolic regulators to provide an efficient reprogramming of metabolic networks with a minimal number of steps.
Experimental evolution of model organisms under novel conditions is a versatile approach for understanding the evolutionary dynamics of adaptation and the functional constraints that shape the physiological evolvability of an organism. Typically, microbial model organisms are selected for adaptation to a single or a few distinct resources [6]–[11], to antibiotics [12], [13], or to temperature [14], [15]. Experiments in such relatively simple selective environments have shown that during adaptation to a single resource, the evolving population typically climbs a single-peak fitness landscape in an incremental manner with diminishing returns epistasis [16]–[18]. Even in such simple environments, resource partitioning or spatial or temporal heterogeneity can lead to the evolution of different specialists and complex ecological interactions [10], [11], [19]–[26].
To better understand the evolutionary and adaptive dynamics in ecologically complex environments, we focused on resource availability and conducted selection experiments in very nutrient-rich conditions. Unlike adaptation to a single limiting resource that is often conceptualized as a single fitness peak, a wealth of resources will potentially present abundant peaks in the fitness landscape. Because resources may differ only slightly, the selection differences can be very small and are reflected as very modest fitness peaks. As a consequence, selection will be weak and lead to an increase in genetic variation through the accumulation of mutations, though the fixation of any specific mutation would be unlikely. Identifying adaptive mutations in such genetically diverse populations can be difficult. However, we reasoned that adaptive mutations should evolve repeatedly in independent populations, while neutral or deleterious mutations should not show any discernible degree of parallelism. Parallel evolution has been readily observed in nature regardless of the ecological complexity [11], [27], [28]. Such adaptive, phenotypic convergence can be based on different underlying genetic changes, such that adaptive, parallel changes can occur in the same gene, or in different genes of the same pathway or functional group [12], [15].
In an environment where selection is weak and selective differences are small, it is hard to imagine a scenario where an individual can quickly accumulate mutations along a multi-protein pathway that will lead to increased fitness. Instead, mutations in regulatory genes such as transcriptional or translational regulators that can simultaneously affect many operons or entire pathways could produce much larger benefits and circumvent potential complications from epistatic interactions among different mutations. One example of a gene with such large pleiotropic effects is the global stress response regulator rpoS, which is activated during late exponential and stationary phase [29]. Mutations in rpoS are often among the first mutations to evolve during experimental evolution of E. coli [30] and have been routinely observed in different selective conditions [31]–[33]. These mutations lead to changes in the stress response and nutrient acquisition [34], change the stress induced mutation rates [35], [36] and increase long-term viability [37]. Knocking out rpoS leads to a down-regulation of the starvation stress response and efflux pumps, and to increased nutrient efficiency via the up-regulation of proteins such as porins. The trade-off of stress resistance and nutritional competence was termed the SPANC balance (self preservation and nutritional competence) by Ferenci [34]. While the prevalence of rpoS knockout mutants is low in wild isolates [38], considerable variation in rpoS expression has been observed among wild strains [39]. We therefore hypothesized that mutations in global regulators could be especially beneficial in ecologically complex, nutrient-rich environments that induce weak selection.
In contrast to previous experiments where laboratory adapted strains were evolved in rich media commonly used in the laboratory [40], we isolated naïve strains from their natural habitat, the gut of healthy humans, and used rich media as novel, selective environments. In the gut, E. coli and C. freundii interact with hundreds of other strains as well as their human host. More than 90% of these commensal gut bacteria are anaerobes, which convert non-digestible complex carbohydrates into short-chain fatty acids and produce the simple mono- and di-saccharides favored by E. coli [41], [42]. In return, facultative anaerobes like E. coli and C. freundii play an important role in maintaining a low-oxygen environment. We chose two complex media (BBL BHI and LB Miller) that differed primarily in the composition and amounts of amino acids, vitamins and carbohydrates (S1 Text). In addition to the populations selected in complex media (two genotypes in two environments resulting in four treatments, Fig. 1), we also performed a control experiment, where we reduced selection as much as possible by daily bottlenecking the population to a single cell, the approach commonly used for mutation accumulation (MA) experiments [43]–[45]. In mutation accumulation experiments, independently evolved lines are expected to accumulate a random set of mutations with little to no parallel evolution.
We used a powerful combination of population proteomics and population genomics to reveal phenotypic convergence to identify potential biochemical mechanisms of adaptation. Despite the complexities imposed by the tremendous amount of underlying genetic diversity accumulated during adaptation to complex nutrient rich environments, we identified clear genomic signatures of adaptation across and within independently evolved populations. Strikingly, changes in the global regulators arcA and rpoS evolved consistently, while changes in other global regulators were largely absent. Subsequent proteomic and carbohydrate analysis of populations adapting to BHI showed increased abundance of enzymes associated with the TCA cycle and amino acid metabolism to make use of abundant amino acids, resulting in the secretion of the polyamine putrescine as a nitrogen sink. Thus in complex media, where the adaptive landscape is relatively flat and has many potential modest peaks requiring many changes to produce a substantive increase in fitness, the “go to” strategy may be to use global regulators such as arcA and rpoS to overcome epistasis by changing the regulation of whole metabolic pathways in a coordinated manner. This allows populations to rapidly reprogram resource utilization and to adapt to complex fitness landscapes in a much smaller number of moves.
We isolated E. coli RU1 and C. freundii RU2 (S1 Figure) de novo from the gut flora of healthy humans using only two overnight growths on agar to reduce any selection prior to our adaptation experiment. For each species, 12 populations were established and allowed to adapt over a minimum of 500 generations to media and conditions that were very rich in resources and substantially different than the environment of the human gut (Fig. 1). We chose LB and BHI as novel environments since they are likely to be very different from the gut resource base, but still support robust growth of both ancestral strains (S1 Table).
Over the course of the selection experiments, we observed modest but significant changes in various fitness components, consistent with adaptation under weak selection conditions. While lag time decreased in most treatments, maximum growth remained constant in LB and decreased in BHI (S1 Table). In all but one treatment, LB-evolved C. freundii, the populations significantly increased their stationary phase density (OD600) indicating enhanced abilities to utilize the resources efficiently. While we observed significant changes, the differences between ancestor and evolved populations were modest and consistent with permissive environments inducing weak selection pressures (S1 Table, S1 Text).
As a consequence of weak selection, considerable genetic variation evolved over the course of our experiment. This was evident both at the phenotypic as well as the genotypic level. We observed considerable phenotypic variation in colony size and in the ability to utilize arabinose (Fig. 2), in redox activity, in exopolysaccharide content and loss of motility (S2A-D Figure, respectively). Interestingly, evolved E. coli populations had at least one colony among the 8 colonies assessed per population that lost motility, but only one single C. freundii colony out of all the colonies assessed (two sets of 96 colonies in total) lost motility (S1 Text).
To assess the evolved genetic variation and identify adaptive mutations, we sequenced the evolved populations and identified mutations in coding regions that occurred at a minimum frequency of 0.05 in a population. Two BHI-evolved populations (E.coli BHI5 and C. freundii BHI20) could not be aligned properly and were omitted from further genomic analyses. The number of mutations ranged from 29 to 725 per population. The number of mutations per population did not differ significantly among the E. coli populations evolved in LB or BHI (Fig. 3; Table 1). Two populations evolved to become mutators in each environment (LB4, LB11, BHI6, and BHI10, S1 Text). If the mutator populations are excluded, the average number of mutations between the LB and BHI-evolved populations was reduced, although there was still no significant difference in the number of mutations across environments. In contrast, the LB-evolved C. freundii populations accumulated significantly more mutations than the BHI-evolved populations. Overall, the number of mutations differed significantly both between media and species (Full factor ANOVA with Media and Strain as fixed factors: Media F1,42 = 15.1, p = 0.0004, Species F1,42 = 4.5, p = 0.039, Media×Species F1,42 = 9.5, p = 0.0036). While synonymous mutations can have fitness effects [46], [47], we focused our analyses on non-synonymous mutations, which include SNPs, insertions, and deletions. The number of non-synonymous mutations ranged from 5 to 198 in a population, with more mutations arising in the LB than in BHI in the E. coli population (Fig. 3, Table 1). Excluding the mutator populations reduced the average non-synonymous mutations per population further (LB: 21±15, BHI: 12±7). Among the C. freundii populations, the average number of non-synonymous mutations was significantly higher in the LB-evolved populations than in the BHI-evolved populations. Again, we observed significant differences among media and species (Media = F1,42 = 23.8, p<0.0001, Species: F1,42 = 13.2, p = 0.0007, Media×Species: F1,42 = 9.4, p = 0.0037).
The accumulation of largely non-adaptive mutations complicates the identification of adaptive changes within a single, polymorphic population. However, we expected that important adaptive trajectories would exist across independently evolved populations. Therefore, we focused our analyses on parallel, non-synonymous mutations that evolved consistently across populations, both within and across species and media. Most mutations occurred in only one or a few populations, consistent with the presence of large non-adaptive genetic variation (Fig. 4). Strong parallel evolution across environments and species occurred in the global regulator arcA, which acquired mutations in all 24 LB-evolved populations and in nine of eleven BHI-evolved E. coli populations (Fig. 5). The probability that mutations evolved in the same gene in 24 independently evolved populations at random is very small considering that E. coli RU1 and C. freundii RU2 had 4565 and 5068 annotated genes, respectively (p = (1/4565)12 *(1/5068)12) and suggests that these mutations are adaptive. Surprisingly, C. freundii adaptation to BHI did not implicate arcA. The second most commonly mutated gene was the global stress response regulator rpoS [29], [48], which had mutations in nine of 23 E. coli and eight of 23 C. freundii populations. None of the mutation accumulation lines had mutations in arcA or rpoS.
Besides arcA and rpoS, only a few other genes acquired mutations in replicated populations, and unlike arcA and rpoS, these other mutations occurred only within a treatment and not across species and selective environment. No other mutation evolved with any degree of parallelism in the E. coli populations. Among the C. freundii populations, mutations in a gene encoding the Valine-Glycine Repeat Protein G, vgrG, a homolog to the tailspike of bacteriophage T4, and in a gene encoding adenosylmethionine-8-amino-7-oxononanoate aminotransferase, an enzyme involved in biotin biosynthesis, occurred in almost all evolved populations in both selective environments, while mutations in different mobile elements, in the peptide deformylase, the methionyl-tRNA formyltransferase and the sodium/glutamate symport protein were only common among the LB-evolved C. freundii populations (for further details see S1 Text).
The highly parallel evolution of mutations in arcA and rpoS combined with their global effects suggests that these mutations are driving adaptation in these complex selective environments. Mutations in these genes were very common with multiple different alleles co-occurring within the same population. The cumulative frequencies of arcA mutations in particular reached high frequencies in LB (average 0.75±0.08 (mean and 95%CI) across 24 populations (Fig. 6). In only one population did we observe the fixation of a single arcA mutation (LB5).
We observed 46 unique mutations in arcA, both within and among populations (Fig. 7). Strikingly, none of these mutations introduced a stop codon or a frame shift; 44 of these 46 unique mutations were non-synonymous substitutions, one mutation resulted in a C-terminal deletion of three amino acids, and one mutation was an insertion of one amino acid. To independently confirm some of the mutations identified from population genomics, we directly sequenced arcA from eight single colonies isolated from six of the LB-evolved E. coli populations. We were able to confirm eleven of the 46 mutations identified in the whole population samples (L2: I122M, Y137C, I22S, L4: N116T, L6: R16H, A76T; L8: E94K; L10: A25T, G59S, L50Q; and L12: L50Q). In addition, we identified two new mutations (L8: G62D and 218ΔTPE; and L12: G62D) suggesting that our cutoff of 5% in the deep sequencing population analysis still missed many arcA variants. Each clone had only one mutation in arcA, suggesting that the one mutation was sufficient to achieve a beneficial effect.
The response regulator arcA is part of the two-component arcAB signal transduction system. The membrane bound sensor kinase ArcB phosphorylates ArcA (ArcA-P) in response to a variety of environmental challenges to maintain redox and metabolic homeostasis [49]–[51]. Mutations in the sensor kinase arcB could also affect the regulation of arcA. We therefore examined the whole genome sequencing data for mutations in the sensor kinase arcB and found mutations in six of the BHI-evolved (BHI1, BHI2, BHI3, BHI4, BHI9 and BHI10) and in one of the LB-evolved E. coli populations (LB8). Among the C. freundii populations, mutations in arcB were less frequent, with only one population evolving mutations in arcB in each environment (LB26 and BHI24). As with arcA, all mutations were non-synonymous, though one mutation did result in a frame shift. Importantly, neither arcA nor arcB evolved mutations in the mutation accumulation lines suggesting that changes in arcAB are under selection and not random.
The amount of variation in rpoS mutations was not quite as dramatic as in arcA, but substantial nonetheless. In the LB-evolved E. coli populations we predominately observed SNPs, while in the BHI-evolved E. coli populations mutations in rpoS resulted in stop codons, large deletions or frame shift mutations, suggesting a loss of function (S3 Figure). Among the five LB-evolved and three BHI-evolved C. freundii populations with mutations in rpoS, only one population (LB31) acquired a mutation resulting in a stop codon while the rest acquired substitutions. To confirm some of these mutations independently, we sequenced rpoS from eight single colony isolates for three of the LB-evolved populations (LB6, LB8 and LB12) and confirmed the A199T mutation in population LB8 as well as a new rpoS mutation (Y283C) in LB6.
Unlike arcA, we observed many loss of function mutations in rpoS, which is consistent with previous selection experiments where knock-out mutations evolved relatively rapidly under different selective conditions [31]–[33]. RpoS levels could also be attenuated through changes in the regulation of its expression. The expression of rpoS is repressed during exponential growth and activated upon starvation during late exponential and stationary phase using different transcriptional and translational mechanisms. Transcription of rpoS is up-regulated through spoT/(p)ppGpp and BarA/UvrY and repressed by arcAB, while translation is up-regulated by two small RNAs, DsrA and RprA and repressed by a third small RNA ArcZ [29]. Presumably, mutations in any of these genes could also affect the up-regulation of rpoS and lead to reduced expression, resulting in similar phenotypes as the knockout mutants. We did not observe any mutations in spoT, though mutations in this gene evolve readily in minimal media [6], [52], [53]. Four of the LB-evolved C. freundii populations (LB27, LB32, LB35, LB36), however, had a frame shift mutation in barA/uvrY. Three of these populations did not have a mutation in rpoS, suggesting that the effect of loss of function mutations in barA/uvrY could lead to reduced transcription of rpoS and result in a similar phenotype to rpoS knockout mutations. Mutations in the small RNAs DsrA and RprA could also lead to reduced translation of rpoS and result in a similar phenotype as rpoS knockout mutants. We looked for mutations in DsrA and RprA in E. coli and did not detect any mutations in these small RNAs. While we also searched for the small RNAs in C. freundii using sequences retrieved from Citrobacter, we were unable to locate the two small RNAs in our reference genome.
Linking mutations to functional changes across regulatory sequences is more difficult as it requires excellent annotation and understanding of the transcriptional regulators. Nonetheless, it is certainly likely that changes in regulatory regions could provide adaptive changes. To test whether arcA or rpoS expression were altered by mutations within their regulatory regions, we examined the 500 nucleotides preceding the start codons of these two genes and found no mutation in any of the evolved populations or the MA lines.
Population-level proteomic analysis of the BHI-evolved E. coli populations showed significant changes in protein abundance between the ancestor and evolved populations as well as a remarkable degree of parallelism among the evolved populations. We observed significant and highly parallel decreases of ArcA abundance in the evolved populations and increases of proteins of the TCA cycle, amino acid metabolism and transporters (Fig. 8). We identified 4469 unique peptides in 39 samples (ancestor and twelve evolved populations with three replicates each), corresponding to 488 proteins (see S1 Text for more details). Quantitative analysis of the 488 proteins revealed 166 proteins that were significantly different between the ancestor and the evolved populations (p<0.01; log2-fold change>±0.7). Of those, 58 proteins decreased and 108 proteins increased significantly over the course of the selection experiment (S2 Table). All observed proteins associated with the TCA cycle (aconitate hydratase, isocitrate dehydrogenase, 2-oxoglutarate dehydrogenase, succinyl-CoA ligase, succinate dehydrogenase, fumarate hydratase, and malate dehydrogenase) and glyoxylate shunt (isocitrate lyase) significantly increased (Fig. 8). The up-regulation of the TCA cycle and the decrease in ArcA is consistent with previous studies that observed increased flux through the TCA cycles in arcA knock-out mutants [50], [54]. RpoS was not detected in our proteomic analyses in any of the samples and therefore we cannot draw any conclusions about its abundance.
Reduced starvation stress is associated with increased nutrient acquisition and metabolism and reduced stress responses [34] – conditions we expected in our resource rich environments. Altered resource utilization can be developed by increasing C/N acquisition via the up-regulation of porins, which allow nutrients to flow through the outer membrane, and a concomitant decrease of the effluxers that provide protection from toxins during starvation stress [34], [55]. Consistent with increased C/N acquisition from amino acids and small peptides, we observed significant increases of peptidases (alpha-aspartyl dipeptidase peptidase E, peptidase B, and methionine aminopeptidase) and of proteins associated with ABC transporter systems responsible for the transport of amino acids or peptides (glutamate aspartate (GltI), lysine-arginine-ornithine (ArgT), glutamine (GlnH), histidine (HisJ) and oligopeptide sytems (OppA)), and carbohydrates (galactose/methyl galactoside (MglB), ribose (RbsB), maltose/maltodextrin (MalE)).
Genomic analyses suggested some loss of function among specific efflux pumps consistent with low stress conditions. We identified mutations in several RND efflux pumps including cmeA and cmeB that are found in multiple copies within the E. coli and C. freundii genomes. Mutations in cmeA and cmeB ranged in frequency from 0.05 to 0.41 and occurred in 15 of 24 populations across both environments and organisms, suggesting that decreases in CmeA and CmeB function are under selection during adaptation. Eight out of twelve LB-evolved E. coli populations acquired mutations in either cmeA or cmeB. Mutations in cmeA that resulted in likely loss of function (all either insertions, deletions or SNPs to stop codons) evolved in four populations (LB5, LB9, LB11 and LB12), while five different populations had mutations in one of the cmeB copies (LB1, LB4, LB5, LB7 and LB8). Mutations in cmeA and cmeB were not as prevalent among the LB-evolved and completely absent among the BHI-evolved C. freundii populations. One LB-evolved C. freundii population acquired a substitution in cmeB (LB25), one had an insertion (LB34) and a third population had an insertion in the RND efflux transporter (LB2). One MA line acquired an insertion in both cmeA and cmeB. The cmeA and B mutations resulting in loss of function mutations support the SPANC balance conditions of low starvation stress and increased nutrient uptake and decreased efflux.
While rpoS mutants are predicted to have a decreased stress response, our proteomic data (Fig. 8, S2 Table) suggested that changes in the stress response, were more nuanced and that some stress pathways, such as the starvation and acid stress responses were up-regulated while others such as protein unfolding stress were diminished. Across the twelve BHI-evolved E. coli populations, we observed decreases in chaperones associated with protein folding stress (DnaK) and heat shock proteins (GroES), and in proteins involved in the oxidative stress response through glutathione (glutaredoxin 2 and 3, and glutathione peroxidase). Conversely, proteins involved in the oxidative stress response through thioredoxin (thioredoxin reductase, universal stress proteins AEFG, superoxide dismutase, glutathione S-transferase), acid stress (HdeAB) and another heat shock protein (HchA/Hsp31) increased.
The up-regulation of the TCA cycle and the increased amino acid acquisition and metabolism could lead to increased production of ammonia or polyamines to maintain nitrogen homeostasis. To test this hypothesis, we determined whether the evolved populations produced and secreted more polyamines or ammonia. We began by testing the pH of spent media after 24 hours of growth, and observed a significant increase in pH from 8.1 to 8.4 in the BHI-evolved E. coli populations (S1 Table) compared to the ancestor. Similarly, the pH also increased significantly in the LB-evolved C. freundii populations, but not in the other two treatments. Increased pH is consistent with proteomic data that suggested significantly increased TCA cycle activity and amino acid metabolism. The breakdown of amino acids by the decarboxylation of ornithine or of arginine to agmatine can result in the production of the polyamine putrescine [56]. Indeed, putrescine was significantly higher in the spent media of the BHI-evolved E. coli population compared to the ancestor (t-test: t = 6.08, df = 22, p<0.0001), but not in the cell extract (t-test: t = 0.3, df = 20, p = 0.76)(S4 Figure). While we only have quantitative data for the BHI-evolved E. coli populations, the odor of the C. freundii populations at stationary phase suggested that they, too, all produced and secreted increased amounts of putrescine.
The natural world presents organisms with complex and variable environments. Resources often range from rich and varied, to poor and limiting. An abundance of new, but usable, resources may induce very weak selection pressures and result in a complex multi-peaked adaptive landscape, where most single nucleotide changes or mutations to specific components of a metabolic pathway would not generate enough fitness gains to facilitate rapid success. One path to adaptation would be to change the global regulators that control the management of metabolic flux to provide a simple “one-step” adaptation for the entire physiology of the organism. Adaptation by such a one-step mechanism constitutes a higher order ‘metabolic selection’ that allows the organism to capture larger gains in fitness and circumvent the complications of multi-gene epistasis. To test this idea, we used wild isolates of E. coli and C. freundii and investigated their adaptive responses under weak selection as they were moved from their natural habitat, the human gut, to a rich and markedly different resource base. We found that, as expected, weak selection induced by rich complex environments resulted in large genetic variation and likely allowed even deleterious mutations to persist. We observed a striking diversity of phenotypes across all populations. Underlying genetic diversity could be observed readily as a tremendous variation in colony sizes and physical appearance on different indicator agar plates, as well as loss of motility (Fig. 2 and S1 Text). Whole population sequencing of the evolved populations identified arcA and rpoS as the targets of selection. Whole population proteomics of the BHI-evolved E. coli populations showed that these populations up-regulated several amino acid and carbohydrate transporters to move abundant nutrients into the cell and up-regulated the proteins of the TCA cycle needed to use them efficiently (Fig. 9). We also observed significantly increasing putrescine production consistent with increased utilization of amino acids as C/N sources. The combination of whole genome sequencing and whole population proteomics proved to be a powerful approach for the mapping of genotypic changes to biochemical mechanisms that, in turn, produce altered phenotypes.
To identify common adaptive strategies, we focused on mutations that arose repeatedly in independently evolved populations. The two most common targets of selection were arcA and rpoS, both global regulators with large pleiotropic effects. Our overall picture for adaptation is one in which the adaptation through mutations in the global regulators arcA and rpoS drive the large metabolic changes essential for adaptation to nutrient rich environments under these selection conditions. Both of these global regulators affect up to 10% of the genes within their host genome [29], [49], [50]. Mutations in arcA evolved consistently in the majority of the populations. ArcA consists of two domains, the receiver domain (residues 1–123) that includes the site of phosphorylation (Asp54) [57]-[59] and a DNA binding domain (124–238) [59]. Phosphorylation stimulates formation of an ArcA-P dimer that binds to a variety of specific DNA sequence motifs with high affinity to repress or activate transcription of up to 229 operons directly or indirectly in response to the environment [49], [50], [54]. The majority of the diverse mutations in arcA were found in the receiver domain (Fig. 7). Mapping these mutations onto the three-dimensional crystal structure (1XHE) of the receiver domain revealed that the vast majority of mutations are in surface positions and solvent accessible loops (S5 Figure), with only a few mutations mapping to the hydrophobic core. While it is likely that some of these mutations could result in a complete loss of function, the likelihood that all 46 mutations do so is slim. It is interesting that all but two mutations in arcA were SNPs and the two exceptions were an insertion and a deletion at the C-terminus and likely resulting in a largely functional protein. This suggests that a complete loss-of-function that eliminates arcA function is not as beneficial as modifying its activity and is consistent with our proteomic data that showed a decrease in ArcA levels rather than a complete absence in ArcA. Mutations within the receiver domain could decrease ArcA signaling by a number of mechanisms including: 1) reducing ArcA stability; 2) decreasing the extent of ArcA phosphorylation during signaling; 3) increasing the rate of dephosphorylation; or 4) decreasing the extent of phosphorylation-dependent oligomerization. Only a few mutations mapped to the DNA binding domain but these could also alter ArcA function by decreasing DNA binding or any of the aforementioned mechanisms for altered receiver domain function. Mutations to arcA were also observed in previous selection experiments, notably during adaptation to glucose limited media [60], [61] and LB [40], [62]. Interestingly, in those studies mutations in arcA only evolved in the aerobic cultures, suggesting that oxygen deprivation and anaerobic stress were not the driver for arcA mutations.
Unlike arcA, it was striking that all mutations to rpoS in the BHI-evolved E. coli populations and only one mutation in the other three treatments appeared to produce a complete loss of function. One explanation could be the differences in the composition of the media, mainly the presence of glucose in BHI. Carbohydrates and glucose in particular are utilized first before switching to amino acids [63]–[65]. This is reflected in a diauxic growth pattern with a second lag and exponential growth phase. The depletion of carbohydrates could induce the RpoS regulated starvation stress response, which might delay the transition to amino acid metabolism. Losing RpoS function might therefore be beneficial to a fast switching response to other nutrient sources. In LB we never observed diauxic growth patterns, which is consistent with the low concentration of carbohydrates in the media. Mutations in rpoS have also been shown to improve longevity during stationary phase [37]. The populations reached stationary phase within eight hours in LB and as a consequence, these populations remained in stationary phase much longer. SNPs that improve persistence in stationary phase could therefore be selected in LB.
Proteomic analyses of BHI-evolved E. coli populations showed increased abundance of enzymes of the TCA cycle, which is consistent with the known phenotypes of arcA and rpoS knockout mutants based on flux analyses [50], [54], [66]. Knocking out either rpoS or arcA resulted in two-fold increases in metabolic flux through the TCA cycle, while knocking out arcB did not have an effect on the flux through the TCA cycle, consistent with our observation that mutations in arcB were rare [54], [66]. While proteomics and flux analyses use different measurements, the effect of mutations in arcA and rpoS seem very similar. This is even more remarkable considering that our proteomic analyses are based on polymorphic populations and the small changes in proteomics are population averages.
In addition, we see strong up-regulation of peptidases, amino acid metabolism, amino acid transporters and oxidative stress responses, indicating that these populations are metabolizing the media at an increased rate (Fig. 8 and 9, S2 Table). Again, this pattern is consistent with previous metabolic studies [54], [66]. ArcA has been shown to either directly or indirectly regulate many operons such as amino acid and polyamine production, beta-oxidation of fatty acid and operons encoding pathways for the utilization of aromatic compounds and peptides [49]. Knock-out mutants of rpoS not only had increased TCA cycle activity but also increased amino acid metabolism [66].
The arginine, asparagine and glutamine metabolism pathways and the TCA cycle feed into the urea cycle. While we did not observe changes in expression of enzymes of the urea cycle, we observed a significant increase in the production and secretion of putrescine in BHI-adapted population. The polyamine putrescine can be produced during the breakdown of amino acids by the decarboxylation of ornithine or by the decarboxylation of arginine via agmatine [56]. Arginine decarboxylase has been proposed to localize in the cell envelope, where it converts exogenous arginine to putrescine via agmatine [67]. Because we observed a significant decrease of agmatinase in the evolved populations, it seems more likely that putrescine is produced from ornithine, a component of the urea cycle. In the adapted populations, putrescine might be acting as a nitrogen sink for catabolism of amino acids via the urea cycle or, alternatively, may help with the increased oxidative stress resulting from increased metabolism. There are many roles for polyamines in metabolism including as C/N sources, oxidative stress response, and signaling [68]–[73], so an understanding of increased secretion of putrescine by our adaptive populations will require further biochemical studies.
We expected that parallel evolution could also evolve along pathways and lead to phenotypic convergence by mutating different genes along the pathways [12], [15]. Indeed, we did see some convergence in the regulation of rpoS, where not all populations had mutations in rpoS and instead had mutations in genes that regulate the expression of rpoS. We interrogated the genomic data for such phenotypic convergence by analyzing parallel evolution for different functional categories, but did not observe any evidence for phenotypic convergence at different levels of increasing complexity (see S1 Text, S6–S8 Figure). Nonetheless, we did see a high degree of parallel evolution among the populations when we grouped the proteins with significant changes to pathways. This is even more remarkable considering that we performed our proteomic analyses on whole, polymorphic populations and as such only measure the average change of a populations compared to the ancestor. This parallelism shows a clear response to selection. The global up-regulation of metabolism and the lack of clear phenotypic convergence of mutations along the pathway further support our assertion that mutations in arcA and rpoS drive adaptation to the rich selective conditions and lead to the observed metabolic changes. These relatively small changes observed in the proteomic data also further support our previous observations that very small biochemical changes can have large fitness effects [74] and are likely very relevant to adaptation in nature.
The consistent evolution of mutations in global regulators arcA and rpoS that each affect expression of about 10% of the genome supports the model in which adaptation evolved through the evolution of a few mutations with large beneficial effects. Instead of acquiring beneficial mutations in every gene involved in the TCA cycle and the various amino acid metabolism pathways, acquiring mutations in regulators affecting all these genes simultaneous is undoubtedly more efficient. Selection studies in other, less complex environments also implicated mutations in global regulators that lead to stable coexistence and polymorphic populations suggesting that mutations in global regulators are beneficial in different environments [60], [61]. In contrast to those earlier studies under glucose-limiting conditions, mutations to arcA and rpoS arose very rapidly and repeatedly across our populations. Diverse phenotypes and genotypes suggest that our populations are polymorphic as well. For example, the sequential utilization of carbohydrates and amino acids [64], [65] could select for different mutations specialized to either carbohydrate or amino acid utilizations, similar to the stable coexistence of different temporal specialist observed in previous studies [10], [19], [20], [25]. It is possible that arcA and rpoS mutations provide selective benefits in different phases of the growth cycle and the coexistence of these mutations in the populations could be an indication of such temporal and potentially nutritional specialization.
By investigating adaptation of wild organisms to resource rich environments we have shown that adaptation occurs within five hundred generations through mutations in global regulators, leading to increased rates of metabolism. Mutations to global regulators might be more common during selection in permissive environments. At niche boundaries such as a thermal limits, single mutations could greatly increase the fitness of an organism [75], mostly because fitness at the niche boundaries can be dramatically reduced compared to the niche optimum [76], [77] and thus a small number of mutations or even a single mutation can result in substantial fitness gains.
Both rpoS and arcA have been linked to virulence [78], [79]. Our findings suggest that it is important to appreciate the role of laboratory adaptation when evaluating strains for pathogenicity, especially in light of the fact that rpoS loss-of-function mutations evolve readily in the laboratory but are not found in natural populations [38]. The transition from the natural habitat to laboratory conditions suggests how we might improve experimental evolution studies that require handling and adaptation of pathogens as well as provide a starting point for forensic attribution of strains during outbreaks of novel pathogens.
E.coli RU1 (hence forth referred to as E. coli) and C. freundii RU2 (hence forth referred to as C. freundii) were isolated from the stool of healthy humans (S1 Text, S5 Figure). To minimize any potential for adaptation during the isolation process, we plated stool samples on MacConkey agar plates. After a single overnight growth, half of a single colony was flash frozen at −80°C in Trypticase soy broth with 20% Glycerol (BD, USA) and the other half was used for phenotypic strain characterization. Strain identification by 16S sequencing was done from the frozen sample. All experimental evolution studies started from the frozen primary isolates by using a single clonalized colony derived from the initial snap frozen isolate. The identity of the wild "un-adapted" strains was based on the results of API 20 E (Biomerieux, USA) test strips and species-specific PCR (using forward primer: AGAGTTTGATCMTGGCTCAG, reverse primer: GWATTACCGCGGCKGCTG).
Assays were performed either at the population level or the single colony level. Cells or populations were grown in their selective media (LB or BHI) and grown in liquid media or plated on agar plates made with their selective media, unless otherwise stated. Adaptation to the selective environments was assessed as changes in lag time and growth rate by measuring OD600 over 24 hours of growth in liquid media following Walkiewicz et al. [74]. To test for changes in the pH of spent media, we grew the populations to stationary phase and measured the pH of the media after removing the cells. To test for genetic variation within the populations, we plated the populations at low density on tetrazolium arabinose plates and observed considerable variation in both colony size and in the ability to utilize arabinose. We plated the populations on the selective media supplemented with agar and isolated eight randomly chosen colonies from each of the 12 populations per treatment. These test sets of 96 individual isolates per species and environment were used for three phenotypic assays: 1) the redox state by plating on methylene blue (0.065 g/liter); 2) differences in exopolysaccharides by plating on Congo Red (0.15 g/liter); and 3) loss of motility by plating cells on soft agar (0.25% DIFCO). For more information see S1 Text.
Cell extracts and spent media samples were prepared by growing the evolved populations and ancestral populations to stationary phase, separating the cells from the supernatant by centrifugation and inactivating remaining cells with 70% (v/v) ethyl alcohol.
|
10.1371/journal.pmed.1002718 | A new aging measure captures morbidity and mortality risk across diverse subpopulations from NHANES IV: A cohort study | A person’s rate of aging has important implications for his/her risk of death and disease; thus, quantifying aging using observable characteristics has important applications for clinical, basic, and observational research. Based on routine clinical chemistry biomarkers, we previously developed a novel aging measure, Phenotypic Age, representing the expected age within the population that corresponds to a person’s estimated mortality risk. The aim of this study was to assess its applicability for differentiating risk for a variety of health outcomes within diverse subpopulations that include healthy and unhealthy groups, distinct age groups, and persons with various race/ethnic, socioeconomic, and health behavior characteristics.
Phenotypic Age was calculated based on a linear combination of chronological age and 9 multi-system clinical chemistry biomarkers in accordance with our previously established method. We also estimated Phenotypic Age Acceleration (PhenoAgeAccel), which represents Phenotypic Age after accounting for chronological age (i.e., whether a person appears older [positive value] or younger [negative value] than expected, physiologically). All analyses were conducted using NHANES IV (1999–2010, an independent sample from that originally used to develop the measure). Our analytic sample consisted of 11,432 adults aged 20–84 years and 185 oldest-old adults top-coded at age 85 years. We observed a total of 1,012 deaths, ascertained over 12.6 years of follow-up (based on National Death Index data through December 31, 2011). Proportional hazard models and receiver operating characteristic curves were used to evaluate all-cause and cause-specific mortality predictions. Overall, participants with more diseases had older Phenotypic Age. For instance, among young adults, those with 1 disease were 0.2 years older phenotypically than disease-free persons, and those with 2 or 3 diseases were about 0.6 years older phenotypically. After adjusting for chronological age and sex, Phenotypic Age was significantly associated with all-cause mortality and cause-specific mortality (with the exception of cerebrovascular disease mortality). Results for all-cause mortality were robust to stratifications by age, race/ethnicity, education, disease count, and health behaviors. Further, Phenotypic Age was associated with mortality among seemingly healthy participants—defined as those who reported being disease-free and who had normal BMI—as well as among oldest-old adults, even after adjustment for disease prevalence. The main limitation of this study was the lack of longitudinal data on Phenotypic Age and disease incidence.
In a nationally representative US adult population, Phenotypic Age was associated with mortality even after adjusting for chronological age. Overall, this association was robust across different stratifications, particularly by age, disease count, health behaviors, and cause of death. We also observed a strong association between Phenotypic Age and the disease count an individual had. These findings suggest that this new aging measure may serve as a useful tool to facilitate identification of at-risk individuals and evaluation of the efficacy of interventions, and may also facilitate investigation into potential biological mechanisms of aging. Nevertheless, further evaluation in other cohorts is needed.
| Aging is one of the leading risk factors for most chronic diseases; therefore, measuring aging has important applications for clinical, basic, and observational research.
Persons of the same chronological age may vary in their rate of aging, suggesting that chronological age is an imperfect proxy of biological aging.
Based on traditional clinical chemistry biomarkers, we recently developed a novel aging measure, Phenotypic Age, which can differentiate mortality risk among persons at the same chronological age.
However, little is known about the applicability of this new aging measure for differentiating morbidity and mortality risk across diverse subpopulations such as healthy and unhealthy groups, distinct age groups, and persons with various race/ethnic, socioeconomic, and health behavior characteristics.
We calculated Phenotypic Age for 11,432 adults aged 20–84 years and 185 oldest-old adults top-coded at age 85 years from NHANES IV and assessed its association with morbidity and mortality.
We found that, overall, Phenotypic Age was highly predictive of mortality even after adjusting for chronological age. This mortality prediction was valid across different subpopulations, stratified by age, race/ethnicity, education, disease count, health behaviors, and cause of death. We also observed a strong association between Phenotypic Age and the disease count a person had, after adjusting for chronological age.
Phenotypic Age can facilitate identification of at-risk individuals for a number of diverse conditions and causes of death. Further, it captures risk stratification in both the healthiest and the unhealthiest populations.
In clinical research, it may serve as a useful tool for evaluating intervention efficacy, avoiding the need for decades of follow-up.
Phenotypic Age may also be applicable to basic and observational research, shedding light on factors that alter the pace of aging, and facilitating investigation into potential biological mechanisms and environmental stressors.
| Rapid population aging represents a major public health burden, as aging is one of the leading risk factors for most major chronic diseases [1,2]. As a result, preventive strategies and interventions that promote healthy aging are critical. While everyone ages, the rate at which aging occurs is heterogeneous, and between-person variations in the pace of aging manifest as differences in susceptibility to death and disease. Thus, differentiating aging in individuals of the same chronological age, particularly in early life, will facilitate secondary and tertiary prevention through earlier identification of high-risk individuals or groups. However, a key issue remains in how to measure aging. Further, to be applicable to the clinical setting, such assessment should be easy to conduct using existing instruments, must do a better job at capturing risk stratification than current tools, and should be able to differentiate risk prior to manifestation of disease or disability.
One method for determining whether a person appears younger or older than expected on a biological or physiological level is to compare observable characteristics, reflecting functioning or state, to the characteristics observed in the general population for a given chronological age. A number of aging measures have been proposed using molecular variables, the most prominent being epigenetic clocks (expressed as DNA methylation age, in units of years) [3] and leukocyte telomere length [4]. We and others have previously shown that while these measures are phenomenal age predictors—especially DNA methylation age—their associations with aging outcomes above and beyond what is explained by chronological age is weak to moderate [5–11]. Conversely, aging measures based on clinically observable data, or phenotypes, tend to be more robust predictors of aging outcomes [12–15]. The differences in prediction between these 2 types of measures could reflect that molecular measures may only capture 1 or a small number of changes involved in the multifactorial aging process, while on the other hand, clinical measures may represent the manifestations of multiple hallmarks of aging occurring at the cellular and intracellular levels [12,13,15–18]. While composite scores based on traditional clinical chemistry measures are not mechanistic, their better performance and relative affordability and practicality compared to current molecular measures may make them more suitable for evaluating the effects of aging interventions on an organismal scale, and/or identifying groups at higher risk of death and disease.
Among the existing clinical measures, the majority were generated based on associations between composite variables and chronological age—with no integration of information on how the variables influence morbidity and mortality. Given that individuals vary in their rate of aging, chronological time is an imperfect proxy for building an aging measure [19]. Recently, we developed a new metric, Phenotypic Age (in units of years), that incorporates composite clinical chemistry biomarkers based on parametrization from a Gompertz mortality model [12]. Rather than predicting chronological age—as previous measures have done—this measure is optimized to differentiate mortality risk among persons of the same chronological age, using data from a variety of multi-system clinical chemistry biomarkers. In general, a person’s Phenotypic Age signifies the age within the general population that corresponds with that person’s mortality risk. For example, 2 individuals may be 50 years old chronologically, but one may have a Phenotypic Age of 55 years, indicating that he/she has the average mortality risk of someone who is 55 years old chronologically, whereas the other may have a Phenotypic Age of 45 years, indicating that he/she has the average mortality risk of someone who is 45 years old chronologically.
The goal of this study was to evaluate the applicability of this measure by (1) assessing whether it is a robust predictor of all-cause mortality compared to traditional risk factors, (2) establishing how it relates to various causes of death and/or comorbid conditions, and (3) determining generalizability through assessing whether this new measure is predictive of long-term mortality risk in a variety of subpopulations, e.g., various age groups, racial/ethnic groups, persons with various socioeconomic status (SES), persons with various smoking/drinking habits, disease-free individuals, and groups with various disease counts.
We previously developed Phenotypic Age using data from NHANES III (the third National Health and Nutrition Examination Survey) (1988–1994) [12]. The independent validation sample used here was from NHANES IV (1999–2010, n = 14,008). We excluded participants with missing data on biomarkers or who did not complete at least 8 hours of fasting prior to blood sampling (n = 1,368), with missing data on follow-up time (n = 15), or who did not have survey weights (n = 1,008). The final analytic sample included n = 11,432 adults aged 20–84 years (S1 Table) and 185 oldest-old adults top-coded at age 85 years. On average, the persons excluded tended to be older (2.5 years on average) and were 40% more likely to self-identify as non-Hispanic black. Details of recruitment, procedures, population characteristics, and study design for NHANES are provided through the Centers for Disease Control and Prevention [20] (https://www.cdc.gov/nchs/nhanes/index.htm). Briefly, NHANES is an ongoing program by the National Center for Health Statistics involving a series of independent, nationally representative cross-sectional surveys designed to assess the health and nutritional status of adults and children in the US. It began in the early 1960s focusing on different population groups and health topics and became a continuous program that has had a changing focus on a variety of health and nutrition measurements to meet emerging needs since 1999. Using both at-home interviews and examinations performed at a mobile examination center, NHANES collects a wide range of information (e.g., via demographic, socioeconomic, dietary, and health-related questions, and medical and physiological measurements) from a nationally representative sample each year in counties across the country [20]. NHANES is approved by the National Center for Health Statistics Research Ethics Review Board, and all participants provide informed consent. Data used in this study are de-identified and publicly available (https://www.cdc.gov/nchs/nhanes/index.htm). This study received approval from the Yale Human Investigation Committee on 15 November 2017 following an expedited review.
Mortality follow-up was based on linked data from records taken from the National Death Index through December 31, 2011, provided through the Centers for Disease Control and Prevention [20]. Data on mortality status and length of follow-up (in person-months) were available for nearly all participants (n = 15 with missing data on follow-up time). Out of 9 underlying causes of death and an “other” category that were provided in the linked data, 7 were used to assess cause-specific mortality in our study—heart disease, cancer, chronic lower respiratory disease, cerebrovascular disease, diabetes, influenza or pneumonia, and nephritis/nephrosis. Alzheimer disease was not considered in the cause-specific analysis due to the small number of deaths assigned to this cause. Accidents were not assessed due to the fact that many may not be age-related, and it is impossible to differentiate age- versus non-age-related accidental death.
We calculated Phenotypic Age in accordance with the method described previously [12]. Briefly, Phenotypic Age is calculated using chronological age and 9 biomarkers (albumin, creatinine, glucose, [log] C-reactive protein [CRP], lymphocyte percent, mean cell volume, red blood cell distribution width, alkaline phosphatase, and white blood cell count) that were selected using a Cox proportional hazard elastic net model for mortality based on 10-fold cross-validation. The algorithm for calculating Phenotypic Age is based on parametrization of 2 Gompertz proportional hazard models—one fit using all 10 selected variables, and the other fit using only chronological age. The resulting final equation for calculating Phenotypic Age is as follows:
PhenotypicAge=141.50+ln[−0.00553×ln(1−xb)]0.09165
where
xb=−19.907−0.0336×albumin+0.0095×creatinine+0.0195×glucose+0.0954×ln(CRP)−0.0120×lymphocytepercent+0.0268×meancellvolume+0.3356×redbloodcelldistributionwidth+0.00188×alkalinephosphatase+0.0554×whitebloodcellcount+0.0804×chronologicalage
Finally, we calculated a measure, Phenotypic Age Acceleration (PhenoAgeAccel), defined as the residual resulting from a linear model when regressing Phenotypic Age on chronological age. Therefore, PhenoAgeAccel represents Phenotypic Age after accounting for chronological age (i.e., whether a person appears older [positive value] or younger [negative value] than expected, physiologically, based on his/her age).
Age categories, race/ethnicity, education, body mass index (BMI), disease count, smoking status, and drinking habits were considered for stratified analyses. Four age categories (20–39, 40–64, 65–84, and 85+ years) and 3 racial/ethnic groups (non-Hispanic white, non-Hispanic black, and Hispanic) were considered. Note that for presenting Kaplan–Meier curves, we used a different set of 4 age categories (20–49, 50–64, 65–74, and 75–84 years) to demonstrate the robustness of the results. A 4-category education variable was used to approximate SES. Categories included less than high school (HS), HS/general educational development (GED), some college (having attended college but not receiving at least a bachelor’s degree), or college (having a bachelor’s degree or higher). BMI was calculated as weight in kilograms divided by height in meters squared. Underweight was defined as BMI < 18.5 kg/m2, normal was defined as 18.5 ≤ BMI < 25.0 kg/m2, overweight was defined as 25.0 ≤ BMI < 30.0 kg/m2, and obese was defined as BMI ≥ 30 kg/m2. Chronic diseases included 10 coexisting self-reported conditions: congestive heart failure, stroke, cancer, chronic bronchitis, emphysema, cataracts, arthritis, type 2 diabetes, hypertension, and myocardial infarction. Based on the disease counts, we created a variable with 5 categories—no disease, 1 disease, 2 diseases, 3 diseases, and 4 or more diseases (with the last two categories combined in subgroup analyses). Three smoking status categories were created, which included never smokers (<100 cigarettes during one’s lifetime), former smokers (100 or more cigarettes during one’s lifetime, but not actively smoking during recent time frame), and current smokers (ongoing smoking habit). Two drinking variables were created—a binary binge drinking indicator (in which binge drinking was defined as having 5+ alcoholic beverages at a time at least once per month) and a 6-category alcohol intake variable (never, none in past year, <1 drink per month, 1–3 drinks per month, 1–3 drinks per week, 4+ drinks per week). All the information was collected through a questionnaire or physical examination at the time of survey.
In this study, when comparing the predictive performance of Phenotypic Age with that of traditional risk factors, we not only considered the individual biomarkers that were already included in Phenotypic Age, but also considered disease count, BMI, total cholesterol, and systolic blood pressure, given that they are commonly considered risk factors for death and disease in both observational studies and clinical practice [21–25]. Data on total cholesterol were obtained from blood analyses, and data on systolic blood pressure were obtained from examination at the time of survey.
The analytic plan for this study is briefly described in Fig 1. Using data from NHANES IV, age-stratified ordinary least squares regression models were first used to estimate the association between disease count and Phenotypic Age within 3 age categories (20–39 years, 40–64 years, and 65–84 years). Based on these regression equations, we then estimated the incremental increase in PhenoAgeAccel for participants in each of the disease count categories in comparison to participants with no disease.
Next, a parametric proportional hazard model (Gompertz distribution) was used to assess the association between Phenotypic Age and all-cause mortality, with adjustment for chronological age and sex. To further evaluate robustness, age-stratified models and a model that excluded short-term mortality (within 5 years after baseline) were also run to ensure the mortality prediction was not driven by older ages and/or an end-of-life phenotype. Participants were then grouped into quintiles for PhenoAgeAccel, so that the highest quintile represented individuals who were most at risk of death for their age—i.e., those whose Phenotypic Age was the highest relative to their chronological age. We then plotted Kaplan–Meier curves for persons in the highest 20% versus the lowest 20%. We also compared predicted median life expectancy at age 65 years by sex and the 5 quintiles for PhenoAgeAccel. Next, receiver operating characteristic (ROC) curves were used to compare the 10-year mortality risk prediction of Phenotypic Age to predictions based on individuals’ clinical chemistry biomarkers and routine risk assessment tools (e.g., based on systolic blood pressure, the biomarkers, and BMI).
Cause-specific mortality risk as a function of Phenotypic Age was assessed via Fine and Gray’s competing risk models [26]. To determine whether Phenotypic Age could differentiate risk in population subgroups (e.g., healthy participants), we conducted the all-cause mortality analysis again by race/ethnicity, education, disease count, BMI, smoking status, and drinking habits.
Participants aged 85+ years (oldest-old adults) were excluded from all prior analyses given that age was top-coded (i.e., everyone aged 85+ years was coded as being age 85 years) for identity protection; therefore, to test mortality associations in this group, we used 2 parametric proportional hazard models (Gompertz distribution), one adjusted for sex and another with adjustment for sex and disease count, rather than chronological age (unknown).
All analyses were performed using R version 3.4.1 (2017-06-30) and STATA version 14.0 software (StataCorp, College Station, TX).
The basic characteristics of the study participants are shown in S1 Table. The mean age of the 11,432 adults was 45.5 years, and about half of the sample were women (50.8%). Young (20–39 years) and middle aged (40–64 years) adults accounted for 40% and 45%, respectively. Three-quarters of participants self-identified as non-Hispanic white, about 11% were non-Hispanic black, and 13% were Hispanic. One-quarter of participants had a college degree, about 30% had some college education, one-quarter had a HS education, and about 19% had not graduated from HS or received a GED. Half of the sample were never smokers, while the other half were approximately equal parts former and current smokers. Approximately 15% had binge drinking tendencies over the past year. Finally, proportions of normal BMI, overweight, and obese were each about one-third.
Fig 2 presents the disease counts overall and by age category. Approximately two-thirds (64%) of the study participants were disease-free at their interview, while 22% reported having been diagnosed with 1 chronic disease, 9% reported 2 diseases, 3% reported 3 diseases, and only 2% reported at least 4 coexisting chronic diseases. As expected, the majority (87%) of young adults (aged 20–39 years) were free of disease, compared to 59% of middle aged (40–64 years) and only a quarter (28%) of older adults. Additionally, 7% of older adults had 4 or more chronic diseases, while only 1% of middle aged adults and essentially no young adults reported 4 or more disease diagnoses.
Fig 3 shows the correlation between Phenotypic Age and chronological age, as well as the distribution of PhenoAgeAccel—the residual of Phenotypic Age regressed on chronological age. Phenotypic Age and chronological age are highly correlated; part of this is due to the fact that age is in the Phenotypic Age measure. Consistent with many of the previous aging measures, we also observed that the Phenotypic Age of young adults tended to be overestimated, while the Phenotypic Age of older adults tended to be underestimated. Given that the Δ for Phenotypic Age and chronological age would be biased by age, we estimated the residual for Phenotypic Age, referred to as PhenoAgeAccel. A score of 0 suggests a Phenotypic Age that is consistent with what is expected based on an individual’s chronological age, whereas a positive value suggests that the person has clinical chemistry biomarkers that characterize an older person, and a negative value suggests the person has the clinical chemistry profile of a person younger than expected. While the measure is fairly normally distributed, most of the outliers tend to be in the positive (older) direction.
Fig 4 shows predicted increases in PhenoAgeAccel for each disease count category, compared to persons with no diagnosis of disease. Overall, participants with disease had older Phenotypic Age compared to those without disease. For instance, among young adults, those with 1 disease were on average 0.2 years older phenotypically than disease-free persons, and both those with 2 diseases and those with 3 diseases were about 0.6 years older phenotypically. In middle aged adults, compared to those who were disease-free, those with 1 disease had a Phenotypic Age that was on average 0.2 years older, those with 2 diseases had a Phenotypic Age that was 0.3 years older, those with 3 diseases had a Phenotypic Age that was 0.6 years older, and those with 4 or more diseases had a Phenotypic Age that was 0.7 years older. Finally, for older adults, Phenotypic Age increased consistently as a function of disease count, with those reporting 1 disease having a Phenotypic Age that was on average 0.1 years older than disease-free participants, those with 2 diseases having a Phenotypic Age 0.2 years older, those with 3 diseases having a Phenotypic Age 0.4 years older, and those with 4 or more diseases having a Phenotypic Age 0.6 years older.
Table 1 shows the association between Phenotypic Age and all-cause mortality, based on proportional hazard models with Gompertz distribution. In the full sample, each 1-year increase in Phenotypic Age (after adjusting for chronological age) increased the risk of mortality by 9% (hazard ratio [HR] = 1.09, 95% CI = 1.08–1.10). When restricting the sample to participants who survived at least 5 years after baseline, we found consistent results, such that each 1-year increase in Phenotypic Age was associated with an 8% increase in mortality risk. When examining mortality within age-stratified groups, we found that Phenotypic Age was predictive in all age groups, such that each 1-year increase in Phenotypic Age was associated with a 13% increased mortality risk in young adults, a 10% increase in middle aged adults, and a 8% increase in older adults. Finally, we found that, on average, females were phenotypically younger than males (β = −1.34, P < 0.001); therefore, we compared sex-stratified models of all-cause mortality associations and found identical results for both sexes (HR = 1.09, 95% CI = 1.07–1.11).
As shown in Fig 5, we found that those with the highest Phenotypic Ages relative to their chronological ages had much steeper declines in survival over the approximately 12.5 years of follow-up. Interestingly, the high-risk groups (highest 20% of PhenoAgeAccel) appeared to have mortality rates that were similar, or in some cases higher, than those of persons in the low-risk groups (lowest 20% of PhenoAgeAccel) who were 10 years older chronologically. For instance, among persons aged 50–64 years at baseline, about 25% of the high-risk group had died after 10 years of follow-up. Conversely, among persons aged 65–74 years, only about 20% of those in the low-risk group had died after 10 years of follow-up. For persons aged 65–74 years in the high-risk group, about half had died after 10 years, compared to only about 67% of the low-risk group who were aged 75–84 years at baseline.
Fig 6 presents predicted median life expectancy at age 65 years by sex and the 5 quintiles of PhenoAgeAccel. Results showed that 65-year-old females in the lowest quintile (low-risk, or healthiest) had a predicted median life expectancy of about 87 years, while females in the highest quintile (high-risk, or unhealthiest) had a predicted life expectancy of just over 78 years. Similarly, 65-year-old males in the lowest quintile had a predicted median life expectancy of about 85 years, while males in the highest quintile had a predicted life expectancy of just under 76 years.
ROC curves (Fig 7) revealed that Phenotypic Age, with an area under the curve (AUC) of 0.88, significantly outperformed the individual clinical chemistry measures and other risk factors. The next highest performing measures were chronological age, with an AUC of 0.86; disease count, with an AUC of 0.71; and serum creatinine, with an AUC of 0.71. Four measures had AUCs between 0.60–0.69 (red blood cell distribution width, fasting glucose, systolic blood pressure, and albumin), 5 had AUCs between 0.50–0.59 (mean cell volume, lymphocyte percentage, CRP, alkaline phosphatase, and white blood cell count), and 2 had AUCs less than 0.50 (total cholesterol and BMI).
As shown in Table 2, we reexamined the ROC curves using various combinations of variables, with and without Phenotypic Age included. We found that in all cases, Phenotypic Age contributed additional predictive power to all models. More interestingly, Phenotypic Age alone was more predictive of 10-year mortality than a model that included chronological age, demographics (race/ethnicity and sex), SES (education), and disease count. It was only when chronological age, demographics, SES, disease count, and health behaviors (smoking, alcohol intake, binge drinking, and BMI) were all included in a single model that the AUC started to approach the AUC for Phenotypic Age alone. Nevertheless, Phenotypic Age still added substantial predictive power when included with those variables, suggesting that it captures something above and beyond what can be explained for mortality risk by demographics, SES, disease, and health behaviors.
As shown in Table 1, as expected, there were large frequency differences between the disease-specific causes of death, with the numbers of deaths ranging from 15 (nephritis/nephrosis) to 227 (cancer). Nevertheless, although Phenotypic Age was trained to predict all-cause mortality (which was heavily skewed towards cardiovascular and cancer deaths), we found that Phenotypic Age was predictive of disease-specific mortality including heart disease, cancer, chronic lower respiratory disease, diabetes, influenza/pneumonia, and nephritis/nephrosis, with exception of cerebrovascular disease mortality (HR = 1.03, 95% CI = 0.98–1.09). HRs were the highest for diabetes and nephritis/nephrosis, suggesting that a 1-year increase in Phenotypic Age relative to chronological age increases the risks of death from these causes by about 20%. For the other causes (aside from cerebrovascular disease), a 1-year increase in Phenotypic Age increased risk by between 7% (cancer and chronic lower respiratory disease) and 12% (influenza/pneumonia).
Given the need to identify aging measures that are generalizable across various populations, we examined all-cause mortality associations using stratified models. In general, we found consistent associations regardless of the subgroup (Table 3). Consistent with the HR for the overall population (Table 1) of 1.09, HRs from stratified models ranged from 1.04 (persons with 3+ diseases) to 1.15 (underweight persons). When all variables, such as age, race/ethnicity, sex, education, smoking, and drinking, were adjusted for, Phenotypic Age remained significantly associated with mortality (HR = 1.06, P < 0.001).
Additionally, given the importance of identifying at-risk persons as early as possible, we evaluated whether Phenotypic Age was associated with all-cause mortality among persons who appeared clinically healthy (defined as having no disease and normal BMI). As shown in Table 3, among those healthy participants (n = 1,906), we observed that a 1-year increase in Phenotypic Age was still associated with an 8% increase in all-cause mortality risk.
Table 4 provides the mortality association in oldest-old adults. We found that regardless of adjustment, Phenotypic Age was associated with mortality in this subpopulation, although to a lesser degree than in the full population (unadjusted model: HR = 1.05, 95% CI = 1.01–1.08; disease-adjusted model: HR = 1.05, 95% CI = 1.02–1.08).
Since another similar aging measure—Levine Biological Age (BioAge), which utilizes the Klemera and Doubal algorithm—currently provides one of the most accurate mortality predictors [13], we performed an additional analysis comparing the associations and predictions of Phenotypic Age to those of this measure. The results are provided in S1 Appendix, S2–S4 Tables, and S1 and S2 Figs. Overall, our results suggested that Phenotypic Age and Levine Biological Age were largely comparable, but Phenotypic Age performed better in the healthy subpopulation (e.g., those having no disease and normal BMI).
In a nationally representative US adult population, we showed that our new measure of aging—Phenotypic Age—was highly predictive of mortality even after adjusting for chronological age. Overall, we found that the mortality prediction of this measure is valid across different stratifications, particularly by age, disease count, health behaviors, and cause of death. For instance, Phenotypic Age is strongly associated with all-cause mortality in multiple age groups, including young adults, middle aged adults, and older adults. Moreover, the effect sizes seem to decrease with age, which may suggest that in younger groups, when the risk of death is low, variations in physiological status—as captured by PhenoAgeAccel—may play a bigger role in who lives longer. Conversely, in older adults, for whom the risk of death increases, mortality may be more stochastic. Nevertheless, we were able to determine that this measure was not just capturing an end-of-life or critically ill status, given that it remained predictive of mortality after excluding participants who had not survived for at least 5 years after baseline.
The finding that Phenotypic Age was predictive of mortality among both healthy and unhealthy populations even after adjusting for chronological age is novel. Many of the measures of aging, such as those based on deficit accumulation [14,27], include measures of morbidity in their construction, and thus it is impossible to disentangle aging and disease, or determine the usefulness of such measures in healthy populations. Belsky et al. evaluated aging measures, including Levine Biological Age, in a cohort study of young adults who were mostly disease-free [15,16]. However, the outcomes available were mostly restricted to functional assessments, which may mean something different in younger adults than they do in older populations. Conversely, in this study, we were able to show that Phenotypic Age was predictive of all-cause mortality among disease-free, healthy adults across the age spectrum. This suggests that Phenotypic Age is not simply a measure of disease or morbidity and instead may be a marker that tracks the effect of aging before diseases become clinically evident. This suggests that in a clinical setting, PhenoAgeAccel could be used to stratify risk among persons who otherwise “appear” healthy.
As expected of an aging biomarker, PhenoAgeAccel also tracks multimorbidity. We observed a strong association between the number of diseases a person reported being diagnosed with and his/her Phenotypic Age relative to his/her chronological age. Despite relatively small sample sizes, in general, PhenoAgeAccel appeared to increase as a function of disease count, suggesting that among persons of the same age, the more coexisting diseases a person has, the phenotypically older he/she appears—based on clinical biomarkers. Nevertheless, PhenoAgeAccel predicted risk of death significantly better than disease count, suggesting that it is capturing information beyond a person’s number of coexisting conditions. This is further supported by the significant association of PhenoAgeAccel with mortality in oldest-old adults—a population with high disease prevalence—and, more importantly, this association remained even after adjusting for disease count.
The efficacy of Phenotypic Age for assessing mortality risk in the general population, as well as multiple subpopulations that are heterogeneous in age and health status, provides strong evidence of its suitability for applications in both the clinical setting and research in the biology of aging. For instance, the generalizability of Phenotypic Age in assessing the risk of various aging outcomes may facilitate identification of at-risk individuals for a number of distinct conditions. Phenotypic Age may also be a useful marker for evaluation of interventions—particularly those concerned with prevention via delaying disease pathogenesis [18,28–30]. Aging changes are hypothesized to begin as early as conception [31]—preceding disease—thus interventions to slow aging will be most effective for reducing disease incidence if started early in the life course prior to significant accumulation of aging-related damage. Our findings suggest that Phenotypic Age is in line with the Geroscience paradigm, which stipulates that “aging is the greatest risk factor for a majority of chronic diseases driving both morbidity and mortality” [32,33]. Therefore, measures such as Phenotypic Age that capture pre-clinical aging as well as future morbidity/mortality risk could facilitate evaluation of intervention efficacy, while avoiding the need for decades of follow-up [28]. While research to develop interventions that target the aging process is ongoing, our paper provides a potential end point for which they can be evaluated. Further, this metric may also shed light on factors that alter the pace of aging, facilitating investigation into potential biological mechanisms and environmental stressors.
Despite the promising applications of Phenotypic Age, one limitation of this study is the lack of longitudinal data for either Phenotypic Age or disease incidence. As such, we were unable to confirm whether higher PhenoAgeAccel is predictive of disease accumulation (e.g., among persons with 1 disease, whether PhenoAgeAccel predicts who will develop a second comorbid condition). We were also unable to distinguish the mortality risks associated with (1) the rate of change in Phenotypic Age (true acceleration) versus (2) the baseline level of Phenotypic Age relative to chronological age.
In conclusion, our study shows that after adjusting for chronological age, Phenotypic Age, a novel clinically based measure of aging, is predictive of remaining life expectancy in a nationally representative population. Importantly, its prediction is robust to population characteristics—it is a reliable mortality predictor regardless of the age or health status of the population being assessed. Further, this measure captures both all-cause and disease-specific mortality, and is also strongly associated with the number of comorbid conditions. These findings suggest that this new aging measure may serve as a useful tool to facilitate identification of at-risk individuals and evaluation of intervention efficacy. Nevertheless, further evaluation in other cohorts is needed.
|
10.1371/journal.pgen.1006757 | Brg1 chromatin remodeling ATPase balances germ layer patterning by amplifying the transcriptional burst at midblastula transition | Zygotic gene expression programs control cell differentiation in vertebrate development. In Xenopus, these programs are initiated by local induction of regulatory genes through maternal signaling activities in the wake of zygotic genome activation (ZGA) at the midblastula transition (MBT). These programs lay down the vertebrate body plan through gastrulation and neurulation, and are accompanied by massive changes in chromatin structure, which increasingly constrain cellular plasticity. Here we report on developmental functions for Brahma related gene 1 (Brg1), a key component of embyronic SWI/SNF chromatin remodeling complexes. Carefully controlled, global Brg1 protein depletion in X. tropicalis and X. laevis causes embryonic lethality or developmental arrest from gastrulation on. Transcriptome analysis at late blastula, before development becomes arrested, indicates predominantly a role for Brg1 in transcriptional activation of a limited set of genes involved in pattern specification processes and nervous system development. Mosaic analysis by targeted microinjection defines Brg1 as an essential amplifier of gene expression in dorsal (BCNE/Nieuwkoop Center) and ventral (BMP/Vent) signaling centers. Moreover, Brg1 is required and sufficient for initiating axial patterning in cooperation with maternal Wnt signaling. In search for a common denominator of Brg1 impact on development, we have quantitatively filtered global mRNA fluctuations at MBT. The results indicate that Brg1 is predominantly required for genes with the highest burst of transcriptional activity. Since this group contains many key developmental regulators, we propose Brg1 to be responsible for raising their expression above threshold levels in preparation for embryonic patterning.
| Brahma-related-gene-1 (Brg1) is a catalytic subunit of mammalian SWI/SNF chromatin remodeling complexes. Loss of maternal Brg1 protein arrests development in mice at the 2-cell stage, while null homozygotes die at the blastocyst stage. These early requirements have precluded any analysis of Brg1’s embryonic functions. Here we present data from X. laevis and X. tropicalis, which for the first time describe a role for Brg1 during germ layer patterning and axis formation. Brg1-depleted embryos fail to develop past gastrulation. Genome-wide transcriptome analysis at late blastula stage, before the developmental arrest, shows that Brg1 is required predominantly for transcriptional activation of a limited set of genes involved in pattern specification processes and nervous system development shortly after midblastula transition. Mosaic analysis by targeted microinjection defines Brg1 as an essential amplifier of gene expression in dorsal (BCNE and Nieuwkoop center) and ventral (BMP/Vent) signaling centers, being required and sufficient to initiate axial patterning by cooperating with canonical Wnt signaling. Since Brg1-dependent genes share a high burst of transcriptional activation before gastrulation, we propose a systemic role for Brg1 as transcriptional amplifier, which balances the embryonic patterning process.
| Vertebrate BAF protein complexes remodel chromatin with the mutually exclusive help of Brahma (brm) or Brahma-related gene 1 (brg1) ATPase subunits. SWI/SNF complexes are known to participate broadly in nucleosome-based aspects of DNA metabolism in normal and malignant cells [1–4], but their specific ATPase subunits designate them for different functions. Brm-/- mice are viable although heavier than normal, suggesting Brm to be a negative regulator of cell proliferation [5]. In contrast, brg1-/- mice die during early embryogenesis and brg1 heterozygotes are predisposed to exencephaly and tumor formation [6]. These results suggest unique functions for BAF complexes carrying the different ATPases. Brg1 containing BAF complexes become further subspecialized in a tissue specific manner by association with co-factors of the BAF60 protein family during cell differentiation [7, 8]. Specific functions have also been described in murine embryonic stem cells, where a specialized esBAF complex containing Brg1, Baf155 and Baf60a regulates aspects of ES self renewal, pluripotency and cell priming for differentiation [9–11].
How these findings for esBAF relate to normal mouse embryogenesis is not fully clear. Embryos lacking maternal Brg1 protein arrest at two cell stage and are compromised in zygotic genome activation [12], while embryos, lacking only zygotic Brg1 protein, die before implantation [6]. Since ATP dependent chromatin remodelers, including Brg1, are conserved among vertebrates [10, 13] deeper insight into Brg1’s developmental functions could be derived from non-mammalian vertebrate model organisms. Although Brg1 is expressed throughout development, several reports from Xenopus and Zebrafish have shown only relatively late requirement of Brg1 in development, i.e. during differentiation of heart, neural plate and brain [13–15]. We had obtained precedence for specific involvement of chromatin remodelers in developmental processes as early as germ layer formation in Xenopus, where Mi2-beta/NuRD remodeling activity is needed to position the boundary between mesoderm and neuroectoderm [13]. These findings let us expect also earlier functions for BRG1/BAF complexes. In search for such early functions, we have investigated the transcriptional and embryonic consequences of Brg1 depletion in the closely related species X. tropicalis and X. laevis.
To generate a Brg1 loss of function situation in Xenopus, we designed three Morpholino (MO) oligonucleotides against mRNAs of both X. laevis brg1 homoeologs (Fig 1A). We determined their relative translation blocking activities in X. laevis embryos with a recombinant brg1/luciferase transcript, containing ~700bp of the brg1 cDNA sequences with the morpholino targeting regions fused in frame to the luciferase ORF. The blocking efficiencies of the Morpholinos increased about three-fold from 5’ to 3’ direction on the target mRNA. BMO1 had the strongest effect and reduced luciferase activity approximately seven-fold (S1 Fig). Based on these results, we selected BMO1 and BMO2 for further analysis.
We investigated the consequences of systemic Brg1 protein knockdown in X. tropicalis embryos, where the target region for BMO1 and BMO2 is conserved (Fig 1A). To achieve a homogenous protein knockdown, Morpholinos were injected four times into the animal pole region at the two- to four-cell stage (“radial” injection type). In titration experiments, we determined a dose of 30 ng BMO1/embryo to reduce Brg1 protein levels to one-third, while even 60ng of BMO2 reduced them only two-fold (Fig 1C). Whereas more than 90% of the control morphants survived until hatching, the majority of the BMO1 injected embryos died during gastrulation (Fig 1B). Consistent with less efficient Brg1 depletion, BMO2 morphant embryos died later than BMO1 morphants and survived better (Fig 1B). The survival rate of CoMO and BMO1 morphants was comparable until late blastula, although the Brg1 protein levels were already diminished in the latter case to about 30% of CoMO injected embryos (Fig 1C). The residual BRG1 protein is very likely of maternal origin [14] and therefore insensitive to MO knockdown. Notably, immunostaining for activated Caspase-3 showed no signs of apoptosis at early gastrula stage indicating that morphant blastulae are healthy and initiate gastrulation without visible defects or delay (S2 Fig).
Bulk zygotic transcription commences in Xenopus at the midblastula transition (MBT), about three hours before gastrulation starts. Because under our conditions BMO1 morphants die mostly during gastrulation or neurulation, but not before, it was possible to assess the consequences of Brg1 protein knockdown on embryonic transcription at late blastula. Using the same conditions described above, we compared CoMO and BMO1 injected X. tropicalis embryos by microarray analysis. Although these conditions were ultimately lethal, at the investigated late blastula stage more than 90% of the mRNAs were expressed at normal levels. A total of 872 transcripts responded to the Brg1 protein knockdown, with 211 of them being upregulated, and 661 being downregulated, relative to control morphants (Fig 1D, S1 and S2 Tables). Gene Ontology analysis revealed an enrichment for the terms “chromatin assembly”, “cellular complex assembly”and “macromolecular complex assembly” in the upregulated cohort (S3 Fig, panel A). In contrast, the downregulated responders were strongly enriched in several GO terms related to various developmental and pattern specification processes (S3 Fig, panel B). Here the most enriched term was “nervous system development”, consistent with a known requirement for Brg1 during vertebrate neural differentiation [14–17]. Nineteen of the 61 genes from this GO category were reduced in our genome-wide data set already at the blastula stage, i.e. before neural plate formation. From these 19 genes, fifteen were found reduced by independent qRT/PCR analysis (Fig 1E). In addition we reproduced the microarray results for a variety of important developmental regulatory genes by qRT/PCR (S3 Fig, panels C and D). Furthermore, we investigated by whole-mount RNA in situ hybridization (WMISH) the expression patterns of genes involved in neural induction at late blastula, confirming the microarray results for downregulated foxD4l1 and noggin expression, and unaffected zic2 expression in BMO1 morphants (S3 Fig, panels E-K’). These independent analyses confirmed the microarray data in a robust manner. In summary, our results indicate an essential function for Brg1 protein before the onset of gastrulation, detailing primarily an enhancement of gene transcription.
We repeated the morphological analysis in X. laevis by injecting radially the BMO1 morpholino at 60 ng, 40 ng and 20ng per embryo, together with a fluorescent lineage tracer (S4 Fig). At the two higher doses, BMO1 injections caused again embryonic death in the majority of the embryos before late neural tube stage (NF22; S4 Fig panel C). All remaining embryos were arrested in gastrulation (S4 Fig, panel A), occasionally surviving in this state until the heartbeat stage (NF34; S4 Fig, panel B). At a dose of 20 ng, more than half of the embryos survived until NF22; while about 80% of the survivors did not finish gastrulation, 20% became arrested at the open neural plate stage and remained in this condition until NF34 (S4 Fig, panels B and C). In summary, embryonic survival is correlated with the BMO1 dose. Moreover, the formation of dorso-anterior structures is still completely blocked under conditions (20ng BMO1/radial), in which the majority of the embryos survives beyond neurulation. Therefore, we wondered, whether local ablation of Brg1 protein by a reduced Morpholino dose might improve overall embryonic differentiation and thus provide information on developmental pathways requiring Brg1 activity. We tested this hypothesis in X. laevis embryos by targeted microinjections.
When CoMO or BMO1 oligonucleotides (10ng/embryo) were injected at the 4-cell stage into the marginal zone of either the two dorsal (DMZ), or the two ventral (VMZ) blastomeres, most of these embryos survived until tadpole stage. β-Galactosidase staining for coinjected nlacZ mRNA confirmed the correct targeting of the injections and demonstrated viability of the injected cell progeny. DMZ injected control morphants grew up into phenotypically wildtype tadpoles (Fig 2A). In contrast, 80% of DMZ injected BMO1 morphants displayed stunted antero-posterior body axes with severely truncated heads carrying the nlacZ stain (Fig 2B and 2F). The remaining 20% of BMO1 morphants showed no morphological abnormalities. The major phenotype was a specific consequence of Brg1 protein depletion, since dorso-anterior structures were largely rescued by coinjection of wildtype human brg1 mRNA that contains four mismatches in the BMO1 target region. The heads of these rescued embryos contained well-developed eyes with lenses and recurrent retinal pigment (Fig 2C). Interestingly, this rescue of the BMO1 morphant phenotype was neither achieved with human brm nor Xenopus iswi mRNAs (S3 Table). When we inspected VMZ injected embryos, their overall morphology was much less affected. Although we have not investigated any internal organs, these embryos were at least capable to develop a well-structured body axis including heads (Fig 2D, 2E and 2G). A smaller fraction (30%) of the BMO1 morphants was weakly anteriorized, displaying enlarged heads and eyes, but concomitantly deficient in posterior tissues such as the fin (Fig 2E’ and 2G).
Several conclusions can be drawn from these results. First, the presence of β-Gal positive cells in tadpoles demonstrates that BMO1 injections are not cell-lethal per se. Second the inability of Brm and Iswi to rescue the BMO1 phenotype argues for distinct remodeling events that specifically require Brg1 protein. Finally, the observed morphological phenotypes suggest that Brg1 is involved in axis formation. Particularly on the dorsal side of the embryo, Brg1 protein seems required to unfold the dorsalizing gene expression program (DGEP) during germ layer patterning.
This assumption was investigated by several experiments. First, we coinjected Xenopus brg1 and nlacZ mRNAs into one ventral blastomere at the 4-cell stage of wildtype embryos. At the tadpole stage, these embryos had formed with high penetrance a secondary axis rudiment, which contained somites with differentiated muscle tissue (S5 Fig). Moreover, the chordin gene is one of the early developmental regulators, downregulated in radial X. tropicalis BMO1 morphants (S3 Fig, panel C, S2 Table). This gene was ectopically induced by ventral injections of human brg1 mRNA (S6 Fig). Notably, hBrg1 also efficiently restore dorso-anterior development in DMZ injected BMO1 morphants (Fig 2C and 2F). These two observations indicate that Brg1 protein overexpression can initiate de novo formation of axial structures, apparently through activation of DGEP genes like chordin.
Normally, formation of dorsal structures is initiated by maternal Wnt/β-Catenin signaling on the prospective dorsal side of the embryo [18–20]. Coinjection of β-catenin mRNA at a dose, which alone was insufficient to cause morphological consequences, reestablished quite efficiently head structures, including eyes, as well as longer body axes and tails in DMZ-injected BMO1 morphants (Fig 3A–3D). When injected ventrally, β-Catenin frequently induced a secondary embryonic anlage with complete heads, which was reduced to single-axis status by coinjection of BMO1 (Fig 3E–3H). These last experiments demonstrate a cooperation between canonical Wnt signaling and Brg1 in early embryonic patterning, which had not been observed before.
At blastula stage, DGEP is initiated by two newly induced, local signaling centers–the dorso-animal Blastula-Chordin-Noggin-Expressing (BCNE) Center and the overlapping, but more vegetally located Nieuwkoop Center (NC). Both regions contribute to head-formation and are induced by maternal WNT-signaling [21]. The BCNE signature genes, including nodal3.1/nr3, chordin and noggin, encode secreted BMP inhibitors. All these genes, in particular chordin, were downregulated by radial Brg1 protein knockdown in X. laevis (Fig 4, panels A-C). Expression of the BCNE genes was largely restored by coinjection of human Brg1 mRNA (Fig 4A”–4C”), consistent with the previously described morphological rescue of dorso-anterior tissues (see Fig 2). The downregulation of chordin and noggin mRNAs matches our results in X. tropicalis (S3 Fig, panels C, and G to H’). We note that the expression of nr3 was upregulated in X. tropicalis blastulae (S3C Fig), while it is downregulated in the X. laevis BCNE. Possible mechanisms for this species-specific difference are discussed later.
The NC marker gene sia1 is a direct Wnt-target, a key regulator of axial patterning, and is needed for proper gene expression in BCNE and Spemann’s organizer [22–24]. Notably, expression of sia mRNA in radial BMO1 morphants was not downregulated compared to control morphants (Fig 4D and 4D’), consistent with the results for sia1 and sia2 from the microarray analysis in X. tropicalis (S3C Fig). The insensitivity of sia1/sia2 genes excludes a role of Brg1 protein as general coactivator of Wnt/β-Catenin signaling in Xenopus, which had been suggested by earlier studies [25]. Notably, Brg1 is involved in both direct (nr3; [26]) and indirect [chd, nog; ref.[27]] Wnt-mediated gene activation events in the BCNE, which is consistent with the morphological phenotypes observed in DMZ-injected BMO1 morphants.
The process of germ layer patterning, which defines the future body plan, occurs during gastrulation and requires Spemann’s organizer, which overlaps with and succeeds the BCNE and NC territories. Cells of the organizer are the first ones to involute during gastrulation. They secrete a panoply of proteins, which inhibit BMP as well as Wnt and Nodal signaling pathways [28–30]. These organizer properties generate gradients of signaling activities that dynamically establish gene expression domains of appropriate size within the morphogenetic field of the forming germ layers [31, 32]. Since we had discovered that BCNE gene expression depends on normal BRG1 protein levels, we sought to extend the analysis to gene expression domains of organizer and non-organizer mesoderm.
We evaluated the expression of critical regulatory factors by WMISH in early to mid gastrula stage (NF10.5 to NF11), i.e. before development becomes typically arrested in X. laevis BMO1 morphants (S4 Fig). Among the organizer genes was the BMP inhibitor nr3 [33], which at blastula stage was downregulated in the BCNE. The nr3 mRNA levels were reduced in most gastrulae within its normal domain (Fig 5A–5C). The otx2 gene is first transcribed in the organizer and specifies at later stages anterior tissues in all three germ layers [34]. Upon Brg1 knockdown, the intensity of otx2 staining and the size of its domain were reduced. This was true both for preinvoluted otx2 expression at the blastopore lip as well as for the involuted part, where otx2 mRNA is confined to a narrow stripe in BMO1 morphants (Fig 5D–5F). Finally, foxA4 mRNA was frequently downregulated in its proper domain at the lip (Fig 5G–5I). Some non-organizer genes in the neighboring dorso-lateral mesoderm were also misregulated. This included the homeobox genes vent1 and vent2, which mediate the ventro-posteriorizing activity of BMP ligands [35]. In BMO1 morphants, transcripts from both genes invaded the organizer territory (Fig 5K–5P). This dorsal expansion of vent gene expression indicates a severe functional impairment of Spemann’s organizer [36]. In addition, the muscle regulatory genes myoD and myf5 were reduced (S7 Fig, panels A-F), while gsc, t/bra and xpo were unaffected (S7 Fig, panels G-S). Also chordin transcription was unimpaired in the organizer, despite the fact that it is downregulated in the BCNE region at blastula stage (compare S7K–S7M Fig with Fig 4A). While chordin is activated by maternal Wnt signaling in the BCNE, it is controlled in the organizer by additional regulators including nodal signaling and the mesodermal transcription factors gsc and not [37, 38]. These findings suggest a context-dependent role for Brg1 in target gene regulation. In summary, a systemic depletion of BRG1 protein leads to a misbalance in dorso-ventral patterning and an unusual coexistence of non-organizer (vent1/vent2) and organizer transcripts within the dorsal blastopore lip.
Our morphological and molecular analyses defined the earliest defects in BRG1-depleted embryos to the late blastula stage, when the BCNE center is established. Kuroda and colleagues have demonstrated by tissue transplantation that the BCNE contributes to brain formation [21]. Therefore, we decided to address by orthotopic transplantation, whether the absence of eyes and forebrains in BMO1 morphants arises autonomously from the BCNE region, or results from a defective crosstalk between germ layers. For this purpose we transplanted wildtype or morphant BCNE grafts into wildtype host embryos (Fig 6). Grafts were marked by fluorescent dextran to distinguish them from host tissues (experimental workflow see S8A Fig). At the tadpole stage, two-thirds of the embryos transplanted with a WT-BCNE had generated tadpoles with heads containing well-developed eyes with lenses (n = 14/21; Fig 6A–6A”). In contrast, almost 90% of the embryos transplanted with a morphant BCNE lacked eyes completely or had only remnants of retinal pigmentation without lenses (n = 22/25; Fig 6B–6B”). The morphological differences between the two conditions were significant (Fig 6C). To visualize the major brain domains we stained the transplanted tadpoles for otx2 mRNA (S8 Fig, panels B-E). Half of the morphant transplants showed a strong reduction in otx2 expression, in which forebrain, midbrain and hindbrain areas were collapsed to an amorphous tissue mass. This result was particularly obvious in specimen, in which the lineage tracer of the transplanted BCNE populated only part of the brain. In these cases, otx2 staining was structured comparatively normal in the host-derived parts of the brain, while it was severely reduced in the transplanted area (S8 Fig, compare panels C, D with C’ and D’). The clearest results were observed for otx2 expression in the retina, which was present in 90% of the WT transplants, but only in 25% of the morphant transplants. In summary the results from this experimental series indicate an autonomous defect within the Brg1-depleted BCNE region for neuronal differentiation and brain patterning, which cannot be compensated by secreted factors from the wildtype host environment, including mesodermal Chordin (S7K–S7M Fig).
We have demonstrated that several BCNE genes, in particular chordin, are specifically downregulated in BMO1 morphants and are responsible for defective head formation. However, the global transcriptome analysis had revealed a much larger number of genes responding to BRG1 knockdown, suggesting that other regions of the embryo also contribute to the BMO1 phenotype. In a new series of experiments, we compared side by side the consequences of dorso-animal (”DA”/BCNE center) with dorso-vegetal (“DV”/Nieuwkoop Center) blastomere injections at the eight-cell stage. The majority of control morphants developed completely normal in both types of injections (Fig 7, panels A, D, C and F). As expected, DA-injections of BMO1 resulted in embryos with shorter, tailless axes, and strongly reduced heads (Fig 7, panels B and C). Targeting of the BMO1 to DV-blastomeres maintained the length of the main body axis much better than DA-injections, but still reduced head and eye formation in a large fraction of the embryos (Fig 7, panels E and F).
RNA in situ hybridisation and ß-Galactosidase staining at the late blastula stage confirmed that both DA and DV blastomeres contribute to the BCNE expression zone, as it has been described before [39, 40]. Consequently, chordin and noggin mRNAs were downregulated with both injections within the overlap (S9 Fig, panels A-J). The genes hhex and cer1 are expressed in the dorso-vegetal region and are known to promote anterior development and head formation, respectively [41, 42]. DA-injections of BMO1, which do not overlap with the hhex and cer1 expression domains, had no effect on these mRNAs (Fig 7, panels G-L). In contrast, DV-injections strongly downregulated both genes in a statistically significant manner (Fig 7, panels M-Q). Most importantly, the results from targeted 8-cell injections demonstrate an additional role for Brg1 in the Nieuwkoop Center, the prospective anterior endoderm region of the embryo, where it is required for hhex and cer1 transcription.
The morphological and molecular analysis of its protein knockdown phenotype demonstrated Brg1 to be essential for embryonic vitality and germ layer patterning. Targeted injections of BMO1 oligo to different regions of the embryo support this conclusion in a consistent manner and revealed functional connections of Brg1 to several signaling pathways (Wnt/Bmp) and embryonic regions (BCNE/anterior mesendoderm). In search for a common denominator of Brg1 function, which could explain both the diverse impact on embryonic gene expression at late blastula stage (>800 altered transcripts) and the developmental functions in dorsal and ventral signaling centers, we investigated the gene response in Brg1 morphants in relation to the zygotic genome activation (ZGA) at MBT. Precedence for this assumption comes from work in mice, where absence of maternal BRG1 protein has been reported to cause developmental arrest at 2-cell stage and to impair transcription during ZGA [12].
While originally identified as global onset of zygotic transcription, MBT is now recognized as a continuous reorganization of the embryonic mRNA pool from early to late blastula stages, consisting of a major turnover of maternal mRNAs, coupled to broad, but not genome-wide initiation of transcription [43–45]. As shown in Fig 8A, three prototypic transcript classes can be operationally defined—i) maternal mRNAs, whose abundance declines; ii) transcripts with relatively constant abundance, and iii) mRNAs with increasing abundance through de novo transcription. We decided to characterize the three mRNA classes through global transcriptome analysis at immediate pre-MBT (NF8) and late Blastula stages (NF9) in X. tropicalis embryos (see S10 Fig, panels A and B for details). We then compared these data with the transcriptional changes observed in the BMO1 morphants at late blastula.
About 104 genes are expressed at the two time points (Fig 8B). After setting a threshold for transcript levels with an adjusted p-value ≤0.05, the abundance of 1357 transcripts changes between pre-MBT and late blastula. Decreasing levels characterize 761 transcripts as maternal mRNAs, whereas 596 transcripts classify as zygotic mRNA due to their increasing abundance (Fig 8B and S4 Table). By plotting the difference in mRNA levels between Brg1 morphants versus control morphants, we found that transcripts from the 596 genes activated at the MBT responded much stronger to the Brg1 knockdown than other transcripts (Fig 8C). A Gene Ontology search associated these zygotic transcripts with “Regulation of Transcription”, “RNA Metabolic Process”, and “Pattern Specification Process” (S10 Fig, panel C). When genes of the GO-term “Pattern specification process” were ranked in a heat map according to their magnitude of transcriptional activation, many of them were significantly downregulated by Brg1 protein knockdown (red asterisks, Fig 8D). A similar correlation between Brg1 dependence and transcriptional activation at MBT was found for the GO-term “Nervous System Development” (S10 Fig, panel D). In both cases, Brg1-dependent genes were enriched in the upper half of the heat maps, where genes with the highest fold-activation are located. The correlation was even more striking, when only the top-activated zygotic RNAs (≥5.65-fold increase after MBT; n = 324 genes) were considered—over 40% of these genes were Brg1 dependent (Fig 8E). Based on this analysis Brg1 protein is needed to amplify a transcriptional burst at MBT, which is necessary to initiate embryonic patterning.
To investigate potential developmental functions for Brg1 between MBT and neurula stages, we have carefully optimized the conditions for Brg1 protein depletion using three overlapping Morpholino oligonucleotides. These MOs differed significantly in their ability to block Brg1 protein synthesis, with the most upstream located targeting site of BMO3 having the least effect. The binding site of the Morpholino oligo used in earlier studies [46] starts 15 nucleotides upstream of BMO3 and does not overlap with BMO1/-2 target regions, presumably resulting in suboptimal targeting of brg1 mRNAs. We consider this the most likely explanation for the much earlier and more severe effects we report in this study. The demonstrated reduction of endogenous Brg1 protein levels in combination with the robust rescue of both morphological and molecular aspects of BMO1 morphants by human brg1 mRNA identifies the reported phenotype as a specific consequence of the Brg1 protein knockdown. Furthermore, both Brm and Iswi fail to compensate the Brg1 deficiency, suggesting specific remodeling events for Brg1-containing SWI/SNF complexes as the underlying molecular cause. These findings are in agreement with other reports [46–48].
Our results extend in a significant manner the current knowledge about the role of Brg1 during vertebrate embryogenesis. Transcriptome analysis from X. tropicalis late blastula stage (NF9) indicates that nearly 9% of transcripts (Fig 1D) were sensitive to Brg1 depletion. By comparison of the pre- and post-MBT transcriptomes we classified almost 600 transcripts to be de novo expressed at MBT, a number well in agreement with current estimates of X. tropicalis zygotic genome activation [47, 48]. The majority of these newly activated genes is Brg1-sensitive, detailing a significant impact for Brg1 on the first wave of zygotic gene expression of the embryo. In general we found that the gene response to Brg1 depletion is conserved between the two frog species. One clear exception is nr3/nodal3.1, which in the absence of Brg1 is downregulated in X. laevis, but upregulated in X. tropicalis (Figs 4, 5 and S2). Such a differential response could involve a species-specific modulation of the functional outcome of Brg1 activity on nr3 gene transcription, and/or reflect differences in the transcriptional regulation of these orthologs. Indeed, there is evidence for chromosomal rearrangements and gene amplifications at the nr3/nodal3.1 gene locus [49] which may have changed the regulation of nr3 transcription in the X. laevis L genome.
The short time span between MBT and late blastula (~3 hrs) implies that many, if not all misregulated genes are direct targets of Brg1-SWI/SNF. We expect this assumption to be valid for up- and downregulated gene cohorts alike, based on abundant evidence implicating Brg1 complexes with both gene activation and repression. Generally, the outcome of its action on a target gene is dictated by Brg1’s protein partners within SWI/SNF multiprotein complexes and by the gene-specific context. The many protein interactors of Brg1 and the versatile effector functions of SWI/SNF complexes pose a remarkably difficult challenge to investigate Brg1 acitivity on the mechanistic level in the embryo. Notably, Brg1-SWI/SNF complexes are found associated with both active and repressed chromatin elements [50] through interactions with DNA, histone modifications, and a large number of specific transcription factors (for detailed information on Brg1-SWI/SNF see reviews in [2, 4, 17]. In our study, we have not investigated genes that become upregulated/derepressed in BMO1 morphants. However, recent studies indicate that in human and mouse ESCs, BRG1-SWI/SNF complexes repress enhancer elements of lineage specification factors by modulating H3K27 acetylation or methylation levels [51, 52]. Based on this evidence, some of the genes, which are repressed by Brg1 in Xenopus before gastrulation, may also be important for embryonic development.
Highly informative for us were the GO-terms enriched in the downregulated gene cohort. They guided our analysis to establish Brg1 protein–and by inference corresponding SWI/SNF complexes–as regulator of early embryonic patterning. In evaluating the role of Brg1, one needs to distinguish carefully early and late phenotypes. The aberrant patterns of gene expression in BCNE, Nieuwkoop center and Spemann’s organizer are immediate consequences of the lack of Brg1 protein by temporal linkage with ZGA. Some phenotypes assessed late, like the axial defects characterizing the DMZ-injected BMO1 morphants, can still be attributed to direct effects, since the state of the underlying regulatory gene network is fixed at gastrulation. This has been demonstrated by classical experiments, in which the normal body plan of a tadpole is severely and irreversibly altered through treatments between the late one-cell to 32-cell stage, which modulate maternal Wnt-signaling activity [18, 53] The situation is different for late phenotypes that arise from defective cell differentiation over time, such as the malformed brains in embryos transplanted with a morphant BCNE (Fig 6) or injected dorso-animally with BMO1 (Fig 7). Since Brg1 is continuously expressed in the neuroectoderm and neural crest, these phenotypes may indicate a later requirement for Brg1 in brain and retina development [50].
What are the main developmental functions of Brg1 and how are they implemented in the process of embryogenesis? While previous studies implicated Brg1 in neuroectoderm differentiation [15–17, 54], our results demonstrate that this ATPase is required for neural plate formation. Programming the neural ground state involves a conserved gene regulatory network [24, 54]. Essential members include FoxD4I1, Zic1, Zic3 and Iroquois2, which are all significantly downregulated in Brg1 depleted embryos at late blastula (Figs 1 and S2, S2 Table). The installation of this network requires inhibition of Bmp signals in the prospective neural plate through secreted antagonists, three of which (chordin, noggin, nr3,) are expressed first in the BCNE region and subsequently in the organizer under the control of maternal Wnt/β-Catenin signaling [31, 55]. For these genes, our data define Brg1 as a coactivator of Wnt signaling, which helps install DGEP as first step to the formation of dorso-anterior tissues. Whether Brg1 could be even sufficient to induce DGEP alone, as suggested by secondary axis formation and ectopic chordin induction after ventral overexpression of Xenopus and human Brg1 proteins, is not clear at the moment. The ventral side of the embryo contains residual amounts of maternal Wnt11 protein, with which Brg1 might cooperate [20, 56].
In dorso-vegetal cells, the combination of maternal Wnt signaling plus high Nodal signaling establishes the Nieuwkoop Center [57]. This mode of regulation separates it from the BCNE, even though it overlaps with the BCNE region by cell lineage [39, 40]. The multifunctional BMP inhibitor cerberus, expressed in the Nieuwkoop Center is also BRG1-sensitive, as shown by dorso-vegetal BMO1 injections. Taken together, these results have identified three different embryonic territories to require Brg1 activity, and prevent a simple assignment of Brg1 function to either ventralizing or dorsalizing gene expression programs.
Interestingly not all Wnt targets depend on coactivation by Brg1-SWI/SNF, for instance the transcription factor gene sia1 (Fig 4D). This helps to explain the peculiar observation that in BMO1 morphants chordin mRNA is selectively reduced in the BCNE, but is unaffected in the organizer (compare Figs 4 and S7). According to current models, maternal Wnt signaling induces transiently sia1 expression (Brg1-independent; Fig 4D), which in turn activates chordin transcription (Brg1-sensitive; Figs 1, 4, S6 and 7) at the blastula stage (see Carnac et al., Development 1996); subsequently, chordin transcription is maintained through Nodal signaling (Brg1-independent; see Figs 5 and S7) in the gastrula organizer [27]. This model would predict that sia1 maintains chordin in a Nodal-responsive state without transcriptional activation. A poised state is also a common theme for Wnt target genes in Xenopus, which are frequently bound by β-Catenin without eliciting a transcriptional response [58]. Transcriptional activation of Wnt targets has been proposed to occur through context-specific mechanisms, downstream of β-Catenin binding to chromatin. One such context could be recruitment of a Brg1-SWI/SNF chromatin remodeler. It should be highly informative to identify mechanisms, by which Brg1-SWI/SNF recognizes its targets at the blastula stage.
Gene expression domains in the gastrula organizer are generally less affected in BMO1 morphants than in the BCNE. Nevertheless, our data indicates a third function for Brg1 in activating transcription of the core BMP synexpression group (bmp4, vent1 and vent2; Fig 5, S2 Table), which specifies ventro-posterior tissues [59]. Interestingly, while their mRNA levels are globally reduced in radial BMO1 morphants at late blastula (S2 Table), the overlapping vent1/vent2 expression domains are expanding into the organizer field. Here, they become coexpressed with DGEP genes like gsc and otx2 (which specify anterior position) and the bmp antagonists chordin and nr3 (Figs 5 and S7). This highly unusual pattern of genes in the organizer had also been generated by simultaneous knockdown of three BMP antagonists (chordin, noggin, follistatin) [36]. It indicates a significant weakening of the organizer function, which is reflected in our data by reduced expression of nr3 and otx2 within the organizer, and of the myogenic bHLH transcription factors myoD and myf5 in non-organizer mesoderm (Figs 5 and S7). That the reduced bmp4/vent1/vent2 expression indeed contributes to the aberrant body plan, is apparent in radial morphants injected with 20ng BMO1 (S4 Fig). Although their blastomeres received the same amount of BMO1 oligo as embryos, which were injected only in DMZ or VMZ (i.e. 5ng/blastomere), the radial morphants were arrested at gastrulation and thus morphologically much stronger affected than DMZ or VMZ morphants (compare S4B Fig with Fig 2). It is plausible to assume that these perturbations in the organizer and non-organizer activities, together with reduced cerberus transcription in the Nieuwkoop Center are responsible for the massive loss of dorso-anterior structures seen in DMZ-injected BMO1 morphants (Figs 2 and 3) and for the posteriorized character of Brg1 depleted embryos, arrested permanently at the gastrula stage (S4 Fig). The invasion of vent1/vent2 gene expression into the organizer field suggests a basal failure in the cell determination process, which normally prevents activation of non-compatible gene expression programs in cells [60]. Indeed, non-compatible gene expression could mount a substantial problem given that almost 900 genes are misregulated in Brg1 morphants.
In summary, we have identified at least two independent functions for Brg1-SWI/SNF, which are essential for Xenopus embryonic patterning, namely being i) selective coactivator of maternal Wnt signaling on the prospective dorsal side of the embryo and ii) coactivator of the core bmp synexpression group on the prospective ventral side. However, more than 800 genes are misregulated in Brg1 depleted embryos at the blastula/gastrula transition. Is there a common denominator to explain Brg1’s impact on development? By quantitative filtering of global mRNA fluctuations at MBT, we have shown that Brg1 is predominantly required for genes with the highest burst of transcriptional activity. Mechanistically, Brg1-SWI/SNF could be involved to catalyze transitions from transcriptional silent to active chromatin states at promoters or facilitate long range interaction between distal enhancers and promoters [61], which in general become engaged at the blastula/gastrula transition [62]. Since many of the bursting genes are key developmental regulators, this may put BRG1 in a key position to raise their expression above threshold levels in preparation for the embryonic patterning process. Notably, both mathematical models and experimental evidence have detailed an enormous self-regulatory capacity within embryonic fields [63–65]. Why BRG1-depleted embryos cannot compensate a quantitatively insufficient ZGA and fail to restore axis formation through self-regulation, constitutes a key question for the future.
We have shown that BRG1 protein is essential for early Xenopus development. BRG1 is involved in de novo activation of transcription at the Midblastula transition and is needed to achieve the transcriptional amplitude of genes with the highest-fold activation. Among these bursting genes are many regulators of embryonic axis formation. Targeted depletion of BRG1 protein levels results in the specific downregulation of key genes of the BCNE Center (chordin, noggin), the Nieuwkoop Center (hhex, cer) and the bmp-controlled ventral signaling territory (vent1, vent2). We propose that Brg1 fulfills a systemic function for late blastula stage transcription in preparation of embryonic pattern formation.
Animal work has been conducted in accordance with Deutsches Tierschutzgesetz; experimental use of Xenopus embryos has been licensed by the Government of Oberbayern (AZ: 55.2.1.54–2532.6-7-12).
The ORF of X. laevis brg1 cDNA was generated by PCR from overlapping ESTs (Genbank acc. Nrs. AW766934, BG234591, BQ7288178) and subcloned into pCS2+. The full-length cDNA was verified by sequencing and deposited into GenBank (AY762376). For testing morpholino targeting efficiencies, the cDNA region from -77 to +617 of X. laevis brg1 (“BISH”) was fused in frame to the luciferase ORF in a gateway-compatible pCS2+ vector. All primer sequences are provided in the supplemental data section, S5 Table, part a. Open reading frames of human brg1, brm and X. laevis iswi (kind gift from Anthony Imbalzano and Paul Wade) were sub-cloned into pCS2+ for in vitro transcription. Capped mRNA for microinjection was synthesized as described [66].
Three antisense morpholino oligonucleotides against the translational start site of Brg1 mRNAs were purchased from GeneTools: All three are fully complimentary to transcripts from both X. laevis homeologs (NM_001086740.1 and BG554361); BMO1 and BMO2 also match perfectly the mRNA sequence of the S. tropicalis homolog (BG554361). BMO1: 5’- CCATTGGAGGGTCTGGGGTGGACAT-3’; BMO2: 5’-CAGGGAGAAGATCCAGTCACTGCTA-‘3; BMO3: 5’-GACATCACTGCAGGGAGAAGATCCA-‘3. The unrelated standard control Morpholino served as control for specificity.
Morpholino targeting efficiencies was determined in vivo with a Brg1-luciferase fusion mRNA. Xenopus laevis embryos were radially injected at the 2 cell stage with individual morpholino oligonucleotides (60ng/embryo). At the 8 cell stage, the four animal blastomeres were superinjected with synthetic BISH-luciferase mRNA (25pg/embryo). The embryos were cultivated until gastrulation (NF11) and luciferase activity was measured with Dual-Luciferase® Reporter Assay System (Promega). Samples consisted of cleared protein lysates from 5 pooled embryos per condition.
X. laevis and X. tropicalis eggs were collected, in vitro fertilized, microinjected and cultivated following standard procedures. Embryos were staged according to Nieuwkoop and Faber (1967). Radial injections were performed at the 2–4 cell stage, targeted injections were performed either at the 4 cell, 8 cell or 16 cell stage. For lineage tracing they were either injected with Alexa Fluor-488 Dextran, Alexa Fluor-594 Dextran (Invitrogen) or with 25-100 pg/blastomere of either nuclear lacZ or eGFP mRNA. Tissue transplantations were carried out in 0.8x MBS + Gentamycin in agarose-coated dishes. The transplant was kept in place with a cover slip for one hour, after which the transplanted embryo was transferred to a new dish in 0.1x MBS + gentamycin.
Whole-mount RNA in situ hybridizations were performed as described [67]. Embryos were photographed with a Leica M205FA stereomicroscope. For immunocytochemistry anti-active Caspase3 antibody (1:20000, Promega) and anti-rabbit alkaline phosphatase-conjugated (1:1000, Chemicon) secondary antibody was used.
For the production of xBrg1 specific monoclonal antibodies, N-terminal domain (amino acid 202–282) was cloned into pGEX4T3 expression vector (Amersham), expressed in E. Coli and purified to immunize rats. For quantitative measurement of Brg1 knockdown 15 embryos or eggs of each condition were collected and lysed in 75 μl NOP buffer [68] and centrifuged for 20 min at 14000 rpm to remove yolk plates. The supernatant was mixed with Roti®-Load 1 (Roth) and loaded onto an 8% SDS-PAGE. The separated proteins were blotted on nitrocellulose membrane and blocked for minimum 1 h at RT in 5% milk in PBSw. The membrane was incubated over night at 4°C with anti-Brg1 mab 3F1 (1:3) and as loading control anti α-tubulin (1:8000, Sigma). As secondary antibody the LiCor α-rat 800 and α-mouse 700 (1:10000, respectively) was used. The membrane was developed using the LiCOR system and the intensities were measured and quantified against the loading control.
Total RNA of 10 embryos was extracted using Trizol (Ambion) and phenol/chloroform. The RNA was precipitated with 70% Isopropanol and cleaned using the RNeasy Cleanup Kit (Qiagen) including DNAseI-on-column digestion. For qPCR analysis 1 μg of total RNA was transcribed with the DyNAmo CDNA Sythesis Kit (Bioenzym). For qPCR 5–20 ng cDNA was mixed with the Fast SYBR Green Master mix (Applied Biosystems) and amplified with a Lightcycler (Roche). Primer sequences are given in S5 Table, part b.
For comparative MicroArray analysis X. tropicalis embryos were injected radially at the 2–4 cell stage with 30ng BMO1 or 60ng CoMO and cultivated until late Blastula stage. Per condition, 10 embryos were collected and RNA was extracted as described above. For the pre/postMBT Microarray, wildtype X. tropicalis embryos were collected. Ten embryos were pooled per sample. After fertilization we collected one sample in the 4cell stage as negative GS17 control. For the preMBT time-point we collected embryos from around the ~1000 cell stage and then every 20–25 minutes for approximately 60-100min. We choose 20–25 min breaks depending on how long the embryos need in average for the first cell divisions. 120–150 minutes after collecting the last preMBT sample we started again to collect every 20–25 min samples until we observed dorsal lip formation. For all samples we extracted RNA the way it was described before. In order to find the preMBT sample closest to the MBT we performed qPCR analysis with MBT-marker gs17. For the preMBT time-point we took the last sample without gs17 expression. As postMBT we choose the sample ~40min before dorsal lip formation, in accordance to the developmental age, at which the comparative Microarray analysis was performed.
The quality of the extracted RNA was controlled for both experimental setups with the Bioanalyzer and handed to the “Facility of Functional Genomics” at the Gene Center, Munich for microarray performance on an Affymetrix Xenopus tropicalis genome Array.
Microarray preprocessing was conducted separately for the two experimental sets (Brg1 knockdown and MBT) using R/Bioconductor (www.bioconductor.org). If not indicated otherwise, we used standard parameters in all functions calls. Expression values were calculated using ‘gcrma’. Probe sets were kept for differential expression analysis if there were more ‘present’ calls (calculated using ‘mas5calls’) in one of the treatment groups than non-‘present’ calls, if their expression level variance was higher than zero across all arrays and if the probe set had an Entrez identifier annotation according to the Entrez database with a date stamp of 2011-Mar16. One gene to many probe set relationships were resolved by retaining only the probe set with the highest interquartile range across all arrays. Differential expression statistics were obtained using a linear model (library ‘limma’). A significant response was defined if the adjusted p-value was smaller than 0.05.
For all embryonic quantitative analysis (morphological phenotype, WMISH, qRT/PCR) SEM are displayed and the statistical analysis was performed using two-tailed, Paired Student’s t-test. For transcriptome analysis see microarray section.
|
10.1371/journal.pcbi.1000922 | The Distinct Conformational Dynamics of K-Ras and H-Ras A59G | Ras proteins regulate signaling cascades crucial for cell proliferation and differentiation by switching between GTP- and GDP-bound conformations. Distinct Ras isoforms have unique physiological functions with individual isoforms associated with different cancers and developmental diseases. Given the small structural differences among isoforms and mutants, it is currently unclear how these functional differences and aberrant properties arise. Here we investigate whether the subtle differences among isoforms and mutants are associated with detectable dynamical differences. Extensive molecular dynamics simulations reveal that wild-type K-Ras and mutant H-Ras A59G are intrinsically more dynamic than wild-type H-Ras. The crucial switch 1 and switch 2 regions along with loop 3, helix 3, and loop 7 contribute to this enhanced flexibility. Removing the gamma-phosphate of the bound GTP from the structure of A59G led to a spontaneous GTP-to-GDP conformational transition in a 20-ns unbiased simulation. The switch 1 and 2 regions exhibit enhanced flexibility and correlated motion when compared to non-transitioning wild-type H-Ras over a similar timeframe. Correlated motions between loop 3 and helix 5 of wild-type H-Ras are absent in the mutant A59G reflecting the enhanced dynamics of the loop 3 region. Taken together with earlier findings, these results suggest the existence of a lower energetic barrier between GTP and GDP states of the mutant. Molecular dynamics simulations combined with principal component analysis of available Ras crystallographic structures can be used to discriminate ligand- and sequence-based dynamic perturbations with potential functional implications. Furthermore, the identification of specific conformations associated with distinct Ras isoforms and mutants provides useful information for efforts that attempt to selectively interfere with the aberrant functions of these species.
| The proto-oncogene Ras mediates signaling pathways controlling cell proliferation and development by cycling between active and inactive conformational states. Mutations that affect the ability to switch between states are associated with over 25% of human tumors. However, despite much effort, details of how these mutations affect the fidelity of activating conformational transitions remain unclear. Here we employ extensive molecular dynamics simulations combined with principal component analysis to investigate whether the subtle differences among functionally distinct isoforms and oncogenic mutants are associated with detectable dynamical differences. Our results reveal that wild-type K-Ras, the most prevalent isoform in a number of cancers, and mutant H-Ras A59G are intrinsically more dynamic than wild-type H-Ras. Furthermore, we have observed the first spontaneous GTP-to-GDP transition of H-Ras A59G during unbiased molecular dynamics simulation. These results indicate that key changes in sequence can lead to different dynamic properties that may be relevant for the unique physiological and aberrant functions of Ras isoforms and mutants. Furthermore, the current results shed further light on the conformational transition mechanism of this important molecular switch.
| Ras proteins couple cell-surface receptors to intracellular signaling cascades involved in cell proliferation, differentiation and development. Signal propagation through Ras is mediated by a regulated GTPase cycle that leads to active and inactive conformations with distinct affinity for downstream effectors. Regulatory proteins including guanine nucleotide exchange factors (GEFs) and GTPase-activating proteins (GAPs) stimulate the intrinsically slow GTPase cycle promoting proper signal flow. Ras mutants with an impaired GTPase activity that are insensitive to the action of GAPs and GEFs result in prolonged downstream signaling associated with oncogenic cell growth in diverse human cancers and leukemia [1], [2].
Ras genes encode multiple isoforms of which H-, N-, and K-Ras are the most abundant. K-Ras can be found as two splice variants termed K-Ras4A and K-Ras4B. Although these isoforms share a high degree of similarity (over 90% sequence identity), their physiological functions are not necessarily equivalent [3]–[9]. K-Ras4B is essential for normal mouse embryogenesis and development, whereas H-, N-, and K-Ras4A are dispensable when K-Ras4B is present [4], [5]. K-Ras plays a unique role in cardiovascular homeostasis as mutant mice with their K-Ras gene modified to encode H-Ras, exhibit dilated cardiomyopathy associated with arterial hypertension[10]. Furthermore, mutations in K-Ras occur most commonly in human cancers and developmental diseases, including pancreatic, colorectal, lung, cervical and hematological cancer, Noonan syndrome, and Cardio-facio-cutaneous syndrome [11]–[13].
The unique functions of Ras isoforms are mediated by their preferences for different binding partners [14]. Thus an understanding of functional fidelity requires a detailed structural and dynamical characterization of each isoform. The recently solved atomic structure of K-Ras4B, hereafter referred as K-Ras, revealed a high degree of similarity to previously solved H-Ras structures (94% sequence identity and 1.03 Å Cα RMSD over the 166 residues of the catalytic domain) (Fig 1). Given these small differences it is currently unclear how distinct binding preferences might arise.
Small changes in sequence can lead to different dynamic properties, which are manifested as only subtle changes in the average structure observed by x-ray crystallography. Recently we reported the observation of spontaneous nucleotide-dependent transition during unbiased molecular dynamics (MD) simulation of the oncogenically active H-Ras G12V variant [15]. This study suggested the existence of a lower thermally accessible energetic barrier between inactive and active states of this variant that renders it prone to adopt an active conformational state. This dynamic effect is not apparent from comparing the crystallographic structures of H-Ras G12V to wild-type H-Ras. Since such a single residue substitution on H-Ras could have dramatic effects on its dynamics, we hypothesize that a few residue substitutions among isoforms might influence their structural dynamics and hence their preferences and affinities to ligands and binding partners.
Previous classical and accelerated MD simulations of H-Ras have successfully characterized its dynamic features and proposed a reaction path for the ligand-associated conformational changes [16]–[18]. Furthermore, classical MD simulations of homology-built K-and N-Ras in the nucleotide-free state have suggested enhanced flexibility of K-Ras relative to other isoforms [15]. However, it is important to determine if this conclusion holds also in the presence of the nucleotide and is not affected by the initial structure. Since downstream effectors bind to active, GTP-bound Ras, the first part of our present study has been focused on investigating the dynamics of GTP-bound Ras isoforms.
Here we employ multi-copy MD simulations to investigate whether the subtle differences between H- and K-Ras isoforms are associated with detectable dynamical differences that might have potential functional consequences. We first conducted an expanded bioinformatic analysis of available Ras crystallographic structures and found that K-Ras has similar conformational features to the H-Ras A59G mutant. We further performed MD simulations on active, GTP-bound wild-type K-Ras, H-Ras, and H-Ras A59G. Indeed, we observed that both the wild-type K-Ras and H-Ras A59G variant are similarly more dynamic than wild-type H-Ras. However, wild-type K-Ras also has dynamic features that are similar to those of wild-type H-Ras, hence wild-type functions are preserved. The different dynamic features between wild-type K- and H-Ras may also provide clues on distinct preferences for binding proteins and hence subsequent downstream signaling functions. Another major contribution of this study is the first report on the spontaneous GTP-to-GDP transition of the H-Ras A59G variant. The H-Ras A59G variant crystallized with a GTP-analog, when simulated with a GDP, is capable of spontaneously adopting GDP-bound conformations within 20 ns. The atomic details of this transition are important for understanding Ras signaling and function.
We first examined the structure of K-Ras in relation to other available Ras experimental structures. This analysis revealed a similarity to a A59G H-Ras variant. We then performed multiple MD simulations on wild-type K-, H-Ras, and H-Ras A59G. The Ras isoforms exhibited differences in their active, GTP-bound conformational dynamics. H-Ras A59G variant simulated with GDP was able to achieve a spontaneous GTP-to-GDP transition.
To investigate the relationship of K-Ras to other available Ras structures, we compiled a crystallographic ensemble comprising 51 chains from the 47 unique Ras structures in the RCSB PDB [19], [20]. Principal component analysis (PCA) was used to characterize inter-conformer relationships. Despite the inclusion of ten new chains from six recently-solved structures, we obtained a similar distribution of conformers along the dominant principal components (PCs) as obtained in an earlier study [15]. This indicates the robustness of the distribution to the inclusion of new structures and the suitability of the method for describing new inter-conformer relationships. Two major clusters are again evident along PC1 corresponding to distinct GTP and GDP bound conformations (Fig 2), confirming the previous observation that different chemical molecules at the active sites are associated with different global Ras conformations [15]. Note that some GTP/GTP-analog-bound structures were not situated within the main GTP-cluster. These structures have a mutation at the P-loop or switch regions. For example, H-Ras G12V, H-Ras A59G, and H-Ras Y32C (PDB: 2Q21, 1LF0, and 2CL0 respectively) did not reside in the GTP-cluster. Interestingly, the wild-type K-Ras conformer also resides outside the main GTP and GDP clusters in close proximity to the crystal conformer of H-Ras A59G mutant. Both structures are indeed similar in terms of Cα RMSD (0.68 Å).
While the first principal component revealed distinct clusters of Ras conformations corresponding to different bound-nucleotides (Fig 2), the third principal component showed distinct clusters of Ras conformations corresponding to mutant states (Fig 2). Ras crystal conformers with mutations at critical residues 32, 59, 61, and 118, affecting GTP hydrolysis, emerged far from the wild-type H-Ras conformers along the third principal component. Two H-Ras mutant conformers, G12P and G60A were found to reside within the majority wild-type cluster. Interestingly, these mutations are non-oncogenic. For example H-Ras G12P is known to possess wild-type like intrinsic GTP hydrolysis and GDP dissociation functions [21]. In summary, the current principal component projections based on the Ras crystallographic ensemble are robust to the inclusion of new structures and facilitate the discrimination of key ligand- and sequence-based structural perturbations.
To further probe the conformational dynamics of the different Ras isoforms we performed multi-copy molecular dynamics simulations on the following systems: wild-type K-Ras, wild-type H-Ras, and H-Ras A59G. Although the starting wild-type K-Ras and the wild-type H-Ras structures are similar in overall conformation (Cα RMSD is 1.03 Å), the active site of wild-type K-Ras is more similar to that of H-Ras A59G (Cα RMSD is 1.16 Å) than wild-type H-Ras (Cα RMSD is 1.74 Å) (Fig S1).
Comparison of the wild-type K-Ras and H-Ras trajectories in the GTP-bound form revealed variations in their dynamic behaviors, consistent with previous MD simulations of homology-built K-Ras and wild-type H-Ras in the nucleotide-free state [15]. Interestingly, wild-type K-Ras is similar to H-Ras A59G in sampling a wider region of conformational space proximal to the H-Ras G12V variant crystal conformer (Fig 3). Previous MD simulations of H-Ras G12V also displayed enhanced sampling when compared to wild-type H-Ras [15]. In terms of RMSD, the MD conformers of wild-type K-Ras and H-Ras A59G appear to be more similar to each other than to the MD conformers of wild-type H-Ras (Fig S2). The same trend is apparent for the canonical switches. For example, across the three sets of MD trajectories of the wild-type H-Ras, switch 1 RMSD values are generally higher than those of switch 2 (Fig S3) whereas the opposite is true for K-Ras and H-Ras A59G. Comparing RMSF values from the respective GTP-bound simulations indicates that wild-type K-Ras and H-Ras A59G are indeed more dynamic than wild-type H-Ras (Fig 4). These differences are significant as indicated by paired student's t tests (P<0.01). Particularly noteworthy are residues at the crucial switch 1, switch 2, loop3, α3 helix, and loop 7 regions, all of which are more flexible in K-Ras and A59G than in wild-type H-Ras.
To identify regions undergoing correlated motions we analyzed dynamic cross-correlation maps for the three sets of trajectories. The correlated motion between loop 3 and α5 helix is absent in H-Ras A59G but present in the wild-type K- and H-Ras (Fig 5). This suggests that the substitution at position 59 in switch 2 affects the correlated motion of structural elements that are sequentially and spatially distant. The loop 3 and the C-terminal of α5 helix are parts of switch 3 [22] and correlated motion between loop 3 and α5 was reported in previous accelerated MD simulations of wild-type H-Ras [23]. On the other hand, wild-type H-Ras and H-Ras A59G exhibit a similar pattern of correlated motions between α2 and α3-loop 7 (Fig 5). The α3-loop 7 region has recently been reported to function as an allosteric site regulating switch 2 ordering [24].
Residue Q61 of switch 2 is essential for GTP hydrolysis with its side chain involved in orienting and activating a nucleophilic water molecule [25]. To monitor the conformation of Q61, we tracked the distances between the γ-phosphate and Q61 side chain atom NE2. Their distances are greater in wild-type K-Ras than in wild-type H-Ras, although not as large as in H-Ras A59G simulations (Fig 6). Such displacement of Q61 toward R68 and away from the catalytic site likely contributes to the much slower intrinsic GTP hydrolysis of H-Ras A59G when compared to wild-type H-Ras [26].
Y64 is more distant to the γ-phosphate in both wild-type K-Ras and H-Ras A59G than in wild-type H-Ras (Fig 6). In simulations of wild-type H-Ras, the side chain of Y64 was persistently oriented and positioned close to the γ-phosphate of GTP (distance between Y64 OH atom and γ-phosphate ranged from 9.0 to 13.9 Å, with an average value of 11.7 Å). In wild-type K-Ras, the distance varied between 9.2 and 25.7 Å, with an average value of 22.4 Å. In A59G, this distance ranged from 19.8 to 25.4 Å, and the average was 22.1 Å. GTP hydrolysis can only occur in conformations with the Y64 side chain being close to the γ-phosphate. Though the average distance of Y64 side chain and γ-phosphate in wild-type K-Ras is as long as in H-Ras A59G, wild-type K-Ras can also adopt conformations in which their Y64:OH and γ-phosphate distance is as short as that of wild-type H-Ras. Fig 7 depicts the change in the orientation of K-Ras Y64, the shortening of Y64:OH - γ-phosphate distance, and the motion of α2 helix toward the nucleotide-binding site.
The crystallographic structure of H-Ras A59G, on which the current simulations are based, was proposed to represent the conformation of wild-type H-Ras following β/γ-phosphate bond breakage but before γ-phosphate dissociation [26]. This mutant was also shown to have an impaired intrinsic GTP hydrolysis activity [26]. When we replaced the original GTP-analog with GDP and performed unbiased 20-ns MD simulation, we observed a spontaneous GTP-to-GDP transition. This spontaneous GTP-to-GDP transition complements our previous characterization of the spontaneous GDP-to-GTP transition observed in the H-Ras G12V variant [15]. The GTP-to-GDP transition has been previously analyzed in some studies [18], [27], [28] that employed biased calculations, such as targeted MD, which introduces external constraints to drive the transition. Different solvation conditions of explicit [18] or implicit solvent [27], were also used. Our study provided new insight on the GTP-to-GDP transition based on unbiased MD simulation in explicit solvent, which better mimics natural physiological conditions [29].
We note that only one of our three multi-copy simulations managed to reach the GDP-cluster of the Ras crystallographic ensemble within 20-ns. Comparing the residue-wise flexibility from this simulation to the other two sets, indicates that the GTP-to-GDP transition involves higher RMSF at the loop 2, loop 4, and α2 regions, as well as lower RMSF at the loop 3 region (Fig 8). Below we focus on this transitioning in order to obtain atomic descriptions of the conformational conversion process.
When the MD-derived conformers are projected onto the first two principal components (PC) obtained from analyzing the Ras crystallographic ensemble, the GDP-bound H-Ras A59G MD conformers sample the GDP cluster of crystallographic conformers (PC 1: 15 to 20, PC 2: −5 to 0) (Fig 9A). Over time, the H-Ras A59G MD conformers evolve to resemble the GDP-bound more than the GTP-bound H-Ras A59G crystal structure (Fig 9B). Although both the GTP-to-GDP or GDP-to-GTP conformational changes involve rearrangements of the canonical switch regions [26] and partial unfolding and re-folding of helix 2, in each case multiple pathways are possible [27]. In the simulation, in the first 4 ns switch 2 has slightly higher RMSD values than switch 1 (Fig 9C) but the relative RMSD of the switches remains stable until 11.2 ns. Following this, the switch 2 RMSD increases significantly before the switch 1 RMSD catches up surpassing the switch 2 RMSD at about 13.8 ns. After a decrease in RMSD (∼16–18ns), the switches evolve again as the N-terminal of α2 helix unwinds (Fig 10A, Video S1). The unwinding of the α2 helical turn (residues 66–69) was also reported in a previous biased MD study of GTP-to-GDP conformational transitions [18]. Interestingly, despite some differences in the sequence of events, these results are similar to those reported based on the computation of minimum energy paths (MEPs) [27].
The side chain of Y32 lies across the nucleotide binding pocket in GTP-bound conformations but is displaced away from the binding pocket in GDP-bound H-Ras A59G and wild-type H-Ras crystal structures [21], [26], [30]. After 11.5 ns of simulation, Y32 moves away from nucleotide and comes closer to Y64 from 14 ns onwards (Fig 10B). The role of Y32 was also highlighted in the Y32F mutant with dysfunctional GTP hydrolysis [31]. Preceding the unwinding of helix 2 in switch 2, the Q61 of switch 2 orientates away from the switch 1 (Fig 10C). However, no significant change was seen in the distance between G60 and nucleotide preceding the unwinding of α2 (Fig 10C) as reported in [18]. This is to be expected as we commenced simulation form the A59G crystal structure conformation in which its G60, a possible hinge [32], had already moved away from the nucleotide to occupy the space made available by the A59G substitution. Next, the Y71 and R68 of switch 2 form close contacts with E37 of switch 1, stabilizing the protein during the α2 unwinding. The unwinding of α2 re-orientates switch 2 and brings Y64 closer to E37 of switch 1 (Fig 10C).
The MD conformers at different time frames also indicated interesting differences in their torsional angles with respect to the GTP-analog-bound versus the GDP-bound H-Ras A59G crystal conformers (Fig S4). The torsional angle differences were emphasized by the residues of loop 2 and loop 4 form switch 1 and 2 regions respectively. These regions exhibit correlated motions (Fig S5) and enhanced flexibility (Fig 8B) highlighting their roles in binding nucleotide and facilitating the GTP-to-GDP transition.
Combining the current results with previous reports [15], [18], [23], [27] on the reaction paths for the transition between the active and inactive states of Ras, we arrive at the following conclusions. (i) Conformational changes accompanying GTP hydrolysis (GTP-to-GDP) and nucleotide exchange (GDP-to-GTP) involve common intermediates that are represented by the A59G structure; (ii) there is a barrier to reach this intermediate upon GTP hydrolysis, which is consistent with the 24.3 kcal/mol activation energy estimated by the minimum energy paths method for the GTP-to-GDP transition of wild-type H-Ras [27]; (iii) that an unbiased simulation of H-Ras A59G led to a transition from the GTP to the GDP state may suggest that this point mutation lowers the barrier between the intermediate and the GDP state; and (iv) once the large barrier between the GTP state and the H-Ras A59G-like intermediate is crossed, the remaining (multiple) barriers to the GDP-bound form can be crossed relatively easily. Finally, it is noteworthy that the PCA analysis enabled us to effectively cluster the diverse Ras crystal structures into two major clusters interspersed by intermediates that are predominantly populated by mutants. Some of these mutant structures are susceptible to nucleotide-associated transitions.
Oncogenic mutations may interfere with the intrinsic ability of Ras to hydrolyze GTP or lead to insensitivity to the action of GAPs. The loss or deregulation of intrinsic GTP hydrolysis seems to be sufficient for causing diseases [12], but GAP-insensitivity is still rescuable in some variants. For example, the H-Ras G12P and G12A variants are GAP-insensitive, but their intrinsic GTPase function and their ability to release GDP only modestly differ from those of wild-type [21], hence these mutants do not transform cells. The intrinsic GTP hydrolysis activity appears to be biologically sufficient to stop continuous downstream signaling [12], [33]. Therefore, structural changes that have the potential to affect the intrinsic GTPase function of Ras may also modulate its aberrant functions. Thus, the enhanced dynamics of K-Ras (and G59A H-ras) in functionally crucial regions may have an important implication for its normal and tumorigenic functions, which may partly explain the more frequent discovery of K-ras in cancer cells than any other Ras isoforms [34].
We speculate that the observed differences in dynamics may be related to the measured differences in the intrinsic GTPase activity of the three proteins. The intrinsic GTP hydrolysis rate of K-Ras (1.2×10−4 s−1 [33]) is roughly half the rate of the wild-type H-Ras (2.7×10−4 s−1 [24]) whereas that of H-Ras A59G (0.35×10−4 s−1 [26]) about eight times smaller. This may suggest that in the absence of GAP, K-Ras and H-Ras A59G spend more time in the GTP-bound conformational state than the wild-type H-Ras. The enhanced dynamics of K-Ras and H-Ras A59G may make the GTP-bound state entropically favored. This is consistent with a previous MD simulation on H-Ras G12V which revealed that the continuously changing active site makes GTP hydrolysis difficult [35]; the resulting elevation in the population of GTP-bound conformations will then lead to increased signaling through this oncogenic variant. However, caution is required in interpreting the existing rate constants as the measured intrinsic GTP hydrolysis rate is dependent on the GTP concentration present in the assay [25]. We thus believe that the current data would be even more useful if used in conjunction with GTP hydrolysis rates measured under identical experimental conditions.
We have performed multiple unbiased MD simulation on wild-type K-Ras, H-Ras A59G and wild-type H-Ras. Results from these simulations support the observation that the active, GTP-bound wild-type K-Ras and H-Ras A59G variant are more dynamic than wild-type H-Ras. We observed spontaneous GTP-to-GDP transition during an unbiased MD simulation of H-Ras A59G. The approach of multivariate clustering of crystal structure conformations to reveal differences due to ligands and sequences highlights intermediate conformers, such as the H-Ras G12V [15] and A59G in this study, that facilitate faster sampling of large scale conformational transitions. We speculate that Ras variants which fall outside the major GTP- or GDP-clusters may be intrinsically more susceptible to transition than wild-type H-ras. Further systematic investigation of the structural and dynamic properties of Ras isoforms and mutants is likely to be informative for drug development, particularly drugs which selectively target distinct structural conformations associated with specific Ras isoforms or mutations. Structure based drug design efforts in this direction are currently underway.
The Bio3D package [36] was used to retrieve homologous structures to K-Ras (PDB: 2PMX) from the PDB, perform principal component analysis (PCA) and additional trajectory analysis as described in [15], [23]. Core positions were first obtained from analysis of the crystallographic ensemble. Iterated rounds of structural superposition were used to identify the most structurally invariant region of the Ras structure. This procedure, implemented in the Bio3D package, entailed excluding those residues with the largest positional differences, before each round of superposition, until only the invariant “core” residues remained [36], [37]. The structurally invariant core was used as the reference frame for structural alignment of both crystal and simulation structures. Next, the Cartesian coordinates of aligned Cα atoms were used to define the elements of a covariance matrix. The covariance matrix was then diagonalized to derive principal components with their associated variances. After PCA, Ras crystal structures and MD conformers were projected on the subspace defined by principal components with the largest variances.
Systems for molecular dynamics simulations were prepared from high-resolution crystal structures of wild-type K-Ras, wild-type H-Ras, and H-Ras A59G variant (PDB: 2PMX, 1QRA, and 1LF0 respectively). Each system was simulated with Mg2+GDP and Mg2+GTP. All simulations were performed with the AMBER 10 package [38]. The LEaP module was used for model construction, adding missing atoms to initial coordinates, and including parameters for guanine nucleotides [39] and Mg2+ ion. The protonation states for all titratable residues were determined using PDB2PQR [40]. The systems were neutralized using charge-neutralizing counter ions at pH 7, followed by explicit solvation with TIP3P water molecules with the buffering distance set to 10 Å. Energy minimization with decreasing constraints on the heavy atoms' positions, constant volume heating (to 300 K) was carried out for over 10 ps, followed by constant temperature (300 K) and constant pressure (1 atm) equilibration for additional 200 ps. The production simulation, using ff99SB force field [41], was conducted with time step of 2 fs, using the isobaric-isothermal ensemble at 300 K, 1 atm, and long-range non-bonded interactions with 10 Å atom-based cutoff. Electrostatic interactions were evaluated using the Particle-Mesh Ewald sum [42]. The SHAKE algorithm was used to constrain all covalent bonds involving hydrogen atoms. To enhance sampling and improve the statistical accuracy of the simulations, three independent MD simulations were performed on each system using different random initial velocities. Each simulation was performed for 20 ns resulting in 60 ns cumulative simulation time for each system.
|
10.1371/journal.pcbi.1002665 | Quantitative Predictions of Binding Free Energy Changes in Drug-Resistant Influenza Neuraminidase | Quantitatively predicting changes in drug sensitivity associated with residue mutations is a major challenge in structural biology. By expanding the limits of free energy calculations, we successfully identified mutations in influenza neuraminidase (NA) that confer drug resistance to two antiviral drugs, zanamivir and oseltamivir. We augmented molecular dynamics (MD) with Hamiltonian Replica Exchange and calculated binding free energy changes for H274Y, N294S, and Y252H mutants. Based on experimental data, our calculations achieved high accuracy and precision compared with results from established computational methods. Analysis of 15 µs of aggregated MD trajectories provided insights into the molecular mechanisms underlying drug resistance that are at odds with current interpretations of the crystallographic data. Contrary to the notion that resistance is caused by mutant-induced changes in hydrophobicity of the binding pocket, our simulations showed that drug resistance mutations in NA led to subtle rearrangements in the protein structure and its dynamics that together alter the active-site electrostatic environment and modulate inhibitor binding. Importantly, different mutations confer resistance through different conformational changes, suggesting that a generalized mechanism for NA drug resistance is unlikely.
| The capacity of the influenza virus to rapidly mutate and render resistance to a handful of FDA approved neuraminidase (NA) inhibitors represents a significant human health concern. To gain an atomic-level understanding of the mechanisms behind drug resistance, we applied a novel computational approach to characterize resistant NA mutations. These results are comparable in accuracy and precision with the best experimental measurements presently available. To the best of our knowledge, this is the first time that a rigorous computational method has attained the level of certainty needed to predict subtle changes in binding free energies conferred by mutations. Analysis of our simulation data provided a thorough description of the thermodynamics of the binding process for different NA-inhibitor complexes, with findings that in some cases challenge current views based on interpretations of the crystallographic data. While we did not find a generalized mechanism of NA resistance, we identified key differences between oseltamivir and zanamivir that discriminate their responses to the three mutations we considered, namely H274Y, N294S and Y252H. It is worth noting that our approach can be broadly applied to predict resistant mutations to existing and newly developed drugs in other important drug targets.
| Current plans for managing future influenza pandemics include the use of therapeutic and prophylactic drugs, such as zanamivir [1] and oseltamivir [2], that target the virus surface glycoprotein neuraminidase (NA) [3]. Inhibition of NA reduces the spread of the virus in the respiratory tract by interfering with the release of progeny virions from infected host cells. A handful of drug-resistant strains have recently emerged due to antigenic drift [4], [5], [6]. NA in these strains contains a series of mutations that do not significantly alter its function, yet render it resistant to inhibition. These mutations lead to a small (1–3 kcal/mol) decrease in the high-affinity binding of these inhibitors that is sufficient to restore in vivo viral propagation. Understanding how different NA mutations confer drug resistance is a critical step in discovering new drugs to safeguard against future influenza pandemics.
NAs from different influenza subtypes exhibit a variety of resistance mutations and these mutations can affect inhibitors differently. For example, the R292K mutation in N2 NAs confers resistance to oseltamivir [7], but in highly similar N1 NAs such mutation remains drug sensitive [8]. These and other complex patterns of resistance can only be explained by the interactions between the binding site and the inhibitors. Previous biochemical [9] and structural studies [10] have implicated the rearrangement of certain binding-site residues as the mechanism of drug resistance in NA. For example, bulky substitutions at H274 result in a conformational shift of the neighboring E276, which alters a hydrophobic pocket that specifically disrupts oseltamivir binding. While such structure-based explanations are plausible, a critical evaluation of these hypotheses requires atomic-scale models that accurately reflect the microscopic structural mechanisms guiding NA-inhibitor interactions.
X-ray crystallography provides high-resolution structures of NA-inhibitor complexes. Although such structures are vital to our understanding of NA-inhibitor interactions, the atomic coordinates themselves lend little direct insight into the underlying thermodynamics of drug resistance. There are numerous examples of crystal structures of proteins with drug resistance mutations, such as of HIV-1 protease [11], that show only minor structural differences when compared to the drug-sensitive wild type (WT) structure and do not reveal any readily apparent mechanism of resistance. Numerous drug resistance mutations in NA fall outside of the immediate binding pocket, and structures of the drug-resistant H274Y and N294S mutants co-crystallized with oseltamivir and zanamivir reveal binding-site conformations that are virtually identical to WT [10]. Molecular simulations that rigorously model the microscopic structure and thermodynamics [12], [13], [14] of NA-inhibitor interactions may provide insight into the mechanisms of drug resistance that elude traditional structure-based approaches.
Accurately modeling the thermodynamic consequences of mutations that alter protein function, such as in drug resistance, is a major challenge in structural biology. The change in binding free energy associated with a drug resistance mutation is a result of systemic shifts across the totality of structural conformations that impact which biochemical interactions are accessible in the wild-type and the mutant protein systems. Due to the staggering conformational complexity of a protein-inhibitor complex, direct and exhaustive modeling of this entire system is computationally unfeasible. To overcome such difficulties, two types of approaches for predicting free-energy changes from point mutations have been developed: empirical approaches, which apply highly trained score functions that approximate the free energy of a given structure, and simulation-based approaches, which combine extensive stochastic sampling with statistical mechanics-based calculations to estimate free energies. These approaches have been reviewed extensively elsewhere [15], [16], [17].
While empirical approaches have been moderately successful at identifying mutations along interfacial residues that disrupt binding, they fail to identify the numerous mutations outside of the interface where the effects are presumably smaller [18]. Even the most rigorous simulation-based methods currently available, such as Thermodynamic Integration (TI) and the closely related Free Energy Perturbation (FEP) [12], [13], [19], [20], [21], [22], may lack the accuracy and precision to assess small changes to otherwise large binding free energies. These methods, which, in theory, should capture the thermodynamic effects of protein mutations, have been applied to compute absolute binding free energies of several small molecules to wild type and mutant enzymes, including T4 lysozyme and NA [23], [24], [25], [26]. However, straightforward applications of these techniques to large, complex systems are hampered by significant sampling issues. These issues are particularly severe in systems with hindered conformational transitions associated with ligand binding, which often render the resulting absolute binding free energy calculations unreliable [27], [28], [29], [30]. Conventional methods for calculating relative binding free energies across a series of related compounds avoid many of the sampling issues associated with absolute binding free energy calculations [31], however, they are typically not directly applicable to assessing the effects of mutations on binding of the same compound.
Successful modeling of the thermodynamics of large, complex systems, such as NA, requires careful selection of both the conformational sampling strategy and the appropriate reference states in order to obtain precise and accurate estimates of free energy changes. We recently described a novel implementation of the Hamiltonian Replica Exchange (HREX) molecular dynamics (MD) method [31] that uses an alchemical thermodynamic pathway to arrive at reliable free energy calculations. Here, we adapted this approach to incorporate residue mutations into the thermodynamic cycle. Instead of estimating changes of binding free energies of different compounds with respect to the same protein, we estimated free energy changes for mutating a residue in the bound and unbound wild type protein. We applied this method to several such pathways to predict the binding free energy changes (ΔΔG) of a set of mutations in H5N1 NA that have been experimentally tested for drug resistance. We successfully identified drug resistance mutations in NA using a judiciously chosen thermodynamic path within the HREX framework.
For this work, we adapted the criterion introduced by Kortemme et al. [32] to classify a mutation as drug resistant when its calculated ΔΔG exceeded +1 kcal/mol. Based on this criterion, the experimentally observed NA mutations N294S, H274Y, and Y252H reveal different resistance patterns with respect to oseltamivir and zanamivir [10]. We explored the capabilities of our approach and alternate ones, including those from previously published work [33], [34], to produce accurate and precise ΔΔG estimates consistent with the experimental data [10].
Analysis of over 15 µs of aggregate MD simulation data revealed that different mutations confer resistance through different conformational changes in the active site. Unexpectedly, we found no evidence supporting the previously reported role of hydrophobic interactions with the oseltamivir tail [10]. Instead, we hypothesize that drug resistance arises from rearrangements of several charged residues that alter the electrostatic environment within the binding site and disrupt inhibitor binding. The complexity of the observed structural perturbations highlights the importance of atomic-level structural details and suggests that identification of a generalized theory of resistance is unlikely.
We computed relative instead of absolute binding free energy changes using Single Reference Thermodynamic Integration (SRTI) [31]. Computing relative ΔΔGs requires measuring the free energy change along an alchemical thermodynamic path linking the WT to the mutant protein for the ligand-bound and ligand-free states independently, which requires only a partial ‘decoupling’ of the mutating residues and/or ligand along that alchemical path. In contrast, absolute ΔΔG computations entail measuring the free energy change along an alchemical thermodynamic path connecting the ligand-bound and ligand-free states, which requires a complete decoupling of the ligand from the protein [35]. Previous MD simulations of NA [36], [37] revealed substantial binding-induced conformational changes along a 150-residue loop. A complete decoupling of the ligand [25] would necessitate extensive sampling of this large conformational transition, making reliable free energy predictions practically impossible. By avoiding the need to explicitly model this binding-induced conformational change, the relative SRTI approach is better suited for ΔΔG calculations for NA.
Determining the molecular mechanisms of NA drug resistance involves identifying key protein structural features that underlie the thermodynamic differences in inhibitor binding observed in the simulation data. Such features may include changes in biochemical interactions in the NA-inhibitor complex, systematic shifts in the NA structure, and even subtle differences in the overall dynamics between WT and drug-resistant NA. A visual comparison between the crystal structures of NA in complex with zanamivir and oseltamivir revealed few apparent differences in NA-inhibitor interactions. Therefore, we analyzed the structural data derived from the SRSM/HREX simulations in order to identify reliable structural differences between WT and drug-resistant mutant trajectories.
Fig. 2 illustrates representative structures from the WT and drug-resistant mutant trajectories for zanamivir and oseltamivir, confirming the x-ray crystallography findings that the most prominent binding interactions are preserved. The negatively charged carboxyl group of both inhibitors maintained interactions with a basic triad formed by R118, R292, and R371. The positively charged ammonium and guanidinium groups of oseltamivir and zanamivir, respectively, maintained salt-bridges with the acidic E119, D151, and E227 residues (E227 is not displayed in Fig. 2 for purposes of clarity). Finally, the polar tail of zanamivir maintained some of the hydrogen bonds with R224, E276, and E277 in both WT and mutant forms. The long-range nature of these electrostatic interactions and the highly flexible nature of the binding site suggest that NA-inhibitor binding is highly sensitive to subtle, systematic rearrangements of the electrostatic environment caused by mutations beyond the immediate binding site. Our analysis identified several such rearrangements that may be critical to drug resistance.
We used MD simulations and statistical mechanics to quantify the effect of drug resistance mutations in NA on the ΔΔG of oseltamivir and zanamivir binding. We found that implicit solvent-based methods, such as MM-GBSA, and empirical approaches, such as Rosetta, were largely unable to predict drug resistance. However, careful use of thermodynamic-integration-based approaches successfully predicted binding affinities with chemical accuracy. Ultimately, the SRSM/HREX approach yielded the most accurate and precise ΔΔG values compared with those obtained experimentally. The SRSM approach minimized the degree of decoupling between the real states and the unphysical reference states, while HREX significantly enhanced conformational sampling as a result of exchanges between the TI simulation windows. Together, the SRSM/HREX approach successfully sampled a thermodynamic path between WT and mutant NA which circumvented conformational sampling barriers that significantly impeded conventional MD simulations to yield highly reliable free energy calculations. The additional computational cost associated with using HREX was practically negligible compared to SRSM because the time required for both types of runs is roughly equivalent. Finally, we must point out that the computation of ΔΔGs using SRSM (or SRSM/HREX) is computationally demanding. To evaluate the six ΔΔG values for NA with their corresponding standard errors, we were required to carry out a minimum of 36 runs, each 4ns-long and involving 31 replicas, for an aggregate simulation time of ∼4.5 µs. The whole analysis presented here required over 15 µs of aggregated MD simulations.
We analyzed trajectories from the SRSM/HREX simulations in order to identify the structural and energetic mechanisms underlying the computed ΔΔGs. We identified a number of subtle, systematic, rearrangements in the extensive hydrogen bonding and electrostatic interactions in the inhibitor binding site in the drug resistant H274Y and N294S mutations that were largely absent in the drug-sensitive Y252H mutation. Although the exact nature of these electrostatic rearrangements varied for each drug and mutation, we hypothesize that these rearrangements in the binding pocket form the basis of drug resistance in NA. This is in contrast with the previous interpretations of the experimental structures that suggested changes in the size and hydrophobicity of the binding pocked as the primary mechanism for resistance [10].
Our study marks the most extensive use to date of molecular dynamics and thermodynamic integration on a large, pharmaceutically relevant system and demonstrates that a rigorous, computationally intensive approach can be successfully applied to studying the thermodynamic mechanisms underlying protein function that can elude traditional structure-based crystallography approaches.
Coordinates of the protein systems were derived from the crystal structures of NA (PDB codes: 2HTY, 3CL0, 3CL2, and 3CKZ) [10]. A detailed description of the setup is provided in SI Section 1b in Text S1.
We used SRTI to calculate the relative free energy difference between a RS and a given end state of a system. Unless stated otherwise, all the simulation details were the same as described previously [31]. The end states in our simulations were NA variants, either free or bound to an inhibitor. To enhance sampling between the states, we employed HREX.
To run the MD simulations, we employed the GROMACS program version 4.0.5. Production runs were 4 ns long for each SRTI window and the coordinates of the system were recorded every 500 steps for subsequent analyses. SRTI simulations augmented with HREX were run using m = 31 windows and replica exchanges were attempted every 500 MD steps. A total of 4,000 attempted exchanges were produced, which resulted in 4 ns-long simulations per window.
The SRMM approach (Fig. 1) allows for simultaneous comparison of binding free energy changes between all pairs of proteins and ligands. To implement this approach, we designed a common RS for all proteins and ligands for the bound and unbound state. Portions of all three mutating residues and the ligand were decoupled in these simulations. The details are provided in SI Section 1i and Fig. S2A in Text S1.
The SRSM approach (Fig. 1) computed ΔΔG between WT and a specific mutant for each ligand. To implement this approach we constructed specific reference states for each mutant and ligand in the bound and unbound state. Only a single amino acid was decoupled in these simulations. The details are available in SI Section 1j and Fig. S2B in Text S1.
The MM-PBSA/GBSA method [14], as implemented in Amber10, was used to obtain additional estimates of the changes in binding free energy based on SRSM trajectories. Additional details are provided in SI Section 1k in Text S1.
RosettaInterface [32] uses computational mutagenesis to predict the change in binding free energy of a protein-protein interaction associated with point mutations. Details on the implementation of RosettaInterface for protein-ligand interactions are provided in SI Section 1l in Text S1.
|
10.1371/journal.pcbi.1006802 | Efficient algorithms to discover alterations with complementary functional association in cancer | Recent large cancer studies have measured somatic alterations in an unprecedented number of tumours. These large datasets allow the identification of cancer-related sets of genetic alterations by identifying relevant combinatorial patterns. Among such patterns, mutual exclusivity has been employed by several recent methods that have shown its effectiveness in characterizing gene sets associated to cancer. Mutual exclusivity arises because of the complementarity, at the functional level, of alterations in genes which are part of a group (e.g., a pathway) performing a given function. The availability of quantitative target profiles, from genetic perturbations or from clinical phenotypes, provides additional information that can be leveraged to improve the identification of cancer related gene sets by discovering groups with complementary functional associations with such targets. In this work we study the problem of finding groups of mutually exclusive alterations associated with a quantitative (functional) target. We propose a combinatorial formulation for the problem, and prove that the associated computational problem is computationally hard. We design two algorithms to solve the problem and implement them in our tool UNCOVER. We provide analytic evidence of the effectiveness of UNCOVER in finding high-quality solutions and show experimentally that UNCOVER finds sets of alterations significantly associated with functional targets in a variety of scenarios. In particular, we show that our algorithms find sets which are better than the ones obtained by the state-of-the-art method, even when sets are evaluated using the statistical score employed by the latter. In addition, our algorithms are much faster than the state-of-the-art, allowing the analysis of large datasets of thousands of target profiles from cancer cell lines. We show that on two such datasets, one from project Achilles and one from the Genomics of Drug Sensitivity in Cancer project, UNCOVER identifies several significant gene sets with complementary functional associations with targets. Software available at: https://github.com/VandinLab/UNCOVER.
| Sequencing technologies allow the measurement of somatic alterations in a large number of cancer samples. Several methods have been designed to analyze these alterations, but the characterization of the functional effects of such alterations is still challenging. A recent promising approach for such characterization is to combine alteration data with quantitative profiles obtained, e.g., from genetic perturbations. The analysis of these data is complicated by the extreme heterogeneity of alterations in cancer, with different cancer samples exhibiting vastly different alterations. This heterogeneity is due, in part, to the complementarity of alterations in cancer pathways, with alterations in different genes resulting in the same alteration at the functional level. We develop UNCOVER, an efficient method to identify sets of alterations displaying complementary functional association with a quantitative profile. UNCOVER is much more efficient than the state-of-the-art, allowing the identification of complementary cancer related alterations from genome-scale measurements of somatic mutations and genetic perturbations.
| Recent advances in sequencing technologies now allow to collect genome-wide measurements in large cohorts of cancer patients (e.g., [1–6]). In particular, they allow the measurement of the entire complement of somatic (i.e., appearing during the lifetime of an individual) alterations in all samples from large tumour cohorts. The study of such alterations has lead to an unprecedented improvement in our understanding of how tumours arise and progress [7]. One of the main remaining challenges is the interpretation of such alterations, in particular identifying alterations with functional impact or with relevance to therapy [8].
Several computational and statistical methods have been recently designed to identify driver alterations, associated to the disease, and to distinguish them from random, passenger alterations not related with the disease [9]. The identification of genes associated with cancer is complicated by the extensive intertumour heterogeneity [10], with large (100-1000’s) and different collections of alterations being present in tumours from different patients and no two tumours having the same collection of alterations [10, 11]. Two main reasons for such heterogeneity are that i) most mutations are passenger, random mutations, and, more importantly, ii) driver alterations target cancer pathways, groups of interacting genes that perform given functions in the cell and whose alteration is required to develop the disease. Several methods have been designed to identify cancer genes using a-priori defined pathways [12] or interaction information in the form of large interaction networks [13, 14].
Recently several methods (see Section Related work) for the de novo discovery of mutated cancer pathways have leveraged the mutual exclusivity of alterations in cancer pathways. Mutual exclusivity of alterations, with sets of genes displaying at most one alteration for each patient, has been observed in various cancer types [7, 11, 15, 16]. The mutual exclusivity property is due to the complementarity of genes in the same pathway, with alterations in different members of a pathway resulting in a similar impact at the functional level, while mutations in different members of the same pathway may not provide further selective advantage or may even lead to a disadvantage for the cell (e.g., in synthetic lethality). Even if mutual exclusivity of alterations is neither a sufficient nor a necessary property of cancer pathways, it has been successfully used to identify cancer pathways in large cancer cohorts [15, 17, 18].
An additional source of information that can be used to identify genes with complementary functions are quantitative measures for each samples such as: functional profiles, obtained for example by genomic or chemical perturbations [19–21]; clinical data describing, obtained for example by (quantitative) indicators of response to therapy; activation measurements for genes or sets of genes, as obtained for example by single sample scores of Gene Set Enrichment Analysis [22, 23]. The employment of such quantitative measurements is crucial to identify meaningful complementary alterations since one can expect mutual exclusivity to reflect in functional properties (of altered samples) that are specific to the altered samples. For example, consider a scenario (Fig 1) in which there are two altered molecular mechanisms: one that is altered in almost all samples and one that is altered in much fewer samples, but is related to the response to a given therapy (for example by interacting with a drug target). Methods that ignore therapy response information will report the first mechanism as significantly altered, while the second mechanisms, altered in a smaller fraction of all samples, is identified only by considering the therapy response information.
Several recent methods have used mutual exclusivity signals to identify sets of genes important for cancer [24]. RME [25] identifies mutually exclusive sets using a score derived from information theory. Dendrix [26] defines a combinatorial gene set score and uses a Markov Chain Monte Carlo (MCMC) approach for identifying mutually exclusive gene sets altered in a large fraction of the patients. Multi-Dendrix [27] extends the score of Dendrix to multiple sets and uses an integer linear program (ILP) based algorithm to simultaneously find multiple sets with mutually exclusive alterations. CoMET [18] uses a generalization of Fisher exact test to higher dimensional contingency tables to define a score to characterize mutually exclusive gene sets altered in relatively low fractions of the samples. WExT [18] generalizes the test from CoMET to incorporate individual gene weights (probabilities) for each alteration in each sample. WeSME [28] introduces a test that incorporates the alteration rates of patients and genes and uses a fast permutation approach to assess the statistical significance of the sets. TiMEx [29] assumes a generative model for alterations and defines a test to assess the null hypothesis that mutual exclusivity of a gene set is due to the interplay between waiting times to alterations and the time at which the tumor is sequenced. MEMo [17] and the method from [30] employ mutual exclusivity to find gene sets, but use an interaction network to limit the candidate gene sets. The method by [31] and PathTIMEx [32] introduce an additional dimension to the characterization of inter-tumor heterogeneity, by reconstructing the order in which mutually exclusive gene sets are mutated. None of these methods take quantitative targets into account in the discovery of significant gene sets and sets showing high mutual exclusivity may not be associated with target profiles (Fig 1).
[33] recently developed the repeated evaluation of variables conditional entropy and redundancy (REVEALER) method, to identify mutually exclusive sets of alterations associated with functional phenotypes. REVEALER uses as objective function (to score a set of alterations) a re-scaled mutual information metric called information coefficient (IC). REVEALER employs a greedy strategy, computing at each iteration the conditional mutual information (CIC) of the target profile and each feature, conditioned on the current solution. REVEALER can be used to find sets of mutually exclusive alterations starting either from a user-defined seed for the solution or from scratch, and [33] shows that REVEALER finds sets of meaningful cancer-related alterations.
In this paper we study the problem of finding sets of alterations with complementary functional associations using alteration data and a quantitative (functional) target measure from a collection of cancer samples. Our contributions in this regard are fivefold. First, we provide a rigorous combinatorial formulation for the problem of finding groups of mutually exclusive alterations associated with a quantitative target and prove that the associated computational problem is NP-hard. Second, we develop two efficient algorithms, a greedy algorithm and an ILP-based algorithm to identify the set of k genes with the highest association with a target; our algorithms are implemented in our method fUNctional Complementarity of alteratiOns discoVERy (UNCOVER). Third, we show that our algorithms identify highly significant sets of genes in various scenarios; in particular, we compare UNCOVER with REVEALER on the same datasets used in [33], showing that UNCOVER identifies solutions of higher quality than REVEALER while being on average two order of magnitudes faster than REVEALER. Interestingly, the solutions obtained by UNCOVER are better than the ones obtained by REVEALER even when evaluated using the objective function (IC score) optimized by REVEALER. Fourth, we show that the efficiency of UNCOVER enables the analysis of large datasets, and we analyze a large dataset from Project Achilles, with thousands of genetic dependencies measurements and tens of thousands of alterations, and a large dataset from the Genomics of Drug Sensitivity in Cancer (GDSC) project, with hundreds of drug sensitivity measurements and tens of thousands of alterations. On such datasets UNCOVER identifies several statistically significant associations between target values and mutually exclusive alterations in genes sets, with an enrichment in well-known cancer genes and in known cancer pathways.
This section describes the problem we study and the algorithms we designed to solve it, that are implemented in our tool UNCOVER. We also describe the data and computational environment for our experimental evaluation.
The workflow of our algorithm UNCOVER is presented in Fig 2. UNCOVER takes in input information regarding 1. the alterations measured in a number of samples (e.g., patients or cell lines), and 2. the value of the target measure for each patient. UNCOVER then identifies the set of mutually exclusive alterations with the highest association to the target, and employs a permutation test to assess the significance of the association. Details regarding the computational problem and the algorithms used by UNCOVER are described in the following sections. The implementation of UNCOVER is available at https://github.com/VandinLab/UNCOVER.
We tested UNCOVER on a number of cancer datasets in order to compare its results to the ones obtained without using the target, to state-of-the-art algorithms, and to test whether UNCOVER allows the analysis of large datasets. In particular, we first assessed the impact of the target values on the results of UNCOVER. We then compared UNCOVER with REVEALER using four datasets described in [33] as well as the GDSC project dataset described above. We then used simulated data to asses the performance of UNCOVERin finding groups of alterations associated with a target. We then performed a scalability test using a large dataset from the Achilles project and alterations from the Cancer Cell Line Encyclopedia (CCLE). Finally, we used UNCOVER to analyze a drug sensitivity dataset from the GDSC project.
We ran UNCOVER on the GDSC dataset for k = 3 and compared the results obtained when the target values are not considered in the analysis, obtained running UNCOVER ILP with k = 3 while setting the target values to 1 for all the samples considered in the analysis of a target (S1 Table). The latter analysis corresponds to the extraction of sets with high mutual exclusivity (e.g., by [34]). As expected, the solutions obtained in the two cases are very different: the solution obtained without considering the target values has one alteration in common with the solution obtained by UNCOVER using either positive or negative values of the target for only 11 targets of the 265 in the GDSC dataset, and for no target the solutions share more than 1 alteration. An example of the solutions obtained target using UNCOVER and without considering the target values are shown in Fig 3. We observe that while the solutions obtained considering the target values display an association with the target profile (positive or negative), the solution obtained when the target values are not considered, while covering a large set of samples, does not display any positive or negative association with the target profile. To asses the association between target values and alterations more consistently we calculated the point biserial coefficient [40] for all 265 solutions. The coefficient varies between −1 and +1 with 0 implying no correlation. The average value obtained when ignoring the target is −0.02 with standard deviation 0.05, while the the average value obtained by UNCOVER is 0.20 with standard deviation 0.05. These results show that a mutual exclusivity analysis that disregards the values of the target does not identify sets of mutually exclusive alterations associated with target values. In addition, the genes in solution identified by considering the drug target have a much more significant enrichment in known cancer genes, as reported in [11], than the genes in solution identified disregarding the values of the target (p = 3 × 10−12 vs p = 10−2).
We run the greedy algorithm and the ILP from UNCOVER on the same four datasets considered by the REVEALER publication [33]. We used the same values of k used in [33], that is k = 3 for all the datasets, except from the KRAS dataset where k = 4 was used. For each dataset we recorded the solution reported by the greedy algorithm, the solution reported by the ILP, the value of the objective functions for such solutions and the running time to obtain such solutions. For ILP solutions, we also performed the permutation test (see Materials and methods) to compute a p-value using 1000 permutations. The results are reported in Table 1, in which we also show the results from REVEALER (without initial seeds). Fig 4 shows alteration matrices and the association with the target for the solutions identified by UNCOVER.
We can see that the greedy algorithm identifies the same solution of the ILP based algorithm in three out of four cases, and that the runtime of the ILP and the runtime of greedy algorithm are comparable and very low (< 40 seconds) in all cases. In contrast, the running time of REVEALER is much higher (> 1000 seconds in most cases). (We included all preprocessing in the reported UNCOVER runtimes in Table 1 to ensure a fair comparison with REVEALER; not including preprocessing our running times are all under 10 seconds). Comparing the alteration matrices of the solutions by UNCOVER and the ones of solutions by REVEALER (S1 Fig) we note that alterations in solutions by UNCOVER tend to have higher mutual exclusivity and to be more concentrated in high weight samples than alterations in solutions by REVEALER. As expected, the value of the objective function we use is much lower for solutions from REVEALER than for solutions from our algorithm.
We then compared the solutions obtained by our algorithms with the solutions from REVEALER in terms of the information coefficient (IC), that is the target association score used in [33] as a quality of the solution. Surprisingly, in two out of four datasets UNCOVER, which does not consider the IC score, identifies solutions with IC score higher (by at least 5%) than the solutions reported by REVEALER. For the other two cases, in one dataset the IC score is very similar (0.50 vs 0.49) while in the other case the IC score by REVEALER is higher (0.7 vs 0.67) but the solution reported by REVEALER differs from the solution reported by UNCOVER by 1 gene only. Interestingly, the latter is the only case where the solution from the ILP has a p-value > 0.1 (p < 0.03 in all other cases), and therefore the solutions (by our methods and by REVEALER) for such dataset may be, at least in part, due to random fluctuations of the data.
In terms of biological significance, in most cases the solutions by UNCOVER and by REVEALER are very similar, with cancer relevant genes identified by both methods. For NFE2L2 activation, both methods identify KEAP1, a repressor of NFE2L2 activation [41]. For MEK-inhibitor, both methods find BRAF, KRAS, and NRAS, three well known oncogenic activators of the MAPK signaling pathway, which contains MEK as well. For KRAS essentiality, both methods report mutations in KRAS in the solution. For β-catenin activation, both methods identify CTNNB1 mutations and APC mutations, that is known to be associated to β-catenin activation [42]. These results show that UNCOVER identifies relevant biological solutions that are better than the ones identified by REVEALER when evaluated using our objective function and also when evaluated according to the objective function of REVEALER with a running time that is on average two orders of magnitude smaller than required by REVEALER. Since UNCOVER and REVEALER consider two different objective functions, it is unclear whether the improvement in running time comes from differences in implementation choices or from a inherently different computational complexity. However, since UNCOVER’s objective function is easier to compute than REVEALER’s objective function, we believe that the use of our objective function plays an important role in the efficiency of UNCOVER.
We also compared the solutions obtained by UNCOVER and by REVEALER on the GDSC dataset (S2 Table). For both algorithms we obtained the solutions for k = 3. For UNCOVER, we considered the solution returned by the ILP. For REVEALER, we could only obtain solutions for 246 targets, since for the other targets REVEALER terminated with an error message. Due to the high running time of REVEALER, we only obtained sets of alterations associated with positive values of the target (Table 2). For 33 targets the solution by UNCOVER and the solution by REVEALER share 1 alteration, while for 33 targets the solution by UNCOVER and the solution by REVEALER share 2 alterations; for no target UNCOVER and REVEALER report the same solution. This shows that the two methods identify completely different solution in most (> 73%) of the cases. We compared the solutions obtained by UNCOVER and by REVEALER using the IC score considered by REVEALER but not from UNCOVER: surprisingly, in more than 50% of the cases (113 out of 208) the IC score of the solution from UNCOVER is higher than the IC of the solution from REVEALER. On the other hand, for all targets the solution by REVEALER is worst than the solution by UNCOVER when the UNCOVER objective function is considered. We also compared UNCOVER and REVEALER evaluating the association between target values and alterations in the solutions using a measure of association that is not considered by the two algorithms. In particular, we considered the point biserial correlation coefficient [40]. In more than 95% of the cases (199 out of 208) the point biserial correlation coefficient between the solution from UNCOVER and the target is higher than the point biserial correlation coefficient between the solution from REVEALER and the target, that is, the solution from UNCOVER has an higher association with the target than the solution from REVEALER. On average, the solution from UNCOVER has a point biserial correlation coefficient that is 37% higher than the point biserial correlation coefficient of the solution from REVEALER. Moreover, the average effect size of solutions from UNCOVER is more than 80% higher than the average effct size of solutions from REVEALER (Table 2). In addition, the genes in solutions from UNCOVER have a much higher enrichment (p = 3 × 10−13; 7-fold enrichment) for known cancer genes than solutions from REVEALER (p = 2 × 10−4; 3-fold enrichment). Analogously, more KEGG pathways display a significant enrichment in genes from UNCOVER solutions than from REVEALER solutions (22 vs 11). We also compared the running time of the two methods: UNCOVER required 3 hours to complete the analysis, while REVEALER required 9 days. Overall, these results show that UNCOVER obtains better results than REVEALER not only in terms of the UNCOVER objective function but also in terms of the score from REVEALER as well as in terms of a independent measure of association, while being 70 times faster than REVEALER.
For each combination we generated 10 simulated datasets as described in Materials and methods. Each dataset contains a planted set of 5 alterations associated with the target. We used both the greedy algorithm and the ILP from UNCOVER with k = 5 to attempt to find the 5 correct alteration, and evaluated our algorithms both in terms of fraction of the correct (i.e., planted) solution reported and running time.
As shown in Fig 5, the greedy algorithm is faster than the ILP for all datasets, and the difference in running time increases as the number m of samples increases, with the runtime of the greedy algorithm being almost two orders of magnitude smaller than the runtime of the ILP for m = 1000 samples. In addition, for a fixed number of samples and alterations, the running time of the greedy algorithm is constant, that is it does not depend on the properties of the planted solution, while the running time of the ILP varies greatly depending on these parameters. For m = 10, 000 samples the running time of the ILP becomes extremely high, so we restricted to consider only two sets of parameters (p − n = 0.95 and p − n = 0.2). In this case the ILP took between 44 minutes and 7 hours to complete, while the greedy algorithm terminates in 5 minutes.
In terms of the quality of the solutions found, as expected the ILP outperforms the greedy (Fig 6) but the difference among the two tends to disappear when the number of samples is higher. In addition, since the ILP finds the optimal solution, we can see that for a limited number of samples we may not reliably identify the planted solution with 200 samples unless the planted solution appears almost only in positive targets and in almost all of them (p − n = 0.95), while for m = 1000 we can reliably identify the planted solution using both the ILP and the greedy algorithm even when the association with the target is weaker (p − n = 0.6). When m = 10, 000, both the ILP and the greedy algorithm perform well in terms of the quality of the solution: the ILP finds the correct alterations on every experiment and the greedy identifies the whole planted solution in all experiments but one for p − n = 0.2, for which it still reports a solution containing 4 out of 5 genes in the planted solution.
These results show that for a large number of samples the greedy algorithm reliably identifies sets of alterations associated with the target, as predicted by our theoretical analysis, and is much faster than the ILP. For smaller sample size the ILP identifies better solutions than the greedy and has a reasonable running time.
The efficiency of UNCOVER renders the analysis of a large number of targets, such as the ones available through the Achilles project, possible. After preprocessing the dataset included 5690 functional phenotypes as targets, and for each of these the CCLE provides alteration information for 205 samples and 31137 alterations. In total we have therefore run 11380 instances (i.e., 5690 targets screened for positive and for negative associations) looking for both positive and negative association with target values. Since the number of samples (205) is relatively small, we have run only the ILP from UNCOVER on the whole Achilles dataset and looked for solutions with k = 3 genes. The runtime of UNCOVER to find both positive and negative associations, including preprocessing, is 24 hours. Based on the runtime required on the instances reported in [33] (see the Section Comparison with REVEALER), running REVEALER on this dataset would have required about 5 months of compute time.
To identify statistically significant associations with targets in the Achilles project dataset we used a nested permutation test. We first run the permutation test with 10 permutations on all instances (i.e., on all targets for both positive association and negative association). We then considered all the instances with the lowest p-value (1/11) and performed a permutation test with 100 permutations only for such instances. We the iterated such procedure once more, selecting all the instances with lowest p-value (1/101) and performing a permutation test with 1000 permutations only for such instances. For positive association we found 60 solutions with p-value < 0.001, and for negative association we found 102 solutions with p-value < 0.001. The solutions with p-value < 0.001 (with 1000 permutations) are reported in S3 Table. See S2 Fig for some corresponding alteration matrices.
The genes in the solutions by UNCOVER with p-value 1/1001 are enriched (p = 2 × 10−12 by Fisher exact test; 8 fold enrichment) for well-known cancer genes. We also tested whether genes in solutions by UNCOVER (with p-value 1/1001) are enriched for interactions, by comparing the number of interactions in iRefIndex [43] among genes in such solution with the number of interactions in random sets of genes of the same cardinality. Genes in solutions by UNCOVER are significantly enriched in interactions (p = 7 × 10−3 by permutation test; 2 fold enrichment). In addition, the genes in solutions by UNCOVER are also enriched in genes in well-known pathways: 12 KEGG pathways [44] have a significant (corrected p ≤ 0.05) overlap with genes in solutions by UNCOVER and four of these (endometrial cancer, glioma, hepatocellular carcinoma, EGFR tyrosine kinase inhibitor resistance) are cancer related pathways. In addition, the targets (i.e., genes) with solutions of p-value 1/1001 are enriched (p = 5 × 10−4 by permutation test; 6 fold enrichment) for interactions in iRefIndex and for well-known cancer genes (p = 2 × 10−12 by Fisher exact test; 8 fold enrichment) as reported in [11]. These results show that UNCOVER enables the identification of groups of well known cancer genes with significant associations to important targets in large datasets of functional target profiles. For example, for target (i.e., silenced gene) TSG101, related to cell growth, UNCOVER identifies the gene set shown in Fig 7 as associated to reduced cell viability. ERBB2 is a well known cancer gene and CDH4 is frequently mutated in several cancer types, and both are associated to cell growth.
We use UNCOVER to analyze the GDSC project data, identifying sets of alterations associated with drug sensitivity. After preprocessing, the dataset included 64144 alterations and 265 targets, and for each of these the number of cell lines with available data varied between 240 and 705. In total we have therefore run 530 instances (i.e., 265 targets screened for positive and for negative associations) looking for both positive and negative association with target values.
We used the UNCOVER ILP for all instances to obtain solutions with k = 3 genes. For each solution, we use 100 permutations to compute its p-value. For positive association we found 51 solutions with p-value < 0.01, and for negative association we found 41 solutions with p-value < 0.01. We used the following procedure to focus on the most significant solutions: we run UNCOVER with k = 4 and computed the p-values for the solutions using 100 permutations; we then identified targets whose solution for k = 3 have p-value < 0.01 and are contained in the solution for the same target with k = 4 and have p-value p < 0.01 for k = 4. In total, this procedure identifies 23 solutions for positive association and 22 solutions for negative associations. These solutions are reported in S4 Table.
The genes in the solutions identified as above are enriched (p = 9 × 10−10 by Fisher exact test; 20 fold enrichment) for well-known cancer genes, as reported in [11]. We also tested whether these genes in solutions are enriched for interactions, by comparing the number of interactions in iRefIndex [43] among genes in such solution with the number of interactions in random sets of genes of the same cardinality. Genes in solutions by UNCOVER are significantly enriched in interactions (p = 2 × 10−2 by permutation test; 6 fold enrichment). In addition, these genes are also enriched in genes in well-known pathways: 21 KEGG pathways [44] have a significant (corrected p ≤ 0.05) overlap with genes in solutions by UNCOVER and 19 of these are cancer related pathways (e.g., ErbB signaling pathway) or related to drug resistance (e.g., EGFR tyrosine kinase inhibitor resistance).
For Palbociclib, UNCOVER identifies RB1 mutations, GRB7 amplifications, and RB1 deletions with significant association with reduced sensitivity to drug. RB1 is a well known cancer gene. The alterations are shown in Fig 3a. While RB1 mutations and RB1 deletions are significantly associated when considered in isolation (the association of single alterations with drug sensitivity and the drug targets have been obtained from https://www.cancerrxgene.org/), GRB7 amplification is not associated with the target values when considered in isolation. GRB7 encodes a growth factor receptor-binding protein that interacts with epidermal growth factor receptor (EGFR). Both RB1 and EGFR are related to the cell cycle pathway, that is the pathway target of the compound, and the drug targets (CDK4, CDK6) as well EGFR are members of the PI3K-AKT pathway. For Sunitinib, UNCOVER identifies mutations in SETD2, ARHGAP19, and RB1, with significant association with reduced sensitivity to drug. The alterations are shown in Fig 8a. RB1 is a well known cancer gene and SETD2 has tumor suppressor functionality. None of these alterations have significant association with drug sensitivity when considered in isolations. RB1 and SETD2 are involved in protein localization to chromatin, and ARHGAP19 is part of Rho mediated remodeling. For PLX-4720-2, UNCOVER identifies mutations in BRAF, CD244, and ARSB with significant association to increased sensitivity to drug. The alterations are shown in Fig 8b. BRAF is a well-known cancer gene; it is the target of the compound and BRAF mutations have significant association to increased sensitivity to the compound, while the other two alterations do not. BRAF and CD244 are part of natural killer cell mediated cytotoxicity pathway, while ARSB is involved in the regulation of cell adhesion, cell migration and invasion in colonic epithelium [45], and is also part of metabolism related pathways. For VX-11e, UNCOVER identifies mutations in BRAF, KRAS, and NRAS, with significant association to increased sensitivity to drug. The alterations are shown in Fig 8c. Only BRAF mutations have significant association with the target when considered in isolation. The pathway target for the compound is the ERK MAPK signaling pathway, to which all three alterations are related. All three genes have well identified roles in cancer. These results show that UNCOVER enables the identification of groups of relevant genes, many related to cancer, with significant associations to important targets in large datasets of drug sensitivity profiles.
In this work we study the problem of identifying sets of mutually exclusive alterations associated with a quantitative target profile.
We provide a combinatorial formulation for the problem, proving that the corresponding computational problem is NP-hard. We design two efficient algorithms, a greedy algorithm and an ILP-based algorithm, for the identification of sets of mutually exclusive alterations associated with a target profile. We provide a formal analysis for our greedy algorithm, proving that it returns solutions with rigorous guarantees in the worst-case as well under a reasonable generative model for the data. We implemented our algorithms in our method UNCOVER, and showed that it finds sets of alterations with a significant association with target profiles in a variety of scenarios. By comparing the results of UNCOVER with the results of REVEALER on four target profiles used in the REVEALER publication [33] and on a large dataset from the GDSC project, we show that UNCOVER identifies better solutions than REVEALER, even when evaluated using REVEALER objective function. Moreover, UNCOVER is much faster than REVEALER, allowing the analysis of large datasets such as the dataset from project Achilles and from the GDSC project, in which UNCOVER identifies a number of associations between functional target profiles and gene set alterations.
Our tool UNCOVER (as well as REVEALER) relies on the assumption that the mutual exclusivity among alterations is due to functional complementarity. Another explanation for mutual exclusivity is the fact that each cancer may comprise different subtypes, with different subtypes being characterized by different alterations [27]. UNCOVER can be used to identify sets of mutually exclusive alterations associated with a specific subtype whenever the subtype information is available, by assigning high weight to samples of the subtype of interest and low weight to samples of the other subtypes. In addition, while we consider a penalty based on mutual exclusivity, other types of penalties may be used to identify sets of alterations associated with a target profile. The study of the theoretical properties of the problem and the analysis of the results with different penalties are interesting directions of future research.
|
10.1371/journal.pntd.0000651 | Structural Characterization of CYP51 from Trypanosoma cruzi and Trypanosoma brucei Bound to the Antifungal Drugs Posaconazole and Fluconazole | Chagas Disease is the leading cause of heart failure in Latin America. Current drug therapy is limited by issues of both efficacy and severe side effects. Trypansoma cruzi, the protozoan agent of Chagas Disease, is closely related to two other major global pathogens, Leishmania spp., responsible for leishmaniasis, and Trypansoma brucei, the causative agent of African Sleeping Sickness. Both T. cruzi and Leishmania parasites have an essential requirement for ergosterol, and are thus vulnerable to inhibitors of sterol 14α-demethylase (CYP51), which catalyzes the conversion of lanosterol to ergosterol. Clinically employed anti-fungal azoles inhibit ergosterol biosynthesis in fungi, and specific azoles are also effective against both Trypanosoma and Leishmania parasites. However, modification of azoles to enhance efficacy and circumvent potential drug resistance has been problematic for both parasitic and fungal infections due to the lack of structural insights into drug binding.
We have determined the crystal structures for CYP51 from T. cruzi (resolutions of 2.35 Å and 2.27 Å), and from the related pathogen T. brucei (resolutions of 2.7 Å and 2.6 Å), co-crystallized with the antifungal drugs fluconazole and posaconazole. Remarkably, both drugs adopt multiple conformations when binding the target. The fluconazole 2,4-difluorophenyl ring flips 180° depending on the H-bonding interactions with the BC-loop. The terminus of the long functional tail group of posaconazole is bound loosely in the mouth of the hydrophobic substrate binding tunnel, suggesting that the major contribution of the tail to drug efficacy is for pharmacokinetics rather than in interactions with the target.
The structures provide new insights into binding of azoles to CYP51 and mechanisms of potential drug resistance. Our studies define in structural detail the CYP51 therapeutic target in T. cruzi, and offer a starting point for rationally designed anti-Chagasic drugs with improved efficacy and reduced toxicity.
| Chagas Disease is caused by kinetoplastid protozoa Trypanosoma cruzi, whose sterols resemble those of fungi, in both composition and biosynthetic pathway. Azole inhibitors of sterol 14α-demethylase (CYP51), such as fluconazole, itraconazole, voriconazole, and posaconazole, successfully treat fungal infections in humans. Efforts have been made to translate anti-fungal azoles into a second-use application for Chagas Disease. Ravuconazole and posaconazole have been recently proposed as candidates for clinical trials with Chagas Disease patients. However, the widespread use of posaconazole for long-term treatment of chronic infections may be limited by hepatic and renal toxicity, a requirement for simultaneous intake of a fatty meal or nutritional supplement to enhance absorption, and cost. To aid our search for structurally and synthetically simple CYP51 inhibitors, we have determined the crystal structures of the CYP51 targets in T. cruzi and T. brucei, both bound to the anti-fungal drugs fluconazole or posaconazole. The structures provide a basis for a design of new drugs targeting Chagas Disease, and also make it possible to model the active site characteristics of the highly homologous Leishmania CYP51. This work provides a foundation for rational synthesis of new therapeutic agents targeting the three kinetoplastid parasites.
| Chagas Disease, a potentially lethal tropical infection, is caused by the kinetoplastid protozoan Trypanosoma cruzi, which is spread by blood-sucking reduviid insects [1]. It is the leading cause of heart failure in Latin America, with an estimated to 8–10 million people infected [2]. The parasite invades and reproduces in a variety of host cells, including macrophages, smooth and striated muscle, fibroblasts and neurons. Disease progression is marked by an initial acute phase, which typically occurs in children, followed by a symptom-free intermediate phase. A chronic phase leading to GI tract lesions and heart failure often ensues. Current chemotherapy options are limited to nifurtimox and benznidazole, which have been in use since the late 1960s and are compromised by adverse side reactions and low efficacy in chronic disease [3], [4]. A need for drugs with more consistent efficacy and less toxicity is manifest.
With an essential requirement for ergosterol [5] and an inability to survive solely on cholesterol salvaged from the host, T. cruzi is vulnerable to inhibitors of the sterol biosynthesis enzyme 14α-demethylase (CYP51) [6], [7]. Disruption of CYP51 results in alteration in the ultrastructure of several organelles, decline of endogenous sterols in the parasites, and an accumulation of various 14α-methyl sterols with cytostatic and cytotoxic consequences [8]. The broad spectrum antifungal drug posaconazole (Noxafil; Schering-Plough) [9], which targets CYP51, is poised for clinical trials against T. cruzi [6], [10], [11]. Posaconazole is capable of inducing parasitological cure in a murine model of both acute and chronic Chagas Disease, curing between 50–100% of animals in the acute phase of infection, and 50–60% of animals chronically infected [7], [11]. However, the high manufacturing cost of posaconazole and the requirement for administration via oral suspension simultaneously with a fatty meal or nutritional supplement to enhance absorption may limit its use in treating chronic T. cruzi infections [12]. The search for CYP51-specific compounds that are easier to synthesize and better absorbed upon oral administration continues [13]–[17].
To rationalize protein-ligand interactions for new inhibitors in T. cruzi, homology modeling based on the x-ray structure of CYP51 from Mycobacterium tuberculosis (CYP51Mt) [18]–[20] has been used [14], [15], [17]. But CYP51Mt has only 27% sequence identity to the T. cruzi enzyme and is unusually exposed to the bulk solvent at the substrate binding site. This structural peculiarity largely excludes the functionally important BC-loop from protein-inhibitor interactions and thus limits the utility of CYP51Mt as a model for a Chagas Disease target. The CYP51 BC-loop residue 105 (numbering according to T. cruzi and T. brucei CYP51) is indispensable in the discrimination of the species-specific sterol substrates in T. cruzi and T. brucei [19], [21]. Also, a critical mutation hot spot [22], the well conserved BC-loop residue Y116 was reported to be involved in fungal drug resistance, inhibitor binding, and the catalytic function of CYP51 in Candida albicans (Y132, according to C. albicans numbering) [22]–[27], Histoplasma capsulatum (Y136, according to H. capsulatum numbering) [28], and in the causative agents of zygomycosis in humans, Rhizopus oryzae and Absidia corymbifera [29]. It may therefore play a similar role in T. cruzi.
Here we report the crystal structures for the CYP51 target in T. cruzi (CYP51Tc) (resolutions 2.35 Å and 2.27 Å) and that of the closely related CYP51 ortholog from Trypanosoma brucei (CYP51Tb) (resolutions 2.7 Å and 2.6 Å), each bound to an anti-fungal triazole drug, either fluconazole or posaconazole. T. brucei is a protozoan parasite closely related to T. cruzi [30] and the agent of another lethal tropical disease, African Sleeping Sickness. In contrast to T. cruzi and Leishmania spp., it is not clear if the sterol biosynthesis pathway can be targeted in T. brucei. Each parasite has a different life-cycle and different sterol requirements. Although the insect (procyclic) form of T. brucei can undertake de novo sterol biosynthesis, the latter is apparently suppressed in the bloodstream form in the mammalian host, which is supported by receptor-mediated endocytosis of host low-density lipoproteins that carry phospholipids and cholesterol esters [31]. Nevertheless, CYP51Tc and CYP51Tb do share 83% sequence identity, a fact which has been crucial for successfully determining their crystal structures and makes it possible to extrapolate structural features learned from one enzyme toward the other. Furthermore, the Leishmania CYP51 are 72–78% identical to that of T. cruzi and T. brucei, so they too can now be modeled to facilitate drug discovery and development.
By trial-and-error we empirically identified the protein N-terminal modification that eventually led to CYP51 crystals of sufficient quality to determine the x-ray structure. To improve our chances for success, we did the work in parallel on CYP51 proteins from Trypanosoma cruzi and Trypanosoma brucei. Five different expression vectors were designed in this work for each CYP51 ortholog to eliminate a stretch of hydrophobic residues which presumably mediate association of the proteins with the endoplasmic reticulum (ER). In their place we introduced hydrophilic or charged sequences at the N-terminus (Table 1). His6-tag (CYP51Tc) or His8-tag (CYP51Tb) was introduced at the C-terminus to facilitate purification. Coding sequences were sub-cloned between the NdeI and HindIII restriction cloning sites of the pCWori vector [32] and in this form used to transform Escherichia coli strain HMS174(DE3). The original coding sequence for CYP51Tb contained an internal NdeI site at 345 bp which was silenced by QuickChange site-directed mutagenesis (Stratagene) using forward GGGGTTGCCTATGCTGCC and reverse CCCCAACGGATACGACGG PCR primers. DNA amplification reaction: 5 min at 94°C, annealing for 1 min at 50–60°C, extension for 1.5 min at 72°C, for 30 cycles, followed by extension for 10 min at 72°C. The highest expression levels were achieved and the best crystals were obtained from the expression constructs modified by replacing the first 21 residues upstream of K22 with the fragment MAKKKKK. Subsequently, based on the analysis of the packing interactions in the crystal, three consecutive glutamate residues, E249-E251, were replaced in CYP51Tb with alanine by site-directed mutagenesis (Stratagene) using forward GCGCGGCTGCTGCTGTCAACAAGGACAGC and reverse GCGCGAGCAGCAGCCTTTCGAGCAATGAT PCR primers. DNA amplification reaction: 5 min at 94°C, annealing for 1 min at 45–65°C, extension for 1.5 min at 72°C, for 35 cycles, followed by extension for 10 min at 72°C. This CYP51Tb variant was used to generate the CYP51Tb-posaconazole crystals. The identity of all resulting vectors was confirmed by DNA sequencing.
Screening of crystallization conditions was routinely performed following purification of protein variants using commercial screening kits available in high throughput screening format (Hampton Research), a nanoliter drop-setting Mosquito robot (TTP Labtech) operating with 96-well plates, and a hanging drop crystallization protocol. Optimization of crystallization conditions, if required, was carried out manually in 24-well plates at 23°C. Proteins were from 1.0–1.8 mM frozen stocks in 20 mM Tris-HCl, pH 7.2 (CYP51Tb) or pH 8.0 (CYP51Tc), 10% glycerol, 0.5 mM EDTA, and 1 mM DDT. The CYP51Tb triple mutant E249A/E250A/E251A was used to obtain CYP51Tb-posaconazole crystals. Prior to crystallization proteins were diluted to 0.1–0.2 mM by mixing with 50 mM potassium phosphate at appropriate pH, supplemented with 0.5 mM (CYP51Tb) or 0.1 mM (CYP51Tc) fluconazole. Dilution in the absence of fluconazole or phosphate caused fast precipitation of protein samples. Posaconazole was prepared as 10 mM stock solution in DMSO and has been used at final concentration of 0.2 mM. Protein-posaconazole mix was incubated at 4°C for one hour prior to crystallization. Crystals of CYP51Tb–fluconazole complex grew from 15% ethylene glycol and 0–3% acetonitrile. Crystals of CYP51Tb–posaconazole complex grew from 6% PEG 4000, 2% tacsimate, pH 8.0, and 2% DMSO. Crystals of CYP51Tc–fluconazole grew either from 40% polypropylene glycol 400 and 0.1 M Tris-HCl, pH 6.0 (PDB ID 2WUZ), or from 25% PEG 4000 and 0.1 M Bis-Tris, pH 5.5 (PDB ID 2WX2), the latter being harvested directly from the Mosquito 200-nl drop. Prior to data collection, the crystals were cryo-protected by plunging them into a drop of reservoir solution supplemented with 20–24% ethylene glycol or 20% glycerol, and flash-frozen in liquid nitrogen.
All native and two-wavelength anomalous dispersion x-ray diffraction data were collected at 100–110 K at beamline 8.3.1, Advanced Light Source, Lawrence Berkeley National Laboratory, USA. Anomalous diffraction data were collected from one CYP51Tb crystal at two wavelengths, one corresponding to the median between the Fe peak and the inflection point and the other at 375 eV higher (Table 2). Data indexing, integration, scaling, phasing, and density modification were conducted using the ELVES automated software suite [34] (Tables 2 and 3).
CYP51Tb–fluconazole data processed in P3121 with Rmerge of 6.5% allowed for location of a single Fe atom. Initial phases with an overall figure of merit of 0.26 were improved by solvent flattening (mean figure of merit 0.85 after solvent flattening) to provide an interpretable electron density map (Table 2). Automated model building using BUCCANEER [35] placed the polyalanine backbone for 84% of the residues in the asymmetric unit. The remaining residues were built manually with COOT [36], alternated with TLS and positional refinement using REFMAC [37], [38]. The structure was refined to 3.2 Å with the R and Rfree values of 32.0% and 38.0%, respectively. Although showing up largely as a polypeptide backbone at low resolution, this T. brucei structure served as a search model for molecular replacement in determining the x-ray structure for CYP51Tc using 2.35 Å native data processed as P21 with Rmerge of 11%. Two CYP51Tc molecules were placed in an asymmetric unit. Manual model building with COOT [36] alternated with TLS and positional refinement using REFMAC [37], [38] resulted in the final CYP51Tc structure with the R and Rfree = values of 21.7% and 27.5% and the Ramachandran statistics of 93.8% residues in preferred regions, 5.2% in allowed regions, and 1% (9 residues) outliers, as calculated by COOT. NCS restrains were applied at all stages of the refinement. The refined CYP51Tc structure was used as a starting model against both the 2.27 Å native data for CYP51Tc and 2.7 Å native data for CYP51Tb, which allowed the majority of the CYP51Tb side chains to be built in. At that time, refinement of CYP51Tb converged with R and Rfree of 21.0% and 27.4%, respectively. Ramachandran statistics indicate 91.2% residues in preferred regions, 5.9% in allowed regions, and 2.9% (13 residues) outliers. Refinement of 2.27 Å CYP51Tc data converged with R and Rfree of 19.3% and 27.3%, respectively, and the Ramachandran statistics of 95.6% residues in preferred regions, 3.5% in allowed regions, and 0.9% (9 residues) outliers.
Analysis of the crystallographic symmetry packing interactions in the CYP51Tb-fluconazole complex revealed contacts between the triplets of glutamate residues D249-D251 situated in the GH-loop. To reduce electrostatic repulsion, all three residues were replaced with alanine. Although this modification did not improve resolution of the CYP51Tb-fluconazole crystals, the triple mutant was more amenable to co-crystallization with posaconazole. The CYP51Tb coordinates refined to 2.7 Å served as a search model to determine CYP51Tb-posaconazole structure using 2.6 Å native data processed as C2 with Rmerge of 8.5%. Four protein molecules were placed in an asymmetric unit. Refinement converged with R and Rfree of 19.1% and 26.4%, respectively and the Ramachandran statistics of 95.3% residues in preferred regions, 4.0% in allowed regions, and 0.7% (13 residues) outliers. In all structures, side chains not visible in the density were modeled as alanine (Table 3).
Binding of posaconazole to CYP51Tc was predicted by molecular docking using the 2WUZ structure. Docking was carried out using GLIDE (version 5.0) [39]. The docking protocol was validated by re-docking of fluconazole, which reproduced the binding mode observed in the crystal structure. The protein was initially prepared by the Protein Preparation Wizard module using default options. Hydrogen atoms were added to the complex structure, followed by a restrained minimization using the OPLS2005 force field. The Receptor Grid Generation module was then employed to prepare a rigid receptor grid centered at M360, which contains the entire binding tunnel of the energy minimized complex, for subsequent docking. The three-dimensional structure of posaconazole was generated by the Ligprep module with the OPLS2005 force field. Computational docking was performed using GLIDE in standard precision (SP) mode, and binding affinities were estimated as GLIDE score. The characteristic coordination between the heme group and the ligand was modeled by applying a constraint at the Fe3+ ion of the heme group that imposed interaction with one of the nitrogen atoms from the ligand's triazolyl ring. Since posaconazole (molecular weight = 700.8 g/mol) is significantly larger and longer than fluconazole (molecular weight = 306.3 g/mol) (Fig. 1), the van der Waals radii of the ligand were softened by a scaling factor of 0.6 in the initial docking calculation, which predicted two binding poses with similar Glide scores of −9.23 and −9.60. The binding model was further refined by relaxing the binding tunnel in the presence of posaconazole. Side chains of residues within 4 Å from the docked posaconazole were optimized by performing side-chain refinement with Prime (version 2.0) [40]. The resulting complex was used to re-dock posaconazole with van der Waals radii scaled by the default value of 0.8. Consistent with the initial calculations, the second-round docking also predicted the same binding orientations with favorable and similar GLIDE scores (pose 1 = −11.07; pose 2 = −10.70).
Protein Data Bank: coordinates and structure factors have been deposited with accession codes 2WUZ, 2WX2, 2WV2 and 2X2N.
By trial-and-error, the highest expression levels and best crystals for both CYP51Tc and CYP51Tb were obtained from the expression constructs modified by replacing the first 21 residues upstream of K22 with the highly positively charged fragment MAKKKKK (Table 1). The triple E249A/E250A/E251A CYP51Tb mutant was based upon this N-terminally modified construct. The UV-vis spectra of purified proteins revealed features characteristic for homogeneous and normally folded P450 (Fig. 2). We first determined the crystal structure for CYP51Tb using anomalous dispersion of the heme iron. Although largely a backbone trace at 3.2 Å resolution, this structure served as a search model for molecular replacement in determining the CYP51Tc–fluconazole structure at 2.35 Å, which was used as a search model against the 2.27 Å CYP51Tc–fluconazole data to reveal an alternative conformation of fluconazole bound in the active site. This same CYP51Tc structure was used as a model against the 2.7 Å CYP51Tb–fluconazole data. Refined to 2.7 Å CYP51Tb coordinates subsequently served as a search model for determining the 2.6 Å CYP51Tb–posaconazole structure.
CYP51Tc and CYP51Tb have a common P450 protein fold characterized by the sets of the α-helices and β-sheets highlighted in Fig. 3A, B. The T. cruzi and T. brucei structures superimpose with r.m.s.d. of 0.89 Å for Cα atoms, with the most pronounced differences in the region encompassing the F and G helices and the loop between them (Fig. 4A). By contrast, Trypanosoma CYP51 enzymes do not superimpose nearly as well with bacterial CYP51Mt (r.m.s.d. of 1.83 Å) (Fig. 4B), being more similar to their human counterpart (CYP51h), based on both backbone similarity (r.m.s.d. of 1.45 Å) and solvent exposure at the active site (Fig. 4C). All three eukaryotic enzymes lack the extreme bending of the I-helix that is associated with CYP51Mt, resulting in their active sites being more isolated from the bulk solvent. The structured BC-region in Trypanosoma CYP51 includes the B'-helix encompassed by the short η-helices blocking access to the active site from the bulk solvent. Seven residues from the BC-region, V102, Y103, I105, M106, F110, A115 and Y116, are part of the active site in CYP51Tc and CYP51Tb, which is consistent with our previous observation that a series of CYP51 inhibitors reported elsewhere [16] have higher binding affinities toward Trypanosoma CYP51 compared to CYP51Mt, where these residues do not participate in the active site due to the “open” conformation of the loop. The two CYP51Tc–fluconazole structures reported here superimpose with r.m.s.d of 0.68 Å, revealing some conformational differences in the F-helix and the BC-loop, which may account for the distinct fluconazole binding modes and result in re-packing of protein molecules in the crystal lattice (Table 3).
In the CYP51Tb-posaconazole structure, four protein molecules in the asymmetric unit superimpose with the r.m.s.d. within of 0.5 Å, revealing virtually no conformational variations. However, posaconazole samples two distinct conformations due to the long tail swinging ∼7–8 Å in the hydrophobic mouth of the substrate binding tunnel (Fig. 3B). The entrance to the tunnel is marked by a patch of the hydrophobic residues (colored yellow in Fig. 3B), which apparently guide access of the sterol substrates to the active site.
As expected, fluconazole is bound in the active site by coordination to the heme iron via the aromatic nitrogen atom of a triazole ring and by multiple van der Waals and aromatic stacking interactions (Fig. 5). All residues within 7 Å of fluconazole (Fig. 6A) are labeled with blue triangles in Fig. 7. The 2,4-difluorophenyl moiety is enclosed in the pocket formed by the heme macrocycle, the aromatic side chains of Y103, F110, Y116 (BC-loop) and F290 (I-helix), and aliphatic side chains M106, A287 and A291. Although fluconazole occupies the same pocket in both CYP51Tc structures, it adopts two conformations that differ by the 180° flipping of the 2,4-difluorophenyl moiety. Orientation 1 is observed both in the 2.27 Å CYP51Tc–fluconazole structure reported in this work (PDB ID 2WX2) (Fig. 5A) and in the CYP51Mt-fluconazole complex reported elsewhere (PDB ID 1EA1) [18]. The same conformation is adopted by the 2,4-difluorophenyl ring of posaconazole in the CYP51Tb-posaconazole complex in all four molecules in the asymmetric unit. In orientation 1, Y103 makes a 2.7 Å H-bonding contact to the main chain amide group of M360. A 180° flipped orientation of the ring, orientation 2, is observed in the 2.35 Å CYP51Tc–fluconazole structure (PDB ID 2WUZ) (Fig. 5B). As evidenced by the residual Fo-Fc electron density map calculated for the orientation 1 (pink mesh in Fig. 5B), the 2-fluoro substituent of the fluconazole difluorophenyl ring in 2WUZ must point toward the heme macrocycle. A 2.6 Å H-bonding contact between the 2-fluoro substituent and the hydroxyl group of Y103 may help to stabilize orientation 2, which appears to be less sterically favorable than orientation 1.
Perhaps both ring conformations co-exist in the CYP51Tc-fluconazole complex, possibly correlated with the conformation of the BC-loop which affects H-bonding pattern of Y103. In orientation 1, the H-bond between the 2-fluoro substituent and Y103 is broken due to the 3.5 Å reorientation of Y103 toward M360 resulting in the 2.7 Å H-bond to its amide nitrogen and in the flipping of the 2,4-difluorophenyl ring into a sterically more favorable orientation with the fluorinated edge facing away from the heme macrocycle (Fig. 5B). Crystallization conditions may have served to shift the equilibrium by stabilizing one of these states. The entire fluconazole molecule is shifted about 1.5 Å between the two CYP51Tc structures, which may be related to the low efficacy of this drug against T. cruzi. Given that the CYP51Tb active site is virtually identical to that of T. cruzi, the same equilibrium would be expected to occur in T. brucei. However, we could not observe this phenomenon as CYP51Tb-fluconazole complex has been co-crystallized under a single set of conditions with one molecule in the asymmetric unit.
The CYP51Tc structures revealed a 42 residue-long hydrophobic tunnel connecting the chamber adjacent to the heme with the protein surface (Fig. 6B). Residues constituting the tunnel in addition to those interacting with fluconazole are labeled with green triangles in Fig. 7. The mouth of the channel is surrounded by residues I45, I46, G49, K50, I209, P210, H458, and M460, which may delineate the substrate/inhibitor entry site in eukaryotic CYP51 (Fig. 8). This entry mode would be in contrast to that in CYP51Mt, where it most likely occurs through the open BC-loop. The tunnel-forming residues are invariant between CYP51Tc and CYP51Tb with the exception of four conservative substitutions at positions 46, 105, 215 and 359. The residue at position 105 (highlighted cyan in Fig. 7) is known to dominate substrate specificity with respect to the methylation status of the C-4 atom in CYP51 sterol substrates. I105 in T. cruzi allows efficient conversion of C4-dimethylated 24-methylenedihydrolanosterol while the bulkier F105 in T. brucei favors C4-monomethylated norlanosterol [19], [21]. Phenylalanine in CYP51Tb protrudes further into the active site than isoleucine in CYP51Tc (Fig. 6A), potentially resulting in interference with the 4β-methyl group of the sterol substrate. However, F105 does not interfere with either posaconazole or fluconazole binding.
Comparison of the residues constituting the tunnel in CYP51Tc with the human counterpart, CYP51h, indicates that two residues, H236 and H489 (numbered according to the human sequence and highlighted yellow in Fig. 7), protrude into the tunnel near the opening, reducing both its size and hydrophobicity. As they are present exclusively in mammalian orthologues [41], H236 and H489 may partly account for the selectivity of azole drugs toward pathogenic fungi and protozoa. In accord with this hypothesis, proline corresponding to H236 in pathogenic fungi is among hot spots that confer resistance to posaconazole in Aspergillus fumigatus (P216) [42] and Candida albicans (P230) [43].
The hydrophobic tunnel in CYP51Tb accommodates the antifungal drug posaconazole in either extended or bent conformations. The 2.6 Å structure of the CYP51Tb-posaconazole complex revealed four protein monomers in an asymmetric unit with posaconazole coordinating to the heme iron in a manner similar to that of fluconazole with the fluorinated edge of the 2,4-difluorophenyl ring facing away from the heme macrocycle, and the long substituent tail extending into the hydrophobic tunnel. Electron density is well defined for the Fe-coordinating head of the posaconazole molecule in all four monomers but somewhat fades out toward its long tail (Fig. 8A). Thus, the terminal 2-hydroxypentan group is defined in none of four monomers. Three from the four monomers (chains A, B, and C) accommodate posaconazole in bent conformation while in the monomer D posaconazole is in extended conformation. Conformational variability of posaconazole is enabled by the interconversion of the piperazine six-membered ring between the chair and twisted boat conformations. The latter serves to accommodate the bend. Electron density is best defined in monomer B, where the terminal phenyl-2-hydroxypentan-triazolone group of posaconazole lies within 6 Å of protein residues I209-P210-A211 and V213-F214 which are invariant between CYP51Tc and CYP51Tb. P210, the mutation hot spot in fungi, is situated right in the bend of the posaconazole molecule (Fig. 9A). In the extended conformation in monomer D, the phenyl-2-hydroxypentan-triazolone group swings toward residues I45-I46 (Fig. 9B). Remarkably, points of posaconazole contact in the tunnel mouth are among mutation hot spots in azole resistant isolates of pathogenic fungi A. fumigatus [42], [44]–[48] and C. albicans [43], [49] (Fig. 8B and C).The scattered Fo-Fc electron density map in the monomers A and D (Fig. 8A) suggests possible interconversion of the posaconazole conformers in dynamic equilibrium, meaning that the phenyl-2-hydroxypentan-triazolone group dangles in space within the tunnel mouth (Fig. 3D). Given the high sequence and structural similarities between CYP51Tc and CYP51Tb, similar dynamics would be expected in the CYP51Tc-posaconazole complex.
The x-ray structures of the CYP51 therapeutic targets determined in this work are intended for use in rational drug design. We also apply computational methods to explore binding modes of known chemical structures as well as to generate new scaffolds based on the configuration of the CYP51 binding sites. Considering the differential geometries of the host and pathogen binding sites, we aim to develop a pool of highly selective molecules with no cross-reactivity to human CYP51. As a first step, we docked posaconazole into the CYP51Tc active site and compared the docking poses with the experimental structure of CYP51Tb-posaconazole complex. Two poses with similar docking scores were identified for posaconazole by GLIDE [39], differing primarily in the orientation of the 2,4-difluorophenyl ring (Fig. 10). Interestingly, the long posaconazole tail docks in a mode more similar to the CYP51Tb-posaconazole complex defined in this work rather than that in the recently deposited T. cruzi structure (PDB ID Code: 3K1O). Given that the protein-posaconazole interactions in the tunnel are of hydrophobic/aromatic stacking nature (Fig. 9), this ambiguity is not surprising.
Another source of docking ambiguity arises from the binding predicted for the 2,4-difluorophenyl substituent. In the better scoring pose 1 (highlighted yellow in Fig. 10), the 2,4-difluorophenyl ring binds in the experimentally observed orientation 1. In the slightly lower scoring pose 2 (highlighted pink), the 2,4-difluorophenyl ring is bound in a different pocket formed by the residues M106, E205, L208, F290, T295, L358 and M460, suggesting an additional cavity in the CYP51 active site suitable for drug targeting. This pose is achieved via flipping of the central furan ring to which all the substituents are attached. Thus, in addition to the experimentally observed binding ambiguity of the long substituent tail, conformational ambiguity of the difluorophenyl ring is predicted by the docking calculations and perhaps will be observed in future structures of CYP51 in complex with inhibitors similar to posaconazole.
The rapid development of azole resistance in T. cruzi observed in vitro suggests that the same may occur in patients [50]. Although no data are available on the development of posaconazole resistance in Chagas Disease patients, studies conducted on fungal infections indicate that posaconazole resistance occurs mainly by a mechanism involving mutation of the cyp51 gene [42], [51], [52]. Posaconazole appears to be less susceptible to the efflux pumps that confer resistance to some other azoles [43], [51], [53]. Mapping mutations in cyp51 genes in clinical posaconazole resistant isolates on the CYP51 structure, points to the tunnel entrance as a mutation hot spot. Mutations of G54, P216 and M220 in clinical isolates of A. fumigatus [42], [44]–[48] (corresponding to G49, P210 and F214, respectively, in CYP51Tc and CYP51Tb) and of A61 [49] and P230 [43] in clinical isolates of C. albicans (I45 and P210, respectively, in CYP51Tc and CYP51Tb) map directly to the tunnel mouth (Fig. 8B and C). Mutations of G54 in A. fumigatus to arginine or tryptophan associate with moderate and high levels of resistance, respectively, and confer cross-resistance between itraconazole and posaconazole [44]. Mutations of M220 confer cross-resistance to all azole drugs tested, including itraconazole, voriconazole, ravuconazole and posaconazole [54], [55] and therefore may interfere with the entry of the drugs. In accord with this assumption, posaconazole is reported to induce resistance to all azole drugs in Candida parapsilosis in vitro [51]. The alarming perspective emerging from antifungal therapy efforts must be taken into consideration when designing anti-Chagasic drugs targeting CYP51Tc. Thus, the terminal phenyl-2-hydroxypentan-triazolone group in posaconazole may play an important role in pharmacokinetics rather than in the interactions with the target, and yet these interactions seem to induce resistance which otherwise could probably be avoided.
In summary, the x-ray structures of Trypanosoma CYP51 enzymes reported here open new opportunities for rationally designed inhibitors against therapeutic targets in important human pathogens. The structures provide templates for developing CYP51 inhibitors with improved efficacy and resistance properties that are structurally and synthetically simpler than posaconazole. By utilizing the differential geometries between host and pathogen CYP51 binding sites, it maybe possible to create new drugs with minimized toxicity and host-pathogen cross-reactivity. In addition, the posaconazole binding mode offers insights into the development of drug resistance in pathogenic fungi, implying that an analogous mechanism may be implicated in protozoan pathogens. The reported structures also provide a good template for drug design targeting Leishmania CYP51. However, drug development must take into account the properties and accessibility of the compartment where these parasites reside. Unlike T. cruzi, Leishmania amastigotes replicate in the acidic environment (pH ∼5) of the phagolysosomal vacuoles in macrophage cells [56], [57], imposing different requirements on the physicochemical properties of CYP51 inhibitors targeting leishmaniasis.
|
10.1371/journal.pgen.1005717 | An Empirical Bayes Mixture Model for Effect Size Distributions in Genome-Wide Association Studies | Characterizing the distribution of effects from genome-wide genotyping data is crucial for understanding important aspects of the genetic architecture of complex traits, such as number or proportion of non-null loci, average proportion of phenotypic variance explained per non-null effect, power for discovery, and polygenic risk prediction. To this end, previous work has used effect-size models based on various distributions, including the normal and normal mixture distributions, among others. In this paper we propose a scale mixture of two normals model for effect size distributions of genome-wide association study (GWAS) test statistics. Test statistics corresponding to null associations are modeled as random draws from a normal distribution with zero mean; test statistics corresponding to non-null associations are also modeled as normal with zero mean, but with larger variance. The model is fit via minimizing discrepancies between the parametric mixture model and resampling-based nonparametric estimates of replication effect sizes and variances. We describe in detail the implications of this model for estimation of the non-null proportion, the probability of replication in de novo samples, the local false discovery rate, and power for discovery of a specified proportion of phenotypic variance explained from additive effects of loci surpassing a given significance threshold. We also examine the crucial issue of the impact of linkage disequilibrium (LD) on effect sizes and parameter estimates, both analytically and in simulations. We apply this approach to meta-analysis test statistics from two large GWAS, one for Crohn’s disease (CD) and the other for schizophrenia (SZ). A scale mixture of two normals distribution provides an excellent fit to the SZ nonparametric replication effect size estimates. While capturing the general behavior of the data, this mixture model underestimates the tails of the CD effect size distribution. We discuss the implications of pervasive small but replicating effects in CD and SZ on genomic control and power. Finally, we conclude that, despite having very similar estimates of variance explained by genotyped SNPs, CD and SZ have a broadly dissimilar genetic architecture, due to differing mean effect size and proportion of non-null loci.
| We describe in detail the implications of a particular mixture model (a scale mixture of two normals) for effect size distributions from genome-wide genotyping data. Parameters from this model can be used for estimation of the non-null proportion, the probability of replication in de novo samples, the local false discovery rate, power for detecting non-null loci, and proportion of variance explained from additive effects. Here, we fit this model by minimizing discrepancies with nonparametric estimates from a resampling-based algorithm. We examine the effects of linkage disequilibrium (LD) on effect sizes and parameter estimates, both analytically and in simulations. We validate this approach using meta-analysis test statistics (“z-scores”) from two large GWAS, one for Crohn’s disease and the other for schizophrenia. We demonstrate that for these studies a scale mixture of two normal distributions generally fits empirical replication effect sizes well, providing an excellent fit for the schizophrenia effect sizes but underestimating the tails of the distribution for Crohn’s disease.
| While genome-wide association studies (GWAS) have discovered thousands of genome-wide significant risk loci for heritable disorders, including Crohn’s disease [1] and schizophrenia [2], so far even large meta-analyses have recovered only a fraction of the heritability of most complex traits. Some of this “missing heritability” may be due to rare variants of large effect, epistasis, copy-number variation, epigenetics, etc. However, recent work utilizing variance components models [2–5] has demonstrated that a much larger fraction of the heritability of complex phenotypes is captured by the additive effects of SNPs than is evident only in loci surpassing genome-wide significance thresholds. Thus, the emerging picture is that traits such as these are highly polygenic, and that the heritability is largely accounted for by numerous loci each with a very small effect [5, 6]. In this scenario, instead of estimating effect sizes individually, it is useful to characterize the distribution of effect sizes for choosing significance thresholds, for estimation of power, for the computation of an individual’s overall genetic risk for a disease, and for the identification of disease mechanisms that can be used for the development of effective treatments.
Effect size distributions can be estimated directly from the genotype-phenotype data [3, 7–10] or from the summary statistics produced from GWAS analyses [11, 12]. In this paper we focus on estimation of effect size distributions from summary statistics, produced from fitting a regression model for each single nucleotide polymorphism (SNP) individually. A Wald test statistic (“z-score”) is computed from the regression of each SNP to test its association with the phenotype of interest. A SNP is often declared significant if the p-value of its test statistic surpasses a Bonferroni-inspired threshold of 5 × 10−8. Note, within this typical GWAS hypothesis testing framework, the effect size for a given SNP computed from massively univariate test statistics is a weighted combination of effects from all SNPs that it is in linkage disequilibrium (LD) with (see [13] as well as S1 Text for more details).
An implicit assumption in GWAS hypothesis testing is that SNP test statistics come from a mixture distribution of zero (null) and non-zero (non-null) effect sizes [14], though this mixture distribution is not usually explicitly modeled. The values of parameters from such a mixture distribution characterize important aspects of the genetic architecture of a phenotype, including the proportion of non-null effects, the variance explained per non-null locus, and the amount of inflation in the null distribution [15]. Mixture model parameters can also be used to compute other quantities of interest, including estimates of the probability of replication in a de novo study, the posterior probability that a given SNP is null or has a negligible effect conditional on its observed z-score (i.e., the local false discovery rate), and the power to detect susceptibility loci for a given study sample size. These parameters are also closely related to the proportion of the phenotypic variance explained by the additive effects of common variants and upper limits on the accuracy of polygenic risk scores [12, 16]. Information such as LD or the functional role of SNPs can be incorporated into the model to provide characterizations of the genetic architecture of complex disorders that do not implicitly assume that all SNPs are a priori exchangeable [17, 18].
In this paper we implement a simple scale mixture of two normals distribution to model GWAS z-scores. Test statistics corresponding to “null” associations are modeled as random draws from a normal distribution with zero mean; test statistics corresponding to “non-null” associations are also modeled as random draws from a normal distribution with zero mean but with larger variance. The proportion of tests corresponding to null associations is also estimated. (This model has a Bayesian interpretation, and the methods proposed are “empirical Bayes” because the prior probability of being null is estimated from the data [19].) A closely related model has been previously proposed for GWAS effect sizes using genotype-phenotype data [10].
We derive the connection between this mixture model and the finite-sample probability of replication in de novo samples, the local false discovery rate, and the power for detecting a specified proportion of the phenotypic variance due to additive effects of genetic loci for a given local false discovery rate. The mixture model is fitted using a resampling-based procedure applied to meta-analysis sub-study z-scores. By repeatedly and randomly partitioning the sub-studies into disjoint training and replication samples, we obtain nonparametric smoothed estimates of replication effect sizes and variances that are scaled estimates of their conditional posterior expectations (given the observed z-scores) with respect to a simple measurement model. We then fit a parametric scale mixture of two normals models that minimizes the sum of squared discrepancies with these nonparametric estimates.
We demonstrate this statistical framework in simulations and on meta-analysis z-scores from Crohn’s disease [1] and schizophrenia [20] GWAS. We show that the scale mixture of two normals model provides an excellent fit to the posterior effect size means and variances for the schizophrenia data, while capturing the general behavior (though underestimating the tails of the effect size distribution) for Crohn’s disease. We conclude that, despite having very similar estimates of variance explained by genotyped SNPs, Crohn’s disease and schizophrenia have a broadly dissimilar genetic architecture due to differing mean effect size and proportion of non-null loci. Finally, we examine the effects of LD on effect size distributions estimated from GWAS summary statistics, both analytically and in simulation studies.
Crohn’s disease (CD) is a type of inflammatory bowel disease that is caused by multiple factors in genetically susceptible individuals. Estimates of narrow-sense heritability for CD are h2 ≈ 0.50 [21]. The variance captured by the additive effects of genotyped SNPs using a liability model assuming an underlying normal distribution for additive per allele risk effects has been estimated at h chip 2 = 0 . 22 [22]. The CD data consist of N = 942,772 SNP z-scores from a GWAS meta-analysis of eight sub-studies on a total of n = 23,671 subjects (7,352 cases) [1]. Sub-study z-scores are available at http://www.ibdgenetics.org/downloads.html. Before running the resampling algorithm, SNPs were randomly pruned for approximate independence, so that LD ≤ 0.20 between any pair of SNPs, resulting in N = 97,855 SNPs.
Fig 1 shows the resampling means and variances of replication z-scores as a function of training z-scores for the CD meta-analysis sub-studies, based on all 70 possible partitions of sub-studies into four training and four replication datasets. Also plotted are the predicted replication conditional means and variances from the best fitting scale mixture of two normals model. The nonparametric and model-based estimates show good agreement except in the tails (absolute discovery z-scores > 3). Lack of fit is due to larger effect sizes in the tails than is predicted by the mixture model. Stated differently, the distribution of effect sizes has a larger kurtosis than can be captured by the two-component mixture. This results in conservative estimates of replication effect sizes, replication probabilities, and local fdr for SNPs in this part of the distribution. Other authors have proposed a scale mixture including more than two components (e.g., [10]), which could be implemented within our resampling-based algorithm at the cost of two parameters per additional mixture component.
The estimated non-null proportion is π ^ 2 = 0 . 0008 indicating that almost 0.1% of the 97,855 approximately independent SNPs fall within in the “large effects” category. The standard deviation for small effects is σ ^ 1 = 0 . 008, and the standard deviation for large effects is σ ^ 2 = 0 . 078. The estimated null standard deviation is σ ^ 0 = 0 . 991, or slightly below the theoretical null standard deviation. Note, the “empirical null” variance [23] is approximately given by σ ^ 0 2 + 2 p ¯ ( 1 - p ¯ ) n σ ^ 1 2 = 1 . 08, where n is the effective sample size of the study and p ¯ is the mean minor allele frequency. As indicated by the small but non-zero estimate of σ1, there is a positive slope through the origin in the plot of replication effects (upper left panel of Fig 1), indicating that even very small z-scores tend to replicate at a higher rate than expected by chance. Thus, it is more appropriate to state that replication z-scores show a mixture of “small” and “large” replicating effects rather than “null” and “non-null”. Small replicating effects could potentially be due to population stratification or to weak yet pervasive LD with causal effects (see S1 Text).
The estimated number of large effect SNPs among the 97,855 is given by N π 2 ^ = 76. There are 45 SNPs declared significant using a local fdr threshold of 0.05, which corresponds to SNPs with p-values ≤ 9.8 × 10−8. Thus, the CD meta-analysis is currently powered to detect approximately 60% of large effect SNPs using a local fdr threshold of 0.05.
Note, the presence of correlation among genetic loci due to LD is important for the interpretation of parameters in the mixture model. For example, the proportion of large effects π2 is dependent on the level of pruning, with π2 being larger in unpruned data and lower in data pruned for approximate independence. This is because large effects tend to be in higher total LD with other SNPs, and hence a higher proportion of these are eliminated during random pruning. One explanation why large-effect SNPs tend to have higher total LD is that these SNPs tag larger genomic regions and hence have a higher probability of tagging causal effects (see [13] and the S1 Text). Another possible explanation, not mutually exclusive with the first, is that SNPs that fall in functional genomic categories (e.g., within genes) are enriched for causal effects and that these categories also tend to be in regions of higher total LD [17, 18]. The balance between these two explanations determines how much π2 is over-estimated using unpruned loci or under-estimated using loci pruned for independence, relative to the underlying and unknown proportion of causal effects. While we perform random pruning to approximate independence here, the efficient and accurate handling of the effects of LD-induced correlation and blurring of effect size distributions is an area of on-going research.
Schizophrenia (SZ) is known to be highly polygenic and has an estimated narrow-sense heritability h2 ≈ 0.8 [24]. The additive variance captured by SNPs using a liability model has been estimated at h chip 2 = 0 . 23 [25], close to that of CD. The SZ data analyzed here consist of N = 2,558,411 association z-scores from a GWAS meta-analysis of 52 sub-studies with n = 82,315 total subjects (35,476 cases) [26]. The full study meta-analysis statistics are available at http://www.med.unc.edu/pgc/downloads. PGC analytic datasets can be obtained by application to the controlled-access NIMH Genetics Repository. Data were randomly pruned for pairwise LD ≤ 0.20, leaving N = 129,973 roughly independent SNPs. The resampling procedure was run over 100 iterations, with random splits of the sub-studies into differing proportions (30%,40%, and 50%) for training and the remaining proportion as replication data.
Fig 1 shows the empirical replication means and variances of z-scores, as a function of training z-scores, for the SZ meta-analysis sub-studies, based on the split-half samples. The predicted replication conditional means and variances show an excellent fit to the nonparametric estimates. The estimated non-null proportion is π ^ 2 = 0 . 012, indicating that about 1.2% of the pruned SNPs are in the large effect class. Thus, in terms of the proportion of large effect SNPs in pruned data, SZ is almost fifteen times more polygenic than CD. At the current effective sample size there are 15 SNPs with local fdr ≤ 0.05, or 1% of the estimated N π ^ 2 = 1,516 large-effect SNPs. The null standard deviation is estimated to be σ ^ 0 = 1 . 01, very close to the theoretical null. The standard deviation for large effects is σ ^ 2 = 0 . 020. Despite being more polygenic, large effect SNPs in SZ on average account for only 7% of the phenotypic variance accounted for by large effect SNPs in the CD data.
The standard deviation for small effects is σ ^ 1 = 0 . 007, and hence the empirical null variance is approximately σ ^ 0 2 + 2 p ¯ ( 1 - p ¯ ) n σ ^ 1 2 = 1 . 32. Since σ ^ 1 > 0, as with CD there is a positive slope through the origin of the replication z-scores as a function of discovery z-scores (upper right panel of Fig 1) which scales with the size of the training sample (see S1 Fig). This is in contrast to what would be expected if the observed z scores were a mixture of true null (exactly zero) and non-null (non-zero) effects (S2 Fig), in which case there would be no positive slope through the origin.
For the SZ data, parameter estimates from the scale mixture of normals model were used to compute the probability that a SNP will replicate given its observed training z-score, as given in Eq (15). Fig 2 displays the resampling-based replication rate and model-based replication probabilities for the CD and SZ meta-analyses, for resampling performed using 30% and 50% of the data in the training sample and the remainder in the replication sample. Fig 2 shows good agreement of the resampling-based replication rates with the mixture model-based replication probabilities for SZ. For CD, model-based replication probabilities underestimate the resampling-based replication rates in the tails, again due to excess kurtosis not captured by the two scale mixture components.
The results displayed in Fig 2 do not constitute a true replication analysis, since the entire set of 52 studies was used to estimate the mixture model parameters. To assess true replication, we divided the sub-studies into disjoint “discovery” and “replication” samples. For the discovery sample, we computed the meta-analysis z-scores and local fdrs using summary statistics from 26 randomly selected sub-studies, consisting of 17,691 cases and 24,683 controls on the same set of N = 129,973 SNPs pruned to pairwise LD ≤ 0.20. For the replication sample we computed the meta-analysis z-scores using the remaining 26 studies, with 17,785 cases and 22,156 controls. We defined replication for a locus as having a one-sided replication p-value ≤ 0.05 and discovery and replication z-scores having the same sign. Other definitions of replication can be easily implemented. Replication proportions and mean predicted replication probabilities using Eq (15) are displayed in Fig 3. While replication proportions are noisy due to small numbers of SNPs in most fdr bins ([0, 0.1): 6, [0.1, 0.2): 0, [0.2, 0.3): 1, [0.3, 0.4): 3, [0.4, 0.5): 3, [0.5, 0.6): 7, [0.6, 0.7): 10, [0.7, 0.8): 31, [0.8, 0.9): 132, [0.9, 1.0]: 129,780), they generally track the predicted replication probabilities, showing some evidence, however, that predicted replication probabilities may be somewhat lower than actual replication rates. A downward bias in predicted replication probabilities could be caused by under-fitting the extreme tails of the distribution; this could potentially be rectified by adding one or more normal mixtures over the current two.
For a given threshold it is possible to estimate the proportion of posterior expected additive variance explained by SNPs selected using a given significance threshold. Let c > 0 be a given significance threshold, so that any SNP |Z| ≥ c is declared significant. Let zi and δi denote the Wald statistic of the ith SNP with effect size δi as given in Eq (3). The proportion of genetic variance explained by these SNPs based on the scale mixture of two normals model is approximately
h c 2 ≈ ∑ | z i | ≥ c E ^ { δ i 2 ∣ Z i = z i } ∑ i = 1 N E ^ { δ i 2 ∣ Z i = z i } (1)
where E ^ { δ i 2 ∣ Z i = z i } is estimated via Eq (13), substituting estimates θ ^ for θ. This estimate relies on the assumption that the average LD of SNPs declared significant is roughly the same as the average LD of all SNPs, or that SNPs are first pruned for approximate independence. We can also modify Eq (1) to give the proportion of variance due to large effects accounted for by SNPs declared significant
h c , 1 2 ≈ ∑ | z i | ≥ c E ^ 1 { δ i 2 ∣ Z i = z i } ∑ i = 1 N E ^ 1 { δ i 2 ∣ Z i = z i } (2)
where E1 denotes the posterior expectation due to large effects [27].
Using the parameters from the model-based fits, we can compute power for discovery when SNPs are declared non-null based on local fdr or p-value cut-offs. It is convenient to express power as the proportion of the genetic variance due to additive effects discovered for a given threshold. For example, the 45 SNPs with fdr ≤ 0.05 in the pruned CD data account for 55% of the genetic variance due to additive common effects in the pruned sample, including both large and small replicating effects. However, these loci account for 83% of variance due to large effects alone. Power estimates for CD are conservative, since the tails of the distribution are somewhat underestimated by the mixture of two normals model. In the SCZ data, the 15 SNPs with fdr ≤ 0.05 account for 3% of the variance due to the additive effects of all common variants, but 34% of the variance due to large effects alone. The difference in power between the two disorders is due to the more polygenic nature of SZ compared to CD, combined with its much smaller average size per “large-effect” SNP.
Fig 4 displays the power for discovery for a genome-wide significance threshold of p ≤ 5 × 10−8 for increasing effective sample sizes for both CD and SZ. The z-scores are corrected using λGC as defined in [28]. For example, for CD the current sample size results in 69% of the variance due to large effect discovered; doubling the sample size for CD would result in the discovery of almost 91%. In contrast, using the same threshold for SZ, the current sample size uncovers SNPs accounting for only 26% of the large effect variance. The sample size would have to be increased 32-fold to detect 90% of the variance due to large effects, despite the fact that the current sample size of the SZ study is already much larger than that of the CD. One reason for the slow increase in power is that the median of the z2 distribution is inflated by both small and large effect variances, and hence the genomic inflation factor λGC [28] grows as a function of effective sample size n.
We conducted a series of Monte Carlo simulation studies to evaluate the performance of the fitting algorithm under different values of the parameters and departures from the standard meta-analysis assumptions (I)-(III) (see Models section) on the nonparametric estimates given in Eq (16) as well as the scale mixture of normals model parameters θ = {π1, σ0, σ1, σ2}, where π1 is the proportion of small effects, and σ0, σ1, and σ2 are the standard deviations of the null, small effect, and large effect (normal) distributions, respectively, as given in Eq (11). The results of these simulations are presented in S3–S7 Figs in the S1 Text section.
The estimates θ ^ produced by minimizing the quadratic estimating equations given in Eq (18) are in general unbiased and exhibit low variability across iterations of the simulations for a wide variety of parameter settings (S3 Fig). In S4 Fig and S6 Fig, we show the impact of large random departures from assumption (II): common minor allele frequencies (MAFs) across sub-studies; large departures from assumptions (I) and (III) will have similar effects. In these simulations, the estimated non-null proportion π ^ 2 is largely unaffected, σ ^ 0 is slightly elevated, and σ ^ 1 and σ ^ 2 are substantially decreased from the true values. In the scenario of large random departures from the overall mean values of the parameters, a random effects meta-analysis is more appropriate [29].
In the S1 Text section we also present simulations demonstrating the effects of LD on the distribution of effect sizes produced from massively univariate regression analyses typical of most GWAS. As described in [13] and in the S1 Text, LD “blurs” effect sizes from multiple loci, i.e., the expected effect size of a given locus produced from a univariate regression is a weighted sum of effects from all loci it is in non-zero LD with.
In this paper we derive the connection between a simple (four parameter) scale mixture of two normals model for effect size distributions and several quantities of interest in genome-wide studies. Specifically, parameter estimates from such a mixture model can be used to compute the proportion of genotyped SNPs with “large” effects, the local false discovery rate, probability of replication in a de novo sample, and power for discovery expressed as proportion of chip heritability explained for a given sample size and significance threshold. Effect size estimates can also be used for applications such as computation of polygenic risk for disorders (see S1 Text for how posterior effect sizes can be used in this fashion). Estimated effect sizes are shrunk empirically via the resampling process, and hence are free from the Winner’s Curse.
Direct observation demonstrates that for the schizophrenia GWAS data the scale mixture of two normals model provides a very good fit to nonparametric replication z-scores. The fit to the Crohn’s disease data is not as good, since the tails of the distribution are underestimated. This can be remedied by adding more components to the scale mixture, with two additional parameters per component. Derivations of local fdr, replication probabilities, and power presented in the Models section can be extended to more than two components. Underestimating the tails of the effect size distribution leads to conservative estimates of replication probabilities, local fdr, and power for discovery.
An interesting aspect of using the resampling-based fitting procedure is the ability to separate the null standard deviation σ0 from the standard deviation σ1 of small but replicating effects, which are confounded in non-resampling based fitting algorithms for mixture models employing the “empirical null” (e.g., [23]). Small replicating effects which scale with sample size could potentially be due to residual population stratification or to weak yet pervasive LD with causal effects. The later case would suggest that weak LD with causal variants may be a significant source of variation in tests statistics, as discussed in [13]. (Note, however, that [13] does not model the distribution of effect sizes and hence does not assess differential effects of LD on null vs. non-null loci.) An important consequence of the presence of small and large effects whose variances scale linearly with effective sample size is that the genomic inflation factor λGC [28] also grows as a function of sample size. It has been argued that the distribution of non-null effects substantially accounts for the observed genomic inflation in large GWAS [15, 26]. While our results are consonant with this fact, we here make a more fine-grained distinction between genomic inflation due to small and that due to large replicating effects. To the degree that small effect inflation is considered spurious, performing no genomic inflation control whatsoever would appear to be overly liberal.
A weakness of the resampling procedure is that the quadratic estimating equations do not produce accurate confidence intervals for parameter estimates. This is due to the complicated correlation structure among terms in the estimating equations induced by the presence of LD in the SNPs and by the overlap in randomly resampled estimates. In theory it is possible to obtain the overall effective degrees of freedom of the estimates by computing the mean induced correlation which can then be used to adjust the length of standard confidence intervals. Non-resampling based mixture model algorithms also exist that estimate the non-null distribution using likelihood-based flexible regression fits (e.g., see [23]), and we are currently developing a fully Bayesian alternative that models the non-null distribution as a location mixture of B-spline densities with mixture weights that can depend on LD and multiple genic annotation categories. These non-resampling based algorithms can provide accurate confidence intervals for parameters assuming the data are first pruned for approximate independence.
Another disadvantage of the proposed algorithm is that splitting studies into disjoint training and replication sets leads to lower power to estimate the non-null component of the mixture when the sample size is small, where “small” depends on the level of polygenicity and the average size for non-null effect. As such, the resampling-based algorithm depends on a fairly sizable signal in the GWAS data so that the parameters π2 and σ2 can be estimated.
In general, it is crucial to consider the impact of LD on the massively univariate regression estimates common to standard GWAS analyses, since regression weights b ^ have expectations that depend heavily on the LD structure (see S1 Text). In particular, the expectation of b ^ i is equal to the causal effect of the ith SNP plus a weighted sum of all the causal effects it is in LD with (see S1 Text for details and [13]). The effects of LD on nonparametric estimates of the effect size distribution, and hence also on estimates of parameters from the scale-mixture of normal model, can be profound. Simulations (S5 Fig and S7 Fig) also show an over 20-fold increase in π2 estimates from the generative model compared to the distribution of observed z-scores. These simulations present a worst-case scenario for inflation of π2: no pruning, all causal effects are in the middle of large LD blocks, and every other SNP in the block is null. In reality, LD blocks containing functional genomic regions appear to have a higher proportion of non-null effects than can be explained by inflation of statistics due to LD alone [17]. LD pruning would also lower the estimate of π2 much closer to the causal proportion. The efficient and accurate handling of the effects of LD on effect size distributions is an area of active research.
For the jth subject, j = 1,…, n, the genotype-phenotype data consist of {xj, yj}, where xj is the vector of mean-centered allele counts from N assayed bi-allelic loci (SNPs) and yj either is a continuous response, or yj ∈ {0, 1} for case-control data, where 0 denotes control and 1 denotes case status. Let X = (ξ1, …, ξN) be the n × N matrix of allele counts, where ξi is the n × 1 column vector of allele counts for the ith genetic locus. (Thus, the jth row of X is given by x j T, where superscript T denotes the transpose of a vector or matrix.) Under Hardy-Weinberg Equilibrium (HWE), the elements of ξi are distributed as centered binomial random variables, Bin(2, pi) − 2pi, where pi is the effect allele frequency for the ith SNP. In the sequel, we assume pi is known, ignoring uncertainty due to estimation, which has no impact on the asymptotic results.
Let b ^ i denote the regression coefficient of ξi on the outcome vector y = (y1, …, yn)T. In this paper, we assume that the vector of regression coefficients b ^ = ( b ^ 1 , … , b ^ N ) T is produced using massively univariate linear (for continuous) or logistic (for dichotomous) regressions. However, the resampling methodology described below is applicable to any regression coefficient estimates b ^, including, for example, best linear unbiased predictors (BLUPs) from random-effects models [3, 30, 31], which may provide better localization of effects. We describe the effects of LD on univariate estimates b ^ both analytically and in Monte Carlo simulations in the S1 Text.
The regression coefficient estimates b ^ are used to produce an N-dimensional vector of Wald test statistics (“z-scores”)
z ≃ n C b ^ ,
where C is an N × N diagonal matrix and ≃ denotes asymptotic equality as the effective sample size n goes to infinity. The diagonal entries c i i = 2 p i ( 1 - p i ) / σ i 2, where pi is the effect allele frequency for the ith SNP and σi is the residual standard deviation (for linear regression) or equal to 1 (for logistic regression). Thus
z ≃ n C b ^ = n diag 2 p i ( 1 - p i ) σ i 2 b ^ = n δ + ω (3)
where δ = (δ1, …, δN)T and ω = (ω1, …, ωN)T are N-dimensional vectors such that
δ i = 2 p i ( 1 - p i ) σ i 2 E { b ^ i } = 2 p i ( 1 - p i ) b i σ i (4)
and ω i ∼ N ( 0 , σ 0 2 ). Here, bi denotes the expectation E { b ^ i }, and normality of ω follows from a large sample approximation (see S1 Text). We assume that the effect sizes are exchangeable with δi ∼ g(δi), where g is an (unknown) marginal density. The theoretical value of the variance σ 0 2 = 1; however, σ 0 2 may be greater than 1 in the presence of the population substructure such as cryptic relatedness [28], and in the model fitting algorithm described below σ 0 2 is estimated from the data. In the remainder of the paper, we define the effect size of the ith SNP as δ i = 2 p i ( 1 - p i ) b i / σ i.
Often the data available from large GWAS meta-analyses are the z-scores from the individual sub-studies, rather than the full genotypic and phenotypic data. In this scenario, it is possible to use the proposed re-sampling based algorithm using z-scores from the individual studies. Suppose the data (X, y) are partitioned into K disjoint independent samples (sub-studies) {(Xk,yk)∣k = 1, …, K}, each with effective sample size nk. The kth sub-study is used to compute an N-dimensional vector of SNP regression weights b ^ k. The z-scores from each sub-study are given by
z k ≃ n k C k b ^ k ,
where C k = diag { 2 p k , i ( 1 - p k , i ) / σ k , i 2 } is an N × N diagonal matrix and σ k , i 2 is the residual variance in the ith regression (for continuous outcomes) or 1 (for logistic regression on discrete outcomes). If for k = 1, …, K, i = 1, …, N, we assume (I) σ k , i 2 = σ i 2; (II) effect allele frequencies pk,i = pi; and (III) b k , i = E { b ^ k , i } = b i; then, the diagonal entries c k , i i = c i i = 2 p i ( 1 - p i ) / σ i 2 and
zk≃nkCkb^k=nkδ+ωk=δk+ωk,
where δ k ≡ n k δ and δ = (δ1, …,δN)T, with δ i = 2 p i ( 1 - p i ) ( β i / σ i ). Thus, δk differs across sub-studies only in the multiplicative factor n k. Assumptions (I)–(III) should be approximately valid if the sub-studies can be considered random draws from the same population. Note, assumptions (I)–(III) are also necessary for meta-analyses to be valid; hence, the assumptions necessary for the random partitioning algorithm proposed below are precisely the standard assumptions used in GWAS meta-analyses [32]. Alternatively, if there are random departures from assumptions (I)–(III), a meta-analysis treating sub-study z-scores as random effects could be performed [29].
If the sub-study z-scores {z1, …,zK} are given, the overall meta-analysis z-scores can be computed as a weighted sum [32]
z = ∑ k = 1 K n k z k ∑ k = 1 K n k = n δ + ω , (5)
where n = ∑ k = 1 K n k and ω = ∑ k = 1 K n k ω k / n and again w i ∼ N ( 0 , σ 0 2 ). In both the Crohn’s disease and the schizophrenia GWAS examples, meta-analysis z-scores are produced using fixed-effects methods, as in their original papers [1, 26].
The N × 1 vector of effect sizes δ is of fundamental interest in GWAS analyses, closely related to power for discovery, proportion of chip heritability discovered, the probability that a SNP is null given its observed z-score, and polygenic risk estimation. As above, let z = (z1, …,zN)T denote the N-dimensional vector of z-scores, where n is the effective sample size of the study. From Eq (3), these z-scores are derived from the simple measurement model z = n δ + ω, where δ is the N × 1 are random draws from an unknown effect size distribution independent of the ω i ∼ i i d N ( 0 , σ 0 2 ).
Since the δi are not observed directly, we are interested in the marginal posterior distributions of δi given the observed test statistic zi. For many uses it is sufficient to obtain the posterior means (E { n δ i ∣ z i }) and variances (Var { n δ i ∣ z i }), for i = 1, …, N. By Theorem 11.1 of [23] (p. 221), these are given by
E { n δ i ∣ z i } = z i + σ 0 2 d d z log { f ( z i ) } , Var { n δ i ∣ z i } = σ 0 2 1 + σ 0 2 d 2 d z 2 log { f ( z i ) } , (6)
where f(zi) is the common marginal probability density function (pdf) of the zi and σ 0 2 is the variance of ωi. This result is quite general, essentially requiring only that δi and ωi are independent and ω i ∼ N ( 0 , σ 0 2 ) [23].
A commonly employed Bayesian framework assumes that some proportion of the tests are generated under the null hypothesis (i.e., δi ≈ 0) and that the complement are generated under the non-null hypothesis (i.e., δi¬≈0) [27]. To formalize this model, let (Zi, Hi) be exchangeable random variables, i = 1, …,N, where as usual Zi denotes the test statistic for the ith test, and Hi ∼ Bernoulli(π2) is an indicator of whether the ith test is null (Hi = 1) or non-null (Hi = 2), and hence π2 denotes the proportion of non-null effects, i.e., the a priori probability that a given hypothesis test is non-null. The marginal density of Zi is given by
f ( z i ) = π 1 f 1 ( z i ) + π 2 f 2 ( z i ) , (7)
where π1 = 1 − π2 is the null proportion, f1 is the null density, and f2 is the non-null density. Under the assumptions following Eq (3), the non-null density f2 is the convolution of a normal density with mean zero and variance σ 0 2, denoted by ϕ ( · ∣ 0 , σ 0 2 ), with the (as yet) unspecified non-null density g of δ.
The two-group mixture model given by Eq (7) is the foundation for the Bayesian interpretation of the false discovery rate [19, 33]. In particular, Efron [19] defined the local false discovery rate (fdr) as the posterior probability that Hi = 0 given Zi = zi. By an application of Bayes’ Rule to Eq (7), the fdr is derived as
fdr ( z i ) = Pr ( H i = 0 | Z i = z i ) = π 1 f 1 ( z i ) f ( z i ) . (8)
The local true discovery rate for the ith SNP is then defined simply as tdr(zi) = 1 − fdr(zi), the posterior probability that an effect is non-null given its observed test statistic zi. Local fdr can be used as a thresholding technique by selecting SNPs corresponding to fdr(zi)≤ α for some choice of cut-off, say α ≤ 0.05, or equivalently, selecting those SNPs for which tdr(z)>1 − α.
There is a close connection between Eq (6) and the fdr defined in Eq (8). By Corrollary 11.3 of [23] (p. 223), these are given by
E { n δ i ∣ z i } = - d d z log { fdr ( z i ) } Var { n δ i ∣ z i } = - d 2 d z 2 log { fdr ( z i ) } . (9)
For a given model E ( θ ) and V ( θ ) for the distribution of effect size expectations and variances, we can estimate parameters θ by utilizing Eqs (6) and (17). Specifically, we enter the model-based predictions (dependent on parameters θ) into quadratic estimating equations that solve for parameter estimates minimizing the differences between the empirical and model-based replication expectations and variances. For the scale mixture of normals model, Eqs (13) and (14) are entered into the quadratic equations.
Q ( θ ) = ∑ m = 1 M ∑ ρ ∈ R E { Z ¯ r , m } - ρ μ ( Z m , n ρ , p ¯ ∣ θ ) 2 + Var { Z r , m 2 } - ρ σ 2 ( Z m , n ρ , p ¯ ∣ θ ) - σ 0 2 ) 2 , (18)
where E { Z ¯ r , m } is the nonparametric posterior mean estimate of the mth bin I m given in Eq (17), Var { Z r , m 2 } = E { Z 2 ¯ r , m } - E { Z ¯ r , m } 2 is the nonparametric variance estimate, Zm is the midpoint of I m, p ¯ is the average effect allele frequency, and ρ = nr,ρ/nρ is the ratio of the effective sample size of the replication sample (nr,ρ) over the effective sample size of the discovery sample (nρ). The advantage of varying ρ is the ability to observe the effects of changing sample size on the effect size distribution and finite-sample replication rates.
In the real applications below, we keep ρ = 0.5 for the Crohn’s disease data, and we vary ρ between 0.3 and 0.5 in the schizophrenia data. Monte Carlo simulations in the S1 Text section use split-half samples (ρ = 0.50). For all analyses, the bin width h was chosen such that there were M = 201 bins equally-spaced bins spanning the range of z-scores. The values for c are chosen to span the entire range of observed z-scores in the given analysis. For the Crohn’s disease example c = 17, for schizophrenia example c = 12, and in the simulations c = 10. Note, the mean allele frequency p ¯ is used in place of the actual frequency for computational efficiency. Actual values of pi could be incorporated by binning with respect to effect allele frequency in addition to binning by discovery z-score; however, in practice this appears to have little effect on estimates.
Eq (18) is minimized over the parameter space θ using a simplex algorithm to produce estimated values θ ^ ≡ { π 1 ^ , σ 0 ^ 2 , σ ^ 1 2 , σ ^ 2 2 } that can then be used to estimate posterior effect sizes, the finite-sample probabilities that SNPs will replicate given their observed z-scores, and the local false discovery rate.
The resampling and fitting algorithm is available in R and Matlab scripts, along with code to generate synthetic sub-study GWAS z-scores, at https://sites.google.com/site/covmodfdr/.
|
10.1371/journal.pntd.0001283 | The Short Non-Coding Transcriptome of the Protozoan Parasite Trypanosoma cruzi | The pathway for RNA interference is widespread in metazoans and participates in numerous cellular tasks, from gene silencing to chromatin remodeling and protection against retrotransposition. The unicellular eukaryote Trypanosoma cruzi is missing the canonical RNAi pathway and is unable to induce RNAi-related processes. To further understand alternative RNA pathways operating in this organism, we have performed deep sequencing and genome-wide analyses of a size-fractioned cDNA library (16–61 nt) from the epimastigote life stage. Deep sequencing generated 582,243 short sequences of which 91% could be aligned with the genome sequence. About 95–98% of the aligned data (depending on the haplotype) corresponded to small RNAs derived from tRNAs, rRNAs, snRNAs and snoRNAs. The largest class consisted of tRNA-derived small RNAs which primarily originated from the 3′ end of tRNAs, followed by small RNAs derived from rRNA. The remaining sequences revealed the presence of 92 novel transcribed loci, of which 79 did not show homology to known RNA classes.
| Chagas' disease is a major health problem in Latin America and is caused by the protozoan parasite Trypanosoma cruzi. T. cruzi lacks the pathway for RNA interference, which is widespread among eukaryotes, and is therefore unable to induce RNAi-related processes. In many organisms, small RNAs play an important role in regulating gene expression and other cellular processes. In order to understand if other small RNA pathways are operating in this organism, we performed high throughput sequencing and genome-wide analyses of the short transcriptome. We identified an abundance of small RNAs derived from non-coding RNA genes, including transfer RNAs, ribosomal RNAs as well as small nucleolar RNAs and small nuclear RNAs. Certain tRNA types were overrepresented as precursors for small RNAs. Further, we identified 79 novel small non-coding RNAs, not previously reported. We did not identify canonical small RNAs, like microRNAs and small interfering RNAs, and concluded that these do not exist in T. cruzi. This study has provided insights into the short transcriptome of a major human pathogen and provided starting points for further functional investigation of small RNAs and their biological roles.
| Trypanosoma cruzi is a protozoan parasite and the causative agent of Chagas' disease, which has substantial health and socioeconomic impact in Latin America [1]. Treatment is currently restricted to a small number of drugs with insufficient efficacy and potentially harmful side effects.
The genome of T. cruzi strain CL Brener is complex in terms of sequence repetitiveness and is a hybrid between two diverged haplotypes, named non-Esmeraldo-like and Esmeraldo-like: we refer to them here as non-Esmeraldo and Esmeraldo. Taken together, both haplotypes [2] sum up to approximately 110 Mb distributed over at least 80 chromosomes [3]. Similar to other trypanosomatids, genes are organized into co-directional clusters that undergo polycistronic transcription. Gene rich regions are frequently interrupted by sequence repeats, which comprise at least 50% of the genome [2]. Several gene variants occur in tandemly repeated copies which often collapse in shotgun assemblies [4]. The genome of a different, non-hybrid strain named Sylvio X10 was recently sequenced and partially assembled, showing a core gene content highly similar to CL Brener [5].
The T. cruzi life cycle is complex and consists of several distinct life stages, morphological states and hosts [1]. To achieve successful completion of the life cycle, the parasite must rapidly adapt to different environments by regulating its gene expression [6]. Transcription in T. cruzi often, but not exclusively, starts at strand switch regions [7], [8], where long transcripts are produced by RNA polymerase II and matured via trans-splicing and polyadenylation [9]. There is to date no definite model of how and if transcription is regulated, as RNA polymerase II promoters for protein-coding genes have not been identified. Thus, it is thought that gene expression is mainly regulated at the post-transcriptional level [9].
RNA interference (RNAi) and related pathways are widespread in animals and other metazoans, participating in a wide range of cellular processes; from chromatin organization to silencing of genes and selfish elements. RNAi relies on small RNA molecules, approximately 20–30 nucleotides in length, to trigger target silencing. In eukaryotes, several different types of small RNAs have been identified. Of these, the best characterized are microRNAs (miRNAs) and small interfering RNAs (siRNAs). See Table 1 for a summary of small RNAs discussed in this study. Two proteins are required for small RNA biogenesis and function: Dicer and Argonaute. Among protozoan parasites, the RNAi machinery has either been lost or retained. T. cruzi have lost the canonical RNAi machinery, which has been confirmed both functionally [10] and from the genome sequence [2], although RNAi is functional in certain other trypanosomatid species [11], [12], [13]. In the African trypanosome Trypanosoma brucei, convincing evidence has shown the presence of an active RNAi machinery (see [14] for references) and more recently pseudogene-derived small RNAs, which were reported to suppress gene expression [15]. A similar situation has been observed in the Leishmania genus. Leishmania braziliensis possesses a functional RNAi-pathway [16], whereas other members of this genus do not [12]. Analyses of the T. cruzi genome have revealed lack of both Dicer and Argonaute homologs. However, similar to other trypanosomatids, T. cruzi possess a protein with a solo Piwi domain, but without a PAZ domain. The biological role of this protein is presently unknown, although it has been suggested to represent a member of a novel Argonaute subfamily [17].
To date, little is known about the presence of small non-coding RNAs (sncRNAs) in trypanosomatids, which do not depend on RNAi. Recently, one study reported the prediction of sncRNAs in the trypanosomatids [18], providing evidence of yet uncharacterized sncRNAs in these species. However, comparative genomics suffers from the limitation that it does not facilitate identification of species-specific small RNAs or regulatory elements. Furthermore, another study described small-scale sequencing of small RNAs in T. cruzi, and reported a population of tRNA-derived small RNAs, which was linked to cellular stress [19]. Moreover, studies from another unusual eukaryote, Giardia lamblia, have shown that sncRNA can be highly diverged from metazoan sncRNA [20], [21] and therefore escape detection using homology searches. Lessons from non-protozoans have taught that novel sncRNAs are often in low abundance and avoid detection using conventional techniques [22], which do not sequence deep enough to capture the full complexity of the transcriptome.
In order to obtain a more complete picture of the short T. cruzi transcriptome, we have performed unbiased deep sequencing and genome wide analyses of the short transcriptome from T. cruzi epimastigotes. The data indicated the existence of an abundance of small RNAs derived from non-coding RNAs and a number of novel expressed loci in the genome.
The sequences have been deposited in the DNA Data Bank of Japan under the accession number DRA000396 and are also available for download from http://www.ki.se/chagasepinet/ncrna.html.
Epimastigotes from T. cruzi CL Brener were grown exponentially at 28°C in liver infusion tryptose (LIT) media [23] supplemented with 10% FBS (Gibco) and streptomycin/penicillin (Gibco), pH 7.3. Total RNA was extracted using the TRIzol method (TRI Reagent, Sigma) following manufacturers' instructions. The total RNA was converted to cDNA using a standard protocol and size fractioned using a polyacrylamide gel. The sequencing library was generated according to the manufacturers' instructions and sequenced with a 454 instrument (GS20 FLX).
The sequence data was stripped of the 3′ ‘CCA’ extension and aligned with the T. cruzi genome assembly [3] using the Burrows-Wheeler Aligner (BWA) [24]. BWA was configured to allow up to two mismatches. Repetitive elements were identified using RepeatMasker [25] and RepBase [26] (r16.01). Only repeats longer than 40 bp were considered.
Identification of novel expressed genomic loci was performed using clustering of reads which could not otherwise be assigned an identity. Clustering was done on reads satisfying the following criteria: i) they do not have an annotation (i.e. tRNA, etc); ii) have only one valid alignment in the genome (single mapping); iii) have an overlap of at least one base with another read. Subsequently, the resulting clusters were filtered using the following criteria: i) a cluster should contain at least six reads ii) at least two reads should be unique. The resulting novel non-coding RNAs were manually examined and assigned putative identities using BLAST.
A sequence database of trypanosomatid genes was established by extracting sequences containing ‘trypanosoma’ in the header line from the GenBank non-redundant database. Statistical evaluation and charts were performed using the R platform. Homology searches were done using NCBI BLAST.
Prediction of tRNA secondary structure was performed using tRNAScan-SE [27] and visualized using VARNA [28]. Prediction of secondary structures of novel small RNAs was done with Vienna RNA [29]. Analysis of putative microRNA targets was performed using TargetScan [30] and GoTermMapper [31]. Scripts were written in Perl and are available on request.
Stem-loop real time-PCR experiments and primer design were performed as described in [32]. The quality of RNA samples was assessed on an 1.5% TBE-agarose gel. All RNA samples were treated with DNAse I (Fermentas) previous to the reverse transcriptase reaction. SnoRNA or 5S rRNA was used as reference RNA for qualitative/quantitative experiments. The following reagents were mixed and subjected to Pulsed RT reaction; 60 ng of T. cruzi RNA (per RT reaction), 0.5 mM dNTP, 10X First Strand Buffer, 5 mM MgCl2, 10 mM DTT, RNAseOUT, 50 units Superscript III RT (Invitrogen) and 1 µl of SL-RT specific primer, using the conditions: 30 min at 16°C followed by 60 cycles of 30 s at 30°C, 30 s at 42°C, 1 s at 50°C and a final step of 5 min at 85°C. RNAse H was added and incubated for 20 min at 37°C. The real time PCR reactions were performed in triplicates using 1 µl cDNA, 300 nM forward specific and reverse universal primers and 2X SYBR Green Master Mix (Roche). Cycling conditions: 5 min at 95°C, 40 cycles 95°C-5 s, 60°C-20 s, 72°C-1 s followed by dissociation curve analysis in Strategene Thermal Cycler Mx3000P. Experiments were repeated at least twice per each of two biological samples. Ct values were normalized against 5S rRNA values and the abundance ratio was calculated for each individual Ct value and mean as well as standard deviation were calculated and graphed using SigmaPlot 9.0. The 5S rRNA (Tc00.1047053509455.160) and C/D snoRNA (Tc00.1047053510739.50) were used for normalization. Primers used are listed in Table S3.
An epimastigote cDNA library was size fractioned on a polyacrylamide gel and sequenced using 454 sequencing [33] (Materials and Methods). Sequencing resulted in a total of 582,243 reads (101,284 unique) with a size range between 16 to 61 nucleotides (Figure 1A). The median sequence length of the library was 38 nt. A total of 12.2% (71,309/582,243) reads occurred as single copy, whereas the remaining reads had a variable copy number between 2 and 41,929. The selected size range should contain only non-coding RNA (ncRNA), as there are no known protein-coding genes in this size range. Further, this size range was selected to avoid spliced leader RNAs. However, degradation products from transcriptional turnover could be present in the sample. Based on two observations we conclude that degradation products were not contaminating the library; i) degradation fragments should exhibit a random distribution pattern in protein-coding genes, which was not the case, ii) ribosomal RNA constitute the bulk (>80%) of cellular RNA, which was not observed in the sequence data.
The sequence data was separately aligned with each of the T. cruzi CL Brener haplotypes; non-Esmeraldo and Esmeraldo (Figure 1BC, Materials and Methods). In addition, reads were aligned with a 38 million base pair collection of unassigned contigs (Figure 1D), which mostly consists of repeats [3], [34]. This resulted in a total of 90.7% aligned reads (528,228/582,243), or expressed in unique reads, 74.0% aligned reads (75,024/101,284) (Table 2). Slightly more reads were aligned with non-Esmeraldo, owing to the more complete status of this haplotype assembly compared to Esmeraldo; however the length distribution of the aligned reads were similar (Figure 1BC), indicating both haplotypes might generate similar RNA populations.
A total of 9.2% (53,646/582,243) of the reads could not be aligned with the genome using the default alignment procedure, raising the question if these reads are biologically derived or technical artifacts. The following scenarios are possible; i) unaligned reads are technical artifacts or enriched with sequencing errors, ii) unaligned reads represent small RNAs derived from unfinished parts of the genome sequence, iii) small RNAs have been subjected to chemical modification and RNA editing events. As the T. cruzi CL Brener genome sequence is not complete [2], [4] it remains possible that at least some small RNAs are derived from unassembled regions. To investigate this, unaligned reads were mapped to the genomic shotgun reads from the genome project, which provided alignment to 0.49% (2860/53,646) of the unaligned reads. Examination of a limited number of reads, that failed alignment, found homology to tRNALys. As these reads occurred in a high copy number (∼300) and mismatches were located in the anticodon loop, this makes it possible that mismatches are not sequencing errors but rather modified nucleosides misinterpreted by the sequencer.
In order to differentiate between known and unknown RNA species in the library, we categorized reads into classes using genome annotations. Alignment coordinates were superimposed on genome annotations and each read was categorized into one of the categories in Table 3 if it completely or partially overlapped with the annotation. In cases where a tRNA, snRNA or snoRNA was overlapping a protein-coding gene, the ncRNA gene was preferentially selected. To further improve the classification, reads without annotation were queried against a database of various trypanosomatid sequences (Materials and Methods).
For reads with a single alignment location (single mappers), 97.4% (378,446/388,551) of the reads in non-Esmeraldo and 96.7% (157,280/162,622) in Esmeraldo were found to correspond to small ncRNAs (sncRNAs) derived from tRNA, rRNA, snRNA and snoRNA (Table 3). tRNA-derived small RNAs (tsRNA) was found to be the most abundant type in the library, composing at least 65.3% (380,191/582,243) of the total sequence data, which we further describe in the next section. This result suggests that the vast majority of small RNA species in T. cruzi epimastigotes are derived from known ncRNA classes.
About 2–5% of the aligned sequences could not be classified into known ncRNA classes. It should be noted that this fraction might not represent the entire abundance of novel sncRNA in T. cruzi, as some sncRNA might only be present in a specific life stage or under a certain physiological condition. Furthermore, long novel ncRNAs (>61 nt) could exist [18], as was for example reported in Leishmania infantum [35].
A total of 19,893 reads aligned with unassigned contigs, out of which 78% (15,506/19,893) represented reads that aligned with rRNA genes (Table 3). A minor fraction consisted of reads that aligned with tRNA (13%) and snRNA/snoRNA (1%). This is consistent with the fact that few rRNA genes have been properly assembled [2], [3].
For both non-Esmeraldo and Esmeraldo, a total of 69.1% (282,036/408,008) of the small RNAs were assigned to the tRNA category (considering single mapping reads), despite the fact that the library was size selected for sequences shorter than 61 nt and mature tRNAs are between 70–80 nt. A closer inspection revealed the presence of tRNA-derived small RNAs (tsRNAs), a phenomenon reported previously in higher eukaryotes [36], [37] and lower eukaryotes [38], [39], [40]. However, the physiological role, if any, of tsRNA is not well defined (for review and discussion see [41], [42], [43], [44]).
T. cruzi tsRNAs were first reported by Garcia-Silva et al. [19], who found tsRNA to be recruited to cytoplasmic granules and increase under stress conditions. The authors employed a 20–35 nt cDNA library and sequenced 348 clones, and found that 26% of the clones were derived from tRNA and 60% from rRNA. The study also showed a higher representation for 5′end tRNA derived small RNAs, which may be explained by the relatively low number of clones sequenced in this study.
In our library, tsRNA had a median length of 38 nt and 88.9% (250,920/282,036) were derived from the 3′ end of tRNAs (Figure 1E, Figure 2, Table 4). Moreover, 75.3% (189,116/250,920) of the 3′-derived reads contained a ‘CCA’ nucleotide extension; indicating that the majority of 3′ tsRNA are derived from mature tRNA species, as the ‘CCA’ addition is post-transcriptionally added in eukaryotes. However, we cannot rule out that the remaining reads did not lose the ‘CCA’ extension during sample or library preparation.
The median length of 38 nt is consistent with the current view of bisectional cleavage of mature tRNA. Despite this, we also identified shorter tsRNA (<25 nt) albeit in lower frequency; a total of 1605 tsRNA were 24 nt or less and primarily derived from tRNAGlu, tRNAAsp, tRNATyr, tRNAVal and tRNAArg (Figure S1). Interestingly, the shorter tsRNA were more often derived from the 5′ arm. Most tRNA isoacceptors were found to be precursors for tsRNA, but with relative different amounts (Figure S1). The most abundant tsRNA were derived from the 3′ arm of tRNAHis and occurred in 41,929 copies and contained the ‘CCA’ extension (Table 5, Figure S2). Interestingly, the 3′/5′ ratio of tsRNA was not equal for all tRNA isoacceptors (Table 4). For example, tRNAGln showed more tsRNA derived from the 5′ arm.
A recent study reported the cloning and characterization of tsRNA in the primitive eukaryote Giardia lamblia (G. lamblia) [38], showing that tsRNA are abundantly expressed during the encystation stage and are ∼46 nt long. Consistent with T. cruzi tsRNAs, G. lamblia tsRNAs are derived from most tRNA isoacceptors and predominantly from the 3′ arm. In G. lamblia, tsRNAs from tRNAAsp and tRNAGly were the most frequently cloned, which may indicate species or life stage specific isoacceptor preference.
If tsRNA would represent degradation products from tRNA-turnover, it would be expected to find a correlation between the RNA fragment amount and the expression levels of tRNA genes. In the absence of tRNA expression data, we utilized the amino acid usage from the predicted proteome and compared it with the observed tsRNA expression. We found no correlation between the observed tsRNA expression and the amino acid usage (Pearson's correlation, r = −0.05), nor was there a correlation between the genomic copy number of tRNA and tsRNA expression (Pearson's correlation, r = 0.08), suggesting that T. cruzi tsRNAs are not random degradation products from tRNA turnover. As we observed a very high expression of tsRNA from certain tRNA isoacceptors (e.g. tRNAHis, tRNAArg and tRNAThr), but almost no expression from others (tRNAPhe and tRNAAsn), this implies tsRNA are differentially expressed in T. cruzi. Furthermore, we performed experimental validation of a few selected tsRNA (Figure S3).
Consistent with previous reports [36], [37], [38], [39], we found that the cleavage site was present within the anticodon loop of mature tRNAs (Figure S2). For shorter tsRNAs, the cleavage site was present in the two other loops, but primarily in the loop of the T-arm. This suggests endonucleolytic cleavage as the responsible mechanism behind tsRNA generation. The precise cleavage supports the idea that tsRNA are generated by a distinct mechanism rather than random degradation. However, as shorter tsRNA were observed, these might require both endonucleolytic cleavage and exonucleolytic trimming in their biogenesis pathway.
We observed tsRNA with and without a CCA 3′ terminus, thus the process of tsRNA formation likely targets both pre-tRNAs and mature tRNAs, and therefore takes place either in the cytosol or nucleus, as only mature tRNAs are imported into the mitochondria [45]. An early study by Zwierzynski et al. reported 3′ CCA activity in nuclear extracts [46], raising questions about the subcellular location of tsRNA biogenesis. The key enzymes involved in tsRNA biogenesis remain to be identified; however, it remains clear that this mechanism is independent of Dicer.
It has been hypothesized that tsRNAs inhibit protein synthesis either by depleting the cellular tRNA pool or by a more intrinsic mechanism involving a protein repression complex [43], albeit there is to date no definite evidence. tsRNAs have been associated with Piwi and Argonaute complexes [47], [48], [49], suggesting that it may guide degradation of target transcripts in RNAi-positive organisms. A recent study reported tsRNAs to guide tRNase Z-mediated cleavage of engineered target sequences and possibly endogenous transcripts [50], which further supports the idea of these species as functional entities.
Small nucleolar RNAs (snoRNAs) are present throughout eukaryotes and guide enzymatic modifications of target RNAs in the nucleolus, and can be subdivided into C/D and H/ACA classes based on sequence motifs. Recently, snoRNA-derived small RNAs (sdRNA) have been reported in animals [37], [51], [52] and in the protozoan G. lamblia [53] and are thought to be generated by a Dicer-dependent mechanism [52]. Metazoan sdRNAs are predominantly ∼17–19 nt and ∼30 nt and generated either from the 5′ (C/D type snoRNAs) or 3′ ends (H/ACA type snoRNAs) [52]. In both humans and G. lamblia snoRNA-derived small RNAs have been implicated to have miRNA-like functions [53], [54].
We found that 0.26% (1413/528,228) of the total data was represented by snoRNA-derived small RNAs, with a median length of 35 nt, similar for both C/D and H/ACA (Figure 1G, Figure 2). The observed length of sdRNA is different from metazoan sdRNA and both types were found to have similar number of reads (n = 770 and n = 643 reads for C/D and H/ACA snoRNA respectively). We did not observe the positional bias towards the 3′ end which has been reported for mammalian sdRNA, or a specific alignment pattern suggestive of regulated cleavage. These findings suggested that the observed sdRNAs were generated by a different mechanism compared to those found in metazoans, or less interestingly, represent degradation or break-down products.
A total of 0.53% (2839/528,228) reads were derived from small nuclear RNAs (snRNAs) which were distinct from snoRNAs, with a median length of 40 nt. Interestingly, 82.1% (2333/2839) of the snRNA derived small RNAs were from snRNA U4 and U5. Two snRNA-derived reads occurred in a high copy number (∼100 copies).
Small RNAs derived from ribosomal RNA have received less attention but are known to exist [37], [40] and have been reported to increase as a response to oxidative stress. Here, 17.2% (91,206/528,228) of the aligned sequences represented small RNAs derived from ribosomal RNAs (rsRNAs). rsRNAs could be grouped into three different subpopulations based on their length distribution (Figure 1F); one population with an average length of 20 nt, a second population with an average length of 33 nt, and a third longer population with an average length of 46 nt. Complete rRNA genes are not present in the current assembly [2], [3] and it is therefore difficult to conclude if the small RNAs represent degradation products or not. However, the copy number of rRNA-derived small RNAs was highly variable; ranging from 1 (n = 6337 reads) to >100 (n = 117 reads), which suggests a mechanism of non-random degradation.
A total of 1.69% (8964/528,228) of the aligned reads were not derived from known tRNA, rRNA, snRNA, snoRNA or repeats, of which 17.4% (1565/8964) aligned with protein-coding genes and the remaining with intergenic regions (Figure 1HI). In order to find novel ncRNAs, we performed clustering of reads with overlapping alignments (Materials and Methods).
These criteria formed 92 loci, consisting of a total 7805 reads (Table 6, Table S1), of which 13 loci were identified as known non-coding RNAs using homology searches, which have been missed in the present genome annotation. The remaining 79 loci did not fall into known ncRNA classes and had an average length of 54 nt. None of these had homology to any known RNA class in Rfam or GenBank, albeit seven displayed partial sequence similarity with protein-coding genes and pseudo genes. We performed secondary structure prediction [29] of these unknown RNAs; 26 did not fold at all, 35 folded into non-hairpin structures and 18 folded into hair-pin structures according to predictions. Next we compared the 79 candidates to ncRNAs previously reported from comparative genomics [18], but failed to find overlap between the two sets of candidates. This result does not exclude the possibility that the previously reported ncRNA are correct, as only 20% (15/72) was in the size range of our library. Finally we queried our 79 novel ncRNA candidates against other trypanosomatid genomes (T. brucei, T. vivax, T. congolense, Leishmania spp.) to test if these sequences are conserved among other trypanosomatids; however, no full length matches were found. These findings suggested that novel RNAs, as identified here, are specific for T. cruzi rather than ubiquitous among trypanosomatids.
The remaining 1159 reads did not pass the criteria for clustering and had a median length of 24 nucleotides. These reads were subsequently queried with BLAST against a trypanosomatid sequence database (Materials and Methods); 335 reads displayed homology to trypanosomatid rRNA genes and 819 with homology to protein-coding genes. For reads with alignment to protein-coding genes we observed no statistical overrepresentation of antisense alignments, and as these did not derive from known ncRNA, the following scenarios are possible; i) small RNAs with homology to protein-coding genes are spurious transcriptional products, or debris from mRNA turnover, without biological significance, ii) small RNAs with homology to protein-coding genes are a result of regulated or non-regulated mRNA turnover with biological significance, iii) small RNAs with homology to protein-coding genes are transcribed from the genome and not derived from mRNA. To address these questions, functional studies will be needed to answer whether these small RNAs are biologically active or debris from the normal cellular turnover.
MicroRNA (miRNA) is a class of regulatory small RNAs that fine tune gene expression in metazoans. One attractive hypothesis is that intracellular parasites utilize the host microRNA pathway to change the cellular environment for its own needs. Partial evidence exists from Cryptosporidium parvum and Toxoplasma gondii that this may take place [55], [56], [57]. None of the small RNAs showed complete or partial homology when compared with human microRNA sequences from [58]. Next, we performed putative target site prediction of the 819 small RNAs. The putative ‘seed region’ (nt 2–8) was extracted from each of the 819 small RNAs and queried using standalone TargetScan against 23-way UTR alignments. A conserved target site was required to be present in the following 7 genomes; Homo sapiens, Mus musculus, Rattus norvegicus, Gallus gallus, Macaca mulatta, Pan troglodytes and Canis lupus familiaris. As a result, a total of 3230 putative target genes were identified. Subsequently, a slimmed gene ontology was used to group the identified genes into a more narrow set of categories. Interestingly, 33% (1063/3230) of the genes grouped into ‘cellular nitrogen compound metabolic process’ (GO:0034641), raising the possibility that parasites may modulate the immune response by interfering with the host production of nitric oxide. Furthermore, ‘immune system process’ (GO:0002376) contained 7.7% (250/3230) of the genes. One hypothesis derived from this bioinformatic prediction is that T. cruzi manipulates the host cell environment by secretion of oligonucleotides that mimic human microRNAs.
Repeats are an inherent feature of most eukaryotes and have been attributed as an important driving force behind genome evolution [59]. T. cruzi have a significant part of its genome devoted to repeats; inactive and active retrotransposons, microsatellites and large gene families, often arranged in tandem. At least two types of non-Long Terminal Repeat (LTR) retrotransposons, designated CZAR and L1, are potentially active in the T. cruzi genome [60]. The CZAR element consists of two open reading frames and represent a site-specific retrotransposon that inserts into spliced-leader genes [60]. Small RNAs have been implicated in the protection against retrotransposons in both metazoans and protozoa [22]. However, it is presently unknown how RNAi-negative protozoa, such as T. cruzi, protect themselves against the potentially disruptive effects of transposition events. This intriguing question motivated us to look for evidence of small RNAs that target or transcribe from retrotransposons and other repeats.
Initially, the T. cruzi genome was searched with RepeatMasker [25] in combination with RepBase [26] to identify all known instances of mobile elements and satellite repeats, which resulted in 13 different types of repetitive elements covering 11–12% of the genome (Table S2). Twenty base pairs flanking each side of a repeat instance was included. To add more confidence to the analysis, we decided to maximize the number of useable reads by including those that go to multiple locations (multi mapping reads). We used a similar approach to what was described in [61], where a particular read was allowed mapping to more than one location, but only to one type of element. Reads going to more than one type of element or outside of repeats were removed. This resulted in a total of 0.13% (782/582,243) of reads from the library that aligned with various repetitive elements (Table S2). This suggests that if any of these small RNAs have a role to inhibit or block transposition events, these are present in a very low amount.
We found that CZAR contained the highest amount of aligned reads (n = 446), despite the fact that this element only covered 0.21% of the genome. Several instances of the CZAR element were found to have reads in the 5′ or 3′ termini, or in the close vicinity. As reads were mostly found to align sense, these may represent initiation fragments from the transcription of these elements, supporting the idea that at least some CZAR elements are actively transcribed in the genome.
The SIRE and TcVIPER have been suggested to represent two classes of dead elements [60]. A low number of reads aligned with TcVIPER (n = 25) and SIRE (n = 2), possibly suggesting that some transcription of these elements might occur despite their inability to transpose.
TcSAT1 is a ∼200 bp satellite repeat and comprises ∼5% of the current draft genome sequence [62]. Conflicting data exist regarding the transcription of TcSAT1, where Northern blot hybridization experiments indicated no transcription, whereas nuclear-run-on assays and microarrays indicated active transcription (see [62] for references). We found 150 reads aligned with this repeat element, which may represent degradation fragments or small RNAs derived from longer transcripts.
Overall, we observed no overrepresentation of antisense reads in any class of repeat elements. However, it is possible that an antisense inhibitory mechanism is present albeit in a very low abundance, which would require deeper sequencing and a more narrow size fraction to be captured. Finally, it is also possible that T. cruzi does not use small RNAs to control transposition.
We validated the presence of 12 small RNAs that were found to be abundant in the sequencing data; six tsRNAs (derived from tRNAAla, tRNATyr, tRNATrp, tRNAGlu, tRNAAsp and tRNAThr), four rsRNAs and a repeat-derived small RNA (Figure S3). Validation was performed by Stem-loop Real Time PCR [32], which has previously been used to detect microRNAs and is more sensitive than Northern hybridization [63]. Of the 18 selected small RNAs, 12 could be amplified (Figure S3AB). A tsRNA derived from tRNAHis was also detected among our samples (data not shown), however, due to primer dimer formation it could not be properly quantified by real time PCR analysis.
To obtain a measure of abundance, the signal intensity from the real time PCR was normalized using full length rRNA and C/D snoRNA (Materials and Methods). A tsRNA derived from tRNAAsp gave the strongest signal of the tested tsRNAs (Figure S3AC). The other five tsRNAs displayed a similar level of expression as the snoRNA-control used in these experiments. Three tsRNAs (derived from tRNAAla, tRNAGlu, tRNAAsp) have been detected previously in the T cruzi clone Dm28c by Northern hybridization [19], thus indicating their presence in independent strains.
Of the four tested rsRNAs, rsRNA-2 gave the strongest signal (Figure S3BC). Interestingly, several small RNAs derived from non-coding RNAs can be aligned with protein coding genes in the anti-sense direction. For example, rsRNA-3 and rsRNA-4 can be aligned with two distinct protein-coding genes, along with several other putative small RNAs. A similar situation occurs with MASP genes, where small RNAs derived from the repeat element TREZO [64] can be aligned close to the MASP 3′ UTR, which is the most conserved region among these genes [65]. We validated a small RNA derived from the TREZO element, showing that the abundance is similar to that of the snoRNA control (Figure S3). TREZO elements cover ∼1–2% of the genome, exhibit site-specificity for insertions and are transcribed [64], although this is the first report to show they generate such small RNAs. Their putative influence on the MASP family expression needs to be further investigated.
In this study, we analyzed the short transcriptome of Trypanosoma cruzi using unbiased deep sequencing and provided a glimpse into the diversity and abundance of small RNAs in this species. Despite the fact that T. cruzi lacks RNA interference, our deep sequencing led to the identification of several new types of small RNAs which have not previously been reported in this important organism. The most common RNA species were small RNAs derived from transfer RNAs, followed by small RNAs derived from ribosomal RNAs. Only 1% of the small RNAs in the library were derived from small nuclear RNAs and small nucleolar RNAs. Our deep sequencing effort confirms that, similarly to other protozoan species and mammalian cell lines, T. cruzi accumulates RNA species from tRNA, rRNA as well as snRNA and snoRNA. A selected set of small RNAs was validated using real time PCR and found to be consistently present in different biological samples, although further experimental work will be needed to provide functional insights into the putative roles of some of these small RNAs. Our sequencing data provide a substantial number of follow-up candidates which might be suitable for detailed experiments.
We found no evidence of canonical small non-coding RNAs (i.e. microRNA and siRNA) as often found in metazoans; an expected finding, consistent with the absence of the RNA interference machinery and confirms the results from previous studies showing that canonical microRNAs do not exist in Trypanosoma cruzi. About 1.69% of the small RNAs in the library were unknown, and we identified 92 novel expressed loci, of which 79 lacked conserved sequence or structural motifs. However, it should be noted that the small RNAs reported in this study may not reflect the complete repertoire, as certain small RNAs may have a life stage specific expression or otherwise only be expressed under a certain physiological condition.
Further sequencing efforts will be needed to elucidate the complete set of small RNAs and to completely distinguish biologically non-stable intermediates from stable RNAs. Furthermore, it remains to be elucidated whether small RNAs are generated by a distinct mechanism or produced by RNA decay, although the latter does not exclude the possibility that small RNAs have a functional role. Currently we are undertaking deep sequencing of a smaller size fraction to further understand the composition and complexity of the short transcriptome in this peculiar organism.
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.