doi
stringlengths 28
28
| title
stringlengths 19
311
| abstract
stringlengths 217
5.08k
| plain language summary
stringlengths 115
4.83k
| article
stringlengths 3.87k
161k
|
|---|---|---|---|---|
10.1371/journal.pntd.0004143 | A Comparative Study of Peripheral Immune Responses to Taenia solium in Individuals with Parenchymal and Subarachnoid Neurocysticercosis | The ability of Taenia solium to modulate the immune system likely contributes to their longevity in the human host. We tested the hypothesis that the nature of the immune response is related to the location of parasite and clinical manifestations of infection.
Peripheral blood mononuclear cells (PBMC) were obtained from untreated patients with neurocysticercosis (NCC), categorized as having parenchymal or subarachnoid infection by the presence of cysts exclusively within the parenchyma or in subarachnoid spaces of the brain, and from uninfected (control) individuals matched by age and gender to each patient. Using multiplex detection technology, sera from NCC patients and controls and cytokine production by PBMC after T. solium antigen (TsAg) stimulation were assayed for levels of inflammatory and regulatory cytokines. PBMC were phenotyped by flow cytometry ex vivo and following in vitro stimulation with TsAg.
Sera from patients with parenchymal NCC demonstrated significantly higher Th1 (IFN-γ/IL-12) and Th2 (IL-4/IL-13) cytokine responses and trends towards higher levels of IL-1β/IL-8/IL-5 than those obtained from patients with subarachnoid NCC. Also higher in vitro antigen-driven TNF-β secretion was detected in PBMC supernatants from parenchymal than in subarachnoid NCC. In contrast, there was a significantly higher IL-10 response to TsAg stimulation in patients with subarachnoid NCC compared to parenchymal NCC. Although no differences in regulatory T cells (Tregs) frequencies were found ex vivo, there was a trend towards greater expansion of Tregs upon TsAg stimulation in subarachnoid than in parenchymal NCC when data were normalized for the corresponding controls.
T. solium infection of the subarachnoid space is associated with an enhanced regulatory immune response compared to infection in the parenchyma. The resulting anti-inflammatory milieu may represent a parasite strategy to maintain a permissive environment in the host or diminish inflammatory damage from the host immune response in the central nervous system.
| Cysticercosis is a parasitic disease caused by the larval cysts of the pork tapeworm Taenia solium following accidental ingestion of tapeworm eggs. Neurocysticercosis is a leading cause of seizures and epilepsy in developing countries. The clinical manifestations depend on many factors including the number, size, location, stage of the parasites and host age, gender and immune responses. However, the most important determinant is whether the infection is restricted to the brain parenchyma (associated with seizures disorders but generally with a good prognosis) or if it involves the subarachnoid spaces in the brain (associated with an exuberant growth of the parasite, intracranial hypertension, and often resulting in mortality). In the current study, we characterized the relationship between T. solium specific immune responses and cyst location. We show that patients with subarachnoid disease have higher suppresive immune responses than the parenchymal group, potentially representing a parasite strategy to survive. Our findings implicate immune responses to be among the mechanisms that underlie the diverse clinical manifestations in neurocysticercosis.
| Human cysticercosis involving the central nervous system (CNS), or neurocysticercosis (NCC), is one of the most common neurological infections in developing countries, accounting for as much as a third of adult onset epilepsy in endemic regions [1–3]. The infection results from the accidental ingestion of Taenia solium eggs. Following release in the duodenum and activation by digestive enzymes, oncospheres rapidly migrate into the blood and eventually reach the brain and other tissues. In the CNS, oncospheres transform into metacestodes or cysticerci and can survive intact for many years [3,4]. Disease occurs with an initial asymptomatic phase and frequently progresses to a symptomatic phase with a variety of symptoms including headaches, increased intracranial pressure and seizures. Progression to symptomatic NCC is usually associated with larval degeneration resulting from parasite aging, progressive destruction from inflammatory host immune responses over time, or rapidly after anti-helminthic treatment [4].
NCC demonstrates a remarkably heterogeneous presentation with a wide variety of clinical manifestations. The symptoms and signs of infection vary depending on cyst size, viability, and/or location within the brain [3–5]. Previous studies correlating differences in immunological responses to different clinical manifestations of NCC have frequently been inconsistent in the inclusion criteria of cases [6–8]. Studies that investigated the relationship between immune responses and disease manifestations of NCC have generally classified their study populations based on symptomatology [6,9–12]. A significant issue with classifying disease by the presence or absence of symptoms is that individuals can remain symptomatic, even with severe symptoms, well after the death of the parasite [4]. Thus, the study groups included individuals with heterogeneous disease states, hence similarly heterogeneous immune responses [6,8,10,12]. In addition, in most previous studies the patient populations were studied also differed in terms of anthelmintic and corticosteroids treatment [6–8]. Furthermore, a wide variety of parasite antigens have been used to analyze the specific immune response resulting in further heterogeneity in the host immune response. Host factors such as age, gender and the nature of the immune response can also influence the clinical presentation of NCC. Thus, independent of gender, older NCC patients frequently have a predilection for multiples parasite and vesicular cysts. While an increased inflammatory response, reflected by high leukocyte counts in CSF and tomographic evidence of inflammation, is seen more commonly in female NCC patients [13,14].
The prognosis of cerebral cysticercosis generally correlates with the varying duration and severity of the disease that is dependent on whether the cyst location is in parenchymal sites, subarachnoid spaces, or in the major sulci [4]. In the present study, we compared host immune responses in subarachnoid disease, which is characterized by an exuberant or unrestrained growth of the parasite membrane, and parenchymal disease, associated with granulomas and calcification that would presumably sequester the parasite. Further, by matching patients with an uninfected control group by age and gender, we were able to relate the difference between individuals in the two infected subgroups could be compared directly to a reference group, normalizing for within group variability. Thus, the goal of this study was to correlate the T. solium-specific immune response in neurocysticercosis with disease manifestations as determined by cyst location. We focused on the immune responses in circulating immune cells with the hope that a better understanding of the peripheral immune response may reflect and provide context to local immune responses around the cyst in the CNS.
For this study, we enrolled 32 patients diagnosed with NCC based on criteria specified in the General protocol for Investigation in Neurocysticercosis for the Instituto Nacional de Ciencias Neurológicas, Lima, Peru (protocol # 01034), that included clinical evaluation, neuroimaging, and serological testing. The patients were classified by the location of cysticerci in parenchymal (n = 16) or subarachnoid (n = 13) spaces based on imaging data. Three subjects who had both parenchymal and subarachnoid cysts were excluded from the analysis, resulting in a study size of 29 individuals. To control for exposure to other helminths and immunologically relevant factors, each patient (n = 29) was matched to an uninfected subject by age and gender. The control group consisted of individuals resident in a region non-endemic for cysticercosis, and all were seronegative for cysticercosis and T. solium taeniasis by the standard diagnostic test, an enzyme-linked immune transfer blot (EITB) [15]; no imaging data were obtained from these uninfected individuals. None of the participants had received steroids or anti-parasitic treatment at the time of blood sampling. Fresh blood specimens were collected from patients and control subjects to obtain mononuclear cells and sera and NCC patients were treated with anthelmintics and anti-inflammatory therapy, if appropriate, in follow up after blood sampling.
The study was approved by the Ethics Committee of both the Universidad Peruana Cayetano Heredia and the Instituto Nacional de Ciencias Neurológicas. The Institutional Review Board (IRB) of the Universidad Peruana Cayetano Heredia, Lima, Peru approved the protocol (protocol # 54702) following the principles expressed in the Declaration of Helsinki (1975). Informed consents were signed by all subjects. NCC patients were under continuing care by staff physicians, and received standard of care treatment after sample collection
T. solium cysticerci were obtained from naturally infected pigs, purchased from cysticercosis-endemic rural regions of Peru. Cysts recovered from pig muscles were washed several times with sterile phosphate buffered saline (PBS), then homogenized and sonicated at 4°C. Cell-free extracts were obtained after centrifugation at 500 g for 10 min. Proteins were quantified using the BCA assay (bicinchoninic acid protein assay; Pierce Biotechnology, Inc., Rockford, IL).
Fresh human peripheral blood was obtained and used as a source of PBMC, separated by differential centrifugation using Ficoll-Paque Plus (GE, Piscataway, NJ), according to the manufacturer’s protocol. Cells were set up for in vitro culture in 24-well plates at a concentration of 1x106 cells/well (for cell phenotyping) and 96-well plates at 2x105 cells/well (for measurement of cytokine production) in RPMI 1640 medium (Gibco) with 5% heat-inactivated human serum, penicillin and streptomycin, in one of the following conditions: (1) medium alone, (2) T. solium antigen (TsAg): 20 and 40 μg/ml and (3) Mycobacterium tuberculosis purified protein derivative (PPD; 5μg/ml) (Mycos Research) was used as a positive control for measuring an in vitro immune response. Culture reagents had minimal lipopolysaccharide (LPS) contamination as assessed by LAL (Limulus amebocyte Lysate test; sensitivity 0.03EU/ml). Cell cultures were incubated at 37°C for 3 days (cell phenotyping) or 48 h (cytokine production).
The frequency of T cell subpopulations (expressing CD3+, CD4+ or CD8+ cells), NK cells (CD3-CD16+CD56+), B cells (CD19+) or regulatory T cells (Tregs: defined as CD4+CD25+CD127low/-Foxp3+) were determined for each study subject using standard phenotyping protocols provided by the manufacturers. For this purpose, the following monoclonal antibodies were purchased from Becton Dickinson (San Diego, CA) or eBioscience (San Diego, CA): FITC-conjugated anti-CD3 (clone HIT3a) and anti-CD25 (clone M-A251); PE-conjugated anti-CD16 (clone 3G8) and anti-CD127 (eBioRDR5); PerCP-conjugated anti-CD4 (clone S3.5) and anti-CD19 (clone SJ25-C1); APC-conjugated anti-CD8 (clone RPA-T8), anti-CD56 (clone MEM-188) and anti-FoxP3 (clone 236A/E7).
The cellular phenotype in PBMC was determined by flow cytometry ex vivo and following 3 days of culture in medium with or without parasite antigens. After washing, mononuclear cells were stained for specific surface molecules and finally fixed. For FoxP3 intracellular staining, after CD4, CD25 and CD127 surface staining, cells were fixed/permeabilized with FoxP3 staining buffer set (eBioscience) and incubated with the FoxP3 antibody. After 30min of incubation at 4°C cells were washed once with permeabilization/wash buffer and resuspended in 1% paraformaldehyde in PBS. For each sample of PBMC, at least 10,000 target events were acquired on a FACS Calibur (Becton Dickinson) and analyzed using Cellquest software (Becton Dickinson).
Sera and supernatants from PBMC cultured for 48 h at 37°C with different concentrations of T. solium antigen were assayed for cytokines: IL-1β, IL-12, interferon gamma (IFN-γ), IL-4, IL-13, IL-10, Granulocyte colony-stimulating factor (G-CSF), vascular endothelial growth factor (VEGF), IL-1rα, IL-2, IL-5, IL-6, IL-7, IL-8, IL-9, IL-15, IL-17, eotaxin, basic FGF, GM-CSF, IP-10, MCP-1, MIP-1α, MIP-1β, PDGF-B, RANTES, and TNF-α using Luminex xMAP technology with a Bio-Plex Human cytokine 27-Plex kit. Cytokine concentrations were expressed as pg/ml using manufacturer-provided standards.
A patient:matched control ratio (r) was calculated for each cytokine concentration and the frequency of each cell sub-population to normalize the patient’s data with the matched control. Statistical significance of differences between patients and controls were tested using a one sample t-test for the difference of ratios from 1.0. The Mann Whitney U rank sum test was used to compare the immune response frequencies of NCC patients with parenchymal vs subarachnoid cyst location. The Wilcoxon signed rank test was used to compare TsAg vs medium for the effect of antigen stimulation. A P-value of <0.05 was considered statistically significant. All analyses were corrected for multiple comparisons by the software (GraphPad Prism V6) and the corrected p-values are used in data interpretations.
NCC patients were seropositive on EITB and were assigned to the parenchymal or subarachnoid disease groups based on their imaging studies as described by Garcia et al.[16]. Patients with parenchymal NCC were included only if they had two or more viable cysts, to minimize the chances of misdiagnosis, and patients with subarachnoid disease were included in this group if they had parasitic lesions only in the basal subarachnoid spaces or in the Sylvian fissure, outside the brain parenchyma. Patients with subarachnoid NCC had parasite membranes and/or cysts identified in the extraparenchymal spaces on MRI imaging, but none had cysts in the brain parenchyma. Three subjects (among the 32 study participants) who had both subarachnoid and parenchymal cysts were excluded for analysis, along with their matched controls. Uninfected control subjects were seronegative on EITB.
All patients with parenchymal cysts (n = 16) had at least two viable brain cysts. Eight of these patients (8/16, 50%) also had additional calcified cysts and one (1/16, 6%) also had a degenerating cyst. However, none of these patients had subarachnoid disease apparent on their brain MRI examinations. In the subarachnoid NCC group (n = 13), five patients (5/13, 38.5%) also had parenchymal calcifications, but none had cysts in the parenchyma (S1 Table).
The cytokine concentrations (pg/ml) in NCC patients were normalized to the values in the matched control as a ratio of patient:matched control and the mean values (± standard error of the mean (SEM)) of the parenchymal NCC patients were compared to those of subarachnoid NCC patients (Fig 1). Compared to patients with subarachnoid infections, patients with parenchymal cysts had higher levels of the circulating pro-inflammatory cytokines IL-12 (patient/control ratio (r) = 6.7±2.9 for parenchymal and r = 0.79±0.23 for subarachnoid disease), IFN-γ (patient/control r = 2.8±1.2 for parenchymal and r = 0.94±0.16 for subarachnoid disease; p<0.05, Fig 1A). Strong trends towards higher levels of IL-1β (p = 0.052), IL-8 (p = 0.074) and IL-9 (p = 0.067) were also apparent in parenchymal NCC compared to subarachnoid NCC (Fig 1D). Individuals with parenchymal cysts also showed higher Th2 cytokines: IL-4 (patient/control r = 2.0±0.3 for parenchymal and r = 0.94±0.08 for subarachnoid disease) and IL-13 (patient/control r = 2.4±0.4; for parenchymal and r = 0.84±0.14 for subarachnoid disease p<0.05; Fig 1B), and higher levels of the growth factors G-CSF (patient/control r = 2.0±0.3 for parenchymal and r = 1.05±0.19 for subarachnoid disease) and VEGF (patient/control r = 2.0±0.4; for parenchymal and r = 0.77±0.21 for subarachnoid disease; p<0.05, Fig 1C). Similar trends towards higher levels of IL-5 (p = 0.66) and eotaxin (p = 0.096) were observed in parenchymal vs. subarachnoid NCC (Fig 1D). No significant differences were noted between the two groups for PDGF-B, IL-1rα, IL-6, IL-7, IL-10, IL-17A, basic FGF, IP-10, MCP-1, MIP-1α, MIP-1β, RANTES, and TNF-α in the serum. No IL-2, IL-15 and GM-CSF were detected in sera from any of the study groups (Fig 1D). In contrast to the normalized cytokine data discussed above, when cytokines in serum expressed as pg/ml were compared between the NCC patients and controls, patients with subarachnoid disease show higher levels of IL-1α, IL-8 (S1A Fig; p<0.05) and IL-5 (S1B Fig; p<0.05) than patients with parenchymal NCC and controls subjects, whereas MCP-1, MIP-1α, RANTES (S1C Fig), IL-4 (S1B Fig), PDGF, G-CSF (S1D Fig) were higher (p<0.05) in subarachnoid patients than in the parenchymal group.
The optimal antigen concentration for PBMC stimulation was determined using fresh PBMC from subjects without NCC to select a dose range with low non-specific stimulatory effects. For in vitro cell proliferation (measured by thymidine incorporation into DNA), 20 to 40 μg/ml of TsAg was determined to be optimal with the best stimulation indices (SIs) at 6 days (SI at 20 μg/ml: 6.8±2.1 and 40 μg/ml: 12.5±5.1) (S2 Fig). Subsequently, these TsAg concentrations were employed for in vitro stimulation of PBMC from patients with NCC for comparison of antigen-driven cytokine secretion between patients with parenchymal and subarachnoid disease.
To assess if TsAg-induced cytokine production in vitro differed between the two clinical forms of NCC, we evaluated the cytokine levels in supernatants of PBMC cultures from parenchymal and subarachnoid patient groups, normalized for their matched controls, following in vitro stimulation with parasite antigen. We tested a total of 27 cytokines and found that only IL-10 values differed significantly between the two groups of patients (S1 Table and Fig 2). After 48h of Ag stimulation, when parenchymal and subarachnoid group were compared for IL-10 production, PBMC from the subarachnoid group elicited a higher IL-10 production (patient/control r = 62.99±20.6) in response to TsAg (40 μg/ml) than the parenchymal group (patient/control r = 9.4±3.8; p<0.05; Fig 2). There were strong trends towards higher levels of TNF-α (p = 0.051) and lower levels of IL-5 (p = 0.075) in parenchymal NCC compared to subarachnoid NCC (S2 Table). In contrast to the normalized data, when the cytokines levels expressed in pg/ml were directly compared between the study groups (S3 Fig), PBMC from patients with parenchymal disease TsAg produced higher levels of TsAg-driven IL-2, IL-4, IL-5, IL-8, IL-10, IL-12, IL-13, MCP-1, MIP-1β and TNF-α than patients with subarachnoid disease and control subjects (S3 Fig; p<0.05).
We phenotyped a number of lymphocyte subsets in order to identify the PBMC population that may have expanded in response to T. solium exposure and may, thus, have roles in regulating inflammatory responses in the two NCC study groups, with either parenchymal or subarachnoid NCC. These data were also normalized for matched controls as we did for cytokine data in serum and supernatants (above), to eliminate the influence of age and gender. The cell subpopulations enumerated included T cells, B cells and NK cells (CD16+CD56+) (Fig 3A). The only lymphocyte subpopulation that was found to differ significantly between the parenchymal NCC and subarachnoid NCC groups was the CD16/CD56 positive (NK) subset, indicating an expansion of these cells in the latter group (Fig 3B). As with the differences in cytokine levels in serum and supernatants of PBMC, we found that differences in frequency data for cell phenotypes between parenchymal and subarachnoid NCC were altered when analyzed after normalization with corresponding controls (See S4 and S5 Figs), suggesting that age and gender affected the expression of these parameters.
When the frequencies of Treg cells, normalized for gender and age, were compared (Fig 4), differences between subjects with parenchymal and subarachnoid NCC ex vivo (S4A Fig) or after TsAg stimulation did not achieve statistical significance (Fig 4C). However, there was a strong trend towards higher frequencies of CD4 cells ex vivo (p = 0.057; S4B Fig) and Treg cells (p = 0.051; S5F Fig) in parenchymal NCC compared to subarachnoid disease. Taken together, these data suggest that parenchymal NCC is associated with a predominat pro-inflammatory response. Further, in vitro stimulation with parasite antigens led to an expansion of Treg cells that, perhaps due to the small sample size, did not achieve statistical significance. Notably, in the case of antigen-stimulated responses, normalization with matched control subjects did not change the conclusions from the non-normalized frequency data (Figs 4C and S5F).
Cerebral cysticercosis is a remarkably complex and heterogeneous infection that results in a wide spectrum of clinical manifestations and a variable prognosis. A major determinant of the severity of NCC and the nature of it’s clinical manifestations is the parasite location: parenchymal cysts are frequently associated with seizures but generally have a better outcome following treatment than does extraparenchymal disease, which is commonly associated with serious neurological and vascular complications such as intracranial hypertension, mass effects, strokes and frequently, with fatality [3–5,12]. The present study was designed to investigate the relationship between the T. solium specific-immune responses and disease manifestations, with a goal of understanding the mechanisms underlying clinical heterogeneity in NCC. Based largely on the prominent symptomatology and inflammatory reaction in the CSF in subarachnoid NCC (compared to parenchymal cysts) we hypothesized that subarachnoid NCC would be associated with a pro-inflammatory and regulatory immunological state compared to parenchymal NCC. This was a possible consequence of encapsulation of the parasite in parenchymal disease by the inflammatory capsule and sequestration from an activated immune system [11]. However, our findings did not support our proposed model of immune responses in the two forms of NCC.
Contrary to our a priori hypothesis the major, and unexpected, finding in our study was that subarachnoid NCC was associated with a less inflammatory and stronger regulatory immune response to parasite antigens. The key observation was that circulating cytokine levels indicated a higher pro-inflammatory status in parenchymal disease, reflected in IL-12 and IFN-γ levels (Fig 1A), compared to subarachnoid disease. Interestingly, this was also accompanied by higher levels of IL-4, IL-13 and growth factors GM-CSF and VEGF (Fig 1B and 1C) suggesting a global immune activation. This was unexpected because the high parasite burden and exuberant growth of the parasite in subarachnoid disease suggested, to us, a stronger inflammatory response than the constrained growth of the parasite in parenchymal locations, in what is considered an immunologically privileged organ. However, patients with subarachnoid cysts demonstrated a systemic cytokine profile similar to that of the control subjects (as shown the patient/control ratio ≤1) with lower levels of circulating pro-inflammatory mediators, circulating Th2 (IL-4 and IL-13; Fig 1B) and higher levels of TsAg-driven IL-10 in vitro than those with parenchymal cysts (Fig 2). Of note, higher levels of inflammation in pericystic tissues surrounding parenchymal cysts than around meningeal cysts have also been observed in a pig model of neurocysticercosis, when naturally infected pigs were treated with praziquantel triggering an acute host response (Cangalaya et al., submitted for publication), and in experimental intracerebral infections with T. solium in rats [17].
The prominent regulatory environment we observed in subarachnoid NCC may result from one or more immunological mechanisms. One explanation could be that in subarachnoid disease, because of greater parasite growth in the unconstrained subarachnoid space, a higher local antigen load than occurs in the smaller, constrained cysts located in the parenchyma. In this setting, a higher IL-10 levels can be induced, as reported by Chavarria et al. [11] in CSF from patients with subarachnoid NCC disease with severe symptoms. It is also possible that greater access to the parasite antigens in the subarachnoid location (compared to parenchymal cysts) may promote a stronger or preferentially immunosuppressive response in extraparenchymal sites, either directly through parasite-derived molecules, or indirectly through manipulation of the host immune system. This mechanism could explain the relative downregulation of inflammation in subarachnoid NCC, which is associated with high parasite burdens in locations accessible to the immune effector cells, whereas in parenchymal disease, encapsulation of the parasite and the consequent sequestration of parasite-derived molecules may inhibit the downregulatory stimulus that potentially protects the parasite. However, in the current cross-sectional study we were unable to directly test this hypothesis.
Some of our observations differ from reports from other investigators who have characterized the immune responses to parasite Ags in NCC patients. Recently, the observation of expanded Tregs subpopulations in NCC patients and after in vitro stimulation with T. solium antigens, has lent support to the notion that Tregs may help in limiting the inflammatory response [10,18–21]. The current, carefully controlled study showed a strong trend towards higher frequencies of Treg cells in subarachnoid NCC compared to parencehymal disease after in vitro stimulation (p = 0.051; S5F Fig). Unfortunately, our analysis does not allow us to determine functional differences in the Treg cell populations between the study populations. Functional analysis of these cells or other cellular sources of regulatory cytokines (such as macrophages and Th2 cells) might help to explain their role in the generating the stronger regulatory state associated with subarachnoid NCC. Additionally, our finding of an enhanced regulatory profile in subarachnoid disease is also in contrast to previous studies that reported an increased inflammatory cytokine response in the subarachnoid space, using CSF, compared to the parenchyma [11,12,21]. We believe that the homogeneous clinical features, careful matching with controls for age and gender and absence of treatment in our group of patients support the validity of our observations and may explain the differences found in the immune profile in this study compared to previous reports.
A remarkable aspect about the suppressive or regulatory immune response observed in the subarachnoid NCC patients in our study was that a high IL-10 production in SA patients was seen mainly after stimulation with high concentrations of TsAg (40 μg/ml; Fig 2); this suggests that strong suppressor activity may be restricted to the proximity of cyst where a levels of released parasite antigens are high. Evidence for immunomodulation by parasite antigen that induced CD4+CD25+FoxP3+ Tregs to produce regulatory cytokines such as IL-10 has been reported with Schistosoma mansoni antigens [22] and also in infections caused by Echinococcus [23,24] and T. solium [10,18].
We analyzed fresh blood/serum samples from well-defined patients at presentation to a clinical center, before they had initiated steroid or antiparasitic treatment to minimize within-group heterogeneity in disease manifestations, a factor that has compromised some previous studies that relied on symptomatology for disease classification. Disease categories in this study were based strictly on the location of the cysticerci (in the parenchymal or subarachnoid space), defining two populations, one with parenchymal disease alone, and a second with subarachnoid disease without parenchymal cysts. In addition the patients were matched by age and gender to a control group, making for more robust comparisons by controlling for immunologically mediated environmental factors and for other helminth infections. We believe that this approach allowed us to have well-controlled and clinically uniform patient groups.
Although we limited our analysis of cell populations to circulating immune cells, there is good reason to expect that the differences in circulating cell subpopulations between the study groups likely reflect those in the CSF compartment, particularly in subarachnoid infections. This has been reported in a previous study comparing blood and CSF cell phenotype from NCC patients [10]. The blood brain barrier has been shown to be compromised around degenerating cysts which would allow two-way trafficking of cells between the circulation and the CNS [25]. Furthermore, in clinical settings and future investigations, blood samples are vastly more accessible than CSF samples from patients, and not technically limited by small numbers of cells present in CSF.
The systemic immune responses, as seen from the cytokine profile analysis of sera from NCC patients with cysts in the parenchyma, elicits not only a broad repertoire of inflammatory cytokines and Th2 responses but also of growth factors (G-CSF and VEGF) (Fig 1C). During cyst development, the survival and growth of the parasite is dependent on vascularization and angiogenesis, and an increased production of VEGF could be beneficial for this process by promoting the development of capsule that functionally isolates the parasite, and may protect it from interactions with damaging responses [26].
Although the analysis of cytokine levels and responses to TsAg stimulation show that patients with parenchymal NCC disease exhibited a higher systemic inflammatory profile compared to subarachnoid NCC disease, few differences were observed in the frequencies of a number of lymphocyte subpopulations ex vivo, with or without normalization with matched controls. The lack of differences between parenchymal and subarachnoid NCC in the phenotypic analysis does not rule out functional differences between lymphocyte subpopulations in the two groups. In this regard, we did find that Ag stimulation induced a greater expansion of NK cells in the subarachnoid NCC patients than in the parenchymal patients (Fig 3B). This expansion may represent a reaction to the downregulatory stimulus to T cells associated with the parasite in subarachnoid locations, although the NK cells role in inflammatory neurological disorders is still unclear [27]. Interestingly, differences in the outcome of comparisons when data were normalized or not normalized for matched controls strongly suggest that matching of patients and control groups is important and appropriate in this disease, and that the age and gender influence the immune response to this parasite significantly.
The marked differences in pathophysiology between parenchymal and subarachnoid infections, with chronic inflammation and progressive disease being very frequent in subarachnoid cases, suggested to us that subarachnoid disease would induce a more pro-inflammatory response than would parenchymal disease. Our data shows that the location of infection influences the nature of the peripheral immune response, and that unexpectedly and perhaps counter-intuitively, subarachnoid disease, the form that presents clinically with inflammatory pathology in the CNS has a more regulatory immune environment than parenchymal disease. In this type of cross-sectional study design it is not possible to determine if the regulatory state in subarachnoid disease is the consequence of an anti-inflammatory (or regulatory) response to a stronger inflammatory response than that seen in parenchymal NCC. Understanding the immune pathogenesis of NCC might in the future lead to more specific treatment by modeling the immune response.
Armando E. Gonzalez, DVM, PhD; Victor C.W.Tsang, PhD (Coordination Board); Herbert Saavedra, MD; Manuel Martinez, MD; Manuel Alvarado, MD (Instituto Nacional de Ciencias Neurológicas, Lima, Perú); Manuela Verastegui, PhD; Mirko Zimic, PhD; Javier Bustos, MD, MPH; Cristina Guerra, PhD; Yesenia Castillo, MSc; Yagahira Castro, MSc (Universidad Peruana Cayetano Heredia, Lima, Perú); Maria T. Lopez, DVM, PhD; Cesar M. Gavidia, DVM, PhD (School of Veterinary Medicine, Universidad Nacional Mayor de San Marcos, Lima, Perú); Luz M. Moyano, MD; Viterbo Ayvar, DVM (Cysticercosis Elimination Program, Tumbes, Perú); Theodore E. Nash, MD; John Noh, BS, Sukwan Handali, MD (CDC, Atlanta, GA); Jon Friedland (Imperial College, London, UK).
|
10.1371/journal.pgen.1006233 | The DUF59 Containing Protein SufT Is Involved in the Maturation of Iron-Sulfur (FeS) Proteins during Conditions of High FeS Cofactor Demand in Staphylococcus aureus | Proteins containing DUF59 domains have roles in iron-sulfur (FeS) cluster assembly and are widespread throughout Eukarya, Bacteria, and Archaea. However, the function(s) of this domain is unknown. Staphylococcus aureus SufT is composed solely of a DUF59 domain. We noted that sufT is often co-localized with sufBC, which encode for the Suf FeS cluster biosynthetic machinery. Phylogenetic analyses indicated that sufT was recruited to the suf operon, suggesting a role for SufT in FeS cluster assembly. A S. aureus ΔsufT mutant was defective in the assembly of FeS proteins. The DUF59 protein Rv1466 from Mycobacterium tuberculosis partially corrected the phenotypes of a ΔsufT mutant, consistent with a widespread role for DUF59 in FeS protein maturation. SufT was dispensable for FeS protein maturation during conditions that imposed a low cellular demand for FeS cluster assembly. In contrast, the role of SufT was maximal during conditions imposing a high demand for FeS cluster assembly. SufT was not involved in the repair of FeS clusters damaged by reactive oxygen species or in the physical protection of FeS clusters from oxidants. Nfu is a FeS cluster carrier and nfu displayed synergy with sufT. Furthermore, introduction of nfu upon a multicopy plasmid partially corrected the phenotypes of the ΔsufT mutant. Biofilm formation and exoprotein production are critical for S. aureus pathogenesis and vancomycin is a drug of last-resort to treat staphylococcal infections. Defective FeS protein maturation resulted in increased biofilm formation, decreased production of exoproteins, increased resistance to vancomycin, and the appearance of phenotypes consistent with vancomycin-intermediate resistant S. aureus. We propose that SufT, and by extension the DUF59 domain, is an accessory factor that functions in the maturation of FeS proteins. In S. aureus, the involvement of SufT is maximal during conditions of high demand for FeS proteins.
| Iron-sulfur (FeS) clusters are inorganic cofactors that are used for diverse cellular processes including cellular respiration, DNA replication and repair, antibiotic resistance, and dinitrogen fixation. A failure to properly assemble FeS clusters in proteins results in widespread metabolic disorders, metabolic paralysis, and oftentimes cell death. Therefore, the biosynthesis of FeS clusters is essential for nearly all organisms. Proteins containing DUF59 domains are widespread in Eukarya, Bacteria, and Archaea. Proteins containing DUF59 domains have roles in FeS cluster assembly, but the function(s) of the DUF59 domain is unknown. Moreover, the function(s) of proteins containing DUF59 domains are largely unknown. Staphylococcus aureus SufT is composed solely of a DUF59 domain, which provides a unique opportunity to examine the role(s) of this domain in cellular physiology. In this report we show SufT to be an accessory factor utilized in FeS cluster assembly during conditions imposing a high-demand for FeS proteins. We also show that deficiencies in the maturation of FeS proteins result in alterations in the ability of S. aureus, an epidemic human pathogen, to form biofilms, produce exoproteins, and resist antibiotic stress.
| Iron (Fe) is an essential nutrient for nearly all organisms. Fe is acquired from the environment and is transported into cells using specific uptake systems. Studies have shown that ~80% of the intracellular Fe is located in inorganic cofactors, called iron-sulfur (FeS) clusters, and heme in a respiring microorganism [1].
The metabolisms of most organisms are highly reliant on FeS cluster chemistry and a failure to properly assemble FeS clusters in proteins can result in widespread metabolic disorders, metabolic paralysis, and cell death [2,3,4]. FeS proteins function in diverse metabolic processes including environmental sensing[5], carbon transformations [6], DNA repair and replication [7,8], RNA metabolism [9], protein synthesis [10], nucleotide, vitamin, and cofactor synthesis [11,12,13], and cellular respiration [14,15,16]. FeS clusters are typically found in proteins as [Fe2S2] or [Fe4S4] clusters, but the use of complex FeS clusters has evolved for processes such as dinitrogen [17], carbon monoxide [18], and hydrogen metabolism [19].
Iron and sulfur (S) ions are often toxic to cells resulting in the evolution of tightly controlled mechanisms to synthesize FeS clusters from their monoatomic precursors [20,21]. Three FeS cluster biosynthetic systems (Nif, Suf, and Isc) have been described in Bacteria and Archaea for the synthesis of [Fe2S2] and [Fe4S4] clusters [22,23,24]. Bioinformatic analyses suggest that the Suf system is the most prevalent machinery in Bacteria and Archaea and perhaps the most ancient [25].
The Suf, Nif, and Isc systems utilize a common strategy to synthesize FeS clusters. First, sulfur is mobilized from free cysteine (typically), using a cysteine desulfurase enzyme and subsequently transferred to either a sulfur carrier molecule (SufU or SufE) or directly to the synthesis machinery [24,26,27]. Monoatomic iron and sulfur, along with electrons, are combined upon a molecular scaffolding protein (SufBD in S. aureus) to form an FeS cluster [28]. The FeS cluster can be transferred directly from the scaffold to a target apo-protein or it can be transferred to a carrier molecule that subsequently traffics the cluster to a target apo-protein and facilitates maturation of the holo-protein [29]. Nfu and SufA serve as FeS cluster carriers in Staphylococcus aureus [4,30]. Nfu is necessary for virulence in models of infection [4,31]
Most studies on bacterial FeS cluster assembly have been conducted using Escherichia coli and Azotobacter vinelandii. E. coli encodes for both the Suf and Isc systems [22] whereas A. vinelandii encodes for the Isc and Nif systems [32]. In contrast, few studies have been conducted on FeS cluster assembly in gram-positive bacteria such as Bacillus subtilis or S. aureus, which encode for only the Suf system [4,27]. Recent findings suggest that SufCDSUB are essential for S. aureus viability, confirming that Suf is the sole FeS cluster biosynthetic machinery used under laboratory growth conditions [4,33,34].
Dioxygen can accept electrons from cellular factors resulting in the spontaneous generation of reactive oxygen species (ROS) such as hydrogen peroxide (H2O2) and superoxide [35,36,37]. FeS clusters are among the primary cellular targets of H2O2 and superoxide [38,39]. ROS readily oxidize solvent exposed [Fe4S4]2+ cofactors of enzymes such as aconitase (AcnA) [38,39]. Oxidation results in conversion to an inactive [Fe3S4]1+ cluster that can be repaired back to the active [Fe4S4]2+ state using Fe2+ and an electron [40]. Studies have implicated roles for cysteine desulfurase (IscS) and the putative Fe donors CyaY, YtfE, and YggX in the repair of oxidized clusters [40,41,42]. Cells also employ mechanisms to physically protect FeS clusters. The Shethna protein shields the FeS cofactor of dinitrogen reductase from dioxygen exposure [43]. Alternatively, protein domains can be situated in a manner that prevents oxidants from interacting with the FeS cluster. The pyruvate:ferredoxin oxidoreductase (PFOR) from Desulfovibrio africanus was found to have greater stability in the presence of dioxygen, relative to alternate PFOR enzymes, due to the presence of a domain that prevents the interaction of oxidants with its [Fe4S4]2+ cluster [44].
We have identified an open reading frame (ORF) in S. aureus that is often associated with the suf operon in a number of bacterial and archaeal genomes. The ORF (SAUSA300_0875) encodes for a protein composed solely of a DUF59 domain and is annotated as SufT since it is often found in operons with a cysteine desulfurase (i.e. SufS) [45]. In eukaryotic cells, the CIA2 (also identified as Fam96a/b or AE7) FeS cluster assembly factor(s) contain a DUF59 domain [46,47]. CIA2a and CIA2b act downstream of the cytosolic iron-sulfur assembly (CIA) machinery and are required for the maturation of FeS cluster proteins. A DUF59 domain is also present in the Arabidopsis thaliana chloroplast FeS cluster carrier, HCF101, which is required for photosystem I maturation [48].
S. aureus is a leading cause of human infectious disease related morbidity and mortality worldwide. S. aureus forms surface associated communities referred to as biofilms that are critical for S. aureus pathogenesis and biofilm associated cells serve as the etiologic agents of recurrent staphylococcal infections (reviewed here [49]). S. aureus also secretes a variety of toxins and enzymes into its extracellular milleu that are critical for biofilm formation, host colonization, nutrient acquisition and survival in the human host (reviewed here [50]). About 60% of the secretome consists of peptide toxins (phenol soluble modulins (PSM's), which have multiple key roles in pathogenesis [51,52].
Since the 1990s the proportion of infections caused by community-associated methicillin resistant S. aureus (CA-MRSA) has been steadily increasing and has now reached near epidemic levels [53]. Vancomycin is a glycopeptide antibiotic that has traditionally been regarded as a last-resort drug for the treatment of MRSA infections [54]. Strains have recently emerged that display intermediate (vancomycin intermediate-resistant S. aureus; VISA) or high (vancomycin resistant S. aureus; VRSA) levels of resistance towards vancomycin [54,55]. Among the characteristics of VISA strains are decreased activity of peptidoglycan hydrolases and alterations in their cell wall that results in increased resistance to the lytic enzyme lysostaphin [55].
S. aureus provides an excellent model to assess the role of the DUF59 domain (SufT) in cellular physiology. In this report we present phylogenetic analyses indicating a widespread distribution for SufT and conservation of SufT homologs in bacterial and archaeal taxa that utilize the Suf system. These analyses also suggest that sufT was recruited to the neighborhood of sufBC over evolutionary time and for the most part retained. The bioinformatic analyses led us to hypothesize that SufT has a role in the maturation of FeS proteins. Results demonstrate an involvement of SufT in the maturation of FeS proteins during conditions imposing a high demand for FeS proteins. Moreover, epistasis studies show that the nfu and sufT mutations display synergy and the introduction of nfu in multicopy partially corrects the phenotypes of a sufT mutant. Deficiencies in the maturation of FeS proteins also result in increased biofilm formation, decreased exoprotein production, and the appearance of phenotypes consistent with vancomycin-intermediate resistant S. aureus (VISA). We propose that SufT functions as an auxiliary factor for the maturation of FeS proteins with maximum usage during conditions of high FeS cofactor demand.
Of the 1669 complete genome sequences available as of October 2011 and compiled as part of our previously published work on the evolution of Suf [25], 1092 (65.4% of total) encoded for SufBC. Among these genomes, 761 (69.7% of total) encoded for SufT. Of the 1669 genomes, 68 genomes contained sufT, but not sufBC. Five genomes contained sufT, but not sufB, iscU, or nifU, which encode for FeS cluster scaffolding molecules. These genomes were all from lactobacilli and the sufT homologues are in apparent operons with the genes encoding for either anaerobic ribonucleoside-triphosphate activating enzyme or serine dehydratase, which are FeS cluster-requiring enzymes [11,56].
Among the 761 genomes that encoded for sufT and sufBC, 374 of the sufT homologs were localized with sufBC (suf operon associated) and 387 sufT homologs were not associated with sufBC (non-suf operon associated). Maximum likelihood phylogenetic reconstructions of SufT (unrooted) and SufBC (rooted), followed by overlays of suf-operon associated and non-suf operon associated sufT, indicate that sufT has been recruited to and lost from the suf operon multiple times during its evolutionary history (Fig 1). However, the overall trend appears to be retainment once sufT was recruited to the suf operon. Mapping of the association of sufT with the suf operon on the SufBC tree indicates that sufT was not associated with the suf operon early during the evolution of taxa that used the Suf FeS cluster biosynthetic system and that it was recruited to the operon recently in its evolutionary history. Each SufT homolog identified contained a conserved cysteine residue, which was previously shown to be hyper-reactive [57], but described FeS cluster-binding motifs were not recognized.
Of the total (n = 761) identified SufT homologs, the predominant structure contained only the DUF59 domain (S1 architecture; ex. S. aureus SufT), but 198 encoded for additional N- and C-terminal motifs represented by nine primary modular structures (Fig 2A). The most prevalent modular structure was the S2 architecture (n = 88), with a N-terminal motif that did not display homology to previously described domains. SufT within the S5 architecture (n = 5) contained a N-terminal domain with homology to U-type FeS cluster scaffolds while SufT within the S7 architecture (n = 3) harbored a N-terminal domain with homology to Rieske iron-oxygenase ferredoxins. Finally, SufT within the S9 architecture (n = 1) contained a N-terminal domain with homology to serine acetyltransferases (CysE). Characterization of the C-terminal motifs also revealed variation that was represented in four unique modular structures. These were characterized as SufT with C-terminal domains that have homology to PaaJ or acetyl-CoA acetyltransferase domains (S3 architecture, n = 75), P-loop NTPase domains (S4 architecture n = 20), DUF1858 domains (S6 architecture, n = 4) and co-enzyme pyrroloquinoline quinone synthesis protein D (PqqD) domains (S8 architecture, n = 2). The S2-S9 architectures were mapped on the phylogenetic reconstruction of core DUF59 (N- and C-terminal motifs were pruned from alignment block) in order to determine if the modules are randomly distributed over the tree or if they are phylogenetically clustered. The overall pattern of clustering of the modular structures on the tree (Fig 2B) indicates that once these modules were fused to an ancestor of a given DUF59 containing protein, they were largely retained. This suggests that the N- and C-terminal motifs, and presumably their functionalities, are under strong selective pressure.
We created and characterized a S. aureus ΔsufT mutant to test whether SufT has a role in the maturation of FeS proteins.
A S. aureus ΔacnA strain is defective in utilizing glutamate as a source of carbon (S1A Fig) [58,59]. Nfu has a role in the maturation of AcnA in S. aureus [4]. The Δnfu and ΔsufT strains displayed growth defects in chemically defined media supplemented with glutamate as a carbon source (hereafter 20AA glutamate medium) (Fig 3A), but the defect of the ΔsufT strain was less severe than that of the Δnfu strain. The WT, Δnfu, and ΔsufT strains had similar growth profiles in defined medium containing glucose as a carbon source (hereafter 20 AA glucose medium) (S1B Fig).
AcnA activity was assessed in the WT, ΔsufT, and Δnfu strains across growth. AcnA activity was decreased in strains lacking Nfu or SufT (Fig 3B). The decreased AcnA activity in the ΔsufT strain could arise due to one of four scenarios: 1) decreased transcription of acnA, 2) decreased abundance of AcnA, 3) decreased occupancy of the [Fe4S4] cluster upon AcnA due to the decreased transcription of genes encoding FeS cluster biogenesis factors, or 4) decreased cluster occupancy upon AcnA due to the absence of SufT.
Transcriptional activity of acnA was increased in the ΔsufT strain (S2 Fig). This suggested that decreased AcnA activity in the ΔsufT strain was not the result of altered acnA transcription (S2 Fig). We constructed acnA::TN strains containing a plasmid with a acnA_FLAG allele under the transcriptional control of a xylose inducible promoter (pacnA). Introduction of pacnA allows for the control of acnA transcription and the simultaneous determination of AcnA_FLAG abundance. The acnA::TN ΔsufT strain was genetically complemented by re-introduction of the sufT allele at a secondary chromosomal location (sufT+). AcnA activity and AcnA abundance was assessed in the acnA::TN, acnA::TN ΔsufT, and acnA::TN ΔsufT sufT+ strains containing pacnA. AcnA activity was ~2-fold lower in the acnA::TN ΔsufT strain compared to the acnA::TN when activity was normalized to AcnA abundance in the same cell-free lysates (Fig 3C). This phenotype was genetically complemented.
Suf is encoded by the sufCDSUB operon in S. aureus. The transcriptional activity of sufC was increased (~2-fold) in the Δnfu strain and mildly, but consistently, increased in the ΔsufT strain (Fig 3D). Similar results were obtained in exponential and stationary growth. From Fig 3 we concluded that the absence of SufT results in decreased occupancy of the [Fe4S4] cofactor upon AcnA.
Synthesis of the branched chain amino acids (BCAA) leucine and isoleucine requires the FeS cluster containing dehydratase enzymes isopropylmalate isomerase (LeuCD) and dihydroxyacid dehydratase (IlvD), respectively [60,61]. Strains lacking either SufT or Nfu displayed growth defects in defined medium lacking leucine (Leu) and isoleucine (Ile) (hereafter 18AA glucose medium) (Fig 4A), but displayed a growth profile similar to WT in 20AA glucose medium (S1B Fig).
We constructed leuC::TN, leuC::TN ΔsufT, leuC::TN ΔsufT sufT+, ilvD::TN, ilvD::TN ΔsufT, and the ilvD::TN ΔsufT sufT+ strains carrying plasmids with either leuCD or ilvD under the transcriptional control of a xylose inducible promoter (pleuCD and pilvD). The activities of LeuCD and IlvD were decreased in strains lacking SufT and these defects were restored by genetic complementation (Fig 4B and 4C). We concluded that SufT is utilized in the maturation of multiple FeS cluster requiring enzymes.
Staphylococcus aureus is a facultative anaerobe and can respire upon dioxygen or nitrate as terminal electron acceptors or grow fermentatively [62]. The acnA::TN and acnA::TN ΔsufT strains containing pacnA were cultured aerobically, as well as anaerobically in the presence or absence of nitrate before determining AcnA activity. The ΔsufT mutant had lower AcnA activity during respiratory growth, but AcnA activity was restored during fermentative growth (Fig 5A). Microaerobic conditions also mitigated the growth defect of both the Δnfu and ΔsufT strains in 18AA glucose medium (S3 Fig).
Fermentative growth imposes a decreased demand for FeS clusters [63]. By inference, fermentative growth should result in decreased transcription of genes encoding for FeS assembly factors. Consistent with this prediction, the transcriptional activities of sufT, nfu, and sufC decreased when aerobically cultured cells were shifted to an anaerobic (fermentative) environment (Fig 5B).
We examined whether SufT functions to protect the AcnA FeS cluster via physical exclusion of dioxygen. Cell-free lysates were generated from the acnA::TN and acnA::TN ΔsufT strains containing pacnA. AcnA activity was assessed at periodic intervals before and after exposure of lysates to dioxygen. Dioxygen exposure resulted in decreased AcnA activity in both the parent and ΔsufT mutant (Fig 5C), but the rate of decrease was statistically indistinguishable between the strains.
Fermentatively cultured cells exposed to dioxygen (reaeration) increased sufC transcription suggesting that the resumption of respiratory processes results in an increased demand for FeS clusters (Fig 6A and [4]). The transcription of sufT was also increased (~2.5-fold) upon reaeration (Fig 6A).
The role of SufT in the maturation of AcnA upon reaeration was assessed. The acnA::TN and acnA::TN ΔsufT strains containing pacnA were cultured fermentatively before one set of the cultures was exposed to dioxygen while the other set was incubated anaerobically (as previously described [40]). AcnA activity increased by ~30% in the parental strain upon dioxygen introduction (Fig 6B). In contrast, AcnA activity decreased by ~20% in the ΔsufT mutant. The use of protein synthesis inhibitors allowed for the conclusion that the increased AcnA activity in the parental strain upon reaeration was due to de novo protein synthesis. These findings led to the conclusion that SufT has a role in FeS cluster assembly in cells attempting to resume respiratory processes, and thereby facilitates the adaptation of cells to shifts in dioxygen tensions.
Reactive univalent species can damage or destroy solvent exposed FeS clusters [4,38,39]. We found that the ΔsufT, and sodA::TN (encoding for the dominant aerobic superoxide dismutase [64]) strains displayed decreased growth in the presence of paraquat, a redox cycling molecule that leads to increased accumulation of intracellular ROS (Fig 7A). However, the phenotype of the ΔsufT mutant was less severe than that of the sodA::TN strain.
The acnA::TN and acnA::TN ΔsufT strains containing pacnA were cultured, challenged with paraquat, and AcnA activity was determined. Challenging cells with paraquat resulted in ~15% and ~45% decrease in AcnA activity in the parent and ΔsufT mutant, respectively (Fig 7B).
The alkylhydroperoxidase system (Ahp) functions as an intracellular H2O2 scavenger and a S. aureus strain lacking Ahp accumulates intracellular ROS [4,65]. AcnA activity was assessed in the WT, ΔsufT, ahp::TN, and ahp::TN ΔsufT strains. AcnA activity was decreased ~25–30% in both the ahp and sufT strains and by ~75% in the ahp sufT double mutant strain (Fig 7C).
Four explanations could underlie the decreased AcnA activity observed in a ΔsufT strain upon ROS toxification: 1) the ΔsufT strain has decreased activities of ROS scavenging enzymes, 2) SufT is necessary for the repair of FeS clusters inactivated by ROS oxidation, 3) SufT is involved in physically shielding and/or excluding ROS from the enzyme active site and preventing damage, or 4) there is an increased need for SufT in FeS cluster assembly.
The activities of the ROS scavenging enzymes catalase (Kat) and superoxide dismutase (Sod) were similar in the WT and ΔsufT strains across growth (Fig 7D, S4 Fig). The acnA::TN and acnA::TN ΔsufT strains containing pacnA also displayed similar levels of Sod activity, both before and after paraquat treatment (S5 Fig).
We examined whether SufT is capable of physically shielding FeS clusters from univalent oxidants [43,44]. Cell-free lysates from the acnA::TN and acnA::TN ΔsufT strains containing pacnA were exposed to varying concentrations of H2O2 and AcnA activity was determined one minute post treatment. AcnA activity decreased with increasing H2O2 concentrations, but the decrease in AcnA activity was similar in the parent and ΔsufT mutant (Fig 7E).
Brief exposure to H2O2 can convert the active [Fe4S4]2+ cluster in AcnA into the inactive [Fe3S4]1+ cluster. This can be repaired to the [Fe4S4]2+ state by Fe2+ and an electron [40]. Cell-free lysates from the acnA::TN and acnA::TN ΔsufT strains containing pacnA were exposed to H2O2. One-minute post challenge, the stress was terminated and reactivation of AcnA activity by factors in the lysate was monitored over-time. The rate of AcnA reactivation was similar in the parent and ΔsufT mutant (Fig 7F). From Fig 7 we concluded that SufT is involved in the de novo assembly of FeS clusters in cells experiencing ROS stress.
The phenotypic abnormalities of the ΔsufT mutant were exacerbated during respiration, during resumption of respiration in fermenting cells, and upon ROS challenge (i.e. conditions imposing a high demand for FeS assembly). The transcription of core genes required for FeS assembly increased upon challenge with ROS or resumption of respiration [4].
We tested the hypothesis that SufT is required for FeS cluster assembly during conditions imposing a high demand for FeS clusters. Growth was monitored in either 20AA glutamate medium, or defined medium containing glutamate as a carbon source and lacking leucine (Leu) and isoleucine (Ile) (hereafter 18AA glutamate medium). Growth in 18AA glutamate medium would impose a simultaneous requirement for the AcnA, LeuCD, and IlvD enzymes, and by inference, exert an increased requirement for FeS clusters. The ΔsufT strain displayed a growth defect in 20AA glutamate medium (similar to Fig 3A; however the magnitude appears lower here due to the scale) and this defect was exacerbated upon culture in 18AA glutamate medium (Fig 8A).
The acnA::TN and acnA::TN ΔsufT strains containing pacnA were cultured in the presence or absence of varying concentrations of xylose followed by assessing AcnA activity. The difference in AcnA activity between the parent and ΔsufT mutant increased in synchrony with increasing inducer concentrations (Fig 8B and 8C).
We next monitored sufT transcriptional activity with respect to the demand for FeS clusters using the acnA::TN strain carrying pacnA, as well as the sufT transcriptional reporter. The transcriptional activity of sufT increased in synchrony with increasing inducer concentrations (Fig 8D).
Mycobacterium tuberculosis contains a DUF59 containing protein (Rv1466) that is part of the suf operon and is essential for viability (Fig 1C and [66]). We examined whether Rv1466 could compensate for the loss of SufT in S. aureus. Rv1466 has a ~20 amino acid N-terminal extension when compared to the S. aureus SufT. Codon-optimized rv1466 and a truncated version of rv1466 (trunc_rv1466) were introduced upon a multi-copy plasmid into the S. aureus ΔsufT strain and phenotypes were examined. The presence of trunc_rv1466, but not rv1466, rescued the growth defect of the ΔsufT strain in 18AA glutamate medium (Fig 9A). The presence of trunc_rv1466, but not rv1466, displayed a dominant effect and inhibited growth of the ΔsufT strain in 20AA glucose medium (Fig 9B).
Epistatic relationships between sufT, nfu, and sufA were investigated by phenotypically examining mutant strains lacking one, two, or all three maturation factors. The ΔsufA strain did not display a defect in AcnA activity, relative to the WT strain, and the ΔsufA ΔsufT double mutant phenocopied the ΔsufT strain (Fig 10A). The phenotypic effects of the ΔsufA and Δnfu mutations displayed an additive effect. AcnA activity in the Δnfu mutant was ~65% of WT while the activity in the ΔsufA ΔsufT double mutant was ~50%. AcnA activity was near the limit of detection in the Δnfu ΔsufT double mutant (~2%). The Δnfu ΔsufT ΔsufA triple mutant had AcnA activity similar to the Δnfu ΔsufT strain. AcnA activity in the acnA::TN Δnfu ΔsufT strain containing pacnA was also nearly undetectable relative to its isogenic parental strains (Fig 10B). This suggested that the low AcnA activity in the Δnfu ΔsufT strain was not solely the outcome of decreased acnA transcription.
Growth was examined in media that impose varying demands for FeS proteins (20AA glucose, 20AA glutamate, or 18AA glutamate media). The ΔsufA strain did not display a growth deficiency in any of the media examined (Fig 10C–10E). The ΔsufA ΔsufT double mutant phenocopied the ΔsufT strain in 20AA glucose and 20AA glutamate medium, but the effects of the mutations were additive in 18AA glutamate medium. The Δnfu ΔsufA double mutant phenocopied the Δnfu strain in 20AA glucose and 20AA glutamate media, but the effect of the mutations were additive in 18AA glutamate medium. The phenotypes of the Δnfu and ΔsufT mutations displayed synergism. The Δnfu ΔsufT double mutant displayed a severe growth defect in each media examined. The Δnfu ΔsufT ΔsufA triple mutant strain largely phenocopied the Δnfu ΔsufT strain in each media.
The Δnfu ΔsufT double mutant also displayed severe growth defects in complex medium. Growth of S. aureus in tryptic soy broth (TSB) results in the consumption of glucose, the release of fermentative byproducts such as acetate, and acidification of the medium [67,68] followed by the uptake of the fermentative byproducts resulting in alkalization of the growth medium. Therefore, the pH and acetate profile of the spent medium correlates with the cells ability to uptake and utilize fermentation products [67,68,69]. We monitored optical densities, pH of the spent medium, and acetate concentrations in the spent medium over time in cultures of the WT, ΔacnA, Δnfu, ΔsufT, and Δnfu ΔsufT strains. The Δnfu ΔsufT double mutant and ΔacnA strains displayed pronounced differences during post-exponential growth reaching lower final optical densities (S6A Fig). The pH of the medium from the Δnfu ΔsufT and ΔacnA mutants did not re-alkalinize (S6B Fig) nor was acetate utilized (S6C Fig).
The interactions amongst sufT, nfu, and sufA were further examined by introducing each gene upon a multi-copy plasmid (psufT, pnfu and psufA, respectively) and assessing whether they impart phenotypic suppression to the ΔsufT or Δnfu strains. The ΔsufA strain did not have decreased AcnA activity, and therefore, suppression was not examined in this strain.
The presence of psufA appeared to increase AcnA activity mildly in both the WT and ΔsufT strains, but a statistically significant phenotypic rescue was not observed (S7A Fig). AcnA activity decreased in the Δnfu strain carrying psufA. AcnA activity was increased in the ΔsufT strain carrying pnfu (increase of ~250%), while the presence of pnfu had little effect on AcnA activity in the WT (Fig 11A). The presence of psufT slightly decreased AcnA activity in the WT, while it did not alter AcnA activity in the Δnfu strain (S7B Fig).
Growth profiles of the WT and ΔsufT strains carrying empty vector or pnfu were examined in 20AA glutamate medium. The presence of pnfu partially mitigated the growth defect of the ΔsufT strain in 20AA glutamate medium (S8 Fig).
The phenotypes of the ΔsufT strain were mitigated during fermentative growth, which imposes a low demand for FeS clusters. We reasoned that Nfu is utilized to fulfill the demand for FeS cluster assembly in the ΔsufT strain during fermentative growth. After fermentative culture the acnA::TN Δnfu ΔsufT strain containing pacnA displayed levels of AcnA activity that were near the limit of detection (~2%), whereas the acnA::TN ΔsufT and acnA::TN Δnfu strains had AcnA activity similar to the parent (Fig 11B). Microaerobic growth in 18 AA glucose medium was also examined. The Δnfu and ΔsufT strains displayed growth profiles that did not significantly deviate from that of the WT (Fig 11C). However, the Δnfu ΔsufT double mutant displayed a large growth defect.
From Figs 10 and 11, S7 and S8 Figs, we concluded that 1) the phenotypic effects of the nfu and sufT mutations are synergistic, 2) overproduction of nfu partially alleviates the phenotypes of the ΔsufT strain, and 3) either Nfu or SufT is sufficient for AcnA maturation during fermentative growth.
Biofilm formation and exoprotein production were assessed in strains lacking FeS cluster assembly factors. Agr is the dominant activator for transcription of exoproteins and toxins, as well as the phenol soluble modulins (PSMs). Therefore, an Δagr strain was included as a positive control [51]. A strain lacking AcnA has been proposed to have increased Agr activity [70]. Since a Δnfu ΔsufT strain phenocopied the acnA::TN mutant, the acnA::TN strain was also examined. Exoproteins were extracted from the spent medium supernatant and analyzed using SDS-PAGE. S. aureus encodes for eight PSMs that are small peptides comprising ~60% of the total exoproteome and are visualized on SDS-PAGE as one band [51]. The Δnfu ΔsufT, Δnfu ΔsufT ΔsufA, and the Δagr strains were deficient in exoprotein production (Fig 12A). For ease of comparative analyses, only the band corresponding to PSMs is displayed.
Static growth of WT in TSB does not induce biofilm formation, and therefore, biofilm formation was examined in biofilm inducing medium (Fig 12B and 12C, [71]). Biofilm formation was also assessed in strains lacking Agr and SigB, which negatively and positively influence biofilm formation, respectively [72,73]. Strains deficient in the maturation of FeS proteins displayed varying degrees of biofilm formation. The Δnfu ΔsufT double mutant displayed the largest increase in biofilm formation (~4.5 fold). The acnA::TN strain formed biofilms at a similar extent as the WT (Fig 12B and 12C).
We examined vancomycin sensitivities of strains lacking FeS cluster assembly factors. The Δnfu ΔsufT double mutant displayed a large increase in resistance towards vancomycin during growth (Fig 13A). In growth inhibition curves we found that the Δnfu ΔsufT strain was not completely resistant towards vancomycin, but rather, it displayed an inhibition response more characteristic of vancomycin-intermediate resistant Staphylococcus aureus (VISA) (S9 Fig and [74]).
Vancomycin resistant strains display alterations in their cell walls resulting in increased resistance towards lysis by lysostaphin [55,74]. The Δnfu ΔsufT double mutant displayed the greatest resistance towards lysis by lysostaphin (Fig 13B).
Decreased activity of peptidoglycan hydrolases is a hallmark of VISA strains [55,74]. Peptidoglycan hydrolase activity was monitored using zymographic analysis upon heat-killed WT cells as a substrate. The Δnfu ΔsufT double mutant displayed the largest alterations in the activities of peptidoglycan hydrolases (Fig 13C).
Staphylococcus aureus SufT is composed solely of a DUF59 domain. Alternate proteins containing DUF59 domains participate in FeS cluster assembly, but the function(s) of the DUF59 domain itself has not been described [46,47,48]. The goals of this study were to determine if SufT has a role in FeS cluster assembly, and if so, begin to dissect its in vivo functional role.
Phylogenetic analyses found that sufT was recruited to the same chromosomal location as sufBC, and once recruited, it was largely retained. These findings suggested that sufT was recruited to the operon to refine the functionality of Suf-mediated FeS cluster assembly. Amongst the genomes analyzed, only five organisms encoded for SufT, but not the FeS cluster scaffolding proteins SufB, IscU, or NifU. The five organisms identified were lactobacilli and within these genomes the SufT homolog was located within apparent operons that encode known FeS cluster requiring proteins. The informatics and phylogenetic findings strongly suggested a role for SufT in FeS cluster assembly.
The S. aureus ΔsufT strain displayed physiological abnormalities consistent with SufT having a role in the maturation of FeS proteins. Further, the phenotypes of the S. aureus ΔsufT strain closely resembled those of a strain lacking the FeS cluster carrier Nfu [4]. Aside from a role in de novo FeS cluster assembly, alternate possibilities for the observed deficiencies manifest in the ΔsufT strain were considered. The ΔsufT strain did not have altered H2O2 or superoxide scavenging activities. SufT was not required for the physical exclusion of H2O2 from the AcnA active site or the repair of the H2O2 damaged FeS cluster upon AcnA. These findings suggested that SufT likely functions in the de novo assembly of FeS clusters upon apo-proteins.
Genes encoding for proteins with functional overlap often display synergistic (superadditive) phenotypic effects when the gene products are absent or non-functional [75]. The phenotypes associated with nfu and sufT were synergistic. This was most evident during fermentative growth where there is a lower demand for FeS clusters. The phenotypes of the Δnfu and ΔsufT strains were nearly indistinguishable from the WT strain, but the Δnfu ΔsufT double mutant displayed a large growth defect and exhibited AcnA activity near the limit of detection. Introduction of nfu in multicopy to the ΔsufT strain led to partial mitigation of the phenotypes of this strain. Taken together, these findings led to the conclusion that both SufT and Nfu function as non-essential, accessory factors in the maturation of FeS proteins. Lending further support to this conclusion, subsequent to our informatics analyses, the genome of Oligotropha carboxidovorans was sequenced and found to encode for a protein consisting of a fusion of the N-terminus of Nfu and SufT (Locus tag: OCA5_c02770).
SufT, Nfu, and SufA are auxiliary FeS cluster maturation factors leading to the question of why S. aureus encodes for three such factors. The simplest explanations are that there is a degree of specificity for each auxiliary factor with respect to their target apo-proteins or that they have different functions. Vinella et al. have recently proposed an expanded model, which visualizes a dynamic cellular network of proteins that varies with growth stage or growth condition allowing for rapid calibration to alterations in the cellular demand for FeS protein maturation [76]. During such a scenario, certain auxiliary proteins and pathways would be preferred during normal growth and alternate auxiliary proteins and pathways during stress conditions.
The findings presented herein are consistent with the model proposed by Vinella et al. [76]. During routine aerobic growth, Nfu was the dominant auxiliary factor required for the maturation of AcnA. However, upon the overproduction of AcnA, the need for SufT for AcnA maturation was increased. The cellular need for SufT was also increased when cells were resuming respiration, toxified with ROS, or grown in 18AA glutamate medium; three conditions that impose a high demand for de novo FeS cluster assembly. The transcriptional activity of sufT also increased as the cellular demand for FeS clusters increased. These findings lend strong support to a model wherein SufT is a dominant factor involved in the maturation of FeS proteins in cells experiencing a high demand for FeS clusters. The epistasis experiments further strengthen the idea that certain accessory proteins are preferentially utilized when confronted with a high demand for FeS clusters. SufA was dispensable for growth under all conditions tested. However, SufA dependent phenotypes were manifest in strains lacking either Nfu or SufT and simultaneously cultured upon a medium imposing a high demand for FeS proteins. Therefore, we propose that SufA facilitates FeS protein maturation in S. aureus under conditions imposing a very high demand for FeS clusters. It is tempting to speculate that cells encode for multiple accessory maturation factors to respond to a gradation of demand for FeS cluster assembly, however, this awaits further experimentation.
It is currently unclear what genetic or biochemical elements dictate the increased usage of SufT or SufA upon increased FeS cofactor demand. Possible explanations include different functionalities, increased stability of a particular factor under stress conditions, or an increased rate of FeS cluster synthesis or FeS protein maturation under select cellular conditions. A similar scenario has been described to exist between the Suf and Isc FeS cluster biosynthetic machineries. In Escherichia coli, Suf is preferred under ROS stress and Fe limiting conditions, whereas Isc is the preferred FeS assembly system during conditions imposed by routine laboratory cultivation [77,78].
What is the role of SufT in FeS cluster assembly? The genetic findings presented make it tempting to speculate that SufT functions in the carriage of FeS clusters, but further biochemical analyses will be necessary to make this conclusion. It also worth noting that the SufT homologues analyzed in Fig 1 contain only one strictly conserved cysteine residue. With the exception of monothiol glutaredoxins, described FeS cluster carriers contain two or more cysteines utilized in FeS cluster ligation [79].
Biofilm formation and exoprotein production are critical in the infectious lifecycle of S. aureus [49,50]. We previously found that a strain lacking Nfu is attenuated for virulence in models of infection [4]. In this report we found that a strain that was crippled in its ability to maturate FeS proteins displayed significantly increased biofilm formation and decreased exoprotein production. Vancomycin is a last resort drug in the treatment of CA-MRSA infections and the genetic and molecular mechanisms underlying resistance to vancomycin are an active area of research [54]. Strains defective in FeS protein maturation also displayed an intermediate resistance to vancomycin and multiple phenotypes associated with VISA strains.
The Δnfu ΔsufT strain phenocopied a ΔacnA strain in growth experiments, but it did not phenocopy this strain in phenotypes involved in virulence. S. aureus encodes for the FeS cluster utilizing two-component regulatory system (TCRS) AirSR [5]. AirSR alters the transcription of genes encoding for peptidoglycan hydrolases, as well as those required for biofilm formation [5,80]. AirR directly binds to the promoter region of Agr [80]. AirSR is also implicated in vancomycin resistance and a strain lacking AirSR displays VISA like phenotypes [80]. Therefore, the accumulation of apo-AirSR in the Δnfu ΔsufT strain may underlie the virulence phenotypes witnessed. An alternate explanation is that the altered Agr activity in these strains results in altered virulence phenotypes. Apart from its roles in toxin production and biofilm formation, Agr has also been implicated in modulating vancomycin resistance in S. aureus [51,81,82]. Regardless of the mechanism(s) underlying the phenotypes presented, these findings highlight the importance of efficient FeS cluster assembly for multiple phenotypes critical for pathogenesis and antibiotic resistance.
In summary, we have identified a role for SufT, and by extension DUF59, in the maturation of FeS proteins. We propose a model wherein SufT is an auxiliary FeS protein maturation factor whose usage is selectively increased during growth conditions necessitating increased FeS cluster assembly in S. aureus. An increased demand for FeS clusters may have been an evolutionary driving force to recruit sufT to the suf operon thereby increasing the efficiency and control of de novo FeS cluster assembly.
Restriction enzymes, quick DNA ligase kit, deoxynucleoside triphosphates, and Phusion DNA polymerase were purchased from New England Biolabs (Ipswich, MA). The plasmid mini-prep kit, gel extraction kit and RNA protect were purchased from Qiagen (Hilden, Germany). Lysostaphin was purchased from Ambi products (Lawrence, NY). Oligonucleotides were purchased from Integrated DNA Technologies (Coralville, IA) and sequences are listed in S1 Table (oligonucleotides used in this study). Trizol (Life Technologies), High-Capacity cDNA Reverse Transcription Kits (Life Technologies), and DNase I (Ambion) was purchased from Thermo Fisher Scientific (Waltham, MA). Tryptic Soy Broth (TSB) was purchased from MP Biomedicals (Santa Ana, CA). An acetic acid quantification kit was purchased from R-BioPharma (Darmstadt, Germany). Unless specified all chemicals were purchased from Sigma-Aldrich (St. Louis, MO) and were of the highest purity available.
Unless otherwise stated, the S. aureus strains used in this study (listed in Table 1) were constructed in the S. aureus community-associated USA300 strain LAC that was cured of the native plasmid pUSA03, which confers erythromycin resistance [83]. The USA300 LAC genome differs from USA300_FPR3757 only by a few single nucleotide polymorphisms [84,85]. Unless specifically mentioned, S. aureus cells were cultured as follows: 1) aerobic growth at a flask/tube headspace to culture medium volume ratio (hereafter HV ratio) of 10; 2) anaerobic growth at a flask/tube headspace to culture medium volume ratio of 0, as described earlier [4]; 3) in 96-well microtiter plates containing 200 μL total volume (detailed procedure below). Liquid cultures were grown at 37°C with shaking at 200 rpm unless otherwise indicated. Difco BioTek agar was added (15 g L-1) for solid medium. When selecting for plasmids, antibiotics were added at the final following concentrations: 150 μg mL-1 ampicillin (Amp); 30 μg mL-1 chloramphenicol (Cm); 10 μg mL-1 erythromycin (Erm); 3 μg mL-1 tetracycline (Tet); 125 μg mL-1 kanamycin (Kan); 150 ng mL-1 anhydrotetracycline (Atet). For routine plasmid maintenance, liquid media were supplemented with 10 μg mL-1 or 3.3 μg mL-1 of chloramphenicol or erythromycin, respectively.
Escherichia coli DH5α was used as a cloning host for plasmid constructions. All clones were passaged through RN4220 and transductions were conducted using phage 80α [86]. All S. aureus mutant strains and plasmids were verified using PCR or by sequencing PCR products or plasmids. All DNA sequencing was performed by Genewiz (South Plainfield, NJ).
Unless otherwise stated, JMB1100 chromosomal DNA was used as a template for PCR reactions. To create the ΔsufT deletion strain (JMB1146), approximately 500 base pairs upstream and downstream of sufT gene (SAUSA300_0875) were amplified using PCR with primer pairs 0875up5EcoRI and 0875up3NheI; 0875dwn5MluI and 0875 dwn3BamHI (S1 Table). Amplicons were gel purified and fused using PCR and the 0875up5EcoRI and 0875 dwn3BamHI primers. The resulting amplicon was gel purified, and digested with BamHI and SalI, followed by a ligation into similarly digested pJB38 resulting in pJB38_ΔsufT. The plasmid pJB38_ΔsufT was isolated and subsequently transformed into RN4220 before transducing into JMB1100. A single colony was inoculated into 5 mL of TSB-Cm and cultured overnight at 42°C followed by plating 25 μL on TSA-Cm to select for colonies containing a single recombination event. Single colonies were inoculated into 5 mL of TSB medium and were grown overnight, followed by a dilution of 1:25,000 before plating 100 μL onto TSA containing Atet to select against plasmid containing cells. Colonies were screened for Cm sensitivity and for the ΔsufT mutation using PCR.
The sufT::tetM strain was created by digesting the pJB38_ sufTΔ with MluI and NheI and inserting the tetM gene between the upstream and downstream regions of sufT. The DNA encoding for Tet resistance (tetM) was amplified using PCR with Strain JMB1432 as a template and the G+tetnheI and G+tetmluI primers before digesting and ligating into similarly digested pJB38_ΔsufT. The resulting plasmid (pJB38_ΔsufT::tetM) was passaged though E. coli, before it was transformed into RN4220. The ΔsufT::tetM mutant was constructed as described above.
Plasmids for genetic complementation, transcriptional analyses, and insertion of epitope tags to allow protein detection by western blots were constructed by subcloning digested PCR products into similarly digested vectors or by using yeast homologous recombination cloning (YRC) as previously described [87,88]. The pLL39_sufT and pCM28_sufT plasmids were created using the 0875_5BamHI and the 0875_3SalI primer pair. The pCM11_sufT was created using the 875gfpKpnI and 875gfpHindIII primer pair. The pCM11_acnA was made using the AcnApHindIII and AcnApKpnI primer pair. The Mycobacterium tuberculosis rv1466 was codon optimized and synthesized by Integrated DNA technologies (IDT; Coralville, IA) and cloned into pCM28 using the native S. aureus sufT promoter using YRC. The full-length construct was constructed using amplicons generated using the following primer pairs: pCM28YCC and Ycc875p3; ycc875p5 and 875pMT3; 875pMT5 and 875pCM28 3. The truncated version was created using the same primers except MT875trunk5 and MT875trunk3 replaced ycc875p5 and Ycc875p3, respectively.
Growth was assessed in 200 μL cultures grown at 37°C in 96-well plates using a BioTek 808E Visible absorption spectrophotometer. Culture optical density was monitored at 630 nm. The staphylococcal-defined medium has been described previously [4]. Strains cultured overnight in TSB were inoculated into minimal medium or TSB to a final optical density (OD) of 0.025 (A600) units. For assessing nutritional requirements, cultures were harvested and treated as above, except that the cell pellet was washed twice to prevent carryover of rich medium components. For aerobic growth the shake speed was set to medium. For microaerobic growth the plate was incubated statically.
The four growth medium formulations utilized for nutritional analyses were: 1) 20AA glucose medium, containing the 20 canonical amino acids and 14 mM glucose as a source of carbon; 2) 18AA glucose medium, containing 18 canonical amino acids and lacking leucine and isoleucine and 14 mM glucose as a source of carbon; 3) 20AA glutamate medium, containing the 20 canonical amino acids and 44 mM glutamate as a source of carbon, and 4) 18AA glutamate medium, containing 18 canonical amino acids and lacking leucine and isoleucine and 44 mM glutamate as a source of carbon.
To examine vancomycin sensitivity, cultures were inoculated into TSB in the presence or absence of varying concentrations of vancomyin (0.025–1.5 μg/mL). Growth inhibition was assessed after 4 hours of growth. Paraquat sensitivity assays were conducted upon solid tryptic soy broth agar (TSA) plates containing 0 or 30 mM of paraquat. Overnight cultures (~18 hours of growth) were serial diluted in 1X phosphate buffered saline and 10 μL of each dilution was placed on plates of the solid medium. The plates were incubated at 37°C for 15 hours before the growth was assessed.
Strains cultured overnight in TSB-Erm medium were diluted into fresh TSB-Erm medium to a final OD of 0.1 (A600) and cultured, with shaking, at a HV ratio of 10. At periodic intervals culture density and fluorescence were assessed as described previously [4]. Fluorescence data were normalized with respect to a strain not carrying a GFP-based transcriptional reporter to normalize for background fluorescence values. The resulting data were normalized to the culture OD. Finally for ease of comparative analyses the data were normalized relative to the wild-type (WT) strain, or as specified in the figure legend.
Anaerobic culture conditions were achieved as described earlier [4,89]. Cells were cultured to exponential growth, aerobically, as described above. The cultures were then split and one set of cells was cultured at a HV ratio of zero in capped microcentrifuge tubes and anaerobiosis was verified by the addition of 0.001% resazurin to control tubes [4,89].
mRNA abundances of genes were examined from a previously described cDNA library [4].
Strains were cultured overnight in TSB and cells were harvested by centrifugation. Cell pellets were washed twice with 1X phosphate buffered saline and resuspended in lysis buffer (recipe above) in the presence of 5 μg/mL of lysostaphin. The lysostaphin mediated decrease in optical densities (A600) was recorded periodically.
Protein concentration was determined using a copper/bicinchonic acid based colorimetric assay modified for a 96-well plate (47). Bovine serum albumin (2 mg/mL) was used as a standard. Western blot analyses were conducted as described previously [4,88].
Strains cultured overnight in TSB (~18 hours) were diluted into fresh TSB to a final OD of 0.1 (A600). Periodically, aliquots of the cultures were removed, optical density was determined, and the cells and culture media were partitioned by centrifugation at 14,000 rpm for 1 minute. Two mL of either the culture supernatant or sterile TSB, which served to provide a pH reading for the point of inoculation, were combined with 8 mL of distilled and deionized water and the pH was determined using a Fisher Scientific Accumet AB15 pH mV Meter. The concentration of acetic acid in spent media was determined using the R-Biopharm Enzymatic BioAnalysis kit following the manufacturer's suggested protocol.
Biofilm formation was examined as described elsewhere, with minor changes [71,97]. Briefly, overnight cultures were diluted into biofilm media (TSB supplemented with 3% NaCl and 0.5% glucose), added to the wells of a 96-well microtiter plate and incubated statically at 37°C for 22 hours. Prior to harvesting the biofilms, the optical density (A590) of the cultures was determined. The plate was subsequently washed with water, biofilms were heat fixed at 60°C, and the plates and contents were allowed to cool to room temperature. The biofilms were stained with 0.1% crystal violet, washed with water, destained with 33% acetic acid and the absorbance of the resulting solution was recorded at 570 nm and standardized to an acetic acid blank and subsequently to the optical density of the cells upon harvest. Finally the data were normalized with respect to the WT strain to obtain relative biofilm formation.
Spent medium supernatants were obtained from overnight cultures, filter sterilized with a 0.22 μm (pore-size) syringe filters, and standardized to culture optical densities (A600). Zymographic analyses of bacteriolytic proteins were conducted using standard methods described elsewhere [98] and samples were separated upon a 12% SDS gel incorporated with 0.3% (vol/vol) heat killed USA300_LAC cells [98]. To determine exoprotein profiles, the spent media supernatant was concentrated using standard trichloroacetic acid precipitation. The resultant protein pellets were resuspended in laemelli buffer and equal volumes were separated upon a 12% SDS gel.
The taxonomic distribution of Suf was determined via BLASTp analyses of publically available genome sequences in October of 2011 as part of a previous study [25]. This distribution of Suf was characterized using the KEGG gene viewer [99], with manual verification using BLASTp or using sequence alignments. 1094 genomes out of a total of 1667 genome sequences (65.6% of total) encoded for SufBC. Genomes that encoded for SufBC were then screened for the presence of SufT using BLASTp. sufT was considered to be associated with the suf operon if they were within four open reading frames from sufBC and appeared to be transcribed from a common promoter.
SufT sequences were compiled and aligned with ClustalW specifying default parameters [100]. The aligned sequences were manually truncated to the minimal SufT sequence or positions 1 to 99 of SufT from Thermoplasma acidophilum (Kegg ID: Ta0200). Phyml was used to reconstruct the evolutionary history of the SufT alignment block specifying the Blosum62 substitution model and gamma distributed rate variation [101]. The topology of the tree was evaluated using Chi2-based likelihood ratio tests. The phylogenetic reconstruction was projected with the Interactive Tree Of Life (Itol) web program [102].
The N- and C-terminal sequences that were pruned from the alignment block were subjected to BLASTp against the Conserved Domain Database (CDD) using an evalue of 0.01 [103]. Identified motifs in both N- and C- terminal motifs were compartmentalized into modular structures based on the presence of unique sequence motifs. These N- and C-terminal motifs were mapped onto the SufT phylogenetic tree using the Itol program. Furthermore, SufT was mapped onto a concatenated SufBC phylogenetic tree using the Itol program. The concatenated SufBC tree was constructed as previously described [25].
|
10.1371/journal.pcbi.1000619 | Bistability of Mitochondrial Respiration Underlies Paradoxical Reactive Oxygen Species Generation Induced by Anoxia | Increased production of reactive oxygen species (ROS) in mitochondria underlies major systemic diseases, and this clinical problem stimulates a great scientific interest in the mechanism of ROS generation. However, the mechanism of hypoxia-induced change in ROS production is not fully understood. To mathematically analyze this mechanism in details, taking into consideration all the possible redox states formed in the process of electron transport, even for respiratory complex III, a system of hundreds of differential equations must be constructed. Aimed to facilitate such tasks, we developed a new methodology of modeling, which resides in the automated construction of large sets of differential equations. The detailed modeling of electron transport in mitochondria allowed for the identification of two steady state modes of operation (bistability) of respiratory complex III at the same microenvironmental conditions. Various perturbations could induce the transition of respiratory chain from one steady state to another. While normally complex III is in a low ROS producing mode, temporal anoxia could switch it to a high ROS producing state, which persists after the return to normal oxygen supply. This prediction, which we qualitatively validated experimentally, explains the mechanism of anoxia-induced cell damage. Recognition of bistability of complex III operation may enable novel therapeutic strategies for oxidative stress and our method of modeling could be widely used in systems biology studies.
| The levels of reactive oxygen species (ROS) that are generated as a side product of mitochondrial respiratory electron transport largely define the extent of oxidative stress in living cells. Free radicals formed in electron transport, such as ubisemiquinone, could pass their non-paired electron directly to oxygen, thus producing superoxide radical that gives rise to a variety of ROS. It is well known in clinical practice that upon recommencing oxygen supply after anoxia a tissue produces much more ROS than before the anoxia, and the state of high ROS production is stable. The mechanism of switching from low to high ROS production by temporal anoxia was unknown, in part because of the lack of detailed mathematical description of hundreds of redox states of respiratory complexes, which are formed in the process of electron transport. A new methodology of automated construction of large systems of differential equations allowed us to describe the system in detail and predicts that the mechanism of paradoxical effect of anoxia-reoxygenation could be defined by the properties of complex III of mitochondrial respiratory chain. Our experiments confirmed that the effect of hypoxia-reoxygenation is confined by intramitochondrial processes since it is observed in isolated mitochondria.
| The pathologic consequences of anoxia–reoxygenation, including the oxidative stress associated with increased production of reactive oxygen species (ROS) in mitochondria, form the basis of major diseases, including heart disease, age-related degenerative conditions and ischemic syndrome in reperfusion [1]. The use of novel antioxidants, which addresses the consequences of oxidative stress, has proven to be effective in organ preservation [2], but there is no doubt that a better understanding of the causes of elevated ROS production during the anoxia/reoxygenation would help to introduce novel strategies addressing the primary events of this clinical phenomenon.
Paradoxically, ROS production increases during severe hypoxia despite of decrease of oxygen concentration; this ROS increase acts as a metabolic signal for cell adaptation to oxygen deficiency [3]. Moreover, when cells that have been exposed to anoxia are returned to their normal oxygen supply, the rate of ROS production, instead of returning to the low baseline, greatly increases, often leading to cell death. This phenomenon of reperfusion injury after ischemia has been well known [4],[5], but its mechanism remains unclear and it is analyzed here.
Mitochondrial respiratory electron transport, which is schematically shown in Figure S1, is generally accepted as a process related with ROS production in living cells. In particular, o-site of quinol oxidation in complex III (Qo) is one of the most frequently considered sites of superoxid anion generation [3],[6],[7], and the considered here mechanism for the stimulation of ROS production by anoxia is related to this site. The proposed mechanism is based on our analysis of the well-known Q-cycle (ubiquinone (Q) oxidation/reduction) mechanism of electron and proton transport performed by complex III in the mitochondrial respiratory chain [8]–[13], as schematically shown in Figure 1. The main components of complex III are cytochrome b, containing two hemes characterized by low (bL) and high (bH) midpoint potentials, Rieske protein containing an iron-sulfur redox center (FeS), and cytochrome c1. The FeS center of the Rieske protein accepts one electron from a bound ubiquinol (QH2) producing a highly reactive anion radical of ubiquinone (Q−) known also as semiquinone radical (further referred as SQ) and releasing two H+ to the cytosolic side. The electron accepted by the FeS center then is delivered to c1 and passes further downstream in the electron transport. The semiquinone radical is normally transformed into ubiquinone as it delivers its unpaired electron to cytochrome bL, which then passes it to bH. However, there is a probability that the semiquinone radical delivers its unpaired electron directly to oxygen, producing a superoxide radical [14]–[16]. Semiquinone oxidation by any of these two mechanisms transforms it into ubiquinone, which then dissociates from complex III, binds at the matrix side and receives two electrons from cytochrome bh that are derived from oxidation of two QH2 molecules as described above. This process subsequently generates semiquinone and QH2 again, taking protons from the matrix. The dissociation of the newly produced QH2 and its subsequent binding at the cytosolic side starts the cycle again.
A great deal of problems in studying ROS production in respiratory chain induced by anoxia arises because of the lack of tools for systematic theoretical analysis of this complicated system. Various combinations of oxidized and reduced states of the electron transporters create 400 possible different redox states for complex III, which are formed in the process of electron transport and which should be taken into account in an analysis of electron transport and the related ROS production. The high levels of complexity has, until now, precluded understanding of complex III functionality. To overcome this problem, we have developed a special software tool that takes into account all possible redox states of complex III. The software automatically constructs a system of differential equations describing their evolution based on several rules defined by the types of reactions between the electron transporters, as outlined in Methods and in more detail in Text S1. Our current approach to simulating mitochondrial respiration has inherited the main principles of automated construction of large equation systems used for stable isotope tracer data analysis [17]–[19].
Using this tool for the simulations of evolution of all redox states of complex III, we found that complex III has a property of bistability: it can persist in two different steady states at the same set of parameters; evolution to one or another steady state is defined only by the initial state of the system. One such state is characterized by a high ROS production rate. The system can be switched to this latter state either by an increase in succinate supply or by a decrease of oxygen availability, and can persist in it after a return to the initial conditions. This behavior explains the mechanics of paradoxical increase in ROS generation induced by anoxia and its further increase after return to a normal oxygen supply. Experimental data presented here qualitatively confirmed such a bistable behavior.
The model constructed, as described in Methods, is based on the generally accepted Q-cycle mechanism of complex III operation and predicts that it can evolve to a one of two different steady states. The direction of complex III evolution depends on the initial state of the system. If initially it is in highly reduced state, characterized by high levels of ubiquinol and semiquinone, it remains in this state, as shows thick gray lines in Figures 2A and 2B. However, if the initial levels of reduction is lower, the system evolves to another steady state, the same for a variety of initial states, as Figures 2A and 2B show. Thus, the Q-cycle mechanism defines bistability of complex III operation: with the same parameters it could function in a mode characterized by either low or high SQ levels (low or high ROS producing states respectively).
Various factors could trigger complex III from one mode to another, specifically, the availability of succinate could be a triggering factor. Figure 3A shows the time course of the transition from low to high ROS producing steady state, induced by an increase in succinate concentration (eq (17)). The capacity for ROS production is reflected in the concentration of free radicals, which can directly interact with oxygen, such as the semiquinone radical (SQ) bound to the Qo site of complex III [14],[20]. An increase in succinate supply, and respective reduction of free ubiquinone (Q) to ubiquinol (QH2), switches the system from low ROS producing state, characterized by low levels of SQ, to the high ROS producing state, characterized by high levels of SQ. In accordance with mass conservation, reduction of Q to QH2 results in the deficiency of Q; this decreases electron flow and promotes reduction of cytochrome b. The QH2 bound at Qo site can freely deliver its first electron to the cytochrome c1, but cannot deliver the second electron to cytochrome bL because latter is already reduced. As a result, highly active SQ radicals are accumulated.
The continuum of steady state SQ levels and electron flow dependent on succinate supply is shown in Figures 3B and 3C. If the system is initially in a low ROS producing state, it switches to the high ROS production when the succinate concentration reaches a certain threshold value as Figure 3B shows. However, if the succinate concentration decreases back when the system initially is in high ROS producing state, it does not follow the same pattern, as Figure 3C shows. This example demonstrates hysteresis in complex III behavior.
The value of transmembrane potential is essential for such hysteretic behavior. As Figure S2 shows, the region of bistability is clearly distinguishable at transmembrane potential of ∼200 mV, which corresponds to the state 4 of mitochondrial respiration. The fall down of transmembrane potential to below 150 mV (as could be when the addition of ADP switches mitochondria to the state 3 of respiration), switches complex III from high to low ROS producing state. At high transmembrane potential the region of bistability persists over a large variation of model parameters as the sensitivity analysis shown in Text S1 and Figures S3 and S4 indicates. A switch from one steady state to another one essentially redistributes the fractions of various redox states of the complex as Figure S5 shows.
The electron flux in high ROS producing steady state, being restricted by the deficiency of Q, is low; maintaining it requires low succinate supply and this steady state persists even if substrate supply decreases until it falls down below the minimal threshold, as displayed in Figure 3C. The reduction of molecular oxygen by semiquinone radicals transforms them into Q, producing an acceptor able to take electrons from bh, and thus to activate the Q-cycle. Thus, high ROS production rate, in a way, is a means to return back to low ROS production. Even a low rate of such an electron leak to oxygen helps the system to revert to low ROS production (see sensitivity analysis presented in Text S1). Normally direct transfer to oxygen insignificantly contributes to the total electron flow, but in the case of extremely high complex II activity, if it reduces all available Q, the electron flow to complex IV would equalize the ROS production rate.
The predicted phenomenon of bistability explored in Figure 3 could be observed experimentally, as Figure 4A shows. The isolated rat brain mitochondria incubated with succinate (state 4 of respiration) are in a high ROS producing state (blue trace “ros”), however, temporal presence of ADP, which is rapidly transformed into ATP, switches them to a low ROS producing mode (red trace “ros,ADP”). When ADP is present, ATP synthesis lowers membrane potential (shown in reverse direction as measured by quenching of fluorescence, red trace “mp,ADP”). However, after complete transformation of the added ADP into ATP, mitochondria again come back to state 4 of respiration, the membrane potential increases to the same levels as that without ADP (blue trace “mp”), but ROS production remains to be much lower (red trace “ros,ADP”). ADP only switched respiratory chain from high to low ROS production state, which could be maintained under the same conditions of incubation. The addition of ATP to the medium had no effect on ROS production (blue trace “ros”). In this case ROS production was the same as in the absence of ATP (Figure S6).
Figure 4B shows two branches of steady state ROS production rates measured before addition of ADP (stars) and after its consumption (squares) (as Figure 4A explains). Solid lines generated by computer simulation show that our model has reproduced these experimental data after implementing the Michaelis' dependence of complex II activity on succinate concentration (eq (17)). Thus, isolated mitochondria show the predicted bistable behavior (Figure 3) and the model simulates the particular experimental data (Figure 4B). In these experiments 5 mM of pyruvate were present at variable concentrations of succinate to ensure that overall substrate supply is not a limiting factor. In parallel experiments it was shown that at different succinate concentrations membrane potential was invariable (not shown). It should be noted that the parameters used as characteristics of succinate oxidation (Vm, Km) in complex II, in fact characterize several processes including succinate transport. In our in vitro experiments the amount of succinate must be sufficiently high to maintain the rate of its transport needed for ubiquinone reduction. The intramitochondrial succinate concentration, which in vivo is produced in TCA cycle and which is sufficient to trigger high ROS production, could be much less than that necessary to be added externally. The actual parameters, which govern the oxidation of succinate produced in the matrix in vivo, could be different, and since phenomenologically the mechanism admits the forward and reverse switch between the steady states (Figure 3), it could be expected under some conditions, such as ischemia, which could induce several fold increase in intracellular succinate content and, as a consequence, increase in FADH2 [21]. Another factor, which substitutes succinate in triggering ROS production in vivo, is reduction of ubiquinone by complex I, whose substrate, NADH, is accumulated during anoxia [21].
The lack of oxygen itself can induce a switch to high ROS production if initially respiratory chain is in a low ROS production mode as Figure 4A shows. For this simulation the electron flux from complex III to complex IV was described as Michaelis' function of oxygen concentration (eq (18)). Parameters were chosen so that oxygen availability limits the flux when oxygen concentration falls to about 0.1% O2, (in accordance with [13],[22]). The decrease of oxygen availability at the excess of succinate results in a reduction of ubiquinone and cytochrome b, and, finally, an accumulation of SQ radicals. As Figure 5A shows, the accumulation of SQ radicals proceeds in conjunction with limiting the electron flow by the lack of oxygen. Apparently, the physiological response to mild hypoxia, which could be observed at 5% O2 [23], is related with another mechanisms.
If after passing through nearly anoxic conditions oxygen availability increases, the system remains in a high ROS producing state. Thus, the functional organization of electron transport in complex III, which has the property of bistability, explains the apparent paradox of increased ROS production induced by anoxia and re-oxygenation.
Figure 5B shows that anoxia induces a switch to a high ROS producing mode in isolated brain mitochondria, which were switched to a low ROS producing mode by temporal presence of ADP. Experiment was started with an open cuvette and normal oxygen supply by an addition of mitochondria to the incubation medium containing 1 mM ADP. Low value of membrane potential is indicative of mitochondria being initially in state 3. Upon conversion of all present ADP into ATP, mitochondria transited into state 4; an increased membrane potential is indicative of that transition. During the first stage of experiment mitochondria released low, but measurable amounts of ROS. An anoxic state was achieved by bubbling of the cuvette with N2 for 2–3 min and then closing it with a Teflon lid. Slow depolarization of mitochondria and their persistence in the depolarized state for the duration of closure indicated an achievement of anoxic state. Notably, the ROS signal stayed flat during that stage. At the third stage of the experiment, the resumption of oxygen supply was achieved by opening the cuvette and bubbling it with air. At this stage mitochondria regained membrane potential up to the initial level thus showing their metabolic integrity. Significantly higher than at initial normoxic stage amounts of ROS were released upon the resumption of oxygen supply.
Although the damaging effect of reoxygenation after anoxia is well known in general, presented here experimental data for the first time directly demonstrate anoxia-induced increase in ROS at the levels of isolated mitochondria.
While at constant oxygen concentration the levels of SQ at Qo site could be considered as a relative rate of ROS production, the change of oxygen concentration must also affect it. Figure 6 shows the prediction of ROS production rate when the oxygen concentration changes as Figure 5A shows. For the calculations of ROS production eq. (16) was used with various KROS. Figure 6 indicates that the increase in ROS production could be detectable despite the decrease of oxygen concentration, if KROS is in the range of Km for respiration. Increase of KROS results in the decreases ROS production before it starts to affect the respiration rate that induces increase of SQ. In this case the change of SQ would not be well detectable. Thus, if experimentally ROS production was found to increase under anoxic conditions [3], it means that KROS must be in the range of Km for respiration or lower.
The algorithms for automated construction of large systems of differential equations, which are presented in Methods, allowed us to simulate in detail the Q-cycle mechanism of complex III operation and to find that the property of bistability is inherited from this mechanism. Due to this property it can operate in a low or high ROS producing mode at the same values of parameters, which explains the difference in ROS generation by respiratory chain before and after an episode of anoxia. The presented experiments qualitatively confirm the existence of two different ROS producing steady states in isolated mitochondria at the same microenvironmental conditions and a possibility to switch between them by the temporal presence of ADP and/or episodes of anoxia.
The following biochemical events were found to underlie the transition from low to high ROS producing state: (i) ubiquinone pool is reduced to ubiquinol, (ii) lack of oxidized form (ubiquinone) restricts cytochrome bH oxidation and results in bH and bL reduction, which, in turn, (iii) restricts the oxidation of semiquinone radicals, thus increasing their concentration. If substrate concentration is then decreased below a threshold, the remaining electron flow is sufficient to maintain this high ROS producing state and therefore low and high ROS producing states could exist at the same set of parameters.
The increase of ROS production in tissues caused by ischemia-reperfusion is a well known clinical phenomenon; it is of great importance in various systemic diseases and organ transplantation. However, it was not known, which system is responsible for the switch from normally low ROS producing state to high ROS production, and how such a switch proceeds. We predicted and experimentally confirmed that complex III could respond to anoxia-reoxygenation in the same manner as it was observed at tissue levels. This indicates that the revealed mechanism could constitute the main part of tissue response to anoxia-reoxygenation. It does not exclude, though, a possible contribution of other parts of respiratory chain and/or other parts of metabolic and signaling pathways as modifiers of the analyzed phenomenon or even as other individual sources of ROS; rather, our findings open a new direction in investigation of the molecular mechanisms of bistable behavior that underlies the signaling effect of anoxia and oxidative stress caused by reoxygenation. In living cells ischemia combines several factors affecting ROS production, such as reduction of electron transport chain, anoxia, converting ATP into ADP and AMP, and deamination of the latter resulting in decrease of cellular adenilate pool [24]. Various combinations of these factors could create various types of behavior. Understanding their underlying mechanisms will improve the clinical care of diseases related to anoxia and also the quality of organ transplantation.
The proposed method for kinetic description of multienzyme complexes inherited from the established method of automated equation construction, which we developed for a different area of investigation, namely the analysis of stable isotopic isomers distribution in the intermediates of central carbohydrate metabolism [17]–[19]. Thus, it is not restricted by an area of mitochondrial respiration and could be further modified for applications in different areas, for instant for the analysis of metabolic signalling, which requires a description of multiple states of receptor molecules. Using this method will allow further expanding the model, including in it the other parts of electron transport chain and central metabolism, which provides substrates for mitochondrial respiration, ion transport that defines transmembrane potential, etc. Thus, the proposed method of differential equations construction opens a new direction for most detailed mathematical analysis of complex bioenergetic systems.
The model simulates the reactions performed by complex III, which overall expression is:The two electrons taken from QH2 then are transferred to cytochrome c and further to complex IV and oxygen in conjunction with translocation of four more protons. Although the presented model does not consider the details of this latter process, it accounts for it stoichiometry.
In the course of complex III reaction its core (consisted of cytochrome b with its two hemes, bH and bL, cytochrome c1, and Fe-S containing center of Rieske protein) bind and dissosiate ubiquinone/ubiquinol either in the matrix (Qi) or cytosolic (Qo) side of the inner mitochondrial membrane. This binding of quinones produce four different species of the complex:(1)(2)(3)(4)The electron transporters consisting the species (1)–(4) could be either in reduced or oxidized state. Various combinations of redox states of transporters produce many forms of each specie. The concentrations of such forms, which are transformed one into another during the complex operation, are the variables of the model. All the concentrations calculated for different redox forms of the same specie are stored in a specific array. To facilitate the references to variables inside the program, the two redox states of each transporter represented either by the digit “0” for oxidized state or by “1” for the reduced one. For instance, the binary “0011” (or decimal “3”) with respect to the array of core (1) refers to the concentration of a specie with oxidized bH and bL and reduced c1 and FeS. On the other hand, these digits together form a binary number, which signifies that the concentration of this redox form occupies third position (starting from 0) in this array. Thus, the program uses such integer number for two purposes: to designate a specific redox form and in the same time to refer to the place of its concentration in the array reserved for the concentrations of all possible redox forms.
The total number of redox forms for a specie is defined by the number of transporters constituting it, being 2n, where n is the number of transporters. In this way, core (1) has 16 different redox states varying from 0000 to 1111; species (2) and (3) have 64 states each, varying from 000000 to 111111, and, respectively, the specie (4) has 256 redox states. The program constructs differential equations for each of these redox states based on the algorithms that reflect the redox reactions between transporters in accordance with the known biochemical mechanism of electron transport in complex III. As is schematically shown in Figure 2, the model takes into account 12 types of reactions, nine of them constitute the proper Q-cycle and three others are just related with complex III activity: superoxid anion formation (10), QH2 supply by succinate dehydrogenase, and cytocrome c1 oxidation by cytochrome c.
The algorithms for automated calculation of reaction rates for each redox form participated in thece 12 types of reactions is described in details next.
Reaction 0 transports the first electron of QH2 bound on the positive (p-, or cytosolic) side to Fe3+ of Rieske protein and releases two protons to the intermembrane space.(5)The program simulates this reaction for all redox states of complex III capable to perform it, i.e. for the forms with attached QH2 at Qo site and oxidized Fe-S center of Rieske protein. In binary terminology described above for species bH-bL-c1-FeS-Q-Q the reaction (5) could be written as follows:(5a)here x designate any redox state of a respective transporter, 0 or 1. The expressions for reaction rates according to the mass action law are:(6)The program passes through the array for species (2), checks if the position of variable satisfy the requirements for substrate given in (5a) (that three last digits are 011), calculates the position of respective product in the array (three last digits are 101 and the digits designated as x are the same as for chosen substrate), then calculates fluxes (6) and adds them to the derivatives for the substrate and product. For the derivatives the program also has an array, where positions are organized in the same way as for concentrations.
For the other species capable to perform this reaction, Qi-Qi-bH-bL-c1-FeS-Qo-Qo, it could be written asand reaction rates expressed similar to (6).
The model takes into account the ratio of forward and reverse rate constants, which could be defined from the difference of midpoint redox potentials (Em):(6a)where n is the number of electrons transported, F = 96500 c/mol is the Faraday constant, R = 8.3 J/(mol×K) is the gas constant, T = 298 K is temperature. The derivation of this equation is based on the basic thermodynamic principles as is described in details in Text S1. The reported values for Em at pH 7 vary as much as 312 mV, [25] or 280 mV, [26] for Fe3+/Fe2+ and 200 mV, [27] or 300 mV, [28] for Q-/QH2; this variability is reflected in the variability of value expressed by (6a) from 0.6 to 18.
Reaction 1 transports the electron received by Fe2+ of Rieske protein from QH2 further to c1:All the species (1)–(4) with reduced FeS center of Rieske protein can perform this reaction:Reaction rates for core species:(7)The ratio of forward and reverse rate constants is defined by an equation similar to (6a):(7a)Midpoint potentials, for Fe3+/Fe2+ of Riske protein is 312 mV [25] and for c1 it is 341 mV [29]. This difference (ΔEm = 29 mV) results in three-time difference in forward and reverse rate constants.
Reaction 2 transports the second electron of Qo to bL:Only the species with ocupied Qo site perform this reaction:Reaction rates for bH-bL-c1-FeS-Qo-Qo:(8)The reported Em (−60 mV for bL [30] and −140 mV for Q/Q− [28]) giving ΔEm = 80 mV according to the equation similar to (7a) define the ratio of forward to reverse rate constant as 25.
Reaction 3 transports electron from bL to bHAll species with reduced bL and oxidized bH perform this reaction:Reaction rates for core species:(9)Since in this reaction electron moves in the direction from positive outer side to negative matrix side of the membrane, the rate constants depend of electric transmembrane potential [31].where ΔΨ is transmembrane potential.
Here the ratio of forward and reverse k0 is defined by midpoint redox potential difference (similar to (7a)). The reported values are −58 mV for bL [30] and 61 mV for bH [30]. This difference (ΔEm = 119 mV) defines the forward rate constant to be two orders of magnitude higher than the reverse rate constant.
Reaction 4 transports first electron from bH to Q on the n-side:This reaction is performed by the species with ubiquinone bound at Qi site:Reaction rates:(10)For bH we accepted Em = 61 mV, reported in [30]; for Q/Q−, the reported Em vary from 90 mV [30] to 45 mV [32]. To start, we accepted the value which is close to Em for bH, in this case the forward and reverse rate constants are equal.
Reaction 5 transports second electron from bH to Q− on the n-side:This reaction can be performed by species that contain semiquinone bound on Qi site:Reaction rates:(11)For bH Em = 61 mV [30]; for Q−/QH2 data in literature vary from 16.5 mV [30] to 150 mV [32]. To start, for this transition we also accepted the value close to Em for bH, in this case the forward and reverse rate constants are equal.
Reaction 6. Ubiquinol binding to and dissociation from the complex III at the cytosolic side (Qo) of the inner mitochondrial membrane. This is a way of transition between species: core (1) and species with occupied Qo site (2), and also (3) and (4):(12)The expressions for forward and reverse reaction rates are:(12a)Simulating this and other binding/dissociation reactions the program works with two arrays, since substrate and product are belong to different species. This does not change much the principles of automated calculation described for the reaction 0. In the same way the program checks the requirements for substrate position (in this case all redox form satisfy it) calculates the position of product specific for each redox form of substrate, calculates the rates (12a) and adds them to the respective derivatives.
For the reaction (12)As a starting point we assumed that Cxxxx = Cxxxx11 when QH2 = Q. In this case Kd = 2 nmol/mg of protein.
Reaction 7. Ubiquinone binding and dissociation to the complex III at the matrix side of the inner mitochondrial membrane:(13)The expressions for reaction rates:(13a)In simulations we used the same value:Reaction 8. Ubiquinone dissociation and binding to the complex III at the cytosolic side of the inner mitochondrial membrane:The expressions for reaction rates:(14)Reaction 9. Ubiquinol dissociation and binding to the complex III at the matrix side of the inner mitochondrial membrane:The expressions for reaction rates:(15)Reaction 10. ROS production in the model.
Semiquinone bound to the Qo site of complex III could be oxidized by molecular oxygen, thus producing superoxide radical:This reaction is considered as unidirectional:(16)For relative ROS production kROS was taken to be 1 and KROS is specified in the legend to Figure 6.
Reaction 11. Succinate oxidation in complex II resulted in ubiquinone reduction (Q to QH2), which rate V is described by the Michaelis' equation:(17)where KQ = 0.5 nmol/mg of protein, KS = 0.25 mM of protein, Vm = 1.88 nmol/mg of protein/s, and S is external succinate concentration. Thus, equation (17) represents succinate dehydrogenase activity in a generalized form not going into details as in above reactions.
Reaction 12. Electron transport from complex III to cytochrome c and further to complex IV.
All species (1)–(4) contribute to this rection:The expression for this reaction rate, which in fact integrates all the steps of electron transition from cytochrome c1 to oxygen, takes into account oxygen concentration:(18)here O2 is a fraction of dissolved oxygen with respect to its content at equilibrium with atmospheric oxygen at normal pressure, k is reaction rate constant, Km is Michaelis constant for interaction with oxygen. These parameters were adjusted so that the model describes limitation of electron flow by oxygen, which was found experimentally. According to the experimental data [22],[13], the electron flow is not limited by O2 availability until the oxygen concentration falls to about 1 µM (<0.1% O2). The model reproduces these data if Km = 0.03, and k = 25 s−1. These parameters were used to calculate the oxygen dependence of semiquinone radical content presented in Figure 5A of the main text.
Complex III content is 0.4 nmol/mg of protein [33], and total ubiquinone content is 4 nmol/mg of protein [33].
The system of differential equations describes the evolution of concentrations (C) of all redox forms, which the program references as is described above. Derivative of each redox state is calculated as a sum of all the reaction rates of influx to or efflux from the considered state, which are calculated in accordance with the equations (6–18). For each type of reaction there is a special function, which, once called for a specie, passes from redox state 0 to the last one checking if the redox conditions permitting the reaction are satisfied and, if yes, calculates the flux and adds it (with respective sign) to the derivatives of concentrations of reagents. In this way, to simulate the operation of respiratory complex it is necessary just to check whether the functions simulating all the reactions performed by the species (1)–(4) are called, and these functions automatically would calculate the derivatives for all 400 redox states of complex III.
For the numerical solutions of differential equations the program implements 5th order Runge-Kutta, Bulirch-Stoer, implicit Bulirch-Stoer, Rosenberg methods [34]. The program and brief user guide is free available in http://www.bq.ub.es/bioqint/selivanov.htm
Similar technique is described in more detail elsewhere [35],[36].
Isolation of rat brain mitochondria - All procedures involving animals were approved by the Children's Hospital of Pittsburgh and were in compliance with “Principles of Laboratory Animal Care” and the current laws of the United States. Rat brain mitochondria were isolated from the cortex of adult Wistar rats. After removal, tissue was minced and homogenized in ice-cold isolation buffer I (IB I) which contained: 225 mM mannitol, 75 mM sucrose, 5 mM HEPES buffer (pH adjusted to 7.3 with KOH), 0.1 mg/ml fatty acid free BSA, 1 mM tetrapotassium EDTA and 12% Percoll. The homogenate thus obtained was carefully layered on the top of a discontinuous gradient of Percoll (24% and 42%) prepared using the same buffer. The preparation was then centrifuged at 31,000×g for 10 min. The fraction containing the mitochondria located between 42% and 24% Percoll was carefully withdrawn by a syringe and washed from Percoll twice by pelleting in IB I. The resulting mitochondrial suspension was diluted in isolation buffer II (IB II), which was same as IB I, except for the concentration of EDTA (0.1 mM) and lack of albumin, and spun down at 12,000×g for 10 min. The deposit of mitochondria was homogenized in IB II at a final protein concentration of ∼20 mg/ml and stored on ice until use. The protein concentration in the mitochondrial samples was determined using a Protein Assay kit (Pierce Chemical Company, Rockford IL) according to the manufacture's instructions. Mitochondria prepared in this way were active for at least 4–5 hours, as determined by their ability to maintain a stable transmembrane potential in the presence of oxidizable substrates.
Hydrogen peroxide measurements were performed in a stirred cuvette mounted in a Shimatzu RF-5301 spectrofluorimeter maintained at 37°C. Mitochondria (0.2 mg/ml of protein) were added to the incubation medium that contained: 125 mM KCl; 2 mM KH2 PO4; 2 mM MgCl2; 10 mM Tris;10 mM HEPES (pH 7.0); 100 µM EGTA; 5 mM pyruvate and particular concentrations of succinate (0.05–5 mM) as the oxidizable substrates. Reporting fluorescent dye for H2O2 was Amplex red (2 µM) which increased fluorescence upon oxidation to resorufin in the presence of 1 U/ml of horseradish peroxidase (HRP) as previously described [36]. Measurements were carried out at excitation/emission wavelengths of 560 nm (slit 1.5 nm)/590 nm (slit 3nm), respectively. Amounts of H2O2 released by mitochondria were estimated by constructing calibration curves using known H2O2 concentrations in the standard incubation buffer together with Amplex red and HRP, but without mitochondria. Normoxia experiments were performed in a conventional square (10×10 mm), open cuvette. Anoxic experiments were performed in a cuvette with a narrow rectangular neck (10×3 mM). Anoxic conditions were made by bubbling the cuvette with N2 and, once the reaction media was depleted of O2, closing it with a Teflon lid. Resumption of oxygen supply was achieved by opening the cuvette and bubbling it with air.
Mitochondrial transmembrane potential, ΔΨm - was estimated using fluorescence quenching of the cationic dye safranine O. Since polarized mitochondria have a negative charge inside, positively charged molecules of safranine O are accumulated inside the matrix; increase in dye concentration inside the matrix leads to fluorescence quenching , thus a decrease in fluorescence corresponds to an increase of membrane potential. The excitation wavelength was 495 nm (slit 3nm) and emission 586 nm (slit 5nm), and the dye concentration used was 2.5 µM [36].
All procedures involving animals were approved by the Children's Hospital of Pittsburgh and were in compliance with “Principles of Laboratory Animal Care” and the current laws of the United States.
|
10.1371/journal.pcbi.1000477 | Can Molecular Motors Drive Distance Measurements in Injured Neurons? | Injury to nerve axons induces diverse responses in neuronal cell bodies, some of which are influenced by the distance from the site of injury. This suggests that neurons have the capacity to estimate the distance of the injury site from their cell body. Recent work has shown that the molecular motor dynein transports importin-mediated retrograde signaling complexes from axonal lesion sites to cell bodies, raising the question whether dynein-based mechanisms enable axonal distance estimations in injured neurons? We used computer simulations to examine mechanisms that may provide nerve cells with dynein-dependent distance assessment capabilities. A multiple-signals model was postulated based on the time delay between the arrival of two or more signals produced at the site of injury–a rapid signal carried by action potentials or similar mechanisms and slower signals carried by dynein. The time delay between the arrivals of these two types of signals should reflect the distance traversed, and simulations of this model show that it can indeed provide a basis for distance measurements in the context of nerve injuries. The analyses indicate that the suggested mechanism can allow nerve cells to discriminate between distances differing by 10% or more of their total axon length, and suggest that dynein-based retrograde signaling in neurons can be utilized for this purpose over different scales of nerves and organisms. Moreover, such a mechanism might also function in synapse to nucleus signaling in uninjured neurons. This could potentially allow a neuron to dynamically sense the relative lengths of its processes on an ongoing basis, enabling appropriate metabolic output from cell body to processes.
| Neurons have extremely long axonal processes that can reach lengths of up to 1 meter in human peripheral nerves. The neuronal cell body response to nerve injury is dependent on signals carried by molecular motors from the lesion site in the axon. The distance between the injury site and the cell body influences the type of response, suggesting that neurons must be able to estimate the distance of an axonal injury site, although how they do this is unknown. We have used a computational approach to model intracellular distance measurement after nerve injury. The models show the feasibility of a mechanism based on a rapid, near instantaneous, signal carried by action potentials in the nerve, followed by multiple slower signals carried on molecular motors. Such a mechanism can enable a neuron to discriminate between distances as close as 10% of total axon length. The model provides insights on retrograde injury signaling in neurons, including the biological relevance of the mechanism over different scales of nerves and organisms. Moreover, if similar mechanisms function in synapse to nucleus signaling in uninjured neurons, this could enable estimation of relative process lengths, thus guiding metabolic output from cell bodies to axons.
| Neurons extend extremely long axonal processes that can exceed the diameter of the cell body by 4–5 orders of magnitude. This poses a unique challenge for intra-cellular signaling, since nerve cells require efficient transport mechanisms to move macromolecules and metabolites from the cell body to neurite terminals and back over distance. This communication problem becomes especially acute in the context of nerve injury, when the axon needs to provide the cell body with accurate and timely information regarding the site and extent of axonal damage [1]. Cell body responses to axonal injury are diverse, ranging from functional repair to cell death, and depend on both the intrinsic regeneration capacity of the neuron and responses to the local environment [2]–[4].
The distance of the lesion site from the cell body is one of the factors determining neuronal responses to injury. For some populations of neurons, a more proximal axotomy leads to greater regenerative response by the cell body ([5]–[7] and references cited therein). Lesion distance was also shown to influence specific molecular responses to injury, including activation of cell body kinases [8] and up-regulation of growth-associated genes [5], [9]–[12]. Interestingly, the precise effect of lesion distance on neuronal response may differ in diverse neuronal populations. For example, an optic nerve lesion study reported that the number of regenerating retinal ganglion cells is inversely correlated with distance of the lesion from the optic disc [13]. In long neurons from two species of fish, lesions close to the cell body induce death, while beyond a certain lesion distance neurons regenerate [14],[15]. Moreover, the lag time for initiation of regeneration in these neurons is tightly correlated with lesion distance [14],[15]. Taken together, these findings demonstrate that neurons from different functional classes and species have the capacity to differentiate between lesion sites at different locations in their axons.
Early workers in the field proposed a number of hypotheses to explain disparate cell body responses to differently located axonal lesions [16],[17]. Diffusion mediated signaling is not likely to function efficiently over the requisite distances [18], and other mechanisms like signaling waves [19] or spatial gradients of protein abundance [20] have not been demonstrated to occur over axonal distances. On the other hand, two long distance signaling mechanisms have been characterized in nerve injury paradigms- a rapid electrophysiological signal of short duration [21] and a second slower wave of signals transported on molecular motors [1],[22]. Motor-driven signaling has emerged as a versatile mechanism for long distance communication along nerve axons [23],[24], and in this study we have used computer simulations to examine the possibility that it can provide lesion distance information in injured neurons. The analyses support feasibility of a multiple signals model, wherein distance information is inferred from the time delay between the arrival of an electrophysiological fast signal and slow signals carried by the molecular motor dynein. The simulations indicate that this mechanism can enable nerve cells to distinguish between distances of 10% or more of their total axon length.
Inferring the distance traveled by a given signal can rely on two types of mechanisms, either quantifying chemical gradients over distance, or measuring the time delay between initiation of the signal and its arrival in the detection region. Although chemical gradients play central roles in biological systems, diffusion-based gradients cannot be established over axonal distances within a biologically relevant time frame after injury [25]. Thus, we examined mainly the second possibility, namely that a time delay between the initiation of a chemical signal in an injured axon and its arrival at the cell body can be interpreted by the cell as representing the distance traveled by the signal. In order for such a mechanism to work, it requires two reference points: an early time point representing the initiation of the signal, and a later time point representing the arrival of the signal. The latter requires a detection system that responds to the arrival of an amount of signal defined by a specified threshold, while for the former requirement, we hereby suggest two models that can in principle define signal initiation:
Both models are based on measuring the arrival of a sufficient amount of the slow signal, defined as a fraction of 500 in silico particles moving in a Matlab-defined simulation environment (see methods). Since we do not have any data regarding the signal concentration required for initiating a response, our models explore a series of thresholds, defined as fractions of arriving signal from the total signal generated at the injury site. These sensitivity thresholds range between 1% and 90% of the injury signal (i.e., a 1% detector will respond to 1% of the originally generated signal, whereas accumulation of 90% of the signal is required for response of a 90% detector).
Retrograde transport along the microtubule cytoskeleton in nerve axons is almost entirely dynein-based, thus the basic assumption in our models is that the slower signals are carried retrogradely as part of a dynein based complex. Dynein velocities have been measured in diverse systems from isolated molecules in vitro to in intact cells, and by different methods including direct imaging or end-point accumulation, leading to reports of a range of velocities from ∼0.5 µm/sec to ∼5 µm/sec [23],[27]. Our models require inputs of velocity distributions (rather than average velocities), and we therefore extracted velocity distributions from two experimental data-sets, one based on in vitro analyses of movement of individual dynein–dynactin–GFP complexes [28], and another that utilized cellular imaging of the retrograde transport of a GFP-labeled endosome marker in embryonic motor neurons [29]. For both data sets, a curve fit procedure was applied (see methods) resulting in a distribution function. Based on these distribution functions, random velocity values were assigned to migrating particles simulating dynein-trafficked retrograde signals. Figure 2 depicts the experimental data and the fitted distribution functions for both data sets. Unless otherwise specified, all simulations utilized the distribution function derived from the data of Ref. [28].
As depicted in Figure 3A, our initial model system performs a comparative measurement. Two injuries are performed in two distinct cells in-silico. In one cell the injury is introduced in a proximal location, and in the other cell a distal injury is performed. In response to each of the two injuries, a slow signal emanates from the injury sites, propagating retrogradely towards the cell body. The system then measures the time delay between the fast and slow signals from the proximal location (Δt1) and the time delay between the fast and slow signals from the distal location (Δt2). The difference between Δt2 and Δt1 reflects the system's ability to distinguish between the two locations: the larger this difference, the better the system in terms of distance measurement.
The simulations explore the influence of two parameters on system performance: the distance between the two injuries (hereafter referred to as injury displacement), and the total distance between the distal injury and the detector (L, or total distance). Figure 3B depicts a schematic representation of two cases, one in which the displacement is small and the total distance is relatively large, and one in which the displacement is relatively large compared to the total distance. The intuitive prediction is that a biological system will find it more difficult to distinguish between the two injury sites in the former case rather than in the latter.
In order to assess consistency of model performance, we repeated each such in-silico experiment 100 times. In each such repetition, the same detector sensitivity, same total distance from cell body, and same injury-displacement distance were used. Differences between repetitions emerge solely from random assignment of dynein velocities to the slow signal particles. Figure 3C depicts two sets of such 100 repeats for the two-signals model (the same procedure was applied to the two-detectors model, data not shown). In the first case (Figure 3C, left panel), the parameters that were chosen were: L (total distance) = 60 cm, Disp (injury displacement) = 0.5 cm, and the detector-sensitivity threshold was set to 30%. Each vertical bar (dark blue) represents a single repetition of the simulation. The value obtained for each repetition represents the time difference Δt2−Δt1 in minutes. The negative bars observed for approximately one tenth of the repeats indicate that for these specific simulations the system infers mistakenly that the distal injury site is closer to the cell body than the proximal injury site. In another ∼10% of the repeats, the measured Δt2−Δt1 time difference in signal arrival is less than an hour. Since the different molecular events involved in both generating the retrograde injury signals at the site of injury and interpreting them at the cell body may take about an hour [30]–[32], such a time difference in signal arrival might be below the resolving power of an injured neuron (i.e., even though the signal from the proximal injury site traveled up to an hour less than the signal from the distal site, the accompanying events of signal production and/or processing may exceed this time difference, thus making it biologically irrelevant). Moreover, despite the fact that these are 100 repeats of the same injury and displacement distances, reproducibility of the measurement is clearly very poor. Thus, at least for this 0.5 cm displacement distance that is two orders of magnitude smaller than the 60 cm total injury distance, the initial model cannot discriminate between locations of the two injury sites. In the second case (Figure 3C, right panel), we set L to 5 cm and Disp to 3.5 cm, using the same sensitivity threshold of 30%. In this case, the system provided a consistent set of measurements, all ranging around 16–18 hours.
Figure 3C depicts two extreme examples of model performance for two distinct combinations of total distance/injury displacement. In order to conduct a systematic exploration of model performance, we extended this analysis to cover a wide range of distance-displacement combinations. For each distance-displacement combination, we performed 100 simulation repeats as described above. From each such set of 100 simulations we discarded the worst 5%, and then chose the minimal Δt2−Δt1 time-difference value out of the remaining 95% of the repetitions (Figure S1, red circle). Note that in the examples of Figure 3C this minimum is a negative value for the left panel, while in the case of the right panel the minimum value is approximately 1000 minutes.
We then used the collection of minima points to plot a 3D graph in which the X and Y axes represent injury displacement and total distance, respectively, and the Z axis represents the minimal Δt2−Δt1 time difference value for each X–Y combination (Figure S1, lower panel). Such graphs can be used to answer two basic questions regarding the models- first, can a given model distinguish between two distinct injury locations. This is determined by setting a cutoff for system failure due to either mistaken identification of the distal injury site as being closer than the proximal (resulting in a negative Z axis value), or a time delay that is too small to enable a differential biological response. Since differential biological responses to injury typically require transcription and translation, for purposes of the modeling the system cutoff was defined as a time delay of at least 60 minutes.
The second issue addressed by the 3D plots is whether a given model is consistent, i.e. will it provide a similar assessment for the same injury displacement, regardless of its distance from the cell body? This is reflected in the smoothness of the graph. In an ideal system, the time difference in the arrival of a signal that travels a distance x and a signal that travels a distance x+Δx should remain constant, regardless of the value of x. Thus for an ‘ideal’ 3D graph (Figure S2), straight lines along the X axis indicate consistency (i.e., for a given value of injury displacement, the time difference (Z) should remain the same at all total distance values). In order to assess the smoothness of a 3D graph plotted from the simulations, we use a root mean square deviation (RMSD) measurement. Given two sets of n points v and w, the RMSD is defined as follows:When calculating the RMSD for a model-generated graph compared to an ideally smooth graph, the lower the obtained RMSD value, the closer the graph to the ideal, hence the effects of changing parameters and models can be inferred from their comparative RMSD values.
Systematic exploration of the two-signals model showed that although it can function over part of the total distance/injury displacement combinations, the system failed over a significant portion of the distances range tested (Figure 4A, Figure S3). Furthermore, for a given injury displacement, the time-delay measurements did not show consistency over increasing total-distance values. For example, the ability of the system to detect an injury displacement of 8 cm decays with distance along the axon, and is essentially lost at total distances of 70–80 cm and above. RMSD values for a wide range of detector sensitivity thresholds indicate that the system performed better at sensitivity settings of up to 30%, and worsened significantly in the range from 40% to 80% (Figure S3).
Performance of the two-detectors model was much poorer, and in the best case the system detected injury location differences for only approximately one third of total distance/injury displacement values (Figure 4B, Figure S4). Unfortunately, the RMSD measurement seems to be uninformative for comparing different permutations of the two-detectors model. Rather than reflecting model performance, RMSD values reflect the ‘gap’ between the sensitive detector and the insensitive detector. The larger the difference between the thresholds of the two detectors, the larger the time difference between the distal and proximal locations. Thus, two 3D graphs that are similar in terms of smoothness, but differ in their Z values (time differences) will yield different RMSD values (Figure S4).
Since model performance in two signals or two detectors mode was not satisfactory, we modified the two-signals model to include several slow signals rather than a single slow signal, and assume that an effective response is triggered when a subset of these signals arrives at the cell body (Figure 5). From a biological point of view, this may reflect a situation in which there are several dynein-carried signals. We further assume that as far as detector-sensitivity is concerned, there is no significant difference between the signals (i.e., in terms of our model they utilize similar detection systems). The rationale behind this modification is that in a noisy system, multiple measurements are expected to be more accurate than a single measurement. In its original configuration, in order for a distal injury to be identified by the system as a proximal one, it was sufficient that a small fraction of the slow signal particles emanating from the distal site would randomly acquire higher velocities than the signal particles originating from the proximal point. In order for a similar phenomenon to occur in the multiple signals system, the distal point needs to randomly “win” not only once, but in several slow-signal velocity acquisitions. Figure 6 compares the performance of a system with six slow signals, of which any three will initiate a response, versus performance of the previously described system with a single slow signal. A significant improvement is observed in consistency (graph smoothness), together with a marked increase in the total distance and injury displacement ranges for which the system attains a successful outcome (Figure 6A and Figure S5). RMSD values are also significantly improved (Figure S5).
We considered examining a similar extension of the two-detectors model to multiple detectors. However, whereas extending the two-signals model to multiple signals did not require any new (and unjustified) assumptions regarding system parameters, a similar extension of the two-detectors model requires overly speculative assumptions. Consider, for example, a system with three kinds of detectors with sensitivity thresholds s1, s2, and s3, where s1<s2<s3 (i.e., s1 is the most sensitive detector). The limiting determinant of system performance will have to be the time delay between activation of s1 and s3 – having s2 as an intermediate detector will not influence the result, unless one assumes preferential effects of such intermediate detectors. In the absence of any data, such speculative configurations may be completely detached from biological reality. Nonetheless, we did try to modify some quantitative features of the slow signal, in order to check whether the poor performance of the two-detectors model results from the specific biological datasets that were used in this work. We applied the following modifications to the model: (i) using a uniform distribution of velocities instead of the data-based Gaussian distributions, (ii) using velocities 1–3 orders of magnitude faster than the data-based velocities, and (iii) using wider and narrower velocity distributions (obtained by modifying the parameters of the curve-fit functions described in the Methods section below). None of these modifications yielded any significant improvement in model performance (data not shown). It therefore seems that the the two (or multiple) signals model is qualitatively superior to the two-detectors model, and the difference in model performances cannot be attributed to a quirk of specific model configuration.
As noted above, we used two sets of dynein velocity measurements for our modeling work: a data-set from Ross et al. [28], representing velocities of isolated dynein-dynactin complexes in vitro, and a data set from Deinhardt et al. [29], based on tracking of GFP-labeled tetanus toxin in live motor neurons. The average dynein velocity measured by Deinhardt et al. was higher than the average dynein velocity measured by Ross et al. – 1.3 µm/sec and 0.45 µm/sec, respectively, and the velocity distributions of Deinhardt et al. spanned a broader range (Figure 2). As a consequence, time delays between the arrival of signals from distal and proximal locations in simulations based on the Deinhardt et al. data were smaller than in simulations based on the Ross et al. data, and simulations based on the Deinhardt et al. data were more susceptible to noise (Figure 7). Thus, in a system configuration integrating five out of ten signals (a model configuration based on multiple slow signals – see also Figure 5 and accompanying text above), simulations based on the Ross et al. data yield satisfactory results over a broader combination of distances and injury displacements than simulations based on Deinhardt et al. 's data (Figure 7). Nonetheless, increasing the number of signals and detector sensitivities for the Deinhardt et al. data show the same trends for improvement as demonstrated for simulations based on Ross et al. (data not shown). Thus, it is reasonable to assume that optimal results can be obtained also from relatively noisy motor behavior given a sufficient number of signals and appropriate detector sensitivity.
The distance between the site of injury and the cell body seems to have a significant effect on a neuron's ability to recover from mechanical injury [14]–[17],[33],[34]. Furthermore, there are both qualitative and quantitative aspects to this distance effect. In specific neuron types, once distance between cell body and site of injury drops below a certain lower threshold, no regeneration occurs, whereas above this threshold the probability of regeneration increases continuously with the increase in distance between the cell body and site of injury [14],[15]. In other neuronal populations, a more proximal axotomy leads to greater regenerative response by the cell body [5]–[7]. Despite the clear biological significance of injury distance in neural tissues, the mechanism by which distance from the site of injury is measured is unknown, and the degree of precision required from such a measurement is not clear.
In this work, we aimed at providing a theoretical framework for examining how intracellular distance measurement might be accomplished at the cellular level within a neuron. Computer simulations based on existing biological data were used to examine these concepts, and to assess their plausibility. Nonetheless, we are fully aware that the results and conclusions presented in this paper were derived from models that are abstractions of the real biological system, although we tried to keep speculations regarding the the mechanisms driving the behavior of these models to the bare minimum. We should also note some of the limitations of our approach, thus for example dynein velocity might be influenced by the type of cargo [35]. Although this was not factored into our models, the analyses show that the differences between the two velocity distributions used for model simulations do not affect key qualitative behaviors of the system (Fig. 7). Another issue not explicitly modeled is processivity of the dynein motor, namely the propensity of the motor to stall, or to move over limited distances in the opposite direction [28],[36]. In the above described simulations, signaling molecules were assigned a given velocity, and they continued moving retrogradely with that velocity throughout the entire simulation. We carried out initial tests of the effects of motor pausing behaviors by running simulations at which in each time step 30% of the particles were randomly selected to remain in the same position until the next time step (Figure S6). This modification did not seem to have any significant effect on model performance. We further examined the effect of switching velocities in the model by re-assigning velocities to 10% of the molecules once per 100 time steps (a typical simulation is of the order of 104 time steps). As can be seen in Figure S7, this modification improved the performance of the system in terms of failure percentage. This can easily be understood by considering that if a given signaling molecule undergoes velocity switches for sufficient time, eventually the velocity of each molecule will converge to the average velocity of the entire population, decreasing noise in the system. Thus, our main findings without considering the possibility of velocity switching may actually reflect a worst-case scenario.
Despite the above caveats, the modeling shows that in principle a set of dynein-mediated signals can provide intracellular distance information in an injured neuron. Furthermore, we did not have to add any “external players” to or impose speculative mechanisms on the model. Both the fast electrical signal and the slow chemical signal have been characterized in the context of nerve cell injury [21]. Moreover, such a mechanism might also function in synapse to nucleus signaling in uninjured neurons if a neurotransmitter or other synaptic stimulation elicits electrical (fast) signals concomitantly with dynein-based (slow) signals. Such a scenario has actually been reported for the neurotrophin BDNF, which elicits both rapid electrophysiological signals [37] and dynein-transported signaling endosomes [38]. NMDA receptor signaling provides another example, transmitting both acute electrophysiological signals [39] and activating macromolecule transport by importins and dynein [40],[41]. If such signaling systems are indeed used to sense synapse to nucleus distance, this would allow autonomic length measurements of neuronal processes on an ongoing basis, which in turn could guide metabolic output from neuronal cell bodies to processes.
The existence of cellular mechanisms that detect time delays between signaling events has been shown to exist in diverse biological systems (e.g. [42],[43]). Even the expansion of the model to multiple slow signals reflects the existence of multiple signaling complexes which are retrogradely transported by dynein [1],[22],[24]. The proposed model can fit a large range of nerve lengths, covering a diversity of organism sizes. Finally, the models allow two firm conclusions that might be testable experimentally in the future; first, that use of multiple and partly redundant signaling entities provides a more robust distance assessment mechanism measurement than a single signal, and second that distance detection resolution is proportional to neurite length (Figure 8). It will be intriguing to follow experimental testing of these ideas in the future.
In order to produce a velocity distribution function, a data fit procedure was applied to the two experimental datasets used in this work. Both datasets were obtained from analyses of the movement of individual molecular complexes – either in vitro [28] or in live neurons [29]. In both cases, the authors reported their results as the relative occurrence of given ranges of velocities (e.g., 5% of the observations were at velocity ranges of 0–0.2 microns/sec) over a non-exhaustive number of molecular complexes (148 discrete complexes in Reference [28], 126 in Reference [29]). Thus, the reported velocity sets are not an ideal representation of velocity distribution, but rather an experimentally limited sampling. The curve fit procedure allowed us to compute a continuous function which could then be used to randomly assign velocities to the signaling molecules in each simulation round. For this purpose, we used a built-in Matlab script (fminsearch) based on the Nelder-Mead method [44],[45]. Since our model does not acount for zero velocities, we introduced a slight modification to the Gaussian function, thus requiring the velocity distribution function to intersect with (0,0). The curve fit function that was used was of the form:Thus, for the velocity X = 0, the function yields zero occurrences.
Goodness-of-fit was assessed by calculating the root mean square deviation (RMSD) between the observed data points and the values predicted by the calculated function:where the values are given in terms of percentage-of-occurrence of given dynein velocities (see Fig. 2). This measurement provides an estimate for the average distance between a given data point and the calculated curve.
For the Ross et al. data set, the following results were obtained:RMSD = 3.1%
For the Deinhardt et al. data the following results were obtained:RMSD = 0.61%
Molecular transport complexes are represented as moving particles. Each such particle has a location in space, and it can move according to its velocity. This approach also allows extension of the model in the future to include additional molecular properties and experimental data. In our model, a signal is composed of 500 moving particles. In order for a signal to achieve its effect, a minimal fraction of the signal should arrive at the detector, this is presented as detector sensitivity in % in the results section. The influence of various sensitivity thresholds was examined during simulations.
All simulation scripts were written in MATLAB, and simulation executions were performed on the Wiccopt cluster (hosted by The Weizmann Institute's computing center) to allow parallel executions of simulations which varied in initial parameter settings. The Cluster's nodes consist of machines with: 2 quadcore xeon CPU's, 1 quadcore xeon CPU, 2 dualcore AMD opteron, and 1 dualcore AMD opteron.
|
10.1371/journal.pcbi.1007076 | Telling ecological networks apart by their structure: A computational challenge | Ecologists have been compiling ecological networks for over a century, detailing the interactions between species in a variety of ecosystems. To this end, they have built networks for mutualistic (e.g., pollination, seed dispersal) as well as antagonistic (e.g., herbivory, parasitism) interactions. The type of interaction being represented is believed to be reflected in the structure of the network, which would differ substantially between mutualistic and antagonistic networks. Here, we put this notion to the test by attempting to determine the type of interaction represented in a network based solely on its structure. We find that, although it is easy to separate different kinds of nonecological networks, ecological networks display much structural variation, making it difficult to distinguish between mutualistic and antagonistic interactions. We therefore frame the problem as a challenge for the community of scientists interested in computational biology and machine learning. We discuss the features a good solution to this problem should possess and the obstacles that need to be overcome to achieve this goal.
| In the late 1960s, Mark Kac asked, "Can one hear the shape of a drum?" challenging readers to reconstruct the geometry of the drum from a provided list of overtones. Physicists and mathematicians found that, although in general, one cannot hear the shape of a drum, many of its properties, such as its area and perimeter, can be "heard". Here, we ask whether the type of interaction being represented in a network can be "seen" when inspecting its shape—for example, whether we can distinguish networks reporting interactions between plants and their pollinators (mutually beneficial) from those representing interactions between plants and their herbivores (beneficial only to the herbivores). We show that many types of nonbiological networks can be easily separated based on structural properties, whereas determining the type of interaction represented by an ecological networks is harder. We therefore turn this problem into a challenge for the scientific community. We argue that solving this problem would greatly benefit the field, and we discuss what the consequences of a failure might be.
| Since the early days of the field, ecologists found themselves detailing the multitude of interactions occurring between different populations, such as predation, parasitism, and herbivory (all "antagonistic" interactions) or pollination, seed dispersal, and symbiosis ("mutualistic" interactions) [1, 2]. To make sense of these data, they built networks in which nodes are species and edges stand for interactions between species. Depending on the interaction being depicted by the edges, we speak of food webs (the edge i → j marks consumption of species i by j), host–parasite networks (j parasitizes i), pollination networks (animal j pollinates plant i), herbivory networks (animal j feeds on plant i), etc. Whereas the earliest published food web dates back more than a century [3], the past 20 years have seen a sharp increase in both the number of networks published and their quality (i.e., higher level of detail, larger number of species reported, and larger number of interactions, often including weighted edges).
Reducing a multidimensional object such as a network to a few numbers is a daunting task, but paralleling the progress of network analysis in other branches of science, ecologists set out to compute summary statistics on the empirical networks they collected; these metrics range from very simple measures, such as the size (number of nodes) and connectance (proportion of realized connections) of the network, to large-scale properties such as modularity (Are networks organized in blocks of dense connectivity loosely connected by few edges? [4]) and nestedness (Can interactions be organized as in a Russian doll—with specialist species choosing interactions among those of generalist species? [5–7]). As in other branches of science, ecologists also investigated degree distributions, motif profiles [8], k-cores [9], and many other network properties.
Here, we ask whether these metrics can be used to characterize the type of interaction being depicted by a network under the hypothesis that the type of interaction would affect network structure in a consistent and detectable way. Take a mutualistic network: both parties benefit from interacting, and they actively seek out this relationship from an evolutionary point of view—for example, plants reward pollinators with nectar, attract them with visual and olfactory cues, etc. Contrast this with herbivory, in which plants are desperate to avoid the interaction and display evolutionary strategies in the form of spines and thorns, toxic compounds, bark, etc. Given the contrasting forces at play, we might hypothesize that these processes give rise to networks with different structure [10, 11].
Previous analysis posited that mutualistic and antagonsitic networks’ properties would differ significantly (for example, mutualistic networks are believed to be "more nested" than antagonistic ones) [10–14]. However, this does not necessarily imply that network types can be separated based on these properties (even if the means of two distributions differ significantly, the overlap could be large enough to render a classification impossible). We frame the problem of detecting the type of interaction being represented in an ecological network by measuring its structure as a challenge to the community of scientists interested in networks, computational biology, and machine learning. We provide a database of networks of unprecedented size and completeness that can be used to test the quality of a solution, and we outline the principal hurdles that need to be overcome in order to solve this problem.
Before examining the details of the challenge, it is important to justify its timeliness. Classifying the type of an ecological network might seem like a pointless exercise: the ecologists that put it together knew perfectly well whether they were recording parasitism or pollination. However, solving this challenge is important for at least three reasons. First, showing that certain properties can be used to distinguish among networks is a post hoc justification for developing these metrics in the first place, thereby proving that they truly measure salient, relevant aspects of network structure. Second, a solution would provide a good justification for the representation of these data as networks—by compiling networks, we can infer properties of the system that would be difficult to assess otherwise. Third, with the advent of cheap, high-throughput molecular techniques, associations between species can be sampled at an unprecedented level of detail—for example, one can sequence the root of a tree to find all its fungal inhabitants, sequence insects to detect their endoparasites, etc. In this way, one could rapidly and reliably build networks of interactions between species—but what type of interactions would these species form? Some of the arbuscular mycorrhizal fungi sampled will be symbionts of the plant, whereas others will be parasites; in fact, the same species could act as either, depending on the ecological and environmental context. If we could find universal indicators of mutualistic versus antagonistic interactions, then we could make an educated guess on the type of interaction represented in these networks by just measuring a few of their properties. Note that, to keep the task as simple as possible, here we ask the broad question, What type of interaction is represented in this network? However, the same question could be asked of each single edge, as ecological networks are typically composed of a variety of interactions [27, 30].
To illustrate the details of the computational challenge with a simple, solvable example, we draw upon a large data set of nonecological networks we have compiled for this purpose. We show that it is quite easy to separate these classes of networks, even using elementary methods. We then turn to ecological networks, showing that the task is much more difficult: not only can we not separate cleanly mutualistic from antagonistic ecological networks, but we cannot even separate ecological networks from the remaining nonecological networks.
We built undirected, unweighted bipartite networks recording the connections between actors and movies (“actor collaboration”), authors of scientific articles and the journals in which they have published (“authorship”), lawmakers votes on laws (“legislature”), microbial organisms and parts of the body where they are found (“microbiome”), and city neighborhoods and crime occurrences (“crimes”). These systems were chosen based on two characteristics: (1) for each category, multiple (100 or more) networks could be built using the same methodology and (2) the networks within each category would display a large range of sizes and levels of connectivity.
We also amassed a large database of more than 500 bipartite ecological networks, representing either antagonistic (host–parasite, host–parasitoid, bacteria–phage, plant–herbivore) or mutualistic (plant–pollinator, plant–seed disperser, ant–plant, anemone–fish) interactions.
The summary statistics on the size and connectivity for each class of networks are reported in Table 1; the details of how the networks were constructed and the references for the published networks are provided in the Supporting information.
Given the great interest in network analysis, there are hundreds of network metrics one could use to attempt to characterize the different classes of networks. When it comes to ecological networks, however, two of their most accessible properties, the size (number of nodes) and connectance (proportion of realized links), should not be used for this purpose. In fact, an ecological network’s size and connectance mostly reflect the experimental design and sampling effort rather than structural differences from other networks. These networks typically integrate information through space (e.g., Where does my ecosystem "end"?) and time (Shall we build a network for each day of sampling? Week? Year?)—sampling more extensively in time and space will surely result in a larger number of species and connections.
Given these limitations, one should choose metrics that either are not influenced by size and connectance or can be rescaled to remove their effects. For the illustrative example we use to detail the computational challenge, we are going to follow the latter approach, though the former would be superior, if we were to find network properties that are insensitive to size and connectivity.
For each network, we measure the two largest eigenvalues of the adjacency matrix, which are associated with important network properties: it is well known that the first eigenvalue (spectral radius) of the adjacency matrix is maximized in perfectly nested networks [15], whereas the second eigenvalue will separate from the bulk of the spectrum in strongly modular networks [16]. Because both eigenvalues are expected to grow whenever we add nodes and connections, we normalize them by computing their expectation under two null models: an Erdős–Rényi, random bipartite graph (in which the number of nodes and connections are preserved, but nodes are connected at random, [17]), and a configuration model (in which the nodes’ degrees are preserved, but again, wiring is random [18, 19]). By computing the relative error we make when using these expectations rather than the observed values, we attempt to remove the trivial effect of size and connectance (Table 2).
Several methods of classifying objects have been developed in the literature on statistics and machine learning, and picking a single method from such an embarrassment of riches is hard. To keep the illustration of the challenge as simple as possible, here we measure the three spectral quantities introduced previously, thereby representing each network as a single point in a three-dimensional space. To better visualize the position of the networks in this space, we perform a principal component analysis (PCA), projecting the network positions on the plane defined by the first two components. Metaphorically, we are mapping out the "network space" such that, if we were to have chosen meaningful quantities, two networks belonging to the same class should be close, and the various network classes should form clusters that are well separated.
The use of a PCA facilitates the illustration of the three properties we believe any good solution to this challenge should possess:
Having demonstrated some level of success in achieving generality, specificity, and scalability, we now turn to the focus of our challenge: applying the method to ecological interaction networks.
When we project our ecological networks onto the principal component space constructed above, we find that they span a large area of the map, overlapping many of the categories that were previously well separated. Moreover, antagonsitic and mutualistic networks overlap substantially (Fig 3). These results persist when only highly replicated networks are used, when ecological networks are utilized in the construction of the principal component space, and when alternative metrics are used to construct the PCA (Supporting information).
The results presented in Fig 3 are somewhat surprising—Why are ecological networks hard to classify when there have been previous publications claiming success? Using our new, larger data set, we can reassess the proposed differences between networks of antagonistic and mutualistic interactions. Using a smaller data set, Thébault and Fontaine [12] had found significant differences in nestedess [20] and modularity [4] between mutualistic and antagonistic networks. We repeat and extend this analysis by looking at additional measures of nestedness, constraining the null model to ensure connectedness, and adding a second null model (Fig 4 and Supporting information). We find conflicting support for the use of these measures in the classification of ecological networks. Indeed, both the size and direction of these differences can be influenced by the choice of metric. Moreover, even in the case of a significant difference, there is substantial overlap in the metric distributions, suggesting that these measures cannot alone reliably distinguish network types (i.e., though the means could be significantly different from a statistical standpoint, the high variance would render a classification based on these features difficult).
For each potential method to solve the problem outlined here, one of the key decisions is which (and how many) metrics to include in the analysis, and a plethora of statistics are currently available. For illustrative purposes, we kept the number of metrics to a minimum, but in the Supporting information, we experiment with many more, obtaining the same qualitative results.
Note that the choice of metrics is not inconsequential: including too few will provide insufficient power to distinguish between similar network types, whereas too many can lead (in some modeling frameworks) to noise or multicollinearity that can obscure meaningful differences. There are also logistical concerns, as some metrics require much more computation time than others. Certain estimates of modularity, for instance, require first identifying the appropriate clustering of the species involved—a computationally intensive problem [23].
Similarly, there are a variety of methods that could be applied to this problem. Although we used simple spectral measures and PCA to render the network space in two dimensions here, a number of techniques, in particular those in the rapidly growing field of machine learning, are arguably better suited to this task. However, although popular machine learning techniques typically require the availability of a properly labeled "training set", here we achieved a good separation between different classes of nonecological networks without supplying any label—we simply mapped out the networks based on their spectral properties.
We used PCA to produce a two-dimensional "map" onto which we can place new networks. Yet PCA has a number of shortcomings. Foremost is that the space constructed varies with the data used. If we were to include more, or different, nonecological networks in the generation phase, we would obtain a different map. Additionally, we have not run statistics on the clustering we observe in principal component space, introducing an element of subjectivity into the analysis as stated. Alternative or additional methods could be implemented that incorporate objective grouping to enhance the robustness of any results.
Finally, we have analyzed unweighted, undirected graphs, despite the fact that many modern ecological networks include measures of interaction strengths [24, 25]. Although weights are likely to help with separating different classes of networks, many of the most popular network properties (e.g., motif profiles) are difficult to extend to weighted networks.
In 1966, Mark Kac asked, "Can One Hear the Shape of a Drum?" Is it possible to infer the shape of a drumhead from a list of the overtones it produces [26]? Although the short answer is no, many important properties of the drum, such as its area and perimeter, can be "heard" distinctly.
Here, we asked whether the type of interaction being represented in a network can be "seen" from its structure. We have shown that the classification is easy for several classes of nonecological networks: even foregoing size and fill, the structure of these networks is distinctive enough to allow for a reliable classification.
Yet the same task is more complicated when examining ecological networks. In this case, the naïve approach that successfully classified nonecological networks fails completely, and adding more metrics or using different approaches seems not to help with the classification (Supporting information). Because of this fact, we framed the problem as a computational challenge to the community of scientists interested in networks and machine learning. We have compiled a large data set that can be used to explore and validate potential solutions, and we have detailed three important properties that need to be fulfilled.
Succeeding in this challenge would have profound implications for the field: although networks are surely convenient objects for storing information about species’ interactions, a positive solution would prove that, by representing ecological communities as networks, we gain insights that would be precluded otherwise. Moreover, as for other areas of machine learning (e.g., face recognition), with the advent of new techniques for the high-throughput production of networks, this question could rapidly move from the purely academic to the applied side. Although current approaches to constructing networks are often laborious, we know what these networks represent. In the future, our ability to produce networks could surpass our ability to identify what interactions represent, leading to novel confusions and challenges. Finally, network metrics that can capture essential aspects of the ecology of these systems would surely deserve a special place in the hearts and minds of ecologists.
But what if a solution to this challenge cannot be found? This failure could be due to a number of reasons. First, unweighted, bipartite graphs representing a single type of interaction might not carry sufficient information for this task. Many of the ecological networks currently being published quantify interactions (e.g., measuring the number of visits a pollinator pays to a plant), and a few started detailing the different types of interactions between the species using multidimensional networks [27–30]. As such, the solution could present itself once one were to extend the appropriate metrics to weighted, multidimensional networks. Second, it could be that ecological networks have greater within-class variation than the other classes examined here. Third, the difference between mutualistic and antagonistic relationships might not be the main driver of network structure, and other, more-meaningful classifications could emerge that would be easy to infer when examining ecological network structure. Ultimately, it is likely a combination of these explanations and others not listed here, the complete description of which remains a topic for future investigation. In any case, there would be much to learn from a failure, and to celebrate in case of success.
|
10.1371/journal.ppat.0030182 | Processing of Nuclear Viroids In Vivo: An Interplay between RNA Conformations | Replication of viroids, small non-protein-coding plant pathogenic RNAs, entails reiterative transcription of their incoming single-stranded circular genomes, to which the (+) polarity is arbitrarily assigned, cleavage of the oligomeric strands of one or both polarities to unit-length, and ligation to circular RNAs. While cleavage in chloroplastic viroids (family Avsunviroidae) is mediated by hammerhead ribozymes, where and how cleavage of oligomeric (+) RNAs of nuclear viroids (family Pospiviroidae) occurs in vivo remains controversial. Previous in vitro data indicated that a hairpin capped by a GAAA tetraloop is the RNA motif directing cleavage and a loop E motif ligation. Here we have re-examined this question in vivo, taking advantage of earlier findings showing that dimeric viroid (+) RNAs of the family Pospiviroidae transgenically expressed in Arabidopsis thaliana are processed correctly. Using this methodology, we have mapped the processing site of three members of this family at equivalent positions of the hairpin I/double-stranded structure that the upper strand and flanking nucleotides of the central conserved region (CCR) can form. More specifically, from the effects of 16 mutations on Citrus exocortis viroid expressed transgenically in A. thaliana, we conclude that the substrate for in vivo cleavage is the conserved double-stranded structure, with hairpin I potentially facilitating the adoption of this structure, whereas ligation is determined by loop E and flanking nucleotides of the two CCR strands. These results have deep implications on the underlying mechanism of both processing reactions, which are most likely catalyzed by enzymes different from those generally assumed: cleavage by a member of the RNase III family, and ligation by an RNA ligase distinct from the only one characterized so far in plants, thus predicting the existence of at least a second plant RNA ligase.
| Interactions of viroids with their host plants are unique because these subviral pathogenic RNAs lack protein-coding capacity. Therefore, hosts must provide all enzymes and auxiliary factors that viroids need for their infectious cycle. Replication of viroids entails reiterative transcription of their single-stranded circular genomes, cleavage of oligomeric strands to unit-length, and ligation to circular RNAs. While cleavage in chloroplastic viroids (family Avsunviroidae) is autocatalytic and mediated by hammerhead ribozymes, where and how cleavage of oligomeric (+) RNAs of nuclear viroids (family Pospiviroidae) occurs in vivo is controversial. We have re-examined this question in vivo, taking advantage that dimeric viroid RNAs expressed transgenically in Arabidopsis thaliana are processed correctly. Together with mapping the in vivo processing site of three members of the family Pospiviroidae, our results with 16 mutants of one of these viroids support that cleavage is directed by an RNA motif conserved in all members of the family, and ligation by an extended conformation containing a motif termed loop E. Both processing reactions are most likely catalyzed by enzymes different from those generally assumed: cleavage by an RNase III–like enzyme, and ligation by an RNA ligase distinct from the only one characterized so far in plants.
| Viroids, plant pathogens with a minimal non-protein-coding circular RNA genome of 246–401 nt [1], are classified into two families. The members of the first, Pospiviroidae, replicate in the nucleus through an asymmetric rolling-circle mechanism, have a central conserved region (CCR), and cannot form hammerhead ribozymes. The members of the second, Avsunviroidae, replicate in the chloroplast through a symmetric rolling-circle mechanism, lack a CCR, and can form hammerhead ribozymes [2–4]. In Potato spindle tuber viroid (PSTVd) [5,6], the type species of the genus Pospiviroid (family Pospiviroidae), the incoming monomeric circular RNA, to which (+) polarity is arbitrarily assigned, is reiteratively transcribed by the nuclear RNA polymerase II into oligomeric (−) RNAs that in turn serve as template for synthesis of oligomeric (+) RNAs. These latter transcripts are then cleaved and ligated to the mature viroid circular RNA [7,8]. In Avocado sunblotch viroid (ASBVd) [9], the type species of the family Avsunviroidae, the oligomeric (−) RNAs generated by a chloroplastic RNA polymerase are cleaved and ligated before serving as template for a second rolling-circle leading to the mature viroid circular RNA. In this family, the oligomeric RNA intermediates of both polarities self-cleave through hammerhead ribozymes [10,11]. In contrast, cleavage and ligation of oligomeric (+) RNAs in the family Pospiviroidae is catalyzed by host enzymes [12–14], which recognize particular RNA motifs.
Early infectivity bioassays with viroid RNAs containing repeated sequences of the upper CCR strand [15–18] implicated these sequences in processing of the oligomeric (+) strands of the family Pospiviroidae through the adoption of either hairpin I, a metastable motif that can be formed by the upper CCR strand and flanking nucleotides during thermal denaturation [19], or through a thermodynamically stable double-stranded structure that can be alternatively assumed by the same sequences of a dimeric (or oligomeric) RNA [15,18] (Figure S1). More recently, in vitro and thermodynamic analyses of the products obtained by incubating a potato nuclear extract with a full-length PSTVd RNA containing a 17-nt repeat of the upper CCR strand have led to the proposal that cleavage of (+) strands is driven by a multibranched structure with a hairpin—different from hairpin I—capped by a GAAA tetraloop conserved in members of the genus Pospiviroid, which subsequently switches to an extended conformation with a loop E that promotes ligation [20] (Figure S1). Similar results have been obtained with a reduced version of this construction containing the GAAA tetraloop and loop E [21]. Loop E, a UV-sensitive motif of RNA tertiary structure that is conserved in PSTVd and members of its genus [3] and exists in vitro [22] and in vivo [23,24], has been also involved in host specificity [25], pathogenesis [26], and transcription [27]. The structural model based on isostericity matrix and mutagenic analyses derived recently for PSTVd loop E [27] can be extended to CEVd. However, the proposed cleavage-ligation mechanism [20] may not apply to other members of the family Pospiviroidae that cannot form the GAAA tetraloop and loop E [3]. Moreover, alternative processing sites in the lower CCR strand, or outside this region, have been observed for several members of the family Pospiviroidae [28–32], with a proposal even suggesting that cleavage could be autocatalytic, albeit mediated by non-hammerhead ribozymes [33].
Here we have re-examined this question in vivo using a system based in transgenic Arabidopsis thaliana expressing dimeric (+) transcripts of Citrus exocortis viroid (CEVd), Hop stunt viroid (HSVd), and Apple scar skin viroid (ASSVd) [34] of the genera Pospiviroid, Hostuviroid, and Apscaviroid, respectively, within the family Pospiviroidae [3]. In addition to mapping what we believe is their major processing site in vivo, data obtained with 16 CEVd mutants support a previous model involving the hairpin I/double-stranded structure formed by the upper CCR strand and flanking nucleotides in cleavage [15], and loop E and flanking nucleotides of both strands in ligation. A corollary of our results is that the RNase and RNA ligase that catalize both processing reactions are most likely different from those generally assumed so far.
We first re-examined where the processing site occurs in vivo. This question could be tackled by mapping the 5′ termini of the monomeric linear (ml) viroid (+) strands isolated from infected propagation hosts (e.g., gynura for CEVd). However, the contribution of nicked byproducts of the monomeric circular (mc) viroid (+) RNA—the most abundant replication product—resulting from in vivo turnover or artifactual degradation during purification has precluded an unambiguous analysis so far. As an alternative, we evaluated the A. thaliana–based transgenic system established recently [34]. More specifically, an A. thaliana transgenic line expressing a CEVd (+) dimeric transcript (dt) [34] was chosen. Northern blot hybridization of RNAs separated by denaturing PAGE showed that plants of this line accumulate, besides the dt viroid (+) RNAs, the mc and ml CEVd (+) forms; however, in contrast to the situation in CEVd-infected gynura, the ml RNA was more abundant than its circular counterpart (Figure 1A and 1C, compare lanes 2 and 4). This indicates that A. thaliana can process correctly the dt CEVd (+) RNAs, with cleavage being more efficient than circularization, thus reducing the contribution of nicked mc (+) RNAs to the population of ml (+) RNAs present in vivo. RT-PCR amplification, cloning, and sequencing of the mc CEVd (+) forms extracted from the transgenic A. thaliana line confirmed that they were full-length [34], and infectious when mechanically inoculated to tomato (unpublished data). Furthermore, northern blot hybridization of another transgenic A. thaliana line expressing a dt CEVd (−) RNA showed low accumulation levels of the mc (+) RNA (Figure 1C, lane 5), indicating that despite not being a typical host, A. thaliana has the enzymatic machinery for replicating CEVd, in line with previous results for HSVd [34].
RNAs from the transgenic A. thaliana line expressing the dt CEVd (+) species were separated by denaturing PAGE, and the positions in the gel of the mc and ml (+) CEVd RNA were inferred using a purified marker stained with ethidium bromide. RNAs migrating in the region of ml CEVd (+) RNA were eluted and examined by northern blot hybridization with a CEVd-specific probe that excluded the presence of other viroid RNA species (unpublished data). Preliminary length estimation of the CEVd-cDNAs from extensions on this RNA with 5′-end labeled primers PI, PIV, and PV, run in denaturing gels in parallel to RNA markers of known size, mapped the processing site around position 100 (Figure S2). Further analysis of the CEVd-cDNAs from extensions on the same RNA with the proximal 5′-end labeled primers PI and PII (Figure 2C), also run in denaturing gels but this time in parallel to sequencing ladders, revealed with both primers single bands corresponding to stops at position G97 (Figure 2A and 2B, lanes 5). These results identified the processing site between G96 and G97, which occupy the third and fourth positions of the tetraloop capping hairpin I (Figure 3), and two central positions of the double-stranded structure that the upper CCR strand and flanking nucleotides can form in di- or oligomeric viroid RNAs (see below). The secondary structure model here presented for hairpin I (Figure 3), with a capping tetraloop [18,35], differs from the original with a capping loop of 14 nt inferred from thermal denaturation studies with PSTVd [19]. As controls, ml and mc CEVd (+) RNAs obtained in parallel from CEVd-infected gynura were also analyzed. Prominent bands resulting from stops at G97 were also observed for the ml CEVd (+) RNA (Figure 2A and 2B, lanes 6), although accompanied by others (particularly in the extension with PII). Some of the extra bands were also observed for the mc CEVd (+) RNA, suggesting that they arise from elements of secondary structure, but others were not, thus supporting the contribution of nicked circular forms to the population of linear forms (Figure 2A and 2B, lanes 7). Altogether, these results identified in CEVd (+) strands a major processing site in vivo located in the upper CCR strand.
To explore how general this finding was, processing was also studied in two additional members of the family Pospiviroidae, each with a characteristic hairpin I/double-stranded structure: HSVd and ASSVd [3,35]. RNA preparations containing the ml HSVd and ASSVd (+) RNAs as the only viroid species were isolated from two transgenic A. thaliana lines expressing the corresponding dt viroid (+) RNAs. However, in contrast to the situation observed in the CEVd-expressing line, these transgenic lines accumulated similar or more mc viroid (+) forms than their ml counterparts [34]. Parallel RNA preparations from HSVd-infected cucumber and ASSVd-infected apple, as well as the purified mc viroid (+) RNAs, were also analyzed. Extension with primer PIII identified the HSVd processing site at G82-G83 and extension with primer PIV identified the ASSVd processing site at G90-A91 (Figure 4). These two sites map, like in CEVd, between the third and fourth nucleotide of the tetraloop capping hairpin I (Figure 3), and at two central positions of the double-stranded structure formed by the upper CCR strand and flanking nucleotides in di- or oligomeric viroid RNAs (see below). It is pertinent in this context to note that the processing site here identified for ASSVd does not coincide with the corresponding site of Citrus viroid-III (also of the genus Apscaviroid) predicted from thermodynamic analysis and comparisons with PSTVd [36]. An alternative hairpin capped by a GAAA tetraloop, the RNA motif proposed to direct cleavage in a potato nuclear extract primed with a ml PSTVd (+) RNA containing a 17-nt repetition of the upper CCR strand [20], can neither be formed by HSVd nor by ASSVd. However, the PSTVd processing site inferred with this in vitro system [20] was in a position equivalent to that mapped here for CEVd with the A. thaliana–based in vivo system.
Collectively, these results strongly suggest a role for the hairpin I/double-stranded structure in directing cleavage in vivo of the oligomeric (+) RNAs in the family Pospiviroidae. If this double-stranded structure is indeed the substrate for the cleavage reaction, the enzyme involved would be most likely an RNase III. Intriguingly, the cleavage sites in each strand of the proposed substrate are separated by two 3′-protruding nucleotides, as also occurs in reactions catalyzed by enzymes of this class (see below).
To gain further support for the RNA motif(s) directing processing in vivo, we determined how 16 different mutations (Table 1) affected cleavage and ligation of dt CEVd (+) RNAs expressed transgenically in A. thaliana. We selected the mutations according to their potential discriminatory effects on: i) the GAAA tetraloop [20], ii) the hairpin I/double-stranded structure formed by the upper CCR strand, and iii) the loop E motif formed by a subset of nucleotides of the upper and lower CCR strands (Figure 5A). It should be noticed that single mutations affect one position in the GAAA-capped hairpin and in hairpin I, but two positions in the double-stranded structure; similarly, the double mutations affect two positions of the hairpin structures, and four positions of the double-stranded structure. For an easier understanding we have clustered the mutations in three groups: those located in central positions of the upper CCR strand, in peripheral positions of the upper CCR strand, and in the lower CCR strand (the effects of the two latter groups will be presented in the two following sections). RNA preparations from the 16 transgenic lines, and from the line expressing wild-type (wt) CEVd, were analyzed by northern blot hybridization after single denaturing PAGE (in which the dt RNAs and their ml and mc processing products are separated) (Figure 5B), or double PAGE (in which better resolution of the ml and mc forms is achieved) (Figure 5C). Mutations in the viroid processing products were confirmed by cloning and sequencing the viroid circular RNA from different A. thaliana transgenic lines.
Mutant #1 (C95→U) has no effect on the stem stability of hairpin I and only debilitates the stem of the GAAA-capped hairpin (a pair G:C is converted into G:U) (Figure 5A). However, in the double-stranded structure, this mutation affects a base pair phylogenetically conserved in the family Pospiviroidae (Figure 3) located in positions very close to the cleavage sites of both strands (Figure 5A). Therefore, if cleavage is directed by the double-stranded structure, changes in these positions are expected to have a negative influence; this was the case, with cleavage being reduced to 38% with respect to wt (Figure 5B).
Results with mutant #2 (G96→A) also support this view because the substitution has no effect on the stem stability of hairpin I and strengthens the stem stability of the GAAA-capped hairpin (a pair G:U is converted into A:U) (Figure 5A). But in the double-stranded structure this mutation affects a base pair also conserved in the family Pospiviroidae and adjacent in both strands to the cleavage sites, which are no longer embedded in an uninterrupted GC-rich helix (Figure 5A). Cleavage was reduced to less than 20% with respect to wt, consistent with a key role of the double-stranded structure in this reaction (Figure 5B).
The corresponding double mutant #3 (C95→U and G96→A) does not essentially alter the stem stability of both hairpin I and the GAAA-capped hairpin but, in contrast to the single mutant #2, restores the stability of the double-stranded structure (two contiguous G:C pairs are substituted by A:U pairs) (Figure 5A). However, cleavage was not restored (Figure 5B), indicating a requirement either for a particular sequence of the two nucleotides preceding the cleavage sites, or for a high thermodynamic stability of the secondary structure in the surrounding region (in which G:C pairs are prevalent).
Mutants #4 (G97→A), #5 (G97→U), and #6 (G97→C), have no influence on the stem stability of hairpin I or weaken the stem stability of the GAAA-capped hairpin (a C:G pair is disrupted). In the double-stranded structure mutations at this position affect nucleotides in both strands adjacent to both cleavage sites, which as in mutant #2 are no longer embedded in a double-stranded region (Figure 5A). Although reduction of cleavage (less than 25% with respect to wt, Figure 5B) supports also the involvement of the double-stranded structure in this reaction, these data can be alternatively interpreted as resulting from a destabilization of the GAAA-capped hairpin. However, cleavage was totally recovered in the double mutant #7 (G97→C and C94→G), in which the stability of the double-stranded structure was restored (these are indeed the nucleotides existing in the corresponding positions of CbVd-1, see Figure 3), whereas the GAAA-capped hairpin was further destabilized, thus providing additional credence to the role of the double-stranded structure in directing cleavage (Figure 5A and 5B).
The seven mutants of this group, despite not affecting nucleotides of loop E (Figure 5A), had a marked negative effect on ligation (Figure 5C). These results indicate that the sequence and/or secondary structure requirements for this reaction are more demanding than those regarding cleavage, and that they include nucleotides apart from those of loop E. The adjacent bulged-U helix (Figure 5A), the stability of which is affected by most of these mutants, appears particularly relevant in this respect.
These mutations, besides covering alternative positions of the upper CCR strand, were anticipated as very informative because most impinge on the GAAA tetraloop capping the hairpin that according to Baumstark et al. [25] directs cleavage, and also because most of these nucleotides form part of the loop E that presumably mediates ligation [20,27] (Figure 6A).
The double mutant #8 (C92→G and G99→C, the rationale for the second substitution is given below), and the single mutants #9 (A100→U), #10 (A100→C), #11 (A101→C), #12 (A102→U), and #13 (A102→C), had in general a mild effect on cleavage. Excepting mutant #9, in which cleavage was reduced to 34% with respect to wt, cleavage of the others was at least 68%, with mutants #8, and #11 to #13, reaching 88%–92% (Figure 6B). Given that the GAAA tetraloop belongs to the family of GNRA tetraloops (in which N is any base and R a purine) [37], these results do not support a role in cleavage of the GAAA-capped hairpin, because changes disrupting interactions critical for the tetraloop stability (mutants #8, and #11 to #13) had essentially no influence on cleavage. In contrast, the six mutants induced a pronounced negative effect on ligation, thus sustaining a critical function of loop E in this reaction (Figure 6C). Indeed, from the structural model derived for loop E of PSTVd [27], and considering that nucleotides critical for this motif are conserved or substituted by others preserving it in CEVd, mutants #9, #10, #12, and #13 all introduce non-isosteric pairs disrupting the loop E structure. However, mutant #11 is predicted to maintain the loop E structure, suggesting that its negative effect on ligation could result from sequence rather than from structural restrictions. Consistent with this view, the nucleotide corresponding to position A101 in CEVd is phylogenetically conserved in the family Pospiviroidae (Figure 3).
On the other hand, mutants #11 to #13 affect minimally the stem stability of hairpin I (particularly of its upper portion because they map outside the 3-bp stem adjacent to the tetraloop) and the double-stranded structure (in which they are outside the GC-rich central region of 10 bp containing the cleavage sites) and, therefore, their effects are consistent with a function of this latter structural motif in cleavage (Figure 6A). The negative effects of mutants #9 and #10 on cleavage are also compatible with this view, because they alter the stability of both the 3-bp stem adjacent to the hairpin I tetraloop and the 10-bp central region of the double-stranded structure, although it is difficult to interpret why cleavage was significantly more reduced in mutant #9 than in #10 (Figure 6A and 6B).
Going back to the double mutant #8, its high cleavage (Figure 6B) can be explained because, despite affecting nucleotides C92 and G99 that form a pair phylogenetically conserved in the hairpin I/double-stranded structure of the family Pospiviroidae, this base pair is just inverted (Figure 6A). In mutant #14 (C92→U), in which the pair between C92 and G99 was substituted by a U:G pair, cleavage still was relatively significant (63%). The differential effect in ligation of these two mutants is intriguing: ligation was essentially abolished in mutant #8, whereas in mutant #14 was close to 10% with respect to wt (the highest value for any of the mutants of the present study) (Figure 6C). It is worth noting that the double mutant #8 affects the nucleotide of the upper CCR strand that upon UV irradiation becomes cross-linked to U266 of the lower CCR strand (data not shown in [22]) and also disrupts a G:C pair of the flanking bulged-U helix, in contrast to the single mutant #14 in which this base pair is substituted by a G:U pair (Figure 6A). These results again underline that ligation is influenced by nucleotides aside from those conserved in loop E.
If only nucleotides of the upper CCR strand direct cleavage, its extension should not be influenced by mutations in the lower CCR strand, which in contrast should reduce ligation particularly if they impinge on nucleotides of the loop E motif that presumably directs this reaction. To test this hypothesis, we constructed two mutants in which U266, the nucleotide of the lower CCR strand that upon UV irradiation becomes cross-linked to G99 of the upper CCR strand (data not shown in [22]) (Figure 7A), was changed: mutants #15 (U266→C) and #16 (U266→A). Extending to CEVd the structural model derived for loop E of PSTVd [27], U266 and A100 in loop E of CEVd should interact via trans Watson-Crick/Hoogsteen edges and belong to the isosteric subgroup I1. In mutants #15 and #16, C266 and A100, and A266 and A100, are predicted to interact similarly; however, they belong to subgroups I2 and I4, respectively, which are non-isosteric with respect to the original I1 and may thus disrupt the loop E structure to some extent [27]. Northern blot hybridization of RNAs from the corresponding transgenic A. thaliana lines showed that cleavage remained essentially unaffected (90%–95% relative to wt), whereas ligation was essentially abolished (Figure 7B and 7C). These results support further the notion that cleavage is determined exclusively by RNA motifs formed by nucleotides of the upper CCR strand and flanking nucleotides, and also show that ligation is determined by nucleotides of loop E and by others of both CCR strands. In particular, the bulged-U helix may play a key role in aligning the termini to be ligated.
Processing of oligomeric (+) RNAs in the family Pospiviroidae entails cleavage to the ml (+) RNA, and ligation of the resulting species to the mc (+) RNA. Hence, the most direct way to identify the processing site is mapping where the ml (+) RNA intermediate is opened. Previous studies have pointed to the upper strand of the CCR, which, due to its strict conservation within each genera of the family, has been long assumed to play an essential role. Data supporting this view include infectivity bioassays with different PSTVd DNA and RNA constructs [17], with longer-than-unit HSVd clones [16], and with CEVd constructs containing sequence repetitions and point mutations in the upper CCR strand [18]. The latter study concluded that processing occurs at one of three consecutive Gs of the upper CCR strand, and advanced hairpin I or an alternative double-stranded structure as the putative RNA motifs directing cleavage (Figure 5A). A critical reassessment of all these data led to a model involving the double-stranded structure in cleavage, although the model did not predict the mechanism of cleavage-ligation or specify the exact processing site [15]. The infectivity-based approach, however, has an important limitation: bioassays do not provide a linear dose-response, being at best semi-quantitative and making it difficult to draw accurate estimations. Reflecting this limitation, other data point to alternative processing sites in the PSTVd lower CCR strand [28]. Furthermore, transcripts with only a 4-nt repetition of the PSTVd upper CCR strand [38] or with the exact unit-length CEVd [39] are still infectious, and another work suggested that the basic requirement for infectivity of a range of unit-length CEVd in vitro transcripts starting at different domains of the molecule is their ability to form a short double-stranded region of viroid and vector sequences at the junction of the two termini [31]. Therefore, at least in some cases, infectivity is independent of duplicated viroid sequences, possibly because the exact full-length sequence is restored by strand switching of a jumping polymerase during transcription [31]. On the other hand, primer-extension on the ml viroid (+) RNAs isolated from infected propagation hosts also has significant constraints (see Results), with this approach having mapped several processing sites in different PSTVd domains [29,30,32]. Finally, conclusions from in vitro assays in which a potato nuclear extract was primed with an ml PSTVd (+) RNA with a 17-nt repeat of the upper CCR strand should be interpreted with caution, because the processing complex formed in vitro may not mimic the corresponding complex in vivo. Moreover, prior to incubation with the nuclear extract, the PSTVd RNA was heated to promote the adoption of a specific secondary structure that may not parallel that existing in vivo [20].
The A. thaliana–based system reported recently [34] circumvents most of these limitations. It is an in vivo system in which the available data indicate that processing is correct: transgenically expressed dt (+) RNAs of typical members of the family Pospiviroidae are cleaved to the ml forms—implying recognition of two identical sites—and then ligated to the infectious mc RNAs, whereas the complementary dt (−) RNAs are not ([34], this work), thus reproducing the situation observed in typical hosts. However, in contrast to typical hosts in which the turnover of the longer-than-unit (+) replicative intermediates is difficult to follow because of their low accumulation and diverse size, the A. thaliana–based system with the viroid-expression cassette integrated in the plant genome provides a constant supply of a size-specific replicative-like intermediate that permits the easy quantification thereof and of its processing products. Moreover, despite typical members of the family Pospiviroidae being able to complete their replication cycle when expressed transgenically as dt (+) RNAs in A. thaliana, the replication level in this non-host plant is rather low (see Figure 1 and [34]), and the ml and mc (+) RNAs can be assumed to come essentially from processing of the transgenically expressed dt (+) RNA. Therefore, the effects of specific mutations in the primary transcript on cleavage and ligation can be evaluated—regardless of whether the resulting products are infectious or not—and it is even possible to identify mutations affecting only ligation.
Our results with the A. thaliana–based in vivo system mapped the cleavage site of CEVd (+) strands at the upper CCR strand, in a position equivalent to that inferred for PSTVd with an in vitro system [20]. However, we consider that the RNA motif directing cleavage in vivo is not the GAAA-capped hairpin proposed previously [20], but the hairpin I/double-stranded structure. The first argument supporting this view is that whereas the cleavage sites of HSVd and ASSVd (+) strands also map at equivalent positions in a similar hairpin I/double-stranded structure, these viroids cannot form the GAAA-capped hairpin. In contrast, examination of the hairpin I/double-stranded structure reveals some appealing features. Hairpin I is composed by a tetraloop, a 3-bp stem, an internal symmetric loop of 1–3 nt in each strand that presumably interact by non-Watson-Crick base pairs [40], and a 9–10-bp stem that can be interrupted by a 1-nt symmetric or asymmetric internal loop [18,35] (Figure 3). Remarkably, these structural features are conserved in the type species of the five genera composing the family Pospiviroidae and additionally: i) the capping tetraloop is palindromic itself, and ii) the two central positions of the tetraloop and the central base pair of the 3-bp stem are phylogenetically conserved (Figure 3) [35]. As a consequence, a long double-stranded structure with a GC-rich central region of 10 bp containing the cleavage sites can be alternatively assumed by the same sequences in a di- or oligomeric RNA (Figure 5A). The second argument supporting the hairpin I/double-stranded structure as the RNA motif directing cleavage derives from the effects on this reaction of mutants affecting differentially this motif versus the GAAA-capped hairpin. Chief among them are mutants #8, and #11 to #13 that, despite disrupting interactions crucial for the stability of the GAAA tetraloop, did not basically modify cleavage. Furthermore, because the ml PSTVd (+) RNA with a 17-nt repeat of the upper CCR strand that was used to prime the potato nuclear extract [20] can also form a fragment of the proposed double-stranded structure containing the cleavage sites, the correct cleavage observed in vitro can be alternatively interpreted as being directed by this structure. Our interpretation of direct effects of the introduced mutations in viroid RNA processing is based on the weak viroid RNA-RNA amplification in A. thaliana and, therefore, side effects of this amplification in cleavage and ligation cannot be totally discarded.
In summary, we believe that the substrate for cleavage in vivo of all members of the family Pospiviroidae is the double-stranded structure proposed by Diener [15], with hairpin I playing a role in promoting the adoption of this structure (see below). Although its existence in vivo remains to be fully demonstrated, we have noticed that the cleavage sites in the double-stranded structure leave two 3′-protruding nucleotides in each strand (Figure 8), the characteristic signature of RNase III enzymes [41,42]. The participation of an enzyme of this class, of which there are at least seven in A. thaliana [43], is consistent with the nuclear location of some of them, which additionally have preference for substrates with a strong secondary structure resembling that of viroids. Moreover, one or more RNase III isozymes should be involved in the genesis of the viroid-derived small RNAs with properties similar to the small interfering RNAs that accumulate in viroid-infected tissues [44–48]. Going one step further, if an RNase III indeed catalyzes cleavage of the oligomeric (+) RNAs of the family Pospiviroidae, the resulting products should have 5′-phosphomonoester and free 3′-hydroxyl termini. Characterization of the ml (+) RNAs from A. thaliana transgenically expressing dt CEVd (+) RNAs shows that this is actually the case (M. E. Gas, D. Molina-Serrano, C. Hernández, R. Flores, and J. Daròs, unpublished data). The adoption in vivo of the double-stranded structure with a GC-rich central region containing the cleavage sites could be promoted by hairpin I because prior work with PSTVd has mapped a dimerization domain at this hairpin [40]. This situation resembles that observed previously in retroviruses in which dimerization, a critical step of their infectious cycle, is mediated by a hairpin with a palindromic loop that can dimerize co- or post-transcriptionally via a kissing loop interaction between two viral RNAs [49]. During transcription of oligomeric (+) RNAs of the family Pospiviroidae, a kissing loop interaction between the palindromic tetraloops of two consecutive hairpin I motifs might similarly start intramolecular dimerization, with their stems then forming a longer interstrand duplex (Figure 8). Part of the negative effects on cleavage of mutants #1 to #6 (Figure 5) could result from weakening dimerization.
Regarding ligation, our results support that the substrate for this reaction in the genus Pospiviroid is the extended conformation containing loop E [20,22] (Figure 8). Therefore, whereas cleavage is solely dependent on the upper CCR strand and flanking nucleotides, ligation is dependent on nucleotides of both CCR strands that encompass those of loop E and others adjacent. This entails a switch between two conformations, one for cleavage and another for ligation, which might be facilitated by the RNA helicase activity associated with some RNase III enzymes [43]. Because within the family Pospiviroidae loop E is only formed in the genera Pospiviroid and Cocadviroid, other genera of this family must have alternative motifs playing a similar role in ligation. Potential candidates are the extended conformation of the CCR with a bulged-U helix conserved in all members of the genera Pospiviroid, Hostuviroid and Cocadviroid, and similar structures in the other genera of this family. Last but not least, the 5′-phosphomonoester and free 3′-hydroxyl termini resulting from cleavage mediated by an RNase III predict the existence of an RNA ligase able to join these ends, which is therefore different from the class represented by the wheat-germ RNA ligase that catalyzes joining between 5′-hydroxyl and 2′,3′-cyclic phosphodiester termini [50]. This latter RNA ligase class has been long regarded as the enzyme involved in circularization of PSTVd (+) strands and, by extension, of other members of its family [29,51]. Our results advise for a reassessment of this long-held paradigm. The A. thaliana–based system is a promising tool for dealing with this and other related questions because it combines the advantages described previously with a broad collection of mutants.
Viroid sequence variants were CEVd (M34917) having a deleted G between positions 70 and 74, HSVd (Y09352), and ASSVd (AF421195). Plasmids pBmCEVdB, pBmCEVdS, and pBmCEVdP contained monomeric CEVd-cDNAs cloned at the BamHI, SacI, and PstI sites, respectively, pBmHSVdE a monomeric HSVd-cDNA cloned at the EcoRI site, and pBmASSVdS a monomeric ASSVd-cDNA cloned at the SalI site. Plasmids containing head-to-tail dimeric cDNA inserts of CEVd, HSVd, and ASSVd have been described previously [34]. Binary vectors for plant transformation were constructed by replacing the β-glucuronidase-cDNA of pCAMBIA-2301 (AF234316) by dimeric head-to-tail CEVd-cDNAs (starting at the PstI site) corresponding to the wt (pCKdCEVd-wt) and 16 mutants (pCKdCEVd-1 to pCKdCEVd-16) (Table 1).
Plasmid pBmCEVdP was amplified with a series of pairs of 5′-phosphorylated adjacent primers that were complementary and homologous to different regions of the wt CEVd sequence, except in some 5′-proximal positions in which changes were introduced to obtain the desired mutants (Table 1). Pwo DNA polymerase was used in the buffer recommended by the supplier (Roche Applied Science). After initial heating at 94 °C for 2 min, the amplification profile (30 cycles) was 30 s at 94 °C, 30 s at 58–68 °C (depending on the predicted melting temperatures), and 3.5 min at 72 °C, with a final extension of 10 min at 72 °C. PCR products corresponding to the full-length plasmid were eluted after agarose gel electrophoresis, ligated, and used for transformation. Incorporation of the expected mutations was confirmed by sequencing. The mutated CEVd-cDNA inserts were PCR-amplified, eluted after non-denaturing PAGE, and ligated to obtain dimeric cDNAs that were cloned in pBluescript II KS (+). Plasmids with dimeric head-to-tail inserts were selected by restriction analysis and subcloned in the binary vector pCAMBIA-2301.
Agrobacterium tumefaciens (strain C58C1) was transformed with plasmids (pCKdCEVd-wt and pCKdCEVd-1 to pCKdCEVd-16) following standard protocols. Transformation of A. thaliana (ecotype Col-0) was performed by the floral dip method using midlog-grown cultures of A. tumefaciens [52], and transgenic plants were selected by germinating the seeds from dipped A. thaliana in plates with 100 μg/ml kanamicine, 300 μg/ml cefotaxime, and 10 μg/ml benomyl.
Total nucleic acids from leaves of CEVd-infected gynura (Gynura aurantiaca DC), HSVd-infected cucumber (Cucumis sativus L.), and transgenic A. thaliana, as well as from fruits of ASSVd-infected apple (Malus pumilla Mill.), were extracted with buffer-saturated phenol and enriched in viroid RNAs by chromatography on non-ionic cellulose (CF11, Whatman) [34]. RNAs from CEVd-infected gynura and ASSVd-infected apple were further fractionated with 2 M LiCl.
RNA aliquots were separated by either single denaturing PAGE in 5% gels with 8 M urea in 1X TBE (89 mM Tris, 89 mM boric acid, 2.5 mM EDTA [pH 8.3]), or double PAGE, first in a non-denaturing 5% gel in TAE (40 mM Tris, 20 mM sodium acetate, 1 mM EDTA [pH 7.2]), with the gel segment containing the monomeric viroid RNAs being cut and applied on top of a second 5% gel with 8 M urea in 0.25X TBE. RNAs were electroblotted to nylon membranes (Hybond-N, Amersham Biosciences), UV-fixed with a cross-linker (Hoefer), and hybridized (at 70 °C in the presence of 50% formamide) with strand-specific 32P-labeled riboprobes obtained by transcription with T3 or T7 RNA polymerases of plasmid pBdCEVdP properly linearized. After washing the membranes, the signals of the dt RNAs and their resulting ml and mc forms were quantified with a bioimage analyzer (Fujifilm FLA-5100). Cleavage and ligation were estimated for each A. thaliana CEVd mutant from the fractions (mc+ml)/(dt+mc+ml) and mc/(mc+ml), respectively, and the results normalized with respect to those of the A. thaliana CEVd-wt (taken as 100%). Two independent plants were analyzed for each transgenic line, with differences in cleavage and ligation being less than 10% in all instances.
Primer extensions were carried out for 45 min at 55 °C, 10 min at 60 °C, and 5 min at 65 °C, in 20 μl containing 50 mM Tris-HCl [pH 8.3], 75 mM KCl, 3 mM MgCl2, 5 mM dithiothreitol, 0.5 mM each of the dNTPs, 40 U of RNase inhibitor (Promega), and 200 U of SuperScript III reverse transcriptase (Invitrogen). The ml and mc viroid RNAs serving as template were obtained by double PAGE and elution. Primers PI (5′-TTCTCCGCTGGACGCCAGTGATCCGC-3′), PII (5′-GCTTCAGCGACGATCGGATGTGGAGCC-3′), PIII (5′- GAGCAGGGGTGCCACCGGTCGC-3′), and PIV (5′-GACTAGCGGCGCGAAGAGTAGGTGG-3′), were 5′-labeled with T4 polynucleotide kinase (Roche Applied Science) and [γ-32P]ATP (Amersham Biosciences). Before reverse transcription, each primer was annealed in water to the purified viroid RNA (10:1 molar ratio) by heating at 95 °C for 5 min and snap-cooling on ice. Reactions were stopped at 70 °C for 15 min, and the products analyzed by PAGE on 6% sequencing gels. The exact size of the extension products was determined by running in parallel sequence ladders obtained with the corresponding primer and a recombinant plasmid containing the monomeric viroid-cDNA insert (Thermo Sequenase cycle sequencing kit, USB). |
10.1371/journal.pgen.1005951 | The Genetic Architecture of Natural Variation in Recombination Rate in Drosophila melanogaster | Meiotic recombination ensures proper chromosome segregation in many sexually reproducing organisms. Despite this crucial function, rates of recombination are highly variable within and between taxa, and the genetic basis of this variation remains poorly understood. Here, we exploit natural variation in the inbred, sequenced lines of the Drosophila melanogaster Genetic Reference Panel (DGRP) to map genetic variants affecting recombination rate. We used a two-step crossing scheme and visible markers to measure rates of recombination in a 33 cM interval on the X chromosome and in a 20.4 cM interval on chromosome 3R for 205 DGRP lines. Though we cannot exclude that some biases exist due to viability effects associated with the visible markers used in this study, we find ~2-fold variation in recombination rate among lines. Interestingly, we further find that recombination rates are uncorrelated between the two chromosomal intervals. We performed a genome-wide association study to identify genetic variants associated with recombination rate in each of the two intervals surveyed. We refined our list of candidate variants and genes associated with recombination rate variation and selected twenty genes for functional assessment. We present strong evidence that five genes are likely to contribute to natural variation in recombination rate in D. melanogaster; these genes lie outside the canonical meiotic recombination pathway. We also find a weak effect of Wolbachia infection on recombination rate and we confirm the interchromosomal effect. Our results highlight the magnitude of population variation in recombination rate present in D. melanogaster and implicate new genetic factors mediating natural variation in this quantitative trait.
| During meiosis, homologous chromosomes exchange genetic material through recombination. In most sexually reproducing species, recombination is necessary for chromosomes to properly segregate. Recombination defects can generate gametes with an incorrect number of chromosomes, which is devastating for organismal fitness. Despite the central role of recombination for chromosome segregation, recombination is highly variable process both within and between species. Though it is clear that this variation is due at least in part to genetics, the specific genes contributing to variation in recombination within and between species remain largely unknown. This is particularly true in the model organism, Drosophila melanogaster. Here, we use the D. melanogaster Genetic Reference Panel to determine the scale of population-level variation in recombination rate and to identify genes significantly associated with this variation. We estimated rates of recombination on two different chromosomes in 205 strains of D. melanogaster. We also used genome-wide association mapping to identify genetic factors associated with recombination rate variation. We find that recombination rate on the two chromosomes are independent traits. We further find that population-level variation in recombination is mediated by many loci of small effect, and that the genes contributing to variation in recombination rate are outside of the well-characterized meiotic recombination pathway.
| Meiotic recombination, the reciprocal exchange of genetic information between homologous chromosomes during meiosis, is necessary for proper chromosome segregation in many organisms [1]. Interestingly, the distribution of meiotic recombination events, or crossovers, varies dramatically in almost all taxa studied to date [2–12]. In addition, crossover frequency varies within and between species and populations in a huge diversity of organisms including humans, chimpanzees, flies, mice, worms, yeast, and many others [3,4,6,8,12–26].
In addition to its role in preserving genomic integrity between generations, recombination is a pivotal force in evolution. Recombination can reduce interference between a genetic variant and the genetic background in which it resides, thereby increasing the efficacy of natural selection [27–29]. Moreover, the exchange of genetic material between homologs creates new allelic combinations and thus contributes to the raw material for the process of evolution. Further highlighting its importance for evolution in general and genome evolution in particular, rates of recombination correlate with numerous genomic features such as the level of DNA polymorphism [30–32], rates of protein evolution [33,34], density of transposable elements [35–38], density of satellite DNA [39,40], and codon bias [41,42].
Given the importance of recombination and the pervasive natural variation in recombination rate, it is perhaps unsurprising that the genetic basis of this variation has been an active area of research for the last decade. With respect to the genetic basis of the distribution of crossover events, the first known determinant of recombination distribution in metazoans was discovered recently [43–45]. This remarkable discovery implicates PRDM9 in determining the locations of meiotic recombination hotspots in both humans and mice. Sequence variation within Prdm9 also modulates hotspot activity in humans [46]. PRDM9 is a histone methyltransferase that catalyzes histone H3 lysine 4 trimethylation [47]. This rapidly evolving protein [48] was first associated with hybrid sterility in rodents [49], and evidence continues to accumulate that it is a major component of recombination hotspot determination in mammalian systems [46,50–55].
Comparatively less is known in other systems such as Drosophila. Several studies have identified sequence motifs associated with recombination events [7,11,12,56–59], but none have been functionally validated to date. Drosophila lacks PRDM9 [48,58], and perhaps relatedly, also lacks the highly punctate recombination landscape seen in mammals. While in humans up to 80% of recombination events fall in 10–20% of sequence [6], crossover distribution in Drosophila is far less heterogeneous [12,60].
Recent work in mammals has also provided insight into the genetic architecture of global recombination rate. RNF212 has been repeatedly associated with natural variation in recombination rate in several systems including humans [61,62], cattle [63], and Soay sheep [64]. Consistent with a role of this protein in modulating recombination rate, RNF212 is essential for meiotic recombination and has a key role in stabilizing meiosis-specific recombination factors in mice [65]. PRDM9 has also been associated with heritable variation in recombination rate in humans and mice [52,66]. Other mediators of recombination rate include REC8 [63], which is a cohesin that is required for proper chromosome segregation in many organisms [67–69]. In humans, inversion 17q21.31, a 900 kb inversion, is associated with increased recombination and reproductive output in European females [70].
The genetic architecture of recombination rate variation outside of mammals remains poorly understood, even in the model organism Drosophila melanogaster. However, it is well-documented that recombination rate is a variable and heritable trait in Drosophila. For instance, classical genetic experiments indicate that the amount of crossing-over as well as the distribution of crossover events can vary among lines of D. melanogaster [12,13,71,72], suggesting population-level variation in this trait. Additionally, genetic control for crossover rate has been suggested by laboratory selection experiments in which recombination rate itself was successfully subject to artificial selection [73–85]. Finally, changes in recombination rate have been shown to evolve as a correlated response to artificial selection on other characteristics, such as sternopleural bristle number [86], DDT resistance [87], geotaxis [88], and resistance to temperature fluctuations [89], which is again consistent with segregating natural variation in recombination rate. Additionally, the observation that modifiers of recombination rate are commonly associated with variants controlling completely unrelated traits suggests that these modifiers are pervasive in the genome and/or may have pleiotropic effects.
To gain the first insight into the genetic basis of population-level variation in recombination rate in D. melanogaster, we used an association mapping approach. We favored an unbiased approach in part because D. melanogaster lacks homologs of the three known determinants of recombination rate in mammals noted above: RNF212, REC8, and PRDM9. We measured recombination rates on both the 3R and X chromosomes in the 205 fully-sequenced inbred lines of the Drosophila melanogaster Genetic Reference Panel (DGRP) [90,91] using a two-step crossing scheme. We find nearly 2-fold variation in recombination rate among lines with a standard karyotype. Unexpectedly, we find that recombination rates are uncorrelated between the X and 3rd chromosomes. We leveraged this pervasive population-level variation in recombination rate for genome-wide association (GWA) mapping to identify dominant or semi-dominant variants associated with phenotypic variation in recombination rate on each chromosome. We selected the top 20 most promising candidate genes associated with recombination rate and subjected these candidates to both gene-level and allele-level functional assessment. Our functional assays implicate five highly promising candidates as novel mediators of recombination rate variation in D. melanogaster: CG10864, CG33970, Eip75B, lola, and Ptp61F. Our results provide new insight into the scale and scope of population level variation in rates of recombination and more importantly implicate new determinants of natural variation in recombination rate in Drosophila.
To assay recombination rate variation in the DGRP, we used a classic two-step crossing scheme (Fig 1). We measured recombination rates in two different genomic intervals: the 20.4 cM interval between ebony (e) and rough (ro) on chromosome 3R and the 33 cM interval between yellow (y) and vermilion (v) on the X chromosome. In total, 506,045 progeny were scored for recombinant phenotypes (217,525 for the e ro interval and 288,520 for the y v interval). On average, each replicate (there were three replicates per DGRP line per chromosome assay) contained ~368 progeny (for the e ro interval) and ~499 progeny (for the y v interval). We first verified that our data conformed to expectations under Mendelian inheritance. Deviations from these expectations would be consistent with viability defects associated with the visible markers used in this study. To do so, for each line we compared the number of wild-type progeny to the number of progeny possessing both markers (either e ro or y v), summing across all three replicates (S1 Table). We also compared the number of recombinant progeny possessing only one marker to the number of recombinant progeny containing only the other marker (either e + versus + ro or y + versus + v) (S1 Table). The null expectation is a 1:1 ratio for the aforementioned pairs of phenotype classes. We used a Bonferroni correction [92,93] with α = 0.05 to correct for multiple tests. When comparing the ratios of the two non-recombinant haplotypes, we find 15 lines that deviate from the expected 1:1 wild-type: e ro ratio (Bonferroni-corrected P < 0.03, all comparisons, G-test) and 8 lines that deviate from the expected 1:1 wild-type: y v ratio (Bonferroni-corrected P < 0.03, all comparisons, G-test). In all but one case, the deviation is in the direction of a relative excess of wild type flies. Only one line deviated significantly in both intervals (DGRP_819), with more wild-type progeny in both intervals. When comparing the ratios of the two recombinant haplotypes, we find that DGRP_31 deviates significantly from the expected 1:1 e +/ + ro ratio (Bonferroni-corrected P < 0.0001, G-test) and that DGRP_819 deviates significantly from the expected 1:1 y +/ + v ratio (Bonferroni-corrected P < 0.0001, G-test).
Similarly, we tested for sex ratio unity by comparing the numbers of female and male progeny. There are no deviations from the expected 1:1 male:female ratio in the 205 lines for the e ro interval (Bonferroni-corrected P > 0.10, all comparisons, G-test). For the y v interval, only two lines significantly deviate from expectation (DGRP_41 AND DGRP_801) (Bonferroni-corrected P < 0.0002, both comparisons, G-test), both in the direction of a relative excess of females.
To assess the consequences of possible viability defects associated with our visible markers on recombination rate estimation, we analyzed correlations between viability defects and recombination. That is, to address whether epistatic interactions between our visible markers and DGRP genotype yield viability defects, we analyzed whether the ratios of the number of males vs. females, + + individuals versus m1 m2 individuals, or m1 + individuals versus + m2 individuals are correlated with our estimates of recombination within the DGRP (S2 Table). Again, each of these ratios should be 1, but could be skewed by viability defects associated with the markers. Our analysis demonstrates that in the y v interval, none of these ratios are correlated with our estimates of recombination rate. For the e ro interval, we observe a weak but statistically significant correlation between the ratio of wild type progeny to e ro progeny and recombination rate. However, no significant correlation is seen between the sex ratio and recombination rate or the ratio of the two classes of recombinants and recombination rate for the e ro interval. These data are consistent with weak epistatic interactions between the e ro genetic background and wild-type genetic backgrounds that yield viability defects.
Overall, however, our data indicate that our assays for measuring recombination rate largely conform to expectations given Mendelian inheritance. There does not appear to be a large systematic bias towards wild-type chromosomes, indicating that there are no major viability defects associated with any of these mutations alone or in the pairs in which they were used for the current experiment. This confirms previous descriptions of these mutants and their lack of viability defects [94,95]. Our analysis does indicate weak viability effects of the e ro background as revealed by epistatic interactions with wild-type genetic backgrounds. As a consequence, the scale and scope of the reported variation in recombination rate may be mis-estimated. Given how weak the viability defects appear to be, we believe any mis-estimation is likely to be small in magnitude.
Following the crossing scheme detailed in the Materials and Methods and in Fig 1, we estimated crossover rate for each DGRP line in the e ro and the y v intervals (S2 Table; S1A and S1B Fig) for three replicates. These replicates are largely consistent with one another (S3 Table; S2A–S2L Fig). Analyzing only lines with a standard karyotype on all chromosomes (n = 112), the average crossover rate for e ro is 20.9 ± 0.2 cM (ranging from 14.2 cM to 26.12 cM) (Fig 2A). This agrees well with the published map distance of 20.4 cM [95]. Among these lines, we observe 1.84-fold variation in mean crossover rate. Analyzing only lines with a standard karyotype on all chromosomes, the average crossover rate for y v is 31.2 ± 0.3 cM (ranging from 23.6 cM to 39.30 cM) (Fig 2B), compared with the published map distance of 33 cM [94]. Similar to the magnitude of population-level variation in recombination rate on 3R, here we observe 1.67-fold variation among these lines in mean crossover rate for the y v interval.
There is significant genetic variation for crossover rate among lines for both intervals (Fe ro = 1.34, Pe ro = 0.038 and Fy v = 3.00, Py v < 0.0001, ANOVA). Using only lines with a standard karyotype (112 lines), we estimated broad-sense heritability (H2) of recombination rate for the e ro interval as 0.12 and for y v interval as 0.41 (Table 1). These results confirm that recombination rate is a heritable trait and has a genetic component. Interestingly, there is no significant correlation between recombination rates in these two intervals (Spearman’s ρ = 0.09, P = 0.36; Fig 2C). Consistent with this, a model fitting effects of line, genomic interval, and line-by-interval interaction effects reveals significant interaction effects (P < 0.0001, ANOVA, S4 Table), indicating that the magnitude of the difference in recombination frequency between the two loci surveyed varies significantly among lines. These analyses illustrate that recombination rate on chromosome 3R and chromosome X, at least in the way they have been assayed here, are independent traits in this panel of flies.
As a widely-used community resource, the DGRP offers a unique opportunity to examine the relationship between recombination rate and other phenotypes because a variety of phenotypes have been surveyed in this panel. We tested whether crossover rates in the e ro or y v interval (of lines with standard karyotypes) were correlated with various traits including organismal fitness. While the majority of correlations were weak and not statistically significant, we elaborate on several interesting significant correlations (S23 Table) in S1 Text.
Recombination is suppressed within inverted regions, and recombination elsewhere in the genome increases through what is known as the interchromosomal effect [96,97]. A large number of the DGRP lines are either homozygous or polymorphic for a chromosomal inversion. To test for the interchromosomal effect, we separated lines with inversions from lines with standard karyotypes and tested whether lines that possessed an inversion somewhere in the genome had higher rates of recombination in our surveyed intervals. Lines with inversions have significantly increased rates of recombination in the y v interval relative to lines with standard karyotypes (35.1 cM vs. 31.0 cM, P < 0.0001, t-test). This trend is echoed in the e ro interval (20.9 cM vs. 20.7 cM) but the difference in recombination frequency between standard and inverted karyotypes is not statistically significant (P = 0.66, t-test). These results are discussed in the context of previous work in S1 Text.
In the DGRP, 108 lines are infected with Wolbachia pipientis [91]. To test for an effect of Wolbachia infection on recombination frequency, we used a linear model (see Materials and Methods) and fit effects of line and Wolbachia infection status for each interval surveyed. Analyzing only lines with standard karyotype, we find there is a significant effect of Wolbachia infection in the y v interval (P = 0.0003, ANOVA), such that Wolbachia-infected lines have a higher crossover frequency (31.8 cM) than uninfected lines (30.0 cM). No effect of Wolbachia infection was found for the e ro interval (P = 0.35, ANOVA). Importantly, estimates of heritability are not driven by Wolbachia infection in either interval (S5 Table).
The continuous variation for recombination rate among lines described above (Fig 2A and 2B) suggests that the genetic architecture of this trait is likely complex and regulated by many independent genetic factors. The observed variation in recombination rate in the DGRP motivates our association mapping approach to more finely define the genetic basis of this trait. To identify genetic variants contributing to variation in recombination rate, we performed genome wide association mapping on the mean crossover rates from the DGRP in the e ro and y v intervals. Note that given the experimental design of our study (Fig 1), we are only able to identify variants that are at least partially dominant in their effects on recombination frequency. Recessive modifiers are not captured in this study, likely yielding underestimates of the scope of natural variation in recombination rate in this system. We did the association mapping in three different ways for each interval because of the inversions segregating in the DGRP and the known effect of inversions on recombination frequency (see [97] for review). Of the inversions segregating in the DGRP, none are on the X chromosome. However, 49 lines contained at least one copy of the C, K, Mo or P inversion on chromosome arm 3R; all four of these inversions span at least part of the e ro interval used to assay recombination rate [99]. We thus exclude these lines when analyzing recombination rate data for the 3R interval. The three datasets used for the 3R analyses were: 1) lines with no inversion on 3R (n = 156), 2) lines with neither 3R inversions nor inversion polymorphisms elsewhere in the genome (n = 130), and 3) lines with the standard karyotype (lines lacking inversions; n = 112). The three datasets used for the X chromosome analyses were: 1) all lines (n = 205), 2) lines without inversion polymorphisms (n = 152) and 3) lines with a standard karyotype (n = 112).
The statistical model used to infer associations assesses and adjusts for significant associations of both Wolbachia status and inversions. For the e ro interval, there is a significant effect of the NS inversion (P = 0.003, ANOVA; Table 2) on crossover rate in the restricted data set that removes lines with inversions on 3R and lines with inversion polymorphisms. For the y v interval, Wolbachia infection is significantly associated with crossover rate in all three of our data sets (P < 0.01, all cases, ANOVA; Table 2). Additionally, inversions t, NS, K, and Mo are all significantly associated with crossover rate in the y v interval (P < 0.05, all cases, ANOVA Table 2). These data are summarized in Table 2.
The full results for all six GWA analyses are presented as supplementary tables (S6–S11 Tables). To generate a list of candidate genes and alleles, we combined the results from the different GWAS for each chromosome interval, using a significance threshold of P < 10−5. For a Venn diagram displaying overlap among the different data sets, see S3 Fig. We tested whether the distribution of these associated variants was significantly different from the null expectation of a uniform distribution across chromosomes (as a function of the number of polymorphisms on each chromosome). Using lines with standard karyotypes, we find that the distribution of associated variants is significantly different from the distribution of variants in the genome for both intervals (P < 0.02, both comparisons, G-tests). It appears that in both intervals, there is an enrichment of associated variants on chromosome 2R (e ro: 63 versus 33; y v: 29 versus 16; observed versus expected).
For the e ro interval, the three GWAS yielded a combined total of 688 unique variants at a nominal significance threshold of P < 10−5. For the y v interval, combining results from all three GWA analyses, we identified 160 unique variants at a nominal significance threshold of P < 10−5. A description of types and locations of these variants is included in S12 Table. There were no variants that overlapped between the two intervals, consistent with the lack of correlation between the two traits. However different variants in the same gene (see below) were shared between the associations found in the two intervals. Variants in 359 genes were implicated as potential candidates from the three e ro GWAS, and variants in 111 genes were associated with recombination rate variation in the y v GWAS. There is very little overlap between these gene lists; a total of fifteen genes showed overlapping (gene-level) associations between the e ro and y v GWAS (bab1, bun, CG4440, CG5953, CG31817, CG32521, CR44199, dnr1, dpr6, Eip63E, Eip75B, Ptp61F, Sec16, Shroom, and SNF4Agamma). The effect sizes for these variants were moderate, averaging ~2.32 cM for both intervals (S4A and S4B Fig). Fig 3A and 3B displays the Manhattan plots and linkage disequilibrium plots for both intervals for the lines with standard karyotypes while S5 and S6 Figs display the same information for the other data sets analyzed.
We sought to functionally assess a subset of the genes identified by our association mapping. We used several criteria to refine our list of candidate associations to a tractable set of 20 candidate genes. First, we restricted our focus to protein-coding genes harboring significantly associated genetic variants. We then integrated the P-value of the association, effect size, and the number of GWAS the gene was implicated in on either or both chromosomes to refine our list of putative candidates. We narrowed our list further by limiting ourselves to genes with documented expression in the ovaries [99–103]. Our final candidate gene list (Table 3) includes eleven genes from the e ro GWA, five genes from the y v GWA and four genes that were found in both. There was more than one significantly associated genetic variant in 8 of our 20 candidate genes (CG1273, CG4440, CG7196, dpr6, Eip75B, jing, Ptp61F and Ubx) with jing and Ptp61F having the most significantly associated variants (17 and 13 respectively). The full list of variants within these genes and associated P-values are listed in S13 Table and the genotypes of each DGRP line at these variants are listed in S14 Table.
If these identified candidate genes mediate recombination rate in some way, we expect that perturbing these genes will affect recombination rate. We used both mutant analysis and RNAi to either knock out or knock down expression of each of these genes, and compared recombination rate in the knock out/down lines to an appropriate genetic background control. We measured recombination rate in the e ro and y v intervals for available mutants and RNAi lines for all 20 candidate genes in the same way as described earlier. We used a combination of P-element insertions, chromosomal deletions, as well as any available RNAi lines. For the RNAi experiments, we used a nanos GAL4 driver, which should target the effects of knockdown to oogenesis. For assessment using the e ro markers, the only line tested that produced a significant difference from control line was a deletion line, Df(3R)ED2 (P = 0.004, Dunnett’s test) (Fig 4A; S15 Table); this line shows a significant increase in recombination frequency relative to the genetic background control. This deletion encompasses 71 full genes and part of 1 additional gene, including two of our candidate genes: cdi and CG10864. It should also be noted that this deletion is on chromosome 3R, spanning the cytological region 91A5 to 91F1 (for reference e is at 93C7-93D1 and ro is at 97D4-97D5). Using the y v markers, seven lines tested show a significant deviation in recombination frequency relative to the appropriate control (Fig 4B; S16 Table). These included alph, CG9650, CG33970, Eip75B, grp, lola, and Ptp61F (P < 0.05, all comparisons, Dunnett’s test). Eip7B, and CG9650 showed a decrease in recombination relative to the control while alph, CG33970, Eip75B, lola, and Ptp61F showed an increase in recombination relative to the control. Interestingly, one P-element insertion in grp showed a significant increase of recombination while a different P-element insertion in grp showed a significant decrease of recombination.
While the mutant/RNAi analysis provides insight into whether the candidate genes function in some way to mediate recombination, we also wanted to test whether these candidate genes show significant differences at the allelic level. We hypothesized that the effects of these genes on recombination rate were mediated by expression level differences and thus tested for differences in gene expression in ovaries between allelic variants of our 20 candidate genes. We measured gene expression as mRNA abundance using quantitative RT-PCR (qPCR). For each of our twenty candidate genes, we selected three DGRP lines containing the major allele and three lines containing the minor allele (S17 Table). For candidate genes that had that multiple significantly associated variants, all attempts were made to include lines in which all minor alleles were present. The genotypes of these lines at the gene surveyed can found in S18 Table. Once a line was selected to assess a candidate gene, it was not used to assess another candidate gene. RNA was extracted from dissected ovaries from virgin DGRP females. The qPCR data (normalized to GAPDH) reveal significant differential expression for 11 of our 20 candidate genes (Fig 5; S19 Table). DGRP lines with the major alleles of CG4440, CG15365, CG33970, and Ptp61F (P < 0.003, all comparisons, t-test) display higher expression levels than lines with the minor alleles. Conversely, DGRP lines with the major alleles of CG1273, CG10864, dpr6, Eip75B, lola, Oaz, and Ubx (P < 0.05, all comparisons, t-test) display lower expression levels than lines with the minor alleles. It should be noted for variants in these eleven candidate genes, all minor alleles are associated with reduced rates of recombination. Comparisons of un-normalized data (given potential concern over unstable housekeeping gene expression [104,105]) largely confirm these results (S20 Table; S7 Fig).
Here we report the largest population-level survey of recombination rate variation in Drosophila to date. We find significant genetic variation for recombination rate in this North American population of D. melanogaster for two independent genomic intervals. At the broadest scope, these data are consistent with previous work from other systems. Indeed, a wealth of data indicate that recombination rate varies between and within populations in species such as Drosophila [12,13], mice [23], and humans [5,20,51,106].
The magnitude of population-level variation in recombination rate exposed by our survey is comparable to what has been previously shown in D. melanogaster. For instance, we observe 1.67 fold-variation for the y v interval, and previous work in this interval shows ~1.2-fold variation [13,72]. Other genomic regions in Drosophila consistently show 1–2 fold variation in crossover frequency among strains [13]. Although measured with a different approach, work from heterogeneous stock mice indicates that crossover frequency varies ~2-fold in both males and females [23]. Work from cattle indicates that average genome-wide recombination rate varies ~1.7 fold in males [63], which aligns well with our survey. Similarly, humans show ~2-fold variation in crossover frequency in both males and females [62,107].
It should be noted that the ~2 fold variation in recombination frequency that we report above is biased downward and is not truly reflective of segregating natural variation in recombination rate in Drosophila. When we include lines with inversions, which clearly segregate in natural populations, we see a much greater span in recombination rates in the DGRP: 5.2-fold for the e ro interval (excepting lines with an inversion on 3R) and 3.5-fold for the y v interval. This range of variation in recombination frequency is remarkable, nearly doubling previous estimates from Drosophila, mouse and humans. However, it also bears mentioning that we cannot exclude that our estimates of recombination may be biased by the weak viability effects associated with our visible markers (see above).
Our results indicate that recombination rate at the two intervals surveyed are uncorrelated in the DGRP. It is certainly possible that the weaker genetic component of phenotypic variation in recombination rate in the e ro interval as compared to the genetic component of phenotypic variation in recombination rate in the y v interval is driving the lack of correlation between recombination rates in the two intervals. In contrast to what we observe here, previous work in humans showed a significant positive correlation between the number of maternal recombination events on individual chromosomes and the number of maternal recombination events in the remaining genome complement for 20 out of 23 chromosomes, as well as a strong, significant correlation for the first eight chromosomes compared to chromosomes nine through twenty-two and the X chromosome [3]. Other work in Drosophila is suggestive that two lines with less crossing over in one interval relative to four other lines generally had less crossing over in other intervals relative to the same four lines [13], though this is anecdotal at best. The putative difference between Drosophila and humans with regard to correlations in recombination rates across chromosomes is interesting, and may point to different genetic architectures of this trait in these systems. Certainly, the molecular mechanics of meiotic recombination have diverged markedly between humans and Drosophila (e.g. [108]) and the recombinational landscapes in humans and flies are qualitatively different as well.
Previous work has estimated heritability for recombination rate in many different species. While estimates of heritability are necessarily population-specific, mammalian estimates encompass a wide range, from as small as 0.14 [109] and 0.30 [110] in humans to as large as 0.46 in mice [23]. In maize, heritability of recombination frequency is considerably higher (broad sense heritability 0.21–0.69; [111]). Insects show a wide range as well, with estimates of narrow sense heritability of recombination rate ranging from 0.16 in Tribolium [112] to 0.27–0.49 in grasshoppers [113]. Early estimates of narrow sense heritability of recombination frequency in Drosophila based on parent-offspring regression are comparable to ours (0.09–0.38; [114]), and selection based approaches yield a narrow sense heritability of 0.12 [79]. That estimates of heritability of recombination are low indicates that much of the observed variation in recombination frequency cannot be ascribed to genetic differences along lines. This is consistent with the remarkable phenotypic plasticity in recombination frequency in Drosophila, evidenced in response to temperature [115–122], maternal age [72,115–117,123–133], nutrition [126,127], parasite pressure [134] and other environmental factors. This phenotypic plasticity could also drive the lower than expected correlations between replicates observed in this experiment (see S3 Table) and also reduce heritability.
Wolbachia pipientis is a common endosymbiont that infects the reproductive tissues of many arthropods [135]. Evidence indicates that over 40% of arthropods are infected with W. pipientis [136–138]. Approximately 29% of Drosophila stocks from Bloomington Drosophila Stock Center [139] are infected, along with 76% of the Drosophila Population Genomics Project (n = 116) [140]. In the DGRP, 108 of 205 (53%) lines are infected with W. pipientis [91]. In Drosophila, there is clear infection in the ovaries [141,142] and infection has been shown to reduce egg production [143].
Interestingly, we see a significant association between Wolbachia infection and crossover rates in the y v interval but not in the e ro interval. This discrepancy between the two intervals surveyed is difficult to explain, and merits further investigation. More curious yet is the contrast with previous results. It has been shown that Wolbachia infection has no effect on rates of crossing over in the w ct interval (18.5 cM) in the laboratory wild-type strain Canton S [144]. The w ct interval is actually within the y v interval surveyed in this study, so the discrepancy between the two studies is puzzling. It may be that the effect of Wolbachia infection on recombination frequency is sufficiently minor that the previous study, using a single genetic background and smaller sample sizes than the present study, was underpowered to detect this small effect (an average increase of 1.8 cM associated with Wolbachia infection in our study). Our results, coupled with previous findings, suggest that W. pipientis might have differential effects on recombination frequencies in different parts of the genome. Testing explicitly for this heterogeneity will be a topic of future exploration. In the future, it will also be interesting to see if infecting DGRP lines with Wolbachia causes an increase of crossover rates and if curing DGRP lines via tetracycline yields a corresponding decrease in crossover rates.
The DGRP allows us to couple phenotypic variation with genetic variation such that the genetic basis of complex traits of interest can be dissected. One benefit of this association mapping approach is that it is unbiased, which means that new genes, outside of known pathways playing a role in the phenotype of interest, can be identified. For example, a recent study using the DGRP dissecting the genetic architecture of abdominal pigmentation yielded associations with several variants in the known pigmentation pathway but importantly, also functionally validated seventeen out of twenty-eight candidate genes that had not been previously associated with pigmentation [145]. Because nothing was known regarding the genetic basis of population-level variation in recombination rate in Drosophila and because Drosophila lacks homologs of all genes associated with recombination rate variation in other systems, we were eager to leverage this unbiased approach to gain novel insight into the genetic architecture of this fundamentally important trait.
Consistent with the power of GWAS to uncover novel genes associated with phenotypic variation, our top candidate genes significantly associated with recombination rate variation contain genes outside of the meiotic recombination pathways, which have been characterized in exquisite detail (see [146] for review). Among the top 20 candidates for functional assessment, seven are computationally predicted genes that have no clearly defined biological function or human orthologs. Interestingly, four of our candidate genes have Cys2His2 zinc fingers (CG9650, jing, lola, and Oaz). This is particularly intriguing due to the link between the zinc-finger domain containing PRDM9 and hotspot determination, and it is tempting to speculate that these proteins bind to DNA and designate crossover sites in a way that is vaguely reminiscent of the role of PRDM9 in mammalian recombination [43–45]. Moreover, the D. pseudoobscura ortholog of Oaz, GA14502, was previously identified as a possible candidate gene involved in recombination as the frequency of its zinc finger binding motif was significantly negatively associated with recombination on a broad scale [58]. Consistent with a role for zinc-finger DNA binding in Drosophila recombination, Trem, which also contains zinc fingers, was recently shown to be necessary along with Mei-W68 and Mei-P22 for the formation of double-strand breaks in Drosophila [147].
We chose two methods for functional assessment of our candidate genes. The first method is a gene-level approach and asks whether perturbation of candidate genes perturbs recombination frequencies. To complement this approach, we also compared expression levels of the different alleles in these candidate genes using qPCR. Significant differential expression of the major versus minor alleles of our candidate genes in the ovaries would be consistent with gene expression differences underlying differences in rates of crossing over.
Overall, there were 5 genes (bru-2, CG4440, jing, MESR3, and pk) which showed neither a change in recombination frequency in the e ro or y v intervals when perturbed nor a difference in expression level between the major and minor allelic variants. However, lack of functional confirmation does not imply that a candidate gene has no role in modulating recombination rate in Drosophila. Indeed, validation of candidate genes is challenging. The effect sizes of the genetic variants are moderate at best (S4A and S4B Fig), making detection of these changes quite difficult in the absence of very large sample sizes. Additionally, recombination rate variation is likely to be a polygenic trait [77,78], and our results confirm this. Further, it has been suggested that in many quantitative traits within the DGRP, there is pervasive epistasis [148,149]. Epistatic interactions may similarly contribute to recombination rate variation in Drosophila. Consistent with this is the observation that for one P-element insertion of grp, there is an increase in recombination relative to the appropriate background and a decrease in recombination rate for another P-element insertion (though we note that this observation is also consistent with variation in allelic effects at a single locus if the two P-elements were inserted into different locations). Finally, the process of recombination is likely to be highly buffered, and one could hypothesize that there is redundancy for maintaining the number of crossovers required. It is also possible that these statistical associations are false positives due to our lenient P-value.
However, integrating across both the gene- and allele-level functional analysis, we find five high quality candidate genes for further investigation. These genes show significant perturbations in recombination frequency relative to the appropriate genetic background control in addition to differential expression specifically in ovaries between allelic variants at these loci. These were CG10864, CG33970, Eip75B, lola, and Ptp61F. Two of these (Eip75B and Ptp61F) were identified in GWAS in both the e ro and y v interval.
CG10864 is involved in potassium channel function [150]. In humans, another protein involved in potassium channel function, KCNQ1, has been shown to somatically imprint regions of the genome with higher rates of recombination [151]. While imprinting appears to be less common in Drosophila females [152], it is unclear if CG10864 is participating in a similar role as compared to KCNQ1.
CG33970 is predicted to be involved with ATP binding and transporter activity [98]. A direct link between ATP binding and meiotic recombination has yet to be shown, but there have been some hints of connections in the literature. For example, mutations in the ATP-binding domain of RecA [153] in Escherichia coli, DMC1 [154], Rad51 and Rad55 in yeast [155,156] and XRCC3 in humans [157] cause defects in homologous recombination and meiosis. While speculative, this gives credence to the idea that the putative ATP-binding ability of CG33970 may contribute to meiotic recombination. Further work is aimed at dissecting this link.
Eip75B (Ecdysone-induced protein 75B) is involved in mediating ecdysone signaling, a steroid hormone. Defective ecdysone signaling affects the early germarium, causing defects with meiotic entry [158]. Interestingly, ecdysone signaling is important for female fertility but not for male fertility [159–161]. Drosophila males do not undergo meiotic recombination [162,163]. It remains to be seen whether the connection between recombination, fertility and ecdysone signaling is merely coincidence; however, the role of Eip75B in oogenesis makes it a particularly exciting candidate for further work.
lola, or longitudinals lacking, is BTB zinc finger-containing transcription factor that is required for axon growth and guidance [164,165]. As noted above, DNA binding ability along with zinc fingers is exciting as a possible link with recombination. The predicted human ortholog, ZBTB46 or BZEL, was shown to repress a desumoylase [166]. Sumoylation has been linked to DNA repair [167] and therefore it is possible that lola is involved in early processes that could ultimately lead to crossover formation.
Ptp61F (Protein tyrosine phosphatase 61F) is a member of the protein tyrosine phosphatase family. Ptp61F is an induced antagonist of the JAK/STAT pathway [168,169] and has been directly implicated in oogenesis [170]. In the female germline, expression of Ptp61F is targeted to the nucleus and cytoplasmic organelles [171] and this gene is required for normal female fecundity [172]. Tentative links between Ptp61F and DNA damage can be made in mammals; Ptp61F is the Drosophila homolog of human PTP1B and knockout PTP1B mice show a higher sensitivity to irradiation and an upregulation of many genes in the DNA excision/repair pathway [173]. Homologous recombination, base excision repair, and nucleotide excision repair are the primary pathways by with DNA damage are repaired in Drosophila. While the role for Ptp61F in meiotic recombination is not obvious, the clear function of this gene in oogenesis coupled with its tentative connection to DNA damage repair is promising.
In conclusion, we have quantified the extent of recombination rate variation in a natural population of D. melanogaster and have shown that genetic background significantly drives phenotypic variation in this critically important phenotype. The magnitude of observed phenotypic variation in recombination rate is large, with almost 2-fold variation present in each genomic interval analyzed. We demonstrate that inversions play a large role in mediating rates of recombination, indicative of the interchromosomal effect, and provide the first evidence that Wolbachia infection can significantly increase rates of recombination. Through our GWA approach, we show that recombination rate is a highly polygenic trait, with many genetic factors of small effect associating with phenotypic variation. We show that a subset of our candidate genes (CG10864, CG33970, Eip75B, lola, and Ptp61F) play putative roles in modulating recombination rate variation in Drosophila through both gene-level and expression-level functional assessment. Future work will be aimed at determining the role of these candidate genes in the molecular process of recombination.
The Drosophila Genetic Reference Panel is a collection of 205 fully-sequenced inbred lines [90,91]. Mated, gravid Drosophila melanogaster females were originally collected in Raleigh, NC, USA in 2003. Their progeny were subjected to 20 generations of full-sibling matings. The resulting inbred lines were then fully sequenced. A total of 4,853,802 single nucleotide polymorphisms (SNPs) and 1,296,080 non-SNP variants were identified among these lines [91].
To assay recombination rate, we took advantage of visible, recessive markers in D. melanogaster. To measure recombination rates on the 3R chromosome, we used a strain marked with ebony (e4) and rough (ro1); these markers are 20.4 cM apart [95]. To measure recombination on the X chromosome, we used a strain marked with yellow (y1) and vermillion (v1); these markers are 33 cM apart [94]. These markers were chosen to examine due to the genetic distance between them, ease of scoring and also their apparent lack of viability defects [94,95]. Each of the doubly marked chromosomes was substituted into a wild-type isogenic Samarkand genetic background, free of P-elements [174], to allow for continuity between assays and to minimize marker genetic background effects.
To assay recombination rate variation in the DGRP, we used a classic two-step crossing scheme (Fig 1). All crosses were executed at 25°C with a 12:12 hour light:dark cycle on standard media using virgin females aged roughly 24 hours. We conducted three replicate assays for each interval (either e ro or y v). For each replicate, all 205 lines were crossed simultaneously to avoid conflating block effects with variation among lines. This yielded three replicate estimates of recombination frequency per line per interval. For the first cross, ten virgin females from every DGRP line were crossed to ten doubly-marked males (either e ro or y v) in eight ounce bottles. Males and females were allowed to mate for five days, after which all adults were cleared from the bottles. F1 females resulting from this cross are doubly heterozygous; these females are the individuals in which recombination is occurring. To uncover these recombination events we backcross F1 females to doubly-marked males. For this second cross, twenty heterozygous virgin females were collected and backcrossed to twenty doubly-marked males. Males and females were allowed to mate for five days, after which all adults were cleared from the bottles. After eighteen days, BC1 progeny were collected, frozen, and scored for sex and for visible phenotypes. Previous work in our lab has demonstrated that freezing flies has no effect on the visible markers we scored. Recombinant progeny were then identified as having only one visible marker (m1 + or + m2). For each replicate, recombination rates were estimated by taking the ratio of recombinant progeny to the total number of progeny. Double crossovers cannot be recovered with this assay, so our estimates of recombination frequency are likely to be biased downwards slightly. The estimated recombination for a given strain and interval was calculated as the average across the three replicates.
Freeze 2 of the DGRP contains information relating to 16 segregating autosomal inversions verified by cytological methods [91]. We therefore performed association mapping in three different ways for each interval. The X chromosome (in this population of flies) lacks inversions while 49 lines contain an inversion on chromosome arm 3R which spans at least part of the e ro interval used to assay recombination rate [98]. We thus completely exclude these lines when analyzing recombination rate data for the 3R interval. The three datasets used for the 3R analyses were: 1) lines with no inversion on 3R (n = 156), 2) lines with neither 3R inversions nor inversion polymorphisms elsewhere in the genome (n = 130), and 3) lines with the standard karyotype (n = 112). The three datasets used for the X chromosome analyses were: 1) all lines (n = 205), 2) lines without inversion polymorphisms (n = 152) and 3) lines with a standard karyotype (n = 112).
To estimate the broad-sense heritability (H2) of recombination rate, we used an ANOVA framework on line means (the average across the three replicates for each line for each interval). The ANOVA followed the form of Y = μ + L + ϵ for each chromosome assayed where Y is recombination rate, μ is the overall mean, L is the random effect of line and ϵ is the residual. Additionally, we ran a similar ANOVA, adding the genomic region as a fixed factor, to test for a significant interaction between line and genomic region. That ANOVA followed the form of Y = μ + L + R + L × R + ϵ, with the terms the same as above and R is the genomic region assayed. To estimate H2, we follow the formula H2 = σ2L / (σ2L + σ2ϵ) where σ2L is the variance component among lines and σ2ϵ is the residual variance or variance component attributed to error. The variance components were calculated using REML. All H2 estimates were calculated using R Statistical Software, v3.2.1 and RStudio v0.99.467.
To test for a significant effect of Wolbachia infection, we used an ANOVA framework as well. The ANOVA follows the form Y = μ + W + ϵ for each chromosome assayed where Y is recombination rate (measured in cM), μ is the overall mean, W is fixed effect of Wolbachia infection status and ϵ is the residual, including all individual measurements.
To identify genetic variants that are associated with differences in mean crossover number in two different intervals of the Drosophila genome, we performed a GWAS using the established web-based pipeline developed by the Mackay lab at NC State University, Raleigh, NC (http://dgrp2.gnets.ncsu.edu/) [90,91]. The first step in conducting the GWAS was to adjust line means for the effects of Wolbachia pipientis infection as well as the presence of inversions that are segregating in the DGRP. The adjusted line means are then used to fit a linear mixed model, Y = Xb + Zu + e. Y is the adjusted phenotypic value, X is the design matrix for the fixed variant effect b, Z is the incidence matrix for the random polygenic effect u and e is the residual. The vector of polygenic effects u has a covariance matrix in the form of Aσ2, where σ2 is the polygenic variance component and A is the genomic relatedness. Additionally, Manhattan plots were constructed using the qqman package in R [175].
As described in the text, we selected 20 candidate genes to functionally assess that contained at least one significantly associated genetic variant within them. We selected these genes based on P-value of the variant located within or near the gene, effect size of the variant, the number of GWAS that a variant within or near the gene was implicated in and available expression data. To functionally explore these candidate genes with respect to their roles in recombination, we took advantage of available P-element insertion lines and chromosomal deletions as well as RNAi lines (S21 Table). Lines containing a P-element insertion or chromosomal deletion (deleting the candidate gene) as well as appropriate controls (genetic background used to generate P-element insertion or chromosomal deletion) were used in the same crossing scheme (Fig 1) detailed above. For the first cross, ten virgin females from every line containing a P-element insertion or chromosomal deletion were crossed to ten doubly-marked males (either e ro or y v) in eight oz. bottles. Males and females were allowed to mate for five days, after which all adults were cleared from the bottles. For the second cross, ten virgin heterozygous females were collected and backcrossed to ten doubly-marked males in vials. Males and females were allowed to mate for five days, after which all adults were cleared from vials. BC1 progeny were collected from each vial, frozen, and scored for sex and for visible phenotypes. For each P-element insertion or chromosomal deletion, there were 30 replicates. For each replicate, recombination rates were estimated by taking the ratio of recombinant progeny to the total number of progeny.
The RNAi lines followed an identical crossing scheme except for the males used in the F0 cross. These males contained the doubly-marked chromosome (e ro) along with nanos GAL4 driver [176,177]. nanos is expressed throughout Drosophila oogenesis [178]. All P-element insertions, chromosomal deletions or RNAi lines were compared to appropriate controls using Dunnett’s Test [179,180] using both the raw recombination proportions as well as arcsined transformed data. Statistics were performed in JMP Pro 11.2.0.
To test the hypothesis that gene expression differences between alleles drive phenotypic variation in recombination rate, we analyzed ovarian mRNA abundance differences between the major and minor allele for each of our 20 candidate genes using quantitative RT-PCR (qPCR). For each candidate gene, three DGRP lines containing the major allele and three DGRP lines containing the minor allele were chosen (S17 Table). For the eight genes that had multiple significant genetic variants associated within the gene region, DGRP lines that contained the most major/minor alleles were selected (S18 Table). For each candidate gene, virgin females were collected from the six DGRP lines contemporaneously to minimize the effects of environmental variation. Females were aged three days in vials with ~0.5 mL of yeast paste. Ovaries were then dissected from anesthetized females in a solution of 1X PBS and stored in Life Technologies RNAlater solution (Life Technologies). For each line, four replicates of ten pairs of ovaries were dissected. Total RNA was extracted from homogenized ovaries using Trizol (Life Technologies) following manufacturer’s instructions. cDNA was generated using Bio-Rad iScript cDNA Synthesis and following manufacturer’s instructions. Primers for candidate genes were generated using FlyPrimerBank [181] (S22 Table). qPCR was run a BioRad CFX384 machine using Bio-Rad iQ SYBR Green following manufacturer's instructions. Four technical replicates for each sample were run on the same 384 plate, minimizing the contribution of between plate variation.
Samples were analyzed using GAPDH for normalization due to its relatively consistent expression [182]. For each candidate gene, there were six lines analyzed, three that contained the major allele and three that contained the minor allele identified in our GWAS. For each line, we collected four biological replicates of RNA. We ran four technical replicates of each RNA sample (converted to cDNA). Therefore, for each line, there are a total of 16 qPCR measurements for the candidate gene of interest and 16 qPCR measurements for the GAPDH control. Measurements from each DGRP line were normalized by dividing by the average Cq value of GAPDH for the corresponding DGRP line, modeled after common normalization procedures [183]. These 96 measurements (48 measurements for the major allele and 48 measurements for the minor allele) were then analyzed by comparing the means of the lines containing the major allele to the means of the lines containing the minor allele via a students t-test using JMP Pro 11.2.0. In addition, the raw Cq values (before normalization) were also analyzed to ensure that potential differential GAPDH expression was not biasing results.
|
10.1371/journal.pbio.1001603 | Effects of Resveratrol and SIRT1 on PGC-1α Activity and Mitochondrial Biogenesis: A Reevaluation | It has been reported that feeding mice resveratrol activates AMPK and SIRT1 in skeletal muscle leading to deacetylation and activation of PGC-1α, increased mitochondrial biogenesis, and improved running endurance. This study was done to further evaluate the effects of resveratrol, SIRT1, and PGC-1α deacetylation on mitochondrial biogenesis in muscle. Feeding rats or mice a diet containing 4 g resveratrol/kg diet had no effect on mitochondrial protein levels in muscle. High concentrations of resveratrol lowered ATP concentration and activated AMPK in C2C12 myotubes, resulting in an increase in mitochondrial proteins. Knockdown of SIRT1, or suppression of SIRT1 activity with a dominant-negative (DN) SIRT1 construct, increased PGC-1α acetylation, PGC-1α coactivator activity, and mitochondrial proteins in C2C12 cells. Expression of a DN SIRT1 in rat triceps muscle also induced an increase in mitochondrial proteins. Overexpression of SIRT1 decreased PGC-1α acetylation, PGC-1α coactivator activity, and mitochondrial proteins in C2C12 myotubes. Overexpression of SIRT1 also resulted in a decrease in mitochondrial proteins in rat triceps muscle. We conclude that, contrary to some previous reports, the mechanism by which SIRT1 regulates mitochondrial biogenesis is by inhibiting PGC-1α coactivator activity, resulting in a decrease in mitochondria. We also conclude that feeding rodents resveratrol has no effect on mitochondrial biogenesis in muscle.
| Studies on cultured muscle cells have shown that treatment with resveratrol, a chemical famously found in the skin of red grapes, stimulates the manufacture of new mitochondria. This has been attributed to the activation of the deacetylase SIRT1 either directly by resveratrol or indirectly via the activation of AMP-activated protein kinase (AMPK). SIRT1 is then thought to deacetylate and activate the transcriptional coactivator PGC-1α, which in turn stimulates mitochondrial biogenesis. It has also been reported that feeding resveratrol to mice increases muscle mitochondria and results in improved running endurance. Here we further analyze the adaptive response of muscle mitochondria to resveratrol treatment to see if it mimics the response to endurance exercise. We find that feeding rats or mice large amounts of resveratrol did not increase muscle mitochondria. In these rodents, the bioavailability of oral resveratrol is low, and the resulting plasma level of resveratrol is far below the concentration required to activate AMPK. Contrary to previous reports we find that deacetylation by SIRT1 decreases PGC-1α activity and results in a decrease in mitochondria; moreover we show that the increase in mitochondria induced in cultured muscle cells by a high resveratrol concentration is due to the toxic activation of AMPK and, in turn, PGC-1α. However, this effect requires resveratrol concentrations that are very much higher than those attained by oral administration, and we conclude that oral resveratrol has no effect on mitochondrial biogenesis in skeletal muscle.
| Resveratrol has been reported to have a number of remarkable effects in mice. These include protection against high-fat-diet-induced obesity and insulin resistance [1]–[3], marked improvements in running endurance and maximal oxygen uptake capacity (VO2max) [3], increased muscle strength [3], improved motor coordination [1]–[3], and antiaging effects [1],[2]. Subsequent studies have shown that resveratrol does not have antiaging effects in mice, as evidenced by no increases in average or maximum longevity [4],[5]. The protection against obesity and insulin resistance was attributed to an increase in, and uncoupling of, mitochondria in brown fat, and the increase in running endurance and VO2max were attributed to an increase in muscle mitochondria [3]. The increase in mitochondria induced by resveratrol was explained by activation of the protein deacetylase SIRT1, resulting in deacetylation and activation of the transcription coactivator PGC-1α [3]. PGC-1α regulates mitochondrial biogenesis [6]. The pharmaceutical agent SRT1720 has also been reported to activate SIRT1, resulting in PGC-1α activation, and an increase in enzymes of the mitochondrial fatty acid oxidation pathway in muscle and improved running performance, muscle strength, and coordination [7]. However, Pacholec et al. [8] have reported that SRT1720 does not activate SIRT1, and that it does not induce an increase in mitochondrial enzymes in mice. Based on studies on yeast and in vitro, it was initially thought that resveratrol directly activates SIRT1 [9]. However, Kaeberlein et al. [10] showed that, although resveratrol binds and deacetylates peptide substrates that contain a Fluor de Lys, it does not bind or deacetylate acetylated peptides lacking the flurophore. They also found that resveratrol has no effect on SIRT2 activity in yeast. Similarly, Bora et al. [11] found that resveratrol activation of SIRT1 was completely dependent on the presence of a covalently attached flurophore. Evidence that resveratrol can activate AMP activated protein kinase (AMPK) [12]–[14] led to further studies that indicated that the activation of SIRT1 by resveratrol is indirect, and is mediated by activation of AMPK [15]. The mechanism by which AMPK is thought to activate SIRT1 is by increasing NAD concentration [15].
We have a long-standing interest in the adaptive responses to endurance exercise, such as running and swimming, which include an increase in muscle mitochondria [16],[17]. Endurance exercise training also results in increases in endurance and in maximal oxygen uptake capacity. Endurance exercise does not, by itself, result in increases in either muscle strength, which occurs in response to heavy resistance exercise, or improved motor coordination, which occurs in response to activities that require various motor skills. A sedentary lifestyle greatly increases the risk of developing obesity, insulin resistance, type 2 diabetes, atherosclerosis, and frailty [18]. Therefore, in addition to being necessary for successful competition in sports, regular exercise is necessary for maintenance of health and functional capacity. Because it is difficult to motivate people to exercise, an effective, nontoxic exercise mimetic—that is, an “exercise pill”—could have great public health value. Therefore, the reports that, in addition to protecting against obesity and insulin resistance, resveratrol feeding mimics not only the adaptive response to endurance exercise but also the adaptations to strength training and motor skill exercise training were of great interest to us. The present study was undertaken to further evaluate the adaptive response of skeletal muscle mitochondria to resveratrol treatment.
Feeding rats resveratrol in a chow diet containing 4 g resveratrol per kg diet [3] for 8 wk had no effect on the expression of PGC-1α or on a number of mitochondrial proteins in rat skeletal muscle as shown in soleus muscle (Figure 1A). A similar lack of effect was found in the gastrocenemius muscle. Feeding rats a high fat diet containing 4 g resveratrol per kg diet also had no effect on the expression of a range of mitochondrial enzyme proteins (Figure 1B). To rule out the possibility that the lack of effect of resveratrol on the mitochondrial content of skeletal muscle in rats was due to a species difference, we fed mice a high fat diet containing 4 g resveratrol per kg/diet as in the study by Lagouge et al. [3]. As in the rats, resveratrol feeding had no effect on the expression of PGC-1α or a number of mitochondrial proteins in skeletal muscle of mice (Figure 1C). To evaluate the possibility that the lack of effect of resveratrol on mitochondrial biogenesis is due to an inadequate increase in plasma resveratrol, we measured plasma resveratrol concentration. Plasma resveratrol concentration at 9:00 am in rats in the fed state averaged 1.56±0.28 µM. This plasma resveratrol concentration is higher than that reported by Lagouge et al. [3] in their resveratrol fed mice, in which the highest concentration attained was ∼0.5 µM.
Most of the information regarding the effects of resveratrol on, and the role of SIRT1 in, the regulation of mitochondrial biogenesis has come from studies on C2C12 myotubes or other cells in culture. Because resveratrol feeding had no effect on mitochondrial biogenesis in laboratory rodents, we evaluated the effect of resveratrol on mitochondrial biogenesis in C2C12 myotubes. The concentration of resveratrol that was routinely used in studies on C2C12 myotubes by Auwerx's group was 50 µM [3],[15], ∼100-fold higher than the highest plasma resveratrol level in their resveratrol fed mice [3]. In our initial experiments we found that 50 µM resveratrol is toxic, with a high proportion of the C2C12 myotubes appearing to be dead or dying after 24 h of exposure to 50 µM resveratrol. That this concentration of resveratrol is cytotoxic was born out by measurements of cytotoxicity (Figure 2A) and of ATP concentration, which was markedly reduced (Figure 2B). Similarly, Zang et al. [13] have reported that exposure of Hep-G2 cells to 50 µM resveratrol for 60 min resulted in an 80% reduction in ATP concentration. The decrease in ATP concentration in cells exposed to a high concentration of resveratrol is mediated by toxic effects on mitochondria, with inhibition of ATP synthase [19] and NADH: ubiquinone oxidoreductase [20]. Numerous studies have shown that concentrations of resveratrol in the 30 to 100 µM range kill a variety of malignant cells [21]. These studies were uncontrolled, and it was assumed that resveratrol specifically kills cancer cells. However, the present finding and that of Zang et al. [13] show that resveratrol at the high concentrations used is also lethal for nonmalignant cells.
In the study in which 50 µM resveratrol increased mitochondrial biogenesis in C2C12 myotubes [3], the investigators used cells that overexpressed PGC-1α. We have observed that myotubes in which PGC-1α is overexpressed have increased resistance to the effect of puromycin (DH Han and JO Holloszy, unpublished findings), suggesting the possibility that overexpression of PGC-1α results in a nonspecific increase in resistance to toxins. We, therefore, evaluated the effect of 50 µM resveratrol in C2C12 cells in which PGC-1α was overexpressed by infection with a virus expressing PGC-1α. As shown in Figure 2C, the toxic effect of 24 h exposure to 50 µM resveratrol on cell viability was markedly reduced. However, there was still a significant reduction in ATP concentration (Figure 2D). Treatment with 50 µM resveratrol for 24 h resulted in an increase in mitochondrial biogenesis in the myotubes in which PGC-1α was overexpressed, as evidenced by increases in the expression of a number of mitochondrial proteins (Figure 3A). All of our subsequent experiments in which 50 µM resveratrol was used were performed on C2C12 myotubes in which PGC-1α was overexpressed.
To evaluate the effect of resveratrol in the absence of PGC-1α overexpression, we tried to identify a resveratrol concentration that induces an increase in mitochondrial proteins in wild type C2C12 cells. Resveratrol concentrations in the 1 µM to 10 µM range did not result in a decrease in ATP concentration (Figure 2B). Although exposure to 20 µM resveratrol for 24 h is less toxic than exposure to 50 µM, it results in a decrease in cell viability (Figure 2A) and a ∼50% decrease in ATP concentration (Figure 2B). Six hours of treatment with 20 µM resveratrol resulted in a smaller decrease in ATP (∼20%), and wild-type C2C12 cells treated with 20 µM resveratrol for 6 h followed by an 18 h recovery period showed no evidence of toxicity. “Training” the wild-type C2C12 cells by exposing them to 20 µM resveratrol for 6 h per day for 3 d resulted in increases in PGC-1α and a number of mitochondrial proteins (Figure 3B), while the same treatment with 1 µM, 5 µM, or 10 µM resveratrol had no effect (Figure S1).
As shown in Figure 4A, treatment with 20 µM resveratrol resulted in increased phosphorylation of AMPK and acetyl-CoA carboxylase (ACC) in C2C12 myotubes. As a first step in evaluating the relative roles of AMPK and SIRT1 in the resveratrol-induced increase in mitochondrial biogenesis, we infected C2C12 myotubes with an adenovirus encoding a dominant-negative AMPK gene construct. That the DN AMPK was effective in blocking AMPK activity is demonstrated by prevention of increases in AMPK and ACC phosphorylation in response to resveratrol treatment (Figure 4B). Blocking AMPK activity prevented induction of an increase in mitochondrial proteins by resveratrol (Figure 4C), showing that AMPK activation is necessary for stimulation of mitochondrial biogenesis by resveratrol.
Jäger et al. [22] have shown that AMPK directly phosphorylates and activates PGC-1α. Canto et al. [15] have interpreted their data to indicate that phosphorylation of PGC-1α by AMPK constitutes a priming event for subsequent deacetylation by SIRT1, and that deacetylation of PGC-1α is a key mechanism by which AMPK triggers PGC-1α activity. To further evaluate the relative roles of SIRT1 and AMPK in the resveratrol-induced increase in mitochondria, we used nicotinamide to inhibit SIRT1 [23]. That 10 mM nicotinamide decreases SIRT1 activity in C2C12 myotubes is evidenced by the finding of increases in the acetylation of p53, which is a SIRT1 substrate [24] (Figure 5A) and of PGC-1α (Figure 5C). Nicotinamide also prevented p53 deacetylation in response to 50 µM resveratrol (Figure 5A). However, we were surprised to find that nicotinamide did not prevent the resveratrol-induced increase in mitochondrial proteins (Figure 5B). Treatment of C2C12 myotubes with 10 mM nicotinamide had no effect on ATP concentration (nicotinamide 5.3±0.l3 µmol/g protein, Control 5.7±0.18; n = 6 per group).
We further evaluated the role of SIRT1 in mitochondrial biogenesis by suppression of SIRT1 activity by adenovirus-mediated expression of a dominant-negative (DN) SIRT1 H355A [23], and knockdown of SIRT1 with a shRNA, in C2C12 myotubes. SIRT1 H355A suppressed SIRT1 activity as evidenced by an increase in PGC-1 acetylation and inhibition of resveratrol-induced PGC-1α deacetylation (Figure 5C). Both the DN SIRT1 and the SIRT1 shRNA resulted in increased PGC-1α coactivator activity, measured in C2C12 myotubes co-transfected with a PGC-1α–GAL4 fusion construct and a luciferase reporter [25], and enhanced the resveratrol-induced increase in PGC-1α activity (Figure 5D). Overexpression of wild-type SIRT1 resulted in PGC-1α deacetylation (Figure 5C), reduced PGC-1α coactivator activity, and prevented the increase in PGC-1α activity induced by resveratrol (Figure 5D).
SIRT1 H355A expression in myotubes resulted in an increase in mitochondrial enzyme proteins (Figure 6A). Expression of SIRT1 H355A in rat triceps muscle by electroporation also resulted in an increase in mitochondrial enzyme proteins (Figure 6B). Furthermore, knockdown of SIRT1 by transfection of C2C12 myotubes with a SIRT1 shRNA brought about an increase in mitochondrial proteins, providing further evidence that acetylation activates PGC-1α (Figure 6C). Expression of DN SIRT1 H355A in C2C12 myotubes had no effect on ATP concentrations (Control 5.7±0.18, DN SIRT1 H355A 6.0±0.3; n = 6 per group). To further evaluate the effect of SIRT1 on mitochondrial biogenesis, we determined the effect of overexpression of SIRT1 by adenovirus mediated infection of C2C12 cells, and electroporation of rat triceps muscle, with a SIRT1 gene construct. SIRT1 overexpression resulted in a decrease in cytochrome c and inhibited the resveratrol-induced increase in cytochrome c in C2C12 myotubes (Figure 6D). Overexpression of SIRT1 in rat triceps muscle resulted in decreases in mitochondrial enzyme proteins (Figure 6E).
Interestingly, the increase in PGC-1α coactivator activity induced by acetylation does not result in an increase in PGC-1α expression (Figure 6). This is in contrast to PGC-1α activation by phosphorylation by AMPK and/or p38 MAPK, which is associated with an increase in PGC-1α expression (Figure 3B) [22],[26]–[29]. A probable explanation for this difference is that AMPK and p38 MAPK do not just activate PGC-1, but also activate the transcription factors that induce increased PGC-1α expression. P38 MAPK phosphorylates and activates ATF2, which binds to a CREB binding site on the PGC-1α promoter, and AMPK and p38 MAPK bring about activation of MEF2, which binds to a MEF2 binding site on the PGC-1α promoter, resulting in increased PGC-1α transcription [27],[30]–[32].
In the present study, resveratrol feeding had no effect on mitochondrial biogenesis in skeletal muscle even though our animals were fed a diet containing the same amount of resveratrol, 4 g/kg diet, as used by Lagouge et al. [3], and more than the dose, 0.4 g/kg diet, used by Bauer et al. [1]. In studies on the effects of resveratrol on cells in culture, concentrations in the 30 µM to 100 µM range have routinely been used [3],[12],[14],[15]. Based on our findings on C2C12 myotubes, the concentration of resveratrol required to induce an increase in mitochondrial biogenesis is above 10 µM, and the data shown by Bauer et al. [1] suggest the concentration of resveratrol needed to activate AMPK in CHO cells is also above 10 µM. The plasma resveratrol concentration in our rats was 1.56±0.28 µM and the highest concentration in the mice of Lagouge et al. [3] was ∼0.5 µM. Thus, a likely explanation for the failure of resveratrol feeding to induce mitochondrial biogenesis in rats and mice in our study is its poor bioavailability.
In our experiments on C2C12 cells, we confirmed the finding of Lagouge et al. [3] that treatment of C2C12 cells with a high concentration of resveratrol results in both PGC-1α activation, evaluated using a PGC-1α–GAL4 construct together with a luciferase reporter, and an increase in mitochondrial biogenesis. The research groups of Auwerx and Puigserver have published a large number of studies, reporting that deacetylation activates and acetylation deactivates PGC-1α [3],[7],[15],[33]–[39]. Phosphorylation of PGC-1α by AMPK results in PGC-1 activation and increased mitochondrial biogenesis [22]. We found that high concentrations of resveratrol activate AMPK in C2C12 cells by a toxic effect on mitochondria that reduces ATP level, and that this is the mechanism by which resveratrol activates PGC-1α. We also found that the concomitant increase in SIRT1 activity, also mediated by AMPK, results in a deacetylation of PGC-1α that causes a blunting of the increase in PGC-1α activity induced by AMPK. This is in contrast to the report by Canto et al. [15] that activation of PGC-1α by AMPK is dependent on PGC-1α deacetylation by SIRT1. In support of this conclusion, they reported that inhibition of SIRT1 with nicotinamide or knock down of SIRT1 markedly reduced PGC-1α activation and attenuated the increase in mitochondrial proteins in response to AMPK activation.
We confirmed that activation of AMPK results in SIRT1 activation, as evidenced by deacetylation of p53 and PGC-1α. We also confirmed that suppression of AMPK activity blocks the increase in mitochondrial proteins induced by resveratrol. However, we were surprised to find that inhibiting SIRT1 with nicotinamide did not prevent the resveratrol-induced increase in mitochondrial proteins in C2C12 myotubes. Furthermore, an increase in PGC-1α acetylation, mediated by suppression of SIRT1 activity using a dominant-negative SIRT1 construct, resulted in an increase in PGC-1α coactivator activity and mitochondrial biogenesis. Knockdown of SIRT1 also increased PGC-1α activity. Further evidence that PGC-1 is activated by acetylation is provided by the findings that overexpression of wild-type SIRT1, resulting in PGC-1 deacetylation, decreases mitochondrial proteins, blunts the resveratrol/AMPK-induced increase in cytochrome c, and reduces PGC-1α coactivator activity. An additional mechanism by which the inhibitory effect of SIRT1 on PGC-1α activity might be mediated is by deacetylation and inactivation of the transacetylase p300 [40]. p300 is a transacetylase that binds to and acetylates PGC-1α [41], and powerfully enhances its coactivator activity [42]. Thus, inactivation of p300, resulting in decreased PGC-1α acetylation, could result in a reduction of PGC-1α activity.
Our findings that SIRT1 activation decreases PGC-1α coactivator activity and that inhibition or knockdown of SIRT1 increases PGC-1α activity are in keeping with data reported by Finkel's group [41]. These investigators showed that SIRT1 binds to and deacetylates PGC-1α, and that increasing SIRT1 expression in PC12 cells results in a ∼25% reduction in O2 consumption, a ∼45% decrease in cytochrome oxidase (COX) subunit 2 expression, and a ∼50% decrease in activity of a GAL4–PGC-1α fusion construct [41]. They also found that overexpression of the transacetylase p300, which activates PGC-1α [42], dramatically increased PGC-1α acetylation [41]. Our findings also confirm the report by Gurd et al. [43] that overexpression of SIRT1 in rat skeletal muscle results in decreased expression of the mitochondrial enzyme COXIV. Gurd et al. [43] also found an inverse relationship between mitochondrial content and SIRT1 content in different types of skeletal muscle and heart muscle.
SIRT1 is induced by, and appears to play a key role in the adaptive responses to, fasting, starvation, and calorie restriction [44]–[47]. The Auwerx and Puigserver research groups have interpreted their findings to indicate that SIRT1 leads to increased mitochondrial biogenesis, which provides a molecular mechanism that allows cells to survive and adapt to periods of nutrient deprivation [35], that SIRT1 activation by SRT1720 mimics low energy levels [7], and that “interdependent regulation of SIRT1 and AMPK provide a finely tuned amplifier mechanism for energy homeostasis under low energy availability” [15]. A key component of this concept is that mitochondrial adaptations induced by increased SIRT1 activity are necessary for the switch from carbohydrate to fat oxidation in response to fasting [7],[35]. What was actually reported is that treatment with resveratrol [3] or SIRT1720 [7] and other interventions that activated SIRT1 [34],[38] resulted in increases in basal oxygen consumption, heat production/body temperature, and protection against weight gain or reduced weight gain despite no decrease in food consumption. This syndrome, which resembles hyperthyroidism, was attributed by the authors to mitochondrial adaptations in brown adipose tissue and is incompatible with the large increase in running endurance reported in these mice [3],[7],[38].
Adaptive responses were selected for because they enhance an organism's chances of surviving environmental changes. Increases in energy expenditure and substrate oxidation resulting in more rapid depletion of energy stores, such as were reported to occur with SIRT1 activation, would be seriously maladaptive responses to fasting, starvation, or CR. Actually, it is well documented that fasting and CR result in decreases in metabolic rate, as reflected in lower resting oxygen consumption and a decrease in body temperature [48]–[51]. With regard to the claim that an increase in mitochondrial fatty acid oxidation enzymes is necessary for the switch from carbohydrate to fatty acid oxidation in muscle [7],[35], no increase in mitochondria is needed. Skeletal muscle has a low rate of energy utilization at rest and contains sufficient mitochondria to make possible a many-fold, acute increase in fatty acid oxidation in response to exercise that greatly exceeds the increase in fat oxidation in muscle in response to fasting. Furthermore, SIRT1-null mice are hypermetabolic, have elevated rates of fatty acid utilization, and readily switch from carbohydrate to fat oxidation in response to fasting [52].
In conclusion, our results show that resveratrol feeding does not induce an increase in muscle mitochondria in rodents. This lack of effect may be due to poor bioavailability, because the plasma levels of resveratrol attained in response to feeding large amounts of resveratrol are far below the concentration of resveratrol required to activate AMPK. This seems fortunate, because the activation of AMPK by resveratrol is mediated by a toxic effect that depletes ATP in cells exposed to AMPK-activating concentrations of resveratrol. Surprisingly, in light of the many studies reporting that deacetylation of PGC-1α results in activation of PGC-1α's coactivator activity, we find that deacetylation decreases, and PGC-1α acetylation increases, PGC-1α activity and mitochondrial biogenesis. Our results indicate that the activation of PGC-1α by resveratrol is mediated by AMPK, and that the activation of SIRT1 by AMPK acts to reduce, rather than induce, this activation.
This research was approved by the Animal Studies Committee of Washington University School of Medicine. Rats were lightly anesthetized during muscle electroporation. Rats were anesthetized with pentobarbital and, after muscles were harvested, were killed by exsanguination while under anesthesia.
Antibodies against cytochrome oxidase subunit I (COXI), cytochrome oxidase subunit IV (COX IV), Core II, Complex III FeS, NADH ubiquinol oxidoreductase (NADH-UO), and succinate ubiquinol oxidoreductase(SUO) ATP synthase alpha subunit #45924 and lipofectamine 2000 were purchased from Invitrogen (Carlsbad, CA). Anti-cytochrome c antibody was obtained from BD Biosciences (San Jose, CA). Antibodies against p53, acetyl-p53, AMP-activated protein kinase (AMPK), phospho-AMPK, acetyl-CoA carboxylase (ACC) and phospho-ACC were products of Cell Signaling technology (Beverly, MA). An anti-SIRT1 antibody #09844 was purchased from EMD Millipore. An anti-PGC-1α c-terminal (777–797) antibody #516557 was purchased from EMD Millipore (Billerica, MA); an antibody against acetylated lysine #9441 was purchased from Cell Signaling (Beverly, MA). Horseradish peroxidase (HRP)–conjugated donkey anti-rabbit IgG and donkey anti-mouse IgG were purchased from Jackson ImmunoResearch Laboratories (West Grove, PA). Enhanced chemiluminescence (ECL) reagents were obtained from Amersham (Arlington Heights, IL). All other reagents were purchased from Sigma (St. Louis, MO).
This research was approved by the Animal Studies Committee of Washington University School of Medicine. Male Wistar rats weighing ∼95 g were purchased from Charles Rivers (Wilmington, MD) and housed in individuals cages. The resveratrol used in the study on rats was purchased from Stryka Botanicals (Hillsborough, NJ). Control rats were fed Purina rodent chow, and the resveratrol-fed rats were given the chow diet containing 4 g resveratrol per kg diet, for 8 wk. Male c57BL/6J mice were purchased from Jackson Laboratory (Bar Harbor, ME), housed 6 per cage, and fed a high fat diet, 50% of calories from fat, or the high fat diet containing 4 g resveratrol per kg diet 3 for 8 wk. The resveratrol used in the study on mice was a kind gift from DSM Nutritional Products (Basel, Switzerland). (The resveratrol used in the study by Lagouge et al. was from Orchid, Chennai, India.)
Rats or mice were anesthetized with sodium pentobarbital 5 mg/100 g body weight. Muscles were dissected out, clamp-frozen, and kept at −80°C until used for assays.
ATP concentration was measured using a luminescence ATP detection assay (ATPlite, Perkin Elmer, Waltham, MA); LDH activity, as an indicator of cytotoxicity, was measured using an LDH-Cytotoxicity Assay Kit (BioVision, Mountain View, CA), according to the manufacturer's instructions.
For expression in skeletal muscle via electroporation, wild-type SIRT1 and dominant-negative SIRT-1 H355A constructs were purchased from Addgene (Cambridge, MA) and inserted into pCDNA3.1 (Invitrogen, Carlsbad, CA). For expression in C2C12 myoblasts by transfection, a gal-4-DBD-PGC-1α plasmid was purchased from Addgene (Cambridge, MA) [53], a 9×gal-4–dependent reporter plasmid was purchased from Promega (Madison, WS), and a LacZ control plasmid was purchased from Invitrogen (Carlsbad, CA). For expression by adenoviral infection in C2C12 myotubes, the adenoviral constructs of pAd-Track Flag-PGC-1α [36], pAd-Track Flag-SIRT1 [23], and pAd-Track Flag dominant-negative SIRT1 H355A [23] were purchased from Addgene (Cambridge, MA). Dominant-negative Myc-AMPKα 2 DNA [54] was PCR cloned and ligated into pAd-Track plasmid. Mouse SIRT1 shRNA (5′-GCCCTGTAAAGCTTTCAGAA-3′) and scrambled control (5′-GATGAAGTCGACCTCCTCAT-3′) sequences were cloned into pRNAT-H1.1/adeno (Genescript, Piscataway, NJ). Recombinant adenoviruses were generated employing a system described by He et al. [55].
Transfection of DNA into rat skeletal muscle was accomplished by using an electric pulse-mediated gene transfer technique [56]. Male Wistar rats weighing ∼60 g were anesthetized with isoflurane gas. A triceps muscle was injected with 100 µg of plasmid DNA containing either empty vector, pcDNA3.1 SIRT1 WT, or pcDNA3.1 Sirt1 H355A in 100 µl saline, using a 27 gauge needle, at a rate of 0.04 ml/min. After injection, an electric field was applied to the triceps muscle using a S88 square-pulse stimulator (Grass) with a 533 model two-needle array (BTX). The electric field application consisted of 8 pulses of 100 ms duration, at a frequency of 1 Hz and amplitude of 100 volts, that were applied perpendicular to the muscles' long axis. Muscles were harvested 14 d after electroporation.
C2C12 mouse myoblasts were grown in DMEM (4.5 g glucose/L, Invitrogen) containing 10% fetal bovine serum, 100 µU/ml penicillin, and 100 µU/ml streptomycin. Differentiation was initiated by switching to medium containing 2% heat inactivated horse serum when the myoblasts were 90% confluent. After 48 h of differentiation, batches of myotubes were infected with adenoviruses expressing (a) Flag-PGC-1α, (b) dominant-negative Myc-AMPKα 2, (c) dominant-negative Flag-SIRT1 H355A, (d) Flag-SIRT1, and (e) SIRT1 shRNA. At 96 h after differentiation, batches of C2C12 myobutes were treated with 20 µM or 50 µM resveratrol or vehicle for the time periods given in the figures, or with 10 mM nicotinamide or vehicle for 24 h.
Homogenates were prepared and Western blotting was performed as described previously [57] using the antibodies described previously [57],[58].
To evaluate the effect of SIRT1 on PGC-1α transcription coactivator activity, C2C12 myoblasts were co-transfected with a gal-4-DBD PGC-1α plasmid, and a 9×gal-4-dependent reporter plasmid, or with a LacZ control plasmid, and with either wild-type SIRT1, dominant-negative SIRT1 H355A, or SIRT1 shRNA-plasmids using lipofectamine 2000. After overnight transfection the culture medium was changed to DMEM containing 10% FBS. Thirty-six hours later, some of the cells were treated with 20 µM resveratrol for 6 h and harvested after a 6 h recovery period. Dual luciferase assays were performed using a kit (Invitrogen) according to the manufacturer's instructions.
Flag-PGC-1α was expressed in C2C12 myotubes by adenoviral infection. To evaluate the effect of SIRT1 on PGC-1α acetylation, the myotubes were co-infected with wild-type SIRT1 or SIRT1 H355A. Forty-eight hours after infection, myotubes were treated with 50 µM resveratrol or vehicle for 18 h. Wild-type C2C12 myotubes were treated with 10 mM nicotinamide for 24 h. The myotubes were then harvested, and cell extracts containing 200 µg of protein were rotated with anti-Flag antibody at 4°C overnight. The following morning, agarose G beads were added and the samples were rotated at room temperature for 2 h. The agarose beads were washed 4 times with PBS and protein was eluted from the beads with 5× SDS buffer, which was boiled for 5 min. PGC-1α was measured with an anti-PGC-1α antibody, and levels of PGC-1α acetylation were then assessed with an anti-acetyl lysine antibody (#9441 Cell Signaling Technology).
Results are expressed as means ± SE. The significance of differences between two groups was determined using Student's t test. For multiple comparisons, significance was determined by one-way analysis of variance followed by post hoc comparison using Tukey significant difference method.
|
10.1371/journal.ppat.1000076 | The Evolutionary Genetics and Emergence of Avian Influenza Viruses in Wild Birds | We surveyed the genetic diversity among avian influenza virus (AIV) in wild birds, comprising 167 complete viral genomes from 14 bird species sampled in four locations across the United States. These isolates represented 29 type A influenza virus hemagglutinin (HA) and neuraminidase (NA) subtype combinations, with up to 26% of isolates showing evidence of mixed subtype infection. Through a phylogenetic analysis of the largest data set of AIV genomes compiled to date, we were able to document a remarkably high rate of genome reassortment, with no clear pattern of gene segment association and occasional inter-hemisphere gene segment migration and reassortment. From this, we propose that AIV in wild birds forms transient “genome constellations,” continually reshuffled by reassortment, in contrast to the spread of a limited number of stable genome constellations that characterizes the evolution of mammalian-adapted influenza A viruses.
| Influenza A viruses are an extremely divergent group of RNA viruses that infect in a variety of warm-blooded animals, including birds, horses, pigs, and humans. Since they contain a segmented RNA genome, mixed infection can lead to genetic reassortment. It is thought that the natural reservoir of influenza A viruses is the wild bird population. Influenza A viruses can switch hosts and cause novel outbreaks in new species. Influenza viruses containing genes derived from bird influenza viruses caused the last three influenza pandemics in humans. In this study, we surveyed the genetic diversity among influenza A viruses in wild birds. Through a phylogenetic analysis of the largest data set of wild bird influenza genomes compiled to date, we were able to document a remarkably high rate of genome reassortment, with no clear pattern of gene segment association and occasional inter-hemisphere gene segment migration and reassortment. From this, we propose that influenza viruses in wild birds forms transient “genome constellations,” continually reshuffled by reassortment, in contrast to the spread of a limited number of stable genome constellations that characterizes the evolution of mammalian-adapted influenza A viruses.
| Low pathogenic (LP), antigenically diverse influenza A viruses are widely distributed in wild avian species around the world. They are maintained by asymptomatic infections, most frequently documented in aquatic birds of the orders Anseriformes and Charadriformes. As such, wild birds represent major natural reservoirs for influenza A viruses [1]–[11] and at least 105 species of the more than 9000 species of wild birds have been identified as harboring influenza A viruses [8],[12],[13]. These influenza A viruses, commonly referred to as avian influenza viruses (AIV), possess antigenically and genetically diverse hemagglutinin (HA) [14] and neuraminidase (NA) subtypes, which includes all known influenza A virus HA (H1–H16) and NA (N1–N9) subtypes. At least 103 of the possible 144 type A influenza A virus HA-NA combinations have been found in wild birds [8],[15].
AIV maintained in wild birds have been associated with stable host switch events to novel hosts including domestic gallinaceous poultry, horses, swine, and humans leading to the emergence of influenza A lineages transmissible in the new host. Adaptation to domestic poultry species is the most frequent [16]–[26]. Sporadically, strains of poultry-adapted H5 or H7 AIV evolve into highly pathogenic (HP) AIV usually through acquisition of an insertional mutation resulting in a polybasic amino acid cleavage site within the HA [15],[25]. The current panzootic of Asian-lineage HP H5N1 AIV appears to be unique in the era of modern influenza virology, resulting in the deaths of millions of poultry in 64 countries on three continents either from infection or culling. There are also significant zoonotic implications of this panzootic, with 379 documented cases in humans, resulting in 239 deaths in 14 countries since 2003 (as of April 2008 [27]). The Asian lineages of HP H5N1 AIV have also caused symptomatic, even lethal, infections of wild birds in Asia and Europe, suggesting that migratory wild birds could be involved in the spread of this avian panzootic [28]–[31]. Concern is heightened since wild birds are also likely to be the reservoir of influenza A viruses that switch hosts and stably adapt to mammals including horses, swine, and humans [3]. The last three human influenza pandemic viruses all contained two or more novel genes that were very similar to those found in wild birds [16],[20],[32],[33].
Despite the recent expansion of AIV surveillance [7],[8],[10],[34],[35] and genomic data [5], [36]–[38], fundamental questions remain concerning the ecology and evolution of these viruses. Prominent among these are: (i) the structure of genetic diversity of AIV in wild birds, including the role played by inter-hemispheric migration, (ii) the frequency and pattern of segment reassortment, and (iii) the evolutionary processes that determine the antigenic structure of AIV, maintained as discrete HA and NA subtypes. Herein, we address these questions using the largest data set of complete AIV genomes compiled to date.
The complete genomes of 167 influenza A viruses isolated from 14 species of wild Anseriformes in 4 locations in the U.S. (Alaska, Maryland, Missouri, and Ohio) were sequenced; viral isolates consisted of 29 HA and NA combinations, including 11 HA subtypes (H1–H8, H10–H12) and all 9 neuraminidase subtypes (N1–N9). These sequences were collected as part of an ongoing AIV surveillance project at The Ohio State University and collaborators in other states (1986–2005) using previously described protocols [39], and more than double the number of complete U.S.-origin avian influenza virus genomes available in GenBank. In total, 1340 viral gene segment sequences (2,226,085 nucleotides) were determined (Table S1) and are listed on the Influenza Virus Resource website (http://www.ncbi.nlm.nih.gov/genomes/FLU/Database/shipment.cgi).
Cloacal samples from wild birds frequently show evidence of mixed infections with influenza viruses of different subtypes by serologic analysis [39]–[41]. Therefore, the isolates chosen for sequence analysis were subjected to sequential limiting dilutions (SLD) [39]. The amplification and sequencing pipeline employed a ‘universal’ molecular subtyping strategy in which every sample was amplified with sets of overlapping primers representing all HA and NA subtypes. In this manner, samples without clear prior subtype information, and/or mixed samples, could be accurately analyzed. Despite performing SLD, 4 samples were shown by sequence analysis to represent a mixed infection (yielding sequence with more than one HA and/or NA subtype. In addition 40 samples had mismatches between the initial antigenic subtyping results (determined on first- or second-egg-passage isolates prior to SLD) and the subtype determined by sequence analysis of cDNA (following one SLD of low-egg-passage isolates) which suggests the possibility of minor populations of antigenically distinct viruses in the low-passage isolate that outgrew the dominant antigenic population in a foreign host system during the SLD or that mixed infections in first egg passage stock caused difficulty in initial subtyping and a dominant strain emerged during SLD (see table of viral isolates at http://www.ncbi.nlm.nih.gov/genomes/FLU/Database/shipment.cgi to examine the discordant results observed). Thus, up to 44 of 167 (26%) of isolates potentially represent mixed infections in the initial cloacal sample. Given the SLD procedure, the true rate of mixed infection, as defined by the presence of >1 HA and/or NA subtype, was likely to be even higher, although mis-serotyping cannot also be ruled out. Sequencing viral genomes directly from primary cloacal material would be the only way to assess the mixed infection frequency, in a manner unbiased by culture, but no such studies have yet been attempted to our knowledge.
For a more comprehensive analysis of AIV diversity, the AIV genomes from this study were compared to other AIV genomes available on GenBank [38]. In total, 452 HA sequences and 473 NA sequences, representative of the global diversity of AIV, were used in phylogenetic analyses. For the internal protein genes (PB2, PB1, PA, NP, M, NS), a subset of 407 complete globally-sampled AIV genomes was used to assess the degree of linkage among gene segments. Phylogenetic trees for the HA alignment (Figures 1a and S1) and NA alignment (Figure 1b and S2) are shown here. Phylogenetic trees for the six other gene segments are presented in Figures S3, S4, S5, S6, S7 and S8.
The topology of the HA phylogeny reflects the antigenically defined subtypes, with some higher-order clustering among them (e.g., H1, H2, H5 and H6; H7, H10 and H15; Figures 1a and S1), as seen previously in smaller studies [14],[42]. Although most subtypes are found in numerous avian species and occupy wide global distributions, this phylogenetic structure indicates that HA subtypes did not originate in a single radiation. More striking was the high level of genetic diversity between different subtypes; the average amino acid identity of 120 inter-subtype comparisons of full-length HA was 45.5%. As expected, inter-subtype comparisons of the HA1 domain exhibited more diversity, with an average inter-subtype identity of 38.5%. In contrast, intra-subtype identity is high (averaging >92%), even when comparing sequences from different hemispheres. Hence, the genetic structure of the AIV HA is characterized by highly divergent subtypes that harbor relatively little internal genetic diversity. However, 4 subtype comparisons show considerably less divergence (76–79% identity); H4 vs. H14, H7 vs. H15, H13 vs. H16, and H2 vs. H5, indicating that they separated more recently (Figure 1; see below).
A similar phylogenetic structure was seen in the NA (Figure 1b and S2), again with evidence for higher-order clustering (e.g., N6 and N9; N1 and N4). In contrast to the HA, however, levels of genetic divergence among the NA types are more uniform, with the 9 subtypes exhibiting an average inter-subtype identity of 43.6% (with an average intra-subtype identity of >89%) and no clear outliers. Hence, no new (detected) NA types have been created in the recent evolutionary past. This correlates with the more uniform distribution of NA than HA subtypes in wild bird AIV isolates [43].
The topology of the NS segment phylogeny was also of note, being characterized by the deep divergence among the A and B alleles as described [44] (Figure S8). Almost every HA and NA subtype of AIV contain both the A and B NS alleles, without evidence of ‘intermediate’ lineages expected under random genetic drift, strongly suggesting that the two alleles are subject to some form of balancing selection. The NS1 protein has pleiotropic functions during infection in mammalian cells, and plays an important role in down-regulating the type I interferon response [45]. Supporting these results are the elevated rates of nonsynonymous to synonymous substitution per site (ratio dN/dS) observed for the NS1 gene in both avian and human influenza viruses [46] suggesting that the NS1 protein has experienced adaptive evolution in both host types. Whether this selection relates to the role the NS1 protein plays in its interaction in the type I interferon pathway is currently unclear.
Far less genetic diversity is observed in the 5 remaining AIV gene segments (PB2, PB1, PA, NP, and M - Figures S3, S4, S5, S6 and S7). Indeed, the extent of diversity in these genes is less than that within a single HA or NA subtype, with average pairwise identities ranging from 95–99%. Our phylogenetic analysis also revealed a clear separation of AIV sequences sampled from the Eastern and Western Hemispheres, as previously noted (3,19), indicating that there is relatively little gene flow between overlapping Eastern and Western Hemisphere flyways. However, despite this strong biogeographic split, mixing of hemispheric AIV gene pools clearly occurs at a low level (see below).
To assess the frequency and pattern of reassortment in AIV, we compared the extent of topological similarity (congruence) among phylogenetic trees of each internal segment. This analysis revealed a remarkably frequent occurrence of reassortment, supporting previous studies on smaller data sets [37],[47]. For example, 5 H4N6 AIV isolates were recovered from mallards sampled at the same location in Ohio on the same morning and in the same trap (Figure 2). For the internal genes, these viruses contained 4 different genome ‘constellations’, with only 1 pair of viruses sharing the same constellation. In the data set as a whole, the large number of different subtype combinations recovered highlights the frequency of reassortment (Figures 1b and S2), and provides little evidence for the elevated fitness of specific HA/NA combinations in AIV isolates from wild birds. That the majority of HA/NA combinations have been recovered [8],[15] also strongly supports the high frequency of reassortment involving these surface protein genes.
Thus, while there is strong evidence of frequent reassortment between HA and NA, we also sought to assess the extent of reassortment among the less commonly studied internal gene segments. A maximum likelihood test of phylogenetic congruence [48] revealed that although the topologies of the internal segment trees are more similar to each other than expected by chance, so that the segments are not in complete linkage equilibrium (in which case they would be no more similar in topology than two random trees), the difference among them is extensive, indicative of extremely frequent reassortment and with little clear linkage among specific segments (Figure 3). Of the 6 internal segments, NS exhibited the least linkage to other genes, falling closest to the random distribution (i.e. possessed the greatest phylogenetic incongruence). This is compatible with the deep A and B allelic polymorphism in this segment. In contrast, the M segment showed the greatest phylogenetic similarly, albeit slight, to the other segments. Overall, however, the relationships between segments are better described by their dissimilarity than their congruence.
Occasional AIV isolates demonstrated hemispheric mixing with reassortment. As reported previously, the majority of such mixing occurs in shorebirds and gulls [36] (with the exception of Eurasian lineage H6 HA genes distributed widely in North American Anseriformes [5] as also revealed in this study). Interestingly, no completely Eurasian-lineage AIV genome has been reported in North America, or vice versa [9],[49]. This suggests that birds initially carrying AIV between the hemispheric flyways have not been identified in surveillance efforts. Most mixed isolates possess only one gene segment derived from the other hemisphere, indicating that there is little or no survival advantage for such hemispheric crossovers in the new gene pool. Since Asian lineage HP H5N1 AIV have been isolated from wild birds in Eurasia [50], concern has been raised over the importation of the virus into North America via migratory birds. Our analyses suggest that enhanced surveillance in gulls and other shorebirds may be warranted, and that with frequent reassortment (see below), entire Asian HP H5N1 AIV isolate genome constellations may not be detected in these surveys.
Overall, 25 of 407 (6%) AIV genomes show evidence of hemispheric mixing, with the phylogenies suggesting a general pattern of viral gene flow from Eurasia to North America: 5 North American isolates possessed two Eurasian-lineage internal gene segments, and 20 carried a single segment. North American isolates possessing a Eurasian-lineage M segment were the most common, seen in 18 isolates (Figure S7), followed by 8 with a Eurasian PB2 segment (Figure S3), four with a Eurasian PB1 segment (Figure S4), and 1 with a Eurasian PA segment (Figure S5). The 18 Eurasian M segments and the 8 Eurasian PB2 segments each form monophyletic groups, suggesting single introductions to North America. In each case, sequences from domestic ducks in China and turkeys in Europe were the closest relatives. It is therefore theoretically possible that some of these introductions may have been derived from imported poultry rather than migratory birds. In contrast, 3 of the 4 Eurasian PB1 and the single Eurasian PA segment in North American AIV contained genes whose closest relatives were in viruses found in red-necked stints from Australia. These small waders are widely migratory, with a range from Siberia to Australasia, and occasionally in Europe and North America. Interestingly, 23 of 25 such mixed genomes were observed in shorebirds along the U.S. Atlantic coast. Unfortunately, no complete AIV genomes are available from shorebirds on the U.S. Pacific coast for comparison.
In theory, two evolutionary models can explain the global pattern of AIV diversity, analogous to the allopatric and sympatric models of speciation. Under the allopatric model, the HA and NA subtypes correspond to viral lineages that became geographically isolated, resulting in a gradual accumulation of amino acid changes among them. Because of physical separation through geographical divergence, there is no requirement for natural selection to reinforce the partition of HA and NA diversity into discrete subtypes by preferentially favoring mutations at antigenic sites. In contrast, under the sympatric model, the discrete HA and NA subtypes originate within the same spatial population, such that natural selection must have reinforced speciation; subtypes that were too antigenically similar would be selected against because of cross-protective immune responses. Therefore, mutations would accumulate first at key antigenic sites, allowing subtypes to quickly diversify in the absence of herd immunity.
The AIV genomic data available here suggest a complex interplay of evolutionary processes. That discrete HA and NA subtypes, as well as the 2 divergent NS alleles, are maintained in the face of frequent reassortment strongly suggests that each represents a peak on a fitness landscape shaped by cross-immunity (Figure 4a). Under this hypothesis, ‘intermediate’ HA/NA/NS alleles would be selected against because they generate more widespread herd immunity, corresponding to fitness valleys. Indeed, it is the likely lack of immunological cross-protection at the subtype level that allows the frequent mixed infections described here (although mixed infections may also occur in young, immunologically naïve birds). Further, in most cases these divergent HA, NA and NS alleles circulate in the same bird species in the same geographical regions, compatible with their divergence under sympatry. In addition, 3 of the most closely related pairs of HA subtypes contain an HA that is rarely isolated or limited geographically or by host species restriction, implying that their dispersion is inhibited by existing immunity; H14 has only been isolated rarely in Southern Russia, H15 only in Australia, and H16 has only been described in gulls. The possible exception is H2–H5, where both subtypes have been isolated from a variety of bird species in a global distribution. Although these may represent more recent occurrences of allopatric speciation, antigenic cross-reactivity between the H2–H5, H7–H15, H4–H14 pairs was recently demonstrated [51], again compatible with the sympatric model. Further support for possible cross-immunity between these subtypes would require experimental challenge studies.
In contrast to the extensive genetic diversity seen in HA, NA and NS, the 5 remaining internal gene segments encode proteins that are highly conserved at the amino acid level, indicating that they are subject to widespread purifying selection. The fitness landscape for these genes is therefore not determined by cross-immunity, but by functional viability, with less selective pressure to fix advantageous mutations (Figure 4b). Further, given such strong conservation of amino acid sequence, large-scale reassortment is permitted as it will normally involve the exchange of functionally equivalent segments, with little impact on overall fitness. These data also suggest that the cross-immunity provided by these proteins is minimal.
Together, these global genomic data provide new insight into the different evolutionary dynamics exhibited by influenza A viruses in their natural wild bird hosts and in those viruses stably adapted to novel species (e.g., domestic gallinaceous poultry, horses, swine, and humans). Based on these analyses, we hypothesize that AIV in wild birds exists as a large pool of functionally equivalent, and so often inter-changeable, gene segments that form transient genome constellations, without the strong selective pressure to be maintained as linked genomes. Rather than favoring successive changes in single subtypes, geographic and ecologic partitioning within birds, particularly within the different flyways, coupled with complex patterns of herd immunity, has resulted in an intricate fitness landscape comprising multiple fitness peaks of HA, NA and NS alleles, interspersed by valleys of low fitness which prevent the generation of intermediate forms (Figure 4a).
In contrast, stable host switching involves the acquisition of a number of (as yet) poorly characterized mutations [24],[33],[52],[53] that serve to separate an individual, clonally derived influenza virus strain from the large wild bird AIV gene pool. Because adaptation to a new host likely limits the ability of these viruses to return to the wild bird AIV gene pool [24],[54], these emergent viruses must evolve as distinct eight-segment genome configurations within the new host. The ability of recent HP H5N1 AIV to cause spillover infections in wild birds is an unprecedented exception. Further, because humans represent a large and spatially mixed population, natural selection is able to act efficiently on individual subtypes [55]. Hence, a limited number of subtypes circulate within humans and evolve by antigenic drift to escape population immunity.
Notably, the recent Asian lineage HP H5N1 AIV strains are intermediate between these two contrasting influenza ecobiologies; a combination of large poultry populations allows natural selection to effectively drive rapid antigenic and genetic change within a single subtype [46],[56], while reassortment with the wild bird AIV gene pool facilitates the generation of new genome constellations [57]–[59]. Similar patterns have also been observed with the widely circulating H9N2 and H6N1 viruses in gallinaceous poultry in Eurasia [60],[61]. Previous analyses have also shown that recent HP H5N1 viruses had the highest evolutionary rates and selection pressures (dN/dS ratios) as compared to other AIV lineages [46]. Consequently, these results underscore the importance of determining the mechanistic basis of how H5N1 has spread so successfully among a diverse range of both domestic and wild bird species.
The genomes of 167 influenza A virus isolates recovered from 14 species of wild Anseriformes located in four U.S. states (Alaska, Maryland, Missouri, Ohio) were sequenced for this study; viral isolates consisted of 29 hemagglutinin (HA) and neuraminidase (NA) combinations, including H1N1, H1N6, H1N9, H2N1, H3N1, H3N2, H3N6, H3N8, H4N2, H4N6, H4N8, H5N2, H6N1, H6N2, H6N5, H6N6, H6N8, H7N3, H7N8, H8N4, H10N7, H10N8, H11N1, H11N2, H11N3, H11N6, H11N8, H11N9, H12N5. Cloacal swabs were collected as previously described [39] from 1986–2005 as part of The Ohio State University's ongoing influenza A virus surveillance activities and in collaboration with many researchers in other states since 2001. A table listing the details of each isolate are available from the Influenza Virus Resource page (http://www.ncbi.nlm.nih.gov/genomes/FLU/Database/shipment.cgi). Avian influenza viruses were originally isolated using standard viral isolation procedures after 1–2 passages in 10-day-old embryonated chicken eggs (ECEs) [62]. Type A influenza virus was confirmed using commercially available diagnostic assays (Directigen Flu A Assay, Becton Dickinson Microbiology Systems, Cockeysville, MD) and isolates were subtyped at the National Veterinary Services Laboratories (NVSL), Animal and Plant Health Inspection Service, United States Department of Agriculture, Ames, Iowa, using standard hemagglutinin inhibition and neuraminidase inhibition testing procedures [51].
Isolates for this investigation were generally selected from the viral archives based on antigenic diversity, clustering of recoveries, no evidence of antigenically mixed subtypes, and distribution over time. First- or second-egg-passage isolates in chorioallantoic fluid (CAF) were rapidly thawed from −80°C to room temperature, vortexed for 30 seconds and centrifuged at 1500 rpm for 10 minutes. Approximately 0.5 ml of CAF was drawn from the vial using a 26-gauge needle and subsequently passed through a 25 mm, 0.2 µm filter. Following filtration, a 10−1 CAF stock dilution was obtained by adding 0.2 ml filtered CAF to 1.8 ml Brain Heart Infusion Broth containing penicillin and streptomycin and vortexed for 30 seconds. Serial dilutions (10−6 maximum) were performed and 0.1 ml of each dilution was inoculated into each of four 10-day-old ECEs. After approximately 48 hours of incubation at 35°C/60% humidity, the inoculated eggs were chilled overnight and CAF was harvested from each egg and tested for hemagglutinating activity. The CAF from the last dilution positive for hemagglutinating activity was tested for the presence of type A influenza virus using the Directigen Flu A or Synbiotics Flu Detect Antigen Capture Test Strips™ (Synbiotics Corp., San Diego, CA). Hemagglutination titer assays were performed and CAF aliquots from the most dilute influenza A positive samples were stored at −80°C. If no endpoint titer was determined, the 10−6 CAF dilution was stored at −80°C and the procedure repeated utilizing 10−4 to 10−9 sequential dilutions.
Viral RNA was isolated from allantoic fluid using Trizol® Reagent (Invitrogen Corp., Carlsbad, CA) and transcribed into 20 µl of cDNA for a subset of samples [63]. Segment-specific universal primers designed to amplify partial and/or full-segments were initially used in RT-PCR assays to assess vRNA quality and RT-PCR primer specificity and sensitivity. Additionally, M13 sequencing tags (F primer: GTAAAACGACGGCCAG; R primer: CAGGAAACAGCTATGAC) were added to each primer set for ease of sequencing RT-PCR products in both forward and reverse directions.
For initiation of a high-throughput sequencing pipeline, a universal strategy for primer design was employed to ensure detection of multiple viral infections within a single sample. Primers were designed to semi-conserved areas of the six internal segments. For the segments encoding the external proteins, primers were designed from alignments of subsets of the 16 HA and 9 NA avian subtypes. Alignments were generated with MUSCLE [64] and visualized with BioEdit [65]. An M13 sequence tag was added to the 5′ end of each primer to be used for sequencing. Four sequencing reactions per run were analyzed on an agarose gel for quality control purposes. The sequence success rate of each primer pair was analyzed relative to the HA and NA subtype. Primers that did not perform well were altered or replaced. All primers and RT-PCR assay cycling conditions are available upon request.
Influenza A virus isolates were amplified with the OneStep RT-PCR kit (Qiagen, Inc., Valencia, CA). Amplicons were sequenced in both the forward and reverse directions. Each amplicon was sequenced from each end using M13 primers (F primer: TGTAAAACGACGGCCAGT; R primer: CAGGAAACAGCTATGACC). Sequencing reactions were performed using Big Dye Terminator chemistry (Applied Biosystems, Foster City, CA) with 2 µl of template cDNA. Additional RT-PCR and sequencing was performed to close gaps and to increase coverage in low coverage or ambiguous regions. Sequencing reactions were analyzed on a 3730 ABI sequencer and sequences were assembled in a software pipeline developed specifically for this project.
Once genomic sequence was obtained for an individual sample, reads for each segment were downloaded, trimmed to remove amplicon primer-linker sequence and low quality sequence, and assembled. A small genome assembly suite called Elvira (http://elvira.sourceforge.net/), based on the open-source Minimus assembler, was developed to automate these tasks. The Elvira software delivers exceptions including failed reads, failed amplicons, and insufficient coverage to a reference sequence (as obtained from GenBank), ambiguous consensus sequence calls, and low coverage areas. The avian influenza A sequences (with GenBank Accession numbers) produced from this ongoing study are available at http://www.ncbi.nlm.nih.gov/genomes/FLU/Database/shipment.cgi. The first 167 avian influenza genomes from this collection were submitted to GenBank and included in this study.
The genomes of avian influenza virus newly determined here were combined with those already available on GenBank, particularly from recent large-scale surveys of viral biodiversity [38]. Sequences from viruses isolated before 1970, which may have been subjected to extensive laboratory passage, were excluded as were the large numbers of H5N1 sequences collected in recent years (a sample of H5N1 genomes, 1997–2005, were included for analysis). In total, 452 HA sequences and 473 NA sequences were used in analyses. For the internal protein-encoding segments (PB2, PB1, PA, NP, M, NS), a total of 407 genomes were analyzed (by considering a common data set we were able to investigate patterns of segment linkage, see below). For each data set, sequence alignments of the coding regions were created using MUSCLE [64] and adjusted manually using Se-Al [66] according to their amino acid sequence. In the case of HA and NA, some regions of the inter-subtype sequence alignment were extremely divergent such that they could not be aligned with certainty (HA signal peptide and cleavage site insertions in HPAI H5 or H7, and variable small stalk deletions in NA). Because of their potential to generate phylogenetic error, these small regions of ambiguity were deleted. This resulted in the following sequence alignments used for evolutionary analysis: PB2 = 2277 nt; PB1 = 2271 nt; PA = 2148 nt; HA = 1683 nt; NP = 1494 nt; NA = 1257 nt; M = 979 nt; NS = 835 nt. All sequence alignments are available from the authors on request. Nucleotide and amino acid identity was calculated using Megalign (Lasergene 7.2, DNAStar, Madison, WI).
Using these alignments, maximum likelihood (ML) trees were inferred using PAUP* [67], based on the best-fit models of nucleotide substitution models determined by MODELTEST [68]. In most cases, the preferred model of nucleotide substitution was GTR+I+Γ4, or a close relative. For each of these trees, the reliability of all phylogenetic groupings was determined through a bootstrap resampling analysis (1000 pseudo-replicates of neighbor-joining trees estimated under the ML substitution model).
We employed a maximum likelihood method to assess the extent of phylogenetic congruence, indicative of reassortment [48]. To reduce any bias in phylogenetic structure caused by geographic segregation, only isolates from North American flyways were used in analyses of the internal gene segments. Briefly, ML trees for each internal gene segment were estimated as described above. Next, the log likelihood (-LnL) of each of the ML trees was estimated on each gene segment data set in turn, optimizing branch lengths under the ML substitution model in every case. The topological similarity between each gene segment tree on each data set was then determined by compared the difference in likelihood among them (Δ-LnL). Clearly, the greater the similarity in topology (congruence) among the trees for each segment, the closer their likelihood scores and so the more likely they are to be linked. To put the distribution of Δ-LnL values in context, we constructed 500 random trees for each data set and optimized their branch lengths in the same manner. If any of the Δ-LnL values among the ML trees falls within the random distribution then we can conclude that the gene segments in question are in complete linkage equilibrium. All these analyses were conducted using PAUP* package [67]. |
10.1371/journal.pcbi.1004346 | The Internal Dynamics of Fibrinogen and Its Implications for Coagulation and Adsorption | Fibrinogen is a serum multi-chain protein which, when activated, aggregates to form fibrin, one of the main components of a blood clot. Fibrinolysis controls blood clot dissolution through the action of the enzyme plasmin, which cleaves fibrin at specific locations. Although the main biochemical factors involved in fibrin formation and lysis have been identified, a clear mechanistic picture of how these processes take place is not available yet. This picture would be instrumental, for example, for the design of improved thrombolytic or anti-haemorrhagic strategies, as well as, materials with improved biocompatibility. Here, we present extensive molecular dynamics simulations of fibrinogen which reveal large bending motions centered at a hinge point in the coiled-coil regions of the molecule. This feature, likely conserved across vertebrates according to our analysis, suggests an explanation for the mechanism of exposure to lysis of the plasmin cleavage sites on fibrinogen coiled-coil region. It also explains the conformational variability of fibrinogen observed during its adsorption on inorganic surfaces and it is supposed to play a major role in the determination of the hydrodynamic properties of fibrinogen. In addition the simulations suggest how the dynamics of the D region of fibrinogen may contribute to the allosteric regulation of the blood coagulation cascade through a dynamic coupling between the a- and b-holes, important for fibrin polymerization, and the integrin binding site P1.
| Fibrinogen, a protein found in the blood of vertebrates, when activated, aggregates and forms fibrin fibers, the basis of a blood clot. Clots are broken down by the enzyme plasmin, which cuts fibrin fibers at specific places, thus helping the regulation of clot persistence. A mechanistic understanding of fibrin degradation by plasmin is still missing. An important determinant of this process might be the flexibility of fibrinogen. The flexible nature of fibrinogen is reported, for example, by the great variety of conformations observed when fibrinogen adsorbs on material surfaces. However, limits in the spatial resolution of these experiments preclude the identification of the atomistic mechanism behind this flexibility. Here, we perform computer simulations that help identifying with atomistic detail large bending motions occurring at a specific hinge on the molecule. We show how these bending motions can explain the variable conformations observed in experiments and how they help exposing sites where plasmin can cut fibrinogen. Furthermore, our simulations let us identify cooperative effects involving several distant parts of fibrinogen that may play a role in the assembly of fibrin fibers. Both the bending and the cooperative effects, thus, represent potential mechanisms for the regulation of blood clotting.
| Fibrinogen (Fg) is a 340kD multi-chain glyco-protein which can polymerize into fibrin, one of the main components of blood clots. Fibrin formation and lysis (fibrinolysis) are tightly controlled processes along the pathway leading to coagulation [1]. Fg, once activated by thrombin, which cleaves the fibrinopeptide A and B (FpA, FpB), exposes specific A- and B-knobs which bind to the corresponding a- and b-holes of neighbor Fg molecules and initiate the fibrin polymerization process. Fibrin is later stabilized by additional non-covalent and covalent interactions. By further interacting with other blood components through its integrin binding sites, fibrin plays an important role in regulating coagulation and immune response. Fibrinolysis on the other hand is effected by plasmin, which cleaves fibrin on specific cleavage points in a well defined temporal sequence [2–4].
The elongated structure of human Fg, as shown by the crystallographic data [5], is formed by two symmetric units which dimerize through a central globular E region. Each symmetric unit (protomer) is constituted by 3 peptide chains Aα, Bβ and γ which depart from their N-terminal region (E region), form an elongated coiled-coil region, and end into two globular domains forming the D region (Fig 1). The C terminal segment of the Aα chain, i.e. the αC region, as well as the N-terminal parts of chain Aα and Bβ, including FpA and FpB, are mostly disordered (thus, not resolved in the crystal).
Although the available crystallographic structures of Fg show a relatively limited variability, atomic force microscopy images of adsorbed Fg on several surfaces reveal a large degree of conformational flexibility. Indeed, the typical tri-nodular structure of Fg, as observed in adsorption studies, where the three nodules correspond to the two D regions and the central E region, is very variable [6], and the angle formed by the three nodules has a wide distribution [7, 8]. The source of this conformational flexibility at the molecular level is not well understood. Early sequence analysis [9] and comparison of several crystallographic structures of Fg [5, 10, 11] suggested the presence of a hinge point in the middle of the coiled-coil regions connecting the E and D regions. However, the role of this hinge point and the extent of flexibility that it confers to the Fg molecule has not yet been described.
Fg is one of the most abundant serum proteins initially adsorbing on foreign surfaces in contact with blood [12, 13], and it plays a crucial role in determining the inflammatory response to the material [14, 15]. In the case of nanoparticles, which have been the subject of intense research for their use in nanomedicine [16, 17], Fg contributes with other serum proteins, upon pre-incubation in blood, to form a protein corona surrounding the particle and determining its fate in its clinical use, i.e., circulation halftime, cell uptake, etc. [18]. Simplified models of Fg have been developed based on its hydrodynamic [19] and adsorption properties [20, 21] and used to study adsorption on solid surfaces. Similar models have been used to study the competitive adsorption of serum proteins on material surfaces and, in particular, on nanoparticles [22]. In these models, as well as in models of fibrin polymerization [23], the internal flexibility of Fg is either ignored or treated approximately, although it may play a very important role especially in the characterization of its hydrodynamic properties.
Fg, in the polymerised form of fibrin, is a structure subjected to mechanical tension. For this reason early simulation work on Fg focussed on its mechanical properties under external stress [24–27]. Here, instead, we report the results of extensive molecular dynamics (MD) simulations performed on Fg in solution. The simulations allow for the identification of large bending motions centered at a hinge point on the coiled-coil region of Fg. We also present an extensive sequence analysis of Fg across vertebrate organisms which suggests that the bending motions associated with the hinge play one or more functional roles. The simulations indicate that one of these roles may consist in the exposure of plasmin cleavage sites on the coiled-coil region. From the simulation results we construct a simplified representation of the internal flexibility of Fg and use it to fit and explain experimental data on conformational distribution of the molecule adsorbed on mica. The results of the fit point to an asymmetry in the adsorption properties of the different sides of Fg, which can be explained by the presence of large charged patches that are unevenly distributed on the surface of the globular domains of the molecule. In addition, the simulation data allow us to characterize the dynamical properties of the D region of Fg involved in fibrin formation and immune response, highlighting the presence of coupled motions between the a- and b-holes and the integrin P1 binding site.
We have performed several atomistic molecular dynamics simulations of Fg, either in its full dimeric state or considering only one of the two symmetric protomers. In either case we have simulated glycosilated and unglycosilated constructs. Each construct has been immersed in a periodic box of water molecules at physiological ion concentration (see Methods section for details). The cumulative time length of the simulations reaches 1.3 μs with several continuous stretches of simulation reaching 0.2 μs.
Fg undergoes large bending motions in all the simulations that we have performed. Principal component analysis (PCA) is used to quantify these motions. The dominant principal components of motions (PCA modes, Fig 2(a)–2(c)) of the Fg protomer are the same in all sampled trajectories as revealed by a large overlap (see Methods section) between the three dominant modes ranging from 0.6 to 0.9 between simulation subsets. In particular, the large overlap observed between the dominant PCA modes in both mono- and di-glycosilated and unglycosilated Fg protomer trajectories shows that the carbohydrate clusters do not affect the large scale dynamics of Fg in solution. Similarly, the dimerization state does not induce any noticeable change in the large scale motion of Fg. Dimer and monomer simulations show consistent hinge bending and the PCA-mode overlap between isolated protomer and dimerised protomer simulations is large. The difference between the dimer and protomer simulations is limited to the dynamics of the dimerization interface, where the absence of the disulphide bridges with the other protomer results in expectedly larger root mean square fluctuations (RMSF) localized to the residues α27–44 β 58–75 and γ 14–18 (see S1 Fig in Supplementary Information). Because of the overlap of the largest PCA modes in the different simulation sets, the analysis presented here is done using all the available data merged together in a single set, which improves the statistical significance of the results. The first three PCA modes span the degrees of freedom associated with bending at a hinge point in the coiled-coil region (Fig 2(a)–2(c)), while the 4th PCA mode is related to a pure torsion of the coiled coil along its axis (not shown). The motions are reversible as shown by the time series of the PCA projections (Fig 2(d)) Lower ranking PCA modes provide smaller contributions to the overall variance so they will not be analyzed further.
The program DynDom [28], applied to the extremal structures observed along the first PCA mode (Fig 2(a)) of the Fg protomer, has been used to identify the regions of the molecule which are more rigid in our simulations, as well as the connecting hinge regions. DynDom reports the presence of two relatively rigid regions, separated by a hinge. The E region and the N-terminal part of the coiled-coil region represent one of the two rigid domains, while the C-terminal part of the coiled-coil region along with the D region represent the second. The hinge region is located approximately in the middle of the coiled-coil region and includes the break in the α-helical structure of the γ chain, which gives rise to a flexible loop (residues γ70–78), along with the neighbor residues on the Aα and Bβ chains (Fig 2(a)–2(c)). The break in the α-helical structure of the γ-chain is facilitated by two proline residues.
The bending around the identified hinge can be described by a bending angle γ and a torsion angle φ defined using groups of atoms from the E region, the hinge region and the D region (Fig 2(a)). The γ and φ angles strongly correlate with the projections along the dominant PCA modes: γ shows a linear correlation coefficient of 0.96 with a linear combination of the first 3 PCA modes (1st-2nd-3rd/10, see Methods section), while φ has a 0.73 linear correlation coefficient with a linear combination of the 1st, 2nd and 4th PCA modes (1st/3+2nd/2+4th/2). Our simulation data show a consistent and significant bending occurring at the hinge region and reaching bending angles below 90 degrees (Fig 2(e)). The time it takes for the Fg structure to reach a bending angle below 110 deg from conformations similar to the crystal structure (bending angle above 150 deg) is 19±1 ns along the trajectories, averaged over the 12 observed events (see Fig 2(d) and S1 Movie in Supplementary Information, for examples). The reverse process occurs twice in the simulations, taking 20 and 26 ns. The simulations of the full Fg dimer do not show significant correlations between the angle values observed at the two hinges.
Comparison of the crystallographic structures of Fg coiled-coil regions from various organisms already hinted at the presence of a flexible hinge [5]. This hypothesis is also supported by hydrogen-deuterium exchange experiments [29]. The latter are in good agreement with our simulations: amino acids from the coiled-coil region with lower helical probability in the simulations (Fig 3(b)) correspond to amino acids with low protection factors in the experiments. Our simulations help to clarify the fact that the presence of a flexible hinge is an intrinsic feature of the coiled-coil region, and not an artefact due to sequence or crystal-packing differences in the compared crystal structures. The hinge is positioned on the non-helical segment of the γ chain (γ70–78), most probably due to the resulting reduction in the stiffness of the coiled coil. This segment is non-helical also in the other crystallized Fg structures [10, 11]. In addition, this segment has markedly helix-breaking features in most of the available Fg sequences from vertebrates that we have analyzed, showing a large density of proline and glycine residues (Fig 3(a)).
As can be seen in Fig 3(a), the appearance of prolines characteristic for the hinge region happens in three stages during evolution. The lamprey (Petromyzon marinus) has no proline, fish (Danio rerio) have a single proline and tetrapods have two or more. These steps coincide with major changes in the clotting cascade, namely the appearance of the intrinsic pathway in jawed vertebrates and the addition of the contact pathway (linking blood clotting and immune response) in tetrapods [30–32]. We used the program DisEMBL [33] to analyze the Fg sequences and identify disordered or loop segments in the coiled-coil region. The region corresponding to the non-helical segment of the γ chain is marked as a hot-loop with high probability in most of the sequences, with the exception of lamprey Fg which, as mentioned above, is known to have a simpler coagulation mechanisms than the other vertebrates [31] (Fig 3(a)). This analysis supports the idea that the non-helical segment of the γ chain provides a function that is strongly conserved across vertebrates. Our simulations suggest that this conserved function is linked to the bending motion of the coiled-coil region.
Besides providing flexibility to the individual Fg molecules as well as the fibrin fibers [5], the bending at the hinge may help expose the plasmin cleavage sites located nearby on the coiled-coil region, an hypothesis already proposed in ref. [9]. Our simulations strongly support this hypothesis showing that the α-helical structure around the plasmin cleavage sites Aα104–105 and Bβ133–134 is partly disrupted by the bending motions, and the exposure to the solvent of the involved peptide bonds increases (Fig 3(c) and 3(d)). The α-helical structure lost during bending is generally replaced by coil structure in our simulations and not by extended β-sheet structure as observed in experiments [34] and simulations [35] of fibrin subject to tension. The observation of the transition to extended β-sheet may then be linked to the presence of tension along the molecule and/or require significantly larger times than those simulated here to occur spontaneously. It is known that the fibrin molecules straighten when they are integrated into protofibrils [36]. This does not mean that the flexibility of individual molecules is lost. The twisting of fibrin fibers [37] compresses molecules in the center of the fiber and stretches them on the perimeter. If the hinge bending is necessary to accommodate such deformation, it is reasonable to believe that the bending motions at the hinge may actually be reduced by tension applied along the fiber axis. Thus, fibrinolysis assisted by the bending motions at the hinge may selectively take place on fibrin molecules subject to reduced tension. This hypothesis is supported by experimental evidence indicating reduced plasmin fibrinolytic effectiveness on fibrin fibers subject to mechanical tension [38].
Bending may also play an important role in fibrin polymerization. According to the recently proposed Y-ladder model of fibrin polymerization [23], in the early stages of the process, the protofibril grows as a single strand where each monomer is connected to the next by a single A-knob-a-hole bond between the E domain of one molecule and the D domain of the next in the chain, giving rise to a Y-ladder structure. The latter is later transformed in a double-stranded protofibril by the formation of the other knob-hole bonds and D-D bonds. Hinge bending as observed in the present simulations can provide the necessary flexibility to accommodate new molecules in the growing fiber. In support of this role, Fig.6C of ref. [8] seems to depict single stranded growth events involving bent fibrin molecules as indicated by arrows 5 and 6. More in general, an imperfect straightening of Fg during integration into a double-stranded protofibril would provide a mechanism for the formation of branch points in the fibrin network.
In the present simulations, several parts of the molecule have been omitted because they are unresolved in the crystal structure, and too long (and disordered) to be meaningfully sampled in the course of our simulations. These include the the βN stretch (residues β1–57) and the αC stretch (residues alpha 201–562). Both parts are mostly disordered (apart from a tendency to form a compact conformation in the C terminal part of the αC domain) as revealed by several experimental techniques, including H/D exchange [30], X-ray crystallography [5, 10, 11] and NMR [39]. In fibrinogen, the compact domains of the αC region from the two protomers are believed to interact with each other and with the fibrinopeptide B (part of the unresolved βB stretch) in proximity of the E region of the molecule, while in fibrin the alphaC domain are believed to move away from the E domain and take part in inter molecular interactions [40]. Although none of these unresolved stretches are known to interact with the coiled coil region of the protein, in principle they could interfere with the bending movements observed in our simulations. The lack of evidence for interactions with the coiled-coil region and the flexible nature of these stretches suggest that their presence will not directly contribute to a stiffening of the hinges of the coiled-coil region. On the other hand, we cannot exclude that their presence may bias bending along particular directions by hindering hinge movements due to excluded-volume effects.
Fg adsorbed on surfaces often forms tri-nodular structures, as reported by AFM experiments [7, 8]. The three nodules correspond to the E and the two globular D regions of the molecule. The distribution of the α angle formed by the three nodules can be used to quantify the flexibility of Fg [7, 8]. The model for Fg emerging from our simulations shows that the flexibility is provided by the presence of the two hinges, while the rest of the molecule does not undergo large conformational changes. Assuming that adsorption does not significantly change this picture, the model can be used to fit the conformational distributions from adsorption data and test various hypothesis on the behavior of Fg. In particular we want to test whether adsorption induces correlations in the behavior of the molecule at the two hinges, which could not be observed in the solution simulations. Such correlations could then be related to specific interactions between the protein and the surface. To this end, we proposed two different models for data fitting: on one hand we used a model where the conformation of one hinge does not affect the conformation of the other, so that the resulting distribution of the experimentally observed α angle, because of the symmetry of Fg, can be written as:
P ( α ) = ∫ w ( γ 1 , φ 1 )w( γ 2 , φ 2 )P γ 1 , φ 1 , γ 2 , φ 2 ( α )d γ 1 d φ 1 d γ 2 d φ 2 (1)
where γ1∣2 and φ1∣2 are the bending and torsion angles at the two hinges, respectively (see Methods section), Pγ1,φ1,γ2,φ2(α) is the distribution of the α angle of adsorbed Fg molecules with given hinge angles γ1, φ1, γ2, φ2, and w(γ, φ) is the statistical weight of the hinge angle pairs. On the other hand, we proposed a more general model which includes possible correlations between the two hinges:
P ( α ) = ∫ W ( γ 1 , φ 1 , γ 2 , φ 2 )P γ 1 , φ 1 , γ 2 , φ 2 ( α )d γ 1 d φ 1 d γ 2 d φ 2 (2)
where W(γ1, φ1, γ2, φ2) is the combined statistical weight of the conformations of the two hinges. The distributions Pγ1,φ1,γ2,φ2(α) were determined using a simple Monte Carlo model (MC) for Fg adsorption (see Methods). The statistical weights w(γ, φ) and W(γ1, φ1, γ2, φ2) for the two models need to be determined by fitting the experimental data. For simplicity, we transformed the integral in Eqs (1) and (2) in a sum over discrete bins defined in the γ − φ space, so that the weight functions w(γ, φ) and W(γ1, φ1, γ2, φ2) become discrete arrays where each element is the statistical weight of the corresponding bin. The fit is done using a maximum entropy approach, where the statistical weights from our simulations (Fig 2(e)) are used as prior knowledge (see Methods for details).
Both models have been used to fit the conformational distribution of Fg adsorbed on mica as observed in AFM experiments [7]. With the independent hinge model, the χ2 of the fit remains above the threshold of 5% confidence level for the given number of degrees of freedom, indicating a poor fit. On the other hand, the general model fits the data very well (see Fig 4(a)). More generally, it is possible to show that the decoupled weight w(γ1, φ1)w(γ2, φ2) of the independent hinge model (Eq 1) would give rise to P(α) distributions peaked at 180° for any weight function w(γ, φ), which is incompatible with the experimental evidence of a deep trough at 180° [7](Fig 4(a)). To see this, we can initially fix the φ angles of the two hinges to the same value so that the E and D domains are on the same plane and we can treat the problem as two-dimensional and get a simple analytical expression of α as function of γ1 and γ2. Then, we can focus on hinges with the same γ angle distribution w(γ, φ0). A single maximum in this distribution will clearly lead to conformations where both hinges sample the same γ angle, which, thanks to the symmetry of the molecule, results in an α angle of 180°. But also distributions w(γ, φ0) with two (or more) equal maxima, will still lead to α-angle distributions with an absolute maximum at 180°, because of the contributions coming from the conformations where the two γ angles sample the same maximum and because the peak at 180°, which has only one tail due to being at the end of the definition interval of the α angle, will grow twice as big as the other peaks due to symmetrical contributions cumulating on the single tail of the peak (see S3 Fig in Supplementary Information, for an illustrative example). Releasing the constraint on φ will not change the situation. The slight drop on the expected counts of the independent hinge model fit at 180° in Fig 4(a) is within the numerical uncertainties introduced by binning the MC data.
An analysis of the fitted parameters of the general model, reveals that the largest contributions to the tail of the distribution (95° < α < = 110°) corresponding to the lower gaussian peak identified in ref. [7]) come from hinge angle pairs with 60° < γ1 ≤ 100° and 100° < γ2 ≤ 140°, that is from fg molecules with one strongly bent and one moderately bent hinge. On the other hand the largest contributions to the 160° peak come from hinge angle pairs with 100° < γ1 ≤ 140° and 140° < γ2 ≤ 180°, that is fg conformations with one moderately bent and one almost unbent hinge.
The failure of the independent hinge model and the success of the general model to fit the observed conformational distribution of Fg adsorbed on mica indicates that the conformation of Fg at one hinge affects the conformation at the other hinge upon adsorption on mica. The most probable explanation for this effect is that the propensity to form interactions with mica is not distributed uniformly on the Fg surface. The optimization of the contact surface with mica on both symmetric protomers possibly induces hinge correlations. Since the correlation is observed on a mica surface, which is charged, it is reasonable to assume that the difference between the sides of the Fg protomer has an electrostatic origin.
We have verified the presence of asymmetrically distributed charged patches on the Fg surface by calculating the electrostatic potential generated by the molecule along the simulations. As shown in Fig 4(b), the calculations identify a large negatively charged patch per protomer located on one side of the D region but absent on the other side. This asymmetrical distribution supports our hypothesis regarding the origin of correlations between the two Fg hinges. It has previously been noted that such patches should contribute to the Fg-Fg association during fibrin fibril formation [41]. Here we can identify the D region patch as the a-hole binding site for the fibrinogen A-knob described in reference [42]. Evidently, the involvement of this part of the molecule in interactions with the adsorbing surface may have consequences with respect to fibrin formation.
Unlike the data from Fg adsorption on mica, the data from Fg adsorption on highly oriented pyrolytic graphite (HOPG) rendered hydrophilic with an amphiphilic carbohydrate-glycine modifier(GM) [8] can be fitted with both the general and the independent hinge model at a confidence level larger than 5% (see S4 Fig in Supplementary Information). This may be due to the differences in the conformational distribution induced by a different surface (for example, the 180° bin on HOPG is significantly more populated than on mica, while the tail of the distribution at small α angles is less populated than on mica) or due to the lower number of bins/histograms used to report the experimental data on HOPG relative to those on mica in ref. [7] and the overall lower number of data binned, both of which provide less stringent constraints to the fit.
Before closing the subsection we would like to note that, the simplified model of Fg flexibility, emerged from the simulations and used in this section to fit experimental data, may help to refine the models used for describing fibrinogen hydrodynamical properties [19], in the context of adsorption on material surfaces and nanoparticles [20–22], as well as fibrin polymerization [23].
Fibrinogen has a multitude of binding partners which help the molecule to carry out its functions. In addition to the plasmin cleavage sites on the coiled-coil region, discussed above, the D region of Fg hosts several other functional binding sites. Two important binding sites are the a- and b-holes, on the γC and βC subunit of the D region, respectively, which bind the A- and B-knobs exposed by thrombin upon cleavage of the respective fibrino-peptides. As mentioned already, these binding sites are important for fibrin polymerization as they help to create lateral non-covalent connections between fibrinogen molecules. Surface plasmon resonance experiments showed that binding of fibrino-peptide analogs to the b-hole increases the binding affinity between the a-hole and soluble fibrin fragments containing the A-knob [43]. Hydrogen-deuterium exchange experiments comparing wild-type Fg with the Bβ235Pro/Leu mutant, showed that the mutation, not only leads to a local increase of the flexibility at the βC-γC interface, but also alters the flexibility of the loops surrounding the a-hole on the γC domain [29]. In addition, it has been suggested that engagement of the B knobs into the b-holes produces a subtle domain rearrangement in the D regions, favoring lateral aggregation of protofibrils [40]. All these information, thus, suggest that allosteric effects can take place in the D region.
Our simulations, analyzed using principal component analysis restricted separately to the two globular subunits (γC and βC), show that the loops surrounding the a- and b-hole (residues γ354–363, γ293–302 and β382–393, β422–432) [44] undergo large correlated fluctuations, as demonstrated by the large weight of those loops in the largest modes of the PCA. To better clarify this aspect, we verified that the dynamics of the a- and b-holes are correlated by the presence of a high correlation pathway (measured using the linear mutual information (LMI) with rcrit ≤ 0.85, see Methods section) connecting them. At the same critical LMI level, several other distant pairs of residues can be connected on the D region including parts of the P1 [45] and P2 [46] integrin binding sites on the γC domain. All the identified pathways pass through the same bottleneck of residues at Bβ204Pro, γ216Gly, γ217His, γ225Glu. More than 75% of all pathways pass through the residues γ200Gly, γ253Trp and γ348Tyr. The identified narrow pathway is shown in Fig 5(a). Some of these residues are in contact with the residues locally perturbed upon the Bβ235Pro/Leu mutation, which may explain why the effects of this mutation reach the a-hole. With the exception of γ217His all residues belonging to the bottleneck of the high correlation pathways are absolutely conserved among vertebrates (sequences discussed above). In addition, the same residues appear to be associated with disease-inducing mutations in humans [47]. These data, taken together, reinforce the hypothesis of an allosteric network connecting the a- and b-holes.
The integrin binding sites represent another important element of the functionalities of Fg, involved in facilitating immune response. Several residues of the γC domain have been implicated in integrin binding [45, 46]. Our simulations show that the accessibility of the P1 region of the integrin binding site is affected by the large scale fluctuations of the D region and of the loops around it. Indeed the three largest PCA modes of the D region involve rigid rotations of the βC relative to γC domains, which directly alter the P1 accessibility. In addition, these modes have large components on two loops surrounding the P1 binding site, γ354–363 and β280–285, whose coordinated fluctuations significantly affect P1 accessibility. The opening/closing mechanism of the P1 cleft described by the largest of the PCA modes is characterized by a change in distance between the integrin binding site P1 and the βC loop B280-B285 from about 1.2nm to 1.9nm (Fig 5(b)).
Extending the PCA analysis to include also the C-terminal segment of the coiled-coil region (residues α125–189, β155–458 and γ100–394) reveals that the relative rotations of the βC and γC domains described above are associated with movements of the C-terminal segment of the coiled-coil region relative to the D domain. In particular, the second largest PCA mode describes a motion where the distance between the βC domain and the coiled-coil region increases in combination with an opening of the b-hole (Fig 5(c)). This mechanism recalls very closely the one which has been hypothesized to explain how the interaction between B-knob and b-hole induces a slight conformational change which favors the lateral aggregation of protofibrils [40].
The classical atomistic molecular dynamics simulations of fibrinogen in solution reveal the extraordinary flexibility of the molecule resulting in large bending motions of the coiled-coil regions, favored by the presence of two hinges. The hinges are linked to a non-helical segment in the coiled-coil region of the γ chain, a feature conserved across vertebrates. This may indicate a possible functional role for the bending motions. In the simulations, the bending of the coiled-coil region helps to expose the early plasmin cleavage sites. A simplified model of Fg flexibility has been derived from the simulations and used to test hypothesis about Fg adsorption by fitting AFM data. This lead to hypothesize correlations between the two hinges upon adsorption on mica. A probable cause for the correlations is an asymmetric distribution of charged patches on the surface of the molecule and in particular on the D globular regions, which is observed in the simulations. We anticipate that the simplified model presented here could lead to more accurate estimates of the hydrodynamic properties of Fg. Furthermore, an analysis of the pathways joining residues with highly correlated motions in the simulations hints at an allosteric regulation of the binding at the a- and b-hole in the D region of Fg.
The simulations are based on the crystal structure of human Fg (PDB ID: 3GHG) [5]. The carbohydrate groups that are only partly resolved in the crystal have been modelled using VMD and introduced in some of the simulations. The unresolved parts of the protein structure (the αC domain and the N terminal segments of all the chains) have not been included in the calculations. Several molecular constructs have been prepared to assess the role of the different components of the Fg molecule. The effects of the carbohydrate chains on the dynamics of Fg have been investigated by simulating both the unglycosylated system, the sytem glycosylated at residue β364 (mono-glycosylated) and the system glycosylated at residue β364 and γ52 (di-glycosylated). Protomer-protomer interactions were investigated by simulating both the full fibrinogen dimer (6 protein chains) and the protomer system (3 chains). Although the single or isolated protomer has never been observed experimentally, the reason for studying it are the following: 1) the two protomers in the dimer are identical, i.e. they will show a similar behavior, 2) since we neglect the unresolved parts of the molecule, the two protomers interact only through the small dimerization interface, i.e. their reciprocal influence is limited, 3) A ring of disulphide bridges covalently bonding the three chains in the N-terminal part of the coiled-coil region dramatically reduces the influences of the dynamics of the dimerization domain on the rest of the molecule (see S1 Fig in Supplementary Information). In addition to that, the simulations of the protomer take much less computational time than the dimer i.e. they can be extended to significantly larger time scales.
Rectangular simulation boxes with explicit TIP3P water [48] and physiological ion concentration (150 mMol [NaCl]) were prepared using VMD [49] (Table 1 for box sizes).
Isobaric-isothermal simulations were set up at a temperature of 310K and pressure of 1atm using NAMD [50] with a Langevin thermostat and a Langevin piston barostat [51, 52] using 200 ps−1 and 100 ps−1 as decay time, respectively. The covalent bonds involving hydrogen atoms were fixed in length and a 2fs timestep was used. The CHARMM22 force field with CMAP corrections [53] was used with its recent extension to carbohydrates [54] in combination with ParamChem (http:/www.paramchem.org) and the CHARMM generalized force field (CGenFF) [55]. This force field has been already tested in a large variety of systems and found very reliable in reproducing several biophysical properties, including also the folding process of proteins [56] where it was shown to provide results very similar to other popular force fields (several versions of AMBER and modifications of CHARMM) in the characterization of the native state of proteins. The van der Waals forces were cut off at 1.2nm while PME was used for long range electrostatic interactions with a grid spacing of 1Å. After energy minimization (NAMD’s conjugate gradient algorithm, 15000 steps) of hydrogen atoms and water molecules, the system was heated and equilibrated for 10ns. Production runs statistics are given in Table 1. We employed collective variable constraints (distanceXY, directiondir or orient) to keep the main axis of the molecule aligned to the simulation box and verified that this had no influence on the overall dynamics by comparing to unconstrained simulations.
To identify the collective motions of the whole Fg molecule and of its subdomains we performed several principal component analyses (PCA) [57] using wordom [58] and GROMACS utilities [59]. DynDom [28] was used to identify rigid domains and hinges of motion. The overlap between spaces spanned by the dominant PCA modes of different simulations was used to quantify the similarity of the observed dynamics [60]. The overlap is defined as:
O ( { x _ i } , { y _ i } ) = 1 n ∑ i = 1 n ∑ j = 1 n ( x _ i · y _ j ) 2 , (3)
where { x ¯ i } and { y ¯ j } are the two subspaces spanned by the principle components x ¯ 1 … x ¯ n and y ¯ 1 … y ¯ n, respectively.
The linear correlation coefficient between two variables is defined as the ratio between the covariance of the two variables and the product of the two standard deviations, r = ∑ ( x i − x ‾ ) ( y i − y ‾ ) / ∑ ( x i − x ‾ ) 2 ( y i − y ‾ ) 2. If P1(t), P2(t), PN(t) are the projections of the trajectory at time t along the 1st, 2nd and Nth PCA components, respectively, a linear combination CP(t) of PCA projections is obtained as CP(t) = ∑j aj Pj(t), where aj are the coefficients of the combination. As mentioned in the Results section, a very large correlation coefficient (0.96) between CP(t) and the γ angle was obtained by using a1 = 1, a2 = −1 and a3 = −0.1 (and all the other aj = 0).
Apart from PCA, as a further measure of correlation between the movement of the residues, the linear mutual information (LMI) [61] was used which is defined as follows:
L M I ( x i , x j ) = 1 2 ( ln [ det C i ] + ln [ det C j ] - ln [ det C i j ] ) (4)
where xi is the position of the ith Cα atom, C i = ⟨ x i T x i ⟩ and Cij = ⟨(xi − ⟨xi⟩)(xj − ⟨xj⟩)⟩
The LMI matrix was calculated for mono-glycosilated fibrinogen simulations. Based on it, a network was built with each node representing a residue. Nodes are connected if their LMI exceeds a threshold rcrit. This parameter was successively reduced until a pathway between the a- and b-holes was identified at rcrit = 0.85. The shortest path connecting the two binding pockets is determined in this network. To save computation time only pathways of a maximum length N where considered such that r c r i t N > 0 . 025. As a control, pathways between all residues that are at least 6nm apart were identified.
The electrostatic potential of the globular domains of fibrinogen was calculated by solving the Poisson-Boltzmann equation with the APBS [62] software over a series of similar aligned structures. Then, the electrostatic potential was averaged over the structures. The time averaged potential for each domain was calculated by averaging the potential at each grid point over all snapshots.
Sequences of Fg were identified by searching the UniProt data base [63]. In this search, sequences of 83 different species were identified that contained the γ chain at least up to the hinge region. Complete sequences of all three chains in the coiled-coil region where found for 33 species. Sequence alignments were performed using Chimera [64], and the analysis of structural disorder was performed using the program DisEMBL [33].
A simplified representation of Fg adsorption has been developed to map Fg conformations represented by the hinge angles to the conformations observed in AFM experiments, characterized by the α angle between the globular domains (see Fig 6). The model for Fg is built around a central rod, representing the stiff coiled-coil regions, which connects the two hinges. The E region is represented as a sphere placed at the center of the rod. The D regions are also represented as spheres connected to the hinges with rods, that can pivot around the hinge. The dimensions of the model components are extracted from the crystal structure (Fig 6). An additional point (x0 in Fig 6) on the surface of the E region is used as a reference for the measure of the torsion angles at the two hinges. Within this simplified representation, adsorption occurs if and only if the distance of the E and the two D regions of a Fg molecule from the adsorbing surface (represented by an ideal plane) is lower than a threshold hmax and both hinges lie above the surface.
We use a simple Monte Carlo (MC) algorithm to generate a large set (4 ⋅ 107) of coarse grained adsorbing Fg conformations. The MC algorithm consists of creating adsorbed conformations by randomly drawing hinge angles for one protomer and placing it such that the first D region as well as the E region contact the surface. The bending angle γ for the second hinge is randomly drawn as well, while the value of φ for the second hinge is calculated to ensure that the second D-region is in contact with the surface. The resulting conformations are accepted in such a way that a uniform distribution of γ and φ is obtained at both hinges. For each accepted conformation we measure the hinge angles, as well as the α angle (Fig 6) between the globular regions, as it would be observed in an AFM experiment, i.e., we project the regions’ centers on the surface and measure the angle between the three points.
To fit the experimentally observed distribution of the α angle we divided the γ − φ-plane characterizing the conformation of one hinge into 12 rectangular bins, indexed accordingly. Each adsorbing Fg conformation can be characterized by the two discrete indexes i and j of the (γ, φ) region where its two hinges occur. We then determine the bending distribution Pij(α) observed in the conformations sampled with MC from each pair of bins i and j of the (γ–φ) plane (see S2 Fig in the supporting information). It is important to note that, not all ij pairs lead to adsorbing structures. We exclude the non-adsorbing ij pairs (bins) from the computation. We then assume that the observed experimental distribution P(α) is a superposition of the distributions from the adsorbing conformations of Fg in each pair of bins. The most general form of superposition is given by P(α) = ∑ij aij Pij(α), which is the discretized version of Eq 2, where the parameters aij replace the continuous weights W(γ1, φ1, γ2, φ2). The aij parameters can then be fitted to the experimental distribution. If the two hinges behave independently, then, given the symmetry of the molecule, an independent hinge model P(α) = ∑ij ai aj Pij(α) should be able to fit the data. This model is the discretized version of Eq 1, where the parameters ai replace the continuous weights w(γ, φ). In this case, we will have only the ai parameters, i.e., one parameter for each bin in γ − φ-space.
Both the general and the independent hinge model are fitted to the experimental data using the maximum-entropy method [65], where the Shannon entropy of the parameters is maximized with respect to the a priori knowledge of the system, with the constraint that the reduced χ2 = ∑α(N(α) − Ntot P(α))2/Ntot P(α) equals 1, where N(α) are the histograms of the α angle from the experiments, and Ntot is the sum of the histograms. The expression that is maximized for the general model is:
J = - ∑ i j a i j log ( a i j / m i j ) - λ( χ 2 - 1 ) - α( F - 1 ) (5)
where mij is the bias distribution for the parameters, F is the normalization function of the parameters F = ∑ij aij and λ and α are Lagrange multipliers to impose the constraints on χ2 and normalization. A similar expression is used for the independent hinge model with the necessary modifications. The bias distribution mij represents the information already known about the parameters before fitting the data. The statistical weight wi of each γφ bin as measured in the atomistic simulations (Fig 2(e)) were used to estimate the mij = wi wj.
|
10.1371/journal.pcbi.1003276 | Novel Methods for Analysing Bacterial Tracks Reveal Persistence in Rhodobacter sphaeroides | Tracking bacteria using video microscopy is a powerful experimental approach to probe their motile behaviour. The trajectories obtained contain much information relating to the complex patterns of bacterial motility. However, methods for the quantitative analysis of such data are limited. Most swimming bacteria move in approximately straight lines, interspersed with random reorientation phases. It is therefore necessary to segment observed tracks into swimming and reorientation phases to extract useful statistics. We present novel robust analysis tools to discern these two phases in tracks. Our methods comprise a simple and effective protocol for removing spurious tracks from tracking datasets, followed by analysis based on a two-state hidden Markov model, taking advantage of the availability of mutant strains that exhibit swimming-only or reorientating-only motion to generate an empirical prior distribution. Using simulated tracks with varying levels of added noise, we validate our methods and compare them with an existing heuristic method. To our knowledge this is the first example of a systematic assessment of analysis methods in this field. The new methods are substantially more robust to noise and introduce less systematic bias than the heuristic method. We apply our methods to tracks obtained from the bacterial species Rhodobacter sphaeroides and Escherichia coli. Our results demonstrate that R. sphaeroides exhibits persistence over the course of a tumbling event, which is a novel result with important implications in the study of this and similar species.
| Many species of planktonic bacteria are able to propel themselves through a liquid medium by the use of one or more helical flagella. Commonly, the observed motile behaviour consists of a series of approximately straight-line movements, interspersed with random, approximately stationary, reorientation events. This phenomenon is of current interest as it is known to be linked to important bacterial processes such as pathogenicity and biofilm formation. An accepted experimental approach for studying bacterial motility in approximately indigenous conditions is the tracking of cells using a microscope. However, there are currently no validated methods for the analysis of such tracking data. In particular, the identification of reorientation phases, which is complicated by various sources of noise in the data, remains an open challenge. In this paper we present novel methods for analysing large bacterial tracking datasets. We assess the performance of our new methods using computational simulations, and show that they are more reliable than a previously published method. We proceed to analyse previously unpublished tracks from the bacterial species Rhodobacter sphaeroides, an emerging model organism in the field of bacterial motility, and Escherichia coli, a well-studied model bacterium. The analysis demonstrates the novel result that R. sphaeroides exhibits directional persistence over the course of a reorientation event.
| The motile behaviour of bacteria underlies many important aspects of their actions, including pathogenicity, foraging efficiency, and ability to form biofilms. The study of this phenomenon is therefore of biomedical and industrial importance, with implications in the control of disease [1] and biofouling [2]. Owing to their small size, bacteria inhabit a world of low Reynolds number, in which viscous forces dominate over inertia [3]. Rotational Brownian motion prevents them from swimming continuously in a straight line, hence many motile species such as the multiflagellate Escherichia coli move in a series of approximately straight ‘runs’, interspersed by reorientating ‘tumbles’ in a process known as taxis [4]. During a run, the flagellar motors in E. coli turn counter-clockwise, causing the helical flagella to form a rotating bundle that propels the cell forward. Tumbles are caused when one or more motors reverse their rotation, which disrupts the flagellar bundle and causes the cell to reorient randomly [4]. A related motile mechanism exists in the uniflagellate bacterium Rhodobacter sphaeroides, in which reorientations are, instead, effected by stopping the flagellar motor [5]. Upon ceasing to rotate, the single sub-polar flagellum [6] undergoes a change of conformation, leading to reorientation by a mechanism that is not yet well understood [7]. The biochemical pathways responsible for chemotaxis in R. sphaeroides are less well studied than those in E. coli, and are known to be more complex [8].
The tracking of bacterial cells, as imaged under a microscope, is a well-established experimental technique for investigating bacterial motility. Such studies have been used to gain biological insight in the case of E. coli [4], [9], Pseudomonas putida [10], Rhizobium meliloti [11], Vibrio alginolyticus [12] and R. sphaeroides [13]. A limitation of cell tracking is that a large number of tracks are required in order to ensure that any inferences drawn from observations are statistically representative of the population. Tracking experiments are therefore often laborious [14]. Earlier experiments involved tracking a single bacterium at a time, either in a fixed field of view [13], or by mechanically shifting the microscope stage to keep the cell in focus [4]. This approach suffers from subjective bias as the experimentalist is required to select which cells to track [14]. More recently, simultaneous multiple target tracking has enabled the measurement of tracks from all bacteria visible in the field of view at any given time [15]. This improves the efficiency of the experimental technique, allowing larger datasets to be obtained. It also reduces sampling bias, as all cells in the field of view are tracked. An experimental method related to tracking is differential dynamic microscopy (DDM), which enables the measurement of the distribution of swimming speeds and the fraction of motile cells in the observed population [16]. DDM records these statistics across very many bacteria, however it is an ensemble method and does not permit the measurement of the motile properties of individual bacteria.
Having acquired experimental tracking data, these must be analysed in order to extract quantities of interest. These include the distribution of swimming speeds [9], [13], [16]–[18], various measures of trajectory curvature [19], [20], turning angles [4], [10], the frequency of reorientations [18], [21], [22] and the extent of accumulation near a surface [23]. The ability to obtain such statistics permits quantitative investigations into the response of bacterial populations to environmental stimuli, in addition to cross-species comparisons and the true variability across a population. The analysis method used to extract statistics of motion from the raw data must be robust to errors in the tracking protocol, for example when cell trajectories intersect and the wrong paths are joined [4], and experimental noise such as errors in finding the centre of a cell. In order to identify reorientation events in bacterial tracks, both manual analysis [9], [22], [24] and heuristic arguments [4], [10], [18], [21], [25], [26] have been used. The former is prohibitively time-consuming when dealing with large datasets and is subjective. Automated heuristic methods may be effective in some cases, however it is important to validate such methods, and to avoid the introduction of systematic bias. To our knowledge, all existing heuristic methods require one or more threshold parameters to be specified. The process of selecting optimal threshold parameters may be automatable, as is the case with the method we use for comparison in our study, however this is not a straightforward task and in most cases no guidelines are given as to how to select optimal values for threshold quantities. For example, the method used by Amsler [21] requires the user to specify a threshold inter-frame angular velocity, above which the bacterium is said to be in a reorientation phase. Furthermore, of all the cited studies, only that of Alon et al. [18] includes an analysis of the sensitivity of the results to the various threshold parameters.
Here, we present novel methods for the automated, non-parametric analysis of large bacterial tracking datasets, based on a two-state model of the observed motion, which is compatible with any form of motile behaviour that is well-approximated by the run-and-stop or run-and-tumble models of motion. The data considered in this study are two-dimensional tracks, but the extension of the methods to three dimensions is straightforward. Our methods take advantage of the availability of non-chemotactic and non-motile mutants to gain empirical knowledge of the appearance of running and stopping phases in the observed motion. The methods are based on a modification to the hidden Markov model (HMM), and are applicable to any bacterial species where such mutants exist and sufficiently long reorientation events are discernible using video microscopy. In addition, we suggest a straightforward method that is applicable in the absence of a non-motile mutant. We use a simulation study to assess the performance of the new methods, and compare them with a heuristic approach. To our knowledge such a systematic comparison of methods has not previously been attempted in this field. In order to demonstrate the wide application of our methods, we apply them to analyse novel R. sphaeroides and E. coli datasets, acquired using a recently developed tracking protocol [27]. We show how our new methods enable us to determine the previously unreported distribution of angle changes during a reorientation in R. sphaeroides, amongst other characteristics of the observed motion.
Bacterial tracks of R. sphaeroides and E. coli were acquired as detailed in Materials and Methods. Figure 1 shows a cartoon illustration of a single track. A bacterium swims in an approximately straight line, enters an approximately stationary stopped phase for some time, then swims off in a new direction. The crosses indicate observations made of the cell centroid at regular intervals, (videos are typically captured at 50 frames per second). The primary focus of this study is the identification of stops as illustrated in Figure 1. This task is complicated by various sources of noise in the data. These include: (i) uncertainty in the position of the centroid of a cell in each image that may cause a track to appear jagged, for example when the cell body rotates whilst swimming; (ii) Brownian buffeting that may also cause departures from straight-line swimming, and lead to stops that are not perfectly stationary; (iii) tracking errors caused by incorrectly linking cells between consecutive frames, or by the disappearance of a cell for one or more frames, that may affect the appearance of a track. The identification of stopping phases in tracks is therefore a challenging process.
Each track generated by the tracking procedure is represented in the form , where designates a two-dimensional position vector at time , and the number of frames in the track is given by . Note that is considered a discrete quantity throughout, as time is measured in numbers of frames. In characterising running and stopping phases, we are concerned not with the positions of cells in each frame, but with the motion of cells between consecutive frames. The information of interest is thus the transitions between consecutive position vectors within a track. These form a list of displacement vectors, with . The framewise speed is defined as the observed speed of travel between two consecutive frames, , where denotes the Euclidean norm. The angle changes between consecutive vectors, henceforth called framewise angle changes, are defined so that gives the difference in polar angle between and .
We assume a two-state model of cell motility, in which each displacement vector, , corresponds to either a running or stopping state. The underlying state at time is denoted , where we use the convention throughout that corresponds to a stop and corresponds to a run, hence for each track a state vector describes the sequence of states. We wish to assign to each displacement vector a probability of being in a running phase, . Note that, since we assume a two-state model, we have .
We use our methods to analyse tracking data from R. sphaeroides and E. coli. In each case, data are obtained from three strains: a wildtype strain, which undergoes discrete running and reorientation phases, a non-chemotactic strain, which is always in the running phase and exhibits no reorientation events, and a non-motile strain, which is unable to propel itself.
There is no well-established gold standard for identifying reorientation events in bacterial tracks; indeed several tracking studies make no attempt to extract quantitative information about the reorientation events in tracks [15], [28], [29], while others use ensemble measures such as angular velocity as a proxy for the rate of reorientation [24], [30]. Various heuristic methods requiring the specification of one or more threshold parameters have been used in tracking studies in bacteria (see the related discussion in Introduction). In this study we compare our methods with that of Taboada et al. [25], which is sufficiently versatile to apply to our current data with little modification. This is henceforth denoted the heuristic method. The focus of the present work is the development and validation of our novel analysis methods, however we note that several other heuristic methods mentioned above may be applicable providing it is possible to automatically optimise the various threshold parameters involved. We do not consider these further as a complete survey of methods is beyond the scope of this paper.
We now describe the heuristic method and the two novel analysis methods considered throughout the rest of this work. In addition, we describe a ‘post-processing’ heuristic that can improve the performance of all of the methods and is particularly effective when combined with the heuristic method.
Prior to applying the heuristic method and our two novel methods to experimental data, we must evaluate and compare their ability to correctly infer stop phases in tracks affected by various levels of noise. A traditional means of evaluating this performance is to compare with the results of manual assignment of stopped phases in real tracks. This approach suffers from several key drawbacks, however. Manual tracking is a time-consuming and often difficult process; the stopped phases in microscope videos are by no means easy to discern unambiguously by eye. In addition, manual assessment of tracks is unavoidably subjective.
Here we use an alternative approach to manual analysis: a simulation study. This is a common means of assessing the performance of automated analysis methods [34], [35], [38]. We assume that experimentally-obtained wildtype tracks are the result of a run and stop velocity jump process [39]. Cells in the running phase travel in straight lines with a constant speed drawn from a Weibull distribution that closely approximates the observed non-chemotactic running speed distribution. After a random, exponentially distributed time interval with mean , cells enter a stopping phase and their speed is set to zero. Cells stop for a random period of time, exponentially distributed with mean , after which they switch to the running phase again with a new, Weibull distributed run speed. A new direction of travel is drawn at each reorientation event from the circular uniform distribution. We also simulate tracks describing the non-chemotactic mutant, in which no reorientation events occur, and the non-motile mutant, which is always in the stopped state. We define the sampling interval to be to match the frame capture rate of the microscope used to obtain experimental movies. We simulate 500 tracks for 250 frames each using the parameter values and . These mean duration values are in close agreement with previous studies of E. coli [4], while the remaining simulation parameters have been chosen to match the experimental protocol used to acquire tracks in this study (see Materials and Methods).
We include a simplified model of the noise in the system by adding a normally distributed perturbation to each coordinate of every recorded position in a track, with zero mean and variance equal to , where is varied to modulate the level of noise applied to the system. A random selection of simulated tracks with varying levels of noise are shown in Figure S5. We note that the use of uncorrelated Gaussian noise to simulate the type of noise exhibited in real experimental data may be an oversimplification, however the nature of the noise present in such cases is unknown and beyond the scope of this study. The true underlying state sequence in the simulations, which is continuous in time, is recorded for later comparison with the state inferred by the analysis methods. In carrying out the steps required to analyse the simulated datasets and compare their performance, we attempt to mimic as closely as possible the process that we use when analysing real data (see Figure 2). We infer the values of all model parameters based on the three simulated datasets; none of the parameters of the true underlying processes are known to the analysis methods.
Before commencing the simulation study, we verify that the methods do not produce spurious results when applied to tracks generated from an incompatible underlying model of motion. This test is carried out by analysing tracks from a non-chemotactic simulated dataset. Such tracks contain no stops; the aim of this initial test is to ensure that the analysis methods do not infer stopping phases falsely. In practice, we find that the optimisation routine fails to find a MLE for the transition rate parameters because the negative log-likelihood is independent of the parameter (see Figure S6 and Text S1 for details). This indicates that the HMM-based methods cannot be applied blindly to tracks that contain no stops.
Figure 4 illustrates the MLE values and 95% two-tailed confidence intervals of the mean running and stopping durations, and , respectively, for a range of values of the noise level, . When the level of added noise is low, the two parameters are estimated correctly by both methods. The MLE value of is overestimated by around 20% by both methods in the absence of noise. In the case of the full HMM method, the MLE value decreases with increasing noise level, which initially causes the estimate to become more accurate. At the highest noise level considered here, the MLE is around 60% of the true value. In contrast, the speed-only method MLE increases with noise level. At the highest noise level, the MLE is around double the true value. The full method estimates the value of accurately throughout the range of noise levels considered, whereas the speed-only method increasingly overestimates the same parameter as the noise level increases. At the highest noise level, the speed-only MLE is around threefold greater than the true value. Since the noise model incorporated in our simulations may differ from the sources of noise in the experimental tracks, the precise quantification of the error in the MLE is not of real interest here. However, this result suggests that parameters estimated from highly noisy data may be unreliable, and that the full HMM method generally provides better estimates.
All of the analysis methods output a run status vector for each track, which is discrete in time. The true underlying state path is, by contrast, continuous in time. In order to facilitate a comparison between the inferred state sequence and the ground truth, we discretise the ground truth over intervals of duration . Any such interval that contains part of a stop phase is designated a stop in the discretised true state sequence. The inferred state sequence is a series of stopping phases and running phases, with the convention that an inferred stop corresponds to a positive result. A false positive (FP) therefore corresponds to an inferred stopping phase where none is present in the true underlying state sequence, while a false negative (FN) corresponds to an inferred running phase where none is present in the true underlying state sequence. Figure 3 illustrates this; compare the true, discretised run status with the inferred run status. There are several discrepancies. A stop lasting two frames is inferred at the start of the track, where none is present in the true state. This is a FP; there is another at around . Conversely, at approximately a true stopping event is missed by the analysis method. This is a FN. As noted previously, the application of the post-processing method with and both greater than one corrects the second FP. For each level of added noise, we compute the mean rate of FPs and FNs as the ratio of the total number of FPs and FNs to the total number of actual stop events in the true underlying state. This is computed as the average over all tracks in the simulated dataset.
Figure 5(a) shows the mean FP and FN rates produced by the three analysis methods. In the case of the heuristic method, we test the results with and without post-processing with . The application of post-processing made no significant difference to the results from the HMM methods (data not shown). A FP rate of one means that the average number of false stops equals the number of true stops, while a FP rate of zero indicates that no FPs are observed. The heuristic method is highly sensitive to low levels of noise, generating significantly higher FP rates than the methods based on the HMM. The heuristic FP rate is reduced somewhat by the application of post-processing, however it still remains significantly higher than either of the HMM methods. The full HMM method has a higher FP rate than the speed-only method, though the discrepancy only becomes large when . The speed-only method has an approximately constant low FP rate throughout the full range of noise levels considered here. In contrast, the speed-only method generates the largest FN rate, with the full HMM and heuristic methods exhibiting a similar, lower FN rate. These results suggest that the full HMM method is better able to identify stops, with the disadvantage that it is also more sensitive to noise and more prone to false positives. On the other hand, the speed-only method detects fewer stops, but makes fewer false declarations.
We further assess the accuracy of the HMM methods in Figure 5(b) by plotting the histogram of all inferred angle changes over the course of a stopping phase (henceforth denoted stopwise angle changes), overlaid with the histogram of stopwise angle changes due to FPs. We use a simulated dataset with an intermediate level of additive noise () for this purpose, as this is similar to the value of the translational diffusion coefficient estimated from the experimental data (approximately ; see Figure S10 and Text S1). The result changes very little for noise levels up to (data not shown). The true underlying distribution of stopwise angle changes is uniform. This figure shows that FPs tend to produce small stopwise angle changes, which introduces some bias into the process. However, the number of FPs is low and the bias is not significant over a range of intermediate noise levels. As Figure 5(c) illustrates, the bias is significantly higher when the heuristic method is used. This study indicates that the novel HMM methods developed here represent a demonstrable improvement over the heuristic method for the identification of stopping phases in tracks. In particular, the level of FPs and degree of systematic bias introduced by the heuristic method are unacceptable, as they would lead us to draw erroneous conclusions from our data.
In this section, we restrict our attention to the HMM-based methods, as the simulation study demonstrated that the FP level is unacceptable using the heuristic method when even low levels of noise are present. Our aim is to demonstrate the broad relevance of our methods to various species of motile bacteria. To this end, we consider two novel datasets, obtained for R. sphaeroides and E. coli as described in Materials and Methods. Results from the analysis of R. sphaeroides are shown in full. Many previous studies have considered the motile behaviour of E. coli [4], [9], [40], therefore for reasons of space we only present the main results from this dataset.
We use the non-chemotactic and non-motile datasets to form the empirical prior in the HMM-based methods. This is achieved by computing the framewise speeds and angle changes in both cases and applying the KDE to estimate the observation pdfs, as described previously. The emprirical prior for the R. sphaeroides dataset is plotted in Figure 6.
The inferred maximum likelihood parameters are shown in Table 1 along with other values reported in the literature. Our simulation study indicated that both HMM-based methods generated MLEs that differed from the true values, with the speed-only method likely to overestimate both and and the accuracy of the full method depending on the level of noise. This is borne out in our analysis, with the speed-only method generating larger MLEs for both R. sphaeroides and E. coli. The discrepancy between the two methods in the inferred transition rates is thus an indication that our estimates of the transition rates should be treated with caution.
A wide range of transition rates have been recorded in the studies cited in Table 1, despite the superficially similar experimental protocols. A few of the many possible explanations include the use of different wildtype strains, small differences in the composition of the motility buffer, and differences in the analysis methods. Comparing with our results, we see that the inferred value of the mean stop duration in R. sphaeroides is in reasonable agreement with the findings of Berry et al. [41]. The results suggest that running phases occur for a shorter mean duration in our datasets than those of Brown [42] or Packer et al. [43], as indicated by the smaller value of . Results for E. coli are in reasonable agreement with those of Berg and Brown [4]. The tethered cell and tracking protocols differ a great deal, as observed by Poole and coworkers [13], who noted that the use of antibody to tether R. sphaeroides to a microscope slide by their flagella substantially reduced their rotation speed and decreased the number of observed stops. This is consistent with our findings, as we estimate a smaller value for , corresponding to shorter runs and an increased number of stopping phases.
Furthermore, we note that our MLEs are computed for pooled data, so that individual variations between tracks are averaged over an entire dataset. There is considerable heterogeneity in switching rates within a bacterial population [43]. However, considering each track separately would result in insufficient data being available for shorter tracks, or those containing no run-stop-run transitions, so we do not consider that problem here. It is for this reason that the estimate of the error in the MLEs is low in comparison with the other results cited; this is because we use bootstrapping of our ensemble sample to generate this estimate (see Materials and Methods for details). The error estimated in our study is therefore a reflection of the nature of the negative log-likelihood surface close to the MLE, rather than an estimate of the deviation between individual tracks. It may be possible to investigate population heterogeneity by applying the HMM-based methods to individual tracks obtained using single-cell tracking methods, as these tracks are generally longer.
In contrast with our simulation study, we have no ground truth with which to compare the result of the analysis of the experimental datasets. Nevertheless, a manual inspection of the inferred state sequence of tracks readily identifies some tracks in which the analysis appears to be successful, in addition to some tracks in which the inferred state sequence is unrealistic. A selection of wildtype R. sphaeroides tracks in which the analysis has been manually identified as successful is shown in Figure 7 (left panel). Several well-defined stopping regions within the tracks have been expanded for greater clarity. Note that, although the speed-only HMM method was used to compute the run probabilities in this figure, the results for these tracks are almost indistinguishable when the full HMM method is used. The track shown in Figure 7 (right panel) arises from a bacterium swimming slowly in an exaggerated helical trajectory, and appears to contain a single genuine stopping event. Both analysis methods incorrectly identify several of the helical turns as stopping phases, leading to an unrealistically rapidly oscillating state sequence. Application of post-processing to either HMM analysis method circumvents this issue. The presence of such a track in the censored dataset motivated a manual examination of all tracks exhibiting either high median curvature or containing a large number of inferred stopping phases. This indicated that, of the 2780 tracks included in the wildtype dataset, fewer than five are clearly identifiable as highly tortuous. Any effects from this minority of tracks, after pooling all analysed data, will be insignificant. A similar outcome is observed in E. coli, although the proportion of tortuous tracks appears to be higher (data not shown). We provide the analogous plot to Figure 7 for E. coli in Figure S11.
In Figure 8(a) we provide a verification of our assumption that wildtype bacterial motility in R. sphaeroides may be approximated as consisting of runs, which are equivalent to those of the non-chemotactic strain, and stops, equivalent to the behaviour of the non-motile strain. This figure shows the observed distribution of framewise speeds in the phases identified as running and stopping by the analysis methods. These are qualitatively similar to those in Figure 6, suggesting that the form of our empirical prior is appropriate. Furthermore, the similarity of the distributions estimated by the speed-only and full methods indicate that the two methods are in close agreement.
Figures 8(b) and 8(c) show the estimated distribution of absolute stopwise angle changes in R. sphaeroides and E. coli, respectively, as computed using the speed-only and full HMM methods without post-processing. Plotting angles rather than absolute angles does not affect the results, as the distribution is symmetric (data not shown). We consider this novel result an important demonstration of the application of our analysis protocol; such a distribution has not been recorded previously for R. sphaeroides. Again, the methodological variants are all in close agreement. The distribution is unimodal, containing a single peak at the origin. We carried out a two-sided Kuiper test [44] on the R. sphaeroides dataset to compare the simulated distribution of inferred stopwise angles (shown in Figure 5(b)) with the experimentally-observed distribution. If these two distributions are similar, we are unable to determine whether the observed experimental distribution is significant, or whether it arises as a result of the bias inherent in our analysis method. Analysis of the experimental R. sphaeroides data indicates that (see Figure S10 and Text S1); we use the conservative value in our simulations. A two-sided Kuiper test reveals that the two distributions differ significantly (, see Text S1 for details of the calculation). The result in Figure 8(b) is therefore more significant than the small bias introduced by the analysis methods, indicating that R. sphaeroides exhibit persistence over reorientation phases.
In this work we have demonstrated the effective application of novel analysis methods based on a modified HMM to tracking data acquired using a simple and relatively inexpensive experimental protocol. The result is a high-throughput method to characterise bacterial motion. We applied our methods to two species of bacteria that exhibit quite different motile behaviour and showed that we are able to estimate certain key distributions, such as the pdf of stopwise angle changes, plotted in Figures 8(b) and 8(c). This result has not been measured before in R. sphaeroides, and provides significant evidence that this bacterium exhibits persistence over reorientation events, which has important consequences for the modelling of their motion, and that of related flagellate bacteria. We note that persistence is a consequence of any reorientation process that occurs over a stochastic duration if some reorientation phases are sufficiently brief that the direction has not been fully randomised. Therefore, we propose that shorter reorientation events in the two species considered here lead to a greater degree of persistence. Testing this hypothesis is the topic of ongoing work.
The stopwise angle change distribution in E. coli (Figure 8(c)) has been measured previously by Berg and Brown [4] (see Figure 3 in that reference for comparison). In contrast with the bimodal distribution centred at approximately found in Berg and Brown's study, we find that the distributions in both E. coli and R. sphaeroides is unimodal and peaked about the origin. In addition, there is no significant difference between the distribution for these two species. For further comparison, Xie et al. measured the distribution of stopwise angle changes in V. alginolyticus, a bacterium that undergoes reversal events, and showed that the distribution is bimodal, with peaks at around 90 and 180 degrees [12]. The difference between the analysis methods used to extract stopping events in our study and that of Berg and Brown may provide an explanation for the discrepancy in our results. In the earlier study, a heuristic method is applied in which the framewise angle change must exceed 35 degrees for more than one frame to be labelled as a stop [4]. This may bias the analysis towards detecting stopping events with larger angle changes.
A further explanation for the discrepancy between this study and that of Berg and Brown may be the substantially different experimental protocols used in the two studies. Berg and Brown track individual bacteria at a frame rate of , while we simultaneously track multiple bacteria at a frame rate of . As a result, our datasets contain significantly more tracks: we analyse 1758 tracks in the E. coli wildtype dataset, compared with the 35 recorded by Berg and Brown [4]. Duffy and Ford [10] more recently used the same tracking apparatus to study P. putida, obtaining 80 tracks. However, the tracks we acquire have a lower mean duration: Berg and Brown [4] present a wildtype track 29.5 seconds in duration; by comparison the mean duration of our tracks is 1.5 seconds in the R. sphaeroides dataset and 6 seconds in the E. coli dataset. This difference in mean track duration is due to the lower magnification used in acquiring the E. coli dataset, in addition to the lower swimming speed of this species.
The duration of tracks is limited by the size of the focal plane and the fact that bacteria may swim out of focus, thus terminating the track. This reduction in track duration is a consequence of the high-throughput, unsupervised protocol used in this study, and is a limitation generally present in many recently-developed multiple cell tracking protocols [15], [29]. Whilst we obtain fewer measurements for each individual, we are able to measure significantly more robust population-wide statistics. As each cell is observed over a randomly-selected time interval in its lifetime, the shorter duration of the tracks has no consequences for our population measurements. Further work is required to determine whether shorter duration tracks reduce our ability to discern variations in the motile behaviour of an individual bacterium. By way of preliminary comment, we note that the appearance of the tracks with the longest duration (around 10 seconds) in the current dataset suggests that the motile behaviour observed in our tracks is not significantly different over a single order of magnitude of timescales. Furthermore, our approach is less subject to bias than a human-operated single-cell tracking protocol, as we image all cells within the field of view and discard tracks using a small number of well-justified censoring parameters. In contrast, any protocol in which the experimentalist may select which cells to track may be systematically biased in favour of a certain, idealised, type of motile behaviour.
A second novel contribution of the present work is the use of a systematic simulation study to validate our analysis methods and compare them with an established method. To our knowledge, no studies have previously compared analysis methods applicable to bacterial tracking data. The comparison indicates that the methods based on the HMM are significantly more robust to noise than the established heuristic method, generating significantly fewer FPs. Furthermore, the simulation study allowed us to determine the extent to which the results are biased by FPs (see Figure 5(b)). We used the results from our simulation study to show that the distribution of stopwise angle changes obtained from experimental data in R. sphaeroides (Figure 8(b)) differs from the distribution of FP stopwise angle changes obtained from simulated tracks (Figure 5(b)) with very high statistical significance. A quantification of the inherent bias in the analysis methodology has not been carried out in previous bacterial tracking studies [4], [45], thus it is unclear to what extent the statistics may be biased. We believe that our simulation approach therefore represents an important advance in the field of bacterial tracking.
An important caveat associated with the high-throughput tracking of many cells simultaneously is the inevitable presence of many tracks that do not appear to conform to the well-studied run-and-tumble model of motility. For example, a non-motile subpopulation has been observed in several similar studies [46]–[48]. Whilst these tracks may be of general interest, the present analysis methods are specifically developed to extract information about bacteria undergoing run-and-tumble motion, hence it is necessary to filter out incongruous tracks. In Materials and Methods, we have presented censoring approaches that mitigate such issues. In particular, the minimum bounding diameter and tortuosity are very useful characteristics for censoring tracks that might otherwise lead to spurious inferences. In particular, we discard the top 5% of tracks, ordered by tortuosity. This approach allows us to apply the same censoring method to multiple datasets without the need to specify multiple thresholds, and therefore permits unbiased comparisons to be made.
Manual inspection of the segmented tracks revealed a selection of tracks in which the new analysis methods appear to have performed well (see Figures 7 and S11). These tracks were manually selected from the dataset because they appear easy to interpret, with clear running and stopping phases. In addition, an example of a helical R. sphaeroides track was shown, for which both analysis methods clearly failed to infer the correct state sequence. The inclusion of post-processing helped to correct the inferred run probabilities.
The HMM approach takes advantage of the availability of non-motile and non-chemotactic mutant strains to obtain empirical prior information on the motion of the bacteria. Such strains are available for many bacterial species not considered in this study, for example Campylobacter jejuni [49], and Caulobacter crescentus [50]. The protocol developed is theoretically applicable to any bacterium that undergoes approximately discrete reorientation events of sufficient duration so as to be captured with a video microscope. It is encouraging that our analysis methods have proved applicable to two very different species of bacteria. There are substantial differences in the reorientation mechanisms of the two species: E. coli undergoes rapid, active reorientation, achieved by the displacement of individual flagellar helices from a peritrichous flagellar bundle, whereas R. sphaeroides reorientates more slowly, by halting the rotation of its single flagellum [5]. The mean stop duration parameter, , is larger in R. sphaeroides, as expected. Further work is required to determine whether all such bacteria are amenable to analysis in this way, however. For example, Bacillus subtilis is believed to accelerate into a running phase [51], which could contravene our two-state model of motion if the acceleration stage is long relative to the timescale of the microscopy.
A further possible application of the methods presented in this study is to the motion of certain eukaryotic species, such as the alga Chlamydomonas, which is known to exhibit motion that is superficially similar to the random swimming of bacteria [52]. However, this alga is approximately an order of magnitude larger than bacteria, and therefore exhibits significantly different properties, such as inertia and spatial sensing. Further work is needed to test whether our methods are applicable to such species.
The methods presented here may also be applied in situations where no mutant strains are available. The motion of non-motile bacteria may be reasonably approximated by a diffusive process, as is the case for the non-motile R. sphaeroides and E. coli in the present study [53]. Furthermore, it is possible to generate an estimate of the behaviour of bacteria in a running phase by manually selecting running phases in a wildtype dataset, although this is a subjective procedure that potentially biases the analysis. Whilst the present study concerns the analysis of a single, identified species of bacteria at any one time, there is also a demand to analyse samples containing multiple unknown bacterial species [54]. Further work is required to determine whether our analysis methods are applicable in these situations. For example, minor modifications should allow the HMM methods to be used to determine the likelihood that a given observed track arises from a reference model of motion.
The current experimental approach produces two-dimensional position coordinates for the cell centroids. We have therefore implicitly projected the true three-dimensional motion of the bacteria swimming in the bulk onto the microscope's image plane. Hill and Häder [55] analysed the effect of projection of tracks onto a two-dimensional plane and concluded that, for their purposes, the error introduced in the observed mean speed is small (). The authors assume an infinite focal depth for their calculation, whereas the focal depth in our setup is small compared to the dimensions of the image plane. We therefore expect the errors caused by projection in our case to be substantially smaller. A further consequence of performing tracking away from a surface within a single focal plane is that bacteria may freely swim out of focus, causing the track to be terminated and leading to tracks of relatively short duration [17]. It is possible to track bacteria in three dimensions, and several groups have made use of various three-dimensional tracking methods to investigate bacterial swimming [4], [10], [28], [29], [45], [56], [57]. The process for obtaining three-dimensional tracks is, however, generally more complex than the method we use and in many cases this leads to a reduced number of tracks available for analysis. Digital holographic microscopy is a promising recent development that could potentially allow the tracking of multiple bacteria simultaneously in three dimensions in a fixed field of view [58]. The HMM-based approaches presented here can be extended in a straightforward manner to deal with three-dimensional data.
Software implementing the methods described in this study is provided in the supporting file Software S1. It is fully documented and written in Python to make it compatible with all major operating systems. The applications of the analysis methods presented here are of potential benefit in a wide variety of bacterial research, including studies of pathogenicity, biofilm formation, and the response of bacteria to chemoattractants and changing environments. In particular, the field of microfluidics is a promising area for further development, as it allows the tracking of bacteria in a well-defined concentration gradient of chemoattractant, as demonstrated by Ahmed and Stocker [17]. In this case, a modification would be required to incorporate the spatial variation of the transition matrix , reflecting the heterogeneous chemoattractant concentration. The ability to quickly assess and compare the motility of a variety of related bacterial strains, or different species, is a powerful addition to the methodological toolbox of the bacteriologist.
Imaging and tracking was performed on three different strains of R. sphaeroides: wildtype (WS8N), a non-motile mutant (JPA467) and a non-chemotactic mutant that is incapable of stopping (JPA1353). Details of the experimental protocol used to create the mutant strains, and the growth conditions, are given in [5]. Some typical raw footage of R. sphaeroides is provided in Video S1. Three strains of E. coli were also used: wildtype (RP437), non-motile (CheY**), and non-chemotactic (ΔCheY). Bacteria were imaged in a homogeneous solution of motility buffer using a tunnel slide. Imaging was performed at 50 frames per second using a Nikon phase contrast microscope with a magnification objective lens in the case of R. sphaeroides and a objective in the case of E. coli. The images are captured in 256 level greyscale, 640 pixels (px) wide and 480 px in height, equivalent to wide and high in the case of R. sphaeroides and twice those dimensions for E. coli. For comparison, a typical R. sphaeroides cell is approximately ellipsoidal, with axial and equatorial diameters of around and , respectively. Imaging was performed with the microscope focused approximately below the top coverslip, and at least this distance from the bottom surface of the microscope slide. This is sufficiently far from either surface that we may neglect surface effects, which are known to cause bacteria to swim in arcing trajectories [45] The observed cells are swimming freely in the medium and may stray out of the focal plane. Typically between 10 and 20 minutes of footage are acquired for each strain, from each of which we obtain between 3000 and 7000 tracks. The tracking procedure is able to cope with a large variation in the density of cells within the field of view, and this value changes depending on the level of magnification used. We typically aimed for around 20–40 cells visible within the field of view in the case of magnification, and 50–80 cells in the case of magnification. Both magnification levels used provided sufficient spatial resolution to find centroids with acceptable accuracy. Further work is necessary to determine whether even lower levels of magnification would allow us to increase the throughput of the experiment without compromising on accuracy. The frame rate of the camera should be sufficiently rapid that reorientation events can be imaged, and preferably so that most events last for greater than a single frame.
We performed cell tracking in two stages. First, in the object detection stage, each frame in a video was processed to establish the centroids of each visible cell. Second, in the data association stage, centroids in each frame were connected to form tracks. The object detection stage is carried out in several steps:
The centroids computed using this method represent the targets present in each frame. The initial background subtraction ensures that any static image artefacts, such as dust on the microscope lens or impurities stuck to the coverslip, are removed from the video. The parameters and were selected separately for each video based on manual verification that the process correctly segmented cells in the images. The values of these parameters were chosen to minimise the number of missed detections, at the expense of producing additional FPs, as the data association routine is robust to high levels of FPs [27]. The minimum cluster size constraint was applied to the region data to remove spurious targets, which are too small to be cells. The minimum cluster size was fixed at px, which is substantially below the mean cross-sectional area of a cell. This resulted in the removal of a significant number of FPs whilst having no effect on true positives. Some errors arise in the process of computing the cell centroid, due to the relatively low contrast of the microscope images. We estimate that such errors should be no greater than half the diameter of a cell body. In order to manually confirm that cell centroid calculation is sufficiently robust for our purposes, tracks from non-chemotactic cells were examined to ensure that they mainly showed smooth swimming, with no overly jagged sections. A further consequence of the low contrast images is that it is not possible to determine cell orientation on this scale; this parameter must therefore be inferred from the angle change between each triplet of consecutive centroids.
The data association method used in this study is a multitarget tracking scheme based on the probability hypothesis density filter. We use an implementation described in [27], which has been applied to microscope videos similar to those used in this study. Video S2 shows the raw microscopy footage of R. sphaeroides overlaid with tracks. As described in Analysis methods, the tracker performs less well when cells are in a stopped phase, as the errors in centroid detection are more significant. Manual inspection of tracks shows well-defined stopping phases in the wildtype strains, however the apparent trajectory during a stop is not accurate. This provides the basis for the modification to the HMM, discussed in the section Hidden Markov model methods.
When optimising the value of the transition parameters and , we require an estimate of the uncertainty in our final MLE. This is achieved using simple bootstrapping [37], in which we resample the tracking dataset by drawing the same number of tracks randomly with replacement. The optimisation procedure is then repeated on the new selected dataset, to achieve a new MLE. This process is repeated for 1000 iterations, after which we sort the bootstrapped MLE transition parameters. We finally use the 2.5th and 97.5th percentile values from the sorted list of and as estimates of the confidence interval.
Preliminary scrutinisation of our R. sphaeroides and E. coli tracking data reveals that a significant proportion of tracks that do not appear to be well described by the run-and-tumble motility described in previous studies [4], [59]. These tracks are either very jagged in their appearance, exhibit unrealistically large movements between frames, or seem to arise from a diffusing object, rather than an actively swimming cell. Possible causes of such tracks include errors in the tracking process, non-motile bacteria, and bacteria with defective motility apparatus. First, the process used to extract tracks from microscope videos may occasionally produce a failed track, for example by linking the trajectories of two different cells, or incorporating a false detection into the trajectory. This is a particular concern if the failed track displays behaviour that differs substantially from the true motion of the observed bacteria, since even a small number of failed tracks may dramatically affect the inferences that are drawn. In order to avoid this issue, tracks containing one or more framewise speeds greater than a threshold value, denoted , are considered to be anomalous and discarded from the dataset. The value of is determined by considering the observed distribution of framewise speeds in the non-chemotactic strain; this gives an indication of the range of speeds exhibited. An upper threshold is then selected that causes outliers to be discarded. In the case of R. sphaeroides, whose mean swimming speed is approximately , we select . The mean swimming speed of E. coli is and we choose . In both cases, is significantly greater than the mean swimming speed. We allow such a large margin for variation in the framewise speed as small errors in consecutive frames can generate large fluctuations in the apparent framewise speed. We do not wish to discard tracks containing a few instances of such inaccuracies, since these quantities will not dominate the population average. This effect is expected to be minor when all tracks in a dataset are considered, and we note that over- and underestimation of the framewise speed are equally probable. Observed framewise speeds above the cutoff value of are unlikely to arise from such a source of noise; these are instead treated as a tracking error and the whole track is discarded.
In addition to tracker errors, a second consideration is the presence of a significant portion of non-motile tracked cells, as is usually observed in experiments of this kind [46]–[48]. Reasons for a lack of motility include cell death, a defective component in the cellular motility machinery, and cell damage due to experimental handling. Figure 9 provides evidence for the presence of a non-motile subpopulation in the non-chemotactic R. sphaeroides strain by comparison with the non-motile strain. As Figure 9(a) demonstrates, the observed distribution of framewise speeds for the non-chemotactic strain is bimodal, with a peak at low speeds that overlaps almost exactly with the non-motile distribution. This suggests that the low speed subpopulation in the non-chemotactic strain is due to non-motile cells. Similarly, in Figure 9(b), non-chemotactic R. sphaeroides bacteria exhibit a bimodal distribution of median curvatures. The subpopulation with higher median curvatures corresponds very closely to the non-motile population.
A third way in which the experimental data differ from the simulated data is the wide range of tortuosities exhibited by real tracks, due to variation within the populations of bacteria being studied. Several tracks appear to be highly tortuous, possibly as a result of bacteria swimming in severely helical paths or with substantial cell body motion. Possible causes for tortuous tracks include damaged or defective flagella, and two bacterial cells swimming whilst stuck together, prior to cell division. None of the analysis methods discussed herein are able to cope with highly tortuous tracks, as these exhibit many large framewise angle changes and low framewise speeds in the running phase. It is therefore challenging to discern stopping phases in such tracks, either automatically or by manual inspection. Tortuous tracks are apparent in the non-chemotactic and wildtype datasets and it is necessary to remove them from the dataset before performing any further analysis.
Our approach to censoring tracks is based on a two-variable representation of a track used by Miño et al. [47]. Each track is summarised in terms of the mean absolute framewise angle change (MAC), and the normalised effective mean speed (NEMS). The NEMS is defined as the ratio of the effective mean speed (EMS) to the mean framewise speed. The EMS is in turn given by the diameter of the smallest circle that encloses the entire track (denoted the minimum bounding diameter, MBD) divided by the total duration of the track. Thus the NEMS takes values between zero and one, and quantifies how straight the track is, with one interpreted as a line that doesn't deviate from a straight path and smaller values indicating increasingly undirected motion.
Miño et al. note that a population consisting of self-propelled particles (which is a good model for motile bacteria) and non-motile diffusing particles exhibits a well-separated bimodal distribution in the MAC-NEMS plot [47]. Figure 10(a) shows such a plot for the non-chemotactic strain of R. sphaeroides, before any censoring. Two modes are clearly visible, one with high MAC and low NEMS corresponding to non-motile cells, and one with low MAC and high NEMS corresponding to motile cells. We use this representation of tracks to determine the effectiveness of our censoring approach.
We also require a measure of the tortuosity of a track, as this is a useful property for the purposes of filtering the dataset. Several methods have been proposed for estimating tortuosity [60]; we employ a method proposed by Lewiner et al., in which a three-point estimator of the curvature of a track is used as a measure of the tortuosity [61]. The curvature is defined for a given position, , , in a track by(13)where the notation is introduced in the Results section and illustrated in Figure 1. The curvature is undefined for the first and last points in a track, as we require three adjacent points to estimate it. We use the median value of the absolute curvature of a track as a summary statistic, as this has been used previously to characterise trajectories [20].
The non-motile tracks are not censored beyond the application of the threshold , as any further censoring would remove all of the remaining tracks. For the non-chemotactic and wildtpe strains we censor tracks in two stages. We first filter out non-motile tracks by imposing a minimum value of for the MBD, and discard tracks whose MBD is lower than this cutoff value. This ensures that tracks that do not cover a sufficiently large region of the field of view are removed from the dataset; in practice, tracks that do not meet this threshold are non-motile or of very short duration. Finally, the top five percent of tracks, ordered by median curvature, are discarded, following Alon et al. [18]. This stage is necessary to remove the remaining non-motile and anomalously tortuous tracks. Discarding an arbitrary proportion of tracks may lead to anomalous tracks remaining in the dataset, or tracks of interest being removed. Nonetheless, this approach has the advantage that the same parameters may be used to censor a wide range of datasets. In this study, for example, we use the same censoring parameters to remove defective tracks from both R. sphaeroides and E. coli tracking data.
Figure 10(b) shows the MAC-NEMS plot for the non-chemotactic R. sphaeroides strain following censoring. The density at high MAC has been filtered out, leaving mainly tracks that lie in the correct region of the plot corresponsing to motile cells. Similar plots for wildtype R. sphaeroides and both non-chemotactic and wildtype E. coli are shown in Figures S7, S8, S9; in all cases, the censoring process removes tracks that lie in the high MAC, low NEMS region.
The number of tracks in each of the datasets before and after the censoring stages is given in Table 2. The censoring stage removes a large proportion of the initial tracks, with most failing on the minimum MBD criterion. This is an important stage of the analysis process, as most of these tracks are due to non-motile cells or very short duration tracks, neither of which are desirable in the final dataset. Figure 10(c) shows a representative sample of tracks before and after the censoring process. The dataset initially contains a large proportion of tracks from non-motile or motility-defective bacteria. After censoring, these tracks have been removed, whilst still retaining longer tracks that exhibit stops.
Post-processing is implemented as follows:
The process is illustrated in Figure 3, in which the short stop inferred at around is removed by the application of post-processing. The relabelling of short runs before short stops introduces a bias towards stops when sustained rapid oscillations occur between the two states (the short run sections will first be converted to stops, resulting in a larger stopped section). We choose to proceed in this fashion as we place greater importance on identifying every stop, possibly at the expense of including some false positives or inferring overly long stopping phases.
|
10.1371/journal.pgen.1003288 | SOX2 Co-Occupies Distal Enhancer Elements with Distinct POU Factors in ESCs and NPCs to Specify Cell State | SOX2 is a master regulator of both pluripotent embryonic stem cells (ESCs) and multipotent neural progenitor cells (NPCs); however, we currently lack a detailed understanding of how SOX2 controls these distinct stem cell populations. Here we show by genome-wide analysis that, while SOX2 bound to a distinct set of gene promoters in ESCs and NPCs, the majority of regions coincided with unique distal enhancer elements, important cis-acting regulators of tissue-specific gene expression programs. Notably, SOX2 bound the same consensus DNA motif in both cell types, suggesting that additional factors contribute to target specificity. We found that, similar to its association with OCT4 (Pou5f1) in ESCs, the related POU family member BRN2 (Pou3f2) co-occupied a large set of putative distal enhancers with SOX2 in NPCs. Forced expression of BRN2 in ESCs led to functional recruitment of SOX2 to a subset of NPC-specific targets and to precocious differentiation toward a neural-like state. Further analysis of the bound sequences revealed differences in the distances of SOX and POU peaks in the two cell types and identified motifs for additional transcription factors. Together, these data suggest that SOX2 controls a larger network of genes than previously anticipated through binding of distal enhancers and that transitions in POU partner factors may control tissue-specific transcriptional programs. Our findings have important implications for understanding lineage specification and somatic cell reprogramming, where SOX2, OCT4, and BRN2 have been shown to be key factors.
| In mammals, a few thousand transcription factors regulate the differential expression of more than 20,000 genes to specify ∼200 functionally distinct cell types during development. How this is accomplished has been a major focus of biology. Transcription factors bind non-coding DNA regulatory elements, including proximal promoters and distal enhancers, to control gene expression. Emerging evidence indicates that transcription factor binding at distal enhancers plays an important role in the establishment of tissue-specific gene expression programs during development. Further, combinatorial binding among groups of transcription factors can further increase the diversity and specificity of regulatory modules. Here, we report the genome-wide binding profile of the HMG-box containing transcription factor SOX2 in mouse embryonic stem cells (ESCs) and neural progenitor cells (NPCs), and we show that SOX2 occupied a distinct set of binding sites with POU homeodomain family members, OCT4 in ESCs and BRN2 in NPCs. Thus, transitions in SOX2-POU partners may control tissue-specific gene networks. Ultimately, a global analysis detailing the combinatorial binding of transcription factors across all tissues is critical to understand cell fate specification in the context of the complex mammalian genome.
| Transcription factors bind DNA in a sequence-specific manner and regulate gene expression patterns in response to developmental cues. Thus, transcription factors often direct a hierarchy of events controlling cellular identity [1], [2]. The HMG box containing transcription factor SOX2 is essential for the development of the epiblast in the early mammalian embryo [3] and for the maintenance of embryonic stem cells (ESCs) in vitro [4]. SOX2 is also necessary for the function and maintenance of neural progenitor cells (NPCs) in the nervous system [5], [6]. Further, SOX2 functions in other adult stem cell and progenitor populations in the gastrointestinal and respiratory tract, as well as in the developing lens, inner ear, taste buds, and testes [7]–[12]. Thus, SOX2 is a critical regulator of distinct stem cell states, but how it can serve this multifunctional role is not fully understood.
In ESCs, SOX2 is a component of the core transcriptional regulatory circuitry that controls pluripotency. Together with OCT4 (Pou5f1) and NANOG, SOX2 binds to the proximal promoters of large cohort of genes with known roles in pluripotency (including Oct4, Sox2, and Nanog) as well as those that function later in development [13]–[16]. These data suggest that SOX2 regulates ESC state by actively promoting pluripotency and by marking the regulatory regions of developmental genes for future activation. Consistent with this, SOX2 can act as a pioneer factor at a subset of genes in ESCs, and can be sequentially replaced by other SOX family members during differentiation, leading to activation of genes [17], [18]. SOX2 is also a critical factor in somatic cell reprogramming, whereby adult cells are converted into a pluripotent ESC-like state by the exogenous expression of a small set of transcription factors [19]–[21], with SOX2 being at the top of a gene expression hierarchy during the late phase of reprogramming [22].
In the central nervous system (CNS), Sox2 is required for proper NPC function during embryonic development and for maintenance of NPCs postnatally [23]–[25]. Specifically, loss of Sox2 in the developing CNS leads to multiple brain defects, including precocious progenitor differentiation and a reduced proliferating cell population in the brain, resulting in perinatal lethality [5], [6], [26], [27]. In contrast, forced expression of Sox2 blocks terminal differentiation of NPCs [26]–[29]. While a critical role for Sox2 in distinct stem cell populations has been firmly established both in vivo and in vitro, the molecular mechanisms by which SOX2 regulates cell type-specific gene expression programs are not clear.
Analysis of genome-wide binding profiles indicates that SOX2 occupies the promoters of thousands of genes [17], [30], however, a direct comparison of SOX2 targets in ESCs and NPCs has not been reported. Emerging evidence indicates that transcription factors drive tissue specific gene expression programs through interactions with distal enhancer elements [31]–[33]. Recent studies have shown that histone modification patterns, specifically monomethylation of lysine 4 of histone H3 (H3K4me1) and acetylation of lysine 27 on histone H3 (H3K27ac), mark distal enhancers [34]–[37]. Using this set of histone marks, we previously identified thousands of enhancer elements in ESCs and NPCs [34]. Thus far, the binding of SOX2 at enhancers has only been clearly demonstrated at a few genes in both ESCs and NPCs. For example, SOX2 occupies the proximal and distal enhancers upstream of the Oct4 promoter in ESCs whereas binding at an intronic enhancer (Nes30) in the Nestin gene was observed in NPCs [14], [38]–[40]. Thus, knowledge of SOX2-bound enhancers in these two cell types will contribute significant new insights into understanding control of cell state.
SOX family members weakly bind DNA and cannot robustly activate transcription alone, suggesting roles for additional partner factors in target selection [41]. Consistent with this, cooperation between SOX and POU transcription factor families has been highly conserved across metazoans where these factors are important regulators of developmental programs [42]. For example, SOX2 cooperates with the Class V POU family member OCT4 in ESCs to maintain pluripotency [13]–[16], however transcription factors that function with SOX2 genome-wide in NPCs are largely unknown. Thus, the identification of factors that bind to genomic sites with SOX2 will also be key to understanding how this master regulator can control distinct phenotypic outcomes.
Here, we defined the genome-wide binding patterns of SOX2 in ESCs and NPCs and show that SOX2 occupied a largely distinct set of genomic regions within promoters and distal enhancer elements in the two cell types. Similar to its cooperation with OCT4 (Pou5f1) in ESCs, we identified the Class III POU transcription factor BRN2 (Pou3f2) as a candidate SOX2 partner factor that co-bound a large fraction of distal enhancers with SOX2 in NPCs. Consistent with a functional role, forced expression of BRN2 in differentiating ESCs led to recruitment of SOX2 to a subset of NPC distal enhancers. This recruitment was associated with changes in chromatin structure, activation of neighboring genes, and ultimately precocious differentiation toward a neural-like state. Further analysis of bound sequences showed differences in the arrangement of a SOX-POU binding in ESCs and NPCs and revealed enrichment for additional transcription factor motifs. Together, these data reveal new insights into how SOX2 can function in a context-dependent manner to specify distinct stem cell states. Our work also has important implications for understanding development as well as the process of somatic cell reprogramming.
SOX2 is a master regulator of pluripotent ESCs and multipotent NPCs, yet how the same transcription factor can specify distinct stem cell states remains an open question. We reasoned that detailed analysis of genomic binding patterns in the two cell types might reveal how SOX2 can regulate diverse gene expression programs. To this end, we differentiated ESCs toward NPCs using established protocols [43], and interrogated SOX2 binding sites by chromatin immunoprecipitation followed by massively parallel sequencing (ChIP-Seq). Analysis of SOX2 binding in genetically identical ESC and NPC lines identified 13,717 and 16,685 enriched regions, respectively (Table S1). Our results were highly consistent with prior work in ESCs [16], however we observed a lower correlation compared to published data sets in neural progenitor cells (Figure S1A and Discussion). We found that >95% of bound regions are unique to each cell type (only 1,274 of the total regions are common to both datasets) (Figure 1A and Figure S1A, S1B). Thus, we identified a union set of 29,128 enriched regions at high confidence and found that SOX2 occupied a largely non-overlapping set of genomic sites in ESCs and NPCs.
SOX2 is thought to bind to regulatory regions of genes with roles in stem cell maintenance and neural differentiation [13]–[16], however, a direct comparison of genome-wide binding in ESCs and NPCs has not been reported. Thus, we first mapped binding sites within 1 kb of a transcription start site (TSS) and found that SOX2 occupied 893 and 3,821 sites within promoters in ESCs and NPCs, respectively (Table S2). While ∼one-third (36%) of bound TSSs in ESCs were common to NPCs, SOX2 largely occupied distinct sites within promoters in the two cell types (Figure S1C). For example, SOX2 occupied the Nanog promoter only in ESCs, while the Egr2 (Krox20) promoter was bound only in NPCs, and a site within the Hdac9 promoter was occupied in both cell types (Figure 1B). Nanog and Egr2 are critical regulators of the ESC state and neural development, respectively, and Hdac9 is a broadly expressed chromatin regulator with a known role in brain development [44]–[48]. Furthermore, we also found examples where SOX2 occupied different sites in ESCs and NPCs but within the promoters of the same gene, such as the Rlim promoter, which encodes a regulator of both X-inactivation and later neural patterning [49], [50] (Figure 1B). Consistent with this, while roughly one-third of TSS-associated regions overlapped in ESCs and NPCs, 58% of the genes bound by SOX2 in ESCs were also NPC targets (Figure 1B, Figure S1C and S1D). These data suggest that SOX2 can utilize different binding sites to regulate genes in a context-dependent manner.
On a global level, SOX2 bound to a set of genes that code for chromatin and transcriptional regulators in both ESCs and NPCs in accordance with previous data [13]–[16] (Figure 1C, 1D and Table S3). While many of these targets were common to both cell types, a large group of chromatin and transcriptional regulators (490) were occupied uniquely in NPCs. Moreover, SOX2 bound more promoter regions in NPCs compared to ESCs and also occupied genes with diverse functions such as RNA splicing, regulation of the ubiquitin cycle, and translation (Figure 1D). While RNA splicing is a general cellular function, alternative splicing is known to play a key role in brain development [51]. For example, in NPCs, SOX2 occupied the promoters of the alternative splicing factors PTB and nPTB, which constitute a molecular switch regulating neuronal commitment [52]. We also found that SOX2 occupied genes displayed higher expression compared to all genes (Figure 1E, 1F and Table S4) suggesting that SOX2 has a positive regulatory role at promoters in each cell type.
While SOX2 occupied proximal promoter regions in the two cell types, the vast majority of bound sites (>93% and >77% in ESCs and NPCs, respectively) mapped greater than 1 kb from annotated TSSs (Figure 2A). Distal enhancers are important non-coding DNA elements that control tissue specific gene expression patterns at variable distances from the promoters they regulate through binding of transcriptional and chromatin regulators [31]–[33]. We previously identified thousands of putative enhancers in ESCs and NPCs by genome-wide analysis of H3K4me1 and H3K27Ac occupancy, two histone marks known to mark distal enhancer elements [34]. SOX2 bound ∼17% (4,947) and ∼24% (6,842) of these putative enhancers in ESCs and NPCs, respectively (Figure 2B and Table S2). Currently, distal enhancers are presumed to regulate the nearest gene [34], [37], and after assigning each enhancer to the nearest upstream or downstream gene, we found that the SOX2-bound enhancers corresponded to 3,372 and 3,990 genes in ESCs and NPCs, respectively (Table S2). While these sites were largely distinct in the two cell types (Figure S2A), ∼44% of genes associated with SOX2 enhancers in ESCs also had a bound enhancer assigned to the same gene in NPCs (Figure S2B) and included many factors with specific roles in neural specification. Notably, analysis of bound enhancers revealed thousands of additional genes that may be regulated by SOX2 in both cell types which would not have been identified by analysis of only TSSs (Figure S2C, S2D). These data are consistent with the idea that, while enhancer utilization is highly cell type-specific, individual genes can be regulated by different enhancers [31], [53].
The pattern of H3K4me1 and H3K27Ac occupancy can distinguish a given enhancer as active (H3K4me1+/−; H3K27Ac+) or poised (H3K4me1+; H3K27Ac−), states which correlate with high expression of a neighboring gene or the potential of that gene to be expressed later during development, respectively [34], [37], [54], [55]. Thus, globally genes nearest active enhancers are expressed at a higher level than those linked to poised elements. By comparison of SOX2-bound regions with the set of active and poised enhancers in our previous study [34], we found that SOX2 occupied 2,100 and 4,037 poised enhancers and 2,847 and 2,805 active enhancers in ESCs and NPCs, respectively (Table S2). Consistent with the idea that enhancers regulate transcriptional output, expression of genes closest to SOX2-bound active enhancers is significantly higher than genes associated with SOX2-bound poised enhancers (Figure 2C).
To gain deeper biological insights, we used the GREAT algorithm to perform Gene Ontology (GO) analysis to determine the function of genes associated with SOX2-bound enhancers. SOX2-bound poised enhancers in ESCs were nearest genes that function in commitment to the neural lineage and morphogenesis and included Jag1, Neurog3, and Nkx2-2, whereas those associated with poised enhancers in NPCs included genes with roles in terminal differentiation into neurons and glia such as Atoh1, Lhx8, Id2 and Id4 (Figure 2D, Tables S5 and S6). Notably, SOX2 bound to active enhancers nearest genes with functions in stem cell development in both cell types. Enriched categories in ESCs also revealed genes that function in early development and axis specification whereas genes linked to active enhancers in NPCs have roles in WNT signaling and neurogenesis (Figure 2E). For example, SOX2 occupied a known enhancer in the 5′ region of the Nanog locus in ESCs [56] and bound to intronic enhancers in Notch1 in NPCs [57], [58], known regulators of pluripotency and neurogenesis, respectively (Figure 2F). Thus, we identified thousands of stage-specific enhancers including many previously known enhancers in both cell types.
Despite the low overlap of SOX2-bound enhancer regions in ESCs and NPCs, genes linked to SOX2-bound poised enhancers in ESCs had functions in neural development, similar to genes linked to SOX2-bound enhancers in NPCs. Thus, we hypothesized that SOX2 might be regulating a subset of targets in both cell types by occupying distinct enhancer elements. Indeed, direct comparison of these genes revealed that ∼50% of genes (821 of 1,654) associated with SOX2-bound poised enhancers in ESCs also had a bound enhancer associated with that gene in NPCs (Figure S2E), despite the regions of SOX2 binding being largely cell-type specific. Importantly, genes where enhancers remained poised showed no significant difference in expression whereas those genes that gained active enhancers were expressed at higher levels (Figure S2F). These data are consistent with the idea that, while enhancer utilization is highly cell type-specific, individual genes can be regulated by different enhancers [31], [53]. Along those lines, using the GREAT algorithm to query the MGI gene expression database, we determined that 2,253 of the 4,037 SOX2-bound poised enhancers in NPCs were linked to genes expressed in the postnatal mouse nervous system (binomial p-value = 1.91e-35) (Table S6). Together, our data support the idea that poised enhancers can predict future developmental potential and suggest that SOX2 regulates a larger network of genes than previously anticipated by binding to distal enhancer elements.
SOX2 binds DNA weakly and it is insufficient to strongly activate transcription without cooperation with additional factors [59]. Consistent with this idea, we identified the canonical SOX2 motif, 5′-CTTTGTT-3′ [60]–[63] as highly enriched in ESCs and NPCs despite the difference in binding patterns (Figure 3A). Thus, we sought to identify additional factors that may function with SOX2 in ESCs and NPCs. SOX2 partners with the Class V POU-domain containing transcription factor OCT4 in ESCs to regulate a large cohort of genes important for pluripotency [13]–[16] however, partner factors in NPCs have not been clearly defined.
Interactions between SOX and POU factors are conserved in all metazoans and play key roles in embryonic development [42], thus, we hypothesized that SOX2 may also function with POU factors in NPCs. To test this, we interrogated a 100 bp window surrounding peaks of SOX2 enrichment in NPCs and determined enrichment for all known vertebrate transcription factor-binding motifs in the TRANSFAC database. Notably, we identified several enriched motifs, including two highly similar motifs recognized by the Class III POU factor BRN2 (Pou3f2) (Figure 3B and Table S7). BRN2 was of particular interest for several reasons. First, our transcriptome analysis showed that Brn2 is highly expressed in NPCs, but not in ESCs (Table S7). Moreover, Brn2 and Sox2 are both expressed in neurogenic regions of the brain and SOX2 and BRN2 are known to co-occupy a small number of loci in this tissue [38], [64], [65]. Like Sox2, Brn2 loss-of-function causes pleiotropic defects and NPC impairment [66]–[69]. Furthermore, Sox2, Brn2, and the forkhead transcription factor Foxg1 are sufficient to reprogram fibroblasts toward a multipotent NPC-like state [70]. These data suggest that transitions in POU partner factors of SOX2 may control cell identity in distinct stem cell populations.
Although neurogenesis and maintenance of cell identity in the brain require BRN2, its target genes in NPCs were not known. To address this, we performed ChIP-Seq and identified 6,574 BRN2 occupied regions in NPCs (Table S1). Similar to SOX2-bound regions, more BRN2-bound regions mapped to previously identified distal enhancers [34] than to promoter regions (Figure S3A). Motif analysis revealed enrichment for a canonical Octamer (OCT) motif (5′-ATGCATAT -3′) [71], [72] within BRN2 bound sites validating the high quality of our data set (Figure S3B).
We next examined the overlap between SOX2 and the two POU factors (BRN2 and our previously published OCT4-ESC dataset [16], Table S1) in ESCs and NPCs. Regions occupied by OCT4 and BRN2 showed little overlap (Figure S3C), indicating that these factors occupied cell-type-specific targets. Our data confirmed that SOX2 and OCT4 co-occupied many genomic sites in ESCs [13]–[16] (Figure 3C and Figure S3D-S3G). For example, SOX2 and OCT4 co-bound the promoter of Fbxo15 and to two putative enhancers of Pax6 that have been previously identified based on evolutionary sequence conservation and histone modification patterns [73] (Figure 3D). Notably, whereas BRN2 was absent from most SOX2-bound promoters in NPCs (Figure 3E), BRN2 occupied a subset of distal enhancers and bound many of these sites with SOX2, including known SOX2-BRN2 targets such as enhancers of Sox2 and Nestin [38], [65] (Figure S3H-S3K). For example, SOX2 and BRN2 co-occupied putative 3′ enhancer regions of Olig1 [74], and a known regulatory region 3′ of the Ascl1 (Mash1) locus [75] (Figure 3F). Together, these data suggest that SOX2 functions with BRN2 at a subset of distal enhancers to regulate target genes in NPCs.
Whereas SOX2-OCT4 bound enhancers associated with genes that have roles in pluripotency and lineage commitment, SOX2-BRN2 enhancers neighbored genes that function in NPC identity. Overall, SOX2 and BRN2 occupied 756 poised and 895 active enhancers in NPCs (Figure 3G). SOX2-BRN2 bound active enhancers correlated with genes that were expressed at higher levels than those associated with poised enhancers (Figure 3H). Further analysis revealed genes linked to active enhancers included transcription factors that play roles in neural development such as Notch1, Rfx4 and Sox2 itself (Figure 3I and Table S8). Interestingly, genes linked to the co-bound poised enhancers in NPCs included regulators of later stages of neuronal developmental such as the pro-neural transcription factor Atoh1 [76], [77] and Dab1, a critical regulator of neuroblast migration [78] (Figure 3J and Table S8). Notably, ∼24% of genes associated with SOX2-OCT4 poised enhancers in ESCs overlapped with genes associated with SOX2-BRN2 bound enhancers in NPCs that included known regulators of neural development such as Atoh1 and Ncam1, despite differences in the bound regions. Thus, SOX2-POU partnerships may control neural development by differentially targeting specific subsets of enhancers in pluripotent ESCs and multipotent NPCs, in order to establish the development potential of this tissue from very early stages of embryogenesis.
The significant overlap between BRN2 and SOX2 in NPCs predicts that BRN2 is also an important driver of neural commitment. To test this idea, we generated ESC lines that harbored a drug-inducible Brn2 transgene (TetO-Brn2) and assayed the potential of these cells to differentiate toward the neural lineage (Figure S4A–S4C). Upon Brn2 induction, ESCs showed distinct morphological changes from round cells that grew in colonies to polarized, Nestin-positive cells at day 1 of differentiation compared to control cells (Figure 4A and Figure S4D). Consistent with these changes, neural lineage genes such as Nestin and Sox1 showed higher expression in ESCs upon Brn2 expression (Figure 4B). Notably, Brn2 induction led to changes in gene expression and cell fate in the absence of additional growth factors whereas control cells did not show significant differences under these conditions. Thus, forced expression of Brn2 can promote differentiation of ESCs toward a neural-like fate.
Our data suggested that POU factor expression may be a key determinant of cell-type-specific SOX2 target selection, so we hypothesized that ectopic BRN2 might be sufficient to recruit endogenous SOX2 to genomic regions de novo. To test this, we collected TetO-Brn2 cells two days after induction (Figure S4D) and performed ChIP-Seq. We identified 12,362 and 8,401 regions occupied by SOX2 and BRN2 in these cells, respectively (Table S1). Similar to SOX2 and BRN2 in NPCs, these factors occupied more distal sites than promoters (Figure S4E). Strikingly, ∼18% (1,034 regions) occupied by BRN2 in the induced ESCs were also bound by BRN2 in NPCs, indicating that ectopic BRN2 retained some of its NPC target specificity. These regions were distal to loci encoding neurodevelopmental regulators such as Ephrin Receptors (Epha3, Epha4, Epha5, Epha7) and transcription factors such as Id2 and Id4 (Table S9).
Importantly, we defined 1,533 regions co-occupied by BRN2 and SOX2 in these cells. Comparison of these regions to SOX2 and OCT4 targets in ESCs and SOX2 in control cells at day 2 (Table S1) revealed 701 (46%) of these sites were bound uniquely by SOX2-BRN2 in the induced cells. These data suggested that BRN2 was necessary for SOX2 binding at these sites (Figure 4C and Figure S4F, S4G). Analysis of enriched GO categories showed that genes closest to these novel targets had roles in the development and function of the nervous system (Figure 4D and Table S9). Notably, 21% of these novel sites (144 regions) were also bound by SOX2 and/or BRN2 in NPCs, including enhancers linked to genes with demonstrated roles in neural development such as Lrrn1 and Abpa2 (X11l) [79]–[81] (Figure 4E). Expression analysis by qRT-PCR of a subset of these TetO-Brn2/NPC, SOX2-BRN2 genes, including Lrrn1, Abpa2, Kirrel3, Cops2, Id4, and Lemd1, revealed that some were significantly induced in TetO-Brn2 cells compared to controls (Figure 4F). Thus, ectopic BRN2 was sufficient to recruit SOX2 to NPC-specific sites and to induce the expression of nearby genes, indicating that POU-factor partners are sufficient to functionally recruit SOX2 to a subset of cell-type-specific target loci.
Given that SOX2-BRN2 binding in NPCs correlated with cell-type-specific distal enhancers, we hypothesize that SOX2-BRN2 might play a role in regulating the state of these elements. Thus, we next examined the distribution of the enhancer chromatin marks, H3K4me1 and H3K27Ac, in TetO-Brn2 and control cells at day 2 in order to determine whether the ectopic binding of these factors could alter local chromatin structure (Table S1). We found that 777 of the 1,533 co-bound sites (∼51%) were coincident with H3K4me1 and/or H3K27Ac regions in TetO-Brn2 cells and 488 of these regions (∼32%) displayed similar patterns in both induced and control cells (Figure 4G). Interestingly, 165 of the co-occupied regions (∼11%) gained H3K27Ac upon Brn2 induction, and were closest to genes involved in neural development such as Atoh1, NeuroD1, and Tcf7l (Tcf3). This included 125 regions (∼8%) that were unmarked (i.e. lacked H3K4me1 or H3K27Ac) in control cells (Figure 4H) and 40 regions (∼3%) that transitioned from a poised state to active (Figure 4I). Thus, ectopic BRN2-SOX2 binding was sufficient to activate both poised and unmarked enhancers, supporting a role for these factors in controlling global gene expression networks by regulating the activity of cis-regulatory elements. Collectively, these data support a role for distinct POU factors in SOX2 binding site selection and gene regulation, and suggests a model by which BRN2 functions with SOX2 to mediate developmental transitions in the neural lineage.
Given that most SOX and POU family members bind highly similar motifs, we hypothesized that distinct motif configurations may explain, in part, the diverse binding patterns in ESCs and NPCs. For example, SOX2 and OCT4 bind DNA in distinct conformations depending on the arrangement of binding sites 82–86 and these configurations have consequences on factor binding and transcriptional outcome [38], [82], [86], [87]. We found that SOX2 frequently occupied sites within 25-bp of OCT4 (∼25%), and that relatively few sites were greater than 100–200 bp from OCT4 (∼8%) (Figure 5A). In contrast, while a significant fraction of regions showed SOX2 and BRN2 bound within 25-bp in NPCs (∼12%), a larger fraction (∼33%) occurred at distances of 100–200 bp. For example, an intragenic region of the Wwc1 locus was bound by SOX2 and BRN2 in NPCs and peaks of enrichment were 100 bp apart (Figure 5B), while in ESCs neither SOX2 nor OCT4 recognized this element. These data indicate that while SOX2 and POU factors occupied similar motifs in ESCs and NPCs, these factors bound to different arrangements of these motifs in a cell type-specific manner.
Many known SOX2-OCT4 target sites comprise a composite SOX-Octamer (OCT) motif, consisting of a 5′-SOX motif followed by a 3′ OCT site [15], [88], [89]. Therefore, we further analyzed the configuration of the SOX2 and OCT motifs by directly inspecting the sequence within the co-occupied regions. SOX-OCT composite motifs can exist in several configurations that were previously termed “canonical”, “order”, “diverging”, and “converging” [86] (Figure 5C). Interestingly, these configurations were shown to determine which combinations of SOX and POU factors could co-occupy a given site. Surprisingly, we observed that the canonical orientation with a 1 bp overlap between the native TRANSFAC motifs was the most highly represented configuration in both SOX2-OCT4 co-bound regions in ESCs (∼23% of motif pairs) and SOX2 and BRN2 co-bound regions in NPCs (∼21% of motif pairs) (Figure 5C). For example, at a locus on chromosome 2 distal to Chd6, SOX2-OCT4 occupied a canonical motif with a 1 bp overlap in ESCs, and SOX2-BRN2 occupied the same site in NPCs (Figure 5D). Thus, SOX2-OCT4 and SOX2-BRN2 prefer the same composite SOX-OCT motif at genomic targets in ESCs and NPCs.
Combinatorial interactions among transcription factors are important for driving specific transcriptional responses [90]–[93]. In ESCs, SOX2 and OCT4 are known to co-occupy genomic sites with a cohort of other transcription factors, including NANOG, SALL4, and TCF7L1 [13]–[16], [94]–[96]. Thus, we sought to identify additional transcription factors that may interact with SOX2 and BRN2 in NPCs. To this end, we analyzed SOX2-BRN2 bound regions for enrichment of known transcription factor motifs (Table S10). To discover factors that may function specifically with SOX2 and BRN2 in NPCs, we contrasted these motifs with those that were enriched in SOX2-OCT4 co-bound regions. Notably, the enriched motifs in SOX2-BRN2 regions corresponded to transcription factors that were highly expressed in NPCs relative to ESCs (Monte Carlo analysis, p-value = 0.03, Materials and Methods) (Figure 6A). For example, NF-I motifs were highly enriched in SOX2-BRN2 regions in NPCs and family members such as NF-Ia, NF-Ib, and NF-Ix were expressed at significantly higher level in NPCs than ESCs (Table S10). NF-I factors have known roles in central nervous system formation and in NPC function [97]. Motifs associated with the RFX family were also enriched in SOX2-BRN2 regions (Table S10). RFX family members play essential roles in early nervous system patterning [98], [99]. While Rfx3, Rfx4, and Rfx7 were expressed at significantly higher levels in NPCs, Rfx2 expression was higher in ESCs (Table S10). Interestingly, a recent proteomic study identified RFX3 and NF-IB as putative SOX2 interaction partners in NPCs [30]. Thus, our analysis has identified additional transcription factors that may regulate specialized gene networks with SOX2 and POU factors in ESCs and NPCs.
We identified 439 SOX2-BRN2-NF-I-motif and 251 SOX2-BRN2-RFX-motif regions in NPCs (see Materials and Methods). Further analysis showed that SOX2-BRN2 regions containing NF-I or RFX motifs were largely exclusive (only 34 common regions) suggesting that SOX2-BRN2 sites could be further classified by interactions with specific sets of transcription factors. Consistent with this observation, SOX2-BRN2 regions containing an NF-I motif were linked to genes with functions in nervous system development and cell growth, including Sox2 and NF-Ib themselves as well as Olig1 and Integrin genes (Figure 6B). In contrast, SOX2-BRN2-RFX-motif regions were linked to a largely distinct set of regulators of neural development including regulators of neuronal apoptosis such as Ntrk2 (TrkB) [100], Ntrk3 (TrkC) [101], and Cdk5r1 (p35) [102], [103], an important process regulating the development of the CNS (Figure 6C). Interestingly, conditional ablation of Sox2 in NPCs is associated with increased apoptosis in the developing brain [6]. Thus, RFX and NF-I family members represent additional candidate partner factors in NPCs that may further contribute to specific regulation at SOX2-BRN2 target genes. Collectively, our work reveals a detailed picture of how SOX2 coordinates gene expression programs during lineage commitment and provides novel insights into the key principles that underpin regulation of diverse stem cell states.
The HMG-box transcription factor SOX2 has critical roles in the function of multiple stem cell types including pluripotent embryonic stem cells (ESCs) and multipotent neural progenitor cells (NPCs). How this master regulator can control diverse transcriptional programs has remained an important and unresolved question in the field. While SOX2 occupied many promoters in both cell types, the major class of genomic elements occupied by SOX2 in ESCs and NPCs were distal enhancers (Figure 1 and Figure 2). While our data displayed high concordance among replicates and with published data sets in ESCs, SOX2 binding in NPCs was less correlated with prior data [17], [30] (Figure S1). This is likely due to the different protocols used to derive and culture NPCs. NPCs with similar developmental potential but distinct molecular profiles exist throughout development [24], and these populations respond differently to external signaling cues present in culture media [104]–[106]. Thus, it is perhaps not surprising that SOX2 binding is more variable in NPCs relative to ESCs.
We derived NPCs directly from genetically identical ESCs allowing us to directly analyze SOX2 binding as these cells transition between states. Several criteria support the high quality of our data. First, we identified many known SOX2 binding sites including promoters and enhancers in both ESCs and NPCs. Second, while many binding sites were distinct, we identified a canonical SOX2 motif as highly enriched in both cell types. Third, SOX2 overlapped significantly with POU partner factors in ESCs and NPCs consistent with the expectation that these transcription factor families function together to regulate developmental progression. In addition, we identified a SOX-OCT composite motif as enriched in these co-bound sites.
SOX2 occupied largely exclusive sites in ESCs and NPCs, despite using the same DNA motif to recognize these genomic targets. Moreover, SOX2 occupied distinct regions in the same promoter and distinct enhancers associated with the same gene. These data indicated that additional factors dictated SOX2 binding site specificity. While SOX2 co-occupied many binding sites with OCT4 in ESCs, partner factors in NPCs have not been well defined. We found the recognition motif for the Class III POU factor BRN2 was enriched in SOX2 bound regions in NPCs. The evolutionary conservation of the SOX-POU interaction, the co-expression of Sox2 and Brn2 in neurogenic regions of the brain, and the neurodevelopmental defects associated with Brn2 loss-of-function suggested that SOX2 and BRN2 together regulate a subset of genes important for neural fate. Consistent with this, we defined a large group of enhancer elements co-bound by SOX2 and BRN2 in NPCs. We identified known functional enhancers bound by SOX2 and BRN2 in NPCs, such as the Nes30 enhancer of the Nestin locus [38], [40] and the 3′ enhancer of the Sox2 locus, SRR2 [5], and extended this list to include hundreds of additional neural-specific enhancers.
Consistent with a positive role in regulating neural cell state, forced expression of Brn2 led to up-regulation of neural markers and to differentiation toward the neural lineage. Our work is in agreement with several studies that have implicated Brn2 as an early marker of neural commitment [40], [107], [108]. Notably, ectopic BRN2 was sufficient to recruit SOX2 to hundreds of novel sites in differentiating ESCs that corresponded to a subset of enhancers also bound in NPCs. The recruitment of SOX2 by BRN2 to specific loci was sufficient to induce expression of nearby genes and to alter chromatin state in some cases. These data are in agreement with the notion that SOX proteins require partner factors to tightly bind to genomic targets and modulate transcriptional outcomes [59]. Interestingly, ectopic expression of OCT4 alone in NPCs was sufficient to reprogram cells into induced pluripotent stem cells, presumably by partnering with endogenous SOX2 [109], consistent with the idea that POU factors can recruit SOX2 to specific targets. Furthermore, ectopic expression of Sox2, Brn2, and the forkhead factor Foxg1 can transdifferentiate fibroblasts to NPC-like cells [70]. Taken together, these data may facilitate efforts to define the minimal set of genes needed to promote the transition from undifferentiated cells to the neural lineage. Thus, our results implicate BRN2 as a SOX2 partner factor and suggest that together these factors are important for neural specification and NPC function.
While the motifs occupied by these factors were highly similar, the arrangement of SOX and OCT motifs in SOX2-POU target sites displayed differences in ESCs and NPCs. Regulation of SOX-POU target genes appears to depend not only on the presence of a SOX and an OCT motif in close proximity to each other, but also on other DNA sequence determinants, including the spacing and orientation of these motifs with respect to each other [82], [84], [87], [110]–[113]. However, these observations related to only a few genes and had not been extended genome-wide. While we found that SOX2-OCT4 and SOX2-BRN2 preferred similar composite motifs when they were bound in close proximity to each other, examination of co-bound regions found that peaks of SOX2 and BRN2 in NPCs were often spaced farther apart than peaks of SOX2 and OCT4 in ESCs. Thus, allosteric interactions between transactivation domains of SOX and POU factors may be key in stabilizing ternary complexes and in setting the stage for additional interactions that determine binding specificity and transcriptional output at target genes [84]–[87], [114], [115].
Combinatorial interactions among transcription factors allow cells to respond to environmental and developmental cues in a tissue-specific manner. A classical example involves the regulation of interferon-β expression through cooperative binding of transcription factors and chromatin proteins to an enhancer, collectively known as the interferon-β enhancesome [116]. In ESCs, SOX2 and OCT4 are known to physically interact with other transcription factors at many loci, including enhancers [14], [94], [117]–[120], suggesting that SOX2-POU factors may also nucleate specific enhancesomes. We identified a set of candidate factors that may interact with SOX2-BRN2 that included RFX and NF-I family members. NF-I factors are expressed in NPCs in vivo and their loss in development leads to defects in central nervous system formation and specifically NPC dysfunction [97], [121]–[124]. RFX family members also play essential roles in proper brain development [99], [125]. For example, RFX4 regulates Sonic Hedgehog (SHH) signaling in the developing nervous system and loss of function resulted in pleiotropic brain defects linked to SHH signaling [99], [125]. Defects associated with conditional ablation of Sox2 in the brain were also shown to be partially mediated by aberrant SHH signaling [6]. Additional studies revealed that SOX2 co-localized with the ATP-dependent histone remodeler CHD7 in NPCs [30]. Thus, interactions with chromatin modifiers or other epigenetic regulators may also be critical for binding site selection and establishment of NPC-specific gene expression programs in response to particular signals.
Recent data showed that SOX2 functions as a pioneer factor in ESCs by marking a subset of genes for activation by other SOX family members, namely SOX4 in the B-cell lineage, SOX3 in NPCs and SOX11 in immature neurons [17], [18]. Interestingly, the POU factor Brn5, like Sox11, is expressed in differentiated cell types in the CNS and thought to play a role in regulating cell state [126]–[130], thus elucidation of BRN5 targets in these cells may reveal another layer of SOX-POU regulation of neurogenesis. Taken together, these data suggest that transitions in SOX-POU partners regulate the earliest stages of development through terminal differentiation. Ultimately, characterization of combinatorial interactions among transcription factors and chromatin regulators at distal enhancers will be central to understanding the complex mechanisms that control cell state throughout development.
ChIP-Seq and Affymetrix microarray data are deposited on GEO database under the accession numbers GSE38850 and GSE35496.
C57/BL6-129JAE (V6.5) mouse embryonic stem cells were cultured in as described [34]. Neural progenitors were derived via in vitro differentiation from V6.5 ESCs as described [43] and cultured on 15 µg/ml polyornithine and 1 µg/ml laminin in N3 medium, supplemented with 5 ng/ml bFGF, 20 ng/ml EGF, and 1 µg/ml laminin. In the presence of growth factors the vast majority of these cells can be labeled homogenously with antibodies against NESTIN, SOX2, and PAX6. Upon growth factor withdrawal, the cells differentiate into TUJ1-positive neurons.
ChIP in NPCs was performed as described previously [131]. Briefly, approximately 5×108 cells were cross-linked and chromatin fractions were sheared by sonication. ChIP-enriched and input DNA were purified and genomic libraries were prepared using the ChIP-Seq Sample Prep Kit (Illumina 1003473) according to the manufacturers protocol (Illumina 11257047) for selecting library fragments between 200 and 350 bp. Samples were run using the GA2X genome sequencer (SCS v2.6, pipeline 1.5).
For ChIP in ESCs, and in TetO-Brn2 cells and control cells were cross-linked and harvested as above. Approximately 5×107 formaldhyde-crosslinked cells were lysed and as above on an IP-Star (Diagenode). Chromatin was sonicated on the Bioruptor (Diagenode) to an average size of 0.2–1 kb. ChIP was performed on chromatin from approximately 5 million cells with 3 µg of antibody (above) using the IP-Star Automated System (Diagenode) and 2.5% of chromatin was used for each whole cell extract (WCE). Following reversal of crosslinks, sample and WCE DNA was purified. ChIP and WCE DNA was dissolved in water and barcoded genomic libraries were prepared using the TruSeq DNA Sample Prep Kit (Illumina) and multiplexed on the HiSeq 2000 (Illumina).
Antibodies used in ChIP experiments are as follows: SOX2 (R and D Systems AF2018 goat polyclonal); BRN2 (Santa Cruz Biotechnology sc-6029 goat polyclonal); H3 (rabbit polyclonal Abcam ab1791) H3K4me1 (rabbit polyclonal Abcam ab8895); H3K27Ac (rabbit polyclonal Abcam ab4729).
Images acquired from the Illumina/Solexa sequencer were processed using the bundled Solexa image extraction pipeline. Sequences were aligned using Bowtie (http://bowtie-bio.sourceforge.net/index.shtml) using murine genome NCBI Build 36 (UCSC mm8) as the reference genome with default settings for mismatch tolerance and non-unique mapping events. Mapped reads were analyzed as described [16]. Briefly, sequence reads were extended 200 bp for transcription factors and 400 bp for histone modifications and allocated in 25 bp bins. Statistically significant enriched bins were identified using a Poissonian background model, with a p-value threshold of 10−8 to minimize false positives. We then used an empirical background model (whole cell extracts (WCE) for transcription factors or pan-histone histone H3 ChIP-Seq (H3) for chromatin marks) that requires bins to be enriched relative to background to eliminate non-random enrichment. Replicate datasets were combined and analyzed in one batch. Previously published datasets for enhancer associated histone marks were analyzed as described [34], [132]. SOX2, BRN2, and OCT4 enriched regions within 1 kb of a TSS were assigned to the associated gene, while bound enhancers were identified as regions that overlap H4K4me1 and/or H3K27Ac regions that are >1 kb from a TSS [34] and were assigned to the nearest gene using the GREAT algorithm for gene ontology studies and using the Galaxy web tool for all other analyses.
ChIP–seq plots for individual genes were generated using the UCSC Genome Browser (http://genome.ucsc.edu/cgi-bin/hgGateway).. wig files were generated from ChIP-Seq reads and density was normalized to reads-per-million. Published datasets were used to correlate SOX2 bound regions to histone modification patterns for enhancer analysis [34], [132].
We used a 1-bp minimum cutoff for the overlap between regions to define common genomic targets, as described throughout the manuscript to define co-bound SOX2-POU sites and sites occupied by SOX2 or POU factors across cell types. Correlation of ChIP-Seq datasets in figure S1 was performed using a similarity metric based on a correlation coefficient [133]. This analysis generates a correlation coefficient between zero and one reflecting the similarity of genomic regions occupied in two datasets.
RNA was isolated using Trizol reagent (Invitrogen) according to the manufacturer's protocol and DNAse treated using the DNA-Free RNA kit (Zymo Research R1028). Samples were then prepared for Affymetrix GeneChip Expression Array analysis. 5 µg total RNA was used to prepare biotinylated cRNA according to the manufacturer's protocol (Affymetrix One Cycle cDNA Synthesis Kit). Samples were prepared for hybridization, hybridized to arrays, and washed according the Affymetrix hybridization manual using the Affymetrix GeneChip Hybridization, Wash and Stain Kit. GeneChip arrays (Mouse 430) were hybridized in a GeneChip Hybridization Oven at 45°C for 16 hours at 60 RPM. Arrays were scanned on a GeneChip Scanner 3000 and images were extracted and analyzed using GeneChip Operating Software v1.4.
To define expression levels of genes linked to bound promoters and enhancers, and to define fold change of expression levels of transcription factors linked to enriched TRANSFAC motifs between ESCs and NPCs, biological replicates were analyzed using the Affymetrix GCOS program and the mean intensity for each probe across three arrays was calculated. Maximum probe mean values for each gene were taken as gene expression levels. Box and Violin plots were constructed depicting median values as the center line, and bottom and top of the box representing the 25th and 75th percentiles, respectively. Whiskers depict+1.5*IQR (interquartile range) for top, −1.5*IQR for the bottom. To define differentially expressed genes, array data was RMA normalized using updated annotation from the BrainCDF the site and remapped from Ensembl Gene ID to Gene Name using Biomart table. For finding differentially expressed (DE) genes, the biological replicates were subjected to moderated welch test (MWT). Genes were called differentially expressed if the MWT FDR<0.05 and the fold change of the mean of the replicates was more than 1.5 fold up or down.
Gene ontology analysis was performed using GOSTAT (http://gostat.wehi.edu.au/cgi-bin/goStat.pl) for genes linked to SOX2 bound promoters or the GREAT algorithm [134] (http://great.stanford.edu/) for regions associated with SOX2 bound enhancers. GOSTAT was performed using the mgi (mouse) GO annotation database for promoter-associated regions. Since GREAT analysis requires inputs in the mm9 genome build, lift-over of mm8 called regions was performed using the Galaxy web tool prior to input into GREAT (main.g2.bx.psu.edu/). In general, terminal “GO Biological Process” terms were presented in figures to maximize the specificity of the information presented. In some cases terminal terms contained few genes and were thus misleading, so more informative parent terms encompassing less specific but more relevant descriptions of biological processes are presented.
MEME (meme.sdsc.edu/) [135] was used to find DNA sequences enriched in SOX2-, OCT4-, and BRN2-bound regions in ESCs and NPCs. Plus/minus 75 base pairs surrounding a subset of the highest peaks of enrichment for each factor (minimum peak height 100 for SOX2, 148 for OCT4, or 225 for BRN2) were input into MEME and motif logos were generated from obtained position weight matrices.
Brn2 inducible ESCs were generated using the “flp-in” system described previously [136]. Briefly, a single copy of a tetracycline inducible mouse Brn2 cDNA were flipped into the Col1a1 locus of KH2 ESCs harboring an M2-rtTA gene in the Rosa26 locus.
Inducible Brn2 and control (KH2) ESCs were passaged off feeders and cultured in ESC medium with 2 µg/ml Dox. Twenty-four hours after passage, cells were culture in N2B27 (without Vitamin A) media without LIF or serum for the duration of the experiment [137]. Gene expression for differentiation markers was assayed by quantitative Real-Time PCR at 24-hour intervals. For immunostaining, cells were fixed with 4% paraformaldehyde in PBS and stained with anti-Nestin (Developmental Hybridoma Bank) and DAPI.
Trizol-isolated RNA from three biologically independent samples was purified, DNAse treated (DNA free RNA Kit, Zymo Research) and reverse transcribed using a First Strand Synthesis Kit (Invtirogen). cDNA was analyzed in triplicate for each biological sample by quantitative PCR using an ABI Prism 7000 (Applied Biosystems) with Platinum SYBR green qPCR SuperMix-UDG with ROX (Invitrogen). All primers used in this study are listed in Table S11. Data were extracted from the linear range, and the standard curve method was used to obtain relative expression values. Technical replicates were averaged and then biological replicates were averaged. Statistical significance was determined using Graphpad Prism to perform an ANOVA with Bonferroni Correction for multiple testing.
Regions of SOX2-BRN2 co-occupancy in TetO-Brn2 cells were defined as above. To define regions of differential chromatin state between TetO-Brn2 and control cells, we first compared H3K4me1 enrichment in these cells to define regions common to both cell types or unique to one or the other. Common regions were merged and treated as one enhancer if detected in both cell types. A similar analysis was performed for H3K27Ac enrichment. SOX2-BRN2 regions were then compared to regions of H3K4me1 and H3K27Ac in order to define SOX2-BRN2 binding events that resulted in changes in chromatin state between TetO-Brn2 cells and controls.
100 base pair windows around the max peak of SOX2-bound regions (in ESCs and NPCs) regions were analyzed for the presence of overrepresented DNA binding motifs. Similarly, 150 base pair windows around the midpoints between the max peaks of SOX2 and OCT4 or BRN2 (in ESCs or NPCs, respectively) in co-bound regions were analyzed for the presence of overrepresented DNA binding motifs. We used a hypothesis-based approach to identify known protein-DNA recognition elements enriched in each dataset. The set of hypotheses are derived from all vertebrate position-specific scoring matrices (PSSMs) from TRANSFAC [138] filtered for sufficient information content (IC>8 total bits). As many of these motifs are redundant, we clustered them based on pairwise distance by KL-divergence of the PSSMs using Affinity Propagation. The TAMO programming environment [139] was used to store the PSSMs and score sequences. We used two approaches to identify overrepresented motifs. All motifs discussed in the paper were found by both methods except for M00145 in SOX2 bound sites in NPCs which was only found by the first approach described below.
In the first approach, we determined whether motifs were overrepresented in a foreground set of all bound regions (SOX2-bound or SOX2-POU co-bound, depending on the analysis) compared to a background set of randomly generated sequences which matched the GC content of the foreground using the Mann-Whitney-Wilcoxon (MWW) ranked sum test. For each independent motif test, sequences were ranked by the maximum motif score in each sequence (across all k-mers in the sequence for a motif of width k). This ranked list was used to compute the U statistic for the foreground set from which we computed a p-value and applied a Benjamini-Hochberg multiple hypothesis correction. Because many motifs in the databases are very similar to each other, we present the motif within each cluster with the most significant p-value.
In the second approach, we determined whether motifs were overrepresented in a foreground set of 1,000 randomly selected bound regions compared to a background set of randomly generated sequences matching the GC content of the foreground using THEME [140]. A β value of 0.7 and 5-fold cross-validation (CV) were used as THEME parameters. Statistical significance of the CV-error was calculated using randomization of 25 trials and multiple hypothesis corrected using the Benjamini-Hochberg procedure. As in the MWW tables, we present the motif within each cluster with the most significant p-value.
Distances between SOX2 bound sites and cofactor bound sites in ESCs and NPCs were calculated as follows. Overlapping regions of SOX2 and POU factors were defined as regions with at least 1-bp of overlap. Peaks from these overlapping regions were then used to define distances between the bound factors. In particular, we calculated distances between SOX2-BRN2 site pairs (NPCs), and SOX2-OCT4 site pairs (ESCs). Site pairs were defined by matching each SOX2 bound site to the closest cofactor bound site within 200 bases. Distance was calculated as the cofactor chromosomal coordinate subtracted from the SOX2 chromosomal coordinate.
Spacing between SOX and OCT sites was determined using a motif-based approach to determine specific spatial arrangement of the motifs in SOX2-OCT4 (ESCs) and SOX2-BRN2 (NPCs) co-bound regions. Max motif scores were calculated as described above and normalized as in Equation 1. Motif matches to SOX were defined as normalized scores greater than 0.85 to a general SOX TRANSFAC matrix, M01308. Similarly, OCT family motif matches were defined as normalized scores greater than 0.85 to a general OCT TRANSFAC matrix, M00342. For each sequence i and motif j, a motif score sij: was calculated. Spacing was defined as the number of base positions between the OCT4 and SOX2 motif matches relative to the SOX2 motif match. OCT-SOX motif pairs were associated with the previously defined “canonical”, “order”, “diverging”, and “converging” orientations [86].
The Mann-Whitney Z-score test result was used to rank all vertebrate TRANSFAC motifs in order of enrichment for SOX2 and OCT4 bound regions in ESCs, and SOX2 and BRN2 bound regions in NPCs. The change in rank (ΔRank) from ESC SOX2-OCT4 bound regions to SOX2-BRN2 bound regions was determined for each motif. Motifs were filtered to include only motifs with a rank less than or equal to 200 in the two ranked lists. Gene expression fold change was determined for each transcription factor associated with at least one TRANSFAC motif. After assigning a pseudocount of 1 to the normalized Affymetrix gene expression values for each transcription factor at the ESC and NPC stages, the log (base 2) of the fold change was calculated. Motifs were mapped to associated transcription factors according to the vertebrate all profile accessible on ExPlain 3.0 containing 656 motifs. The TF with the fold change that best agreed with the ΔRank of the associated motif was chosen as the ‘representative’ factor for the motif (for instance, if ΔRank was negative, the associated transcription factor that had the most negative log-transformed fold change was chosen.) Scaled motif ΔRank values and the associated log-transformed gene expression fold change values were sorted in order of log-transformed gene expression fold change, and viewed in a heatmap (Spotfire, TIBCO). Only motifs that had associated transcription factor expression values were considered.
A spearman correlation coefficient was calculated between the ΔRank values of the motifs and the expression-fold change of their associated transcription factors. This required each motif to be associated with a single transcription factor. In the case where multiple transcription factors are known to bind a single motif, the TF was selected as described in the previous section. The significance of the spearman correlation coefficient was assessed by a Monte-Carlo algorithm. The input to the Monte-Carlo algorithm was a table in which each row of the first column was a motif and each row of the second column was the set of transcription factors known to bind the motif. The column containing the motifs was randomly permuted 100,000 times (thereby randomizing the associations of transcription factors to motifs), and the process of selecting a single transcription factor to be associated with each motif was repeated. After associating each motif with a single random transcription factor, the spearman correlation between the ΔRank of the motifs and the log-fold-change in expression of the transcription factors was computed. The fraction of randomized tables that produced a higher spearman correlation than the original table was reported as the p-value. Only motifs for which the rank in either the SOX2-OCT4 list or the SOX2-BRN2 list was in the top 200 were used. The motifs also had to have at least one associated transcription factor for which gene expression data was available.
Genomic intervals corresponding to enriched transcription factor binding motifs in SOX2-BRN2 bound regions were determined. The single nearest gene to a given region was determined using the GREAT algorithm. Genes associated with a motif having a motif similarity score of equal to or greater than 0.85 (439 NF-I motif regions and 251 RFX motif regions) were used to generate a non-redundant target gene list. This gene list was then used as the input for Ingenuity Pathway Analysis. Ingenuity recognized 431 NF-I associated genes and 249 RFX associated genes. Overlap of genomic regions containing motif sequences was performed using Galaxy.
|
10.1371/journal.pntd.0001916 | Molecular Changes in Opisthorchis viverrini (Southeast Asian Liver Fluke) during the Transition from the Juvenile to the Adult Stage | The Southeast Asian liver fluke (Opisthorchis viverrini) chronically infects and affects tens of millions of people in regions of Asia, leading to chronic illness and, importantly, inducing malignant cancer ( = cholangiocarcinoma). In spite of this, little is known, at the molecular level, about the parasite itself, its interplay with its hosts or the mechanisms of disease and/or carcinogenesis.
Here, we generated extensive RNA-Seq data (Illumina) representing adult and juvenile stages of O. viverrini, and combined these sequences with previously published transcriptomic data (454 technology) for this species, yielding a combined assembly of significantly increased quality and allowing quantitative assessment of transcription in the juvenile and adult stage.
This enhanced assembly reveals that, despite the substantial biological similarities between the human liver flukes, O. viverinni and Clonorchis sinensis, there are previously unrecognized differences in major aspects of their molecular biology. Most notable are differences among the C13 and cathepsin L-like cysteine peptidases, which play key roles in tissue migration, immune evasion and feeding, and, thus, represent potential drug and/or vaccine targets. Furthermore, these data indicate that major lineages of cysteine peptidases of socioeconomically important trematodes have evolved through a process of gene loss rather than independent radiation, contrasting previous proposals.
| Opistorchis viverrini is an important and neglected parasite affecting ∼9 million people in South-east Asia. The parasite has a complex life-cycle which involves an intermediate phase in cyprinoid fishes. Consumption of raw or under-cooked fish infected with the metacercarial (larval) stage of O. viverreni results in infection, with adult worms living primarily in the intra-hepatic bile duct. In addition to the affects of the infection itself, O. viverrini is directly carcinogenic, with up to 70% of infected individuals in endemic regions developing malignant cholangiocarcinomas. Control of the parasite relies exclusively on the use of praziquantel and little is known about the mechanisms through which O. viverrini stimulates carcinogenesis. An improved understanding of the molecular biology of O. viverrini is urgently needed. In our study, we employed RNAseq technology to assess changes in gene transcription during the development of O. viverrini within the definitive host, and significantly improved the characterization of the transcriptome of this parasite. In so doing, we shed new light on the evolution of a major group proteins (i.e., the cysteine peptidases) which, given their important function roles as excreted/secreted molecules, have been proposed as attractive drug/vaccine targets for a wide-range of neglected flukes, including species of Opistorchis, Clonorchis, Schistosoma and Fasciola.
| Parasitic worms of humans and other animals cause diseases of major socio-economic importance around the world. In spite of their significance, many of them have been substantially neglected in terms of research and their control [1]. Among parasitic flukes (i.e., trematodes), the foodborne trematodes, including the major human liver flukes, Opisthorchis viverrini and Clonorchis sinensis, are particularly understudied. In parts of Southeast Asia, including Cambodia, People's Democratic Republic of Laos, Thailand and Vietnam, O. viverrini is estimated to infect ∼9 million people [2], with ∼67 million being at risk of infection [3].
The life-cycle of this parasite is complex, involving multiple intermediate hosts and a prey-to-predator transmission cycle [4]. Briefly, embryonated eggs are shed into the environment in the faeces from the infected definitive host (mainly humans, dogs and cats). After the eggs are shed into water (usually via untreated sewage), they are ingested by freshwater snails (Bithynia spp.) and then hatch in the gut, releasing the motile embryo ( = miracidium), which develops into a sporocyst. Asexual reproduction within the sporocyst gives rise to rediae and then cercariae. The motile cercariae are released from the snail into the aquatic environment. Thereafter, these larvae undergo host finding and must penetrate the skin of a cyprinoid fishes (e.g., Puntius spp.) and encyst as metacercariae within the skin and/or musculature to survive. The piscivorous definitive hosts become infected upon ingestion of such fish in a raw or undercooked state [5]. Following gastric passage, the metacerariae excyst in the duodenum, and the juvenile flukes migrate via the ampulla of Vater and common bile duct to the intra-hepatic bile and/or sometimes into the pancreatic ducts (∼14 days), whereupon they develop into reproductively-active, hermaphroditic adults (∼4 weeks), which release embryonated eggs via bile/pancreatic fluid into chyme and then via host faeces into the freshwater environment, thus completing the life cycle [4].
Although infection is often asymptomatic, signs or symptoms associated with opisthorchiasis can include eosinophilia and, in intense infections, diarrhoea, epigastric pain, anorexia, pyrexia, jaundice and/or ascites [6]. Chronic opisthorchiasis often leads to cholangitis, periductal fibrosis, cholecystitis and/or cholelithiasis, and, in up to 71% of infected humans in endemic areas, can induce malignant cancer (cholangiocarcinoma) [7]. Hence, O. viverrini has been classified as a Group I carcinogen [8]. In endemic regions, sanitation infrastructure is often limited, and cyprinoid fish is consumed in a variety of traditional dishes. For cultural reasons, this fish is often eaten raw, and there is a resistance to recommendations to cook fish to prevent the transmission of opisthorchiasis. Therefore, the only practical measure to reduce the prevalence of cholangiocarcinoma is the treatment of O. viverrini infection with praziquantel [5]. However, the reliance on this sole treatment carries a significant risk that drug resistance will develop against this compound, as has been observed for trematocidal drugs in other flukes [9].
Clearly, understanding the intricacies of the biology of O. viverrini and opisthorchiasis is central to developing new and urgently needed intervention strategies. Yet, in spite of our knowledge of the morphological changes that occur in the parasite throughout its life cycle and its paramount importance as a carcinogen, we know very little about the molecular basis of the developmental biology of O. viverrini, its interactions with its hosts and the pathogenesis of disease, particularly carcinogenesis. The advent of next-generation sequencing and bioinformatic technologies [10], [11] now provides unprecedented opportunities to address some of these key areas. Recently, Young and coworkers [12] characterised the transcriptome of O. viverrini using 454 sequencing technology, which provided a solid basis to explore, for the first time, the transcriptional profiles of different developmental stages of this parasite. Logically extending this work, we now characterize differential transcription between adult and juvenile stages of O. viverrini using the method of RNA-Seq (Illumina technology) [13], and identify key molecules inferred to be associated with development, reproduction, feeding and survival of this neglected parasite.
Hamsters used in this study were maintained at the animal research facilities at the Faculty of Medicine, Khon Kaen University, Thailand. All work was conducted in accordance with protocols approved by the Animal Ethics Committee of Khon Kaen University (reference number #0514.1.12.2/23) in accordance with the ‘Animal for Scientific Purposes Act’ of the National Research Council of Thailand.
Using established methods [14], metacercariae were collected from infected cyprinoid fishes from the Khon Kaen province of Thailand and used to orally infect eight helminth-free, inbred Syrian (golden) hamsters (Mesocricetus auratus) in Khon Kaen University. Hamsters were euthanized 14 and 42 days following inoculation with metacercariae, in order to collect juvenile and adult stages of O. viverrini, respectively. Worms (all stages) were expressed from the intra- and extra-hepatic bile ducts and cultured in vitro for 2 h, to stimulate the regurgitation of their caecal contents [12] prior to being washed extensively in physiological saline (25°C), snap-frozen in liquid nitrogen and then stored at -80°C. The specific identity of the flukes was verified by first isolating genomic DNA [15] and then carrying out PCR-coupled, bidirectional sequencing (ABI 3730×l DNA analyzer, Applied Biosystems, California, USA) of the second internal transcribed spacer (ITS-2) of the nuclear ribosomal DNA [16]. These sequence data were compared by pairwise alignment to published sequence data in the National Center for Biotechnology Information (NCBI) GenBank archive (accessible via www.ncbi.nlm.nih.gov:AY584735).
Total RNA was isolated each from pools of adult (n = 15) or juvenile (n = 40) O. viverrini using the TriPure (Roche) reagent [17] and treated with DNase I (TurboDNA-free, Ambion) according to the manufacturer's instructions. In order to control for variation in transcription due to host-related factors (e.g., immunological/genetic differences), these pools were constructed from individuals isolated from worms collected from each of the infected host animals. Total RNA concentrations were estimated spectrophotometrically, and RNA integrity was verified by agarose gel electrophoresis and using a 2100 BioAnalyzer (Agilent). Polyadenylated (polyA+) RNA was purified from 10 µg of total RNA using Sera-Mag oligo(dT) beads, fragmented to a length of 100–500 nucleotides, reverse-transcribed using random hexamers, end-repaired and adaptor-ligated, according to a recommended protocol (Illumina). Ligated products of 200 bp were excised from an agarose gel, PCR-amplified for 15 cycles, purified over MinElute column (Qiagen) and sequenced (single-end) on a Genome Analyzer II (Illumina).
Following sequencing, the quality of all RNA-Seq data was assessed; only sequences with a PHRED score of ≥30 and a length of ≥40 nucleotides (nt) were retained. RNA-Seq data from the adult and juvenile libraries were combined with 454 data generated from adult O. viverrini in a previous study [12] and assembled using the program OASES v.0.1.21 [18]. The k-mer and coverage cut-off were optimized to achieve the best assembly with the greatest mean length of contiguous sequences and the fewest incomplete transcripts. Reads predicted to represent mitochondrial, ribosomal, host or microbial sequences, based on a BLASTn comparisons with sequences in the NCBI non-redundant database (accessible via www.ncbi.nlh.nih.gov), were removed from the dataset prior to subsequent analysis. Each transcript was conceptually translated in six frames, using customized PERL scripts, with the longest opening reading frame (ORF) being used to define the coding domain and its inferred amino acid sequence.
Following data assembly, the combined 454/Illumina transcripts were subjected to analysis/annotation using an established semi-automated, bioinformatic pipeline [12], [19] and compared (by BLASTx analysis at an E-value cut-off of <10−05, 10−15 and 10−30) with conceptually translated proteins from previously published transcriptomic data for O. viverrini [12], C. sinensis [12], Fasciola spp. [19], [20], as well as coding domains predicted from the genomes of Schistosoma spp. [17], [21], [22]. Functional annotation of each predicted peptide was inferred using multiple methods, including BLASTx comparison (cut-off: E-value: <10−5) with the non-redundant sequence database (March, 2012) available via GenBank (NCBI; http://www.ncbi.nlm.nih.gov/est/) and the UniProt database [23], prediction of conserved protein domains using InterProScan [24], allowing assignments of parental gene ontology (GO) terms (http://www.geneontology.org/) and homology-based mapping to conserved biochemical pathways in the Kyoto Encyclopaedia of Genes and Genomes (KEGG) using the KEGG orthology-based annotation system (KOBAS) [25]. Excretory/secretory (ES) proteins were predicted on the basis of the presence of a signal peptide at the N-terminus and absence of a transmembrane domain using the program PHOBIUS [26] and/or the identification of close homologues (BLASTp analysis: E-value cutoff <10−05) in the signal peptide database (SPD) [27] or a custom-built ES database containing published proteomic data for nematodes [28] and trematodes [17]. Peptidases and their inhibitors were predicted by BLASTx comparison with the MEROPS database (January 2011) [29]. The additional annotation of specific, key protein classes was achieved by BLASTx comparison with the KS-Sarfari and GPCR-Sarfari (http://www.sarfari.org) and the Transporter Classification (TCDB) [30] databases.
The transcriptome assembled here using both 454 and Illumina data was compared qualitatively and quantitatively with the 454-based transcriptome published previously for the adult stage of this species [12]. Both assemblies were compared based on common assembly metrics (i.e., number of contigs, N50 and N90 metrics [28], largest contigs, number of annotated proteins, mean predicted peptide length and largest predicted peptide) using established bioinformatic approaches [28]. One-to-one orthologous transcripts (transcript = a ‘true’ mRNA represented by the contigs constructed during the assembly of the 454 and/or 454+ Illumina data) were identified in the ‘old’ (i.e., 454) and ‘new’ (i.e., 454+ Illumina) assemblies using the reciprocal best-hit method [31] based on the BLASTn algorithm. Using this approach, we identified transcripts common to both datasets and unique to each, and then assessed the functional annotation data available for each of these transcript groups (i.e., common to both, unique to the ‘old’ assembly and novel to the ‘new’ assembly). Where possible, we attempted to assess the support for these BLASTn comparisons by mapping all Illumina reads to the contigs representing these sequences using the Burrows-Wheeler Aligner (BWA) program [32].
Because homopolymer errors (i.e., indels) represent a known limitation of 454 sequencing [33], and had been predicted in the previously published O. viverrini transcriptome [12], we explored the extent to which the addition of Illumina RNA-Seq data included here was able to correct these errors and the frequency with which such repairs improved/restored the predicted ORF of each contig. To do this, we conducted reciprocal pairwise Smith Waterman alignments of each contig in each assembly using the BWA-SW command in the BWA program [32]. These pairwise alignments were filtered for the best (i.e., most similar) alignment for each contig pair, and interrogated for insertions or deletions associated with homopolymers of ≥4 nucleotides (i.e., indels in the 454 only assembly that were corrected through the addition of the Illumina data).
Following the optimization of the ‘new’ transcriptome for O. viverrini, we explored differential transcription between the adult and juvenile stages of this species using the RNA-Seq data for each stage. To account for large differences in the numbers of reads generated for these two libraries [34], we randomly sub-selected sequence reads (n = 7,527,263) from each library and aligned ‘adult’ and ‘juvenile’ read pools to the final transcriptome using the program SOAP2 [32], requiring that each read mapped exclusively to one location in the transcriptome with a minimum alignment length of 40 nt and a maximum of three nucleotide mismatches per read. Relative levels of transcription in adult and juvenile stages were inferred based on the calculation of reads per kilobase per million mapped reads (RPKM) [35]; statistical differences in quantitative transcription between the juvenile and adult stages of O. viverrini was determined using a modified Audic-Claire equation [36] relating to a Bonferroni transformed p-value (i.e., False Discovery Rate) of ≤0.01 and ≥2-fold absolute difference in RPKM levels, as described previously [28].
To assess the expansion of the cysteine peptide domain proteins in O. viverrini relative to the other parasitic trematodes for which extensive transcriptomic/genomic data are available, we compiled all representative sequences in the present optimized transcriptome or available from transcriptomic data for C. sinensis [12] and Fasciola spp. [19], [20] as well as genomic data for Schistosoma spp. [17], [21], [22]. We extracted the nucleotide region encoding the C13 legumain-like (PF01650) or C1 cathepsin-like (PF01650) cysteine peptidase domains from each of these sequences, aligned these data using the program MUSCLE [37] through 50 iterations and manually verified this alignment by visual inspection in BioEdit (http://www.mbio.ncsu.edu/bioedit/). To eliminate redundancy, a complete or nearly complete transcript for each unique C13 legumain-like or C1 cathepsin-like domain detected in each alignment was retained as a representative of all sequences ( = contigs or singletons) assembled in these datasets. To assess evolutionary relationships between and among the C13 legumain-like or C1 cathepsin-like peptidases represented in this consensus alignment, phylograms were constructed by Bayesian inference (BI) using the program Mr Bayes v. 3.1.2 [38] employing the Monte Carlo Markov chain method (nchains = 4) over 1,000,000 tree-building generations, with every 100th tree being saved; 10% of the saved trees were discarded (burnin = 1,000 trees) to ensure stabilisation of the nodal split frequencies, and consensus trees for each peptidase family were constructed from all remaining trees, with the nodal support for each clade expressed as a posterior probability (pp). The consensus trees were generated and labelled in Figtree (http://tree.bio.ed.ac.uk/software/figtree/).
We used an RNA-Seq-based approach to improve the transcriptome of O. viverrini, and to explore differential transcription between the juvenile and adult stages of this parasite. Following sequencing and quality filtering (PHRED quality ≥Q30), we generated 14,862,797 and 7,527,263 single-end sequence reads (mean length: 49 bp) from juvenile and adult O. viverrini, respectively and deposited raw data under the accession number SRA052929 in the sequence read archive of NCBI (http://www.ncbi.nlm.nih.gov/sra). This RNA-Seq data was then combined with 642,918 454-based sequence reads (mean length: 373±133 bp) from a previously published ‘adult’ dataset [12]. This composite dataset was assembled into 24,896 contigs (see Table 1; mean contig length = 1068.76±1284.61 nt, longest contig = 20,661 nt, shortest contig = 100 nt); 17,357 of these contigs had a homologue in available transcriptomic data for C. sinensis [12], 13,035 or 12,480 in Fasciola hepatica or F. gigantica [19], [20], respectively, and ∼13,250 had a homologous sequence in genomic data available for Schistosoma spp. [17], [21], [22] (Table 2).
In a direct comparison with the ‘old’ O. viverrini transcriptome [12], 18,729 of the ‘new’ contigs had a close homologue (E-value cutoff: 1×10−5; Table 2). Based on comparative alignment of these homologous contig pairs (i.e., from the ‘old’ and ‘new’ assemblies), we identified 2,688 insertion and 1,311 deletion events associated with a homopolymeric region of ≥4 nt relating to 3,086 distinct transcripts. In total, 313,515 such homopolymers were detected, suggesting a total indel error rate of ∼1.3% in the previous dataset [12]. Correction of these indel errors coincided with the improved ORF lengths in the new assembly (n = 21,026; ≥50 amino acids [aa] in length), which were significantly longer than those achieved using 454 data alone (mean ORF length: 329 versus 168 aa, respectively; Table S1) and contained more information to facilitate functional annotation (e.g., 2863 unique PFAM domains in the 454+Illumina data assembly versus 2541 such domains in the 454 only dataset; Table S2).
Of the 21,026 contigs inferred to encode a peptide (≥50 aa) in the present assembly, 65.6% had a homologue (E-value cutoff: 1×10−5) in a eukaryote in the non-redundant protein database, with 9,827 predicted to encode at least one conserved protein domain (mean of 2.2 domains per sequence; Table 1) and, on the basis of these data, 37.1% of the predicted peptides could be assigned GO terms. For a more specific annotation, 6,277 peptides had an orthologous match in the KEGG database relating to 2,823 distinct KEGG orthologues and 249 conserved biological pathways (Table 1). Signal peptides and transmembrane domains were identified in 5,441 and 8,204 proteins of O. viverrini, respectively, and a total of 545 was inferred [28] to represent ES proteins (Table 1).
In addition to annotating these transcripts using such generalist resources, we interrogated specialist databases, in order to identify key protein classes, including kinases, transporters and channel proteins, receptors and peptidases, known to have important functional roles and being druggable in many parasitic helminths [39]. Using this approach, we annotated 333 kinases, 2,827 transporter and/or channel molecules, and 699 peptidases (Table S3). Among the peptidases, which are known to play a range of important functional roles in trematodes [40], the cysteine (58% of the peptidases), serine (20%) and metallo- (20%) peptidases predominated (Table S3), with the remaining classes (threonine and aspartate peptidases) being relatively rare. Interestingly, although metallopeptidases were relatively evenly distributed among ∼30 protein families (e.g., M1, M12B and M13), cysteine and serine peptidases were clearly dominated by a small subset of families, including the C1A (‘papain-like’: n = 62 transcripts), C2 (‘calpain-like’: n = 18), C13 (‘legumain-like’: n = 190), C19 (‘ubiquitin-specific’: n = 64), S1A (‘chymotrypsin’: n = 25) and S8A (‘subtilisin-like’: n = 51) peptidases. It is likely that much of this richness relates to alternative splicing of the mRNA. When the transcripts were clustered based on their ‘definitive’ peptidase domains (i.e., the region that allows classification to family), these transcripts were inferred to relate to 9 C1A, 6 C2, 5 C13, 20 C19, 5 S1A and 5 S8A domains based on homology to proteins in the MEROPs database.
The expansion of the C13 ‘legumain’-like molecules (Pfam code: PF01650) was of interest, considering that only one of them was predicted previously for O. viverrini [41] and a single ‘legumain’-like gene has been annotated for each trematode species for which extensive genomic or transcriptomic data are available (see Figure 1), with the exception of F. gigantica for which two such genes have been described [42]. In the present dataset, 190 transcripts had high BLASTx homology to 1 of 5 unique C13 peptidase domains in the MEROPs database. Subsequently, we conducted a multi-alignment of these transcripts and identified 9 unique C13-domain sequence types. It is likely that the discrepancy between the alignment and homology data relates to an under-representation of homologous sequences in the MEROPs database rather than a misidentification of the sequences themselves. However, to assess the support for the annotation of the C13 ‘legumains’, we conducted a phylogenetic analysis of a consensus alignment for each of the 9 unique C13-like domain sequences inferred from the O. viverrini transcripts as well as sequence data for homologues from all other parasitic trematodes for which extensive datasets are available (i.e., C. sinensis, Fasciola spp. and Schistosoma spp.) (see Figure 1). This analysis showed a clear radiation of the C13-like peptidases in O. viverrini, with all of the sequences described here forming a distinct, monophyletic clade. These novel C13 sequences appear to relate most closely to C13s described previously for the Opistorchiidae. The species of Fasciolidae (i.e., F. hepatica and F. gigantica) and Schistostomatidae (i.e., S. haematobium, S. japonicum and S. mansoni) also formed monophyletic clades by family. Posterior probability (pp) support for each of these major clades was 1.00.
In addition to refining the current assembly of the O. viverrini transcriptome, the RNA-Seq data generated here allowed, for the first time, a detailed assessment of quantitative differences in transcription between juvenile and adult stages of this species. Of the 24,896 contigs in the current assembly, 19,283 did not differ significantly in their observed levels of transcription between these two stages (Bonferroni transformed p-value: >0.01; RPKM difference ≤2-fold), whereas 3,020 and 2,593 transcripts were differentially transcribed (p-value≤0.01; RPKM difference ≥2 fold) in juveniles and adults, respectively (Table S3).
A significant percentage (11%) of molecules with increased transcription in juvenile O. viverrini encoded peptides associated with energy metabolism, including oxidative phosphorylation (e.g., cytochrome c oxidase and NADH dehydrogenase (ubiquinone) 1 alpha sub-complex 1), fatty acid metabolism and amino acid (e.g., valine, leucine, isoleucine and lysine) degradation (see Table S4, S5, S6). In the juvenile stage, increased transcription was observed for both secreted and non-secreted cysteine peptidases (e.g., families C13 and C1A; see Table S3 and S7), including molecules known to be involved in protein catabolism, proteolysis and lysosome-specific pathways in other species of fluke [43]. Of the expanded C13 legumain-like sequences predicted, six had increased transcription in the juvenile stage, as did four unique C1 cathepsin-like (PF01650) cysteine peptidase domains, each of which has close homology to cathepsin L genes in Clonorchis and/or Fasciola. To support the annotation of these transcripts, as for the C13 legumains, each unique C1 cathepsin-like domain was aligned with cathepsin L-like sequences from C. sinensis [44], Fasciola spp. [45] and Schistosoma spp. [46], and then clustered by Bayesian inference (see Figure 1). In total, we identified six unique C1 domains using this approach. Although three additional, unique domains were identified based on homology with data in the MEROPs database, these transcripts were not sufficiently complete to allow an assessment of the entire C1 domain and thus were not considered further. Based on these analyses, three of the six putative cathepsin L-like domains clustered with the other representatives of the Opistorchiidae (one juvenile-enriched, one constitutively transcribed and one adult-enriched) and three with the recognized cathepsin Ls from Fasciola (all juvenile enriched). Nodal support for these clusters was high in all instances (pp values ranging from 0.8 to 1.0).
Significantly increased transcription in the adult stage of O. viverrini related to nucleotide processing and oocyte meiosis. For example, GO terms that were highly represented in the adult included nucleoside, nucleobase and nucleotide kinase activity as well as ribonucleotide binding (for adenylate, nucleoside diphosphate and casein kinases) (see Table S4). Transcription linked to nucleotide replication and processing was significantly higher in the adult compared with the juvenile stage (Table S5). In addition, chromosome–specific proteins (e.g., histone-lysine N-methyl transferase and chromosome-transmission fidelity protein), DNA repair and recombination proteins (e.g., meiotic recombination protein and DNA polymerases) as well as proteins involved in the processing of mRNA ( = spliceosome), such as integrator complex subunit (INTS) and splicing factor (e.g., SFRS2), were enriched in adult O. viverrini (Table S3 and S6). Pathways associated with the nucleotide synthesis, including purine and pyrimidine metabolism, were also enriched in this developmental stage (Table S4). Furthermore, transcription associated with oocyte meiosis was higher in the adult stage (Table S3 and S7), and included transcripts encoding adenylate cyclases, ribosomal kinases and egg-specific antigens.
Opisthorchis viverrini (Trematoda; Platyhelminthes) is a socioeconomically important liver fluke that affects ∼9 million people in southeast Asia and, due to poorly understood mechanisms, is one of a small number of parasitic helminths known to directly cause malignant cancer [5]. In an effort to better understand the biology of this neglected parasite, transcripts from the adult stage were sequenced by 454 technology in a previous study [12], yielding ∼55,000 sequences (≥100 nt) predicted to encode ∼49,000 peptides of ≥50 aa. Although this published dataset [12] provided significant, new insights into the transcriptome of O. viverrini, the authors acknowledged that substantial gaps remained to be addressed. Specifically, the published assembly was described as being fragmented and, due to limitations in the sequencing chemistry used, potentially contained homopolymer-associated sequencing errors [12]. In addition, because only the adult stage of O. viverrini was represented and the initial study focused on a qualitative exploration of the transcriptome, quantitative assessment of transcription during the life-cycle of this parasite was not possible at the time.
In an effort to overcome these limitations and enhance the transcriptomic data available for this important parasite, we generated ∼25 million 50 bp (single-end) sequence reads using the Illumina platform, allowing transcription levels to be quantitated in the juvenile or adult stages of O. viverrini. An enhanced (‘new’) assembly was achieved from the combined output of the Illumina platform (including data for each stage) and the raw 454 data generated in the previous study [12]. Comparative analyses conducted here demonstrate that this approach improved considerably the assembly of the transcriptome. Notably, the number of contigs (≥100 bp) in the ‘new’ combined assembly (n = 24,869) was ∼31,000 fewer than in the previous assembly [12], and both the N50 and N90 assembly metrics increased substantially (by 360 and 225% respectively), with reductions particularly in the number of short sequences (i.e, <200 bp: n = 11,177 and 5,175 in the ‘old’ and ‘new’ assemblies, respectively) (Figure 1). This reduction in the number of sequences appears to relate specifically to an enhanced assembly and a reduction of redundancy overall, with 55.9% of the sequences and 80.4% of all contigs from the 454-only assembly having a close match in the new assembly. Indeed, of the remaining 24,330 ‘454-unique’ sequences (contigs or singletons), only 1,480 encoded peptides of ≥50 aa with homology to proteins in other eukaryotes (excluding likely contaminants, such as fungal or vertebrate sequences) in a functional database (Table S8). Importantly, the mapping of the raw Illumina reads to the ‘454-unique’ sequences indicated little coverage in the present dataset (Table S9), suggesting that the transcripts represented by these sequences were of low abundance in the adult and juvenile stages from which mRNA was purified. Notably, the 454 data were generated from a normalized cDNA library, in order to specifically enrich for such lowly transcribed sequences [12].
A comprehensive analysis of the enhanced dataset indicated that the transition from the juvenile to the adult stage of O. viverrini relates primarily to a down-regulation of metabolic pathways (likely in response to the reduced growth demands of the organism) and, predictably, an increase in pathways associated with DNA replication and reproduction. It is likely that the increased transcription of these molecules relates to the production of eggs, sperm and embryos, a hypothesis supported by the increased transcription in adults of key meiosis-related genes such as the meisosi specific serine/threonine kinase mek1. These findings are consistent with those reported previously for this developmental transition in other flukes, such as Schistosoma haematobium [17]. Many of the genes typically associated with highly specific reproductive functions (i.e., spermatogenesis) appear to be constitutively transcribed in the juvenile and adult stages investigated here. Exceptions to this relate primarily to the transcription of genes expressed in the terminal stages of the reproductive process, such as vitelline B (involved in egg-yolk production) [47] and homologues of the tyrosinase genes tyr1 and tyr2, known to be associated with late-phase egg shell synthesis in Schistosoma spp. [48]. Indeed, these genes were among the most highly transcribed genes in adult O. viverrini.
All hamsters used in the present study to produce O. viverrini were helminth-free and infected at the same time point, with the juvenile worms being harvested two weeks following inoculation. The pre-patent period for O. viverrini is at least four weeks [4], and all specimens yielded at two weeks were confirmed to be immature worms. Our data suggest that the transcription of many of the genes associated with the early phases of reproductive process (e.g., spermatogenesis and oogenesis) begins long before the worm reaches adulthood. However, an alternative hypothesis is that these genes have different functional roles at different stages during the life-cycle of O. viverrini.
The enhanced assembly of the transcriptome allowed greater insights into a variety of important and/or druggable groups of molecules, including receptors and transporters, kinases and peptidases. Most notably, we observed a substantial enhancement in the assembly of the complex cysteine peptidase families, including those representing C1 and C13, which are noted for their important functional roles in many helminths [40]. In particular, we assembled 367 transcripts encoding a conserved cysteine peptidase domain based on BLASTx homology, and further characterized particular families of ES molecules based on subsequent phylogenetic analysis of known homologues from other helminths. Based on our analysis, a striking expansion of the C13 legumains in Opisthorchis relative to all other parasitic helminths for which extensive genomic and/or transcriptomic data are available was detected. This expansion appears to relate primarily to an independent lineage positioned close to, but, clearly, distinct from the nearest homologous sequence in another helminth species, being the single C13 legumain-like peptidase identified in C. sinensis, although we did detect also a close orthologue of this sequence. Interestingly, transcription data relating to these sequences suggest that all but two of the C13 legumain-like sequences detected here for O. viverrini were significantly up-regulated in the juvenile relative to the adult stage. The legumains, or asparaginyl endopeptidases (AEPs), are known to cleave peptide bonds on the carboxyl-terminus of asparagine residues [49] and, in S. mansoni, have been implicated in the trans-processing and activation of cathepsin B haemoglobinase, thus enabling the degradation/digestion of host haemoglobin [50]–[52]. The asparaginyl endopeptidase sequences characterized here include typical histidine and cysteine residues, essential for enzymatic activity, suggesting that they are indeed functional. The specific divergence and radiation of these peptides in O. viverrini, and their transcription in the juvenile stage suggests an essential role during a critical phase of development. A likely hypothesis is that these molecules have radiated in O. viverrini to facilitate the exploitation of particular proteins in bile as a novel food source and/or cell detritus resulting from epithelial cell (cholangiocyte) turnover and/or immune cells undergoing diapedesis through the epithelium, particularly in chronically infected animals. The up-regulation of transcripts linked to these proteins is co-incidental with those associated with a variety of proteins involved specifically in metabolic pathways in the juvenile stage, which, at least circumstantially, supports this hypothesis.
Also notable among the cysteine peptidases were the cathepsin L-like proteins. Close homologues of individual sequences defined previously in C. sinensis [44] were identified and clustered using phylogenetic inference. However, intriguingly, we also detected specific, close homologues in O. viverrini of genuine cathepsin Ls reported for Fasciola spp. [43]. These sequences do not segregate into clades by species, rather are positioned on the tree as monophyletic pairs (one O. viverrini sequence grouping with one Fasciola sequence) based on the nearest homologue in Fasciola, suggesting a common evolutionary (i.e., orthologous) origin for each gene, rather than an independent radiation of this group of enzymes, as has been proposed previously for cathepsins of trematodes [43]. Despite this apparent orthologous relationship, we can only speculate as to the specific functional roles of these enzymes in O. viverrini. In Fasciola, cathepsin Ls appear to be directly involved in tissue penetration (cathepsin L3) and the digestion of host proteins during migration and development of the immature stage (cathepsins L1 and L2) and within the bile duct as adults (cathepsin L5) [43]. To our knowledge, O. viverrini does not penetrate host tissues, and, consistent with the understanding of the cathepsin Ls in Fasciola, we find no evidence of a homologue of cathepsin L3 in Fasciola. We did detect, however, putative orthologues of cathepsins L1 and L5. It is likely that these enzymes are critical also for feeding in Opisthorchis. For the O. viverrini cathepsin L1 orthologue, this inference is supported by differential transcription data, with the peptide being significantly up-regulated in the juvenile stage (relative to the adult), consistent with the transcriptional profile known for this molecule in Fasciola spp. [43]. In contrast, in Fasciola, cathepsin L5 has been reported to be up-regulated in adults, with little evidence of transcription in the juvenile stage [43]. However, in O. viverrini, the cathepsin L5 orthologue is clearly transcribed at a higher level in the juvenile stage, suggesting a possible difference in the role of this enzyme in the latter species, despite the apparent orthologous origin. Intriguingly, although the Fasciola-like cathepsin L sequences appear to be enriched in the juvenile stage, the three Clonorchis-like cathepsin Ls, are, generally, represented by a higher level of transcription, and two of them are specifically enriched in the adult stage. Unlike Fasciola, for which juvenile stages migrate through the liver parenchyma, both the juvenile and adult stages of O. viverrini live in the bile ducts of the host. Therefore, it is possible that O. viverrini has developed ‘stage-enriched’ suites of cathepsins that allow both of these stages to fill complementary, but non-overlapping niches within the host (e.g., exploiting similar but distinct food-sources), thus reducing inter-generational competition for resources/nutrients. Clearly, an in-depth exploration of the cathepsin Ls in O. viverrini would enable better insights into the feeding mechanisms of these parasites. Our proposals could be explored at a functional level, either through gene knock-down and knockout approaches, which are already established for some parasitic flukes, including O. viverrini [53], [54] or through the use of free-living ‘model’ flukes, such as Schmidtea mediterranea or Macrostomum ligano, in the same way that C. elegans has been employed as a (surrogate) functional tool for parasitic nematodes [11]. Given the utility of these model flukes as tools to study tissue and nerve regeneration [55]–[57], ongoing efforts to sequence their genomes and transcriptomes (of different developmental stages) should assist in the establishment of effective functional genomic tools for trematodes [58]. The development of such tools would provide major support toward the development of novel interventions against socioeconomically important trematodiases.
Through the use of Illumina-based sequencing, the present study provides a deep insight into the battery of cysteine peptidases that O. viverrini can deploy [43], specifically in relation to the radiation of the C13 legumains and the presence of cathepsin Ls orthologous to those of Fasciola spp. These findings contrast the existing hypothesis for the evolution of the peptidases in flukes (i.e., family-specific radiation/evolution) [43]. The finding that the cathepsin Ls of O. viverrini do not appear to have an homologue in C. sinensis, despite the close biological relationship shared by these species (both being members of the Opistorchiidae), suggests that radiation prior to differentiation of the major trematode families, followed by subsequent gene loss, may also have shaped the cathepsins in trematodes [43]. This interpretation is further supported by the presence of cathepsin F-like proteins encoded in the adult transcriptome of F. hepatica [19]. Notably, previous knowledge and understanding of these molecules in trematodes has largely been based largely on proteomic observations, which tend to be biased toward abundantly expressed proteins [59]. Indeed, taking into account only the highly transcribed sequences in the current dataset, the results of the present study are consistent with the existing hypotheses for opistorchiids (i.e., a reliance on cathepsin Fs, proliferation of cathespin Bs and a single C13 legumain-like homologue) and for the family specific radiation of these enzymes [43]. However, it is well established that one of the strengths of RNA-Seq technology is its ability to resolve both lowly transcribed sequences and alternative splicing events which are often not detectable using traditional technologies (e.g., microarray) or even sensitive proteomic tools [60]. This point is clearly worthy of note, given that RNA-Seq has not be widely deployed for the characterization of other socioeconomically important flukes (e.g., species of Schistosoma, Fasciola and Clonorchis). It may well be the case that similar expansions of the cysteine peptidases and specialization relating to the exploitation of specific food sources has occurred in a range of fluke species, but has, as yet, not been resolved due to the limitations of previous technologies. Clearly, given the key functional roles that many of these molecules play in fluke biology and pathogenesis [40], including in O. viverrini [61], [62], deeper exploration of fluke cathepsins using RNA-seq technology is needed. Coupling such investigations with expanded genomic sequencing of key parasites would provide much greater insight into the role that alternative splicing may play in the transcriptional biology of flukes; an area which to date has been explored only to a limited extent.
|
10.1371/journal.pcbi.1003630 | Quantifying and Predicting the Effect of Exogenous Interleukin-7 on CD4+T Cells in HIV-1 Infection | Exogenous Interleukin-7 (IL-7), in supplement to antiretroviral therapy, leads to a substantial increase of all CD4+ T cell subsets in HIV-1 infected patients. However, the quantitative contribution of the several potential mechanisms of action of IL-7 is unknown. We have performed a mathematical analysis of repeated measurements of total and naive CD4+ T cells and their Ki67 expression from HIV-1 infected patients involved in three phase I/II studies (N = 53 patients). We show that, besides a transient increase of peripheral proliferation, IL-7 exerts additional effects that play a significant role in CD4+ T cell dynamics up to 52 weeks. A decrease of the loss rate of the total CD4+ T cell is the most probable explanation. If this effect could be maintained during repeated administration of IL-7, our simulation study shows that such a strategy may allow maintaining CD4+ T cell counts above 500 cells/µL with 4 cycles or fewer over a period of two years. This in-depth analysis of clinical data revealed the potential for IL-7 to achieve sustained CD4+ T cell restoration with limited IL-7 exposure in HIV-1 infected patients with immune failure despite antiretroviral therapy.
| HIV infection is characterized by a decrease of CD4+ T-lymphocytes in the blood. Whereas antiretroviral treatment succeeds to control viral replication, some patients fail to reconstitute their CD4+ T cell count to normal value. IL-7 is a promising cytokine under evaluation for its use in HIV infection, in supplement to antiretroviral therapy, as it increases cell proliferation and survival. Here, we use data from three clinical trials testing the effect of IL-7 on CD4+ T-cell recovery in treated HIV-infected individuals and use a simple mathematical model to quantify IL-7 effects by estimating the biological parameters of the model. We show that the increase of peripheral proliferation could not explain alone the long-term dynamics of T cells after IL-7 injections underlining other important effects such as the improvement of cell survival. We also investigate the feasibility and the efficiency of repetitions of IL-7 cycles and argue for further evaluation through clinical trials.
| Human Immunodeficiency virus (HIV) infection is characterized by a profound depletion of CD4+ T cell numbers and function. Immune restoration with combination antiretroviral therapies (cART) has substantially improved patients' outcomes. Unfortunately, this restoration may be delayed, notably in patients starting treatment late, and/or incomplete, despite control of the viral replication [1]. Hence, immune therapy may be a complementary intervention to accelerate or improve immune restoration. Interleukin-7 (IL-7) is a cytokine produced by non–marrow-derived stromal and epithelial cells and is required for the development and persistence of T cells in the periphery [2], [3]. IL-7 may enhance thymopoiesis [4]–[6], as well as thymic-independent peripheral proliferation of recent thymic emigrants [7]–[9] and of more mature T cells [7], [9] even in the absence of cognate antigen [9]–[11]. Improved cell survival has also been shown in vivo [11]–[15]. In HIV-infected patients, a strong inverse correlation has been observed between plasma IL-7 levels and CD4+ T cell numbers as well as with CD4+ T cell reconstitution after initiation of antiretroviral therapy [16], [17]–[19]. Increased levels of IL-7 during lymphopenia are thought to be mainly the consequence of a decreased receptor-mediated clearance of IL-7 as the availability of receptors diminishes [20]. In addition, the IL-7 signaling on IL-7 receptor-a-positive (IL-7Ra+) dendritic cells in lymphopenic settings may diminish the homeostatic proliferation of CD4+ T cells [21]. A recent study has suggested that the remaining chronic inflammation in treated HIV-infected patients due to exposure to IL-1β and IL-6 may decrease T-cell sensitivity to IL-7 and therefore a reduced CD4+ T cell reconstitution [22]. Recent analyses of lymph node tissues have shown that collagen deposition may restrict T-cell access to IL-7, resulting in apoptosis and depletion of T cells [23]. This in turn leads to decreased production of lymphotoxin B, a trophic factor for reticuloendothelial cells, leading to their demise and loss of IL-7 producing cells. In summary, the IL-7 effect on CD4+ T-cell homeostasis is highly compromised in HIV infection [24].
The beneficial effect of administration of IL-7 on T cell homeostasis in patients with refractory cancer [9], in SIV infected macaques [15], [25], [26] and HIV infected individuals has been shown through several early trials [27], [28] and observational studies [29]. However, before proceeding with phase II and III trials, several questions remain. From a mechanistic standpoint, the respective contributions of thymic production [9], peripheral proliferation and survival [11] in the observed increase of CD4+ T cell count in HIV-infected patients are unclear. Also, the schedule of IL-7 administration, notably the frequency of cycling needed to reach optimal and durable CD4+ T cell restoration is not defined. Finally, the long-term effects of IL-7 therapy and repeated IL-7 cycles on T cell homeostasis in HIV-infected patients are unknown. To address these questions, we have developed a mathematical model to approximate the effect of IL-7 on CD4+ T cell homeostasis to fit the data from two phase I trials of IL-7 intervention in HIV-infected patients. This analysis is most consistent with a significant additional biological effect (on cell survival and/or thymic production) to the observed transient increase in peripheral cell proliferation. The predictions from the model have been used to explore the feasibility of repeated “maintenance” cycles of IL-7 administration with the aim of maintaining a given level of circulating CD4+ T cells.
Chronically HIV-1 infected patients with CD4+ T cell counts between 100 and 400 cells/µL and plasma HIV RNA<50 copies (c)/mL while on antiretroviral therapy were studied in three phase I/II trials (see Methods and Table 1 for characteristics). In Study I and II, there was a dose-dependent increase of CD4+ T cell count peaking between 14 and 21 days after the initial injection and followed by a steady decline. The peak increase ranged between 152 and 1202 CD4+ T cells/µL in the two studies [28], [30]. A significant increase compared to baseline (and placebo group in Study II) persisted until 12 weeks in the first study (Figure S1) and 52 weeks in the second (Figure 1). The main contributors to CD4+ T cell increase were naive and central memory cells [28], [30]. There was a transient increase of Ki67 expression (a marker of proliferation; see Methods) in all CD4+ T cell subsets during IL-7 administration (Figure 2). The peak of Ki67 expression was observed at the first available measurement after the initiation of IL-7 therapy, which is 14 days in study I and 7 days in study II. At 28 days, Ki67 expression returned to baseline in both studies.
This increase in cell proliferation, observed in parallel with the increase in CD4+ T cells, might be the only significant effect in vivo of the injection of exogenous IL-7. Indeed, IL-7 induces an acute cellular proliferation during a short time period leading to a rapid CD4+ T cell increase, followed by a slow return to baseline levels as CD4+ T cells die, explaining the observed dynamics. However, additional effects, especially on thymic output or cell survival, might exist and slow down the decline of CD4+ T cells. Recent thymic emigrants (defined as CD45RA+CD31high) and the sj/β T cell receptor excision circles (TREC) ratio (in Study II), which are both an indirect measure of thymic output [31], are significantly increased after IL-7 injections [28], [30]. In Study II, we also observed a decrease in PD-1 expression (a marker of cell exhaustion) by CD4+ T cells [30] suggesting an increased cell survival. Although these observations gave some insight in potential effects of IL-7 on T cell homeostasis, they do not quantify the respective contribution of these mechanisms to the observed CD4+ T cell dynamics in blood in terms of input and output of cells. This is why we embarked on a mathematical analysis to test whether the observed peripheral proliferation could explain the CD4+ T cell dynamics after IL-7 injections or if other additional biological mechanism played a significant role.
We used a simple mathematical model to investigate mechanistically the effect of IL-7 on total CD4+ and naive CD4+ T-cell dynamics (see Methods and Figure S2). Modeling CD4+ dynamics and Ki67+ expression by changing the proliferation rate during IL-7 administration provided a fair fit of the data of study II (Figure 3A, plain lines). Interestingly, there was a significant linear increase of estimated proliferation rates according to the dose group (p<0.0001; Figure 4A and 4B). However, we found a better fit of CD4+ dynamics with Model 2 (LCVa −0.173 vs. 0.937; Figure 3A, dashed lines) that includes an effect of IL-7 on proliferation rate during IL-7 administration and on loss rate after IL-7 administration. In addition to the significant dose-dependent increase of proliferation during IL-7, we estimated a decrease of the loss rate of quiescent cells from 0.061 to 0.044–0.049 per day corresponding to an improvement of the life span of about 25% from 16.4 days to 20.4–22.7 days (likelihood ratio test p-value<0.001; Figure 4, Table 2 and Table S1). This result was found with both formulation of IL-7 (with either rh-IL-7 or glycosylated rh-IL7; Table S2). Adding a modification of the constant production of CD4+ (Model 3) rather than a modification of loss rate (Model 2) did not substantially improve the fit to total CD4+ T cell dynamics as shown in Figure 3A where the fits from the two models overlap (see also Figure S3 and S4). In other words, although the Model 2 that includes a modification of quiescent cell loss rate was better from a statistical point of view (LCVa = −0.131 vs. −0.173), it was difficult to distinguish the fits of the two models. Interestingly, all models described correctly the initial increase of CD4+ T cells and thereafter, Model 1 predicted a slower decline of CD4+ T cells than Model 2 and 3. This poorer long-term fit might be explained by a transient effect on the proliferation rate (until day 16) that altered only briefly the equilibrium while the lingering effect on the production or loss rate changed it in the long-term.
To further analyze the potential effect of IL-7 on thymic output, we explored the effect of IL-7 on the naive (CD45RA+CD27+) CD4+ cells (either Ki67+ or Ki67−) using the available data for this subset (until 12 weeks). Here again, we found that an additional effect of IL-7 after its administration either on the thymic production or the loss rate significantly improves the fits compared to a model including only an effect on the proliferation rate (Table S3). Interestingly, we found that the best model was the one including an effect on the thymic production rate of naive cells after IL-7 administration in addition to the proliferation rate (LCVa = 1.705 vs. 1.760 for the model including an effect on the loss rate after IL-7 administration; Table 2 and Table S3). Both models including an additional effect of IL-7 were better than the model with an effect on proliferation only (LCVa = 1.832). However, a change in loss rate of quiescent cells led to a good fit as well and individual fits from Model 2 and 3 were very close as shown in Figure 3B and S5.
Finally, we were interested in the ability of Model 2 (with the effect of IL-7 on proliferation and loss rates) to predict individual responses to IL-7. We made use of data from 12 additional patients (from INSPIRE 2, Table 1) treated with a 20 µg/kg dose as per the INSPIRE study. We used only the first two measurements of total CD4+ T cells and Ki67+ cells to compute the Empirical Bayes estimates for each parameter that could vary between patients. The other population parameters were fixed according to the previous estimations (Table S1). We then predicted the individual CD4+ T cell dynamics until week 12. Most of the observed total CD4+ T cell counts were in the prediction interval (Figure S6). Therefore, the model that includes an effect of IL-7 on proliferation and loss rates led to a good description of the total CD4+ T cell dynamics and a fair predictive ability at the individual level.
To our knowledge, no data exists yet on the effect of repeated cycles of IL-7 in vivo. Therefore, to investigate to what extent IL-7 administration might sustain CD4+ T-cell restoration, we artificially created data and compared different scenarios allowing the IL-7 effect to wane after subsequent injections compared to the initial one (see Methods). In this part, we considered an extended version of the mathematical model that incorporates a homeostatic proliferation (see Methods). This model gave similar results as presented in the previous section (not shown) but more realistic long-term dynamics for repeated IL-7 administrations (total CD4+ T cell counts staying below 1500 cells/µL).
As the dose 20 µg/Kg was the one recommended for further phase II/III studies [30], we simulated CD4+ T cell dynamics with hypothetical repeated cycles of IL-7 administration for each patient who received this dose in the Study I (INSPIRE): namely 14 patients. The parameter κ controlling the proliferation rate was fixed to the same value for each patient (see Methods). The initial CD4+ dynamics was the one predicted by the Model 2 as presented in the previous section. CD4+ T cell count were assumed to be measured every three months and when it dropped below 500 cells/µL a cycle of IL-7 administration was simulated (one injection per week for 3 weeks). The dynamics were therefore based in part on fit to observed data and in part on a predicted response to repeated therapy. We analyzed two primary outcomes: the time spent above 500 cells/µL and the number of cycles (including the first cycle) needed to maintain CD4+ T cell counts above 500 cells/µL over 2 years. A secondary outcome was the median time between two successive cycles.
Simulations were performed according to several scenarios varying from a constant effect after each cycle and decreasing effects on loss and proliferation rates. At each cycle, we assumed that the effect of IL-7 on the loss rate of CD4+ T cells started to decrease 3 (or 9) months after the last injection and disappeared after 1 (or 2) year. Table S3 shows some scenarios ordered from the best to the worst according to the primary outcomes. Where all the IL-7 cycles were assumed to keep 100% of their effects on CD4+ T cell counts (Scenario A-1 and A-2; Figure 5A and Table S4), the intervention was highly effective. CD4+ T cell counts were maintained above 500 cells/µL between 84% and 91% of the time compared to 11.5% when no new IL-7 cycle was administrated during the 24 months of follow-up. Moreover, the time between two cycles of IL-7 was estimated to be greater than 6 months. We also investigated reduced effect of IL-7 on peripheral proliferation and/or loss rate of non-proliferating cells during subsequent cycles. The scenario assuming that during repeated cycles only 50% of the effect on proliferation was effective while the effect on the loss rate was conserved (scenario B-1 and B-2; Figure 5B and Table S4) gave similar results as the scenario assuming a full and constant effect of proliferation and loss rate (Scenario A-1 and A-2; Table S4). Moreover, we observed that the more the effect of IL-7 on the loss rate was reduced, the more often cycles have to be administrated and the shorter the time between two successive cycles (Figure S7). These results suggest that it is more important to keep a strong effect on the loss rate of non-proliferating cells than on proliferation. As shown in Figure S7, when the effect of IL-7 is only maintained on the proliferation rate (i.e. no more effect in the loss rate of non-proliferating cells: αμQ = 0), we predicted the highest number of injections, whatever the size of the effect on the proliferation rate, to sustain CD4 count above 500 cells/µL. Overall, in comparison to no repeated cycle (Reference scenario in Table S4) whatever the assumed effect of IL-7 during successive cycles, the CD4+ T cell count may be durably increased with clinically realistic administration schedules. Therefore, IL-7 repeated cycles seems feasible and efficient.
We report here a mathematical analysis of total and naive CD4+ T cell dynamics in HIV-1 infected patients treated with antiretrovirals who experienced a significant increase of CD4+ T cell counts while receiving IL-7 therapy. We confirm, once again, that IL-7 induces a significantly increased peripheral proliferation of CD4+ T cells as measured by Ki67 expression. However, results presented here extend our knowledge on the in vivo effects of IL-7 by showing that this increased peripheral proliferation alone could not explain the long-term changes in CD4+ T cell number that were observed. An increase of the production rate of naive CD4+ T cells and a decrease of the loss rate for total CD4+ T cells might also contribute to T-cell homeostasis during IL-7 therapy. Importantly, baseline parameters such as naive T cell production (around 9×108 naive cells/day) [32], loss rate of proliferating cells (0.08 day−1) [33] or reversion rate (accounting for duration of division and duration of Ki67+ expression) were in agreement with current knowledge. Furthermore, our mathematical model shows good performance for individual predictions and provides insights on the feasibility of repeated cycles of IL-7 (or long-lasting formulation) for maintaining CD4+ T cell counts in HIV-infected patients. Predictions from the mathematical model underline the importance of an additional effect of IL-7 beyond peripheral cell proliferation for long-term CD4+ T cell responses.
HIV infection leads to a profound disturbance of T cell homeostasis with an increased turnover of these cells [33]–[35]. The naive T cell pool is replenished mainly by post-thymic proliferation in adults [36], [37] but the observed proliferation of naive CD4+ T cells is not enough to prevent the slow decline of these cells in HIV infected patients. Augmenting immunity with exogeneous cytokines has been attempted in HIV infection; and although CD4 T cell numbers were increased with administration of IL-2, two large clinical trials failed to show any evidence of clinical benefit [38]. The failure of IL-2 therapy to confer clinical benefit despite CD4 T cell increases could be attributed at least in part to the regulatory phenotype of the expanded cells [39] and the possibility of enhanced inflammation and coagulation during administration [40]. The effects of IL-7 on the other hand are fundamentally different [13] and the defined role of IL-7 in maintaining T cell homeostasis in health provides rationale for testing its therapeutic administration in HIV infection complicated by immune failure [41].
Our data argue for an increase in cell survival after IL-7 administration. Our results show that CD4+ T cell dynamics are better explained by a decrease of cell loss in addition to the transient peripheral proliferation. This finding is consistent with previous findings that increasing cellular survival through up-regulation of bcl2 expression is a physiological function of IL-7 [13], [42]. Moreover, increases in T cell survival after IL-7 injection have been demonstrated in monkeys using BrdU labeling [15]. Our findings warrant further study to define the precise mechanisms of IL-7 induced cell expansion in HIV infection.
Improvement of thymopoiesis has been reported during exogenous IL-7 administration [5], [6]. However, there is some controversy on the importance of this effect in vivo [9], [43]. Sportes et al. showed a modest increase of absolute numbers of TRECs and a major dilution of TREC content due to peripheral cell proliferation leading to the conclusion that the increase of TCR repertoire diversity is mainly due to the proliferation of recent thymic emigrants. The effects of IL-7 on thymopoiesis may be dependent, on one hand, on the underlying disease (HIV-1 infection, cancer and chemotherapy) and, on the other hand, on the duration of cytokine therapy [9]. In our simulations of repeated IL-7 cycles, we did not consider any effect of IL-7 on thymopoiesis because we favoured the hypothesis of an effect on cell survival according to the rationale above and the slightly better fit. Therefore, if IL-7 administration improved thymopoiesis in addition to peripheral proliferation and enhanced cell survival, the CD4+ T cell response should have surpassed our predictions.
For long-term predictions using simulations, the initial model was extended by adding a homeostatic control of proliferation after IL-7 administration. This can be related to the modulation of IL7Ra expression that prevents uncontrolled proliferation [44], [45]. Furthermore, the effect on T cell loss was also modeled to wane over time. Strikingly, the estimation of the duration of the effect of IL-7 on cell survival was prolonged up to 2 years. Likely, this could not be explained by the pharmacokinetics of exogenous IL-7 that was administrated during two weeks only. However, exogenous IL-7 is known to bind to components of the extracellular matrix resulting in saturation of this tissue compartment followed by a slow release of IL-7, which may exert long-lasting effects [3]. Also, IL-7 could have persistent effects on cellular homeostasis by normalizing tissue architecture through decreasing fibrosis in the gut [46] and in lymph nodes [23], [47], [48], [49] and thus improving cellular access to survival signals. Other potential activities such as effects on cell trafficking [27], [50], IL-7 antibody formation, switch to memory phenotypes [14], [51] or impact on proviral HIV DNA content [52] have not been taken into account in this model. Redistribution of CD4+ T cells to tissues, leading to a transient decrease of CD4+ T cells levels in blood, is mainly observed in the first days after IL-7 administration and should not affect measurements made thereafter. Although neutralizing anti-IL-7 antibodies were not observed in patients following the first cycle of IL-7 [28], their induction after repeated administration could attenuate the effects that we modeled here. For these reasons, new clinical trials are needed to help distinguishing the persistent IL-7 effect(s) involved in CD4 recovery in HIV-infected patients and to propose personalized therapy in the future [53]. Furthermore, we assumed repetition of cycles (i.e. three injections over two weeks) but the repetition of single injections could lead to similar results shown in the simulations if the effect of one injection respect the assumptions made in some scenarios. For instance, the repetition of a single injection may have a reduced effect on proliferation and cell survival compared to a whole cycle but this would still lead to a good maintenance of CD4+ T cell counts.
There are limitations to this study that include the restricted number of harvest times and the lack of validated markers for cellular lifespan. One way to overcome these limitations and to help distinguishing between increased production and increased survival is to perform studies that include in vivo labeling with deuterium and TREC content measurements. Indeed, deuterium labeling is a recent and powerful tool to estimate cell turnover [54] and used in combination with TREC content allows estimation of thymic output [37], [55]. Also, it may have been relevant to distinguish the dynamics in lymph node tissue to better capture long-term effect of IL-7 although data on lymph node tissue would be difficult to obtain. Despite these caveats, the goodness of fit and the predictive capacity of the model provide important insights for further development of IL-7 treatment strategies. We have learned that IL-7 administration leads to a burst of peripheral proliferation that is likely associated with a lingering effect beyond the period of IL-7 administration. We surmise that a durable effect of IL-7 on T cell homeostasis could be achieved after repeated administration but safety and activity need to be confirmed [2], [3].
Data were generated in three phase I/II studies (Table 1). All participating institute's Institutional Review Boards approved the studies and the procedures and all participants provided written informed consent before study participation. The rh-IL-7 study (referred to as Study I) [28] evaluated Recombinant Human Interleukin 7 (rh-IL-7), a nonglycosylated protein composed of 153 amino acids, and included 14 HIV-infected patients receiving antiretroviral therapy whose CD4+ T cell counts were between 100 and 400 cells/µl and whose plasma HIV RNA levels were less than 50 copies/ml. Patients received a total of 8 subcutaneous injections of 2 different doses of recombinant human IL-7 (3 or 10 µg/kg, dose 1 and 2, respectively) 3 times per week over a 16-day period. Eleven repeated measurements of total CD4+ T cells up to 48 weeks and four measurements of Ki67+ positive T cells among CD4+ T cells up to 12 weeks were performed.
The INSPIRE Study (referred to as Study II) [30] evaluated 3 weekly subcutaneous (SC) injections of a purified glycosylated 152 amino acid rhIL-7 (CYT107 over a period of 2 weeks). Three doses were tested: 10, 20 and 30 µg/Kg/week. Seven, 8 and 6 patients received three injections (one per week) in each dose group, respectively. Two HIV-infected patients were randomized per dose level and received a placebo (NaCl). Visits for safety and immunologic evaluation were performed at days 7, 14, 21, 28, 35, week 9 and week 12 and then quarterly up to week 52. In the INSPIRE 2 Study (referred to as Study III), 12 patients received 3 subcutaneous injection of 20 µg/Kg/week CYT107 (one per week over 2 weeks) and were followed up to week 52.
Absolute CD4 T-cell counts, T cell expression of Ki67, and the proportions of naive subsets defined by expression of CD45RA and CD27 (naive: CD45RA+CD27+) were measured in whole blood by flow cytometric assays within 6 hours of blood draw in the Rh-IL7 study [28]. In INSPIRE, naive cells in cryopreserved samples were identified by expression of CD45RA, CD27 and CCR7; in INSPIRE 2 naive cells were enumerated in cryopreserved PBMC by expression of CD45RA and CCR7.
Ki67 is a cellular marker of proliferation [56] and is associated with cell proliferation. It is present during all active phases of the cell cycle (namely G1, S, G2 and M) and it is absent from resting cells (phase G0). Therefore, some of the cells expressing Ki67 are actually in the division phase M and the rest are “on their way” to this phase or very recently in this phase. In this model, we assume that proliferating cells express Ki67 whereas non-proliferating (i.e. resting) cells do not; this is a relatively good approximation. We consider the following mathematical model including two populations of cells (see Figure S2 for a general cartoon of the model): non-proliferating cells (Ki67−, denoted Q) and proliferating cells (Ki67+, denoted P):
Non-proliferating cells (Q) are produced at a constant rate λ. They become Ki67+ at rate π and die at rate μQ. Proliferating cells (P) die at rate μp and lose their proliferation marker Ki67 at rate ρ (cells express Ki67 for 1/ρ days). We assumed that 2ρP cells enter the Q compartment that is two daughter cells are produced after one single cell cycle. The loss rates (μP and μQ) are influenced by cell survival but also by any redistribution between blood and other tissues. Before the first injection at t = 0, both populations are assumed to be at equilibrium (i.e. and ). This model was used for total CD4+ T-cells where Q and P include both naive and memory CD4+ T-cells. Similarly, we used this general model to describe naive CD4+ T-cell dynamics adding a superscript ‘N’ to all parameters (λN, πN, μQN, μPN, ρN) and where QN represent non-proliferating naive CD4+ T-cells and PN proliferating naive CD4+ T-cells.
To model the effect of IL-7, we considered different models allowing some parameters to change over time (i.e. during or after the IL-7 administration). We distinguished three models:
Model 1 includes only an effect on the proliferation rate π during the IL-7 administration
Model 2 includes an effect on π and an additional effect on the loss rate of non-proliferation cells μQ (effect tested during or after the IL-7 administration)
Model 3 includes an effect on π and an additional effect on the production rate λ (effect tested during or after the IL-7 administration).
Models including an effect of IL-7 on the duration of Ki67 expression or on a direct production of Ki67+ cells from the thymus were also tried but did not improve the fit of the data (not shown).
Each parameter is assumed equal to a baseline value (denoted π0, μQ0, λ0) and we tested a possible effect of IL-7 and a possible dose effect. For instance, the proliferation rate π was assumed equal to before (t = 0) and after IL-7 administration (t>τ) and to during IL-7 administration (0<t≤τ). The variable trt indicates if an individual received the placebo (trt = 0) or IL-7 injections (trt = 1), hence the parameter η0 represents the possible effect of IL-7 on the parameter π. Similarly, we tested a possible dose effect via the parameter η1 and the variable dose; dose = 0 if the individual received the placebo and 0.3, 1, 2 or 3 according to the received dose of IL-7. The additional dose effect is here assumed to be linear: the dose 30 µg/Kg will have 3 times more effect than the dose 10 µg/Kg and 10 times more than the dose 3 µg/Kg. The time τ (time of IL-7 effect on the proliferation rate π) was fixed to 16 days and robustness analyses with values between 14 and 16 days have been performed leading to similar results but slightly worse likelihood (not shown).
For the simulations of repeated administrations of IL-7, a homeostatic control of proliferation was incorporated in order to constrain the CD4+ T cell level in a credible range (below 1500 cells/µL). The newly defined rate of proliferation, denoted π*, is defined as follows:where N0 is the baseline CD4 T cell count, Nt is CD4 T cell count at time t and κ is a number lower than 1 estimated on the data. As N0 is relatively low due to lymphopenia, cells will be allowed to proliferate more while the circulating number of CD4+ T cells is still relatively low but as soon as this number deviates too much from the baseline value the proliferation rate is reduced [57]. Several formulations have been tested, including one with Nmax = 1000 referring to a normal healthy value rather than N0 and all formulations led to similar conclusions due to different estimates of the parameter κ (not shown). The parameter κ was estimated by profile likelihood because it was difficult to estimate it at the same time as the other parameters. The parameter κ was therefore fixed at different plausible values and the model was estimated for each value of κ and we kept the value of κ for which the model had the lowest likelihood.
Here, we only considered Model 2 including the effect of IL-7 on the proliferation rate and the loss rate of non-proliferating cells as it was the best model according to real data. Moreover, in this model, the loss rate of non-proliferating cell (μQ) was assumed decreased (to a constant value) after day 16 and during (Tfull-16) days after the injection and then linearly increased to its baseline value after Tend days (see Figure S2). As subsequent cycles of IL-7 might have reduced effects compared to the first one, the proliferation rate (resp. loss rate) of the subsequent cycle is denoted by πsub (resp. μQsub) and keep π (resp. μQ) for the first injection. These rates are defined as πsub = αππ and μQsub = αμQμQ, where απ and αμQ represent the strength of new IL-7 administration (i.e. the percentage of effect of πsub and μQsub compared to the initial injection). Therefore, we assumed for all subsequent cycles a similar reduction of IL-7 effects compared to the first cycle.
Parameters were estimated using maximum penalized likelihood that takes into account unbalanced data due to sparse missing values and the availability of Ki67+ staining up to week 12 (no measurements were available beyond that time). This method can be viewed either as a version of maximum likelihood allowing taking into account prior knowledge (from previous estimates found in published studies), or as an approximation of Bayesian inference [58], [59].
The Ordinary Differential Equations (ODE) system was solved with dsolve from the ODEPACK [60] for stiff system using Backward Differentiation Formula (BDF) methods (the Gear methods). Each parameter (θi) was modeled as the sum of a population (fixed) parameter (β) and a random effect (bi) allowing the parameter to be different from one patient to another: Each random effect was assumed to be normally distributed with a variance to be estimated: . A stepwise selection procedure was used and when the variance of a given random effect was not significant, the parameter was considered as fixed in the next model. We observe the number of proliferating cells (P) and the total number of cells (P+Q) plus a measurement error adding two unknown parameters and . The final models included only two random effects, one for λ (production rate) and one for ρ (the rate at which proliferating cells go back to rest). The best model was selected using an approximation of the likelihood cross-validation criterion (LCVa, [58], [59]) that is based on the likelihood weighted by the number of parameters estimated like the Akaike Criteria (AIC). The lower is the value of the criteria the better is the model. Individual predicted trajectories were computed using the Parametric Empirical Bayes (PEB) for all parameters having a random effect [61].
|
10.1371/journal.pgen.1005611 | Anaplastic Lymphoma Kinase Acts in the Drosophila Mushroom Body to Negatively Regulate Sleep | Though evidence is mounting that a major function of sleep is to maintain brain plasticity and consolidate memory, little is known about the molecular pathways by which learning and sleep processes intercept. Anaplastic lymphoma kinase (Alk), the gene encoding a tyrosine receptor kinase whose inadvertent activation is the cause of many cancers, is implicated in synapse formation and cognitive functions. In particular, Alk genetically interacts with Neurofibromatosis 1 (Nf1) to regulate growth and associative learning in flies. We show that Alk mutants have increased sleep. Using a targeted RNAi screen we localized the negative effects of Alk on sleep to the mushroom body, a structure important for both sleep and memory. We also report that mutations in Nf1 produce a sexually dimorphic short sleep phenotype, and suppress the long sleep phenotype of Alk. Thus Alk and Nf1 interact in both learning and sleep regulation, highlighting a common pathway in these two processes.
| Animal and human studies suggest that sleep has a profound impact on learning and memory. However, little is known about the molecular pathways linking these phenomena. We report that mutations in the Drosophila Anaplastic lymphoma kinase (Alk) gene, an ortholog of a human oncogene ALK, cause increased sleep. ALK is required for sleep suppression in the mushroom body, a structure important for both sleep and memory. ALK generally activates the Ras/ERK pathway, which is negatively regulated by Neurofibromin 1 (NF1). Mutations in Nf1 are the causes of the common neurological disorder Neurofibromatosis type 1 (NF1), which affects 1 in 3,000 live births. We find that male flies lacking the NF1 protein have reduced sleep, a phenotype opposite that of Alk flies. Interestingly, even though mutations in Nf1 don’t always cause short sleep in female flies, they suppress the sleep increase induced by ALK inactivation. Previous studies have shown that Alk and Nf1 play antagonistic roles in learning and that both genes regulate synaptic growth. Thus Alk and Nf1 interact to regulate both sleep and learning, suggesting that the two processes share a common pathway. Our results support a model in which changes in synaptic plasticity during sleep promote learning and memory.
| Sleep behavior is conserved from worms and insects to fish and mammals [1]. Why animals spend a large amount of time seemingly doing nothing, passing on opportunities to forage, hunt or mate, and remaining vulnerable to dangers, is still a mystery. It is hypothesized that a major function of sleep is to maintain brain function, in particular, to ensure synaptic homeostasis of neurons and to consolidate memory [2,3]. Levels of synaptic proteins are associated with sleep/wake states in both flies and mammals [4,5], and sleep deprivation impairs memory formation in a variety of species, including humans [6], mice [7], and Drosophila [8,9]. In addition, some molecules that regulate learning and memory turn out to be required for sleep/wake regulation [10]. However, only a handful of such molecules have been identified and for most it is not known if effects on the two processes are mechanistically linked.
Anaplastic lymphoma kinase (Alk), which encodes a member of the ALK/LTK (leucocyte tyrosine kinase)) family of receptor tyrosine kinases (RTKs), is proposed to play important roles in the nervous system based on its extensive expression in the CNS of both mammals and flies [11–14]. Its in vivo functions are mostly studied in the context of Drosophila development. Together with its secreted ligand Jelly Belly (Jeb), ALK is essential for 1) gut muscle differentiation [15,16]; 2) retinal axon targeting in the optic lobe [17]; 3) growth and organ size regulation [14,18]; and 4) modulation of neuromuscular transmission and synapse growth at larval neuromuscular junctions (NMJ) [19]. However, there is also evidence for a role of Alk in brain plasticity in adult contexts. Adult-specific activation of Alk causes deficits in associative olfactory learning in Drosophila; concordantly, reducing neuronal Alk activity in adult flies enhances olfactory learning [14]. Similarly in mice, loss of ALK function enhances spatial memory and novel object recognition, and reduces anxiety and depression [20,21]. Effects of Alk on learning, at least in flies, are most likely mediated by Ras/ERK signaling. This is supported by an interaction with Nf1, conserved ortholog of the human Neurofibromatosis type 1 (NF1) disease gene, which encodes a GTPase-activating protein (GAP) that negatively regulates Ras/ERK signaling. Specifically, the learning deficit in Nf1 flies is rescued by down regulation of Alk [14].
Based on the role of Alk in neuronal plasticity and learning, we hypothesized that Alk may be involved in sleep regulation. We found that inactivation of ALK causes increased sleep. We probed for the neuronal circuit that underlies Alk’s involvement in sleep and found that inhibiting Alk in the mushroom body induces more sleep, suggesting Alk as a mechanistic link between learning and sleep. In addition, Alk interacts with Nf1 in the regulation of sleep just as it does in the context of learning. Nf1 also have circadian phenotypes [22], but Alk is not required for circadian rhythms, nor does it interact with Nf1 in the circadian regulation of rest/activity rhythms. Thus, interactions between the two molecules are specific for sleep and learning.
Because ALK plays a crucial role in gut development, the null allele Alk1 is homozygous lethal at the early larval stage [23]. However, a temperature-sensitive allele, Alkts, fully complements Alk1 at 18°C, but fails to complement developmental Alk1 lethality at 29°C [19]. We were therefore able to raise Alkts/1 trans-heterozygous or Alkts homozygous flies to the adult stage at 18°C and prevent developmental phenotypes such as changes in body length (S1 Fig). Alk mutants raised in this manner are presumably also spared other developmental defects seen with manipulations of ALK activity, such as altered NMJ structure and function [19] and mis-targeting of retinal neurons in the optic lobes [17].
We assayed sleep using the traditional single infrared beam interruption device (Trikinetics, MA) in adult Alkts/1 and Alkts female flies at the permissive temperature of 18°C and at the restrictive temperature 29°C. Because genetic background has a profound impact on sleep [24], Alk mutants were backcrossed for five generations into a white (w) isogenic background, iso31, a line generated specifically for use in behavioral experiments [25]. At 18°C, control iso31 and Alk flies had very similar sleep patterns. However, acute inhibition of Alk by switching the environmental temperature to 29°C drastically increased sleep in Alk flies as compared to iso31. In iso31 flies, the shift to 29°C initially increased daytime sleep and decreased nighttime sleep, consistent with previous reports of increased siesta at higher temperatures [26]; overall sleep increased on the third day, but there was no net change in sleep over the three day period. Inhibition of Alk increased both day and night sleep. This temperature-sensitive sleep phenotype was reversed by lowering the temperature back to 18°C (Fig 1A). Quantification shows that Alkts/1 flies slept ~51.06±10.09% more than the control iso31 flies during the high temperature shift (6 independent experiments, n = 87 for iso31 and n = 80 for Alkts/1). Similarly, Alkts homozygous flies and flies that harbor an Alkts allele over a deficiency uncovering the Alk gene slept more than control flies at 29°C. There was no difference in total sleep amount between Alkts/1, Alkts, or Alkts/Def flies, suggesting that the restrictive temperature completely abolished ALKts protein function (Fig 1B). Importantly, we were able to rescue the sleep phenotype of Alk mutants by re-expressing Alk transgenically (discussed below).
Though sleep profiles and sleep metrics produced by the standard single infrared beam sleep monitors are reliable and have been published widely, they sometimes overestimate sleep as they miss fly movement away from the infrared beam; the degree of error varies from one genotype to another [27,28]. We therefore assayed sleep in a new multi-beam sleep monitor (Trikinetics, MA), which provides an order-of-magnitude higher spatial resolution compared to the traditional single beam monitors. Measurements by multi-beam monitors validated our results from the single beam method, showing that Alkts/1 females had increased daytime and nighttime sleep at 29°C as compared to iso31 control flies (Fig 1C and S2 Fig). However the long sleep phenotype was less pronounced when assayed with multi-beam monitors, with Alkts/1 female flies showing a 32.3±3.3% increase over iso31 (3 independent experiments, 47 iso31 flies and 48 Alkts/1flies). We also assayed sleep in Alk male flies with both monitor systems, and found that it was similarly increased (S3 Fig). Henceforth we focused our analysis on female flies, typically the gender studied in Drosophila sleep experiments. To exclude the possibility that the apparent increase in sleep was due to locomotion impairment, we measured waking activity, calculated as the average number of beam crossings per waking minute. We found that it did not account for the long sleep phenotype as it was reduced in both long-sleeping Alkts/1 flies as well as normal-sleep Alk1/+ controls (Fig 1D). We also assayed the mobility of Alk flies in an independent negative geotaxis assay that measures the ability of flies to climb vertically when startled [29]. The response of Alkts/1 flies was indistinguishable from that of iso31 flies (Fig 1E), suggesting that Alkts/1 flies have no gross motor defects. To confirm that the increased inactivity in Alk mutants is genuine sleep, as opposed to quiet wake, we assayed their arousability to a mechanical stimulus at different times of day (Fig 2). Similar percentages of previously sleeping iso31 and Alkts flies were aroused by the stimulus at ZT6 and ZT20, when most flies were sleeping. At ZT22, more iso31 flies were aroused than Alkts flies, probably because iso31 flies were transitioning from sleep to wake at this time point. Notably, a higher percentage of the previously awake flies exhibited activity after the stimulus than sleeping flies at all three time points, suggesting that arousal threshold was indeed higher in sleeping flies. We conclude that Alk mutants are bona fide long sleepers, suggesting that ALK functions to inhibit sleep or promote wake.
Following a period of sleep deprivation, flies, like other animals, show a homeostatic response in the form of increased sleep [30]. We wondered whether this homeostatic regulation was disrupted in Alk mutants. We found that after 6 hours of sleep deprivation by mechanical stimulation, both iso31 and Alkts flies increased sleep the following morning (Fig 3A). However, sleep-deprived Alkts flies fell asleep faster than iso31 controls, suggesting sleep pressure was higher in Alkts flies (Fig 3A and 3D).
Similar to the expression pattern of the mouse ALK gene, the Drosophila Alk gene is extensively expressed in the developing and adult nervous system [12,14,23]. Its adult expression includes the mushroom body, the protocerebral bridge, the antennal lobes, the suboesophageal ganglion, the medial bundle and lateral horns [14]. To locate the sleep regulatory function of Alk in the brain, we carried out a brain mini-screen using a series of GAL4 lines to drive UAS-Alk RNA interference (RNAi) in brain circuits. Inducing Alk RNAi with a pan-neuronal driver, elav-GAL4, reduced Alk mRNA level to ~35% of that of control flies and produced longer sleep, suggesting that ALK functions in neurons (S4 Fig). We then screened over 40 GAL4 drivers with diverse neuronal expression patterns (S1 Table and S5 Fig), including those with expression patterns in the known sleep/wake regulating regions [31–39]. 13 GAL4 lines significantly increased sleep relative to controls, when driving Alk RNAi, while c309-Gal4-driven expression of Alk RNAi decreased sleep (Fig 4A). Of the identified sleep-promoting GAL4 lines, most are broadly expressed, overlapping in several regions of the brain, such as the mushroom body, the ellipsoid body, and the pars intercerebralis. However, inhibiting Alk specifically in the pars intercerebralis (InSITE106, Kurs58 and Dilp2-GAL4), or the ellipsoid body (c819, c232 and c107) did not alter total sleep. The amounts of sleep increase as well as the sleep profiles were different between different GAL4 lines (S6 Fig). Most lines showed increased daytime sleep as well as nighttime sleep. However, lines 1471, c320 and 386Y mainly increased daytime sleep and 7Y mostly increased night-time. 386Y caused a delay in sleep at the beginning of the night while driving an increase at other times. c320 increased sleep immediately after lights-on in the morning. These sleep patterns were consistent across repeated experiments.
We found that the long sleep phenotype of Alk mutants could be rescued by restoring functional ALK with targeted drivers that we identified through the RNAi screen. To circumvent the developmental effects of overexpressing Alk, we expressed Alk in adult Alkts mutants only during a period of restrictive temperature at 29° by combining GAL4s with a temperature-sensitive form of GAL80 (tub-Gal80ts) [40]. We tested 386Y, 7Y, 121Y, 1471, 30Y and MJ63, all of which induce longer sleep when driving Alk RNAi. We found that 386Y and 1471 fully rescued the long sleep phenotype of Alkts at high temperature, while 30y and121y produced a partial rescue (Fig 4B). 7Y and MJ63, however, did not rescue the long sleep of Alkts, and neither did any of the drivers that yielded no phenotype with Alk RNAi (S7 Fig). We infer that while Alk expression is necessary, it may not be sufficient for sleep regulation in regions of 7Y and MJ63 expression. The rescue results further confirm that the long sleep phenotype is specific to the Alk gene and involves specific brain regions.
Prompted by the extensive mushroom body expression of most driver hits in our screen, we focused on the function of ALK in the mushroom bodies (MB). When Alk RNAi was excluded from the MB by combining GAL4s with a mushroom body-specific Gal80 transgene (MB-Gal80)[41](Fig 5), it eliminated the sleep increase induced by 386Y and 30Y, suggesting that Alk function is required in mushroom body neurons labeled by these two drivers. In fact, when 30Y was combined with MB-Gal80, it decreased sleep to below those of controls. c309, a driver that caused short sleep with Alk RNAi, also has extensive mushroom body expression. Driving Alk RNAi with c309/MB-Gal80, however, decreased sleep further compared to Alk RNAi driven by c309. These results indicate that the sleep-increasing effects of Alk deficiency occur mainly in the MBs and with 30Y and c309 they are countered by sleep-inhibiting influences of Alk knockdown outside the mushroom body. We thus tested several drivers that have localized expression in the mushroom body and little expression elsewhere, but did not observe any increase in sleep (Fig 4). Of these three drivers, 17D innervates the core of α/β lobes; D52H has sexually dimorphic expression with strong expression in the α/β and the main γ lobe in males but faint dorsal γ expression in the females, which is the gender we used for sleep assays; R71G10 preferentially innervates the γ lobe and R76D11 has strong expression in both α/β and γ; c305a, which innervates α′/β′ lobes in addition to cells in other brain regions [42,43]. Given that of the positive drivers, H24, NP1131 and 1471 label Kenyon cells that project exclusively in the γ lobes, we hypothesize ALK functions to inhibit sleep in a subset of γ lobe neurons that are not targeted by R71G10 and R76D11.
Genetic interactions between Alk and Nf1 in growth and learning processes led us to investigate whether Nf1 is also required for sleep regulation. Interestingly, a prevalence of sleep disturbances has recently been reported in NF1 patients [44,]. We detected considerable variability in total sleep amount in Nf1 mutant flies across experiments. We tested three Nf1 alleles, Nf1P1, Nf1P2, and Nf1c00617. Nf1P1 and Nf1P2 alleles are both assumed to be null because neither expresses NF1 protein and homozygous flies have similar defects in locomotor activity rhythms, body size and learning [22,46, 47]. Although the average total sleep of male flies harboring any two Nf1 alleles was significantly less than that of control flies, we did not observe a consistent difference between Nf1 mutant and control female flies (Fig 6). While Nf1P2/c00617 female flies slept less than controls, sleep amounts in Nf1P1/P2 female flies were generally not different from those of control flies. Furthermore, Nf1P1 and Nf1P2 female flies did not consistently show sleep reduction as compared to iso31 controls in separate experiments. Similar discrepancies were found when sleep was assayed with the multi-beam sleep monitors. However, both Nf1 male and female flies exhibited nocturnal hyperactivity (Fig 6A and 6C), which resulted in an increase in daytime sleep and a decrease in nighttime sleep. Nf1 mutants also showed consistent defects in sleep consolidation. Both daytime sleep and nighttime sleep were highly fragmented in Nf1 males, such that the average sleep bout duration was reduced and the number of sleep bouts increased (S8 Fig). While reductions in average bout duration did not reach significance in females, increased numbers of bouts suggest deficits in sleep maintenance and consolidation. Rescuing Nf1 with pan-neuronal expression of a UAS-Nf1 transgene increased sleep of Nf1 mutants and in fact caused a long sleep phenotype compared to wild-type controls. This did not result from ectopic expression of the transgene as expressing the same UAS-Nf1 transgene in wild-type flies had no effect (Fig 6B). Interestingly, sleep increase was seen with Nf1 rescue in both male and female flies, suggesting that effects of Nf1 on sleep are quite complex.
We then investigated whether Alk and Nf1 interact to regulate sleep. We chose to test Alk with the Nf1P1/P2 allelic combination because Nf1P1/P2 females have normal amount of sleep and so any sleep suppression in the double mutants would not be confounded by additive effects of short-sleeping Nf1 mutants. To sensitize the assay, we compared sleep in Alkts, Alkts/1 or Nf1P1/P2 single mutants and Alkts;Nf1P1/P2 and Alkts/1;Nf1P1/P2 double mutants at three different temperatures that render different dosages of functional ALK. Interestingly, total amounts of sleep in Alk;Nf1 double mutant flies were less than those of flies deficient for Alk alone, and not different from Nf1 single mutants or iso31 control flies (Fig 7). We found that regardless of the severity of the Alk alleles, Nf1P1/P2 completely suppressed the long sleep phenotype. As another piece of evidence for genetic interaction, we found that Nf1P1/P2 also suppressed the long sleep phenotype caused by Alk pan-neural RNAi (S9 Fig). These results suggest that Nf1 interacts with Alk in a sleep regulating circuit.
Nf1 is part of the circadian output pathway that controls rest: activity rhythms [22]. As Alk was found to interact with Nf1 in sleep regulation, as well as in growth and learning, we asked whether Alk is required also for circadian rhythms and whether Alk and Nf1 interact in circadian pathways. We found that Alkts/1 trans-heterozygotes raised at 18°C maintained locomotor activity rhythms at the restrictive temperature of 29°C in constant darkness (Fig 8 and Table 1), indicating Alk is not required to maintain circadian activity. Indeed, the FFT values, a measure of rhythm strength, of Alkts/1 and Alkts/Def flies were higher at 29°C than at 18°C, suggesting that loss of Alk may actually improve rhythms rather than disrupt them. However, inhibiting Alk failed to rescue the circadian defects in Nf1 flies: Alkts/1; Nf1P1/P2 double mutant were arrhythmic, just like Nf1 single mutants. To exclude a requirement for Alk in the development of circadian circuits, we also tested Alkts homozygous flies raised at 25°C, at which temperature Alkts flies have moderate lethality [19]. The partial reduction in ALK function throughout development did not cause arrhythmia nor did it suppress arrhythmia in Nf1P1/P2 flies (Table 1). These results suggest that Alk does not function in the circadian output circuit regulated by Nf1.
Though a few studies implicate Alk orthologs in regulating behaviors such as decision-making, cognition, associative learning and addiction, most functional studies demonstrate various developmental roles for Alk [14,20,21,48–50]. We acutely induce a long-sleep phenotype by taking advantage of a temperature-sensitive allele, Alkts, revealing that Alk regulates sleep directly rather than through developmental processes. We also show mutations in Nf1, a gene encoding a GAP that regulates the Ras/ERK pathway activated by ALK, causes a sexually dimorphic short-sleep phenotype. Thus we establish a novel in vivo function for both Alk and Nf1 and show they interact with each other to regulate sleep.
Many downstream signaling pathways have been proposed for ALK, among them Ras/ERK, JAK/STAT, PI3K and PLCγ signaling [11]. ERK activation through another tyrosine receptor kinase Epidermal growth factor receptor (EGFR) has been linked to increased sleep [36,51], while here we show that Alk, a positive regulator of ERK, inhibits sleep. We note that ERK is a common signaling pathway targeted by many factors, and may have circuit- specific effects, with different effects on sleep in different brain regions. Indeed, neural populations that mediate effects of ERK on sleep have not been identified. The dose of ALK required for ERK activation might also differ in different circuits. Region-specific effects of Alk are supported by our GAL4 screen, in which down-regulation of Alk in some brain regions even decreased sleep. The overall effect, however, is to increase sleep, evident from the pan-neuronal knockdown. We found that the mushroom body, a site previously implicated in sleep regulation and learning, requires Alk to inhibit sleep. Interestingly, the expression patterns of Alk and Nf1 overlap extensively in the mushroom body [14], suggesting that they may interact here to regulate both sleep and learning. However, it was previously shown that Alk activation in the mushroom body has no effect on learning [14]. The mushroom body expression in that study was defined with MB247 and c772, both of which also had no effects on sleep when driving Alk RNAi (Fig 4). The spatial requirement for Nf1 in the context of learning has been disputed in previous studies with results both for and against a function in the mushroom body [14,52]. The discrepancies between these studies could result from: 1) varied expression of different drivers within lobes of the mushroom body, with some not even specific to the mushroom body; 2) variability in the effectiveness and specificity of MB-Gal80 in combination with different GAL4s. We confirmed that our MB-Gal80 manipulation eliminated all mushroom body expression and preserved most if not all other cells with 30Y, 386Y and c309. Future work will further define the cell populations in which Alk and Nf1 interact to affect sleep.
We observed a substantial sleep decrease in Nf1 male flies compared to control flies. However, sleep phenotypes in Nf1 female flies are inconsistent. It is unlikely that unknown mutations on the X chromosome cause the short-sleeping phenotype because our 7 generation outcrosses into the control iso31 background started with swapping X chromosomes in Nf1P1 and Nf1P2 male flies with those of iso31 flies. In support of a function in sleep regulation, restoring Nf1 expression in neurons of Nf1 mutants reverses the short sleep phenotype to long sleep in both males and females. This does not result from ectopic expression of the transgene as expressing the same UAS-Nf1 transgene in wild-type flies has no effect. We hypothesize that Nf1 promotes sleep in some brain regions and inhibits it in others, and sub-threshold levels of Nf1, driven by the transgene in the mutant background, tilt the balance towards more sleep. As reported here, Alk also has differential effects on sleep in different brain regions, as does protein kinase A [33], thus such effects are not unprecedented. We also note severe sleep fragmentation in Nf1 mutants, which suggests that they have trouble maintaining sleep.
The sex-specific phenotypes of Nf1 mutants may reflect sexually dimorphic regulation of sleep. A recently published genome-wide association study of sleep in Drosophila reported that an overwhelming majority of single nucleotide polymorphisms (SNPs) exhibit some degree of sexual dimorphism: the effects of ~80% SNPs on sleep are not equal in the two sexes [53]. Interestingly, sex was found to be a major determinant of neuronal dysfunction in human NF1 patients and Nf1 knock-out mice, resulting in differential vision loss and learning deficits [54]. The sex-dimorphic sleep phenotype in Nf1 flies provides another model to study sex-dimorphic circuits involving Nf1. Interestingly, a prevalence of sleep disturbances have recently been reported in NF1 patients [44,45], suggesting that NF1 possibly play a conserved function in sleep regulation.
An attractive hypothesis for a function of sleep is that plastic processes during wake lead to a net increase in synaptic strength and sleep is necessary for synaptic renormalization [3]. There is structural evidence in Drosophila to support this synaptic homeostasis hypothesis (SHY): synapse size and number increase during wake and after sleep deprivation, and decrease after sleep [55]. However, little is known about the molecular mechanisms by which waking experience induces changes in plasticity and sleep. FMRP, the protein encoded by the Drosophila homolog of human fragile X mental retardation gene FMR1, mediates some of the effects of sleep/wake on synapses [55,56]. Loss of Fmr1 is associated with synaptic overgrowth and strengthened neurotransmission and long sleep. Overexpressing Fmr1 results in dendritic and axonal underbranching and short sleep. More importantly, overexpression of Fmr1 in specific circuits eliminates the wake-induced increases in synapse number and branching in these circuits. Thus, up-regulation of FMR accomplishes a function normally associated with sleep.
We hypothesize that Alk and Nf1 similarly play roles in synaptic homeostasis. They are attractive candidates for bridging sleep and plastic processes, because: 1) Alk is expressed extensively in the developing and adult CNS synapses [14,57]. In particular, both Alk and Nf1 are strongly expressed in the mushroom body, a major site of plasticity in the fly brain. 2) Functionally, postsynaptic hyperactivation of Alk negatively regulates NMJ size and elaboration [19]. In contrast, Nf1 is required presynaptically at the NMJ to suppress synapse branching [58]. 3) Alk and Nf1 affect learning in adults and they functionally interact with each other in this process [14]. It is tempting to speculate that in Alk mutants, sleep is increased to prune the excess synaptic growth predicted to occur in these mutants. Such a role for sleep is consistent with the SHY hypothesis. The SHY model would predict that Alk flies have higher sleep need, which is expected to enhance rebound after sleep deprivation. While our data show equivalent quantity of rebound in Alk mutants, we found that they fall asleep faster than control flies the morning after sleep deprivation (Fig 3), suggesting that they have higher sleep drive. Increased sleep need following deprivation could also be reflected in greater cognitive decline, but this has not yet been tested for Alk mutants. We note that Nf1 mutants have reduced sleep although their NMJ phenotypes also consist of overbranched synapses [58,59]. We postulate that their sleep need is not met and thus results in learning deficits. Clearly, more work is needed to test these hypotheses concerning the roles of Alk and Nf1 in sleep, learning, and memory circuits.
The following lines were used previously in the lab [33, 37, 60]: Elav-Gal4, 201Y-GAL4, c739-GAL4, 238Y-GAL4, c309-GAL4, Kurs58-GAL4, Pdf-Gal4, H24-GAL4, c507-GAL4, 30Y-GAL4, 50Y-GAL4, MJ63-GAL4, c232-GAL4, 104Y-GAL4, 17D-GAL4, Dilp2-Gal4, Tdc2-Gal4, TH-Gal4, Mai301-GAL4, c767-GAL4, 1471-GAL4, ok107-GAL4, c929-GAL4, 53b-GAL4, c320-GAL4 and UAS-GFP.NLS. elav-GAL4; Dcr2 (25750), c305a-GAL4 (30829), c107-GAL4(30823), c819-GAL4 (30849), 121Y-GAL4 (30815), 7Y-GAL4 (30812), 36Y-GAL4 (30819), 386Y-GAL4 (25410), c584-GAL4 (30842), DopR-GAL4 (19491), Ddc-GAL4 (7009), R71G10-GAL4 (39604) and R76D11-GAL4 (39927) were ordered from the Bloomington Drosophila Stock Center. NP1131-GAL4 (103898), NP1004-GAL4 (112440) and NP2024-GAL4 (112749) were ordered from the Drosophila Genetic Resource Center. Tub-GAL80ts [40], MB-Gal80 [41], D52H-GAL4 [42], c687-GAL4 [61], 6B-GAL4 [62], InSITE106-GAL4 [63],Cha-GAL4 [64], Kurs45-GAL4 [65] were gifts. Nf1P1 and Nf1P2 alleles were reported [22] and were outcrossed into an iso31 background for 7 generations. Alkts/CyO was a gift from Dr. J Weiss [13] and was outcrossed into an iso31 background. Alk1/CyO and UAS-Alk/CyO were gifts from Dr. R Palmer and both were outcrossed into an iso31 background [13,16]. The deficiency line uncovering the Alk gene, AlkDef (7888)/CyO, was ordered from Bloomington. Alk RNAi (11446) and UAS-Dcr2(60008) were ordered from the Vienna Drosophila Resource Center.
Reporter GFP expression driven by GAL4 lines was visualized through whole-mount brain immunofluorescence as previously described [60]. Rabbit anti-GFP (Molecular Probes A-11122) 1:1000 and Alex Fluor 488 Goat anti-rabbit (Molecular Probes A-11008) 1:500 were used.
Sleep was monitored as described previously [37]. Flies were raised and kept on a 12h:12h light/dark (LD) cycle at 18°C or 25°C as stated. 3–7 day old flies were loaded into glass tubes containing 5% sucrose and 2% agar. Locomotor activity was monitored with the Drosophila Activity Monitoring System (Trikinetics, Waltham MA), or when indicated with multi-beam monitors (Trikinetics, Waltham MA) that generate 17 infrared beams. Data were analyzed with Pysolo software [5]. For sleep assays with temperature shifts, total sleep amount was averaged for 3 d at the lower temperature before the shift and 3 d at the high temperature. For sleep deprivation experiments, flies were monitored for a baseline day and then sleep deprived on the second day for 6 hours from Zeitgeber Time (ZT) 18 to ZT24 during the night. Sleep was continually monitored for 2 recovery days. Mechanical sleep deprivation was accomplished using a Trikinetics vortexer mounting plate, with shaking of monitors for 2 seconds randomly within every 20 second window for 6 hours. The arousal threshold assay was described previously [66]. A 12oz rubber weight was dropped from 2-inch height onto a rack supporting large DAMS monitors at ZT20. Flies with no activity 5 min before a stimulus and exhibited beam crossings within 5 min after the light pulse were recorded as “aroused”.
Individual male flies were loaded into glass tubes containing 5% sucrose and 2% agar. Locomotor activity was monitored with the Drosophila Activity Monitoring System (Trikinetics, Waltham MA), and analyzed with Clocklab software (Actimetrics, Wilmette). To evaluate ALK’s role in maintaining adult rhythms, all genotypes were raised at 18°C to avoid inactivating ALK during development. 3–7 d old flies were then loaded into glass tubes and entrained for 3 d to a 12h:12h LD cycle, followed by 4 days in constant darkness at 18°C and then 4 days in constant darkness at 29°C. Rhythmicity analysis was performed for each 4 d period. In a separate experiment, iso31, Nf1P1/P2, Alkts and Alkts;Nf1P1/P2 flies were raised and tested at 25°C to evaluate whether inhibiting Alk during development affect rest-activity rhythm. A fly was considered rhythmic if it met 2 criteria: 1) displayed a rhythm with 95% confidence using χ2 periodogram analysis, and 2) a corresponding FFT value above 0.01 for the determined period length.
The negative geotaxis assay was adapted from (Barone MC and Bohmann D 2013). Please see supplemental methods for details.
Data were analyzed and plotted with SigmaPlot and GraphPrism software. One-way ANOVA analysis was done to reveal differences between genotypes in the same experiments and pairwise comparison between genotypes were done with post-hoc analysis as indicated in the figures. ns, not significant; **, p<0.01; ***, p<0.001; ****, p<0.0001.
|
10.1371/journal.pcbi.1007128 | Open collaborative writing with Manubot | Open, collaborative research is a powerful paradigm that can immensely strengthen the scientific process by integrating broad and diverse expertise. However, traditional research and multi-author writing processes break down at scale. We present new software named Manubot, available at https://manubot.org, to address the challenges of open scholarly writing. Manubot adopts the contribution workflow used by many large-scale open source software projects to enable collaborative authoring of scholarly manuscripts. With Manubot, manuscripts are written in Markdown and stored in a Git repository to precisely track changes over time. By hosting manuscript repositories publicly, such as on GitHub, multiple authors can simultaneously propose and review changes. A cloud service automatically evaluates proposed changes to catch errors. Publication with Manubot is continuous: When a manuscript’s source changes, the rendered outputs are rebuilt and republished to a web page. Manubot automates bibliographic tasks by implementing citation by identifier, where users cite persistent identifiers (e.g. DOIs, PubMed IDs, ISBNs, URLs), whose metadata is then retrieved and converted to a user-specified style. Manubot modernizes publishing to align with the ideals of open science by making it transparent, reproducible, immediate, versioned, collaborative, and free of charge.
| Traditionally, scholarly manuscripts have been written in private by a predefined team of collaborators. But now the internet enables realtime open science, where project communication occurs online in a public venue and anyone is able to contribute. Dispersed teams of online contributors require new tools to jointly prepare manuscripts. Existing tools fail to scale beyond tens of authors and struggle to support iterative refinement of proposed changes. Therefore, we created a system called Manubot for writing manuscripts based on collaborative version control. Manubot adopts the workflow from open source software development, which has enabled hundreds of contributors to simultaneously develop complex codebases such as Python and Linux, and applies it to open collaborative writing. Manubot also addresses other shortcomings of current publishing tools. Specifically, all changes to a manuscript are tracked, enabling transparency and better attribution of credit. Manubot automates many tasks, including creating the bibliography and deploying the manuscript as a webpage. Manubot webpages preserve old versions and provide a simple yet interactive interface for reading. As such, Manubot is a suitable foundation for next-generation preprints. Manuscript readers have ample opportunity to not only provide public peer review but also to contribute improvements, before and after journal publication.
| The internet enables science to be shared in real-time at a low cost to a global audience. This development has decreased the barriers to making science open, while supporting new massively collaborative models of research [1]. However, the scientific community requires tools whose workflows encourage openness [2]. Manuscripts are the cornerstone of scholarly communication, but drafting and publishing manuscripts has traditionally relied on proprietary or offline tools that do not support open scholarly writing, in which anyone is able to contribute and the contribution history is preserved and public. We introduce Manubot, a new tool and infrastructure for authoring scholarly manuscripts in the open, and report how it was instrumental for the collaborative project that led to its creation.
Based on our experience leading a recent open review [3], we discuss the advantages and challenges of open collaborative writing, a form of crowdsourcing [4]. Our review manuscript [5] was code-named the Deep Review and surveyed deep learning’s role in biology and precision medicine, a research area undergoing explosive growth. We initiated the Deep Review in August 2016 by creating a GitHub repository (https://github.com/greenelab/deep-review) to coordinate and manage contributions. GitHub is a platform designed for collaborative software development that is adaptable for collaborative writing. From the start, we made the GitHub repository public under a Creative Commons Attribution License (CC BY 4.0 at https://github.com/greenelab/deep-review/blob/master/LICENSE.md). We encouraged anyone interested to contribute by proposing changes or additions. Although we invited some specific experts to participate, most authors discovered the manuscript organically through conferences or social media, deciding to contribute without solicitation. In total, the Deep Review attracted 36 authors, who were not determined in advance, from 20 different institutions in less than two years.
The Deep Review and other studies that subsequently adopted the Manubot platform were unequivocal successes bolstered by the collaborative approach. However, inviting wide authorship brought many technical and social challenges such as how to fairly distribute credit, coordinate the scientific content, and collaboratively manage extensive reference lists. The manuscript writing process we developed using the Markdown language, the GitHub platform, and our new Manubot tool for automating manuscript generation addresses these challenges.
Manubot supports citations by adding a persistent identifier like a Digital Object Identifier (DOI) or PubMed Identifier (PMID) directly in the text so that large groups of authors do not have to coordinate reference lists. When text is changed, Manubot automatically updates the manuscript’s web page so that all authors can read and edit from the latest version. Because manuscripts are created from GitHub repositories, Manubot supports a workflow where all edits are reviewed and discussed, ensuring that the collaborative text has a cohesive style and message and that authors receive precise credit for their work. These and other features support an open collaborative writing process that is not feasible with other writing platforms.
There are many existing collaborative writing platforms (Table 1) [6]. In general, platforms with “what you see is what you get” (WYSIWYG) editors, such as Microsoft Word or Google Docs, require the least technical expertise to use. On the flip side, WYSIWYG platforms can be difficult to customize and incorporate into automated computational workflows. Traditionally, LaTeX has been used for these needs, since documents are written in plain text and the system is open source and extensible. Rendering LaTeX documents requires specialized software, but webapps like Overleaf now enable collaborative authoring of LaTeX documents. Nonetheless, LaTeX-based systems are limited in that PDF (or similar) is the only fully supported output format. Alternatively, Authorea is a collaborative writing webapp whose primary output format is HTML. Authorea allows authors to write in Markdown, a limited subset of LaTeX, or their WYSIWYG HTML editor.
Existing platforms work well for editing text and are widely used for scholarly writing. However, they often lack features that are important for open collaborative writing, such as versatile version control and multiple permission levels. For example, Manubot is the only platform listed in Table 1 that offers the ability to address thematically related changes together and enables multiple authors to iteratively refine proposed changes.
Manubot’s collaborative writing workflow adopts standard software development strategies that enable any contributor to edit any part of the manuscript but enforce discussion and review of all proposed changes. The GitHub platform supports organizing and editing the manuscript. Manubot projects use GitHub issues for organization, opening a new issue for each discussion topic. For example, in a review manuscript like the Deep Review, this includes each primary paper under consideration. Within a paper’s issue, contributors summarize the research, discuss it (sometimes with participation from the original authors), and assess its relevance to the review. In a primary research article, issues can instead track progress on specific figures or subsections of text being drafted. Issues serve as an open to-do list and a forum for debating the main messages of the manuscript.
GitHub and the underlying Git version control system [7,8] also structure the writing process. The official version of the manuscript is forked by individual contributors, creating a copy they can freely modify. A contributor then adds and revises files, grouping these changes into commits. When the changes are ready to be reviewed, the series of commits are submitted as a pull request through GitHub, which notifies other authors of the pending changes. GitHub’s review interface allows anyone to comment on the changes, globally or at specific lines, asking questions or requesting modifications [9]. Conversations during review can reference other pull requests, issues, or authors, linking the relevant people and content (Fig 1). Reviewing batches of revisions that focus on a single theme is more efficient than independently discussing isolated comments and edits and helps maintain consistent content and tone across different authors and reviewers. Once all requested modifications are made, the manuscript maintainers, a subset of authors with elevated GitHub permissions, formally approve the pull request and merge the changes into the official version. The process of writing and revising material can be orchestrated through GitHub with a web browser (as shown in S1 Video) or through a local text editor.
The Deep Review issue (https://github.com/greenelab/deep-review/issues/575) and pull request (https://github.com/greenelab/deep-review/pull/638) on protein-protein interactions demonstrate this process in practice. A new contributor identified a relevant research topic that was missing from the review manuscript with examples of how the literature would be summarized, critiqued, and integrated into the review. A maintainer confirmed that this was a desirable topic and referred to related open issues. The contributor made the pull request, and two maintainers and another participant made recommendations. After four rounds of reviews and pull request edits, a maintainer merged the changes.
We found that this workflow was an effective compromise between fully unrestricted editing and a more heavily-structured approach that limited the authors or the sections they could edit. In addition, authors are associated with their commits, which makes it easy for contributors to receive credit for their work. Fig 2 and the GitHub contributors page (https://github.com/greenelab/deep-review/graphs/contributors) summarize all edits and commits from each author, providing aggregated information that is not available on most other collaborative writing platforms. Because the Manubot writing process tracks the complete history through Git commits, it enables detailed retrospective contribution analysis. These pull request and contribution tracking examples both come from Deep Review, the largest Manubot project to date, but illustrate the general principles of transparency and collaboration that are shared by all open Manubot manuscripts.
GitHub issues can also be used for formal peer review by independent or journal-selected reviewers. A reviewer conducting open peer review can create issues using their own GitHub account, as one reviewer did for this manuscript (https://github.com/greenelab/meta-review/issues/124). Alternatively, a reviewer can post feedback with a pseudonymous GitHub account or have a trusted third party such as a journal editor post their comments anonymously. Authors can elect to respond to reviews in the GitHub issues or a public response letter (https://github.com/greenelab/meta-review/blob/v3.0/content/response-to-reviewers.md), creating open peer review.
Although we developed Manubot with collaborative writing in mind, it can also be helpful for individuals preparing scholarly documents. Authors may choose to make their changes directly to the master branch, forgoing pull requests and reviews. This workflow retains many of Manubot’s benefits, such as transparent history, automation, and allowing outside contributors to propose changes. In cases where outside contributions are unwanted, authors can disable pull requests on GitHub. It is also possible to use Manubot on a private GitHub repository. Private manuscripts require some additional customization to disable GitHub Pages and may require a paid continuous integration plan. See the existing manuscripts for examples of the range of contribution workflows and Manubot use cases.
Manubot is a system for writing scholarly manuscripts via GitHub. For each manuscript, there is a corresponding Git repository. The master branch of the repository contains all of the necessary inputs to build the manuscript. Specifically, a content directory contains one or more Markdown files that define the body of the manuscript as well as a metadata file to set information such as the title, authors, keywords, and language. Figures can be hosted in the content/images subdirectory or elsewhere and specified by URL. Repositories contain scripts and other files that define how to build and deploy the manuscript. Many of these operations are delegated to the manubot Python package or other dependencies such as Pandoc, which converts between document formats, and Travis CI, which builds the manuscript in the cloud. Manubot pieces together many existing standards and technologies to encapsulate a manuscript in a repository and automatically generate outputs.
Manubot does not impose any restrictions on authorship. It allows authors to adhere to the author inclusion and ordering conventions of their field, which vary considerably across disciplines [53]. Some Manubot projects create a table in their GitHub repository to track contributors who did not commit text to the manuscript (https://github.com/Benjamin-Lee/deep-rules/blob/cfb7f744573ca0532a19ca1a8e9473a555cf8eb2/contributors.md). This provides a transparent way to record contributions such as experimental research that generated data for the manuscript and discuss whether they meet that project’s authorship criteria. Contribution transparency helps prevent ghostwriting [54] and is especially important in collaborative writing [55]. Although we recommend authors provide their ORCID and GitHub username, Manubot also supports pseudonyms, pseudonymous GitHub usernames, and authors without an ORCID or GitHub account.
To determine authorship for the Deep Review, we followed the International Committee of Medical Journal Editors (ICMJE) guidelines and used GitHub to track contributions. ICMJE recommends authors substantially contribute to, draft, approve, and agree to be accountable for the manuscript. We acknowledged other contributors who did not meet all four criteria, including contributors who provided text but did not review and approve the complete manuscript. Although these criteria provided a straightforward, equitable way to determine who would be an author, they did not produce a traditionally ordered author list. In biomedical journals, the convention is that the first and last authors made the most substantial contributions to the manuscript. This convention can be difficult to reconcile in a collaborative effort. Using Git, we could quantify the number of commits each author made or the number of sentences an author wrote or edited, but these metrics discount intellectual contributions such as discussing primary literature and reviewing pull requests. Therefore, we concluded that it is not possible to construct an objective system to compare and weight the different types of contributions and produce an ordered author list [56].
To address this issue, we generalized the concept of “co-first” authorship, in which two or more authors are denoted as making equal contributions to a paper. We defined four types of contributions [5], from major to minor, and reviewed the GitHub discussions and commits to assign authors to these categories. A randomized algorithm then arbitrarily ordered authors within each contribution category, and we combined the category-specific author lists to produce a traditional ordering. The randomization procedure was shared with the authors in advance (pre-registered) and run in a deterministic manner. Given the same author contributions, it always produced the same ordered author list. We annotated the author list to indicate that author order was partly randomized and emphasize that the order did not indicate one author contributed more than another from the same category. The Deep Review author ordering procedure illustrates authorship possibilities when all contributions are publicly tracked and recorded that would be difficult with a traditional collaborative writing platform.
Papers with hundreds or thousands of authors are on the rise, such as the article describing the experiments and data that led to the discovery of the Higgs Boson [57] (5000 authors) and the report of the Drosophila genome [58] (1000 authors). Yet the number of people that participated in writing those papers, as opposed to generating and analyzing the data, is not always clear and is likely to be far below the number of authors [59,60]. Manubot provides the scientists involved in large collaborations the opportunity to actively participate, through a public forum, in the writing process.
The open scholarly writing Manubot enables has particular benefits for review articles, which present the state of the art in a scientific field [61]. Literature reviews are typically written in private by an invited team of colleagues. In contrast, broadly opening the process to anyone engaged in the topic—such that planning, organizing, writing, and editing occur collaboratively in a public forum where anyone is welcome to participate—can maximize a review’s value. Open drafting of reviews is especially helpful for capturing state-of-the-art knowledge about rapidly advancing research topics at the intersection of existing disciplines where contributors bring diverse opinions and expertise.
Writing review articles in a public forum allows review authors to engage with the original researchers to clarify their methods and results and present them accurately, as exemplified at https://github.com/greenelab/deep-review/issues/213. Additionally, discussing manuscripts in the open generates valuable pre-publication peer review of preprints [22] or post-publication peer review [16,62,63]. Because incentives to provide public peer review of existing literature [64] are lacking, open collaborative reviews—where authorship is open to anyone who makes a valid contribution—could help spur more post-publication peer review.
The Deep Review was not the first scholarly manuscript written online via an open collaborative process. In 2013, two dozen mathematicians created the 600-page Homotopy Type Theory book, writing collaboratively in LaTeX on GitHub [65,66]. Two technical books on cryptocurrency—Mastering Bitcoin (https://github.com/bitcoinbook/bitcoinbook) and Mastering Ethereum (https://github.com/ethereumbook/ethereumbook)—written on GitHub in AsciiDoc format have engaged hundreds of contributors. Both Homotopy Type Theory and Mastering Bitcoin continue to be maintained years after their initial publication. A 2017 perspective on the future of peer review was written collaboratively on Overleaf, with contributions from 32 authors [67]. While debate was raging over tightening the default threshold for statistical significance, nearly 150 scientists contributed to a Google Doc discussion that was condensed into a traditional journal commentary [68,69]. The greatest success to date of open collaborative writing is arguably Wikipedia, whose English version contains over 5.5 million articles. Wikipedia scaled encyclopedias far beyond any privately-written alternative. These examples illustrate how open collaborative writing can scale scholarly manuscripts where diverse opinion and expertise are paramount beyond what would otherwise be possible.
Open writing also presents new opportunities for distributing scholarly communication. Though it is still valuable to have versioned drafts of a manuscript with digital identifiers, journal publication may not be the terminal endpoint for collaborative manuscripts. After releasing the first version of the Deep Review [10], 14 new contributors updated the manuscript (Fig 2). Existing authors continue to discuss new literature, creating a living document. Manubot provides an ideal platform for perpetual reviews [70,71].
Concepts for the future of scholarly publishing extend beyond collaborative writing [72–74]. Pandoc Scholar [12] and Bookdown [75], which has been used for open writing [76], both enhance traditional Markdown to better support publishing. The knitcitations (https://github.com/cboettig/knitcitations) package enables citation by DOI or URL in R Markdown documents. Examples of continuous integration to automate manuscript generation include gh-publisher (https://github.com/ewanmellor/gh-publisher) and jekyll-travis (https://github.com/mfenner/jekyll-travis), which was used to produce a continuously published webpage (http://book.openingscience.org/) for the Opening Science book [77,78]. Binder [11], Distill journal articles [79], Idyll [80], and Stencila [81,82] support manuscripts with interactive graphics and close integration with the underlying code. As an open source project, Manubot can be extended to adopt best practices from these other emerging platforms.
Several other open science efforts are GitHub-based like our collaborative writing process. ReScience [83] as well as titles from Open Journals, such as the Journal of Open Source Software [52], rely on GitHub for peer review and hosting. Distill uses GitHub for transparent peer review and post-publication peer review [84]. GitHub is increasingly used for resource curation [85], and collaborative scholarly reviews combine literature curation with discussion and interpretation.
There are potential limitations of our GitHub-based approach. Because the Deep Review pertained to a computational topic, most of the authors had computational backgrounds, including previous experience with version control workflows and GitHub. In other disciplines, collaborative writing via GitHub and Manubot could present a steeper barrier to entry and deter participants. In addition, Git carefully tracks all revisions to the manuscript text but not the surrounding conversations that take place through GitHub issues and pull requests. These discussions must be archived to ensure that important decisions about the manuscript are preserved and authors receive credit for intellectual contributions that are not directly reflected in the manuscript’s text. GitHub supports programmatic access to issues, pull requests, and reviews so tracking these conversations is feasible in the future.
In the Deep Review, we established contributor guidelines (https://github.com/greenelab/deep-review/blob/v1.0/CONTRIBUTING.md) that discussed norms in the areas of text contribution, peer review, and authorship, which we identified in advance as potential areas of disagreement. Our contributor guidelines required verifiable participation for authorship: either directly attributable changes to the text or participation in the discussion on GitHub. These guidelines did not discuss broader community norms that may have improved inclusiveness. It is also important to consider how the move to an open contribution model affects under-represented minority members of the scientific community [19]. Recent work has identified clear social norms and processes as helpful to maintaining a collaborative culture [86]. Conferences and open source projects have used codes of conduct to establish these norms (e.g. Contributor Covenant at https://www.contributor-covenant.org/) [87]. We would encourage the maintainers of similar projects to consider broader codes of conduct for project participants that build on social as well as academic norms.
Science is undergoing a transition towards openness. The internet provides a global information commons, where scholarship can be publicly shared at a minimal cost. For example, open access publishing provides an economic model that encourages maximal dissemination and reuse of scholarly articles [18,88,89]. More broadly, open licensing solves legal barriers to content reuse, enabling any type of scholarly output to become part of the commons [90,91]. The opportunity to reuse data and code for new investigations, as well as a push for increased reproducibility, has begot a movement to make all research outputs public, unless there are bona fide privacy or security concerns [92–94]. New tools and services make it increasingly feasible to publicly share the unabridged methods of a study, especially for computational research, which consists solely of software and data.
Greater openness in both research methods and publishing creates an opportunity to redefine peer review and the role journals play in communicating science [67]. At the extreme is real-time open science, whereby studies are performed entirely in the open from their inception [95]. Many such research projects have now been completed, benefiting from the associated early-stage peer review, additional opportunity for online collaboration, and increased visibility [50,96].
Manubot is an ideal authoring protocol for real-time open science, especially for projects that are already using an open source software workflow to manage their research. While Manubot does require technical expertise, the benefits are manyfold. Specifically, Manubot demonstrates a system for publishing that is transparent, reproducible, immediate, permissionless, versioned, automated, collaborative, open, linked, provenanced, decentralized, hackable, interactive, annotated, and free of charge. These attributes empower integrating Manubot with an ecosystem of other community-driven tools to make science as open and collaborative as possible.
|
10.1371/journal.pmed.1002847 | Validation of the prognostic value of NF-κB p65 in prostate cancer: A retrospective study using a large multi-institutional cohort of the Canadian Prostate Cancer Biomarker Network | The identification of patients with high-risk prostate cancer (PC) is a major challenge for clinicians, and the improvement of current prognostic parameters is an unmet clinical need. We and others have identified an association between the nuclear localization of NF-κB p65 and biochemical recurrence (BCR) in PC in small and/or single-centre cohorts of patients.
In this study, we accessed 2 different multi-centre tissue microarrays (TMAs) representing cohorts of patients (Test-TMA and Validation-TMA series) of the Canadian Prostate Cancer Biomarker Network (CPCBN) to validate the association between p65 nuclear frequency and PC outcomes. Immunohistochemical staining of p65 was performed on the Test-TMA and Validation-TMA series, which include PC tissues from patients treated by first-line radical prostatectomy (n = 250 and n = 1,262, respectively). Two independent observers evaluated the p65 nuclear frequency in digital images of cancer tissue and benign adjacent gland tissue. Kaplan–Meier curves coupled with a log-rank test and univariate and multivariate Cox regression models were used for statistical analyses of continuous values and dichotomized data (cutoff of 3%). Multivariate analysis of the Validation-TMA cohort showed that p65 nuclear frequency in cancer cells was an independent predictor of BCR using continuous (hazard ratio [HR] 1.02 [95% CI 1.00–1.03], p = 0.004) and dichotomized data (HR 1.33 [95% CI 1.09–1.62], p = 0.005). Using a cutoff of 3%, we found that this biomarker was also associated with the development of bone metastases (HR 1.82 [95% CI 1.05–3.16], p = 0.033) and PC-specific mortality (HR 2.63 [95% CI 1.30–5.31], p = 0.004), independent of clinical parameters. BCR-free survival, bone-metastasis-free survival, and PC-specific survival were shorter for patients with higher p65 nuclear frequency (p < 0.005). As the small cores on TMAs are a limitation of the study, a backward validation of whole PC tissue section will be necessary for the implementation of p65 nuclear frequency as a PC biomarker in the clinical workflow.
We report the first study using the pan-Canadian multi-centre cohorts of CPCBN and validate the association between increased frequency of nuclear p65 frequency and a risk of disease progression.
| In Canada and the US, prostate cancer is the most commonly diagnosed cancer in men.
Identifying patients with aggressive prostate cancer is crucial for the choice of treatment to increase survival.
Previously, we and others showed that the localization of a biomarker, nuclear factor kappa B (NF-κB) p65, in the nucleus of prostate cancer cells allows the identification of patients with aggressive prostate cancer. More specifically, the nuclear expression of NF-κB p65 is associated with cancer recurrence.
To further validate our findings in a larger, multi-centre cohort of patients, we took advantage of the Canadian Prostate Cancer Biomarker Network (CPCBN) tissue microarrays that include a test (n = 250) and a validation (n = 1,262) series built with radical prostatectomy specimens from prostate cancer patients.
We showed that NF-κB p65 in the nucleus of prostate cancer cells was associated with all evaluated prostate cancer endpoints (biochemical recurrence, development of bone metastases, and prostate-cancer-specific death).
We validated the association between NF-κB p65 nuclear frequency and more aggressive prostate cancer.
Nuclear frequency of NF-κB p65 should help to better identify patients with a higher risk of disease progression, and this could impact patient management.
For the implementation of this biomarker in the clinical workflow, the investigation of nuclear frequency of NF-κB p65 in whole diagnostic samples will be important.
| Prostate cancer (PC) is the most commonly diagnosed cancer in Canadian men [1]. In men with high-risk PC, progression of the disease will lead to biochemical recurrence (BCR), distant metastases, and disease-specific mortality. Up to now, patient prognosis has been based on 3 parameters: preoperative prostate-specific antigen (PSA) level, stage, and Gleason score [2]. However, these are not always sufficient for accurate stratification of patients. The identification of high-risk PC patients is a major challenge for clinicians, and failure to correctly identify these cases leads to disease progression that does not receive the most appropriate management. To more accurately predict PC prognosis, rigorous validation and clinical implementation of new prognostic markers are required [3].
The extensively studied nuclear factor kappa B (NF-κB) pathway is involved in the regulation of inflammation and the immune response [4], and more recently demonstrated its importance in cancer development [5]. Homo- or heterodimers of 5 subunits (p65, c-rel, RelB, p50, and p52) are implicated in the NF-κB signalling pathway to induce the expression of target genes. In the canonical pathway, the inactive form of the p65/p50 dimer is associated with the inhibitor IκB and is retained in the cytoplasm. Once phosphorylated, IκB releases the p65/p50 dimer and is degraded by the proteasome while the dimer translocates to the nucleus, allowing the transcription of target genes. The activation of the NF-κB signalling pathways can result in the progression of several types of cancer, including PC [6].
Previously, we identified the nuclear localization of p65 as a prognostic biomarker in PC patients [7]. We first showed with immunohistochemical staining that nuclear expression of p65 in positive surgical margins of tissue was associated with BCR in a small cohort of 42 patients [8]. This association was also observed in PC tissues using tissue microarrays (TMAs) comprised of specimens from 63 patients [9]. Subsequently, we used a large independent cohort of 1,826 PC patients to validate the relation between nuclear translocation of p65 and increased risk for BCR [10]. More recently, a study in a cohort of 200 PC patients confirmed the frequency of nuclear p65 expression as a prognostic indicator of BCR risk with an immunofluorescence-based approach [11]. These observations were also reported independently by other groups [12–14].
Despite their identification of p65 nuclear expression as a strong predictor of BCR, these studies evaluated p65 in small or single-centre cohorts. To represent the entire population and avoid site-specific biases, multi-institutional cohorts are needed. Overall, several studies have been conducted on promising prognostic biomarkers, but none have been added to current clinical parameters for PC [15]. One of the mandates of the Canadian Prostate Cancer Biomarker Network (CPCBN) is to validate such prognostic biomarkers and accelerate their integration into the clinical workflow to improve PC patient management. For this purpose, the TMA platform of CPCBN includes 2 independent multi-institutional cohorts of PC patients treated by first-line radical prostatectomy (RP) with complete clinical data [16]. The Test-TMA cohort, comprising 250 PC patients, evaluates the prognostic ability of a biomarker. Following conclusive results, the biomarker is next analyzed on specimens from the 1,262 PC patients of the Validation-TMA to identify high-risk PC patients.
In the present study, we validated the nuclear expression of p65 as an independent predictor of BCR in these 2 independent multi-institutional cohorts of PC patients. This is to our knowledge the first study to show the prognostic value of p65 nuclear frequency for the development of bone metastases and PC-specific death.
The TMAs of the CPCBN are composed of RP specimens from 2 independent cohorts of 250 and 1,262 PC patients who agreed to participate in the biobank of 1 of 5 Canadian institutions (Centre de recherche du Centre hospitalier de l’Université de Montréal, Research Institute of the McGill University Health Centre, Centre de recherche du Centre hospitalier universitaire de Québec–Université Laval, University Health Network, and Vancouver Prostate Centre). All patients signed an informed consent for the use of their prostate tissue samples in research. Each biobank received the approval from their local ethics review board for inclusion of prostate tissue samples (n = 300 per site) in the CPCBN resource. Treatment-naïve RP specimens were collected from 1990 to 2011. The first cohort of 250 specimens (n = 50 per site) was named the ‘Test-TMA series’, and the second cohort of 1,262 samples (n ≈ 250 per site) was labelled as the ‘Validation-TMA series’ [16]. For each patient, 0.6-mm cores (3–4 of tumor tissue and 1–2 of benign adjacent tissue) were punctured from formalin-fixed paraffin-embedded blocks and transferred to receiver blocks.
Four-micrometre-thick sections of each TMA block were subjected to antigen retrieval in Cell Conditioning 1 (#950–124, Ventana Medical Systems, Tucson, AZ, US) for 90 minutes and then stained using the BenchMark XT autostainer (Ventana Medical Systems). Pre-diluted anti-NF-kB p65 mouse monoclonal antibody (sc-8008 [F-6], Santa Cruz Biotechnology, Santa Cruz, CA, US) at 1:100 was manually added to the slides and incubated at 37°C for 60 minutes. Antibody binding was revealed using the ultraView Universal DAB Detection Kit (#760–500, Ventana Medical Systems). Counterstaining was achieved using hematoxylin and bluing reagents (#760–2021 and #760–2037, Ventana Medical Systems). Tissues were dehydrated and mounted using Sub-X Mounting Medium (Leica Microsystems, Concord, ON, Canada). All sections were scanned using a VS-110 microscope with a 20× 0.75 NA objective and a resolution of 0.3225 μm (Olympus Canada, Richmond Hill, ON, Canada).
The frequency (0%–100%) of epithelial cells with nuclear p65 expression in benign and tumor cores was assessed by 2 observers using digitalized images. Interclass correlation of the scoring for each core between the 2 observers was 0.88. When more than 1 core per patient was evaluated, the average p65 nuclear frequency for each patient (benign adjacent tissue and tumor) was used for subsequent analyses.
Statistical analyses were performed with SPSS version 23.0 (SPSS, Chicago, IL, US). Interclass comparisons were performed to evaluate the agreement between the 2 observers. The correlation with clinicopathological parameters was estimated with a non-parametric Spearman correlation test. The analysis plan was to evaluate the association of p65 nuclear frequency with PC patient clinical endpoints, which included the BCR, the development of bone metastases, and PC-specific mortality. The cutoff applied for the dichotomization of the data was defined by the median frequency of nuclear p65 expression (3%) in the Test-TMA series. This threshold was then applied to the Validation-TMA series. BCR-free, bone-metastasis-free, and PC-specific survival curves were plotted using the Kaplan–Meier estimator, and the log-rank test was used to evaluate significant differences. The univariate and multivariate proportional hazard models (Cox regression) were used to estimate the hazard ratios (HRs) for nuclear p65 frequency. For multivariate analyses, the serum PSA level prior to RP, the pathological staging of the primary tumor (pT2, pT3, pT4), the Gleason score category (6, 7 [3+4], 7 [4+3], 8+), and the margin status (negative/positive) were included in the model. In the rare occasion where clinical data were missing, the case was withdrawn from the analyses. Results were considered statistically significant at p-values below 0.05.
The STROBE checklist is provided in S1 Table.
The multi-institutional CPCBN cohorts and TMAs have been previously characterized [16]. Pathology review and grading of each core had been performed and showed that Gleason grading of TMA cores agreed with RP specimen grades [17]. Nuclear p65 expression in the CPCBN TMAs was assessed in 2 distinct steps. First, we attained access to the Test-TMA series, a cohort of 250 patients. Next, our results were reviewed by the CPCBN committee for approval of access to the second cohort of 1,262 patients, the TMA-Validation series. Clinicopathological characteristics of PC patients in each cohort are presented in Table 1.
Immunohistochemical staining of both TMA series showed variable levels of nuclear p65 expression in PC specimens, ranging from 0% to 100% of epithelial cells with positive staining. Median frequency of nuclear p65 expression in tumor cores was 3%. Representative images of negative (0%), low (>0% and ≤3%), and high (>3%) expression are presented in Fig 1. We observed a statistically significantly higher level of p65 nuclear frequency in tumor cores compared to benign adjacent cores in both the Test-TMA and the Validation-TMA (Fig 1G and 1H, respectively).
Correlation of p65 nuclear frequency with clinical characteristics was evaluated as continuous or dichotomized data. Threshold was based on the median nuclear frequency of 3% in tumor cores and 1% in benign adjacent cores in the Test-TMA series. The same threshold was applied to the Validation-TMA series. The Test-TMA cohort contained 128 PC patients with >3% positive nuclear expression of p65 in tumors compared to 456 PC patients of the Validation-TMA cohort. Kaplan–Meier survival curves show that nuclear p65 expression tended to be associated with an increased risk of BCR in the Test-TMA cohort (p = 0.06) (Fig 2A), and this association was highly significant in the Validation-TMA cohort (p < 0.001) (Fig 2B). Univariate Cox regression analysis demonstrated that all known predictors of BCR (preoperative PSA level, pathological staging, RP Gleason score, and margin status) were statistically significant in both cohorts. Although univariate regression analyses of both continuous (HR 1.02 [95% CI 1.00–1.05], p = 0.099) and dichotomized (HR 1.54 [95% CI 0.98–2.44], p = 0.063) data of p65 nuclear frequency in tumors did not reach significance in the Test-TMA series (Table 2), a strong association was observed for continuous (HR 1.03 [95% CI 1.02–1.04], p < 0.001) and dichotomized (HR 1.60 [95% CI 1.32–1.94], p < 0.001) data in the Validation-TMA series (Table 3). In a multivariate analysis, nuclear p65 was an independent prognostic parameter for BCR in both continuous (HR 1.02 [95% CI 1.00–1.03], p = 0.004) and dichotomized (HR 1.33 [95% CI 1.09–1.62], p = 0.005) data in the Validation-TMA series (Table 3). In contrast, nuclear p65 in benign adjacent tissues was not associated with BCR (Tables 2 and 3).
The progression of PC leads to the spread of the disease, particularly to bones [18]. To evaluate the prognostic significance of the nuclear localization of p65, we assessed its association with the development of bone metastases. In univariate and multivariate Cox regression analyses, pathological staging and RP Gleason score showed an association with bone metastases in both TMA series (Tables 4 and 5). Although significant in univariate analysis, preoperative PSA level did not remain significant in the multivariate analysis. Univariate analysis on the Test-TMA cohort showed that the continuous value of p65 nuclear frequency was significantly associated with bone metastases (HR 1.06 [95% CI 1.01–1.11], p = 0.023); dichotomized data of nuclear p65 almost reached a significant association (HR 4.14 [95% CI 0.89–19.19], p = 0.069), with an HR of 4.14. The Validation-TMA series confirmed that p65 nuclear expression was a strong predictor of bone metastasis development using both continuous (HR 1.04 [95% CI 1.01–1.06], p = 0.003) and dichotomized (HR 2.13 [95% CI 1.23–3.66], p = 0.007) data. This association remained significant when clinical parameters (Gleason score and pathological staging) were included in the model using the dichotomized data (HR 2.63 [95% CI 1.30–5.31], p = 0.033). Kaplan–Meier estimates also demonstrated that bone-metastasis-free survival was shorter for PC patients with a high p65 nuclear localization in both TMA series (log-rank p = 0.048 for Test-TMA and p = 0.005 for Validation-TMA) (Fig 2C and 2D).
Overall survival of PC patients was assessed only in the Validation-TMA series since the number of events was too small in the Test-TMA series. The percentage of tumor cells showing nuclear expression of p65 was significantly associated with shorter PC-specific survival (p = 0.001) (Fig 2E). Univariate Cox regression analysis showed a significant association of PC-specific mortality with staging (HR 3.26 [95% CI 1.84–5.78], p < 0.001) and RP Gleason score (HR 3.56 [95% CI 2.43–5.20], p < 0.001) (Table 6). Moreover, dichotomization of nuclear p65 frequency in tumor tissues was also significantly associated with disease-specific mortality (HR 3.12 [95% CI 1.55–6.27], p = 0.001), and this association remained independent when combined with clinical parameters (HR 2.63 [95% CI 1.30–5.31], p = 0.033). Overall, our results show that p65 nuclear frequency is an independent prognostic marker for PC.
Integration of powerful prognostic biomarkers in the pathology workflow would help identify patients with an increased risk of developing an aggressive disease. PSA level, stage, and Gleason score are the main prognostic parameters used to identify low-, intermediate-, and high-risk PC. The accuracy of PC stratification would gain from the addition of new prognostic biomarkers. The use of immunohistochemistry is a particularly suitable approach since immunohistochemistry is already used routinely by genitourinary pathologists. Immunohistochemical assays also allow for evaluation of protein subcellular localization [19]. In particular, the combination of immunohistochemistry and digital pathology could facilitate the standardization of biomarker expression analysis and simplify evaluation [20].
The goal of the CPCBN research program is to identify promising markers for integration into the clinical workflow to improve PC patient management [16]. The pan-Canadian multi-centre cohort is an important addition to PC research to support the validation of potential biomarkers that have been reported from studies using small cohorts from single institutions. Previously, we and others have shown that overall p65 expression, or more specifically its nuclear frequency in PC, is associated with BCR in single institution cohorts [8–14,21–23]. Here, we identified a threshold of p65 nuclear expression in a multi-centre cohort of 250 PC patients and validated this prognostic biomarker in a second independent multi-centre cohort of 1,262 PC patients. This study represents the first report of biomarker evaluation using the Test-TMA and Validation-TMA series of the CPCBN and demonstrates the usefulness of the resource for the biomedical community.
In addition to validating nuclear p65 as a predictor of BCR, we found that this biomarker was also associated with bone metastasis development and PC-specific death. These observations, to our knowledge, have never been reported previously and highlight that such findings require large cohorts with longer follow-up, such as the Validation-TMA series, which includes a median follow-up of 71 months, to achieve statistically significant associations. The present study shows that bone-metastasis-free survival and PC-specific survival are shorter for patients with higher p65 nuclear frequency in cancer cells. This biomarker is an independent predictor of prognosis since it remained significant when analyses were adjusted for pathological staging of the primary tumor and Gleason score at RP. In addition, patients with a high frequency of nuclear p65 expression showed a greater risk for PC-specific death than those with lower expression. Preoperative PSA value as a biomarker failed to reach statistical significance for the prediction of bone metastases and PC mortality. These data suggest that p65 could augment the established clinical prognostic markers used to stratify PC risk and could be a useful parameter in pathological practice.
Our results are compatible with the known activities of the NF-κB pathway. Although activation of NF-κB does not induce the development of PC, its expression is associated with the progression of the disease. The inactivation of IκBα, an inhibitor of p65, in the Hi-Myc mouse PC model increases the aggressiveness of the disease [24]. Moreover, activation of NF-κB pathways in PC cells induces the expression of osteoclastogenic genes such as receptor activator of NF-κB ligand (RANKL) and parathyroid hormone-related protein (PTHrP) [25]. These 2 proteins are well known for their contribution to the development of bone metastases. Our study shows that nuclear localization of p65 is an independent predictor of bone metastases and disease-specific death. In addition, constitutive activation of p65 has been identified during progression of PC to castration resistance, a stage when patients no longer respond to anti-androgen therapy [26].
We also hypothesize that nuclear p65 could act as a predictive biomarker for specific treatments. Bortezomib is recognized to block NF-κB pathways through the inhibition of the 26S proteasome [27]. This specific mechanism involves inhibition of IκB degradation, confining NF-κB to the cytoplasm. Currently, bortezomib is used to treat cancers such as multiple myeloma and mantle cell lymphoma, in which NF-κB is highly activated. However, the NF-κB pathway blocker is not yet indicated for phase III studies for PC patients due to the inadequate activity-to-toxicity ratio [28–30]. The recruitment of PC patients with a highly activated NF-κB pathway may be necessary to ensure a higher response rate. Future studies should be considered to evaluate the nuclear frequency of p65 by immunohistochemistry on PC tissues from early phase bortezomib trials to evaluate the theragnostic potential for p65 in this treatment regimen.
In conclusion, we demonstrated the prognostic significance of nuclear p65 in 2 independent PC TMA cohorts that included primary tumor tissues of patients from multiple institutions. To our knowledge, our study is the first to highlight the prognostic ability of nuclear p65 to identify patients with an increased risk of developing bone metastases and PC-specific mortality. The implementation of this biomarker in the clinical workflow would allow genitourinary pathologists and clinicians to improve the identification of patients with high-risk PC.
|
10.1371/journal.ppat.1007923 | Toxoplasma gondii activates a Syk-CARD9-NF-κB signaling axis and gasdermin D-independent release of IL-1β during infection of primary human monocytes | IL-1β is a potent pro-inflammatory cytokine that promotes immunity and host defense, and its dysregulation is associated with immune pathology. Toxoplasma gondii infection of myeloid cells triggers the production and release of IL-1β; however, the mechanisms regulating this pathway, particularly in human immune cells, are incompletely understood. We have identified a novel pathway of T. gondii induction of IL-1β via a Syk-CARD9-NF-κB signaling axis in primary human peripheral blood monocytes. Syk was rapidly phosphorylated during T. gondii infection of primary monocytes, and inhibiting Syk with the pharmacological inhibitors R406 or entospletinib, or genetic ablation of Syk in THP-1 cells, reduced IL-1β release. Inhibition of Syk in primary cells or deletion of Syk in THP-1 cells decreased parasite-induced IL-1β transcripts and the production of pro-IL-1β. Furthermore, inhibition of PKCδ, CARD9/MALT-1 and IKK reduced p65 phosphorylation and pro-IL-1β production in T. gondii-infected primary monocytes, and genetic knockout of PKCδ or CARD9 in THP-1 cells also reduced pro-IL-1β protein levels and IL-1β release during T. gondii infection, indicating that Syk functions upstream of this NF-κB-dependent signaling pathway for IL-1β transcriptional activation. IL-1β release from T. gondii-infected primary human monocytes required the NLRP3-caspase-1 inflammasome, but interestingly, was independent of gasdermin D (GSDMD) cleavage and pyroptosis. Moreover, GSDMD knockout THP-1 cells released comparable amounts of IL-1β to wild-type THP-1 cells after T. gondii infection. Taken together, our data indicate that T. gondii induces a Syk-CARD9/MALT-1-NF-κB signaling pathway and activation of the NLRP3 inflammasome for the release of IL-1β in a cell death- and GSDMD-independent manner. This research expands our understanding of the molecular basis for human innate immune regulation of inflammation and host defense during parasite infection.
| IL-1β is a proinflammatory cytokine that contributes to host defense against infection and is also associated with autoimmune and inflammatory diseases. Our prior research has demonstrated that the intracellular parasite Toxoplasma gondii induces IL-1β release from primary human monocytes during infection. Here we report the novel finding that within minutes of infection, T. gondii activates a spleen tyrosine kinase (Syk), PKCδ, CARD9/MALT-1, and NF-κB signaling pathway that is critical for the production of IL-1β in primary human monocytes. We have also investigated the mechanism of IL-1β release from monocytes. Interestingly, although IL-1β can be released during pyroptotic cell death, which is driven by gasdermin family proteins such as gasdermin D (GSDMD), we have found that T. gondii triggers the release of IL-1β from viable cells, independent of GSDMD, thereby preserving the parasite’s intracellular niche. These studies provide mechanistic insight into the regulation of inflammation and host defense against parasite infection by human innate immune cells.
| Toxoplasma gondii is an obligate intracellular foodborne parasite capable of infecting and replicating in any nucleated cell of its infected hosts [1]. Global estimates suggest that as much as a third of the world population is chronically infected with this parasite and that over thirty million people become ill from T. gondii infections each year [2,3]. While a robust immune response typically controls the infection, T. gondii poses severe health risks to immunocompromised individuals and to the developing fetus during congenital disease [4,5]. In particular, CD4+ and CD8+ T cells and the production of IFN-γ are required for protection against T. gondii infection [6,7]. Innate immune cells also contribute significantly to host defense against T. gondii through the production of IL-12 and cell intrinsic mechanisms of host defense [8]. Monocytes, in particular, are among the first cells recruited to sites of T. gondii infection and are critical for parasite control during both the acute and chronic stages of disease [9–14].
IL-1β is a potent pro-inflammatory cytokine that is induced by infection and injury and coordinates both the innate and adaptive immune responses [15]. Uncontrolled production of IL-1β has been implicated in the pathogenesis of a variety of diseases such as atherosclerosis, arthritis, diabetes, inflammatory bowel disease, and Alzheimer’s disease [16,17], indicating that IL-1β production and release must be tightly controlled to maintain healthy immune function, during both homeostasis and infection.
Myeloid cells, such as macrophages and monocytes, are major producers of IL-1β during infection or injury. Macrophages regulate IL-1β via a two-signal model. The first signal (Signal 1) is typically induced by Toll-like receptor (TLR) engagement and MyD88 signaling that results in NF-κB nuclear translocation and IL-1β transcription [18]. IL-1β is then translated as a biologically inactive pro-protein that cannot bind to the IL-1 receptor until it is cleaved into the mature form by a protease, such as caspase-1. The second signal (Signal 2) activates a multiprotein complex called the inflammasome, of which at least five have been described, which leads to caspase-1 activation and IL-1β cleavage and release [19]. Interestingly, inflammasomes are differentially regulated in macrophages and monocytes [20], and even in human and mouse monocytes: human monocytes activate the inflammasome and release IL-1β in response to LPS alone, using a “one-step” pathway, whereas mouse monocytes stimulated with LPS require an additional Signal 2 for IL-1β cleavage and release [21]. These differences in response to stimulation may reflect unique species- and cell-specific strategies for the regulation and induction of inflammation. Inflammasome activation and IL-1β production are also differentially regulated depending on the nature of the stimulus, which can be as diverse as pathogen infection, microbial products, or sterile inducers of inflammation.
Unlike most cytokines, IL-1β does not possess a signal peptide or traffic through the standard secretory pathway [22]. Instead the best-characterized mechanism of IL-1β release from myeloid cells is through an inflammatory form of cell death known as pyroptosis [23]. Activation of the noncanonical or canonical NLRP3 inflammasome induces pyroptosis through caspase-11- or caspase-1-mediated cleavage of gasdermin D (GSDMD) [24,25]. The cleaved N-terminal domain of GSDMD then inserts into the plasma membrane, where it forms pores through which IL-1β can pass. These pores allow for an influx of extracellular fluid, cell swelling, and eventually pyroptosis, which can release any remaining IL-1β into the extracellular space [26–28]. Recent work has shown GSDMD-dependent pore formation can also mediate IL-1β release from viable “hyperactivated” cells [29], suggesting that GSDMD could serve as a critical mechanistic unifier for the release of IL-1β from both pyroptotic and viable, non-pyroptotic cells.
IL-1β production contributes to host control of T. gondii infection [30–33], and we have previously shown that T. gondii infection of primary human monocytes induces the production of IL-1β transcripts and activation of the NLRP3 inflammasome [34,35]. However, T. gondii infection does not activate any known human TLRs, and the signaling pathways involved in TLR-independent IL-1β production during infection, particularly in human cells, remain poorly defined. In the current study, we demonstrate that primary human monocytes infected with T. gondii produced IL-1β through a Syk-PKCδ-CARD9/MALT-1-NF-κB signaling pathway and activated the NLRP3 inflammasome for IL-1β release from viable cells in a GSDMD-independent manner. Moreover, we have defined differences in the role of Syk in T. gondii-infected compared to LPS-stimulated primary human monocytes: during T. gondii infection, Syk was critical for pro-IL-1β synthesis, whereas in LPS-stimulated monocytes, Syk predominantly mediated IL-1β release. These studies detailing the activation and regulation of the IL-1β pathway during infection and in response to microbial products further our understanding of how primary human immune cells regulate inflammation when activated by diverse stimuli.
To investigate IL-1β production and release from primary human monocytes during T. gondii infection, we isolated monocytes from healthy blood donors as previously described (S1 Fig and [34]) and immediately infected the cells in vitro with GFP-expressing T. gondii or treated them with an equal volume of culture medium (mock). At 4 hours post-infection (hpi), IL-1β released into the supernatant was detected by ELISA, and the response of cells from individual human donors was compared (each dot represents a unique donor) (Fig 1A). The pretreatment of primary monocytes with MCC950, an NLRP3 inhibitor; YVAD, a caspase-1 inhibitor; or KCl, which prevents K+ efflux and inhibits activation of the canonical NLRP3 inflammasome [36], all resulted in a significant decrease in IL-1β release from infected primary human monocytes from multiple independent donors (Fig 1A). Notably, none of these treatments or inhibitors affected the efficiency of infection, as determined by the percent of GFP+ (infected) cells in the culture (S2A Fig) or the ability of the parasite to replicate in and lyse human foreskin fibroblasts (HFFs), as determined by plaque assays (S2B Fig). The transfer of supernatants from T. gondii-infected, but not mock-treated monocytes to the HEK-Blue reporter cell line resulted in reporter cell activation, indicating the release of functional IL-1 from the infected primary monocytes (Fig 1B).
There are three subsets of peripheral blood monocytes that have been described in humans, which are defined by their relative expression of CD14 and CD16 [37] (Fig 1C). Recent publications have shown that the CD14+CD16- inflammatory subset of monocytes is associated with increased and chronic inflammation and the development of arthritis [38,39]. Using intracellular cytokine staining (ICCS) to compare IL-1β production in each subset, we found that T. gondii infection stimulated all three subsets to produce IL-1β by 4 hpi (Fig 1D). We also observed that in each of the five donors analyzed, the CD14+CD16- inflammatory monocytes exhibited the highest percentage of IL-1β+ cells (Fig 1D). Collectively, these data demonstrate that T. gondii triggers the production of IL-1β in all subsets of primary peripheral blood human monocytes and activates the canonical NLRP3 inflammasome for the release of bioactive IL-1 by 4 hpi.
Syk is a tyrosine kinase that is expressed in hematopoietic cells and is involved in NLRP3 activation during fungal infection, viral infection, and in response to LPS stimulation [40–43]. However, the role of Syk in IL-1β regulation during parasite infection is unknown. Interestingly, rapid phosphorylation of Syk at tyrosine 525/526, which is an indicator of activation, was detected by phospho-flow cytometry (Fig 2A) and Western blotting of lysates (Fig 2B) from monocytes that were infected with T. gondii or treated with LPS, as a positive control. The phosphorylation of Syk was reduced in the presence of the Syk-specific inhibitor R406 (Fig 2B). To investigate a potential role for Syk in IL-1β release during T. gondii infection, primary human monocytes were pre-treated with the Syk inhibitor R406 or a vehicle control, and either infected with T. gondii or treated with LPS. LPS stimulation induced significantly more IL-1β release than T. gondii infection, but R406 treatment significantly reduced IL-1β release from primary human monocytes treated with either stimulus at the 4 hour time-point (Fig 2C). Titration of R406 revealed a dose-dependent effect of the Syk inhibitor (Fig 2D). Importantly, pretreatment of monocytes with the R406 inhibitor did not reduce the infection efficiency of the parasite or the GFP median fluorescence intensity (MFI) of monocytes infected with GFP-expressing parasites at either 4 hpi or 16 hpi (S3A Fig). In addition, R406 did not decrease the ability of T. gondii to grow in and lyse HFFs, as measured by plaque assays (S3B Fig), or the viability of monocytes (S3C Fig). Furthermore, an independent Syk inhibitor, entospletinib, which is currently in use in clinical trials for leukemia [44], also reduced IL-1β release from T. gondii-infected monocytes in a dose-dependent manner (Fig 2D), without affecting parasite infection efficiency (S2A Fig) or viability (S2B Fig). Collectively, these data indicate on-target effects of the Syk inhibitors and demonstrate that T. gondii infection induces IL-1β release from primary human monocytes in a Syk-dependent manner.
Syk has been proposed to play two roles in the regulation of IL-1β production in other models: inducing NF-κB activation and IL-1β transcription via the PKCδ-CARD9/MALT-1-NF-κB pathway or indirectly activating NLRP3 inflammasome assembly via ASC phosphorylation and oligomerization [45,46]. In the lysates and supernatants of primary human monocytes pretreated with the Syk inhibitor R406, we observed a marked reduction in both pro- and mature IL-1β protein levels compared to infected monocytes treated with the vehicle control (Fig 3A). In contrast, R406 pre-treatment of LPS-stimulated primary monocytes did not reduce the levels of pro-IL-1β protein in the cell lysates (Fig 3A). Notably, the NLRP3 and caspase-1 inhibitors MCC950 and YVAD, respectively, had no effect on pro-IL-1β synthesis or release from primary monocytes, as expected (Fig 3B). ICCS of primary human monocytes infected with T. gondii or treated with LPS in the presence or absence of R406 indicated that Syk inhibition reduced the percentage of intracellular IL-1β+ T. gondii-infected monocytes but did not decrease the percentage of intracellular IL-1β+ monocytes stimulated with LPS (Fig 3C).
To directly examine the effect of Syk inhibition on IL-1β and NLRP3 transcript levels, qPCR was performed on samples from human monocytes infected with T. gondii or treated with LPS in the presence or absence of R406. These data corroborated the Western blot and ICCS data and demonstrated that R406 decreased IL-1β and NLRP3 transcripts at 1 and 4 hpi in T. gondii-infected monocytes (Fig 3D). Interestingly, R406 treatment also reduced IL-1β and NLRP3 transcripts in LPS-stimulated monocytes (Fig 3D), despite having little to no effect on pro-IL-1β levels in these cells (Fig 3A). Together these data suggest that Syk signaling is critical for production of IL-1β and NLRP3 transcripts in T. gondii-infected primary human monocytes, and therefore appears to act in the priming stage of IL-1β production during infection.
Since primary human monocytes cannot be genetically manipulated (or reliably cultured in vitro for more than ~24 hours), we examined a role for Syk in the human monocytic cell line THP-1. These cells also release IL-1β in response to T. gondii infection, but with delayed kinetics compared to primary monocytes [34,35]. Similar to primary monocytes, T. gondii infection induced Syk phosphorylation in THP-1 cells (S4A Fig), and pre-treatment of THP-1 cells with R406 resulted in the release of significantly less IL-1β than in infected THP-1 cells treated with the vehicle control (Fig 4A). To complement the R406 inhibitor experiments, THP-1 cells were transduced with lentivirus carrying guide RNAs targeting Syk for CRISPR/Cas9-mediated genome editing. As a control, THP-1 cells were transduced with an empty vector (EV) lacking the Syk targeting sequence. Unlike in the wild-type (WT) parental THP-1 cells, Cas9 was detected in the EV control and Syk KO lines, and Syk was absent only from the KO line (Fig 4B). This Syk KO clone harbors mutations in the second SH2 domain, resulting in a frameshift mutation that alters the amino acid sequence of the targeted exon (S4 Fig).
The control EV line and the Syk KO THP-1 cells were infected with T. gondii and examined for IL-1β production. qPCR analysis revealed reduced levels of IL-1β mRNA in the Syk KO THP-1 cells compared to the EV control cells (Fig 4C). Similarly, reduced levels of pro-IL-1β protein were detected in the Syk KO cells compared to the EV control cells during infection (Fig 4D), suggesting that Syk functions upstream of IL-1β transcription and pro-IL-1β protein synthesis in THP-1 cells, similar to the effects of the R406 and entospletinib inhibitors in primary human monocytes. Finally, Syk KO THP-1 cells released less IL-1β in the supernatant, as detected by ELISA, than the control EV cells during T. gondii infection (Fig 4E). Collectively, these data indicate that T. gondii infection induces IL-1β synthesis and release from both primary human monocytes and THP-1 cells in a Syk-dependent manner.
While it has been well documented that LPS activates a MyD88-IRAK1/4-TRAF6 pathway resulting in NF-κB nuclear translocation, Syk has been shown to activate an alternative PKCδ-CARD9/MALT-1-NF-κB signaling pathway [46]. To investigate a potential role for this pathway in IL-1β production during T. gondii infection of human monocytes, we examined the activation of PKCδ and the NF-κB subunit p65 and found that T. gondii infection induced phosphorylation of both PKCδ and p65 in primary human monocytes (Fig 5A and 5B). Treatment of monocytes with R406; Go6983, a PKC inhibitor which is active against PKCδ; MI2, a MALT-1 and CARD9 complex inhibitor; or PS1145, an IKK inhibitor, prior to T. gondii infection all significantly reduced p65 phosphorylation induced by T. gondii infection (Fig 5B), indicating that the inhibitors all targeted a pathway upstream of NF-κB activation. Similar to the Syk inhibitor, the PKCδ, CARD9/MALT-1, and IKK inhibitors all reduced pro-IL-1β protein production in T. gondii-infected primary human monocytes at 4 hpi (Fig 5C). Consistent with these data, IL-1β release was significantly reduced in primary human monocytes treated with these inhibitors compared to vehicle control-treated monocytes during infection, as determined by ELISA (Fig 5D). Importantly, the inhibitors did not reduce infection efficiency at 4 hpi (S2A Fig) or affect the ability of the parasites to replicate in and lyse HFFs at the concentrations used (S2B Fig). Together, these data suggest that primary human monocytes rely almost completely on signaling through a Syk-PKCδ-CARD9/MALT-1-NF-κB signaling pathway for IL-1β production during T. gondii infection.
To investigate roles for PKCδ and CARD9 in IL-1β production during T. gondii infection using a genetic approach, THP-1 cells were subjected to CRISPR/Cas9-mediated genome editing using guide RNAs targeting each of these two proteins. Cas9 protein was detected in the EV control cells and the KO cells, and PKCδ and CARD9 were absent or severely reduced in each of the respective KO populations (Fig 6A). The faint detection of PKCδ in the PKCδ KO population (Fig 6A and 6C) may reflect the fact that these cells represent a mixed population, rather than a clonal KO line. Infection of either the PKCδ KO or CARD9 KO THP-1 cells with T. gondii resulted in reduced IL-1β release compared to infection of the EV cells (Fig 6B). In addition, T. gondii-infected PKCδ KO and CARD9 KO THP-1 cells contained less pro-IL-1β protein in the cell lysates than T. gondii-infected EV THP-1 cells (Fig 6C). These data indicate that both PKCδ and CARD9 contribute to IL-1β synthesis and release from THP-1 cells, consistent with the results obtained using inhibitors of these proteins in primary human monocytes.
The most well characterized mechanism of IL-1β release due to inflammasome activation is an inflammatory form of cell death, marked by cell membrane pore formation, cell swelling, and lysis, termed pyroptosis [23]. To address a potential role for pyroptosis in IL-1β release during T. gondii infection, primary human monocytes were mock treated, infected with T. gondii, or stimulated with LPS or LPS and ATP, as controls. The cells were then stained with propidium iodide (PI), which passes through small pores in the plasma membrane and binds to DNA. At 4 hpi, when bioactive IL-1β release was detected (Fig 1B), the viability of the T. gondii-infected monocyte population, as measured by the percentage of PI+ cells, was indistinguishable from mock-treated cells (Fig 7A), and the addition of the Syk inhibitor R406 did not alter the percentage of PI+ cells. In contrast to T. gondii infection and LPS stimulation, a high level of cell death was detected when cells were treated with the canonical inflammasome activator, LPS and ATP (Fig 7A). Titrating the ATP in LPS-stimulated cells triggered cell death in a dose-dependent manner (S5 Fig). The addition of extracellular glycine, which inhibits ion flux, thereby halting cell swelling and the completion of pyroptosis, reduced IL-1β release from LPS and ATP-stimulated cells but not from LPS-stimulated or T. gondii-infected cells (Fig 7B). In contrast, Syk inhibition with R406 did reduce IL-1β release in the LPS-stimulated and T. gondii-infected cells, without affecting cell death (Fig 7A and 7B). Interestingly, the stimulus of LPS and ATP, which induced cell death, led to the release of more than ten times the amount of IL-1β from primary human monocytes than T. gondii infection or LPS stimulation alone (Fig 7B). These results support the idea that the degree of pyroptosis and IL-1β release during stimulation may relate to the intensity of the stimulus encountered by the cells [47].
Although we did not detect significantly more cell death among T. gondii-infected monocytes compared to mock-treated cells, we formally tested a role for GSDMD by infecting wild-type (WT) and GSDMD knockout THP-1 cells [48] with T. gondii and examining IL-1β release by ELISA (Fig 7C). Notably, the GSDMD KO cells were not impaired in their release of IL-1β during T. gondii infection compared to WT THP-1 cells (Fig 7C), and the viability of these cells was not significantly different than that of mock-treated THP-1 cells at the same time-point (Fig 7C). Furthermore, whereas LPS and ATP stimulation of primary human monocytes led to the cleavage of GSDMD from the full-length 60 kD protein to the N-terminal p30 fragment, neither LPS nor T. gondii infection resulted in increased GSDMD cleavage at 4 hpi, and Syk inhibition with R406 did not affect GSDMD cleavage in the T. gondii or LPS conditions (Fig 7D). These data further support the conclusion that LPS and T. gondii trigger IL-1β release from human monocytes independent of GSDMD cleavage, pore formation, and pyroptosis.
Syk is a tyrosine kinase expressed in immune cells and is typically activated by receptors or receptor-associated adaptor proteins with cytoplasmic immunoreceptor tyrosine-based activation motifs (ITAMs) [49]. Syk is known to be critical for lymphatic development, inflammatory signaling, and inflammasome activation [50]. We now demonstrate that inhibition of Syk in primary human monocytes and genetic deletion of Syk in monocytic THP-1 cells both significantly reduced IL-1β transcripts and pro-IL-1β production during T. gondii infection, indicating a role for Syk in IL-1β synthesis in parasite-infected cells. Among other signaling pathways, Syk can signal through PKCδ and CARD9 to induce NF-κB activation [46,51], and indeed, inhibitors against PKCδ, CARD9/MALT-1, and IKK, or genetic deletion of PKCδ and CARD9 in THP-1 cells revealed the importance of this pathway in IL-1β synthesis in T. gondii-infected monocytes, as depicted in Fig 8. Syk also contributed to NLRP3 transcript induction in response to T. gondii infection, further supporting a role for Syk in the priming of both IL-1β and NLRP3. In contrast, Syk appeared to be less important for the production of pro-IL-1β in LPS-stimulated primary human monocytes, but rather, seemed to contribute more substantially to IL-1β release from these cells. LPS-stimulated monocytes likely rely more heavily on canonical NF-κB signaling downstream of TLR4 and MyD88 for IL-1β transcription, and on Syk for processing and release of IL-1β. Indeed, it has been shown that Syk can lead to activation of the NLRP3 inflammasome through the indirect phosphorylation of the inflammasome adaptor protein ASC [45,52]. ASC phosphorylation induces its oligomerization, facilitating activation of the NLRP3 inflammasome and caspase-1 [53,54]. Collectively, our data indicate that immune cells responding to pathogens harboring multiple PAMPs or vitaPAMPs will likely regulate IL-1β differently than cells responding to a single PAMP or stimulus. For example, primary human monocytes utilize Syk signaling through ERK1/2 to produce IL-1β when dengue virus is in complex with antibody [43], and our research shows that during T. gondii infection Syk activates a separate signaling pathway to reach the same net result. Thus, examining the production and processing of IL-1β and other cytokines during infection with various live pathogens may unveil new pathways of regulation that could be critical for enhancing or dampening inflammation during different types of infections.
In mice, T. gondii is sensed by TLR11/12 recognition of the parasite actin-binding protein profilin [55–57]; interestingly, however, these TLRs are not functional in humans. And although Syk can be activated downstream of TLR4 [41,58], there is no known human TLR that has been shown to recognize T. gondii, suggesting that T. gondii-induced Syk signaling occurs via a different receptor. Recent work has shown that T. gondii-infected cells release alarmin S100A11, which binds to the RAGE receptor on monocytes [59], yet no innate immune sensor has been shown to directly bind to T. gondii PAMPs. Our data suggest that an ITAM-bearing receptor or adaptor protein that activates Syk may serve as a potential recognition receptor, and this possibility is under investigation. We have previously identified a partial role for the parasite-secreted protein GRA15 in IL-1β production in primary human monocytes [35], and Syk signaling may synergize with GRA15 to induce maximal priming of IL-1β production.
Although IL-1β production was detected in all three subsets of monocytes by ICCS, in each donor examined, the CD14+CD16- inflammatory monocytes produced more IL-1β in response to T. gondii than the other monocyte subsets. In human blood, this inflammatory subset of monocytes is present in significantly greater numbers than the other subsets. Recent research indicates that this CD14+CD16- population of inflammatory monocytes is largely responsible for pathogenic inflammation in arthritis and sepsis [38,39], and our data are consistent with these findings. An intriguing possibility is that inflammatory monocytes regulate the expression or function of the receptors or signaling molecules involved in T. gondii-induced IL-1β production differently than the other monocyte subsets, rendering them more responsive to inflammatory stimuli.
Since the discovery that the fungal metabolite brefeldin A, which inhibits conventional protein secretion, did not inhibit IL-1β secretion from stimulated immune cells [22], the potential mechanisms of IL-1β release have been intensely studied. The best characterized mechanism of release occurs through an inflammatory form of cell death marked by cell swelling and lysis, termed pyroptosis [23]. Notably, IL-1β can also be released from viable cells in a pyroptosis-independent manner [29]. Indeed, this is the case with human, but not mouse, monocytes treated with LPS [21]. In 2015, the identification and characterization of GSDMD [24,25], which can be activated by the inflammasome and functions as the effector protein of pyroptosis by forming small pores in the cell membrane [26–28], provided a molecular basis for this inflammatory form of cell death. Interestingly, in the context of T. gondii infection, GSDMD cleavage and cell death did not appear to drive IL-1β release from primary human monocytes, as there was no difference in the percent of viable T. gondii-infected or mock-infected monocytes at 4 hpi, the time-point when functional IL-1β was detected in the supernatant. In addition, glycine treatment, which inhibits ion flux and pyroptosis, had no effect on T. gondii-induced IL-1β release. While it cannot be completely ruled out that a small number of monocytes that die early during T. gondii infection are responsible for all the IL-1β released, this possibility seems unlikely because T. gondii continues to live and replicate within human monocytes for at least another 14 hours after maximal IL-1β release is detected, suggesting that the cells do not die rapidly after infection. In examining a role for GSDMD, we found that the cleaved, active N-terminal fragment of GSDMD was not increased in T. gondii-infected primary monocyte lysates. Finally, GSDMD KO THP-1 cells released comparable levels of IL-1β to wild-type THP-1 cells during T. gondii infection, further suggesting a pyroptosis-independent mechanism of IL-1β release from T. gondii-infected monocytes.
Our current findings support and expand on a threshold model in which the amount of IL-1β production and the mechanism of its release are dependent on the stimulus [47]. The amount of IL-1β released by primary human monocytes during LPS stimulation was significantly higher than that released during T. gondii infection, and LPS and ATP stimulation together induced almost an order of magnitude more IL-1β release than LPS alone. Notably, only LPS and ATP stimulation triggered significantly more cell death than mock-treated monocytes. These data suggest that perhaps different signaling pathways are activated to induce low, medium, and high amounts of IL-1β release, all depending on the stimulus that a cell encounters. The use of a variety of stimuli that can lead to the same response, but perhaps through different mechanisms, will be a valuable tool in developing a more comprehensive understanding of how human immune cells regulate inflammation. This work also demonstrates that IL-1β production can be uncoupled from IL-1β release during LPS stimulation of primary human monocytes and highlights GSDMD-independent mechanisms of IL-1β release in the context of viable cells. Collectively, the current findings not only provide a more detailed understanding of how human innate immune cells regulate inflammation but also shed light on the pathways that contribute to host defense against a parasite pathogen of global importance.
Human whole blood was collected by the Institute for Clinical and Translational Science (ICTS) at the University of California, Irvine from healthy adult donors who provided written informed consent. Blood was collected according to the guidelines of and with approval from the University of California, Irvine Institutional Review Board (HS #2017–3753).
Primary human monocytes were isolated from human whole blood collected by the Institute for Clinical and Translational Science (ICTS) at the University of California, Irvine from healthy adult donors. PBMCs were isolated from whole blood by density gradient centrifugation using lymphocyte separation media (MP Biomedicals, Santa Ana, CA). Monocytes were enriched from PBMCs by counterflow elutriation, as previously described [34], and stained for purity after isolation. This protocol typically resulted in >90% pure monocyte cultures (ranging from 85–95%) based on CD11b+ and CD3−CD20−CD56− staining (S1 Fig). Freshly isolated monocytes were resuspended in RPMI 1640 (HyClone, Logan, UT) supplemented with 2 mM L-glutamine, 100 U/ml penicillin, 100 μg/ml streptomycin, and either 10% heat-inactivated FBS (Omega Scientific, Tarzana, CA) (R-10%) or no serum (R-0%). Monocytes were used immediately after isolation for experiments.
The human monocytic THP-1 cell line and the gasdermin D knock-out (GSDMD KO) THP-1 cells, a gift from Dr. Derek Abbott (Case Western Reserve University) [48], were cultured in R-10% (HyClone, Logan, UT) supplemented with 2 mM L-glutamine, 100 U/ml penicillin, and 100 μg/ml streptomycin. The Syk KO, CARD9 KO, and PKCδ KO THP-1 cells were cultured in R-10% (HyClone, Logan, UT) supplemented with 2 mM L-glutamine, 100 U/ml penicillin, 100 μg/ml streptomycin and 2 μg/ml puromycin.
Human foreskin fibroblasts (HFFs; from the lab of Dr. John Boothroyd, Stanford University School of Medicine) were cultured in D-10% medium: DMEM (HyClone) supplemented with 10% heat-inactivated FBS, 2 mM L-glutamine, 100 U/ml penicillin, and 100 μg/ml streptomycin. T. gondii tachyzoites were maintained by serial passage in confluent monolayers of HFFs. Type II (Prugniaud) [60] parasites constitutively expressing GFP were used.
All mammalian and parasite lines were cultured at 37°C in 5% CO2 incubators. All cultures were tested bimonthly and confirmed to be free of mycoplasma contamination.
Knockout THP-1 cells were generated using the Lenti-CRISPR-Cas9 system. Guide RNAs (sgRNA) targeting human Syk, PKCδ, or CARD9 were cloned into the LentiCRISPR v2 plasmid (Feng Zhang, Addgene plasmid #52961). Virus was generated by transfecting the sgRNA plasmid constructs into HEK 293T cells along with the psPAX2 packaging (Didier Trono, Addgene plasmid #12260) and pCMV-VSVG envelope (Bob Weinberg, Addgene plasmid #8454) plasmids. Viral supernatants collected 48 hr post-transfection were used to infect THP-1 cells by spinfection at 1800 rpm for 1 hr. Single-cell Syk knockout clones were generated by limiting dilution in 96-well plates under puromycin selection. Single-cell clones were sequenced after PCR amplification of a 500 bp region near the Cas9 binding site. Interference of CRISPR edits (ICE) analysis software (Synthego) was used to characterize the indel for each clone. A Syk KO clone (named 1–6) with a biallelic indel-induced frame-shift mutation in the second SH2 domain was used for subsequent experiments. The PKCδ and CARD9 KOs are comprised of mixed cell populations. All the mixed populations and clones were screened via Western blot for the presence of Cas9 and the absence of the gene targeted for deletion. The sequences for the guide RNAs were as follows: Syk: GAAAGAAGTTCGACACGCTC, PKCδ: AGTTCTTACCCACGTCCTCC, and CARD9: ATCGTTCTCGTAGTCCGACA.
Primary human monocytes and monocytic THP-1 cells were resuspended in R-0% or R-10% medium directly after isolation and incubated with small molecule inhibitors or equal volumes of vehicle and incubated for 40 minutes at 37°C. T. gondii-infected HFF were washed with D-10% medium, scraped, and syringe lysed. Lysed tachyzoites were washed with R-0%, passed through a 5-μm filter (EMD Millipore, Billerica, MA), and washed with R-0% medium again. This resulted in parasite cultures that were free of host cell debris and soluble factors. Purified T. gondii tachyzoites were immediately added to host cells at a multiplicity of infection (MOI) of 2. All infections were performed with GFP-expressing type II T. gondii. “Mock” infections were samples in which an equivalent volume of culture medium without parasites was added to the cells.
Cells were stimulated with 100 ng/ml ultrapure E. coli LPS (List Biological Laboratories, Campbell, CA) and 5 mM ATP (Sigma-Aldrich, St. Louis, MO) for the last 30 min of culture, as indicated. “Mock” treatment was the addition of the equivalent volume of media (without parasites or LPS) to cells. At the indicated time point, monocytes were pelleted by centrifugation at 500 x g for 5 min. Collected cells were stained, fixed, or lysed accordingly, as described below.
MCC950 (Adipogen, San Diego, CA), glycine (Fisher Scientific, Waltham, MA) and potassium chloride (Fisher Scientific) were resuspended in deionized water. Go6983 (Selleck Chemicals, Houston, TX), Ac-Tyr-Val-Ala-Asp-chloromethylketone (Ac-YVAD-CMK or YVAD) (Cayman Chemical, Ann Arbor, MI), MI2 (Tocris Bioscience, Bristol, UK), PS1145 (Cayman Chemical), R406, and entospletinib (Selleck Chemicals), were all resuspended in DMSO. Monocytes were treated with the inhibitors or with an equivalent volume of the appropriate vehicle, for 40 min at 37°C and then infected or stimulated as described above. MCC950, Go6983, MI2, PS1145, R406 and entospletinib were all added to infected cells in half log titrations to determine the concentrations of the inhibitors that did not induce cell death or reduce infection efficiency.
HFF were plated on 6-well plates and grown to confluence for two days. HFF were pretreated with specific inhibitors for 40 min and then infected with freshly-lysed T. gondii tachyzoites for 5–7 days, followed by fixation with 10% Neutral Buffered Formalin. Staining was done using 600 μg/ml of Neutral Red solution overnight. The plaques were manually counted and imaged using a Leica DMi8 microscope with a DMC 5400 camera.
To measure infection efficiency, cells were harvested at the time points listed and immediately analyzed by FACS to determine the percent of GFP+ (T. gondii-infected) cells. To measure cell viability, cells were harvested, washed and resuspended in FACS buffer (2% FBS in PBS) containing propidium iodide and analyzed by flow cytometry without fixation. For cell surface staining, cells were blocked with Human TruStain FcX (BioLegend, San Diego, CA) on ice for 10 min and then stained with control Ig or the following anti-human Abs (all from BioLegend, unless otherwise indicated): anti-CD56–allophycocyanin (HCD56), anti-CD11b–PE or, anti-CD14–FITC (M5E2) or -PE/Cy7 (HCD14), anti-CD16-PE/Cy7 or -APC (3G8), anti-CD3–PE (UCHT1), or anti-CD20–PE/Cy7 (2H7). Cells were stained with the Abs on ice for 30 min, washed with FACS buffer, and either run live or fixed with 2–4% paraformaldehyde. For intracellular cytokine staining (ICCS), cells were fixed and permeabilized with 100 μL of BD Cytofix/Cytoperm solution (BD Biosciences, Franklin Lakes, NJ) for 20 minutes on ice. After incubation, cells were washed with FACS buffer containing 0.1% Triton-X, blocked with Human TruStain FcX as described above, stained with control Ig-PE or anti-IL-1β–PE (CRM56; eBioscience, San Diego, CA), control Ig-PE/Cy7 or anti-phospho-Syk (Y525/526)-PE-Cy7 (C87C1; Cell Signaling Technologies, Danvers, MA) Abs for 30 min, and washed with FACS buffer.
Samples were analyzed by flow cytometry on a FACSCalibur flow cytometer using CellQuest software (BD Biosciences). Data were analyzed using FlowJo software (TreeStar, Ashland, OR). Cells were first identified based on their forward and side scatter profile and subsequently analyzed for cell surface marker expression, intracellular cytokine expression, or GFP signal.
At the harvest time point, total RNA was harvested using the RNeasy Kit (QIAGEN, Germantown, MD) and treated with DNase I (Life Technologies, Carlsbad, CA) to remove any contaminating genomic DNA. cDNA was synthesized using the Superscript III First-Strand Synthesis kit (Life Technologies), according to the manufacturer’s instructions, and subsequently used as template in quantitative real-time PCR (qPCR). qPCR was performed in triplicate using a Bio-Rad iCycler PCR system (Bio-Rad, Hercules, CA) and iTaq Universal SYBR Green Supermix (Bio-Rad). Previously published sequences for IL-1β, NLRP3 and GAPDH primers were used [34]. All primer pairs spanned intron-exon boundaries whenever possible and bound to all isoforms of the gene, where applicable. All primers were commercially synthesized by Integrated DNA Technologies (Coralville, IA). qPCR data were analyzed using the threshold cycle method, as previously described [34], and gene expression data are shown normalized to that of the housekeeping gene GAPDH. In all qPCR assays, cDNA generated in the absence of reverse transcriptase, as well as water in the place of DNA template, were used as negative controls, and these samples were confirmed to have no amplification.
Human IL-1β protein released into the supernatant was measured using ELISA MAX Deluxe kits (BioLegend), according to the manufacturer’s instructions. Signal from ELISA plates was read with a Spectra Max Plus 384 plate reader (molecular Devices, San Jose, CA) using SoftMax Pro Version 5 software (molecular Devices), and the threshold of detection was 0.5 pg/ml.
At the harvest time point, cells were lysed by addition of 2X Laemmli buffer containing 10% 2-ME. For experiments in which supernatant was analyzed, serum-free R-0% medium was used during the infection, and supernatant was concentrated using Amicon Ultra Centrifugal filters (EMD Millipore, Burlington, MA), according to the manufacturer’s instructions. Concentrated supernatant was diluted with 2X Laemmli buffer containing 10% 2-ME. Samples were boiled at 100°C for 10 to 15 min, separated by SDS-PAGE, and transferred to polyvinylidene difluoride (PVDF) membranes (Bio-Rad, Hercules, CA) for immunoblotting. Membranes were blocked for 1 h at room temperature (RT) with blocking buffer: 5% non-fat milk or 5% bovine serum albumin (BSA) (Fisher Bioreagents). Membranes were then incubated with primary antibodies diluted in blocking buffer for 1 h at RT or overnight at 4°C. Membranes were probed with antibodies against NF-κB p65 (D14E12; Cell Signaling), phospho-NF-κB p65 (Ser536) (93H1; Cell Signaling), total Syk (2712S; Cell Signaling), phospho(Tyr525/526)-Syk (2711S; Cell Signaling), gasdermin D (NPB2-33422; Novus Biologicals, Littleton,CO), or β-actin (AC-15; Sigma-Aldrich). Membranes were blotted for IL-1β (3ZD from the National Cancer Institute Biological Resources Branch) using the SNAP i.d. Protein Detection System (EMD Millipore), according to the manufacturer’s instructions. Primary Abs were followed by HRP-conjugated secondary Abs (BioLegend), and membranes were developed with SuperSignal West Femto Maximum Sensitivity Substrate (Thermo Fisher Scientific, Carlsbad, CA), Amersham ECL Prime Western Blotting Detection Reagent (GE Healthcare, Little Chalfont, U.K.) substrate or ECL Prime Substrate (Thermo Fisher Scientific). Signal was detected using a Nikon camera, as previously described [34]. Quantification analysis of blots was performed using ImageJ software.
HEK-Blue IL-1 reporter cells (Invivogen, San Diego, CA), which respond to IL-1 binding to the IL-1R, were incubated in D-10% medium supplemented with 2 mM L-glutamine, 100 U/ml penicillin, 100 ug/ml streptomycin, Normocin (100 μg/ml), Hygromycin B Gold (200 μg/ml) and Zeocin (100 μg/ml). For the detection of IL-1 released from THP-1 cells and primary human monocytes, HEK-Blue cells were resuspended at a concentration of 500,000 cells/ml and added to flat-bottom 96-well plates. Supernatants collected from THP-1 cells or primary monocytes were added to the HEK-Blue cells and incubated for 24 hours at 37°C. The HEK-Blue cell supernatant was combined with the QUANTI-Blue detection reagent (Invivogen), incubated at 37°C for 1 to 3 hours, and then quantified with a Spectra Max Plus 384 plate reader (Molecular Devices, San Jose CA) using SoftMax Pro Version 5 (Molecular Devices) software.
Statistical analyses were performed using GraphPad Instat software. Analysis of variance (ANOVA) followed by Tukey’s or Bonferroni’s test, as indicated, were used for comparison between means. Differences were considered significant when the P value was <0.05.
|
10.1371/journal.pbio.1002046 | A Sensory-Motor Control Model of Animal Flight Explains Why Bats Fly Differently in Light Versus Dark | Animal flight requires fine motor control. However, it is unknown how flying animals rapidly transform noisy sensory information into adequate motor commands. Here we developed a sensorimotor control model that explains vertebrate flight guidance with high fidelity. This simple model accurately reconstructed complex trajectories of bats flying in the dark. The model implies that in order to apply appropriate motor commands, bats have to estimate not only the angle-to-target, as was previously assumed, but also the angular velocity (“proportional-derivative” controller). Next, we conducted experiments in which bats flew in light conditions. When using vision, bats altered their movements, reducing the flight curvature. This change was explained by the model via reduction in sensory noise under vision versus pure echolocation. These results imply a surprising link between sensory noise and movement dynamics. We propose that this sensory-motor link is fundamental to motion control in rapidly moving animals under different sensory conditions, on land, sea, or air.
| Bats are extremely skillful aviators: they are able to capture prey and land on targets under challenging flight conditions, as well as maneuver accurately using either echolocation or vision. It remains a mystery, however, how bats—or other flying animals—rapidly translate the noisy incoming sensory information into correct motor commands in order to converge onto a target. To address this question, we developed a sensorimotor control model that explains animal flight guidance and tested it in bats with experiments conducted under dark and light conditions. The model reproduced the bats’ flight trajectory with very high accuracy, suggesting that bats have to estimate not only the angle to target but also changes in the angle over time (angular velocity). Additionally, we demonstrate that the bat must suppress its sensory noise by integrating sensory information over several sonar pulses in order to successfully guide its flight. Comparisons of flight trajectories in light and dark suggest that the surprisingly curved flights exhibited by bats in the dark are due to sensory noise, not motor limitations. We hypothesize that rapidly moving animals must adaptively change their motor control strategy to optimally match the sensory conditions.
| An important open question in neuroscience is how animals transform incoming sensory information into motor commands to control movement [1–6]. This problem is particularly difficult in flying animals, in which sensory input is rapid and motor control must be accurate. Bats are the only mammals capable of true flight [7,8]. Many tasks performed by bats require fine maneuvering toward targets (e.g., catching mobile prey or landing on small stationary objects). To converge to its target, the bat must reduce to zero the azimuth-to-target angle, θ (Fig. 1A). Reducing θ requires careful, well-coordinated control of the forces the bat applies to maneuver. Thus, a major challenge the bat faces is to translate the noisy estimates of θ into motor commands.
A few previous studies tested how bats control their approach [9,10]. These studies focused either on the flight pattern [10–13] or on the echolocation dynamics [9,10,14], but they did not examine the sensorimotor loop [15]—that is, how bats rely on sensory input to control flight motor commands. Here, we aimed to examine: (1) If it is enough for the bat to measure the azimuth-to-target (θ), or does it need additional information in order to apply the correct motor commands to converge onto the target? Specifically, it is well known from control theory that noisy measurements can pose a great challenge in fine guidance tasks, necessitating additional measurements, such as derivatives of θ. (2) Does the bat change its control strategy under different sensory conditions which differ in their noise level, such as in dark versus light, and moreover, do environments with reduced sensory noise affect the motor commands and accordingly the flight? To answer these questions, we developed a control-theory based sensorimotor model that receives noisy sensory input and computes the needed forces based on basic physical principles. This model allowed us to test different control strategies to evaluate the bat’s sensorimotor strategy and to examine its control strategy under different sensory conditions (light versus dark). To test our model, we used behavioral data from Egyptian fruit bats (Rousettus aegyptiacus)—flying mammals that possess an advanced biosonar (echolocation) system [16–21], as well as an excellent visual system [22].
We found that a simple model, which considers only the angle-to-target and its derivative (a “proportional-derivative controller”), was able to reconstruct complex, several-meter-long flight trajectories very accurately—with an average error of only 14.6 cm. We show that, to successfully land in the dark, the bat has to integrate (to average) sensory data from its biosonar system over several hundred milliseconds, in order to overcome sensory noise. We also demonstrate that when flying in light and using vision, bats reduced their flight curvature. This change in movement pattern was fully explained by the model as resulting from the reduced sensory noise when using vision. Taken together, these results imply an important link between sensory noise and movement dynamics in a rapidly moving mammal.
Bats were trained to land on a 10 cm diameter sphere placed at a random position in a 6 × 7 m large room. We started off with experiments in complete darkness, in which the bats had to rely exclusively on biosonar to estimate the angle-to-target θ using acoustic cues extracted from the reflected echoes [14,17,22,23]. In complete darkness, Egyptian fruit bats emit biosonar pulse-pairs at a rate of 7–10 Hz [17,20], and hence in the model we assumed a 10 Hz sensory acquisition rate. Interestingly, bats often did not fly to the target in a direct line, but rather exhibited a curved trajectory, sometimes even circling around the target (Fig. 1B-D, black dashed curves; a total of 222 trials were analyzed). We aimed to elucidate this behavior. Our model received as input only the initial conditions of the real bat: initial flight angle θ0, initial velocity ν0, the initial position of the bat and the target, as well as a fixed controller parameter determined from a training set that contained 80% of all the trials for each bat (see details in the text below and in the Materials and Methods). The model then simulated the full flight trajectory according to a sensorimotor control model and Newton’s laws of motion, until landing on the target (or missing it). The simulated bat began its flight at the real bat’s origin and with the same direction and speed as the real bat. From this instant onwards, at every sensory update (see below) the simulated bat re-estimated the direction and distance of the target (relative to the bat’s new position) and updated its next motor commands accordingly. In our model, the distance to the target was only used to slow down the bat as it approached landing and not to guide it. We assumed that distance is estimated via bio-sonar. The main objective of the sensorimotor control model was to translate the sensory azimuth measurements (direction to target) into motor commands. We tested several paradigms to perform this translation, comparing them to the bats’ behavior (see below).
We simulated several control strategies with increasing complexity; in all models, turning forces F⊥ were applied perpendicularly to the movement trajectory in order to turn (maneuver) the bat based on the sensory input.
We first tested the simplest possible control strategy, the so-called “proportional control” [24–26], in which the turning forces F⊥ at every moment are proportional to the measured angle θ:
F
⊥
=−
k
p
⋅
θ
where kp is a constant, referred to as the “proportional gain” (Materials and Methods). Such a simple proportional controller performed poorly: First, it converged onto the target in only 61% of the complete-darkness trials (136/222). Second, it converged successfully only in trials with “easy” initial conditions, i.e., small initial angle θ0; 88% of the successful trials (120/136) had θ0 < 45° (see, for example, red curve in Fig. 1B). The convergence in these trials is not surprising because the bat started flying already heading towards the target, and thus small corrections were sufficient to reach the target. Third, in the difficult cases with θ0 > 45°, the model failed in 73% of the trials (θ0 > 45° occurred in 59/222 trials, and in 43/59 or 73% of these difficult trials, the model failed to converge; Fig. 1E, compare empty versus full bars to the right of the dashed line [θ0 > 45°]; c.f. Fig. 1C). Failures to converge occurred mostly because the forces applied during flight were too large, leading to large angular velocity and oscillations of θ that did not converge to zero (see S1 Fig., middle row). Thus, a simple proportional controller is not adequate to explain bat guidance of flight.
The failure of the proportional controller implies a need for a stabilizing effect, to counteract these strong forces. A simple stabilizer commonly used in control theory is to add a derivative term, which dampens the overshooting effect of the proportional controller; such a control strategy is called a “proportional-derivative (PD) controller” [24–27]:
F
⊥
=−
k
p
⋅θ−
k
d
⋅dθ/dt
The PD controller proved to be dramatically superior as compared to the simpler proportional controller: the PD controller led to convergence in 100% of the trials (222/222, see examples in Fig. 1B-D, blue lines); it converged successfully even in those trials in which the bat maneuvered strongly (Fig. 1C,D: compare the red lines [proportional controller] that diverged versus the blue lines [PD controller] that converged accurately; see more examples in S2–S7 Figs.). Notably, the PD controller was able to explain even extreme-maneuvering trials in which the bat exhibited a full 360° rotation (Fig. 1D). For each trial we found the model parameters (kp and kd ) that minimized the error between the simulated trajectory and real bat trajectory (see Materials and Methods for error-index definition). We computed the error index for the proportional controller and for the PD controller, for those 136 trials in which the proportional controller converged. The analysis revealed that in 92% of the cases (125/136), the PD controller led to trajectories with a smaller error (Fig. 1F, blue bars). The average reconstruction error of the PD controller over the testing data was extremely small—only 14.6 ± 1.1 cm (mean ± S.E.M.; see Materials and Methods for details on the training and testing procedure; see also S1 Table for data on each individual bat). In contrast, the proportional controller did not converge at all in 39% of the trials (86/222; Fig. 1F, gray bar), and in those trials where it did converge, its trajectory error was, on average, 3.6 times worse than in the PD controller. In some cases (12.5%, 17/136), the PD controller outperformed the proportional controller by more than 10-fold.
Importantly, when we plotted the distribution of the parameters (kp, kd) that best fitted each experiment, we found that bats used one consistent control strategy (Fig. 1G, note the uni-modal distribution of best parameters for each individual bat, peaking at around kd ≈ 2–6 and kp ≈ 1–5). When examining different individual bats, each bat had slightly different optimal parameters (Fig. 1G and S2 Table), but they were all centered within the same limited range in the parameter space. This consistency between and within bats, all of which exhibit high (non-zero) values of optimal kd, argues against the proportional model—and emphasizes the need for differentiating the angle, as implemented in the PD model.
Finally, we tested two alternative two-parameter models, the proportional-integrative (PI) model and the derivative-integrative (DI) model. Both of these models performed worse than the PD model, and one of them performed worse than the Proportional model as well (S8 Fig). We therefore conclude that the superior performance of the PD controller was specifically due to the combination of the proportional and derivative terms (P+D). This combination is particularly effective during rapid maneuvering because the proportional term corrects the error and the derivative term stabilizes the controller.
All the models we discussed so far were deterministic and assumed that the bat can perfectly measure the angle-to-target. In reality, however, all organisms have sensory errors. To assess the effect of this sensory noise, we tested two models of angle-dependent additive Gaussian noise, which mimic the sensory errors found in the auditory system of several vertebrates (S9 Fig. and Materials and Methods) [28–30]. As expected, sensory noise had strong implications for the convergence of the model: in many trials, adding the noise resulted in increased maneuvering errors (Fig. 2A, top, red dashed curves), and oftentimes led to complete failure to converge (Fig. 2B-C, top). The noise impaired convergence in 85% of the 222 trials (189/222). The percent of diverging simulations increased in trials with highly curved turns, as expressed by their small straightness index and large maximal angle θmax (Fig. 2D, see Materials and Methods for definitions of these indices). Even in the trials that did converge despite the noise, the error index increased substantially, on average by more than 2-fold.
The model failed to converge in the presence of noise because the derivative term of the PD controller differentiates noisy measurements, which further amplifies the noise (see Materials and Methods). We therefore hypothesized that the bat must use a noise-suppression strategy and integrate (average) over several sensory measurements to overcome the noise. To elucidate the number of measurements of θ that should be averaged before executing the motor commands, we tested several different integration functions and evaluated their effect on the percent of successful convergences and on the error index. An exponentially decaying integrator (Materials and Methods), which only takes into account the three to four most recent measurements, was found to outperform uniform and linearly decaying integrators (S10 Fig.). This simple exponential integrator exhibited successful noise suppression and reproduced the bat’s flight trajectories with high fidelity (Fig. 2E, arrow; see also Fig. 2A-C, bottom—compare to the failed convergences in the top panels when not using noise suppression). Interestingly, there is a biologically-plausible update algorithm to implement such an exponential integrator, by only requiring the bat to remember the last average θ (see Materials and Methods). Previous work has shown that in stationary micro-bats (Eptesicus fuscus) that perform a detection task, integrating information over five to ten sonar calls substantially improves performance [31]. Our current findings strengthen the need for integration. The shorter integration window that we found in flying bats (three to four sonar pulses; note the optimum in Fig. 2E: arrow) versus stationary bats (five to ten calls [31]) might reflect the need for faster decision-making during flight maneuvers. It might also result from the difference between the bat species used in the two studies, which rely on echolocation to a different extent (Rousettus less than Eptesicus).
An important prediction of our model is that sensory noise determines motor performance. In particular, the model predicts that using an alternative sensory system with less noise would allow the bat to apply stronger turning forces. In the highly visual Egyptian fruit bat [22], visual estimation of θ is less noisy than echolocation-based estimates. This is due to two factors: First, visual angular acuity in these bats is 0.3° [22], 10-fold better than their echolocation-based acuity of 2°–3° [17]; and second, vision has a much higher effective update rate (at least 25 Hz flicker-fusion limit in mammalian vision [32] versus 7–10 Hz echolocation rate). We therefore hypothesized that, because the effective sensory noise is substantially reduced under vision, bats performing a visually guided flight will be able to use stronger maneuvering forces F⊥, which will be expressed by higher gain parameters (kp, kd ) in our model.
To test this hypothesis, we conducted new experiments in which the same individual bats (five of the six bats) flew under a light level that is considered optimal for bat vision (1 lux) [33], while performing the same landing task. We examined here all of these 43 light-based trials, and found that the PD controller used for modeling the dark conditions (Figs. 1 and 2) was inadequate in explaining the light trials (the mean error was 2-fold higher, see Materials and Methods). Therefore, we implemented a new version of the PD controller, where we increased the sensory update-rate from 10 Hz (sonar) to 25 Hz (vision [32]) and also decreased the noise level [22] (see Materials and Methods). This PD model was able to reproduce the trajectories flown in the light with a high fidelity (Fig. 3A, top row, green curves). We found that, as we hypothesized, the simulated bat exhibited significantly larger gain parameters (kp, kd ) in light versus dark (Fig. 3B; 8/10 parameters increased in the five bats; Wilcoxon signed rank test, p < 0.01). Further, the simulated bat exerted significantly stronger forces when flying in the light (Fig. 3C; t-test, comparing flights with θ0 > 45°: p < 10−11 see also Fig. 3D). Moreover, flight trajectories in light conditions were more direct than in darkness, as quantified by their higher straightness index (Fig. 3E, t-test: p < 0.03). Additionally, we found that using an appropriate integration window is even more crucial in light than in dark (the optimum is more pronounced, see Fig. 3F, arrow).
We next tested if the light-based gain parameters could explain trajectories flown in the dark, and vice versa. When we applied on flights performed in light the kp, kd parameters derived from dark trials, the simulations produced trajectories that converged onto the target, but were much less direct (Fig. 3A, top row, blue curves). Conversely, when we applied on dark trials the kp, kd parameters from the light condition, the simulations produced straighter trajectories (Fig. 3A, bottom row, green curves; and Materials and Methods), consistent with the stronger forces applied by the bats in the light. Interestingly, this implies that, in terms of its motor abilities, the bat could have easily flown more directly to the target in the dark—but instead, it flew in a curved manner because of sensory limitations (namely, the higher sensory noise and lower sensory update-rate in the dark).
Taken together, our results demonstrate a simple, biologically-plausible, sensorimotor model that explains flight guidance in bats—that is, how they maneuver towards a desired target in the presence of sensory noise. In our model we concentrate on translating noisy sensory input into motor commands. We ignore noise in the motor and proprioception systems [34], both of which affect the bat’s flight performance in reality. Such noise should be similar in light and in dark and thus will not affect any of our results (aside from making the flight trajectories less smooth).
Our model relies on two key components. The first is (i) integrating (averaging) sensory information over time, in order to acquire a better (less noisy) estimation of the angle-to-target (θ). Importantly, we have found that the bat could apply a very simple biologically-plausible implementation of such an estimator (or integrator), which uses minimal memory and only requires memorizing one previous sensory estimation of θ (see Materials and Methods). And the model relies on (ii) the derivative of θ in addition to θ itself, in order to stabilize the controller. These two components suggest strong limitations on the processing that must be implemented in the bat’s brain, or in the brain of other rapidly moving animals.
We found that two main factors explain why bats fly in curved trajectories towards the target: (i) The initial angle θ0: when the bat starts flying at a non-zero initial angle θ0 (e.g., Fig. 1C), it flies in a curved manner, while if the bat takes off from the wall with a near-zero θ0, it flies rather straight to the target (Fig. 1B). This is especially true in the dark (when sensory noise is higher) and the bat does not apply strong forces to decrease its angle-to-target. (ii) The gain parameters kp, kd of the PD controller: low gain parameters are used when the sensory noise is higher (e.g., in the dark), leading to curved flights; while high gain parameters lead to straighter flights.
Finally, the experiments in the light suggest that, when using vision, bats can exert stronger motor command (stronger forces) and thus fly to the target more directly than when using biosonar. In contrast, in the dark, when there is higher sensory noise and lower sensory update-rate, the bats exert weaker forces—and therefore the angle-to-target converges more slowly. Taken together, these results imply the surprising conclusion that the highly curved flight trajectories often exhibited by bats in the dark are due to sensory limitations—not motor limitations.
All experimental procedures were approved by the Institutional Animal Care and Use Committees of the Weizmann Institute of Science and the University of Maryland, where these experiments were performed.
The full behavioral methods for the dark experiments were described elsewhere [17]. In brief: Six adult Egyptian fruit bats (Rousettus aegyptiacus) were trained to detect, localize and approach a polystyrene sphere (10 cm diameter) that was mounted on a vertical pole positioned inside a large anechoic flight-room (6.4 × 6.4 × 2.7 m). The target’s size resembles the size of some fruits that are commonly eaten by these bats in nature, such as mango. To exclude the possibility of using vision, the target was painted black and the room was in complete darkness (illuminance < 10−4 lux). To prevent use of olfaction, the bats were food-rewarded (with a piece of fruit) only after landing on the target. After every trial, the target was randomly repositioned inside the room, both in the horizontal and in the vertical planes. A total of 253 trials were collected with at least 30 trials per bat. Because our model only addresses azimuth measurements, θ, we excluded all flights in which there was a change of more than 15% in the bat’s elevation (height) over the entire flight. This resulted in 222 trials of darkness flights, at least 20 trials per bat, which were analyzed here. To test the effect of a sensory system with lower noise (vision), additional 43 new trials were conducted in light conditions, using 5 of the 6 individual bats that were tested in the dark. We used a relatively dim light (illuminance = 1 lux), which is considered ideal for bat vision [33], and which is equivalent to light levels under bright full moonlight.
The model consists of two controllers (one angular controller for turning the bat and one thrust controller for forward movement), and a set of differential equations describing the bat’s motion. We used a fixed sensory update rate of 10Hz for echolocation (i.e., 100 ms steps) or 25Hz for vision (i.e. 40 ms steps). At each sensory update step, the model uses the angle-to-target θ, either deterministically or with additive noise (see main text). The noise was added to the value of θ, which was based on a single estimation or on an average of several estimations depending on the exact model (main text). The model then uses this θ as input for the first controller—the angular controller, which calculates the turning force according to one of the control methods described in the main text and below (proportional controller or proportional-derivative [PD] controller). This turning force is then used by the second controller which calculates the forward thrust. The two forces (turning and thrust) are then integrated in the motion equations to calculate the angle-to-target, θ, at the next time step. In between two updates, the simulated bat continues to move with the same forces as used in the most recent calculation.
We calculate the line-of-sight (LOS) distance to target r, the velocity vector ν, the angle θ between w and the LOS, and the angular velocity ω (see S11 Fig.), by solving the following set of differential equations (the motion equations)
r
˙
(t)
=v(t)cosθ
(t)
(1)
v
˙
(t)=−
drag−
angular damping force+
thrust
(2)
θ
˙
(t)=
ω(t)
(3)
ω
˙
(t)=
U
θ
(t)
(4)
where the operator Ẋ represents the derivative with respect to time, d(X)/dt. The forces drag, angular damping, thrust, and the turning force Uθ, which is calculated by the angular controller, are explained below. Additionally, we defined θmax as the maximum angle θ along the flight t > 0.
The acceleration ν̇(t) can be described by the following ordinary differential equation (ODE):
v
˙
(t)=
−drag−
angular damping force+thrust
=−
D
f
v(t)−
D
t
∥
U
θ
(t)∥+
U
v
(t)
(5)
where drag is proportional to the velocity ν(t), and Df is the drag coefficient [11]; and the angular damping force is proportional to the turning force Uθ(t) (i.e., when the bat turns, it slows down), with a coefficient Dt.
The forward thrust of the bat Uν(t) is proportional to the wing beat and is described by
U
v
(t)=
Fsin(2πβt)
(6)
where β is the flapping frequency of the wings, which we set to 10 Hz according to the behavior of the real bats [11], and F is the forward thrust of each wing flap, according to the data from [11]. In addition to the azimuth θ(t), we assume that the bat estimates the distance to the target (for instance, by using bio-sonar). The distance in our model is only used to slow down the approaching velocity when the bat reaches a critical minimum distance to the target. We estimated from our observations this distance to be approximately 0.5 m. Slowing down is modelled by replacing Equation 6 with the damping term Uν(t) = −1.1ν(t).
In order to compare the simulated bat’s motion to our experimental results, we transformed the bat’s egocentric polar coordinates (r,θ) into the Earth Cartesian coordinates x,y using the linear transformation θx = θ + α, where θx is the angle between the vector of flight and the x-axis (see S11 Fig.).
This allows computing the velocity components along x and y through the following equations:
v
x
(t)=v(t)
cos(
θ
x
(t)),
v
y
(t)=v(t)
sin(
θ
x
(t))
(7)
α(t)
=atan2(y(t),
x(t))
,
−
π≤α(t)≤π
(8)
where the coordinates x(t) and x(t) are calculated by integration of νx and νy at every instance (every 1 ms), given the initial conditions x(0), y(0), ν(0) and θ(0), and the function atan2 (y,x) is the Matlab (Mathworks) function for arctan(y/x ), defined in [−π,π]. In the main text we denoted θ(0) by the shortened notation θ0.
Using the linear transformation above, we can then calculate the angular velocity and the angular acceleration as
θ
˙
(t)=
ω
x
(t)−
α
˙
(t)
(9)
ω
˙
(t)=
ω
˙
x
(t)−
α
¨
(t)
(10)
where α̇(t) and
α
¨
(t)
are the first and second time derivatives of α(t) (dα(t)/dt) and d2α(t)/dt2, respectively). Combining Eq. 3 and 9 leads to the set of differential equations
θ
˙
x
(t)=
ω
x
(t)
(11)
ω
˙
x
(t)=
U
θ
(t)+
α
¨
(t)
(12)
In order to compute α̇(t) and
α
¨
(t)
, we first define the relative velocity components in Earth coordinates to be
V
TBx
=
V
Tx
−
V
Bx
V
TBy
=
V
Ty
−
V
By
where VT is the velocity of the target; in our case VT = 0, since we tested only stationary target. VB = ν(t) is the velocity of the bat, and VTB is the relative velocity (target-bat). The components of the velocity vectors can be calculated using VBx = νx(t) and VBy = νy(t) found in Eq. 7. Then the expression for change in the line of sight, α̇(t), is obtained by [27]:
α
˙
(t)=
x(t)
V
B
y
−y(t)
V
B
x
x
(t)
2
+y
(t)
2
(13)
and
α
¨
(t)
is obtained by differentiation of Eq. 13 using the chain rule.
Control theory offers many solutions for how to calculate the amount of forces needed in order to achieve a desired response with respect to performance and stability. In the case of the bat, the objective is to reduce the angle θ between the line of flight and the LOS, by applying turning forces Uθ(t) (in the main text denoted F⊥), thus controlling the angular acceleration of θ. In our case of a stationary target, the control objective is to reduce θ to zero. The simplest control strategy that can achieve this is the so-called proportional control, applying forces that are proportional to the error e(t); in our case of reducing θ(t) to zero, the error e(t) is θ itself. The forces are computed by
U
θ,p
(t)=−
k
p
θ
(14)
where p denotes the proportional controller, and kp is the proportional factor (gain). This angular force Uθ,p(t) is then used in Eq. 4 to calculate
ω
˙
(t)
—the change in angular velocity. A typical problem of this proportional controller is that it does not consider the rate of convergence, and therefore tends to overshoot and often oscillates around the desired angle. In certain cases, the proportional controller may even lead to divergence away from the target (instability). To compensate for this and to stabilize the system, it is common to introduce a derivative element, which takes into account the rate in which the angle θ changes. In the proportional-derivative (PD) controller, the forces are computed as:
U
θ,pd
=−
k
p
θ−
k
d
θ
˙
(15)
where d denotes the derivative term, kd is the derivative gain, and kp is the proportional gain as before. We assume that the bat estimates the angular velocity, for instance by using a simple first order numerical differentiation, such as first order Euler [26], obtaining
θ
˙
(t)=
θ(t)−θ(t−
1)
dt
(16)
where θ(t – 1) is the previous measurement of θ.
Other simple, well-known, two-gain control strategies that we tested were the proportional-integrative (PI) and the derivative-integrative (DI). The PI controller can be written as
U
θ,PI
=−
k
p
θ−
k
i
∫
t
0
t
θdτ
(17)
where the ki is the integral gain. The less commonly used, two-gain parameter DI controller can be written as
To estimate the proportional and derivative gains of the controller (kp and kd, respectively) that best fit the bats’ behavior in the dark, we first simulated 31 values in the parameter space of each gain, ranging from 0 to 16. We validated the model by randomly dividing the dark trials of each bat into 80% training trials and 20% testing trials (80% of the 222 trials were randomly selected with equal probability). We first computed the best mean gain parameters for each bat in the dark using the training set and then tested these parameters on trials from the test set (the remaining 20%). This procedure was repeated 100 times. The average error for the training set was 4.6 cm, while for the test set it was 14.6 cm, suggesting that our model could indeed predict accurately the bats’ behavior. All the errors reported in this article are errors on the testing set. The test values for each bat are provided in S1 Table.
To test whether our model, developed for dark conditions, is accurate for light conditions as well, we simulated the trials conducted in light for each bat using the same model, i.e., 10 Hz update rate of echolocation and the same gain parameters that were used in the dark; we then compared the error index. Few of the trials diverged, and of the trials that did converge, the mean error was 2-folds larger than the error we obtained for the dark trials, especially for large angles (maximum error over 60 cm). We therefore went on to test a different model for the light with a higher update rate (25 Hz) and a different noise model with less noise (see below).
Interestingly, when we used the training procedure in order to optimize the control parameters in the light trials (see the kp and kd values in S2 Table), and then tested these on the dark trials, most of the simulations diverged. Of those simulations that did converge, the error increased from 2.6 cm in the light to 21.50 ±0.6 cm in the dark, supporting our hypothesis that the bat uses a different control model in the light than in dark conditions.
We tested the parameter space kd and kp by simulating the 222 dark trials and 43 light trials, all without noise, over a grid of combinations of kd and kp from 0 to 16. We measured the error index between each simulation and its corresponding real trajectory (see below). The parameter values that minimized the error index are provided in S2 Table.
A trial simulation was stopped when one of the following convergence criteria was satisfied: (1) When the simulated bat hit the target (i.e., reached a 5 cm distance from the center of the 10 cm diameter target). (2) When the simulation lasted more than 7 s. This criterion was set because the average time-to-landing in the real experiments was 1.5 s, and the longest flight duration observed in the 222 trials was less than 4 s.
In order to compare the trajectories of the simulated and the real bat, we first created a joint closed shape that is defined by the two curves (i.e., as illustrated in S12 Fig., the curves that surround the areas A1, A2, and A3). The area inside the closed continuous curve can be calculated by
Since our vectors are in discrete form, the above continuous equation can be approximated to discrete-time form by the following
A
j
=
1
2
[
∑
i=1
m−1
(
x
i
+
x
i+1
)(
y
i+1
−
y
i
)−
∑
i=1
m−1
(
y
i
+
y
i+1
)(
x
i+1
−
x
i
)]
(20)
where m is the number of discrete points in each curve. If the curves intersect, the intersection points are found using linear interpolation between each of the elements composing the curves (S12 Fig.). In those cases, we sum all the closed areas using
∥A∥=
∑
j=1
k−1
A
j
(21)
where k is the number of intersection points (e.g. in S12 Fig., k = 4). The error index is defined by normalizing the total area ∥ A ∥ by the length of the experimental trajectory of the corresponding trial. Simulations in which the bat circled the target prior to landing (see convergence criterion above) were penalized by multiplying the error by a factor of 10, since such a behavior was never observed in the real experiments.
In the dark, we assumed that the bat measures the angle, based on echolocation with a frequency of 10 Hz, i.e., every 0.1 s. We assumed that the noise is θ-dependent [20–22], and we tested two different models. The first noise model has no bias (the measured θ linearly depends on the real θ), and it has a nonlinear noise term that is low at small angles and increases above π / 2 (S9 Fig.—left):
θ
˜
=θ+n
κ
1
sin
2
θ+n
κ
2
(22)
where
θ
˜
is the noisy estimation of θ, κ1 = 0.65, κ2 = 0.15 and n is Gaussian noise (
N(0,1)
).
The second noise model (S9 Fig. - right) has both a bias in the estimation of θ(t) and additive Gaussian noise that is θ-dependent [20–22]:
θ
˜
=asin(θb)+0.4nθ+0.15n
(23)
with the parameters a = 2.22 and b = 0.458, and n is the Gaussian noise. In the light conditions, the simulated bat estimated the angle-to-target θ by using both vision and echolocation (the real bats did not completely stop echolocating at this light level). We assumed that the additional independent visual measurements were acquired at a rate of 25 Hz (see more details in the main text). Additionally, the noise model assumed a measurement of θ, which is approximately 7-fold more accurate at each time step, following previous behavioral and anatomical studies (see main text):
θ
˜
=asin(θb)+0.07nθ+0.02n
(24)
We implemented the exponentially decaying integration function as follows:
θ
¯
=
∑
i=1
N
W
i
⋅
θ
˜
i
∑
i=1
N
W
i
(25)
where
θ
¯
(t)
is the estimated angle θ at the current time
after integration (averaging) over several sensory measurements, and
θ
˜
i
is the measured angle θ at sample time i, with noise and without integration. Each of the previous N sensory inputs was integrated with a weight Wi calculated according to Wi = wi–1 / V and V = (1–wN)/(1–w), and w is the decay parameter, which was estimated in our case to be w = 0.5. The window that was found to best explain the behavior was N = 3–4 (Fig. 2E). We also tested a linearly decaying averaging model, and a uniform averaging model, and found the exponentially decaying average to be superior in terms of successful convergence and accuracy (error index); this testing was done by comparing between the experiments trajectories and 120 simulated trajectories with noisy measurements for each trial (S10 Fig.).
The above integration function requires the memory of at least three past measurements, and to perform a relatively complicated mathematical expression at each time step. A much more practical and biologically plausible method to implement the above integration function is to use the discrete form of a low pass filter [26], which updates the current angle estimation using the previous one according to the weight constant γ
θ
¯
k
=
θ
¯
k−1
(
1−γ
)+γ⋅
θ
˜
k
(26)
where k is the time step and
θ
˜
k
is the current noisy measurement of the angle (Eq. 22–24). Using the above equation, the organism needs only to remember the most recent measurement and the value
θ
¯
k−1
estimated at the previous measurement. The above practical implementation is another presentation of Eq. 25. To see that, we apply Eq. 26 recursively:
θ
¯
k
=γ
θ
˜
k
+
(
1−γ
)
γ
θ
˜
k−1
+
(1−γ)
2
γ
θ
˜
k−2
+
⋯
+
(
1
−
γ
)
k
θ
¯
0
=
∑
i=0
k−1
(1−γ)
i
γ
θ
˜
k−i
+
(1−γ)
k
θ
¯
0
and for the choice of γ = 0.5 and assuming
that the first estimation
θ
¯
0
=0
, we have
θ
¯
k
=0.5
θ
˜
k
+0.25
θ
˜
k−1
+0.125
θ
˜
k−2
+0.0625
θ
˜
k−3
+⋯
(27)
which is corresponding to Eq. 25 with a decay parameter w = 0.5. Indeed, the simulation results of the above equation and an exponentially weighted function (Eq. 25) with three past measurements are very similar.
We have shown (S8 Fig.) that the PD controller was superior in performance to the other two-parameter controllers, i.e. the DI and the PI. It was also superior to the single parameter P controller. We will prove here that the derivative term is essential for stabilization of the guidance controller.
Applying the proportional controller (Eq. 14) to the model dynamics (Eq. 12), we get
ω
˙
x
=
θ
¨
x
=−
K
p
(
θ
x
−
θ
d
)+
α
¨
(t)
(28)
=−
K
p
(
θ
x
−α)+
α
¨
(t)
(29)
since θd = 0. Rearranging the above equation, we get
(
θ
¨
x
−
α
¨
)+
K
p
(
θ
x
−α)=0
(30)
The solution for this differential equation is marginally stable [24,25] and the solution oscillates; i.e., it never converges to an isolated steady state, except when the true solution is a straight line (initial angle to target is identically zero). Applying a proportional-derivative controller, (Eq. 15) we get
ω
˙
x
=
θ
¨
x
=−
K
p
(
θ
x
−α)+
K
d
(
θ
˙
x
−
α
˙
)+
α
¨
(31)
and then we get the second order differential equation
(
θ
¨
x
−
α
¨
)+
K
d
(
θ
˙
x
−
α
˙
)+
K
p
(
θ
x
−α)=
(32)
e
¨
(t)+
K
d
e
˙
(t)+
K
p
e(t)=0
(33)
which is an exponentially stable system for any positive parameters Kp ≥ 0 and Kd > 0 [24,25], and thus the error e(t) = θx – α = θ is reduced to zero exponentially with time. This proves that the derivative term is essential for ensuring convergence of the controller—that is, the PD controller is the simplest stable controller for guidance purposes.
|
10.1371/journal.ppat.1001228 | HIV-1 Envelope Subregion Length Variation during Disease Progression | The V3 loop of the HIV-1 Env protein is the primary determinant of viral coreceptor usage, whereas the V1V2 loop region is thought to influence coreceptor binding and participate in shielding of neutralization-sensitive regions of the Env glycoprotein gp120 from antibody responses. The functional properties and antigenicity of V1V2 are influenced by changes in amino acid sequence, sequence length and patterns of N-linked glycosylation. However, how these polymorphisms relate to HIV pathogenesis is not fully understood. We examined 5185 HIV-1 gp120 nucleotide sequence fragments and clinical data from 154 individuals (152 were infected with HIV-1 Subtype B). Sequences were aligned, translated, manually edited and separated into V1V2, C2, V3, C3, V4, C4 and V5 subregions. V1-V5 and subregion lengths were calculated, and potential N-linked glycosylation sites (PNLGS) counted. Loop lengths and PNLGS were examined as a function of time since infection, CD4 count, viral load, and calendar year in cross-sectional and longitudinal analyses. V1V2 length and PNLGS increased significantly through chronic infection before declining in late-stage infection. In cross-sectional analyses, V1V2 length also increased by calendar year between 1984 and 2004 in subjects with early and mid-stage illness. Our observations suggest that there is little selection for loop length at the time of transmission; following infection, HIV-1 adapts to host immune responses through increased V1V2 length and/or addition of carbohydrate moieties at N-linked glycosylation sites. V1V2 shortening during early and late-stage infection may reflect ineffective host immunity. Transmission from donors with chronic illness may have caused the modest increase in V1V2 length observed during the course of the pandemic.
| The HIV envelope gene (env) encodes viral surface proteins (Env) that are vital to the basic processes used by the virus to infect and cause disease in humans. Adaptations in env determine which cells the virus can infect, and permit the virus to avoid elimination by the immune system. Env is one of the most variable genes known, and it can change dramatically over time in a single individual. However, Env-host cell interactions are complex and incompletely understood, and changes in this viral protein during infection have not yet been systematically described. We examined a large number of env sequences from 154 individuals at various stages of HIV infection but who had never received antiretroviral treatment. We found that the env V1V2 region lengthens during chronic infection and becomes more heavily glycosylated. However, these changes partially reverse during late-stage illness, possibly in response to a weakening host immune system. V1V2 lengths are also increasing over time in the epidemic at large, possibly related to the epidemiology of HIV transmission within the subtype B epidemic. These results provide fundamental insights into the biology of HIV.
| The gp120 portion of the HIV-1 envelope protein (Env) mediates attachment prior to fusion with the host cell membrane during target cell infection. gp120 has five hypervariable regions (V1–V5) bounded by cysteine residues and separated by four relatively “constant” regions (C1–C4) [1]–[3]. Gp120 is notable for its sequence variation, which may arise through recombination and point mutation, as well as by insertion and deletion of one or more nucleotides. Insertion and deletion events (indels) occur throughout env but are maintained through positive selection particularly within the hypervariable loops, which thereby may acquire significant length variation [4], The third hypervariable region is known to encode the primary determinants of coreceptor usage specificity [5]–[7], as well as epitopes recognized by humoral [8], [9] and cellular [10], [11] immune responses. V3 loop sequence variation has been extensively studied, and correlated with changes in host cell range, cytopathogenicity, and disease progression [12]–[14].
The V1V2 region in particular is characterized by a high degree of length polymorphism, sequence variation, and predicted N-linked glycosylation sites (PNLGS) [15]–[20], each of which may affect viral attachment, coreceptor usage and recognition by neutralizing antibodies [20], [21]. Comparison of structural models of gp120 and gp120 bound to CD4 and a chemokine coreceptor have yielded considerable insight into the functional roles played by V1V2 and V3 during viral attachment [22], [23]. In the unbound gp120 conformation, the V2 loop partially obscures V3 and other gp120 residues involved in coreceptor binding. Binding to CD4 induces conformational changes that expose the coreceptor binding site on gp120, including residues from V1V2, V3 and other regions [22], [24].
Numerous studies have suggested that sequence variation in V1V2 influences host cell range and/or syncytium-inducing (SI) phenotype [25]–[31]. For example, Toohey demonstrated that recombinant chimeric clones with a V1V2 region from macrophage-tropic HIV-1 strains replicated efficiently in macrophages, whereas clones with the V1V2 region from lymphotropic strains did not [31]. However, not all studies have been concordant on the role of V1V2 in viral replication kinetics, cell range and transmission [15]–[19], [32]. For example, Pastore showed that sequence changes in V1V2 could rescue otherwise lethal mutations in V3 associated with a change in coreceptor usage [33], and V2 polymorphisms have also been linked with restriction to CCR5 coreceptor usage [16]. In contrast, Wang et al found no relationship between SI phenotype and V1V2 sequence, length, distribution of PNLGS or charge [32].
The V1V2 region also appears to be an important determinant of sensitivity to neutralizing antibodies [34]–[38]. The V1V2 region evolves under positive natural selection in vivo [4], [39]–[41], and an inverse relationship between V1–V4 length and neutralization susceptibility has been demonstrated in subtypes A [20], B [34]–[38] and C [42]. Tellingly, laboratory strains lacking V1V2 may still replicate efficiently in vitro, but appear to be especially sensitive to antibody neutralization [43], [44]. Consistent with this observation, viral strains with shorter and less glycosylated V1V4 regions have been reported to preferentially replicate in subjects newly infected with HIV-1 subtype C [45] (where presumably an effective neutralizing antibody response has not had time to emerge), and similar observations have been made concerning the V1V2 loop in individuals recently infected by HIV-1 subtype A [19]. However, we and others have not observed this effect in HIV-1 subtype B [19], [46], [47].
Despite these reports, the relationship between V1V2 region length polymorphism and disease progression remains unclear. In two small longitudinal studies, elongation of V1 and V2 was noted in long-term nonprogressors (LTNP), but not within individuals progressing rapidly to AIDS [15]–[19]. In a third study, no clear relationship between V1V2 length variation and disease progression was observed [48]. Lastly, some investigators postulate that V1V2 length changes positively correlate with the pace of disease progression [16], [19], while others have suggested that V1V2 length increase may be a correlate of delayed progression to AIDS [18].
Thus, our understanding of the role of the V1V2 loop in influencing HIV pathogenesis remains incomplete and is challenged by several contradictory observations. To more fully characterize HIV envelope subregion variability and to clarify the associations between subregion length variation, glycosylation, and disease progression, we have comprehensively examined length and glycosylation of each gp120 subregion as a function of clinical parameters in a large collection HIV-1 subtype B infected individuals.
This study was performed using publicly available data from the Los Alamos database, and previously unpublished experimental data obtained at the University of Washington. Unpublished data were obtained and analyzed with written informed consent of study participants, and approval by the University of Washington Institutional Review Board.
We analyzed new and published HIV-1 envelope gene sequences and associated clinical data from all available subjects in the Seattle Primary Infection Cohort (PIC) [49], the Multicenter AIDS Cohort Study (MACS) [50], and from the Los Alamos National Laboratories HIV database (HIVDB) (http://www.hiv.lanl.gov/content/hiv-db/mainpage.html) not meeting pre-specified exclusion criteria. Subjects were excluded from this study if younger than 18 years of age or if there was any history of antiretroviral therapy prior to sampling as determined by patient report and clinical records (MACS, PIC) or as indicated in the methods section of published reports (HIVDB), unless otherwise noted. All subjects considered in the cross-sectional and longitudinal analyses were infected with HIV-1 subtype B, except for two subjects infected with HIV-1 subtype A who were included in longitudinal analyses, but were excluded from cross-sectional analyses. (Additional subtypes were considered in analyses of env subregion length change during transmission, presented in Text S1, Section 8). Clinical data retrieved included CD4 count, viral load, time since infection, and treatment history. Sequence data were only accepted if directly derived from plasma or PBMC without an intervening step involving viral propagation in vitro. In some cases, individual authors were consulted to resolve clinical or methodological ambiguities. Accession numbers for published sequences are provided in Table S1. Gene sequence data used in this study are available at http://mullinslab.microbiol.washington.edu/publications/curlin_2010/.
Viral gene sequence data were considered in both cross-sectional (Table 1) and longitudinal analyses (Table 2). The cross-sectional dataset included only plasma and PBMC sequences derived from individuals infected with subtype B (see results, and Table 1). Sequences were triaged by author, database identifier and associated clinical data to exclude duplicate entries. To assess the role of stage of illness on loop length variation, subjects were divided into four non-overlapping groups; group Cx1 subjects were sampled within two months of the estimated time of infection. Group Cx2 subjects were sampled between two months and three years following infection. Group Cx3 subjects were sampled at times >3 years post infection. Group Cx4 was comprised of all individuals meeting 1993 CDC criteria for AIDS when sampling occurred (generally CD4 count <200/mm3), regardless of time since infection.
The longitudinal dataset was derived from 20 subjects infected with subtype B and 2 individuals infected with subtype A, from the PIC cohort and from previous reports [18], [51]–[55], in whom data were available from two or more timepoints (see results, and Table 2). All intra-individual longitudinal comparisons were made between sequences obtained from the same compartment (e.g., plasma vs. plasma). Individuals partitioned into group L1 (N = 15) did not meet criteria for AIDS at any time prior to the final sample (median follow-up 3.25 years, range 1 to 20.8 years), whereas subjects in group L2 (N = 7) were reported to have an AIDS-defining illness or peripheral CD4 count <200/mm3 between the first and second samples (median follow-up 2.75 years, range 2 to 4 years).
Sequences from the PIC and MACS cohorts (Tables 1 & 2) were obtained from plasma or PBMC by standard methods [56], [57], using safeguards to prevent contamination and template resampling [58]. Briefly, PCR amplification was performed using Taq polymerase (Bioline) with primers ED3 and BH2 [59] (first round) followed by ED5 and DR7 (second round) [60]. PCR products were cloned into a TA TOPO vector (Invitrogen) and selected colonies sequenced under contract using Big Dye dye-terminator protocols. Genbank accession numbers pending submission.
Deduced amino acid sequences were aligned using ClustalW [61] and divided into seven subregions; V1V2 (HXB2 nucleotide positions 6615–6812), C2 (HXB2 6813-7109), V3 (HXB2 7110–7217), C3 (HXB2 7218–7376), V4 (HXB2 7377–7478), C4 (HXB2 7479–7556), and V5 (HXB2 7557–7637). Alignments were manually edited and subregion lengths were counted using MacClade. PNLGS were counted using NetNGlyc.1 (http://www.cbs.dtu.dk/services/NetNGlyc/). Coreceptor usage (CCR5 vs. CXCR4 tropism) was predicted for all available subtype B V3 loop sequences, using the Position-Specific Substitution Method (PSSM) [62], Geno2pheno [63] and two other machine learning algorithms [64], [65] (hereafter denoted PSSM, G2P, PGRC and BMLC, respectively). For G2P coreceptor usage predictions, we selected the standard 10% false positivity threshold, and PGRC predictions were based on the support vector machine (SVR) user option. Estimated time since infection was calculated for all data entries. When time was reported as time since onset of symptoms or time post seroconversion (SC), symptoms and seroconversion were assumed to occur at 14 days and 42 days after infection, respectively [66], [67]. Date of seroconversion was assumed to occur at the midpoint between most recent negative serological test and first reported positive test, unless additional information was available.
For cross-sectional analyses, univariate and multivariate regressions were conducted assessing subregion lengths and number of glycosylation sites as a function of time since infection, stage of disease, CD4 count, HIV viral load, adjusting for sample source (plasma vs. PBMC), and date of sampling (calendar year). In regression analyses, to allow direct comparisons of the effect of each variable on V1V2 length and/or glycosylation, we compared β values (i.e., regression coefficients scaled such that each variable is equivalent to having a mean value of 0 and a standard deviation of 1). Generalized estimating equations (GEE) were utilized to account for non-independence of data points [68]–[70], and an exchangeable correlation structure was assumed. This method adjusts for the correlation of multiple sequences nested within a sample as well as multiple samples per patient. As an additional means of verifying that analysis outcomes were not influenced by data linkage, regression analyses were performed on replicate data subsets reconstituted from the original data by random resampling, including analyses on 100 data subsets each obtained by using one randomly selected sequence from each individual (See Text S1 section S2). To ensure that results were not unduly influenced by outlying sequences with extremely short or long loop lengths, analyses were repeated after excluding sequences representing the shortest 5% and longest 5% of the V1V2 loops in the dataset. For the longitudinal dataset, multivariate linear regressions were conducted assessing V1V2 length and number of glycosylation sites as a function of time since infection within a person, and the mean rate of change per year was estimated. Statistical analyses were performed using SAS version 9.1 (SAS Institute, Cary, NC).
We obtained 5185 partial length HIV-1 env gene sequences for cross-sectional and longitudinal analysis by the methods described above (Tables 1 & 2). Sequences were isolated from 475 samples obtained from 154 individuals, including 27 from the MACS, 43 from the Seattle PIC and 84 from the HIVDB. Study subjects resided in North America (N = 116), Western Europe (N = 25), East Africa (N = 2), and Asia (N = 11), contributed a median of 14 sequences (range 1–287) and included persons in stages 1 (N = 41), 2 (N = 62), 3 (N = 40), and 4 (N = 27) of infection (note that some subjects contributing to the longitudinal analysis were included at more than one stage of infection). Sequences were derived from plasma (N = 2495), PBMC (N = 2620) and other sites (N = 70). Sequences were of subtype B (N = 5013) and subtype A (N = 172). All subtype A sequences and sequences derived from sites other than blood were excluded from cross-sectional analyses, but were considered as special cases under longitudinal analyses (sequence data available at: *webaddress pending acceptance*).
In the longitudinal dataset, significant V1V2 length increases between first and second timepoints were noted in 10 of 22 subjects, a significant V1V2 length decrease over time occurred in one subject, and no significant V1V2 length changes over time were seen in the remaining 11 subjects. These findings appeared to vary by stage of infection (t-test p = 0.03). In the 15 patients from the L1 group (individuals not meeting AIDS criteria at any time prior to final sampling), the mean increase of V1V2 length per subjects was 1.69 amino acids per year, and 9 subjects experienced significant V1V2 length increases over time (Figures 4 and 5). In contrast, of the seven subjects in the L2 group (individuals progressing to AIDS between first and final sample), the mean V1V2 length decreased by an average of 0.10 amino acids per year, with only one having a significant trend of increasing length, while one individual showed a significant decrease in length (Figure 6). The distribution of V1V2 length change (increase or decrease) by group was therefore asymmetric (Fisher's exact test, p = 0.02), reflecting a trend of increasing length in asymptomatic individuals (group L1) and stable or decreasing length in individuals with AIDS (group L2) (Table 4). Three subjects in group L1 had extensive longitudinal sampling (Figure 5); in 1362 and Q23 [51], there was a period of V1V2 length stability of approximately 2 years, followed by increase through 4.5 years. V1V2 length increase over time was also seen in CC1. In the case of CC1, a pseudotyped virus was created using the gp120 coding region from the initial timepoint from this individual in a HIV-1 NL4-3 background, and cultured in vitro [54]. In contrast to the patterns observed in vivo, V1V2 length and number of glycosylation sites both declined rapidly over 20 generations in vitro (p<0.001).
We have systematically examined gp120 subregion length variation, and the relationship between length polymorphism, N-linked glycosylation sites, and clinical markers of disease progression. Although V1V2, V4 and V5 all displayed remarkable length heterogeneity, and V1V2, C3 and V4 were also quite variable with respect to glycosylation, the most significant associations between virological and clinical variables localized to the V1V2 region. We found that V1V2 length and glycosylation increased significantly over time during chronic infection, and then declined in late-stage illness. In regression analyses, time since infection was the most influential factor in determining V1V2 length. In addition, there was a modest but significant increase in V1V2 length over the period from 1984–2004. V5 loop length was highly variable, but tended to decrease slightly in length over the course of infection.
In SIV infection, the number of PNLGS in gp120 increases over time in vivo following inoculation of a cell-passaged strain [71]. In one earlier study in humans, Bunnik et al noted expansion in gp120 length followed by contraction over time in 4 of 5 individuals receiving antiretroviral therapy, and similar changes in glycosylation in 3 subjects [72]. Others have noted a relationship between early infection and reduced V1V2 length and glycosylation in subtypes C and A [19], [45]. In contrast, a comparison of early and chronic HIV-1 subtype B sequences from the HIV sequence database failed to reveal any significant difference in V1V2 length [19], suggesting that these effects may be subtype-specific. Data on length/glycosylation changes during transmission have been conflicting. Derdeyn et al [45] demonstrated reduced length and glycosylation in V1–V4 following heterosexual transmission in HIV-1 subtype C. However, Frost et al failed to note similar findings in a study of eight subtype B homosexual transmission pairs [47], and in our examination of these and 10 additional subtype B infected homosexual transmission pairs, we found no consistent pattern of change in V1–V2 or V1–V4 length or glycosylation upon transmission [46].
Interpretation of the data presented here may be affected by several methodological factors. There is probably some variation in the accuracy of the reported time of infection for sequences obtained from previous reports. In some cases, sequences obtained from prior publications may have been obtained under conditions permitting template resampling [73], and a systematic error due to evolving laboratory methods could result in bias. Also, in our analyses, we have not formally corrected for multiple comparisons. Physiological factors are also likely to introduce some noise, particularly in cross-sectional analyses of parameters with respect to time since infection. The individuals included here represent a broad spectrum of clinical scenarios, diverse host immune response profiles and varying disease progression rates. Plasma sequences may receive contributions from both recently infected target cells and older reservoirs, and therefore imperfectly reflect selective pressures prevailing at the time of infection. Finally, length and glycosylation phenotypes are likely to be affected by chance events and unknown factors not considered in our analyses. Therefore, the effects we describe are influential rather than deterministic, and reflect important selective forces that can be discerned against a background of high inter-individual variation.
Despite these limitations, the analyses presented here and the work of others [40], [45]–[47], [72] provide the outlines of an overall pattern characterized by transmission of randomly selected V1V2 loop lengths from viruses present in the donor pool, a brief decline in loop size during the initial months immediately following infection, gradual selection for bulkier V1V2 loops during chronic infection, and finally, reversion to more compact loops during late stage illness. Structural studies [22], [23], neutralization studies [20], [34]–[38], [42], and in vitro data on viruses lacking V1 and V2 [43], [44] suggest that one major function of the V1V2 region may be to permit evasion from humoral immune responses in the host. Thus, the trends outlined above support the hypothesis that HIV populations may evolve to escape humoral selective pressure by increasing V1V2 loop size. According to this view, the newly infected, immunologically naïve host might be expected to harbor relatively short V1V2 loops that eventually lengthen in response to an effective humoral response at some fitness cost (Figure S9). Experimental evidence indicating that relaxation of antibody-mediated selective pressure during early infection is associated with shorter loops is provided by Derdeyn, who demonstrated significantly greater neutralization sensitivity among five recipients during early infection, than in the corresponding donors [45]. The decline in V1V2 size observed in advanced disease probably reflects waning effectiveness of humoral immunity in hosts with late-stage illness and profound immune dysregulation (Figure 7). This decline is also congruent with previous findings of an inverse relationship between the rate of HIV genetic evolution and the rate of CD4 T cell decline in some individuals [74]. The dramatic reduction in V1V2 length associated with transfer to the in vitro environment [54] represents the extreme case of absent host immunity, where viruses without an unnecessarily bulky V1V2 loop achieve maximum replicative fitness. As would be expected, the patterns we observe are most pronounced in plasma sequences, which most directly reflect the selective forces present at the time of sampling. In contrast, a significant increase in V1V2 length over time was not seen in the PBMC compartment. These observations are consistent with the presence of archived genotypes from earlier times during the course of infection within the PBMC compartment. We also note that genotypes present in plasma may emanate from other cellular compartments in addition to PBMC, and may therefore reflect somewhat different evolutionary pressures. However, a considerably greater number of V1V2 sequences were derived from plasma, and sample size may also account for some of the differences observed between these compartments.
Our model may help to explain a failure to find any significant difference in V1V2 length in a comparison of early and chronic HIV-1 subtype B sequences (including sequences from late-stage individuals) [19]. When we reanalyzed the data presented by Chohan [19] after separating subjects with stable chronic illness from subjects with AIDS (Figure S13), we observed a pattern of lengthening over time, followed by decline in late-stage illness, as reported here (See Text S1, section S7). Similarly, we may explain discordant results obtained on V1V2 length variation during transmission of HIV-1 subtypes C and B. While a trend towards shorter loops in recipients was seen in subtype C [45] but not B [46], [47], it is likely for methodological reasons that the subjects studied by Derdeyn were sampled at somewhat later times than those of Frost and Liu. Thus the sequences in the latter two studies would be expected to be a random sampling from the donor pool, while those of Derdeyn might reflect the expected shortening prior to the onset of an effective antibody response. Indeed, when we examine a much larger set of subtype A and C transmission pairs from East Africa with more precisely known sampling times obtained soon after transmission, it is difficult to appreciate any consistent pattern of V1V2 length change (See Text S1, section S8 and Figure S14). Thus there may be no need to infer separate mechanisms for different HIV-1 subtypes and modes of transmission.
In addition, we may also explain a trend of increasing V1V2 length by calendar year. If shorter and less glycosylated V1V2 were always selected during transmission, transmission from donors in early infection would maintain a constant V1V2 length within the epidemic, whereas if all new cases were acquired from chronically infected hosts, this increase of V1V2 length by calendar year could be dramatic. However, most studies suggest that about half of transmission events involve subjects in early infection [46], [75], [76], consistent with the moderate trend we observed. Alternatively, the temporal trends we have observed could represent a gradual adaptation by HIV-1 to host the host environment at the population level, a hypothesis that has been proposed by several investigators with respect to mutational escape from HLA-restricted CTL epitopes [77]–[79].
Finally, our results imply that the polymorphisms seen in V1V2 reflect the ability of the host to mount a meaningful immunological response, rather than virologic features that dictate the course of illness. That is, we argue that V1V2 length change is a consequence of environmental selective pressure rather than a causative factor in disease progression.
|
10.1371/journal.pcbi.1007056 | The breadth of HIV-1 neutralizing antibodies depends on the conservation of key sites in their epitopes | Developing HIV-1 vaccines that trigger broadly neutralizing antibodies (bnAbs) is a priority as bnAbs are considered key to elicitation of a protective immune response. To investigate whether the breadth of a neutralizing antibody (nAb) depended on the conservation of its epitope among circulating viruses, we examined Antibody:Envelope (Ab:Env) interactions and worldwide Env diversity. We found that sites corresponding to bnAb epitopes were as variable as other accessible, non-hypervariable Env sites (p = 0.50, Mann-Whitney U-test) with no significant relationship between epitope conservation and neutralization breadth (Spearman’s ρ = -0.44, adjusted p = 0.079). However, when accounting for key sites in the Ab:Env interaction, we showed that the broadest bnAbs targeted more conserved epitopes (Spearman’s ρ = -0.70, adjusted p = 5.0e-5). Neutralization breadth did not stem from the overall conservation of Ab epitopes but depended instead on the conservation of key sites of the Ab:Env interaction, revealing a mechanistic basis for neutralization breadth that could be exploited for vaccine design.
| So far, no HIV-1 vaccine has elicited broadly neutralizing antibodies (bnAbs) in humans. HIV-1, one of the most rapidly evolving pathogens, is remarkable for its high variability across individuals and adaptability within hosts. We tested the relationship between HIV-1 diversity and neutralization breadth. While bnAbs did not specifically target more conserved regions of HIV-1 Env, we found that the broadest bnAbs relied forcibly more on structural interactions at key sites of the Ab:Env interaction than other Abs. Understanding mechanisms underlying neutralization breadth provides guidelines to design more efficacious vaccines and antibody-based therapeutics.
| There is an urgent need for a vaccine against HIV-1. Since HIV-1 shows remarkable diversity, it is assumed that a vaccine should elicit bnAbs to block the most extensive array of HIV-1 strains[1–4]. Neutralizing Abs develop over the course of HIV-1 infection and there is a continuum in the extent of neutralization breadth developed across individuals, with typically half of a cohort being able to neutralize about half of a virus panel [5]. A number of studies have focused on the fraction of individuals who can develop bnAbs that can neutralize a majority of the viruses in a panel[6–9]. These bnAbs arise after Ab lineages have matured typically over multiple years[5, 10–13]. Highlighting the many paths that can lead to bnAb development, bnAbs have been isolated from individuals with different HIV-1 subtypes and presenting different clinical disease progression profiles. For example, VRC01 was isolated from patient 45, an African-American male who had been infected with HIV-1 subtype B for 11 years at the time of Ab isolation; he was considered a long term non-progressor as his viremia was maintained around 10,000 copies/ml [14]. In contrast, the bnAb CH103 was isolated from patient CH505, a male from Malawi who had been infected with HIV-1 subtype C for 2.5 years when Ab were isolated; patient CH505 was followed for six years and maintained a high median viral load of 173,667 over that time [15]. BnAbs recognize exposed regions of the Env trimer and tend to target five sites: the V1V2-glycan site (e.g. PG9), the V3-glycan site (e.g. PGT128), the CD4 binding site (e.g. VRC01), the gp120-gp41 interface (e.g. 8ANC195, 35O22) and the membrane proximal region of Env-gp41 (e.g. 10E8) [16].
It is generally believed that bnAbs target conserved epitopes on HIV-1 Env trimers [17–20]. Yet, no study has systematically quantified the relationship between the neutralization breadth of bnAbs and the conservation of their respective epitopes on Env. Here we analyzed publicly-available Ab:Env complex structures and characterized how the neutralization breadth of an Ab was influenced by the conservation of its epitope. We describe how neutralization breadth was positively associated with Env epitope conservation only when the epitope conservation was defined by taking into account the strength of the Ab:epitope interaction, i.e. specifically weighting structurally important sites, and not simply sequence conservation.
To characterize the diversity of HIV-1 circulating strains, we created an HIV-1 group M alignment of 239 Env sequences that reflected the global representation of HIV-1 subtypes (gp M), as well as specific datasets for subtype A1 (n = 203), B (n = 1035), C (n = 1184), D (n = 116) and CRF01_AE (n = 577). We analyzed 34 Abs for which neutralization breadth had been measured using a panel of 136 viruses and ranged between 31 and 97% [21] and for which Ab:Env complex structures were available (S1 Table, S1 Fig). For these 34 Abs, the epitope consisted of 8 to 36 sites. We looked at the diversity among group M sequences at each accessible, non-hypervariable site on the surface of Env and found no difference between sites that belonged to Ab epitopes (n = 31 Abs, epitopes of 3 MPER antibodies were excluded as they are partially/totally missing from the Env structure 5FYJ, which was used to define surface sites) and sites that were outside of epitopes: median Shannon entropy = 0.47 vs. 0.32 bits, respectively (p-value = 0.50, Mann-Whitney U-test) (Fig 1). This result was confirmed when the analysis was restricted to the 15 antibodies that showed over 70% breadth: median Shannon entropy = 0.44 vs. 0.32 bits, respectively (p-value = 0.47, Mann-Whitney U-test) (Fig 1B). Thus, bnAbs targeted Env sites that were as variable as other accessible Env sites. To define the epitope diversity, we summed the Shannon entropy of all the epitope sites and adjusted with the mutual entropy of neighbor pairs of sites. This corresponds to the diversity of the whole epitope patch, which we then normalized based on the size of the epitope. There was no relationship between the epitope diversity (or conservation) and the breadth of neutralization of HIV-1 strains by nAbs if we considered the epitope diversity of the whole epitope (Spearman’s ρ = -0.20, adjusted p-value = 0.70) or when the epitope diversity was normalized based on the size of the epitope (Spearman’s ρ = -0.44, adjusted p-value = 0.079; Fig 2A) when using the set of sequences representative of the global HIV-1 distribution. Results were similar when we tested the sequence sets corresponding to different subtypes/CRF: Spearman correlation coefficient ρ ranged between -0.31 and -0.43 with adjusted p-values > 0.078, showing that the finding was not dependent on the dataset tested (group M versus subtype-specific alignments) (Fig 3).
We modified our definition of epitope diversity to integrate structural factors beyond sequence conservation that may be critical to efficient Ab neutralization. These epitope diversity measures weighted each site in the Env epitope by i) the number of Ab atom contacts for each Env epitope residue, ii) the number of atom pairs in contact between the epitope and the Ab, iii) the number of neighboring Ab residues for a given epitope residue in the Ab:Env complex and iv) the reduced accessible surface area after Ab binding. These definitions were also normalized for the size of the epitope (8–36 sites). The results reported below correspond to the nine epitope sites that had the highest number of neighboring Ab residues (we focused on nine sites as it showed the best correlation when comparing between seven and eleven key sites, S2 Fig). Fig 3 shows the effect of weighting by the features described above, we found a negative relationship between epitope diversity and neutralization breadth: the more conserved the structurally weighted Env epitope, the greater the neutralization breadth. If we consider the alignment of 239 representative group M sequences, the Spearman correlation coefficient ρ was -0.70 (adjusted p-value = 5.0e-5) (Fig 2B, S2 Table). The negative relationship between the Env epitope diversity and neutralization breadth was replicated when restricting the analysis to specific subtypes/CRF: considering the top nine target sites, the Spearman correlation coefficient ρ ranged between -0.57 and -0.73 with adjusted p-values ≤ 6.1e-3. We note that the relationship was similar when we considered only the dataset corresponding to the subtype matching the subtype of the infected individual from whom the Ab was obtained (Spearman’s ρ = -0.70, adjusted p-value = 1.7e-3; S3 Fig and S4 Fig). Since most of the 34 Abs targeted the CD4 binding site (CD4bs) (n = 21), we analyzed data separately for these Abs. We found that the relationship between epitope diversity and neutralization breadth was largely driven by CD4bs Abs (Spearman’s ρ ranged between -0.61 and -0.73 with adjusted p-values ≤ 3.1e-3 when analyzing the different sequence sets, S3 Table). However, for the 13 other Abs, the relationship was not improved by the structural-weighting (Spearman’s ρ ranged between -0.39 and -0.66, 0.18 ≤ adjusted p-values ≤ 1.00, S3 Table; the lack of significance may be due to the small sample size).
The fact that epitope conservation failed to strictly derive from sequence conservation but corresponded to a structurally-weighted conservation measure is illustrated by the comparison of the Env epitopes of 3BNC117 and VRC03, two Abs that target the CD4 binding site. The bnAb 3BNC117 neutralizes 82% of HIV-1 strains while VRC03 neutralizes 48% of HIV-1 strains. Because the Env epitopes of 3BNC117 and VRC03 are very similar, the unweighted measure of diversity gave similar diversity values. However, when we considered only the structurally key epitope sites for Ab binding, 3BNC117 engaged conserved sites while VRC03 had many atom contacts with more variable sites such as amino acids 460 and 461 in Env-V5. Hence, the VRC03 epitope had a higher diversity value for its top nine sites than the 3BNC117 epitope and was associated with a more limited breadth of neutralization, highlighting that the mode of interaction between the Ab and epitope and not just the location of the epitope was associated with increased neutralization coverage (Fig 2).
We obtained similar results when we used as a measure of epitope conservation the similarity of a given epitope to its counterpart in the three strains experimentally-defined as the most susceptible to neutralization. The three most susceptible strains were selected using the data from the panel of 136 viruses assayed by Doria-Rose and colleagues [21] (up to five strains were tested before choosing a combination of three strains, S5 Fig). For each Ab, we calculated the fraction of viruses with epitopes matching the epitope in the three most susceptible strains. When we considered whole epitopes, the nAbs that showed a higher fraction of epitopes similar to the three most susceptible strains were associated with increased neutralization breadth (Spearman’s ρ = 0.60, adjusted p = 1.9e-3). As seen above, this relationship was stronger when we focused on the top nine epitope sites (ranked by the number of neighboring Ab residues) in the Ab:Env interaction: Spearman’s ρ = 0.80, adjusted p = 2.7e-7 (Fig 4). Similar to what we noted above, the Spearman’s correlation coefficient was stronger for Abs that targeted the CD4bs (Spearman’s ρ = 0.78, adjusted p = 5.2e-4, S4 Table) than for Abs that targeted other epitopes (Spearman’s ρ = 0.73, adjusted p = 0.070, S4 Table).
One shortcoming of this study is that Env accessibility was measured using Env structures with glycans removed, while we know that HIV-1 Env trimers are covered by a glycan shield of ∼90 N-linked oligosaccharides constituting about half of the Env mass—a key factor in HIV-1 evading humoral immunity[22–24]. We weighted the epitope diversity calculations (described above) for the presence of glycans at specific epitope sites, yet this modification did not reveal any significant difference in the relationship between epitope diversity and neutralization (Fig 4). We compared the 9-, 7- and 5-mannose models sampled in molecular dynamics simulations [25] and found that the median Ab accessibility was diminished by about three-fold when glycans were integrated: 13.6 Å in the 9-mannose model vs. 4.6 Å in the absence of glycans. The 7- and 5- mannose models provided better Ab accessibility than the 9-mannose model (median depth of 11.1 Å, 12.2 Å and 13.6 Å for 5-, 7 and 9-mannose Env models, respectively), indicating that engineering the glycan shield to have only 5 or 7 mannose residues at each glycan site may improve Ab accessibility. This is consistent with the finding that restricting the glycan site to be 5 mannose greatly increased its susceptibility to an array of bnAbs[26].
In summary, we systematically analyzed the interaction between nAbs and their corresponding Env epitopes to identify the mechanistic basis of HIV-1 neutralization breadth. Surprisingly, although it is widely accepted that HIV-1 bnAbs target conserved segment of HIV-1 Env, we demonstrated that bnAbs targeted Env sites that were no more conserved than other accessible, non-hypervariable Env sites and that the breadth of a nAb was not significantly related to the conservation of its epitope among circulating viruses if we used a standard measure of epitope conservation. It is only when HIV-1 conservation was measured by accounting for the structural strength of the Ab:Env interaction that we found a positive relationship between sequence conservation and neutralization breadth. We note, however, that certain factors complicated our analysis. For example, it is difficult to account for the influence of the glycan shield or for the missingness of some structural information such as for MPER antibodies where only a fragment of Env is complexed with the antibody. Future studies will also be needed to evaluate whether non-neutralizing antibodies differ from bnAbs in their mode of interaction with Env.
Our finding has implications for vaccine development. We showed that epitopes were not more conserved than any other non-hypervariable sites at the surface of the prefusion-closed Env trimer, yet the broadest bnAbs showed key interactions at very conserved sites. Our study indicates that targeting conserved epitopes is necessary but not sufficient to promote breadth and that an antibody’s interactions with key conserved sites are primordial to achieve neutralization breadth. This would suggest that targeting conserved epitopes may not be sufficient if there is no further Ab maturation to focus on key epitope sites. Hence, this work suggests Ab breadth could be improved by designing epitope patches on immunogens to favor breadth-promoting Ab interactions. In addition, this knowledge can directly be used to characterize breakthrough and rebound viruses following bnAb-based interventions, providing a novel blueprint to prospectively interpret results of clinical trials that use bnAbs as a therapeutic agent.
An Ab epitope was defined based on the Ab:Env complex structure as the Env sites with heavy atoms (called atoms below and in the main text) that were located within 4 Å of the Ab. Weights were assigned to specific Env sites in different ways, using: i) the number of Ab atoms contacted (w.natom), ii) the number of atom pairs in contact between the epitope and the Ab (w.npairs), iii) the number of neighboring Ab residues in the Delaunay tetrahedralization of Cβ (Cα of Gly) atoms of the Ab:Env complex (w.nnbs), and iv) the reduced accessible surface area after Ab binding (w.asa) [31]. The Delaunay tetrahedralization was obtained using Quickhull [32] and edges longer than 8.5 Å were removed. If multiple complex structures were available for an Ab, the mean weight from the different complex structures was used as the final weight of the epitope. To avoid overweighting, the weight was capped at the 98th percentile of all sites; any site with a weight above the 98th percentile was set to the 98th percentile. The weight of N-linked glycans was scaled such that its 98th percentile was equal to the 98th percentile of the amino acids before adding to the corresponding asparagine.
We defined the epitope diversity as the Shannon entropy [33] of the epitope:
H=∑iHi−∑i,jIi,j
(1)
Hi=−∑kp(k)log2p(k)
(2)
Ii,j=∑m,npi,j(m,n)log2pi,j(m,n)pi(m)∙pj(n)
(3)
where Hi is the Shannon entropy of epitope site i; Ii,j is the mutual entropy between a pair of neighbor sites (i,j); p(k) is the fraction of amino acid k on a site; pi,j (m,n) is the fraction of amino acid combinations (m,n) on a pair of neighbor sites (i,j) (m on site i and n on site j). The summation is over all epitope sites in the first term and over all neighbor sites in the second term of Eq 1. In Eq 2, the summation is over all amino acids. In Eq 3, the summation is over all amino acid combinations on neighbor sites i and j. All neighbor pairs were identified by Delaunay tetrahedralization of Cβ (Cα of Gly) atoms in the ab:Env complex structure. An example based on a toy epitope is provided in the supplementary material to illustrate how the diversity is estimated.
To account for the contribution of specific sites in the Ab:Env complex, we used:
H=∑iwiHi−∑i,jwi,jIi,j
(4)
wi,j=(wiHi+wjHj)/(Hi+Hj)
(5)
where wi, wj and wi,j are the weights assigned to epitope sites i, epitope site j, and a pair of neighbor sites (i,j), respectively. The summation is over all epitope sites in the first term and over all neighbor pairs in the second term of Eq 4. When the epitope diversity is normalized, the weight of each site is adjusted as wi′=wi/∑iwi, in which the summation is over all epitope sites.
Based on the chemical similarity between certain amino acids, we grouped amino acids D and E as ‘a’, R and K as ‘b’, N and Q as ‘n’, L and M as ‘l’, V and I as ‘i', and F and Y as ‘f’ before estimating the Shannon entropy. If N was a potential N-linked glycosylation site, it was flagged as ‘g’.
The epitope similarity between a sequence X and a reference sequence R was defined as:
S(R,X)=−[M(R,R)−M(R,X)]
(6)
M(R,X)=[∑iwi∙Sim(Ri,Xi)]/∑iwi
(7)
where M(R, X) is the match score between R and X, wi is the weight assigned to epitope site i, and Sim(Ri, Xi) is from either the BLOSUM62 [34] or the VTML200 [35] matrix, which describes the similarity between Ri (amino acid on site i of R) and Xi (amino acid on site j of X). The minus sign on the right side of Eq (6) converted a distance to the similarity.
To avoid relying on a single strain as the reference, we selected three references strains (n one to five strains with lowest IC50s from 136 strains were tested). The highest epitope similarity to the three references were used as the epitope similarities to the susceptible strains (S(R,X) = max(S(R1,X),S(R2,X),S(R3,X))). Then, we predicted the breadth of an antibody as the fraction of strains with an epitope similarity below a similarity threshold (TH). The number of resistant viruses predicted by the epitope similarity should be same as the number of resistant viruses determined in the neutralization assays. Thus, we set the threshold such that ∑NIC50≥25μg/ml = ∑NSim<TH, in which the summation is over all 34 antibodies. Specifically, given the 136 viruses tested against 34 antibodies, there were 34×136 = 4,624 epitope similarities. The neutralization assays identified 1,561 = 33.8% of virus-Ab combinations as resistant (with IC50 ≥ 25 μg/ml). Thus, we set the similarity threshold as the 1561th element after all the 4,624 epitope similarities were sorted ascendingly.
Data analysis, visualization and statistical testing were performed in the Python environment[41–47]. Statistical details of analyses can be found in the main text and figure captions where applicable; significance was established at p < 0.05. A link to the data archive and code to reproduce the analysis is provided below.
|
10.1371/journal.pntd.0004821 | Safety Overview of a Recombinant Live-Attenuated Tetravalent Dengue Vaccine: Pooled Analysis of Data from 18 Clinical Trials | A recombinant live attenuated tetravalent dengue vaccine (CYD-TDV) has been shown to be efficacious in preventing virologically-confirmed dengue disease, severe dengue disease and dengue hospitalization in children aged 2–16 years in Asia and Latin America. We analyzed pooled safety data from 18 phase I, II and III clinical trials in which the dengue vaccine was administered to participants aged 2–60 years, including long-term safety follow-up in three efficacy trials. The participants were analyzed according to their age at enrollment. The percentage of participants aged 2–60 years reporting ≥1 solicited injection-site or systemic reactions was slightly higher in the CYD-TDV group than in the placebo group. The most common solicited injection-site reactions were pain. Headache and malaise were the most common solicited systemic reactions. In both groups 0.3% of participants discontinued for safety reasons. The most common unsolicited adverse events were injection-site reactions, gastrointestinal disorders, and infections. Reactogenicity did not increase with successive doses of CYD-TDV. The frequency and nature of SAEs occurring within 28 days of any dose were similar in the CYD-TDV and placebo groups and were common medical conditions that could be expected as a function of age. Baseline dengue virus serostatus did not appear to influence the safety profile. No vaccine-related anaphylactic reactions, neurotropic events or viscerotropic events were reported. In year 3 after dose 1, an imbalance for dengue hospitalization, including for severe dengue, observed in participants aged <9 years in the CYD-TDV group compared with the placebo group was not observed for participants aged ≥9 years. In Year 4, this imbalance in participants aged <9 years was less marked, giving an overall lower risk of dengue hospitalization or severe dengue from dose 1 to Year 4 in the CYD-TDV group. These results have contributed to the definition of the target population for vaccination (≥9 years old) for which CYD-TDV has a satisfactory safety profile. Long-term safety will continue to be monitored in the ongoing follow-up of efficacy trials. Safety and effectiveness in real-life settings will be assessed through post-licensure studies.
| Dengue is a mosquito-borne viral infection in tropical and subtropical regions of the world that causes flu-like illness and, in severe cases, death. Every year, 390 million dengue infections occur worldwide, and over 3.9 billion people are now at risk from the disease. Although mosquito control and improved care have helped reduce dengue’s impact, a vaccine is needed. CYD-TDV, a live-attenuated vaccine, has been shown to be efficacious in preventing symptomatic virologically-confirmed dengue disease in children aged 2–16 years. To help establish its safety, we analyzed integrated data from available clinical trials, with emphasis on safety parameters suggested by the World Health Organization. The analysis included data for 77,234 doses in 26,356 participants aged 2–60 years who have received ≥1 dose. We found that the overall safety profile of the dengue vaccine is satisfactory. The imbalance observed for dengue hospitalization and severe dengue observed in year 3 after the first dose in children aged <9 years who had received CYD-TDV vaccine was less marked in year 4. This imbalance was not observed in participants aged ≥9 years in either year. These results have contributed to the definition of the target population for vaccination, i.e. ≥9 years old for which the CYD-TDV vaccine has a satisfactory safety profile.
| Dengue is a viral infection found in tropical and subtropical regions of the world that causes flu-like illness and can, in severe cases, result in death [1]. The disease is caused by four serotypes of dengue virus (Flaviviridae), DENV1-4 which are transmitted by Aedes mosquitoes. The incidence of dengue has grown dramatically over the last few decades. Results from a recent study suggest that 390 million dengue infections occur worldwide each year of which 96 million have clinical manifestations [2]. Another study estimated that about 3.9 billion people are now at risk [3]. Approximately 500,000 people, mostly children, are hospitalized with severe dengue, and approximately 2.5% of them die [4]. Mosquito control and improved clinical management have helped reduce the burden of dengue, but it remains a major public health concern in Asia and the Americas [4].
Among the candidate dengue vaccines, a recombinant, live-attenuated tetravalent dengue vaccine (CYD-TDV) has been tested in several phase I, II, and III clinical trials since 2002 [5–17]. This candidate vaccine is composed of four recombinant vaccine viruses (DENV1–4), each of which expresses the pre-membrane and envelope genes of a single serotype, with a yellow fever virus (YFV) 17D backbone. Pre-clinical studies have demonstrated that CYD-TDV is genetically and phenotypically stable, non-hepatotropic, and less neurovirulent than the parental 17D YFV [18].
The efficacy and safety, up to 25-months after the first dose of a three-dose vaccination regimen of CYD-TDV at D0, M6 and M12 was assessed in a monocenter phase IIb trial in Asia and two large-scale pivotal phase III efficacy trials in Asia and Latin America [19–21]. In addition, preliminary results from on-going longer-term assessment of vaccine safety in these trials, including an extension trial of the phase IIb, have been reported [22].
The World Health Organization (WHO) states that specific potential safety issues should be taken into considerations in the assessment of candidate dengue vaccines and long-term safety surveillance [23, 24]. These issues include possible vaccine-associated dengue-like disease due to vaccine viremia, sensitization or enhancement of dengue illness, and an increased likelihood of severe disease in young children. In addition, for the CYD-TDV vaccine, viscerotropic and neurotropic disease should be assessed due to the YFV backbone of the CYD-TDV. Here we report a pooled analysis of safety data from 18 phase I, II and III clinical trials that assessed the candidate CYD-TDV vaccine. The specific purpose of this pooled analysis was to increase the power to detect potential safety signals and provide more precise estimates of adverse event (AE) rates than those available in individual trials.
A total of 18 clinical trials were included in this pooled safety analysis (Table 1). In addition, the participants from a phase IIb, proof of concept, efficacy trial in Thailand, and the two phase III large-scale efficacy trials in Asia and Latin America were invited to participate in a long-term follow-up period [22]. During this long-term follow-up the participants did not receive any further vaccine or placebo injections and they were analyzed according to their original group (CYD-TDV or placebo) and their age enrollment.
The CYD-TDV vaccine candidate contained about 5 log10 CCID50 of each live, attenuated dengue vaccine virus (serotypes 1–4) [5, 18]. The vaccine was supplied as a freeze-dried powder and was reconstituted in 0.4% sodium chloride immediately prior to use. In placebo-controlled trials, the placebo was 0.9% NaCl, except in two phase II trials, when the placebo was 0.4% NaCl containing 2.5% human serum albumin [9, 15]. The vaccine and placebo were administered by subcutaneous injection in the deltoid region. In trials using a licensed vaccine as an active control, the vaccines were administered according to the usual route of administration.
In all trials, after each injection, participants (or parents or legal guardians for children) used diary cards to record the occurrence and severity of solicited injection-site reactions (for 7 days after vaccination), solicited systemic reactions (for 14 days), and unsolicited adverse events (AEs; for 28 days).
Solicited reactions were all considered as being vaccine-related whereas the vaccine-relatedness of unsolicited AEs and SAEs was assessed by the investigators. AEs considered as vaccine-related were called adverse reactions (ARs). AEs occurring within 30 minutes of an injection were considered immediate AEs. Serious adverse events (SAEs), including deaths, and their relatedness were recorded, as specified in each protocol, by investigators.
Solicited injection site reactions included pain, erythema, and swelling, and solicited systemic reactions included fever, headache, myalgia, asthenia, and malaise. In the three efficacy trials, reactogenicity data were collected for participants who had been randomized to the immunogenicity/reactogenicity subset [19–21]. The severity of solicited reactions was graded as 1, 2 or 3 (Table 2). Analyses are reported by age group, defined using the age at enrolment, and, in trials with data available for specific subgroups, analyses are also reported by baseline dengue serostatus and post-injection viremia. Although in the original trials the safety events were coded using different versions of MedDRA, the events were re-coded using version 14.0 to ensure homogeneity across the trials.
Allergic reactions and anaphylaxis within 7 days of vaccination and severe dengue disease virologically-confirmed any time after dose 1 were considered as adverse events of special interest (AESIs). Viscerotropic or neurotropic events within 30 days of vaccination (because of the YFV backbone of CYD-TDV) were also considered as AESIs [31, 32]. Episodes of serious dengue disease were defined as acute febrile illness, clinically suspected to be dengue by the investigator before virological confirmation, regardless of severity, but requiring hospitalization (with bed attribution); these events were also recorded as AESIs.
Biological parameters were assessed at pre-specified time points in subsets of participants in three main trials and all secondary trials (Table 1). Based on changes observed in phase I trials and on biological abnormalities that can mimic dengue disease, the parameters assessed were creatinine, liver function markers, i.e., alanine aminotransferase [ALT] and AST, and bilirubin and hemoglobin [Hb], hematocrit, white blood cells [WBCs], lymphocytes, neutrophils, and platelets. Biological parameters collected from participants with symptomatic dengue disease were not included in this pooled analysis. Since the parameters were assayed by local laboratories using local standards and the normal reference ranges varied between trials, these biological data were standardized for use in the quantitative and toxicity grading analyses. For hematology parameters, the normal ranges in the individual studies were used since there was limited variability across laboratories.
Vaccine viremia was assessed at pre-specified time points in nine phase I-III trials, in participants with acute febrile illness in six trials conducted in endemic areas (Table 1). Data after doses 1 and 2 were analyzed since vaccine viremia is mainly observed after these injections. Data were analyzed for time points for which vaccine viremia was consistently collected across studies, i.e., D7 (D5-D11) and D14 (D12-D17) in accordance with WHO recommendations [23]. Vaccine viremia was also assessed in individuals with acute febrile episodes within 28 days after vaccination, in trials performed in dengue endemic regions to determine if the fever was vaccine-related (positive vaccine viremia) or was due to dengue infection, in accordance with WHO guidelines [7, 9, 10, 15, 16, 21, 27]. Individuals with positive vaccine viremia, i.e., ≥LLOQ (lower limit of quantitation) measured by YF RT-PCR (non-serotype specific) or CYD RT-PCR (serotype specific) were considered as viremic [16, 19, 21].
WHO guidelines stipulate that the clinical evaluation of a candidate live-attenuated dengue tetravalent vaccine should provide evidence that immune response to the vaccine does not predispose vaccinated individuals to develop severe dengue during natural infections in endemic regions [23, 24, 33]. Suspected symptomatic dengue cases were detected using a passive surveillance method in eight phase I/II non-efficacy trials and using an active surveillance system in the phase IIb and two of the phase III efficacy trials [7, 9, 10, 15, 16, 19, 21, 28, 34, 35]. Blood samples were taken from individuals with acute febrile illness (see case definition in Table 1) occurring from 28 days post-dose 1 for virological testing. In the pooled analysis, the relative risk (RR) of virologically-confirmed dengue (hospitalized and/or severe as assessed by the Independent Data Monitoring Committee (IDMC), see next section) occurring up to 25-months post-dose 1 in the phase IIb and two of the phase III efficacy trials was assessed by age group. In addition, virologically-confirmed dengue cases were assessed for severity according to the WHO 1997 recommendations (DHF grades 1–4) [36]. These events occurring in the non-efficacy trials with longer-term follow-up data were summarized using counts and percentages.
In addition, a pooled analysis of the RR of dengue hospitalization and severe dengue among those hospitalized during longer-term follow-up the phase IIb (Years 3 and 4 post-dose 1) and two of the phase III trials (CYD14: Years 3 and 4 post-dose 1; CYD15: Year 3 post-dose 1) by age group at enrolment was performed. The RRs for these endpoints were also calculated for the overall follow-up, i.e. from D0.
WHO guidelines stipulate that an independent data monitoring committee (IDMC) should be set up to ensure the participants’ safety and provide an independent assessment of the safety and efficacy data [23, 24, 33]. Separate IDMCs were set up for each phase I trial whereas a global IDMC was set up for the phase II/IIb and phase III trials to ensure a consistent assessment of the safety profile across all trials in the CYD clinical development program. The IDMC regularly reviewed safety data. Fatal, related SAEs and serious AESIs were reviewed as they occurred. In addition, throughout the CYD-TDV vaccine clinical development program, all dengue cases that were virologically-confirmed were blindly-assessed by the IDMC for disease severity according to pre-defined criteria, as described previously (IDMC severe disease) [19, 20].
No formal testing between groups was performed for reactogenicity and all safety parameters, except severe virologically-confirmed dengue disease, although 95% confidence intervals (CIs) were calculated. The analysis set included all participants with available data who had received ≥1 dose of the dengue vaccine or placebo. Participants were analyzed according to the product received. Sub-group analyzes were performed by age group at enrollment (2–8 years, 9–60 years, 9–17 years and 18–60 years) and for virologically-confirmed dengue from the three efficacy trials, analyses were also performed for those aged 9–16 years. Analyses were also performed by dengue serological status at baseline; individuals with neutralizing antibodies above the low limit of detection (≥10 (1/dil)) against ≥1 dengue serotype at baseline were considered as dengue-seropositive and the others were considered as seronegative.
The statistical methods for estimating the annual incidence rates and relative risk (RR) have been described previously [19–22]. The RR function by age was estimated by kernel smoothing using the univariate Epanechnikov kernel method. Age was considered as a moving window, centered on all the possible ages, a0, and with a size of a0-h and a0+h, where h was 2.0; this approach resulted in a smoothed curve. The weight for each subject was maximal when their age was equal to a0, decreased as the window moved to a0-h and a0+h and was zero outside this window.
In the main clinical trials, 26,356 healthy participants aged between 2 and 60 years at enrollment, received at least one dose of CYD-TDV vaccine and were included in the integrated and pooled analyses (Table 3). The participants in the main trials received 77,234 doses of CYD-TDV and 36,006 doses of placebo. Some participants in the control groups of certain trials received licensed vaccines in accordance with the trial protocols (Tables 1 and 3).
Slightly more participants vaccinated with CYD-TDV were female (51.0%) than male (49.0%), and the mean age was 11.7 years (Table 4). Almost one-third of the participants were Asian (32%) and nearly half were Hispanic with mixed ethnic origins (45%). Among the subset of 7,500 participants tested for dengue sero-status at baseline 59.2% were sero-positive. The demographic characteristics were generally similar between the CYD-TDV and placebo groups, within each age group.
Overall, in the main trials, there was a trend for a higher percentage of solicited injection-site reactions in participants vaccinated with CYD-TDV (50.9%; 3177/6243) compared with those who received placebo (40.1%; 1018/2537). The highest rates of injection-site reactions were in children in both the CYD-TDV and placebo groups. In both groups, pain was the most common solicited injection-site reaction in all age groups and tended to be more frequent in the CYD-TDV group than in the placebo group (Fig 1; Table 5). Most solicited injection-site reactions were of grade 1 severity, occurred within 3 days of vaccination, and resolved within 3 days.
There was also a trend for a slightly higher incidence of solicited systemic reactions in the CYD-TDV group (65.7%) compared with the placebo group (57.7%) (Fig 1). Headache and malaise were the most frequently reported solicited systemic reactions in the CYD-TDV and placebo groups (Table 5). The rates of solicited systemic reactions were similar between the age groups except for fever, which was less frequent in adults than in children and adolescents. The highest rate of grade 3 solicited systemic reactions reported for those aged 9–17 years and 18–60 years was headache (6.4%); for children aged 2–8 years it was fever (4.4%). Most solicited systemic reactions occurred within 3 days of injection and resolved within 1–3 days, although fever occurred throughout the solicited period (up to 14 days after vaccination). Overall, the rates of solicited injection site and systemic reactions were higher after the first dose than after the second or third dose (Fig 2).
Ten (0.3%) and three (0.2%) participants reported immediate unsolicited AEs in the CYD-TDV and placebo groups, respectively; seven (0.1%) and one (<0.1%) participants reported immediate unsolicited ARs, respectively. Unsolicited non-serious AEs were reported for just over 40% of the participants in the CYD-TDV and placebo groups for all age groups, except for those aged 18–60 years in the placebo group (Fig 1). These occurred more frequently after the first dose than after the second or third doses (Fig 2).
Unsolicited non-serious ARs were reported for 4.6% and 1.6% of participants in the CYD-TDV and placebo groups, respectively. The most common were injection-site reactions (hematoma), gastrointestinal disorders, and infections (Table 6). Most of these ARs occurred within 3 days of vaccination, resolved within 3 days, and were grade 1 or 2. The nature of non-serious ARs in the placebo group was similar to that in the CYD-TDV group, but they tended to be less frequent. The rates of non-serious ARs were generally lower after the second and third doses than after the first.
In the main trials, 218 (0.8%) participants aged 2–60 years in the CYD-TDV group reported ≥1 SAE up to 28-days post any injection and 935 (3.5%) reported ≥1 SAE between day >28 and 6 months post any injection. In the placebo group, there were 121 (1.0%) and 499 (4.0%) participants, respectively. Among participants aged 9–60 years with ≥1 SAE up to 28-days post any injection six and two in the CYD-TDV and placebo groups, respectively (<0.1%) were considered to be vaccine-related SAEs by the investigator; one other SAE (convulsion in a participant aged 9–17 years who had received the CYD-TDV vaccine) was considered to be vaccine-related by the sponsor, but not by the investigator. SAEs considered as vaccine-related by the investigator in the CYD-TDV group were: urticaria, asthma, acute polyneuropathy, tension headache in participants aged 9–17 years; polymyalgia rheumatica and headache in participants aged 18–60 years. In the placebo group, the related SAEs were visual impairment and pyrexia, both in participants aged 9–17 years. In the period between day >28 and 6 months post any injection, one SAE, miscarriage due to blighted ovum considered to be vaccine-related by the investigator occurred in a participant aged 18–60 years in the CYD-TDV group and none in the placebo group. In participants aged 2–8 years, one related SAE was reported in the CYD-TDV group (acute disseminated encephalomyelitis) and two in the placebo group (7th nerve paralysis and angioedema).
The frequency and nature of SAEs occurring within 28 days of any dose were similar in the CYD-TDV and placebo groups. The SAEs were common medical conditions that could be expected as a function of age. The most frequently reported system-organ class was infections and infestations, followed by injuries and gastrointestinal disorders for those aged 9–60 years (Table 7). Among the participants, 22 and 13 had ≥1 neurological SAE within 30-days post-injection in the CYD-TDV and placebo groups, respectively. None of the SAEs resulted in permanent sequelae or death. In addition, a similar profile was observed in those aged 2–8 years.
The nature of SAEs observed between day >28 and 6 months post any injection was similar to that observed up to 28 days post any injection. In addition, no safety concern were observed in the review of SAEs during longer-term follow-up, (up to year-3 post-dose 1), particularly in the two phase III efficacy trials, in which all SAEs were recorded.
Six and eight deaths were reported within 6 months after any injection in the CYD-TDV and placebo groups, respectively, in participants aged 2–60 years; none were assessed as related to the CYD-TDV vaccine. In the CYD-TDV group the deaths were due to road traffic accidents (n = 3), tracheal injury (n = 1), deliberate poisoning (n = 1) and accidental asphyxia by strangulation (n = 1). In the placebo group the deaths were due to drowning (n = 2), T-cell lymphoma (n = 1), road traffic accident (n = 1), bronchoscopic aspiration (n = 1), head injury (n = 1), lupus nephritis (n = 1) and metastatic osteosarcoma (n = 1). In the period after 6-months post-dose 3; 17 and 6 deaths occurred in the CYD-TDV and placebo groups, respectively; none were judged to be related to the CYD-TDV vaccine.
A total of 71/26,356 (0.3%; 95% CI: 0.21; 0.34) and 33/12,562 (0.3%, 95% CI: 0.18; 0.37) participants in the CYD-TDV and control groups discontinued for safety reasons, including the 14 participants who died (see above). Eight and four participants, respectively, discontinued for SAEs considered related to the vaccine. After the occurrence of the SAE, the individual received no further injections, but continued in the safety surveillance, according to the trial protocol. Trial discontinuations due to vaccine-related SAEs all occurred within 28 days of vaccination, except one which occurred between 7 weeks after vaccination (miscarriage due to blighted ovum) in a participant who had received CYD-TDV.
Baseline dengue virus sero-status did not appear to influence the rates of solicited injection site and systemic reactions, unsolicited non-serious AEs and SAEs in those aged 2–8 and 9–60 years (Fig 3).
Among participants aged 9–60 years in the main trials, 46 (0.7%) and 15 (0.5%) experienced non-serious potential allergic reactions (mainly rashes) within 7 days after any injection in the CYD-TDV and placebo groups, respectively. Only 3 were grade 3 (in CYD-TDV group). Most occurred within 3-days post-injection, and resolved spontaneously or after treatment in ≥5 days. In the CYD-TDV group, 14 participants reported these events after dose 1, 5 after dose 2 and 1 after dose 3. Eight of the events in the CYD-TDV group were assessed as vaccine-related; one was grade 3. Five participants (<0.1%) in the CYD-TDV group experienced serious allergic reactions (four experienced asthma or asthmatic crisis and all had a medical history of asthma, asthmatic bronchitis, or bronchial obstructive symptoms; one experienced urticaria and had a history of allergic rhinitis). One participant in the placebo group experienced asthma. A similar profile for AESIs was observed in participants aged 2–8 years. All participants recovered spontaneously or after medical care. Overall, no severe or serious immediate anaphylactic reactions following CYD-TDV vaccination were reported in any age group.
No confirmed cases of viscerotropic or neurotropic disease were reported.
The reporting of serious dengue disease, occurring at any time during the trials, as an AESI was implemented in the four phase III trials (Table 1). In two of these trials (non-efficacy), CYD17 and CYD32, that enrolled 250 and 715 participants, respectively, no serious dengue disease was reported over the 18-month follow-up period. In CYD14, 38/50 (76%) and 56/64 (88%) of serious dengue disease events in the CYD-TDV and placebo groups, respectively, that occurred up to 25-months post-dose 1 were virologically-confirmed. In CYD15, 14/41 (34%) and 38/50 (76%) serious dengue events, respectively, were virologically-confirmed.
The pooled analyses of biological data available for 676 participants showed that most values were within normal ranges both at baseline and after any CYD-TDV dose. The highest rates of grade 3 biological abnormal values reported were 2.2% for low hemoglobin (15/668) and 1.8% for neutrophils (12/668), with no specific patterns being observed. The abnormal values for 12 participants were assessed as being vaccine-related, but none were reported as an SAE within 28 days after any CYD-TDV dose.
In addition, in individual trials, the biological safety profile of the CYD-TDV vaccine was found to be similar to that of the control groups (placebo or licensed vaccines). The incidence of biological abnormalities was 73.7% in participants with viremia compared with 74.8% in those without viremia.
The pooled analyses of viremia data available for 683 participants showed that 38 subjects (<6.0%) had detectable vaccine viremia after dose 1 or 2 of the CYD–TDV vaccine (34 after dose 1; 4 after dose 2). All levels of vaccine viremia were low. None of the participants with viremia experienced immediate AEs, post-vaccination dengue-like syndrome, AEs leading to trial discontinuation, AESIs, or SAEs. The rates of solicited reactions, non-serious unsolicited AEs, non-serious unsolicited ARs were similar between viremic and non-viremic participants (63.2% vs. 69.3%; 63.2% vs. 63.9%; 13.2% vs. 12.1%, respectively). Overall, no safety concerns were identified in the participants with vaccine viremia.
In six trials, blood samples were collected from participants who experienced an acute febrile episode (as defined in each protocol) within 28 days following dose 1 (n = 113) or 2 (n = 106), to be tested for wild-type and vaccine dengue viremia [7, 9, 10, 15, 16, 21, 27]. Vaccine viremia was detected in only one participant (after dose 1). This participant did not have virologically-confirmed dengue disease and had no reports of safety outcomes.
In the phase IIb and phase III efficacy trials, 89 and 134 individuals in the CYD-TDV and placebo groups, respectively, were hospitalized during the 25-month period after dose 1, with a RR of 0.33 (95% CI: 0.25, 0.43) in vaccine recipients. In these trials, there were 15 and 33 cases of severe dengue disease (IDMC assessment) in the CYD-TDV and placebo groups, respectively, with a RR of 0.23 (95% CI: 0.12; 0.42). In the CYD-TDV and placebo groups, 11 and 10 severe dengue cases, respectively, were reported in those aged 2–8 years; in those aged 9–16 there were 4 and 23 cases, respectively. There was no evidence of increased severity of dengue disease based on the review of the severity of clinical outcomes, biological parameters, vaccine viremia and hospitalization rates [19–21].
In the other phase I/II/III non-efficacy trials with a passive surveillance, very few hospitalized virologically-confirmed dengue cases were reported in the CYD-TDV group up to 6 month post-dose 3. None was assessed as severe by IDMC.
The planned longer-term follow-up for participants in the efficacy trials is on-going. In the follow-up study for the phase IIb trial (CYD23/57), and in the phase III study in Asia (CYD14) data for virologically-confirmed dengue disease hospitalization and severe dengue disease hospitalization are available for two-years of longer-term follow-up (i.e. four-years after dose 1). In the phase III trial in Latin America (CYD15), data are available for one-year of longer-term follow-up (i.e. three-years after dose 1). The participants were originally randomized 2:1 to receive CYD-TDV or placebo.
Only one hospitalized virologically-confirmed dengue case, assessed as non-severe by IDMC, was reported in the CYD-TDV group, during the longer-term follow-up (from 6 months after the last injection up to year 5) in one of the three non-efficacy phase I-II trials.
The cumulative RR for hospitalization for dengue disease by age at enrollment was analyzed from D0 up to year 3 in the phase III trial in Latin America and year 4 in the phase IIb and phase III trials in Asia (Fig 4). The RR for dengue hospitalization or severe dengue disease by age at enrollment decreased to below 1 at about age 5 and the 95% CI was below 1 from age 6 onwards, demonstrating an overall reduction in risk of dengue hospitalization and severe dengue in those aged ≥6 years (Fig 4).
This pooled analyses included data from 13 main and 5 secondary clinical trials with 77,234 doses of the current formulation of the CYD-TDV vaccine (about 5 log10 CCID50 of each of the four live, attenuated dengue vaccine viruses) in individuals aged between 2 and 60 years. The results showed that the overall reactogenicity and safety profile for the CYD-TDV vaccine, including solicited reactions, unsolicited and serious adverse events, viremia and biological parameters was satisfactory and comparable to that for placebo across all age groups (2–8 years; 9–60 years) with similar reported rates and nature of events in the CYD-TDV and placebo groups. The safety profile was not influenced by baseline dengue sero-status or successive doses of the CYD-TDV vaccine. In addition, the reported AEs were transient and mainly mild to moderate. No safety concerns were identified for the frequency and nature of unsolicited AEs and SAEs. The nature of the reported SAEs was consistent with the participants’ ages, with few being considered as vaccine-related. There were no vaccine-related deaths. Likewise, no severe or serious immediate anaphylactic reactions to CYD-TDV were reported, although rash, an event that may indicate an allergic reaction, was reported by a similar percentage of participants in both groups.
Viscerotropic and neurotropic diseases are very rare events that have been reported for YFV vaccines such as Stamaril [37]. The event rate after vaccination with a licensed YFV vaccine has been reported to be 0.4/100,000 doses for viscerotropic events and 0.8/100,000 doses for neurotropic events [38]. Since the CYD-TDV vaccine viruses cannot express the YFV E protein, which is largely responsible for YFV tropism, it is unlikely that they will display the same tropism as YFV and none were reported in any age group [39]. Nevertheless, monitoring for these very rare events will continue after vaccine introduction through post-marketing surveillance in real life settings.
Vaccine viremia due to vaccination with live attenuated tetravalent dengue vaccines is considered by the WHO to be unlikely to cause dengue disease due to poor replication of the attenuated viruses [24]. The results presented here show that vaccine viremia was observed in <6% of the participants with available data, and the safety profiles were similar between participants with viremia and those without viremia. In addition, no cases of virologically-confirmed dengue disease were reported among the participants with viremia.
Several risk factors for severe dengue following natural exposure to dengue infection have been identified, including previous infection with a different serotype [24]. Approximately 2% to 4% of patients who have a secondary infection with a heterologous type of dengue virus develop more severe illness [40]. Several hypotheses have been put forward to explain this phenomenon including antibody-dependent enhancement (ADE) [1, 41, 42]. ADE has been observed in mouse models and monkeys, but only indirect evidence is available for this phenomenon in humans [43–46]. As specified by the WHO guidelines, we assessed if the immune response to the live attenuated tetravalent dengue vaccines predisposed individuals in endemic regions to more severe dengue disease [23, 33]. In this regard, a reduction in the rates of dengue hospitalization or severe dengue was observed in the 25-month period post-dose 1 in participants in the CYD-TDV group compared with those in the placebo group. This reduction in rates of dengue hospitalization and severe dengue has continued up to Years 3 and 4 post-dose 1 in vaccinated children aged ≥9. In contrast, the data up to Year 3 from the phase IIb and phase III trials in Asia showed that there was an imbalance of hospitalization and severe dengue in younger children aged <9 years, mainly driven by participants aged <6 years [22]. However, data from Year 4 in these two trials, that enrolled participants aged <9 years, no longer show this imbalance for the younger participants, with the RR for hospitalization decreasing from 1.57 in Year 3 to 0.54 in the phase IIb trial and from 1.58 to 1.19 in the phase III trial. In addition, there were 12 vs. 0 participants with severe dengue in the CYD-TDV and placebo groups, respectively in Year 3, compared with 10 vs. 6 in Year 4. Within the setting of the clinical trials we have closely monitored safety during pre-defined periods (i.e. yearly). However, it is important to look at the overall safety profile in terms of value of the vaccine which shows that rates for hospitalization and severe dengue from post-dose 1 to the end of Year 4 were lower in the CYD-TDV group compared with the placebo group, overall and by age group (<9 years and ≥9 years).
Some interconnected mechanisms, involving interactions between the infecting virus, pre-existing host immunity and vaccine-induced immune responses, have been proposed to explain the Year 3 observations in participants aged <9 years. [47, 48]. Although there is no conclusive evidence yet to support a particular mechanism for this phenomenon, our observations from the Year 4 data showing a decreased RR would seem to support the hypothesis that clustered vaccination in young vaccines, which may act as a primary-like exposure, would result in an ‘accelerated secondary infection’ in that group compared with the placebo group. In the placebo group, the ‘accelerated secondary infection’ would eventually occur at a later time point, making the observed Year 3 imbalance only temporary [48]. It is essential to note that there were no important differences in the clinical pattern and outcomes of severity (e.g. bleeding, thrombocytopenia, shock, plasma leakage, duration of symptoms, duration of hospitalization), biological parameters and presence of viremia for the cases of dengue hospitalization and severe dengue in all participants irrespective of the age, group and observation period; all subjects with severe dengue fully recovered [22]. In addition, the measurement of 38 cytokines, chemokines and growth factors did not reveal any particular immune risk profile in those who had received CYD-TDV vaccine [49]. We observed no differences in the profiles of these factors measured in acute sera from vaccine and placebo recipients who had been hospitalized for dengue or who had severe dengue, irrespective of trial, observation period, severity and age, which is consistent with the clinical findings and viremia results.
The cumulative RR by age showed that there was an overall reduction in risk of dengue hospitalization and severe dengue in those aged ≥6 years. In a conservative approach, a safety margin has been integrated in the vaccine’s indication which is for subjects aged 9 years or more. At the time of manuscript submission, these safety data have supported the licensure of this vaccine, Dengvaxia, in individuals aged 9 to 45 years in Mexico, the Philippines, Brazil and El Salvador and in individuals aged 9 to 60 years in Paraguay, making it the first vaccine to be licensed for the prevention of disease caused by four dengue virus serotypes. The WHO Strategic Advisory Group of Experts (SAGE) on Immunization reviewed the data for CYD-TDV in April 2016 and recommended countries consider introduction of the vaccine in geographic settings (national or subnational) with high endemicity [50]. A WHO vaccine position paper will be published outlining their recommendations in July 2016. In accordance with WHO recommendations, safety and efficacy follow-up will continue for five years after the third dose in the phase IIb efficacy trial and the two phase III efficacy trials in the context of the post-licensure surveillance [23, 33].
To evaluate the safety and effectiveness including indirect effects of the vaccine in ‘real-life’ setting, the post-licensure plan, summarized in the pharmacovigilance risk management plan includes post-authorization safety and effectiveness studies, which will be planned in close collaboration with national health authorities, in addition to routine pharmacovigilance surveillance. This surveillance, which will be implemented once the vaccine has been introduced, will provide data for larger-scale reactogenicity and safety assessments of the CYD-TDV vaccine in real-world settings, including in populations excluded from clinical trials. The pooled safety database was large enough to allow detection of uncommon adverse events occurring in at least 1 per 1,000 individual. As longer-term post-marketing safety data become available the detection of any rare and unexpected events will be also possible.
The results from this integrated analysis show that the CYD-TDV vaccine has satisfactory, short- and long-term reactogenicity and safety profiles for up to four years post-dose 1 in participants aged 9–60 years.
|
10.1371/journal.pgen.1000267 | Indistinguishable Landscapes of Meiotic DNA Breaks in rad50+ and rad50S Strains of Fission Yeast Revealed by a Novel rad50+ Recombination Intermediate | The fission yeast Schizosaccharomyces pombe Rec12 protein, the homolog of Spo11 in other organisms, initiates meiotic recombination by creating DNA double-strand breaks (DSBs) and becoming covalently linked to the DNA ends of the break. This protein–DNA linkage has previously been detected only in mutants such as rad50S in which break repair is impeded and DSBs accumulate. In the budding yeast Saccharomyces cerevisiae, the DSB distribution in a rad50S mutant is markedly different from that in wild-type (RAD50) meiosis, and it was suggested that this might also be true for other organisms. Here, we show that we can detect Rec12-DNA linkages in Sc. pombe rad50+ cells, which are proficient for DSB repair. In contrast to the results from Sa. cerevisiae, genome-wide microarray analysis of Rec12-DNA reveals indistinguishable meiotic DSB distributions in rad50+ and rad50S strains of Sc. pombe. These results confirm our earlier findings describing the occurrence of widely spaced DSBs primarily in large intergenic regions of DNA and demonstrate the relevance and usefulness of fission yeast studies employing rad50S. We propose that the differential behavior of rad50S strains reflects a major difference in DSB regulation between the two species—specifically, the requirement for the Rad50-containing complex for DSB formation in budding yeast but not in fission yeast. Use of rad50S and related mutations may be a useful method for DSB analysis in other species.
| During meiosis, which creates haploid gametes from diploid cells, recombination between two homologous chromosomes increases genetic diversity and, in most organisms, is crucial for proper segregation of chromosomes into haploid nuclei. To better understand where recombination occurs and why it occurs there, we investigated in fission yeast the initiating step in recombination—formation of DNA double-strand breaks (DSBs). A genome-wide DSB map is crucial to understand how DNA sequence and chromatin structure affect DSB formation and may help answer these questions in other organisms, including humans. Mutants in which DSBs accumulate are particularly useful for determining the DSB distribution. As recently reported, however, in budding yeast the DSB distribution in one such widely used mutant, rad50S, differs markedly from that in a dmc1 mutant, in which DSBs also accumulate and appear to have a more nearly wild-type distribution. We have detected in fission yeast a DNA–protein intermediate of recombination assumed to exist, but never before detected, in a recombination-proficient strain (rad50+). The distributions of this intermediate, and therefore those of DSBs, in rad50+ and rad50S strains are indistinguishable. rad50S-like mutations may also accurately reflect the wild-type DSB distribution in other species and may be particularly useful in species lacking Dmc1 orthologs.
| Sexual reproduction involves the fusion of two gametes to create diploid offspring with equal genetic contributions from each parent. To maintain the proper chromosome number (ploidy), it is therefore necessary for the gametes to be haploid. This is achieved via meiosis, where a single round of DNA replication is followed by two nuclear divisions: in the first division, homologous chromosomes (homologs) separate from each other (Meiosis I), followed in the second division by the separation of sister chromatids (Meiosis II). Meiotic recombination, a highly conserved feature of meiosis, creates between the homologs a physical connection that is necessary in most species for proper homolog segregation during Meiosis I.
Before the first meiotic division, homologs become aligned and then intimately synapsed [1]. During this time meiotic recombination is initiated by DNA double-strand breaks (DSBs), introduced by Spo11 in the budding yeast Saccharomyces cerevisiae or its ortholog Rec12 in the fission yeast Schizosaccharomyces pombe [2]. The DNA ends undergo 5′ to 3′ resection, producing 3′ single-stranded (ss) ends capable of invading intact homologous DNA, with the invaded duplex serving as the template for new DNA synthesis [3]. Resolution of the joint DNA molecules can result in a reciprocal exchange of genetic information, called a crossover, which aids proper homolog segregation at Meiosis I. This exchange of genetic material is also beneficial in that it enhances the genetic diversity of the progeny.
As a type II topoisomerase-like protein, Spo11 (or Rec12) breaks phosphodiester bonds in the two DNA strands and becomes covalently bound to each 5′ DNA end of the DSB [4],[5]. This DNA-protein linkage enables determination of where DSBs are made, by chromatin immunoprecipitation (ChIP) of Spo11 or Rec12 and assay of the attached DNA, e.g. by tiling microarray hybridization. In wild-type cells the Spo11 protein is removed from the DNA end by endonuclease action before strand resection occurs [6]. In rad50S mutants, the bound Spo11 or Rec12 protein is not removed from the DNA ends, repair of the DSB by recombination is blocked, and the protein-bound DSBs accumulate [2]. This has facilitated genome-wide analysis of the DSB distribution in both fission and budding yeast strains with the rad50S mutation. These comprehensive DSB maps revealed in both organisms regions of DNA within which DSBs are made at high frequency, called DSB hotspots [5],7.
Two recent studies [8],[9] described a new technique for genome-wide mapping of DSBs that cast doubt on earlier results where a rad50S mutant was used. It had previously been observed that DSBs in S. cerevisiae strains with a rad50S mutation (or mutations in other genes with similar phenotypes, such as sae2Δ and mre11S) did not show in many regions of the genome the same DSB pattern as that in dmc1 Δ mutants [10]. In addition, the overall genetic map of recombination (crossovers) did not agree in certain intervals with the frequency of DSBs determined in rad50S strains [11]. This led Buhler et al. [8] and Blitzblau et al. [9] to develop a method that enriched for regions of ss DNA formed in a dmc1 mutant; dmc1 mutants lack a protein important for strand exchange, and DSBs with resected ends accumulate in these mutants. The enriched ss DNA was hybridized to a genome-wide tiling microarray of oligonucleotides to identify sites of DSBs. The results showed, in many but not all regions of the genome, a clear under-representation of DSB hotspots in rad50S-like mutants compared to the distributions in the wild type or dmc1Δ mutant, which appear similar by Southern blot analysis. Specifically, the intensity of breakage at some, but not all, DSB hotspots was greatly reduced in the rad50S-like mutants compared to that in the wild type or dmc1Δ mutant. The validity of DSB maps created with rad50S mutants, not only in S. cerevisiae but in other organisms as well, is therefore under new scrutiny.
Our lab has reported that DSB hotspots in S. pombe are preferentially located in large intergenic regions and are widely-spaced – on average there are about 65 kb between hotspots – but these experiments were done with rad50S mutants [5],[12]. We wanted to know if a wild-type (rad50+) strain has a DSB map similar to that seen in rad50S mutants. ChIP experiments to detect the Spo11-DNA covalently linked intermediates in RAD50 strains of budding yeast have not been successful [13],[14; M. Lichten, personal communication], apparently due to the short life-span of the hypothesized Spo11-DNA complexes. By contrast, in fission yeast we were able to detect and analyze the wild-type (rad50+) Rec12-DNA complexes. To our knowledge, this is the first time that this protein-DNA intermediate has been detected in recombination-proficient cells. We report here that the locations of DSBs, measured as Rec12-DNA linkages, across the genome in S. pombe rad50+ meiosis are indistinguishable from those in rad50S strains, although the intensities are lower, as expected due to ongoing DSB repair in rad50+ strains. Therefore, conclusions from our earlier studies using the rad50S mutation are still valid: in particular, DSBs are separated by large distances and are preferentially located in large intergenic regions [5],[12]. However, the genetic recombination maps (crossover distributions) and physical maps (DSB distributions in rad50+ and rad50S strains) display non-congruence in S. pombe. We discuss the significance of these observations for studies of meiotic recombination in S. pombe and in other species, including humans.
We began our comparison of the DSB distributions in rad50+ and rad50S strains by assaying DSBs using standard Southern blots. rad50+ and rad50S strains were meiotically induced (Figure S1), and the DNA was extracted and digested with NotI restriction enzyme to generate large DNA fragments, which were separated by pulsed-field gel electrophoresis. Previous Southern blot analyses of DNA from rad50+ and rad50S strains revealed the same meiotic DSB pattern on the 0.5 Mb NotI restriction fragment J, which includes the well-characterized DSB hotspot mbs1 [5],[12],[15]. Two additional NotI fragments were probed to strengthen this observation; these analyses were of the 0.5 Mb NotI fragment K (Figure 1A and S2A) and the 1.2 Mb NotI fragment D (Figure 1B and S2B). These results revealed that rad50+ strains have on each fragment multiple DSB sites at the same locations as those from a rad50S strain. As expected, in almost all cases the maximal level of the transient DSBs in the rad50+ strain was less than that in the rad50S strain, in which DSBs accumulate. At each hotspot site on these NotI fragments in a rad50+ strain there is a hotspot in the rad50S strain, and vice versa.
We next compared DSB sites in a repair-deficient mutant other than rad50S in which DSBs accumulate. During meiotic recombination in S. pombe there are two mediator complexes that assist the strand exchange protein Rhp51 in strand invasion: Swi5-Sfr1 and Rhp55-Rhp57 [16],[17]. Mutants lacking either complex show reduced recombination and delayed DSB repair, and strains with a mutation in both complexes display recombination defects and spore viability as severe as an rhp51 null mutant [16],[18] but slightly better growth and meiotic induction than an rhp51 null mutant (RWH, unpublished data). Thus, the double swi5Δ rhp57Δ mutant is an ideal candidate for assaying defective DSB repair at a stage later than the rad50S repair defect, allowing for DSB accumulation in a non-rad50S strain. Southern blot analysis of NotI fragments K (Figure 1A), D (Figure 1B), and J (data not shown) from the swi5Δ rhp57Δ mutant revealed a DSB pattern similar to those seen in rad50+ and rad50S strains, except that the broken DNA persisted in rad50S and swi5Δ rhp57Δ mutants but was repaired in wild type. Thus, by Southern blot analysis in the rad50S mutant there is no lack of DSB sites that are present in other mutants, unlike the situation in S. cerevisiae, as noted above [8],[9].
Since Rec12 becomes covalently bound to the DNA ends at a meiotic DSB [5], ChIP of epitope-tagged Rec12 protein without exogenous cross-linking can identify the genomic loci where DSBs occur. Previous ChIP analysis of FLAG-tagged Rec12 in rad50S meiosis showed that DSB hotspots assayed by locus-specific PCR gave a meiosis-specific signal dependent on Rec12 (i.e., DSB formation), while DSB coldspots gave no detectable signal [5]. We wanted to know if it was possible to repeat this analysis in a rad50+ meiosis, or if Rec12 was removed from the DNA too quickly to be detected, as appears to be the case in budding yeast [13],[14]. PCR analysis of two prominent DSB hotspots, ade6-3049 on chromosome III [19] and mbs1 on chromosome I, revealed that DNA isolated 3.5 h after induction of meiosis was considerably enriched by ChIP when compared to 0 h (uninduced) DNA, based on the relative abundance of PCR products. This was true for DNA from a rad50+ strain as well as from a rad50S strain, though as expected enrichment was lower in the rad50+ strain due to ongoing repair of the DSBs (Figure 2). There was no detectable enrichment at the DSB coldspot ura1 (Figure 2C and [5]). In addition, the enrichment at the hotspots in rad50+ was transient: very little signal was detected at 0 h (before DSB formation) or at 6 h after meiotic induction (after DSB repair). The maximal signal was at 3.5 h, which is about the time of maximal DSBs detectable by Southern blots in rad50+ strains (Figure 1 and [5]). This contrasts with a rad50S meiosis, where the PCR assay detects high DNA enrichment at least to 6 h after meiotic induction [5], a reflection of Rec12 remaining bound and the DSBs not being repaired in rad50S meiosis.
As an additional test for Rec12-DNA linkages in rad50+ strains, we treated meiotic extracts with a protease (or not, as a control) and extracted the material with phenol. Protein-linked DNA is removed from the aqueous phase by phenol extraction [20]. A significant fraction of the DNA at the mbs1 and ade6-3049 DSB hotspots was removed by phenol extraction, as expected for DNA covalently linked to Rec12 protein, unless the extracted material was treated with a protease before extraction. This was true for material from both rad50+ and rad50S strains (Figure S3), and contrasts sharply with results from S. cerevisiae, in which no detectable DNA is removed by phenol extraction in RAD50 strains [20]. Our results show that a significant fraction of the DNA at DSB hotspots in S. pombe rad50+ strains remains linked to a protein, likely Rec12.
To extend these observations to the entire genome, we used a genome-wide microarray analysis similar to our previous analysis with rad50S strains [5]. We prepared Rec12-DNA samples from immunoprecipitated (IP) chromatin and from whole-cell extracts (WCE) prepared at 0 and 3.5 h in rad50+ meiosis and at 0 and 5 h in rad50S meiosis. These samples were amplified, differentially labeled, and hybridized to a tiling oligonucleotide microarray (∼44,000 60-mers, “probes,” spaced approximately every 290 bp across ∼12.5 Mb of the non-repetitive S. pombe genome). The relative frequency of Rec12-DNA linkage at each probe position was measured as the median-normalized ratio of IP signal to WCE signal. The 0 h data [log (IP/WCE) values] were normally distributed, as expected for random background data (Figure S4). In contrast, a distinct subset of probes in both the 3.5 h rad50+ and 5 h rad50S data showed elevated non-normal ratios, reflecting genuine enrichment over background. The analysis below is focused on these enriched values.
The data show that the sites of Rec12-DNA linkage, and hence the sites of meiotic DSBs, in a rad50+ meiosis almost completely coincide with those in a rad50S meiosis. The genomic intervals of NotI fragment K and NotI fragment D, analyzed for DSBs by Southern blot analysis (Figure 1A and B), are compared by microarray analysis in Figure 3. There are no significant peaks of Rec12-DNA linkage in rad50+ that are not also in rad50S; this correspondence is true genome-wide, as well (Figure S5). This result is dramatically different from that observed in S. cerevisiae, where multiple genomic regions show many more DSB hotspots in RAD50 (dmc1Δ) meiosis than in rad50S meiosis, as measured by the enrichment for accumulated ss DNA ends [8],[9]. Within these regions, at many DSB hotspots seen in RAD50 (dmc1Δ) the level of breakage in rad50S falls below the authors' definition of a hotspot. We note in particular that the DSB patterns surrounding the centromeres in rad50S and rad50+ strains are indistinguishable in S. pombe (Figures 3 and S5) but markedly different in S. cerevisiae [8],[9].
Closer examination of one hotspot from each of NotI fragments K and D revealed that the shape of the enrichment peaks, considering non-background probes, was essentially identical for the rad50+ and rad50S datasets, but with ∼3-fold less enrichment in the rad50+ experiment (Figure 4). This 3-fold difference is consistent with comparisons of maximal meiotic DSB frequencies in rad50+ and rad50S strains by Southern blot analysis [5; unpublished data] and appears to hold true genome-wide (Figures 5A, S6A and S7A). The matching peak shapes indicate that the Rec12-DNA shear sizes, DSB positions, and relative DSB intensities are nearly identical in the rad50+ and rad50S experiments. This result rules out the possibility that the Rec12-DNA species detected by microarray hybridization in the rad50+ experiment involves a significantly different length of DNA than that in the rad50S experiment. More specifically, we can discount the Rec12-DNA species in the rad50+ experiment being a short Rec12-oligonucleotide released after DSB end-processing [6], rather than the Rec12-DNA intermediate first formed by Rec12 and accumulating in the rad50S background. In fact, such a short Rec12-oligonucleotide would not be amplified and hybridized in the procedure used here. As expected, the lengths of the two strands of DNA extending from one side of the DSBs at one hotspot to a common restriction site were similar (Figure S8), suggesting that at least some of the 5′ ends remain full length (i.e., attached to Rec12).
We analyzed our genome-wide data on a probe-by-probe basis to determine if hotspots of DSBs were at the same positions in both rad50+ and rad50S; i.e., are the probes with high IP/WCE ratios in rad50+ also high in rad50S? For each of the ∼44,000 probes on the microarray, the IP/WCE ratio from the 3.5 h rad50+ DNA was plotted against the IP/WCE ratio of the 5 h rad50S DNA; these are the times of maximal DSB levels in the two strains (Figure 1 [5],[12]). Essentially every probe that showed enrichment (high normalized IP/WCE ratio) in one strain was enriched in the other (Figures 5A and S6A). There is a clear quantitative, positive correlation between the IP/WCE ratios (Figure S9A) of these enriched (DSB hotspot) probes across the two experimental conditions, consistent with the data in Figure 4. Background probes showed no such quantitative correlation. Probes showing enrichment in the rad50S 5 h (induced) DNA (i.e., DSB hotspot probes) showed, however, no significant enrichment in the rad50+ 0 h (non-induced) DNA (Figures 5B, S6B, and S9A), as expected since DSB hotspots are not apparent in the 0 h data (Figures 3 and S5). The subset of enriched probes in the 5 h rad50S and 3.5 h rad50+ conditions was consistent across the two independent inductions of rad50S and rad50+ (Figure S10). These data indicate that there are no obvious regions of the S. pombe genome where DSB hotspots occur in rad50+ but not in rad50S strains.
Compared to the correlation between the S. pombe rad50+ and rad50S meiotic datasets, the correlation between the RAD50 (dmc1Δ) and rad50S enrichment ratios of S. cerevisiae is much weaker (Figures S6C, S6D, and S9B). Among probes showing enrichment, there are many probes that have a higher, and often much higher, enrichment ratio in RAD50 (dmc1Δ) meiosis than in rad50S meiosis, as well as other probes that show similar high enrichment ratios in both. This is expected, given loci where DSBs are frequent in both RAD50 (dmc1Δ) and rad50S meiosis and other loci where DSBs are frequent only in RAD50 (dmc1Δ) [8],[9].
As another way of comparing the meiotically induced rad50S and rad50+ data from S. pombe, we identified regions of significant ChIP enrichment using ChIPOTle [21], with a p value cutoff of 0.001. Due to the accumulation of Rec12-DNA intermediates, Rec12 ChIP enrichment over background should be greater in the rad50S experiments. As the p value that ChIPOTle attaches to peaks is dependent on their degree of enrichment over background, peaks can be detected with greater sensitivity in the rad50S experiments. Therefore, for any given significance threshold, if the same pattern of DSBs occurs in both the rad50S and rad50+ experiments, we expect some peaks (the stronger ones) to be detected in both sets of experiments but other peaks (the weaker ones) to be detected only in the rad50S experiments. This is what we observed. Combining the two independent inductions (Datasets S1 and S2), an average of 10.2% and 5.0% of the genome was enriched (i.e., within ChIPOTle-determined peaks) in the 5 h rad50S and the 3.5 h rad50+ data, respectively, but 4.9% of the genome was enriched in both. Therefore, there is no significant class of peaks identified in the rad50+ data that do not have equivalents in the rad50S data. In contrast, in S. cerevisiae [8] 63% of the genome was enriched in the RAD50 (dmc1Δ) strain, and 32% in the rad50S strain, but 31% of the genome was enriched in both. Therefore, in S. cerevisiae there is a significant class of probes that are enriched only in the RAD50 (dmc1Δ) background, as well as probes that are enriched in both backgrounds.
A simpler consideration of the ChIPOTle analysis leads to the same conclusion. In our S. pombe data, an average of 255 significant peaks was detected in the two 3.5 h rad50+ datasets, and 427 in the two 5 h rad50S datasets. Essentially all (94%) of the rad50+ peaks were present in the corresponding rad50S datasets (i.e., the peaks overlap), but only 48% of rad50S peaks were present in the rad50+ dataset. That is, there are almost no peaks detectable in the rad50+ background that are not detected in the rad50S background. The larger number of peaks identified in the rad50S background is expected from the greater peak detection sensitivity of ChIPOTle using the rad50S dataset, as discussed above. For probes showing enrichment in either the 3.5 h rad50+ or 5 h rad50S datasets, the rad50S enrichment ratio is consistently ∼3 fold higher than the rad50+ enrichment ratio (Figure S7A). In comparison, the data from S. cerevisiae [8] look very different. Here, 95% of 2010 rad50S peaks overlap with RAD50 (dmc1Δ) peaks but only 60% of 1816 RAD50 (dmc1Δ) peaks overlap with rad50S peaks. That is, there is a substantial number of loci (hotspots) where significant DNA breakage is seen in the RAD50 (dmc1Δ) strain but not in the rad50S strain, as well as other loci where significant breakage is seen in both strains (Figure S7B).
Our detection, for the first time, of Rec12-DNA covalent linkages in recombination-proficient (rad50+) cells allowed us to compare the genome-wide distribution of these linkages, and hence meiotic DSBs, in rad50+ strains and the more thoroughly studied rad50S strains. Our results show that the genomic distributions of S. pombe meiotic DSBs in these strains are indistinguishable (Figures 1, 3, 4, S2, S5, and [5],[12]). In addition to confirming our previous meiotic DSB map [5], our results have additional implications about the regulation of DSB formation and differences in this regulation among species, as discussed below.
An analysis of DSBs by ChIP of the Spo11 protein in a rad50+ meiosis in budding yeast has not been successful [13],[14; M. Lichten, personal communication], presumably because Spo11 is rapidly removed from the DSB 5′ ends. The success of our Rec12-ChIP analysis in fission yeast rad50+ strains (Figures 2, 3, and S5) suggests that the Rec12 protein remains linked to DNA for a longer period of time in fission yeast than does Spo11 in budding yeast. However, even in fission yeast, Rec12 appears to be removed in a rather short period – a DNA sample taken 30 min after the 3.5 h DNA sample studied here (Figures 2, 3, and S5) and similarly analyzed on a microarray showed no discernible difference genome-wide from the 0 h pre-meiotic DNA sample (unpublished data). In addition, multiple assays for Rec12-DNA by PCR at selected loci show that the Rec12-DNA species diminishes substantially between 3.5 and 4 h (Figure 2 and unpublished data). Thus, the first step of DSB repair (Rec12 removal) begins about 30 min after DSB formation (which occurs at about 3 h after meiotic induction) and about 30 min before joint DNA molecules (single Holliday junctions) are first detected [22]. The time between DSB formation and joint molecule detection in S. cerevisiae is also about 1 h [e.g., 23]. We infer that in S. pombe the Rec12-DNA complex persists until the nuclease for its removal, perhaps the MRN (Mre11-Rad50-Nbs1) complex, binds and acts on this intermediate. In S. cerevisiae this step may be very fast.
Why does the rad50S mutation behave differently in these two yeasts? The answer may lie in the differential dependence on the MRN (MRX in S. cerevisiae) complex for DSB formation in these two distantly related yeasts. S. cerevisiae rad50Δ and mre11Δ mutants do not form DSBs [24],[25], whereas S. pombe rad32Δ (mre11 homolog) and rad50Δ mutants form DSBs with the same kinetics as rad50S mutants, although none of these mutants repair the DSBs [26]. The dependence on MRX for DSB formation in budding yeast likely reflects its Spo11-dependent binding at sites of DSBs [13], where it is then also in position to quickly remove the Spo11 protein from the DNA. Since fission yeast lacks this MRN requirement for DSB formation, MRN may be recruited only after DSBs are formed, allowing for a greater life-span of Rec12-DNA complexes. The initial steps of DSB repair – the removal of Rec12 (Spo11) and resection to form invasive ss DNA ends [6] – by MRN and other proteins are thought to be similar in both organisms.
The rad50S mutation commonly used in both organisms changes the same amino acid of the protein (Lys81→Ile81) [12],[18],[24],[27], but this rad50S mutant does not form the full number of DSBs in budding yeast [8],[9]. These observations lead us to suggest that in budding yeast, which requires MRX for DSB formation, the rad50S (K81I) mutant is incompetent (or less competent) compared to RAD50 to activate DSB formation at some sites or regions but not at others. Thus, not all hotspots are revealed in S. cerevisiae rad50S (K81I) strains [8],[9]. In dmc1 mutants, a more complete spectrum of hotspots would, in this view, be activated by the wild-type MRX complex, as observed [8],[9]. In contrast, the lack of MRN requirement for DSB formation in S. pombe may be the basis for the rad50S mutation having no discernible effect on the distribution of DSBs in fission yeast. The decision to make DSBs is made before MRN's meiotic activity on DNA, making MRN unnecessary for the formation – but not the processing – of meiotic DSBs. Thus, in S. pombe the entire spectrum of DSBs, with readily detectable Rec12-DNA complexes, is observed. In S. cerevisiae and other species in which Rad50 is required for DSB formation, rad50 mutants with an amino acid substitution other than Rad50 (K81I) [27] and that accumulate DSBs may also allow a full spectrum of DSBs to be observed.
Crossovers arising from meiotic recombination are much more uniformly distributed across the genomes of both fission yeast and budding yeast than are the sites of DSBs observed in rad50S strains [8],[11],[12],[15]. A recent study by Buhler et al. [8] determined that the non-congruence in S. cerevisiae is due at least in part to a lower DSB frequency and more restricted DSB distribution in a rad50S strain than in a dmc1Δ strain, which appears to be more representative of wild-type meiosis. Our results in wild-type (rad50+) S. pombe meiosis reveal the same DSB pattern as that seen in earlier studies of rad50S mutants [5],[12]: meiotic DSBs are preferentially located in large intergenic DNA regions and are separated by long distances (∼65 kb on average) where no DSBs are apparent. Studies of wild-type (rad50+) meiosis have in the past been problematic, primarily because the repair of DSBs in wild-type strains prevents all of the meiotic DSBs from being analyzed and low-level breaks can be missed. While there may be low-level DSBs dispersed across the S. pombe genome and not detected in our analysis, it is clear that there are essentially no DSB hotspots in rad50+ that are not present in rad50S (Figures 3, 5, and S5).
In S. pombe, some intervals with no detectable DSBs nevertheless contain abundant crossovers [5],[12],[15]. The 0.5 Mb region of NotI fragment J on chromosome I has been extensively studied both genetically for crossovers and physically for DSBs [12]. The number of DSBs detected in this interval – about one DSB per four DNA molecules in a meiotic cell – is not enough to account for the crossovers that occur on this fragment – about one per meiotic cell – since there are about three times more intersister (genetically silent) exchanges than interhomolog exchanges, at least at the major DSB hotspot mbs1 on that fragment [15]. In the 57 kb res2 – ura1 subregion of NotI fragment J there are ∼0.08 crossovers, over 10 times more than predicted by the <0.005 DSBs per meiotic tetrad [12; unpublished data]. It had been suggested that crossovers in such regions might arise from ss nicks [15], but since all meiotic crossovers are dependent on Rec12 [28], we would expect even sites of nicks to have Rec12 covalently linked to the DNA and therefore enriched by ChIP. Ludin et al. [29] analyzed by microarrays the genome-wide distribution of Rec12 after it was formaldehyde-crosslinked to DNA and found a more uniform distribution than we find for Rec12 self-linked to DNA. Although much of the Rec12 detected with formaldehyde-crosslinking does not make detectable DSBs, this population of Rec12 may nevertheless be required for crossovers in DSB-poor regions. Although the basis of the DSB–crossover discrepancy remains undetermined, our results rule out one explanation – that DSBs are underrepresented in rad50S strains.
Results from the DSB analysis of a dmc1Δ mutant in S. cerevisiae [8],[9] have brought into question the reliability of DSB maps generated using the rad50S mutation. Our results in S. pombe question whether these findings from budding yeast apply to other organisms. rad50S-like mutations may reveal the wild-type distribution in other species, particularly those in which Rad50 is not required for DSB formation, such as Arabidopsis thaliana, Drosophila melanogaster, Coprinus cinereus, and perhaps Caenorhabditis elegans [30],[31],[32; M. Zolan, personal communication]. In species that appear not to have a Dmc1 ortholog a microarray analysis of DSBs performed with a rad50S-like mutant may be the most feasible method to reveal the DSB distribution. Our results indicate that in these cases the results may reflect those in wild type. Regardless of the genetic background used and methodology chosen, understanding where meiotic DSBs occur and what DNA characteristics influence DSB location remains an important question in understanding the regulation of meiotic recombination.
Strains used were GP1979 (h−/h− ade6-52/ade6-M26 lys3-37/+ +/ura1-171 pro1-1/+ pat1-114/pat1-114 end1-458/end1-458), GP3718 (h+ ade6-3049 pat1-114 rad50S end1-458), GP6203 (h−/h− ade6-3049/ade6-3049 pat1-114/pat1-114 rad50S/rad50S rec12-201::6His-2FLAG/rec12-201::6His-2FLAG +/his4-239 lys4-95/+), and GP6232 (h−/h− ade6-3049/ade6-3049 pat1-114/pat1-114 rec12-201::6His-2FLAG/rec12-201::6His-2FLAG +/his4-239 lys4-95/+). Alleles were described previously [5],[18],[19],[26].
To assess events in S. pombe meiosis, we used strains carrying the temperature-sensitive pat1-114 mutation, which affords high synchrony but has no detectable effect on DSB formation or location [5],[33]. Cultures were grown to mid log-phase and starved for nitrogen to arrest cells in the G1 phase of the cell cycle; nitrogen was restored and the temperature raised to initiate meiosis. Cells were harvested, embedded in agarose plugs, and treated with enzymes to lyse the cells and to partially purify the DNA. After digestion with NotI restriction enzyme, the DNA was subjected to pulsed-field gel electrophoresis and Southern blot hybridization. These methods are detailed elsewhere [12],[34]. The probe used on NotI fragment K (Figure 1A and S2A) extends from bp 3600336 to bp 3601359 on chromosome I; the probe used on NotI fragment D (Figure 1B and S2B) extends from bp 1025344 to bp 1026300 on chromosome I (accession # NC_003424.3).
Strains with both rad50+ and rad50S genetic backgrounds were induced twice. Chromatin was prepared, immunoprecipitated, assayed by locus-specific PCR, and analyzed on microarrays as described [5], except that Agilent Whole Genome 4×44 K S. pombe oligonucleotide microarrays were used. The 0 h and 3.5 h rad50+ DNA and the 5 h rad50S DNA were analyzed on microarrays twice; the 4 h rad50+ DNA was analyzed only once, as was the 0 h rad50S DNA, which confirmed earlier results [5].
Regions of significant enrichment were identified using the Gaussian setting of ChIPOTle (v 1.0) [21] with a p value cutoff of 0.001.
|
10.1371/journal.pbio.2004526 | Zbtb7a is a transducer for the control of promoter accessibility by NF-kappa B and multiple other transcription factors | Gene expression in eukaryotes is controlled by DNA sequences at promoter and enhancer regions, whose accessibility for binding by regulatory proteins dictates their specific patterns of activity. Here, we identify the protein Zbtb7a as a factor required for inducible changes in accessibility driven by transcription factors (TFs). We show that Zbtb7a binds to a significant fraction of genomic promoters and enhancers, encompassing many target genes of nuclear factor kappa B (NFκB) p65 and a variety of other TFs. While Zbtb7a binding is not alone sufficient to directly activate promoters, it is required to enable TF-dependent control of accessibility and normal gene expression. Using p65 as a model TF, we show that Zbtb7a associates with promoters independently of client TF binding. Moreover, the presence of prebound Zbtb7a can specify promoters that are amenable to TF-induced changes in accessibility. Therefore, Zbtb7a represents a widely used promoter factor that transduces signals from other TFs to enable control of accessibility and regulation of gene expression.
| Gene activation is driven by the binding of regulatory proteins to the specific DNA sequences that control each gene. However, these sequences are not always accessible for binding in every type of cell, and so differences in their accessibility can underlie the range of cell types in which particular genes can be activated. Although several cellular processes can alter the accessibilities of these sequences, it is still often unclear how these processes are directed to act at specific genes. We have discovered that the protein Zbtb7a binds near numerous gene-regulatory sequences throughout the genome and that it enables other DNA-binding proteins to trigger changes in their accessibility and to activate nearby genes. However, unlike many other factors that control gene activation, the binding of Zbtb7a alone does not seem to be sufficient to switch on gene expression; instead, its function is required for activation of genes that are independently bound by a specific set of transcription factors (TFs), and it could therefore be considered to “transduce” their gene-regulatory activities. The implication of this is that the presence or absence of Zbtb7a at any gene in a particular cell type may represent one of the aspects that can determine whether that gene is able to be activated or not.
| The expression of protein-coding genes is governed by the combined activities of activating and repressing transcription factors (TFs), which bind in a specific fashion to regulatory sequences in proximal promoter regions and at distal enhancers. The abilities of TFs to bind to their genomic target sequences are limited by the chromatin context, or “accessibility,” at these sites. Accordingly, regulation of the accessibility of promoter and enhancer sequences represents one of the major, critical controls of the specificity of gene expression [1,2]. However, although many chromatin-remodeling complexes capable of regulating accessibility have been described [3,4], the upstream pathways that direct their activities to particular genomic loci and in response to defined signals are much less well characterised.
TFs differ in their requirements for binding to their target sites [5]. In many cases, TFs that are able to bind to less accessible regions can themselves trigger changes in local accessibility and thereby allow the recruitment of “secondary” TFs with more-stringent binding requirements, which can contribute to activation (or repression) of specific target genes [6,7]. This context-dependent mode of activation—which requires the presence of additional promoter/enhancer sequences that can be recognised and bound by secondary TFs—is well-established for the first-binding “pioneer” factors and has been described for other TFs, including NFκB [8,9].
The mechanistic steps underlying the activities of pioneer factors and similarly behaving TFs are still not well characterised [10]. An intriguing question is whether TFs with pioneer-like activity generally act by directly recruiting chromatin-modifying complexes to their binding sites or whether their activities may instead be transduced by preloaded factors at target promoters and enhancers. A corollary of the latter model is that the presence or absence of such “transducer” proteins could predetermine which sites are amenable to remodeling upon TF binding. However, so far, the identities or existence of possible transducers are unknown.
Separating regulation of accessibility from direct gene activation has been experimentally hard to approach because it is often difficult to distinguish whether changes in promoter accessibility represent downstream effects of the transcription process itself. However, we have previously shown that in mouse 3T3 fibroblasts, the principal activating NFκB subunit p65 is able to regulate target gene expression by 2 independent modes: either by direct transcriptional activation—dependent on its TA1 and TA2 domains, which interact with the Mediator complex—or indirectly, by controlling promoter accessibility for the binding of secondary TFs [9]. We term the as-yet undefined region(s) of p65 that mediate this latter activity as “TA3.” In this study, we have exploited this system as a model to identify molecular cofactors involved in accessibility-driven gene activation.
To uncouple direct transcriptional activation from the control of promoter accessibility, we generated a variant form of p65 lacking its 2 direct transcription activation domains. This variant (hereafter “p65 TA3”; Fig 1A) is completely unable to drive direct transcriptional activation of a reporter plasmid containing tandem NFκB-binding motifs linked to a minimal promoter (Fig 1B; in agreement with previous reports [11]). Despite its lack of direct activation capacity, p65 TA3 is able to robustly induce activation of the endogenous Cxcl2 gene and other NFκB targets in 3T3 fibroblasts (hereafter simply “fibroblasts”) upon stimulation of the NFκB pathway by tumour necrosis factor alpha (TNF-α (Fig 1C and see later), albeit to lower magnitudes than those driven by the direct activation domains. Activation of endogenous promoters by p65 TA3 is associated with a strong induction of promoter accessibility, measured by increased DNase-I hypersensitivity (Fig 1D and S1A, S1B and S1C Fig).
These results suggest that the ability of p65 TA3 to induce transcription is “context dependent” and requires the presence of additional sequences within natural promoters, whose regulated accessibility could mediate gene activation. To investigate this, we generated reporter plasmids containing the natural approximately 1 kb promoter sequences from 2 p65 target genes, Cxcl2 and Saa3. Episomal plasmids have been shown to be rapidly chromatinised upon transfection and can mimic normal patterns of accessibility at regulatory elements [12,13] (although precise nucleosome placement may differ from that at endogenous promoters [14,15]). For plasmids containing both promoters, p65 TA3 was able to trigger robust reporter gene activation (Fig 1E). Deletion of the NFκB-binding motifs within the cloned Cxcl2 promoter abolished TA3-driven reporter expression, indicating that activation requires both direct binding of p65 TA3 as well as indirect regulation dependent on additional sequence elements present in natural promoters.
We identified the region within p65 TA3 responsible for its activity using a series of truncations, deletions, and substitution mutations (S1D, S1E and S1F Fig). The minimum active region (p65 amino acids 342–390) is not predicted to encode any known protein structural domains or catalytic activities and shares little primary sequence similarity to other non-NFκB proteins, but it is highly conserved among mammalian homologues of p65 (S1G, S1H and S1I Fig). The TA3 region displays conserved activity when assayed in either mouse or human cells, using the mouse Cxcl2 promoter as a reporter (S1J Fig), demonstrating that the function (and potential interactor[s]) of the TA3 region is conserved between these species. Targeted mutations within the minimal TA3 region impaired its ability to induce activation of the endogenous Cxcl2 promoter by more than 80% (Fig 1F, S1E Fig). Together, our results define the minimal TA3 region of p65 as a separable, conserved, functional element.
To search for factors that are generally required for TF-driven regulation of promoter accessibility and context-dependent gene activation, we used stable isotope labelling of amino acids in cell culture (SILAC) coupled with mass spectrometry (MS) to find proteins that could bind to the TA3 region but not to a functionally impaired mutant form (Fig 1G). We expressed both forms of TA3 as fusions to glutathione S-transferase (GST) in Escherichia coli and used these to enrich for binding proteins in nuclear extracts by affinity chromatography. Taking advantage of the conserved functionality of the TA3 region, we initially used nuclear extracts from TNF-α-stimulated human HeLa S3 cells as a source of potential interacting proteins. Among a small number of proteins with consistently high SILAC ratios (representing the ratio of binding to normal versus functionally impaired TA3), we identified the protein Zbtb7a (Fig 1H; empirical P = 0.0052; false discovery rate [FDR; Benjamini-Hochberg] = 0.0086), which has previously been reported as an interaction partner of p65 [16]. Zbtb7a exhibited the highest specificity for the TA3-containing bait when compared to pull-downs using GST alone (to identify likely promiscuously binding contaminants; S1K Fig); moreover, Zbtb7a also showed high SILAC ratios in pull-downs using mouse fibroblast nuclear extracts. In independent experiments, conventional (nonquantitative) MS analysis of proteins after fractionation by gel electrophoresis was consistently able to detect Zbtb7a-derived peptides in pull-downs using TA3, but with only reduced coverage when using the functionally impaired TA3 mutant (S1K Fig).
We validated the MS results by using the same GST fusion-protein baits to pull down a tagged form of Zbtb7a from transfected cells, followed by detection by western blotting (Fig 1I). Even under these conditions of strong overexpression—which promote low-affinity interactions—recovery of tagged Zbtb7a was still significantly reduced using the functionally impaired mutant of TA3, confirming the strong preference of the interaction for the functional form of TA3. The interaction between endogenous Zbtb7a and p65 proteins could be detected in unmodified fibroblasts by co-immunoprecipitation (S1I Fig).
Despite its high specificity for the functional TA3 region, Zbtb7a-derived peptides were detected by conventional MS with lower coverage than those of several other proteins that interact with the full C-terminus of p65, suggesting that the interaction between TA3 and Zbtb7a may be less structurally robust under the biochemical conditions used. Therefore, to confirm the intracellular interaction of Zbtb7a and the p65 C-terminus, independently of any biochemical extraction conditions, we used bimolecular fluorescence complementation (BiFC). We expressed Zbtb7a as a fusion protein with the N-terminal fragment of the fluorescent protein Venus (V1), together with p65 fused to the Venus C-terminus (V2). Cells expressing fusion proteins individually were nonfluorescent, and only background fluorescence could be detected from cells co-expressing Zbtb7a-V1 together with control protein fragments fused to V2. In contrast, co-expression of Zbtb7a and p65 fused to complementary Venus fragments yielded strong nuclear fluorescence, demonstrating their intracellular interaction (Fig 1J). Interaction of Zbtb7a could also be detected with the isolated C-terminus of p65 (Fig 1J) or with the TA3 region alone (S1M Fig), but not with the unrelated C-termini of other NFκB subunits, cRel and RelB. To address whether the interaction with the p65 TA3 region requires co-assembly with other cellular or nuclear factors, we expressed Zbtb7a fragments in vitro using a cell-free linked transcription/translation system. We could detect a reliable binding of the zinc-finger domain of Zbtb7a to GST-TA3 and not to GST alone (even using higher bait protein amounts; S1N Fig), suggesting that the interaction is likely to be direct and independent of other specific co-interacting proteins (although we cannot rule out the participation of chaperones or other generic factors that may be present in the cell-free expression system).
Knockdown of Zbtb7a using short hairpin RNA (shRNA; described later) abolished the ability of p65 TA3 to induce expression of the endogenous Cxcl2 gene in response to TNF-α stimulation (Fig 1K) but only partially impaired its activation by full-length p65 (S1O Fig), indicating that Zbtb7a is functionally required for TA3 activity.
From these experiments, Zbtb7a emerges as a protein that interacts specifically with the functional form of the TA3 region of p65 and that is required for its context-dependent transcriptional activation. It is therefore a strong candidate as a cofactor that may be mechanistically involved in TF-dependent regulation of target promoter accessibility.
Zbtb7a (also named FBI-1, LRF, or Pokemon) belongs to the POZ-ZF family of proteins—characterised by linked POZ/BTB (pox virus zinc finger/broad-complex, tramtrack, bric-a-brac) and zinc-finger (ZF) domains—and has been previously reported to associate with promoters of several well-studied model genes (reviewed in [17]). To examine the binding specificity of Zbtb7a at the genome-wide scale, we performed chromatin immunoprecipitation sequencing (ChIP-seq) for Zbtb7a in fibroblasts. Enrichment of the Zbtb7a ChIP signal at predicted peaks was highly reproducible between independent replicate experiments, although the magnitude of the detected signal at many peaks was moderate. We could robustly detect Zbtb7a binding at thousands of genomic loci (12,861 predicted peaks; Fig 2A, 2B, 2C and 2D), which predominantly represent promoter and enhancer regions (Fig 2C). Consistent with a role in regulating p65 function, Zbtb7a is strongly enriched at p65 target promoters (Fig 2D, S2A, S2B, S2C and S2D Fig) and at p65-bound intergenic peaks (S2E and S2F Fig), and Zbtb7a is among the top-ranking factors associated with p65-bound promoters among 109 distinct genome-wide ChIP-seq datasets (S2G Fig). However, Zbtb7a is not restricted to NFκB target promoters, and indeed, a large fraction of all genomic promoters are associated with overlapping—or nearby—Zbtb7a binding (up to 17% or 41%, respectively, of all promoters in fibroblasts; Fig 2D), with a preference for guanine-cytosine (GC)-rich and CpG island (CGI)-containing promoters (S2H and S2I Fig). Thus, Zbtb7a binding represents a very prevalent event, occurring at many promoters and suggesting a widespread mechanistic role in gene regulation.
De novo motif discovery at Zbtb7a-bound ChIP-seq peaks and promoters recovered the previously described Zbtb7a binding motif, confirming the ChIP specificity ([18,19]; Fig 2E, S2J Fig). The Zbtb7a motif can be detected at up to 66% of Zbtb7a peaks, indicating that Zbtb7a binding is generally a specific, sequence-encoded feature rather than an indirect association with transcriptionally active or GC-rich genomic regions, and also revealing that Zbtb7a binding is hard-wired into many promoter and enhancer sequences.
The large number of promoters bound by Zbtb7a is much higher than those bound by many conventional TFs and is more comparable to known transcriptional cofactors (S2K Fig). Zbtb7a-bound promoters are enriched for those of expressed genes but also include those of many nontranscribed genes and of genes with very low expression levels (Fig 2F). This is not a consequence of different levels of Zbtb7a occupancy because the magnitudes of the Zbtb7a ChIP signal at bound promoters with different expression levels are comparable (S2L Fig). Therefore, consistent with a role as a regulatory cofactor, Zbtb7a binding is not alone sufficient to autonomously activate gene expression.
To identify pathways and processes associated with possible regulation by Zbtb7a, we examined gene ontology (GO) annotations of genes with Zbtb7a-bound promoters. Zbtb7a target promoters displayed highly significant enrichments for multiple GO terms (Fig 2G), prominently including processes linked to cell proliferation and apoptosis, gene regulation, and signal transduction (including inflammatory and innate immune responses, which encompass many of the NFκB target promoters, as well as several specific developmental signalling pathways [S2M Fig]). These data support the notion that the involvement of Zbtb7a in gene regulation is a feature shared by multiple pathways (in addition to that of NFκB).
We used experimental depletion of Zbtb7a coupled with quantitative genome-wide DNase-I hypersensitivity mapping to examine the involvement of Zbtb7a in regulating promoter and enhancer accessibility.
We depleted Zbtb7a in fibroblasts using stable shRNA (S3A, S3B, S3C, S3D and S3E Fig). In agreement with previous studies [20], we found that even incomplete depletion of Zbtb7a (to roughly 20%–30% of normal protein levels; S3E Fig) resulted in a gradual impairment in cell viability and proliferation. Therefore, to minimise indirect effects, we avoided long-term culture of Zbtb7a-knockdown cells and performed all analyses within approximately 5 to 10 passages. For similar reasons, we also felt that (for these experiments) the knockdown approach is preferable to analysing cell lines from Zbtb7a-knockout mice because it avoids any indirect effects that may arise from the different developmental histories of distinct, independently derived cell lines.
Using DNase-I hypersensitivity mapping, we could detect altered levels of accessibility at a significant fraction of Zbtb7a-associated promoters and enhancers upon knockdown of Zbtb7a (Fig 3A, 3B, 3E, 3F and S3F, S3G, S3H, S3I, S3J and S3K Fig). Accessibility was largely unchanged at Zbtb7a-unbound regions (Fig 3D, 3E and 3F), indicating that this is a direct effect mediated by Zbtb7a and ruling out a nonspecific role for Zbtb7a in controlling overall chromatin structure. Therefore, Zbtb7a is necessary for ongoing regulation of accessibility at many of its binding sites. Regions that exhibit Zbtb7a-dependent changes in accessibility include around 47% of Zbtb7a-bound promoters (or 9% of all genomic promoters) and comprise roughly equal proportions of Zbtb7a-dependent increased and decreased accessibility (Fig 3E and 3F), implying that additional factors or signals may determine the outcome of Zbtb7a-dependent regulation. Notably, at TA3-responsive p65 target promoters, regulation through Zbtb7a almost exclusively mediates increased promoter accessibility (Fig 3G), suggesting that particular biological pathways may predominantly utilise Zbtb7a in a uniform manner.
To determine whether Zbtb7a-dependent increases in promoter or enhancer accessibility act to enable binding of TFs, we analysed DNase-I footprints at known TF binding motifs within Zbtb7a-regulated regions. We first analysed the recognition motifs of known TFs to identify those that are enriched among the complete set of promoters and enhancers that are active in fibroblasts. Although the identities of the specific TFs that bind to each enriched motif are often uncertain (since many motifs may be bound by more than one known or unknown TF), these motifs nonetheless reflect targets of TFs that are relevant for gene regulation in fibroblasts. Many of these motifs are also enriched within the subset of regions exhibiting Zbtb7a-dependent accessibility, suggesting that these regions indeed include functional binding sites for TFs. We next mapped DNase-I cut sites across instances of each motif within regions exhibiting Zbtb7a-dependent accessibility: for many motifs, a clear and significant “footprint” of differential DNase-I sensitivity is apparent at these sequences, indicative of direct protein-DNA binding ([21,22]; S3L Fig). Thus, Zbtb7a-dependent changes to the accessibility of promoters and enhancers expose binding motifs that can be recognised and bound by secondary TFs. Finally, we directly tested the requirement for Zbtb7a for motif-binding by TFs, by analysing the magnitudes of motif footprints in Zbtb7a-knockdown cells. Indeed, for many motifs that are present within Zbtb7a-regulated regions, binding footprints are markedly reduced upon Zbtb7a depletion, indicative of Zbtb7a-dependent TF occupancy (S3M Fig). Consistent with this, Zbtb7a is positioned as one of the most upstream-acting factors in previous network analyses of TF binding [23].
To verify whether Zbtb7a-depenedent regulation of accessibility is required for p65-driven recruitment of specific secondary TFs to NFκB target promoters, we performed ChIP for 2 previously characterised factors, Cebpb and JunD [9]. In both cases, promoter-binding requires stimulation of the NFκB pathway by TNF-α as well as the activity of p65 (S3N and S3O Fig). However, we find that binding to multiple NFκB target promoters is significantly impaired or even abolished upon knockdown of Zbtb7a, confirming that Zbtb7a is required for p65-driven regulation of secondary TF recruitment.
Despite the Zbtb7a-dependent regulation of accessibility at many promoters and enhancers, however, a major fraction of Zbtb7a-bound sites do not exhibit any detectable requirement for Zbtb7a for steady-state accessibility (Fig 3C, 3E and 3F, S3H and S3K Fig; see also section 6 later). These sites include promoters of both nonexpressed and expressed genes and reveal that the simple presence of Zbtb7a is not itself sufficient to trigger changes in accessibility at all sites.
The findings above suggest a model in which promoter- and enhancer-bound Zbtb7a acts as a general cofactor, which is required—but not sufficient—for regulation of local accessibility and gene activation. Moreover, the enrichment of Zbtb7a-bound promoters for a variety of biological processes hints that Zbtb7a may be utilised by a diverse set of TFs, in addition to NFκB p65.
With this in mind, we set out to identify examples of other TFs that may exploit or utilise promoter-associated Zbtb7a in fibroblasts. As a first step, we used de novo motif prediction to search in an unbiased fashion for DNA sequences that are overrepresented at or nearby Zbtb7a-bound promoters or peaks, irrespective of their measured levels of steady-state accessibility regulation by Zbtb7a. Motif prediction was consistent across several matched background sets, including total and CGI promoters, ruling out that enrichment could be biased by the strong overlap of Zbtb7a with promoter regions. We consistently found the motif matching the consensus NFκB binding site (Fig 4A), confirming that this is a feasible approach to reveal the sequence specificities of candidate Zbtb7a-utilising TFs. In addition, we identified several other sequence motifs that are significantly overrepresented at Zbtb7a-associated peaks and promoters. Notably, a number of the most significant de novo motifs closely resembled the consensus binding sequences for known TF families (indicated in Fig 4A), including several involved in developmental and signal-dependent gene activation. We therefore addressed the possibility that one or more TFs recognising each enriched motif may utilise Zbtb7a as a promoter-bound adaptor protein for the regulation of accessibility and/or gene activation, analogously to the TA3 region of p65.
We considered motifs that also displayed preferential enrichment at sites with disrupted accessibility in Zbtb7a-knockdown cells, consistent with a functional link to Zbtb7a in fibroblasts under steady-state conditions (Fig 4A). Following the strategy illustrated in Fig 4B, starting from each de novo motif, we identified candidate TFs with matching known DNA-binding specificities (belonging to the AP1, Tead, Runx, NFκB, Cepb, and NF1 families) and assembled experimentally defined lists of functional target genes using publicly available gene expression datasets generated in mouse fibroblasts. In addition, we constructed a high-confidence list of putative direct targets of NFκB p65 by performing ChIP-seq for p65 and combining this with microarray gene expression analysis in normal and p65-knockout fibroblasts. Finally, for each candidate TF, we analysed changes in target promoter accessibility and target gene expression upon experimental depletion of Zbtb7a, both by genome-wide analyses in Zbtb7a-knockdown fibroblasts and by analysis of individual TF target gene expression by Zbtb7a-knockout cells after experimental restoration of Zbtb7a (Fig 4B).
Using this approach, we were able to identify several TFs belonging to distinct families whose target gene promoters exhibit significantly impaired accessibility in Zbtb7a-knockdown cells (Fig 4C). The fraction of TF target promoters with strong evidence for Zbtb7a-dependent regulation of accessibility ranged from 14% (c-Jun) to 21% (p65), suggesting that Zbtb7a may be utilised at only a subset of TF targets; however, this may be biased by the possible inclusion of some indirectly regulated genes in our functional target lists for each TF. Indeed, when considering only putative direct target genes for p65 defined by ChIP-seq evidence for promoter binding, this fraction rises to 34%; and by restricting the analysis to only putative direct targets whose expression could also be induced by the p65 TA3 region alone, the fraction of Zbtb7a-regulated promoters using the same cutoff rises further to 43%.
We used microarray analysis of Zbtb7a-knockdown cells to interrogate Zbtb7a-dependent changes in gene expression. Similarly to the regulation of accessibility (Fig 3E and 3F), we found that expression levels of individual genes with Zbtb7a-bound promoters could be both reduced but also increased by knockdown of Zbtb7a (S4A and S4B Fig). Zbtb7a-dependent increased expression is generally associated with increased promoter accessibility (S4C Fig), whereas the converse is true for genes that are repressed by Zbtb7a. Promoters that exhibit Zbtb7a-dependent repression in fibroblasts are moderately enriched for a small set of sequence motifs, possibly reflecting TFs that are amenable to, or participate in, repression by Zbtb7a (S4D Fig). To corroborate this, we reanalysed several publicly available gene expression datasets describing the effect of Zbtb7a deficiency: consistent with our findings, in most cases, the proportions of Zbtb7a-activated and -repressed genes were similar (S4E Fig). At Zbtb7a-dependent promoters, we found that the magnitude of Zbtb7a-regulated accessibility correlates strongly with the magnitude of Zbtb7a-dependent gene expression (S4F Fig), consistent with the notion that regulation of promoter accessibility represents a mechanistic step in gene activation.
Using the microarray data, we examined the effect of Zbtb7a depletion on the expression of the experimentally defined target genes of the Zbtb7a-utilising TFs identified above. In line with their impaired promoter accessibility, we found that the expression levels of functional TF target genes were detectably reduced in Zbtb7a-knockdown cells. Furthermore, although the magnitude of this effect was moderate for many TF target genes, it was most pronounced at the subset of target genes for each TF that displayed Zbtb7a-regulated promoter accessibility (Fig 4D).
To independently confirm the link between gene activation by cJun, Tead2, Runx2, Cebpd, and NFκB and regulation by Zbtb7a, we validated a set of Zbtb7a-dependent target genes for each TF using Zbtb7a-knockout fibroblasts. To avoid indirect effects that can arise from differently derived cell lines, we acutely restored Zbtb7a by stable transduction into the Zbtb7a-knockout cells. For all Zbtb7a-utilising TFs, selected target genes exhibited reduced expression in Zbtb7a-knockout cells compared to congenic controls, and their expression was partially or completely rescued by experimental restoration of Zbtb7a (S5 Fig).
The set of Zbtb7a-utilising TFs described here is undoubtedly incomplete because we pursued only the top-ranking motifs associated with Zbtb7a-bound regions, in a single cell type and under steady-state conditions. Nevertheless, our results demonstrate that Zbtb7a is necessary to regulate promoter accessibility and enable optimal gene expression at the target genes of a diverse but specific set of TFs.
A straightforward interpretation is that Zbtb7a is required to transduce activation signals from each TF to downstream effector complexes that mediate changes in chromatin accessibility. However, an alternative hypothesis could be that the observed Zbtb7a-dependent changes in accessibility may instead arise as a downstream consequence of transcriptional activation. To address these possibilities, we used p65 TA3 as a model whereby Zbtb7a-dependent regulation of accessibility can be separated from direct transcriptional activation.
We set up a cellular system in which normal levels of p65 expression can be stably restored into p65-knockout fibroblasts by retroviral transduction and cell sorting (S6A Fig); we then used this system to establish fibroblast cell lines expressing similar levels of normal and engineered variants of p65 (S6B Fig), which can be directly compared to each other and to the parental p65-deficient cells. Using this approach, we performed microarray analysis to examine gene activation by the p65 TA3 variant lacking direct activation domains (Fig 1A). Reconstitution with p65 TA3 is sufficient to restore activation of around half of all p65 direct target genes in TNF-α-stimulated p65-knockout fibroblasts (Fig 5A). For the majority of these TA3-responsive genes, restoration of gene expression is impaired upon simultaneous knockdown of Zbtb7a (Fig 5B), confirming the general requirement for Zbtb7a for context-dependent activation by p65 TA3. As an independent test, and to formally exclude any indirect effects of Zbtb7a knockdown on upstream NFκB pathway activation, we also assayed the ability of an artificially tethered TA3 domain to induce activation of a Cxcl2 promoter–reporter vector in Zbtb7a-knockout cells. We fused the C-terminus (CT) or the TA3 region from p65 to the DNA-binding domain from yeast Gal4 and replaced the NFκB binding motifs in the Cxcl2 promoter with a Gal4 binding sequence (Gal4-UAS). Similarly to the activities of p65 variants on the intact Cxcl2 promoter (Fig 1E), both Gal4-p65CT and Gal4-TA3 were able to induce activation of the Cxcl2 promoter containing a Gal4-UAS site in control fibroblasts, and activation of the plasmid reporter by p65CT (containing the direct activation domains TA1 and TA2) was not detectably impaired by the absence of Zbtb7a. In contrast, however, the activity of Gal4-TA3 was completely abolished in Zbtb7a-knockout cells (Fig 5C).
To separate transcriptional activation from regulation of promoter accessibility, we focused on those p65 target promoters for which the activity of p65 TA3 is insufficient to induce transcription. We observed that, at many of these “non-TA3-responsive” promoters, p65 TA3 is nonetheless able to induce measurable changes in accessibility, in a Zbtb7a-dependent fashion (Fig 5D and 5E). Therefore, these examples establish that Zbtb7a can transduce changes in target site accessibility upstream and independently of transcriptional activation. This conclusion is also supported by the ability of Zbtb7a to regulate accessibility at many enhancer regions (Fig 3F), at which transcription rates are many-fold lower than at promoters [25].
To directly address whether the interaction between a TF activation domain and Zbtb7a drives the regulation of promoter accessibility, we also reconstituted p65-knockout fibroblasts with a form of p65 TA3 bearing the function-impairing mutations that disrupt its interaction with Zbtb7a (Fig 1F, S1E Fig). In these cells, promoter accessibility induced by the TA3 mutant was significantly reduced compared with the level induced by unmutated p65 TA3 (Fig 5F). Lower levels of accessibility can still be observed at some promoters in this situation (S6C, S6D and S6E Fig); however, because the mutations in the TA3 domain reduce but do not abolish binding to Zbtb7a—reduced by approximately 5 to 7 times (based on SILAC ratio; Fig 1H) or by approximately 2 times (based on overexpression and WB; Fig 1I)—it is likely that these residual effects are driven by a remaining weak interaction with Zbtb7a. Nevertheless, these results strongly support a model whereby it is the interaction between TFs (exemplified here by p65 TA3) and Zbtb7a that triggers changes in accessibility (most likely through activation or recruitment of remodeling enzymes; see Discussion).
The strong preference of Zbtb7a for binding to promoters and enhancers (Fig 2A, 2B, 2C and 2D)—and its substantial overlap with NFκB target promoters (Fig 2D, S2A, S2B and S2C Fig) and with target sites for other TFs (Fig 4A)—strongly suggests that interaction with TFs and regulation of accessibility is performed by promoter-bound (or enhancer-bound) Zbtb7a. To test this, we disrupted the consensus Zbtb7a binding motif present in the TA3-responsive Saa3 promoter and assayed its activity in a reporter plasmid. Deletion of the putative Zbtb7a binding site resulted in significantly impaired reporter activation by p65 TA3 (S6F Fig). In a reciprocal experiment, we generated a reporter vector bearing the approximately 1 kb promoter region of the TA3-responsive Gem gene but excluding the consensus Zbtb7a motif, which is present in the first intron. In this case, reporter expression is driven by p65 TA3 only when Zbtb7a binding is experimentally restored (by recruitment to the reporter plasmid as a fusion to the Gal4 DNA-binding domain; S6G Fig). Together, these results support the hypothesis that binding to its target loci is required for Zbtb7a to interact with TFs and regulate promoter accessibility and activation.
What is the role of Zbtb7a at promoters and enhancers that are nonregulated or inactive under steady-state conditions (Figs 2F, 3C, 3E and 3F)? We hypothesised that these may represent sites that are utilised in response to distinct signals, or in different developmental contexts, but that are unbound by the appropriate effector TF(s) in fibroblasts. In this case, the presence of Zbtb7a could serve to impart responsiveness to these elements, in readiness for stimulus- or differentiation-induced TF recruitment.
To test this hypothesis, we again focused on target genes of NFκB p65 as a model. In p65-knockout fibroblasts, NFκB target genes are nonexpressed and require restoration of p65 as well as cellular stimulation to be activated. We therefore examined whether Zbtb7a is already prebound at p65 target promoters in this cellular setting. We performed ChIP-seq for Zbtb7a in p65-knockout and normal fibroblasts, both in resting conditions and after stimulation of NFκB pathway activation using TNF-α. Zbtb7a binding to inactive p65 target promoters was completely undiminished in p65-knockout cells as well as in unstimulated fibroblasts (Fig 6A). Moreover, whereas TNF-α stimulation strongly induced recruitment of p65 to its target sites in promoters and enhancers (Fig 6B), the levels of bound Zbtb7a at the same regions remained unchanged. Therefore, Zbtb7a prebinds to the target promoters and enhancers of its “client” TF p65, independently of p65 binding.
At p65 target promoters, which are normally loaded with prebound Zbtb7a, binding of p65 triggers changes in promoter accessibility and gene activation. We therefore examined whether the presence of Zbtb7a is required for, or facilitates [16], inducible p65 recruitment. We performed ChIP-seq for p65 in Zbtb7a-knockout fibroblasts, and in congenic-control cells, during TNF-α stimulation. The complete absence of Zbtb7a had no detectable effect on stimulus-driven p65 recruitment to >93% of its normal target sites (Fig 6B), ruling out its requirement for NFκB pathway activation and indicating that client TF binding can occur independently of the presence of Zbtb7a.
Despite normal recruitment of p65 to its target regions in the absence of Zbtb7a, p65 is impaired in its ability to trigger changes in accessibility upon Zbtb7a knockdown. In p65-knockout cells, the basal level of accessibility at inactive NFκB target promoters is very low (S7A Fig), but this is induced by restoration of p65 and TNF-α stimulation. Similar to our observations using the isolated TA3 region, simultaneous knockdown of Zbtb7a was able to strongly reduce the ability of reintroduced p65 to trigger increased accessibility at its target promoters and enhancers in p65-knockout cells (Fig 6C, 6D, 6E and 6F, S7 Fig). Therefore, the presence of prebound Zbtb7a at future TF target sites is required to allow newly recruited client TFs to trigger accessibility changes.
Finally, we examined the scope and influence of Zbtb7a-dependent regulation on normal p65-driven gene expression. Like most other TFs, p65 contains autonomously acting, direct transcription activation domains (TA1 and TA2), so it was pertinent to assess the extent to which indirect, context-dependent activation transduced by Zbtb7a can additionally contribute to normal TF-driven gene expression levels. As already shown above (Fig 4D, S1O Fig), overall levels of p65 target gene activation are reduced upon knockdown of Zbtb7a. Furthermore, by dividing p65 target genes according to their activation by p65 TA3 alone (Fig 5A), we found that the majority of TA3-responsive genes are strongly dependent on Zbtb7a, even when assaying the activity of full-length p65 (Fig 6G). In a reciprocal experiment, we reconstituted p65-knockout fibroblasts with a p65 variant containing only TA1 and TA2—and lacking all TA3 regions—and compared its ability to activate p65 target genes to that of intact p65. This form of p65 is able to strongly activate expression from reporter plasmids (Fig 1B and 1E; see also [9,11]). Nevertheless, more than half of TA3-responsive genes displayed significantly reduced expression when driven by the p65 TA1&2 variant (Fig 6H). In summary, TA3- and Zbtb7a-transduced regulation plays an important contribution to p65-driven activation of a subset of its target genes, which is not accomplished by its direct activation domains alone. It seems likely that many of the other TFs that utilise Zbtb7a may have a similar pattern of behaviour.
Together, our data show that Zbtb7a associates with a large set of genomic promoters and enhancers, where it acts as a prebound transducer, which is required for many TFs to induce changes in accessibility (summarised in Fig 7). Using NFκB p65 as a model Zbtb7a-utilising TF, we find that binding of Zbtb7a and of its “client” TF are independent events and that induction of accessibility is triggered by the interaction between a specific TF domain and Zbtb7a. Moreover, at the target genes for a diverse set of TFs, involved in multiple biological processes, Zbtb7a is required for regulation of promoter accessibility and for normal gene expression.
Numerous TFs—including “pioneer” factors—have been shown to be able to induce changes in accessibility at their target sites, but in most cases, it is unclear how this effect is triggered [6–8]. Our results indicate that many TFs may not themselves carry this functionality nor independently recruit and activate the remodeling enzymes required; instead, they rely on the presence of Zbtb7a as a prebound “transducer” at their target sites. This notion implies that the set of target sites amenable to regulation upon binding by a particular TF may be predetermined in a given cell type by the genomic distribution of Zbtb7a, or other “transducer” factors: this could act to control the scope of activity of TFs in a given cell type and might even represent a limitation to the possible targets of pioneer TFs that function using the same (or related) mechanisms.
Zbtb7a has previously been characterised to be involved in both repression and activation of several individual model promoters, with its described effect(s) varying according to the system studied [17]. Our global analysis of Zbtb7a binding and function is in agreement with this: we find that similar proportions of Zbtb7a-regulated promoters appear to be repressed or activated in a Zbtb7a-dependent fashion, and regulation is generally direct in both situations (based on binding of Zbtb7a and enrichment for Zbtb7a sequence motifs). We obtained the same results using independent, publicly available data (S4E Fig). However, our data indicate that Zbtb7a does not behave as a direct transcriptional activator (in line with its lack of any described activation domain)—because its presence at many nonexpressed gene promoters is not alone sufficient to drive gene expression—and that it instead functions by transducing TF-driven changes to promoter accessibility. It remains to be determined whether gene repression by Zbtb7a is accomplished using a similar mechanism. Regulation of accessibility could be envisaged to enable either activation or repression at a given promoter, depending on whether the outcome increases or decreases the availability of specific TF motifs, and according to the downstream recruitment of activator or repressor proteins. Nevertheless, at NFκB target promoters in fibroblasts, Zbtb7a-dependent changes in accessibility are overwhelmingly associated with gene activation (Figs 3G, 5B and 6G), hinting that these promoters may be preconfigured to respond in a consistent fashion. Likewise, the other TFs that we investigated also appear to utilise Zbtb7a predominantly to mediate gene activation, and a similar situation may also apply at genes activated by the TF GATA1 in erythroid cells in which Zbtb7a was previously described to preferentially occupy the promoters of GATA1-activated genes [26] (although a role in repression of other GATA-1 targets has also since been reported [27]).
The structural events that enable Zbtb7a to trigger changes in promoter accessibility remain to be determined. The POZ/BTB domain, present at the N-terminus of Zbtb7a, is a highly conserved interaction and dimerisation domain found in many metazoan DNA-binding proteins. Shortly after its original description [28,29], it was conjectured that many nuclear POZ domain proteins may represent “regulators of chromatin folding rather than direct transcriptional regulators” [30]. This notion fits well with our proposed mode-of-action of Zbtb7a. However, the molecular effectors of Zbtb7a, and indeed of many POZ-ZF proteins, are not well characterised. Most structural insights into POZ function so far have come from studies of the mammalian TF B cell lymphoma 6 (Bcl6), which binds either to NCoR/SMRT or BCoR cofactors through an interaction groove located within the dimerisation interface of its POZ domain [31,32]. Notably, though, the Bcl6-binding domains of SMRT and of BCoR do not share any significant sequence similarity to each other, and moreover, the sequence of the interacting groove within the Bcl6 POZ domain is not conserved in other POZ-ZF proteins [33]. Thus, although the POZ domain may represent a common cofactor-interaction module for POZ-ZF proteins, its binding targets can be highly protein specific, and a single POZ domain can interact with diverse cofactors. The POZ domain of Zbtb7a itself has been reported to interact with several proteins that can act as subunits of nucleosome remodeling complexes [27]. These could provide a direct, mechanistic link to the regulation of promoter accessibility, although it remains to be resolved whether they represent interactions with a single complex or with subunits of multiple, distinct complexes. Prominent among reported interactors are subunits of NuRD (nucleosome remodeling and deacetylase) complexes, which have been best characterised for their role in gene repression [34] (although some reports have found that they can associate with promoters or enhancers of active genes or even function as activators [35,36])—therefore, it remains to be established whether remodeling by variants of these complexes could also be responsible for Zbtb7a-mediated promoter activation.
A key aspect of the “transducer” model of Zbtb7a function is our finding that its gene-regulatory activity is only triggered upon interaction with other TFs, exemplified by the TA3 region of p65. How could such an interaction enable Zbtb7a activity? One plausible possibility is suggested by the known ability the POZ and ZF domains of Zbtb7a to mutually interact [37]. Our in vitro interaction experiments suggest that Zbtb7a contacts the TA3 region through its C-terminal ZF domain (similarly to other POZ-ZF proteins that have been described to interact with collaborating TFs [38–40]), implying that binding of TA3 (or other TF domains) to the Zbtb7a ZF domain may displace the prebound POZ domain, thereby allowing it to bind and recruit other effector molecules. An analogous displacement of prebound p300 has been reported upon binding of the TF Myc to the ZF domain of another POZ-ZF protein, Miz [38]. This model, or related alternatives based on competitive binding to the same domain of Zbtb7a, may also explain how mutations that reduce the strength of binding by TA3 (by approximately 2 to 7 times; Fig 1H and 1I) could result in an all-or-nothing functional outcome. Future studies will be needed to establish whether this or another mechanism underlies the transducer function of Zbtb7a.
Lastly, Zbtb7a has been strongly linked to cancer development and progression. Intriguingly, though, Zbtb7a has been found to behave both as an oncogene and as a tumour suppressor, depending on context [17]. Based on our results, it seems likely that the widespread role of Zbtb7a as a potential cofactor for multiple TFs in diverse pathways may contribute to the pleiotropic effects exhibited by cells with disrupted or misregulated Zbtb7a.
Mouse 3T3 fibroblasts used in these experiments were derived from wild-type, p65-knockout, p53-knockout, and p53-/Zbtb7a-double-knockout mice and were treated with 5 ngml−1 recombinant mouse TNF-α to stimulate NFκB pathway activation (1 hour before analysis, unless otherwise specified). For reporter assays, fibroblasts were transiently transfected by electroporation or using lipofectamine, and HEK-293 cells were transfected using CaPO4. Stably transduced fibroblast cell lines were generated with retroviral vectors using supernatants from transfected Ecotropic-Phoenix packaging cells. Reconstituted p65-knockout fibroblasts were infected with serial dilutions of retroviral supernatants and analysed by flow cytometry; samples with comparable expression levels of co-expressed Tomato protein were chosen and sorted for low and homogeneous Tomato levels, and the levels of expressed p65 variant proteins were assayed by intracellular staining using an antibody specific for the DBD of p65, present in all variants; all cell lines used expressed similar levels of p65 variants to the level of p65 in wild-type fibroblasts (see S6B Fig). Zbtb7a expression was experimentally restored in p53-/Zbtb7a-double-knockout fibroblasts by retroviral infection and sorting for low levels of co-expressed Tomato protein (corresponding to around 10- to 20-fold overexpression of Zbtb7a-encoding mRNA) (see S5A Fig). All cells were grown in DMEM supplemented with 10% fetal bovine serum (FBS). For preparation of nuclear extracts for SILAC MS analysis, dialysed (3.5 kDa cutoff) FBS was used, and lysine- and arginine-deficient DMEM was supplemented either with normal (“light”) or isotopically labelled (“heavy”: lysine 13C6 plus arginine 15N4) amino acids.
All p65 variants were derived from the coding sequence of mouse p65 and were expressed using pMY-ires-Tomato (a retroviral vector driving co-expression of red fluorescent Tomato protein) or pCDNA3. p65 DBD: amino acids (aa) 1–305; p65 TA3: aa 1–440,475–519; p65 TA1&2: aa 1–305,441–549; p65 TA3 mutant: substitution of aa 361–370 in TA3 to alanine residues. The positions of other truncations, deletions, and mutations within p65 are summarised in S1D Fig. To generate Gal4DBD fusion proteins (used in Fig 5C, S1J, S6F and S6G Figs), the DBD of yeast Gal4 (aa 1–147) was fused to the carboxy-termini of p65 (aa 306–549) or TA3 (aa 306–440,475–519), or to full-length Zbtb7a. For BiFC, proteins fused at either the N- or C-termini using a 14aa glycine/serine linker to fragments of the Venus fluorescent protein (V1: aa1–158, V2: aa159–239) were expressed using pCDNA3. Reporter plasmids containing different promoters were based on pGL3-Gfp. “NFκB motif only” promoter: 3x GGGATTCCCC motifs immediately upstream of an artificial minimal promoter (comprising fused fragments from chicken conalbumin and SV40 promoters); Cxcl2, Saa3, and Gem promoters: approximately 1 kb genomic sequences upstream of the TSS (Cxcl2: mm9 chr5:91331907-91332973(+); Saa3: mm9 chr7:53970968-53971803(-); Gem: mm9 chr4:11630585-11631714(+)). The Cxcl2 promoter was modified in some experiments (Figs 1E and 5C) by replacing the 2 NFκB binding motifs (chr5:91332828-91332859(+)) with the Gal4 UAS sequence; the Saa3 promoter was modified (S6F Fig) by disrupting the consensus Zbtb7a binding motif (chr7: 53971038-53971045(+); mutated GGGACCCC to AAGCTTCC [mutations underlined]), as well as a consensus motif within the Gfp coding sequence; a 5x tandem repeat of the Gal4 UAS sequence was cloned into the Gem promoter reporter plasmid (S6G Fig). Zbtb7a was stably knocked down using hairpins directed against either GCACAACTACGACCTGAAGAA or GAAGCCCTACGAGTGTAACAT (within the Zbtb7a CDS), cloned into the retroviral vector pSirΔ-U6CPuro (which drives hairpin expression from the mouse U6 promoter and confers resistance to puromycin). For control experiments, a hairpin directed against GGCACAAGCTGGAGTACAACT (derived from the Gfp CDS) was used. The full-length Zbtb7a coding sequence was expressed in fibroblasts using pMY-ires-Tomato. The Zbtb7a ZNF domain (aa 343–569) was expressed in vitro from pCDNA3. For production of GST fusion proteins, bait protein coding sequences (full p65 CT, p65 TA3, and p65 TA3 mutant) were cloned downstream of the GST coding sequence in pGEX-4T-1.
Polyclonal antibodies against p65 (c20; sc-372), Cebpb (c19; sc-150), JunD (329; sc-74), and the HA epitope (Y-11, sc-805) as well as monoclonal antibodies against p65 (clone F6; sc-8008) and Zbtb7a (clone 13E9, sc-33683) were from Santa Cruz biotechnology.
p65-knockout 3T3s were infected using retroviral knockdown vectors targeting either of 2 regions within the Zbtb7a CDS or a control region (within the Gfp CDS) and were selected for stable transduction using puromycin. For both hairpins targeting Zbtb7a, Zbtb7a mRNA levels were reduced by 50% to 60%, Zbtb7a protein levels were reduced by around 80%, and TNF-α-induced expression of the TA3-responsive p65 target gene Cxcl2 was reduced more than 10-fold. Zbtb7a mRNA and protein levels, and expression of Cxcl2, were unaffected in control cells using the hairpin targeting Gfp.
For analysis of the expression of individual genes, cDNA was prepared from total cellular RNA by reverse transcription using random hexamer primers, and gene expression was assayed using quantitative real-time PCR with gene-specific fluorescent probes. Gene expression was normalised to the expression level of Tbp. For microarray analysis, RNA samples were prepared and processed using Qiagen RNeasy purification kits. Sample processing and microarray hybridisation was performed using standard procedures (KFB, Regensburg).
For ChIP, fibroblasts were washed in PBS and fixed at room temperature with 2 mM di-succinimidy-glutaraldehyde (DSG) for 45 minutes, washed extensively in PBS, and fixed again with 1% formaldehyde for 15 minutes. Cells were then washed extensively in ice-cold PBS, and nuclei were released by incubation for 5 minutes in ice-cold L1 buffer (50 mM Tris [pH 8], 2 mM EDTA, 0.1% NP40, 10% glycerol) followed by 5 minutes of centrifugation at 1,000 g. Nuclei were lysed in L2 buffer (50 mM Tris [pH 8], 5 mM EDTA, 1% SDS) at a concentration of 5 × 107/ml. Chromatin in the supernatant was fragmented to a size range of approximately 300 to 700 bp using a tip sonicator, and insoluble debris was removed by centrifugation. Chromatin was diluted 10-fold in DB (50 mM Tris [pH 8], 5 mM EDTA, 200 mM NaCl, 0.5% NP40) and precleared with 27 μl/ml protein-A sepharose (for ChIP using rabbit IgG) or protein-G sepharose (for mouse and hamster IgG) for at least 1 hour at 4°C. Precleared chromatin was incubated with antibodies at a concentration of 2μg/ml overnight at 4°C, immunoprecipitated for 30 minutes using 10 μl/ml protein-A or -G sepharose, and sequentially washed 6 times with ice-cold NaCl wash buffer (20 mM Tris [pH 8], 2 mM EDTA, 500 mM NaCl, 1% NP40, 0.1% SDS) followed by 3 times with ice-cold TE (50 mM Tris [pH 8], 2 mM EDTA). Immunoprecipitated chromatin was released by incubation at room temperature in buffer EB (50 mM Tris [pH 8], 2 mM EDTA, 2% SDS), and cross-links were reversed by overnight incubation at 65°c. ChIP and input DNA were both purified using Qiagen MinElute PCR purification kits. Recovery at individual promoters was determined using quantitative real-time PCR with amplicon-specific fluorescent probes and was normalised by parallel measurement of input DNA samples. For ChIP-seq, DNA was refragmented as required to achieve a mean size of 300 bp using a water-bath sonicator. Sequencing library preparation was performed using standard procedures (genecore unit, EMBL), and samples were sequenced using single-end reads.
For mapping of DNase hypersensitive sites, fibroblasts were washed twice in ice-cold PBS, pelleted by centrifugation for 5 minutes at 500 g, and resuspended in ice-cold buffer A (15 mM Tris [pH 8], 15 mM NaCl, 60 mM KCl, 1 mM EDTA, 0.5 mM EGTA, 0.5 mM spermidine, 0.3 mM spermine, 2 mM DTT) at a concentration of 5 × 106/ml. Nuclei were released by the addition of an equal volume of buffer A plus 0.2% NP40 and incubation for 8 minutes at 4°C, then washed once in buffer A, and pelleted by centrifugation for 5 minutes at 500 g. Separate pellets of 5 × 106 nuclei were resuspended in 600 μl of 37°C digestion buffer (buffer A plus 6 mM CaCl2 plus 75 mM NaCl) and digested with varying amounts of DNaseI (typically 40, 80, and 160 units of DNaseI per 600 μl, preincubated for 3 minutes at 37°C) for 150 seconds at 37°C. Digestion reactions were stopped by the addition of 700 μl of stop solution (50 mM Tris [pH 8], 100 mM NaCl, 0.1% SDS, 100 mM EDTA, 20 mg/ml RNase A, 0.5 mM spermidine, 0.3 mM spermine) and incubation for 15 minutes at 55°C, followed by addition of 200 μg proteinase K and incubation for an additional 2 hours at 55°C. Digestions were emulsified with an equal volume of phenol/chloroform/IAA, and the aqueous phase containing digested DNA was collected after centrifugation. NaCl was added to a final concentration of 0.8 M, and digested DNA samples were loaded onto a 10% to 40% sucrose gradient and separated by overnight centrifugation at 90,000 g. Fractions were collected and analysed by agarose gel electrophoresis and staining with SYBR green, and fractions containing fragments <600 bp (with an average size of around 200 bp) were pooled and used for sequencing. Sequencing library preparation was performed using standard procedures (genecore unit, EMBL), and samples were sequenced using single-end reads.
Fibroblasts were transfected with reporter plasmids, and HEK-293 cells were cotransfected with BiFC plasmids; the fluorescence intensity of GFP-expressing cells was quantified by flow cytometry after 24 to 48 hours.
GST fusion protein baits were produced in BL21-codon plus E. coli and purified to approximately 90% purity (based on inspection of coomassie blue-stained gels) using glutathione sepharose. For pull-downs of in vitro expressed proteins, Zbtb7a fragments were expressed using T7-coupled transcription and translation (TnT, Promega) in the presence of 35S-methionine. Expressed proteins were incubated with immobilised bait proteins in NETN buffer (100 mM NaCl, 1 mM EDTA, 20 mM Tris, 0.5% NP-40 [pH 8.0] plus protease inhibitors [“complete” Roche]) at 4°C, washed 3 times in ice-cold NaCl wash buffer (250 mM NaCl, 1 mM EDTA, 20 mM Tris, 0.5% NP-40 [pH 8.0]), and bound proteins were directly analysed by SDS-polyacrylamide gel electrophoresis (PAGE). Gels were stained with coomassie brilliant blue R-250, dried, and exposed using a storage phosphor screen for up to 48 hours to detect 35S-labelled proteins. The intensity of the coomassie blue signal was used as a loading reference for bait proteins. For pull-downs of cellular proteins, nuclear extracts were prepared from HeLa S3 cells or 3T3 fibroblasts as described [41] and incubated with immobilised bait proteins for 4 hours at 4°C. Bound proteins were washed 4 times with ice-cold wash buffer (20 mM Hepes [pH 7.6], 1 mM EDTA, 10% glycerol, 0.1% NP40, 100 mM KCl) and released by incubation in elution buffer (20 mM Hepes [pH 7.6], 1 mM EDTA, 5% glycerol, 0.1% NP40, 100 mM KCl, 0.2% sarkosyl) for 2 hours at 4°C. Eluted proteins were separated by SDS-PAGE and analysed by MS. For identification of differentially binding proteins by SILAC, nuclear extracts from cells grown using “heavy” and “light” labelled amino acids were processed in parallel using different baits and mixed immediately after elution. Each nuclear extract was split and processed twice in a “label-swap” setup, so that “heavy” and “light” extracts were each used with each of the 2 baits. Four biological replicates were performed for HeLa S3 cells and for 3T3 fibroblasts, and overall SILAC ratios were calculated as the mean ratio of all replicates in which each protein was detected.
Sequencing reads were aligned to the mouse reference genome (mm9) using bowtie ([42]; with options -v 2 -a -m 5—tryhard), and statistically excess reads mapping to the exact same location (based on the local level of coverage)—representing likely PCR artefacts—were removed. Mapped read coordinates were extended to the mean fragment size of the sample DNA (300 bp for ChIP, 200 bp for DHS recovered DNA, 1 bp for identification of individual DNase cut sites), and the coverage at each genomic interval was calculated as the mean number of overlapping fragments per base pair (allowing fractional contributions from non–uniquely mapping reads) after normalising datasets to a nominal sequencing depth of 20 million reads. Enriched “peaks” were identified in combined datasets using macs1.4 ([43]; with options -p 1e-4—nomodel—shiftsize = 150 [for ChIP] or 100 [for DHS]—keep-dup = all), using input DNA sequencing data as background and with high-confidence peaks defined as those with a predicted −10log10(P value) greater than 50. Zbtb7a ChIP-seq datasets exhibited consistently higher backgrounds and lower peak heights when compared to p65 ChIP-seq datasets, likely reflecting differences in ChIP efficiency; nevertheless, high-confidence Zbtb7a peaks exhibited strong correlations between biological replicate datasets (derived from independently cultured cells; Pearson’s r > 0.87) and between parallel datasets generated from untreated and TNF-α-treated cells (Pearson’s r > 0.91), and the Zbtb7a binding sequence motif could be identified at 66% of Zbtb7a peaks, arguing that the signal is highly specific. DHS datasets were normalised (so that differences in the fraction of reads at background, non-hypersensitive regions do not affect the measurement of accessibility at DHS sites) by scaling the coverage so that the number of read-starts (corresponding to DNase cut sites) mapping under predicted DHS peaks was equal for each sample. Note that this conservative approach is valid if the mean accessibility of all DHS regions genome wide is assumed to be approximately equal in all samples; it will underestimate DHS changes between samples if the mean accessibility changes consistently at all, or a large fraction of, DHS sites.
Differences in gene expression were calculated using the means of microarray RMA-normalised signals from 3 biological replicates per experimental group, and genes defined as differentially expressed were required to be significant at a threshold of P < 0.05 using unpaired, two-tailed Student t tests between groups (based on the described normal distribution of microarray replicate measurements [44]). Subsets of promoters were defined according to criteria based on both microarray expression data and ChIP-seq data: TA3-responsive promoters were defined as those with affymetrix signal difference for (“+p65 TA3”—“p65ko”) ≥0.5 and (“+p65”—“p65ko”) ≥0.4, and with a high-confidence p65 peak within 2 kb. TA3-nonresponsive promoters were those with affymetrix signal difference for (“+p65”—“p65ko”) ≥0.5 and (“+p65 TA3”—“p65ko”) <0.4, and a high-confidence p65 peak within 2 kb. p65 direct target promoters were the union of TA3-responsive promoters and TA3-nonresponsive promoters.
Known and de novo enriched TF motifs at ChIP-seq peaks and promoter subsets were identified using Homer [45]. To analyse DNase footprints at TF motifs, motif instances were first identified within changing DHS regions using Homer, and the patterns of observed DNase cut sites at each nucleotide position surrounding each motif were divided by the mean cut frequencies at all instances of the same surrounding 6 bp sequence across all DHS regions (to avoid any influence of DNase cleavage preferences [22,46]). Note that cut frequencies can be both increased as well as reduced within the TF motif (after correction for cutting biases), likely depending on the exposure of individual nucleotides induced by TF binding, as described [21,47].
To determine the overlap of ChIP-seq data with particular genomic features (Fig 2C), promoters and transcript end sites were defined as the genomic intervals within 1 kb of RefSeq annotated transcript starts and ends, and enhancers were specified as the genomic intervals within 1 kb of the summits of fibroblast H3K4me1 ChIP-seq peaks; the remaining genomic regions were assigned as either exons, introns, or intergenic according to RefSeq annotations, and the fraction of the summits of ChIP-seq peaks that fall within each defined region were calculated.
Promoters with Zbtb7a-regulated promoter accessibility (Fig 3E, 3F and 3G, S3G Fig) were defined as those exhibiting greater changes in DNase hypersensitivity in the 1 kb interval upstream of each TSS than the changes exhibited by 95% or 99% of Zbtb7a-negative promoters, when comparing DHS datasets from normal and Zbtb7a-knockdown fibroblasts. Promoters with Zbtb7a-regulated gene expression (S4A and S4B Fig) were defined as those with an affymetrix signal difference for (“+p65 TA3”—“shZbtb7a + p65 TA3”) ≥0.5 (indicated as “+++” in Fig 6B), or (“+p65 TA3”—“shZbtb7a + p65 TA3”) ≥0.2 (indicated as “+” in Fig 6B), that were also significant at a threshold of P < 0.05 using unpaired, two-tailed Student t tests between groups.
To identify genes whose expression is controlled by TFs belonging to the AP1, Tead, Runx, Cepb, and NF1 families (Fig 4B, 4C and 4D), we retrieved publicly available gene expression datasets from mouse fibroblasts with experimentally manipulated expression of these TFs (consisting of c-Jun-knockout fibroblasts; fibroblasts with induced overexpression of Cebp-α, -β, -δ, or -ε; fibroblasts expressing a constitutively activating form of Tead2; fibroblasts with ectopic expression of constitutively active Runx1, Runx2, or Runx3; and Nf1c-knockout fibroblasts). High-confidence target genes for each TF were defined as the 200 genes exhibiting the greatest TF-dependent change in expression, which were significant at a threshold of P < 0.05 using unpaired, two-tailed Student t tests between groups. Where more than 1 member each TF family exhibited significant enrichment for Zbtb7a-regulated accessibility, the family member exhibiting the greatest fraction of high-confidence target promoters with Zbtb7a-regulated accessibility is shown in Fig 4C and 4D. Note that Nf1c target promoters did not display any enrichment for Zbtb7a-dependent regulation (suggesting that other Nf1 family member[s] or other TF[s] recognising a similar motif may account for the motif enrichment at Zbtb7a-bound regions in Fig 4A) and are therefore not shown in Fig 4C and 4D. Individual TF target genes that exhibited reduced expression in Zbtb7a-knockdown fibroblasts were independently validated by quantitative PCR in p53-/Zbtb7a-double-knockout fibroblasts, with and without restoration of Zbtb7a expression to control for nonspecific variation between independently derived cell lines.
Numerical values underlying figures are reported in S1 Data, and individual data are reported in S2 Data.
Microarray datasets generated in this study are available from the NCBI gene expression omnibus (https://www.ncbi.nlm.nih.gov/geo/) through GEO series accession number GSE97468, and sequencing data are available from the NCBI sequence read archive (https://www.ncbi.nlm.nih.gov/sra/) through the SRA study accession numbers SRP103286, SRP103318, and SRP103319.
Public datasets used to identify high-confidence targets of AP1, Tead, Runx, Cepb, and Nf1 family TFs (Fig 4B, 4C and 4D) (GSE2188, GSE11732, GSE12498, GSE15871, and GSE26205); to determine absolute mRNA levels in 3T3 fibroblasts (Fig 2F & S2I Fig) (GSE39524 [sample GSM970853]); to determine Zbtb7a-dependent increases and decreases in gene expression in diverse cell types (S4C Fig) (GSE41839, GSE70680, GSE74977, GSE24889, and GSE46473); and to define enhancer positions in 3T3 fibroblasts (Figs 2C and 3F, S2G Fig) (GSE32380).
Differences between experimental groups were analysed using two-tailed tests without assumption of equal distribution or variance; full details, including individual tests used and P values, are reported as Supporting information.
|
10.1371/journal.pgen.1007973 | Genome-wide association study reveals sex-specific genetic architecture of facial attractiveness | Facial attractiveness is a complex human trait of great interest in both academia and industry. Literature on sociological and phenotypic factors associated with facial attractiveness is rich, but its genetic basis is poorly understood. In this paper, we conducted a genome-wide association study to discover genetic variants associated with facial attractiveness using 4,383 samples in the Wisconsin Longitudinal Study. We identified two genome-wide significant loci, highlighted a handful of candidate genes, and demonstrated enrichment for heritability in human tissues involved in reproduction and hormone synthesis. Additionally, facial attractiveness showed strong and negative genetic correlations with BMI in females and with blood lipids in males. Our analysis also suggested sex-specific selection pressure on variants associated with lower male attractiveness. These results revealed sex-specific genetic architecture of facial attractiveness and provided fundamental new insights into its genetic basis.
| Facial attractiveness is a complex human trait well integrated into people’s daily life experience with profound influence on human behavior. Despite being widely studied in sociology, psychology, and related fields, its genetic basis remains poorly understood. Using carefully-measured facial attractiveness and dense genotyping data from the Wisconsin Longitudinal Study, we identified novel genes for facial attractiveness, assessed the selection signature, and dissected the shared genetic architecture between facial attractiveness and various human traits. Interestingly, sex-specific genetic architecture of facial attractiveness was a recurrent pattern observed in almost all our analyses. Our results provided new insights into the genetic basis of facial attractiveness and have broad implications for the complex relationships between attractiveness and various human traits.
| Facial attractiveness is a complex human trait of great interest in sociology, psychology, and related fields due to its profound influence on human behavior. Although variability exists across individuals and cultures, it has been suggested that some commonly agreed cues are used by people everywhere to judge facial beauty [1, 2]. As a trait that is well integrated into people’s daily life experience, it is unsurprising that facial attractiveness influences a variety of sociological outcomes. Studies have suggested that facial attractiveness is associated with job-related outcomes [3–6], academic performance [7], and economic mobility [8]. It affects human psychological adaptations and serves as an important influence of mate preference and reproductive success [9–14]. Even attractive babies receive more nurturing from their mothers than unattractive babies [15]. Further, people all over the world highly prize beauty. The annual revenue of the cosmetic industry is around 18 billion dollars in the US [16]. Fashion and beauty dominate daily discussions on traditional media as well as social media posts. Understanding the perception of attractiveness is a great interest in both academia and industry.
The genetic basis of facial attractiveness may provide new and mechanistic insights into this complex human trait. Evidence suggested that attractiveness is heritable and genetic variations explain a substantial fraction of its variability [17, 18]. However, no genetic variant or gene underlying the biology of facial attractiveness has been identified to date. Our understanding of its genetic architecture is certainly far from complete. In this study, we utilized data from the Wisconsin Longitudinal Study (WLS), a longitudinal study of a 1/3 random sample of over ten thousand Wisconsin high school graduates in 1957. Facial attractiveness in WLS was measured by 12 coders using an 11-point rating scale based on each individual’s 1957 high school yearbook photo. These well-characterized facial attractiveness data from WLS have expanded our knowledge on the complex relationship between facial attractiveness and various sociological traits including educational aspirations and occupation [19–22]. Recently, dense genotype data have been made available in WLS. These advances make it possible for the first time to identify specific genetic components associated with facial attractiveness and probe its genetic architecture.
We performed a genome-wide association study (GWAS) on 4,383 samples from WLS to identify single-nucleotide polymorphisms (SNPs) associated with facial attractiveness. In addition, sex-stratified analyses suggested distinct genetic architecture between the perception of male and female attractiveness. Integrated analysis of GWAS results and transcriptomic and epigenetic functional annotations also provided mechanistic insights into how genetics may influence facial attractiveness.
We conducted a GWAS for facial attractiveness on individuals of European ancestry in WLS. After quality control, a total of 3,928 individuals of self-reported European ancestry were included in the discovery stage. Ancestry information was unavailable for a fraction of individuals in WLS. We confirmed the European ancestry for 455 additional individuals using genetic data and used these samples to replicate genome-wide significant findings (S1 Table). In 2004 and 2008, each individual’s high-school yearbook photo was independently rated by 12 coders (6 females and 6 males) who were selected from the same birth cohort as the WLS participants. These scores were normalized into two metrics to represent the average attractiveness ratings from female and male coders on each individual (Methods). We conducted two separate GWAS on all samples using attractiveness scores given by female and male coders as quantitative traits (Figs 1A and S1 and S2). These two analyses are referred to as FC-AS (female coders, all samples) and MC-AS (male coders, all samples) throughout the paper. We identified one genome-wide significant locus at 10q11.22 for FC-AS (rs2999422; Table 1 and Fig 2A). Of note, this locus also passed a more stringent study-wise Bonferroni correction adjusted by the total number of traits we analyzed in this study. The leading SNP at this locus showed consistent effect direction in the replication cohort but did not reach statistical significance, possibly due to the limited replication sample size. However, the p-value became more significant in meta-analysis (p = 6.5e-10), showing strengthened statistical evidence. No genome-wide significant loci were detected for MC-AS. Additionally, we identified three loci showing suggestively significant associations (p<1.0e-6) with FC-AS (6p25.1) and MC-AS (20q13.11 and 2q22.1; S2 Table and S3 and S4 Figs).
Attractiveness is known to have sex-specific associations with various social factors [23–25]. Thus, we hypothesized that different genetic components may be associated with male and female attractiveness and the genetic architecture may further diverge when comparing the perceptions of male and female coders. We conducted four sex-specific GWAS (Figs 1B and S5 and S6) based on female coders and female samples (FC-FS), female coders and male samples (FC-MS), male coders and female samples (MC-FS), and male coders and male samples (MC-MS). We identified one additional genome-wide significant locus at 2p22.2 for MC-FS (rs10165224; Fig 2B). This locus also showed a consistent effect direction in the replication cohort. Meta-analysis further strengthened the statistical evidence and lowered the p-value (p = 2.3e-8; Table 1). Seven loci (i.e. 1q21.3, 5p15.31, 8q24.11, 11p15.2, 12q12, 17p13.3, and 17q11.2) showed suggestive associations in sex-stratified analyses (S2 Table and S7–S10 Figs). Of note, associated loci identified for FC-AS and MC-AS all showed consistent effects in sex-specific analyses (S3 Table). We also formally tested SNP-sex interaction effect for the two genome-wide significant SNPs. With male coded as 1 and female coded as 2, we identified significant interaction effect of sex and rs10165224 for male coders’ ratings (effect = -0.298, p = 6.7e-5). Interaction was nominally significant but much weaker when analyzing ratings of female coders (effect = -0.181, p = 0.015). These results are consistent with effect estimates in the sex-stratified analyses and suggest that rs10165422 has a significantly stronger negative effect on facial attractiveness in females than in males, especially when rated by male coders. Furthermore, we performed an X-chromosome wide association study (XWAS) to investigate sex-specific effects on the X-chromosome. However, no loci reached genome-wide significance in either sex-stratified analysis or meta-analysis (S4 Table).
A total of 80 coders participated in the attractiveness study in WLS (Methods). To investigate the heterogeneity of identified signals, we performed association analyses using attractiveness scores from each coder separately. In order to maintain statistical power in the association analysis based on ratings from each coder, we focused on coders who rated more than 500 male or female samples in WLS (S11 Fig). The genome-wide significant association identified for FC-AS, i.e. rs2999422, showed consistently negative associations with attractiveness scores from all female coders and most male coders (Fig 2C). The genome-wide significant locus in MC-FS analysis, rs10165224, also showed consistency–negative associations for all male coders except one (Fig 2D). Consistent association patterns were also observed for other identified loci (S12 Fig). In addition, we investigated the variability of ratings from different coders (S13 and S14 Figs). All tested correlations were statistically significant after Bonferroni correction. These results suggest that attractiveness ratings from different coders were mostly consistent and the associations identified in our analyses were not driven by coder biases. Rather, they represent genetic associations with the consensus of opinions among coders.
Consistent with many other complex traits [26], top associations identified in our analyses only explained a small fraction of phenotypic variability. We obtained positive estimates of chip heritability for FC-AS (heritability = 0.109, p = 0.230) and MC-AS (heritability = 0.277, p = 0.036) using genome-wide data [27]. However, we note that standard errors for these estimates were high (0.149 for FC-AS and 0.159 for MC-AS) and the GREML algorithm [27] did not converge in sex-specific analyses, likely due to limited sample size. Analysis based on a different method–GEMMA [28]–yielded similar results (S5 Table). Next, we applied linkage disequilibrium (LD) score regression [29] to partition heritability by tissue and cell type. Interestingly, several tissues related to reproduction and hormone production were strongly enriched for heritability of facial attractiveness (S6 Table). Despite not reaching statistical significance after correcting for multiple testing, testis was the top tissue for FC-AS (enrichment = 3.9, p = 0.04) and ovary was the most enriched tissue for MC-AS (enrichment = 4.5, p = 0.032), MC-FS (enrichment = 5.7, p = 0.040), and FC-FS (enrichment = 3.0, p = 0.005). Reproductive organs were not highlighted in male-specific analyses (i.e. MC-MS and FC-MS).
Further, following a strategy proposed in [30], we investigated the relationship between minor allele frequencies (MAF) and minor allele effects on facial attractiveness. We grouped SNPs with MAF between 0.05 and 0.5 into 10 equally-sized bins based on MAF quantiles. Interestingly, minor alleles with low frequencies tend to have negative effects on male facial attractiveness (Fig 3). The mean minor allele effect on FC-MS from SNPs in the lowest 10% MAF quantile was -0.005, implying very strong statistical evidence for its deviation from zero (p = 7.3e-313; two-sided t-test). SNPs in the highest 10% MAF quantile, however, did not show significantly negative associations (mean effect = -4.1e-5, p = 0.493). This hinted at selection pressure on genetic variants associated with negative male attractiveness. The selection signature in females was not as clear.
One genome-wide significant locus at 10q11.22 was identified for FC-AS. The leading SNP at this locus, rs2999422, is located in an intron of pseudogene ANTXRLP1. The closest protein-coding gene is ANTXRL (Figs 2A and S15). This locus has been previously reported to associate with skin pigmentation [31] and transferrin saturation [32]. The genes closest to the three suggestively significant loci for FC-AS and MC-AS, i.e. LRP1B, PTPRT, and LY86 (S3 and S4 Figs), have been reported in multiple association studies. Specifically, LRP1B is a member of the low-density lipoprotein (LDL) receptor family and is associated with body-mass index (BMI) [33], aging [34], and age at menarche [35]; PTPRT is associated with facial morphology [36]; LY86 is associated with waist-hip ratio [37] and hip circumference [38].
Among the loci identified in sex-stratified analyses, one locus at 2p22.2 reached genome-wide significance for MC-FS (Fig 2B). The leading SNP rs10165224 is located in an intergenic region between protein-coding gene CDC42EP3 and RNA gene LINC00211. This locus was known to be associated with height [37, 39]. Genes at the seven suggestively significant loci (S7–S10 Figs) were also associated with various traits related to facial features. Both SPON1 at 11p15.2 and NXN at 17p13.3 are associated with facial morphology [36]. NXN is also associated with vulvitis and vulva disease (MCIDs: VLV008 and VLV036). Additionally, EXT1 at 8q24.11 is associated with obesity [40] and PDZRN4 at 12q12 is associated with BMI [41] and skin pigmentation [31]. Finally, the locus at 1q21.3 contains a large LD block covering multiple genes, among which ANXA9 is associated with melanoma [42].
A few leading SNPs at identified loci are expression quantitative trait loci (eQTL) for nearby genes (S7 Table). The genome-wide significant SNP at 10q11.22 for FC-AS, rs2999422, is an eQTL for ANTXRL across various tissues (minimum p = 1.1e-8); rs17746363 is an eQTL for MED30 in skeletal muscle (p = 2.5e-6); rs2074151 and rs6587551 are eQTL in thyroid for RAB11FIP4 (p = 1.2e-5) and CTSS (p = 5.6e-26), respectively. To systematically utilize multi-tissue eQTL data and better quantify associations at the gene level, we performed cross-tissue transcriptome-wide association analyses for six facial attractiveness traits using the UTMOST method (Methods; S16 Fig) [43]. We identified four significant gene-level associations after correcting for multiple testing: SYT15 at 10q11.22 for FC-AS (p = 1.0e-6), CTSS at 1q21.3 for FC-MS (p = 9.6e-7), RPL22 at 1p36.31 for MC-FS (p = 1.5e-7), and ATAD5 at 17q11.2 for MC-MS (p = 8.9e-7). SYT15 is 700kb upstream of ANTXRL, the genome-wide significant locus for FC-AS. CTSS is located at a suggestively significant locus for FC-MS and is associated with BMI [44]. ATAD5 is 700kb upstream of the suggestively significant locus for MC-MS and is known to associate with many complex traits including height, waist circumference, hip circumference, and BMI [45–47]. RPL22 is a novel association. All gene-level associations for six attractiveness traits are summarized in S8 Table.
Next, we investigated the relationship between facial attractiveness and various related human traits. First, we tested the enrichment for associations with six dermatological traits related to skin and hair pigmentation (S9 Table) among top SNPs identified in the attractiveness GWAS (Methods). Overall, SNPs associated with female coders’ ratings were enriched for associations with hair pigmentation while SNPs for male coders’ attractiveness ratings were more strongly enriched for associations with skin pigmentation. Ten enrichment results reached statistical significance after Bonferroni correction (S10 Table). Specifically, SNPs associated with FC-AS were significantly enriched for associations with multiple hair color traits, i.e. blonde (p = 5.3e-05), light brown (p = 4.2e-4), dark brown (p = 2.8e-4), and red (p = 3.2e-4). Top SNPs for MC-AS were only significantly enriched for skin pigmentation (p = 3.8e-4). A similar preferential distinction between male and female coders was also observed in sex-stratified analyses. Top SNPs for FC-FS were enriched for associations with dark brown hair (p = 8.2e-4) and top SNPs for FC-MS were enriched for associations with blonde (p = 4.4e-5), red (p = 4.2e-4), and black hair (p = 1.7e-5). In contrast, top SNPs for MC-MS were enriched only for associations with skin pigmentation (p = 3.2e-4). No significant enrichment was observed for MC-FS.
To further reveal the polygenic relationship between facial attractiveness and other complex traits, we estimated genetic covariance between facial attractiveness and 50 complex traits with publicly accessible GWAS summary statistics which covered a spectrum of social, psychiatric, anthropometric, metabolic, and reproductive phenotypes (S11 Table). Results for all 300 pairs of genetic covariance are summarized in S12 Table. One pair of traits–female BMI (BMI-F) and MC-FS, showed a strong and negative correlation, and the p-value achieved Bonferroni-corrected significance (covariance = -0.053, p = 4.7e-5). Additionally, three other pairs of traits did not reach statistical significance but showed genetic covariance with Benjamini-Hochberg false discovery rate (fdr) below 0.1, including BMI and MC-FS (covariance = -0.035; p = 6.4e-4), high-density lipoprotein cholesterol (HDL-C) and FC-MS (covariance = -0.058; p = 8.2e-4), and total cholesterol (TC) and FC-MS (covariance = -0.063; p = 5.1e-4). Interestingly, female attractiveness traits were negatively correlated with all three BMI traits, and the correlation signal was the strongest when attractiveness was rated by male coders (i.e. MC-FS) and the BMI analysis was specific to females (Fig 4A). However, such a relationship was completely absent between male attractiveness and BMI. In fact, both FC-MS and MC-MS were positively correlated with BMI traits although the p-values were non-significant. In contrast, genetic covariance between attractiveness and lipid traits was specific to male samples, especially the FC-MS analysis in which female coders’ scores were analyzed (Fig 4B).
Furthermore, we explored if the strong genetic covariance of MC-FS with BMI-F could be explained by a causal relationship between these traits. We performed robust Mendelian randomization (MR) to infer causality (Methods). We identified a negative effect (effect = -1.05, SE = 0.62) from BMI-F to MC-FS with p = 0.099. Our results hinted at a causal effect between BMI-F and MC-FS but the validity of the relationship requires future investigation using larger samples.
Despite tremendous interests from both academia and industry, the genetic basis of facial attractiveness is poorly understood, partly due to the scarcity of well-phenotyped facial attractiveness in large-scale cohorts with genetic information. Carefully-measured human facial attractiveness, in conjunction with dense genotype data in WLS, made it possible to identify specific genetic components for facial attractiveness. In this paper, we conducted a GWAS to identify DNA variants associated with human facial attractiveness. We identified two genome-wide significant loci on 10q11.22 and 2p22.2 and highlighted several genes via eQTL analysis and transcriptome-wide association analysis. Human tissues involved in reproduction and hormone production were implicated in heritability enrichment analysis. Additionally, we identified evidence for shared genetics between attractiveness and other complex traits. Top SNPs for attractiveness were enriched for associations with dermatological traits related to skin and hair pigmentation. Via a genome-wide genetic covariance estimation approach, we identified strong evidence for shared genetic architecture of facial attractiveness with BMI and lipid traits. Of note, sex-specific genetic architecture of facial attractiveness was a recurrent pattern observed in almost all our analyses. The loci that reached significance in analyses based on all samples showed consistent effects between males and females, but sex-specific analyses revealed a list of new loci. The leading SNP at genome-wide significant locus 2p22.2 showed a significant interaction effect with sex. In multi-trait analyses, SNPs associated with female coders’ attractiveness ratings were enriched for associations with hair color while top SNPs for male coders’ ratings were enriched for associations with skin pigmentation. Additionally, only female attractiveness (especially MC-FS) showed strong and negative genetic correlation with BMI while male attractiveness was more strongly correlated with blood cholesterol levels which are known to be involved in the synthesis of testosterone and other steroid hormones [48]. Finally, variants and genes were identified for both male and female attractiveness but the selection pressure on negative associations of male attractiveness seemed to be particularly strong. These results not only provided fundamental new insights into the genetic basis of facial attractiveness, but also revealed the complex relationship between attractiveness and a variety of human traits.
This study was not without limitations. First, although WLS provided a great opportunity to study the genetics of facial attractiveness, the sample size was moderate and we did not find an external cohort to replicate our association findings due to the uniqueness of this phenotype. Although heritability of facial attractiveness has been demonstrated in twin studies [17, 18], we were unable to obtain statistically significant results on chip heritability in our study. Due to weak effect sizes, extreme multiple testing, and ubiquitous confounding, external replication and validation are critical steps in studies of complex trait genetics. In our analysis, we used 455 samples with genetically confirmed European ancestry in WLS to replicate genome-wide significant findings and performed a variety of analyses to assess the heterogeneity of identified associations, including comparing association signals between males and females as well as across different coders. The effect directions of both genome-wide significant SNPs were consistent between the discovery and replication stages and the p-values became more significant in the meta-analysis. Still, spurious associations remain a possibility and the validity of our findings needs to be further investigated using independent samples. Second, attractiveness measurements in WLS were based on high-school yearbook photos. Although it is a common practice to use photos as the basis of attractiveness measurements [11, 18, 49], our phenotyping approach does not cover every aspect of attractiveness and the results need to be interpreted with caution. In our study, each photo was rated by 12 different coders and the rating scores were consistent across coders. These results suggest the robustness of the phenotypic measurements in our study, but many questions remain unanswered. What are the roles of age, physical body shape, facial expression, and make-up in the perception of attractiveness? What is the impact of assortative mating on the genetics of attractiveness [50]? And what is the shared and distinct genetics between attractiveness and closely related facial phenotypes such as symmetry, averageness, and sexually dimorphic features [14]? These are just a handful of questions beyond the scope of this study. We also note that since each yearbook photo in WLS was rated by 6 female and 6 male coders, we were able to obtain robust phenotypic measurements based on male and female coders separately. This stratification proved critical for some analyses we conducted in this study. However, for future replication using other cohorts without sufficient numbers of male and female-specific ratings, it may be necessary to conduct additional GWAS by combining all coders’ ratings. Additionally, we note that both the raters and the people being rated were from one state that was racially and ethnically quite homogeneous and we only included samples with European ancestry in this study. Further, the yearbook photos in WLS were collected more than sixty years ago. It is unclear how well our results can be generalized to other populations, age groups, and generations. If the same high school yearbook photos were to be rated for facial attractiveness by a more ethnically or racially diverse set of raters, and if the findings were to be replicated, then the inference regarding genetic association of attractiveness would be strengthened. Nevertheless, this study was a successful attempt to pin down genetic components of human facial attractiveness. Many of these unanswered questions will be exciting directions to explore in the future. We have little doubt that robust and comprehensive phenotypic measurements, coupled with larger sample sizes from diverse populations, will further advance our understanding of this interesting human trait.
WLS is a longitudinal study of a 1/3 random sample of over ten thousand Wisconsin high school graduates in 1957. Facial attractiveness in WLS was measured based on each individual’s 1957 high school yearbook photo by 12 coders (six females and six males) selected from the same cohort in 2004 and 2008. The subjects in the photos were of the same age and the photos had the same yearbook format. In total, 80 coders were involved in the study and not all photos were rated by the same group of coders. An 11-point rating scale was used to quantify attractiveness. End-points of rating were labeled as “not at all attractive” and “extremely attractive” for 1 and 11, respectively. In this study, we used normalized average ratings from female coders and normalized average ratings from male coders as two quantitative traits for facial attractiveness. Normalization was performed in a prevailing fashion as subtracting mean and then dividing by standard deviation.
Genetic data were obtained from saliva samples in the years 2006 and 2007 using Oragene kits and a mail-back protocol. All participants provided informed consent under a protocol approved by the Institutional Review Board of the University of Wisconsin-Madison. Genotyping was conducted using the Illumina Human Omni Express Bead Chip. 713,014 SNPs were genotyped. The quality control process was previously conducted for a published GWAS on educational attainment [51]. We used genotype data imputed against the Haplotype Reference Consortium (HRC) panel. Individuals were removed if they met one of the following criteria: 1) genotype missingness rate > 0.05; 2) surveyed sex did not match genetic sex; 3) surveyed relatedness did not match genetic relatedness; 4) the individual was an outlier in respect of heterozygosity or homozygosity (F statistic > 0.03 or < -0.03); 5) the individual was an ancestral outlier–we iteratively dropped individuals with nearest neighbor z-score < -5 until no more individuals with a z-score < -5 remain. In addition to these quality control criteria, we only included individuals with available attractiveness ratings, self-reported European ancestry, and birthday between 1937–1940 in the study. SNPs were removed if: 1) call rate < 0.95; 2) Hardy-Weinberg exact test p-value < 1.0e-5; 3) minor allele frequency < 0.01; or 4) imputation quality score < 0.8. After quality control, 7,251,583 autosomal SNPs and 3,928 individuals remained in the discovery stage.
In the analysis using all samples (i.e. MC-AS and FC-AS), we applied linear mixed model (LMM) implemented in the GCTA software [52] to perform association analysis while correcting for relatedness among samples. In addition, sex, round of coding (i.e. was attractiveness rated in 2004 or 2008), dummy variables for birth year were included in the model as covariates. In sex-stratified association analyses, we applied linear regression instead of LMM due to the reduction in sample size and the consequent non-convergence of the restricted maximum likelihood algorithm and add the first two principal components into covariates. We used the prevailing p-value cutoff 5e-8 to claim genome-wide significance and 8.3e-9 as the study-wise significance cutoff to further adjust for 6 traits we analyzed. In addition, we used 1e-6 as a suggestive significance cutoff. WLS data were collected on high school graduates of the same year and distant cousins may be involved due to the study design. To adjust for family structure in linear regression analysis, we kept only one individual in each pair of samples with relatedness coefficient greater than 0.05. Relatedness coefficients among samples were estimated using PLINK [53]. After these additional quality control steps, 1,792 male samples and 2,062 female samples remained in sex-stratified analyses, PLINK was used to perform association analysis with sex, round of coding, birth year, and the first two principal components (PCs) included as covariates. We also used PLINK to test the interaction between genome-wide significant SNPs and sex using all samples. Males were coded as 1 and females were coded as 2. Significance was determined by a Bonferroni-corrected p-value cutoff (i.e. 0.05/4) which adjusted for two SNPs and two traits (i.e. attractiveness ratings based on male and female coders).
We replicated genome-wide significant findings from the discovery stage using WLS samples who did not report ancestry information but had genetically confirmed European ancestry. Scatter plot based on top two PCs for WLS and 1000 Genome samples [54] is shown in S17 Fig. Quality control procedure in the replication dataset is the same with that in the discovery stage. 455 individuals passed quality control and were used to replicate the association of rs2999422 with FC-AS, and 213 female samples were used to replicate the sex-stratified association between rs10165224 and MC-FS. We used linear regression implemented in PLINK to perform association analysis. Inverse variance-weighted method was applied to meta-analyze results from the discovery and replication stages.
SNPs on the X-chromosome were imputed against HRC panel using the Michigan Imputation Server [55]. Variants were removed if 1) missing call rates > 0.1; 2) MAF < 0.005; 3) significant deviation from Hardy-Weinberg equilibrium in women (p<1e-6); 4) imputation quality score < 0.8. 5) located in the pseudo-autosomal regions (PARs), or 6) MAF between males and females was significant (p<0.001). Additionally, individuals were removed if their reported sex did not match the heterozygosity rates observed on chromosome X. After these quality control steps, 156,615 SNPs and 3,921 samples (2,102 females and 1,819 males) were included in our analyses. We used XWAS software [56, 57] to perform sex-stratified tests on X-chromosome. We added the first two PCs calculated from autosomes as covariates to adjust for population stratification. Round of coding, birth year were also included in the model as covariates. Fisher’s method and Stouffer's method implemented in XWAS were used to meta-analyze male and female samples (i.e. FC-AS and MC-AS).
Multi-tissue gene expression and eQTL data were acquired from data portal of the Genotype-Tissue Expression (GTEx) project (https://www.gtexportal.org). We applied UTMOST [43] to perform cross-tissue transcriptome-wide association analysis for six facial attractiveness traits. We used cross-tissue gene expression imputation models trained from 44 tissues in GTEx [58]. Gene-level association meta-analysis was performed using generalized Berk-Jones test [59] implemented in UTMOST software. Statistical significance was determined using a Bonferroni corrected p-value cutoff 3.2e-6.
The GREML method implemented in GCTA software was used to estimate heritability of facial attractiveness [60, 61]. We also used GEMMA as an alternative approach to validate the results [28]. Sex, round of coding, and dummy variables for birth year were included as covariates. We applied stratified LD score regression [62] implemented in the LDSC software to perform heritability enrichment analysis and identify biologically relevant tissues for facial attractiveness. Tissues with sample sizes greater than 100 in GTEx were included in the analyses. In sex-stratified analyses, non-existent tissues (e.g. testis for females and ovary for males) were removed from the analyses. For each tissue, functional annotation was defined as regions near highly expressed genes (within 50,000 bp up- or downstream). We used median transcripts per million (TPM) as the criterion to select top 10% highly expressed genes. We then estimated partitioned heritability using functional annotation for each tissue while including 53 baseline annotations in the model. P-values were calculated using z-scores of regression coefficient as previously suggested [62].
GWAS summary statistics for six dermatological traits in the UK biobank were downloaded from GWAS atlas (http://atlas.ctglab.nl). After clumping the data using an LD cutoff of 0.1, we tested if top SNPs associated with each attractiveness trait (p < 0.05 in attractiveness GWAS) were enriched for SNPs associated with skin and hair pigmentation (p < 0.05 in the corresponding GWAS). We used hypergeometric test to assess enrichment and a Bonferroni-corrected p-value cutoff to claim statistical significance (p<0.05/36 = 0.0014). We used the GNOVA method [63] to estimate genetic covariance between complex traits. Association statistics of six facial attractiveness traits were jointly analyzed with publicly accessible GWAS summary statistics for 50 complex traits (S11 Table). Since samples in WLS were not used in those 50 published datasets, uncorrected genetic covariance estimates were reported in our analyses. Additionally, due to numerically unstable estimates for heritability, we report genetic covariance instead of genetic correlation throughout the paper.
We used MR-Egger [64] approach implemented in ‘MendelianRadomization’ R package to perform causal inference between complex traits. We selected instrumental SNP variables by applying a LD cutoff of 0.05 and a p-value cutoff of 1.0e-9. Based on these criteria, 31 top SNPs for BMI were included in our analysis.
Manhattan plots and QQ plots were generated using ‘qqman’ package in R [65]. Locus plots for GWAS loci were generated using LocusZoom [66].
Genotype data from WLS are available to the research community through the dbGaP controlled-access repository at accession phs001157.v1.p1. Phenotypic data in WLS can be accessed via the WLS data portal (https://www.ssc.wisc.edu/wlsresearch). Summary statistics for facial attractiveness are available at (ftp://ftp.biostat.wisc.edu/pub/lu_group/Projects/Attractiveness).
|
10.1371/journal.pcbi.1004604 | Model-Based Analysis of Cell Cycle Responses to Dynamically Changing Environments | Cell cycle progression is carefully coordinated with a cell’s intra- and extracellular environment. While some pathways have been identified that communicate information from the environment to the cell cycle, a systematic understanding of how this information is dynamically processed is lacking. We address this by performing dynamic sensitivity analysis of three mathematical models of the cell cycle in Saccharomyces cerevisiae. We demonstrate that these models make broadly consistent qualitative predictions about cell cycle progression under dynamically changing conditions. For example, it is shown that the models predict anticorrelated changes in cell size and cell cycle duration under different environments independently of the growth rate. This prediction is validated by comparison to available literature data. Other consistent patterns emerge, such as widespread nonmonotonic changes in cell size down generations in response to parameter changes. We extend our analysis by investigating glucose signalling to the cell cycle, showing that known regulation of Cln3 translation and Cln1,2 transcription by glucose is sufficient to explain the experimentally observed changes in cell cycle dynamics at different glucose concentrations. Together, these results provide a framework for understanding the complex responses the cell cycle is capable of producing in response to dynamic environments.
| The cell cycle is an exquisitely tuned process, alternating between states of cell growth, DNA replication, mitosis, and cytokinesis. While this process is robust, it is also responsive to diverse environmental signals. For example, cell cycle events may be delayed or advanced in response to changes in temperature or nutrient availability. While the molecular mechanisms underlying cell cycle progression have been well-characterised, how these mechanisms are perturbed by a cell’s environment is still not well understood. This problem is made difficult by the dynamically changing nature of the cell cycle itself. In this paper, we tackle this issue by performing a meta-analysis of mathematical models and experimental data describing cell cycle progression in the budding yeast, Saccharomyces cerevisiae. This shows how the timing of perturbations relative to the cell cycle stage (e.g. during DNA replication or mitosis) can give rise to qualitatively different responses. By looking for consistent patterns across multiple models and experimental datasets, we demonstrate how known molecular mechanisms change cell cycle behaviour in different nutrient conditions. This also allows us to make predictions for novel behaviours that can be experimentally tested in the future.
| The cell cycle is the process by which cells alternate replication of their DNA with cell division. As a central process in the life of a cell, it is subject to multiple forms of regulation. These range from hormonal and growth factor signals in higher organisms, down to nutrient and stress signals in micro-organisms. While there has been much progress in understanding the mechanisms driving cell cycle progression, a system-level understanding of how signals regulate this progression has been lacking. In this paper, we investigate the dynamic response of the cell cycle to perturbations. In particular, we apply a combination of computational and mathematical analyses to study how the cell cycle of a particular model organism—the budding yeast Saccharomyces cerevisiae—responds to changes in conditions.
The progression of the cell cycle in S. cerevisiae, as in all eukaryotic cells, can be divided into four phases: the G1, S, G2, and M phases. The G1 and G2 (“gap”) phases mark the pauses between the essential processes of DNA duplication (which occurs during S phase) and segregation (which occurs during M phase). Several checkpoint mechanisms regulate progression through the cell cycle. These checkpoints ensure that progression through the cell cycle occurs only when the cell is in a suitable environment, and has adequately completed the previous stages of its cell cycle. For example, cells in G1 (with unreplicated DNA) must pass a checkpoint, regulated by factors such as nutrient availability and cell size, to go into S phase and begin synthesising DNA [1, 2]. Similarly, cells in the G2 phase must pass through checkpoints to enter mitosis (e.g. the spindle assembly checkpoint).
In S. cerevisiae, progression through the cell cycle is coupled to changes in cell morphology and growth, as depicted in Fig 1. After birth, the cell grows isotropically during the G1 phase. The duration of this phase is strongly correlated with the size of the cell as a result of a “cell size checkpoint” [3]. Beyond this checkpoint, the cell is allowed to pass into S phase. Upon entry into S phase, DNA replication begins, the cytoskeleton is polarised, and a bud forms [4]. Cell growth continues, but with growth directed to the bud. The cell then passes through the G2 and M phases and begins the process of cytokinesis. This results in the bud splitting from the mother cell, producing a new daughter cell.
The prevailing view of the molecular mechanisms underlying cell cycle progression is one of interlocking positive and negative feedback loops which trigger a cascade of transitions in the appropriate sequence [5]. One of the central components in the cell cycle network is cyclin-dependent kinase (CDK), named Cdk1 in S. cerevisiae. A pre-requisite for the kinase activity of CDK during the cell cycle is the presence of cyclins. Different cyclins are expressed in different phases of the cell cycle and lend specificity to the CDK-cyclin complex, allowing regulation of many transcription factors and other processes [6, 7]. These cyclins may be broadly divided into different classes depending on the timing of their expression. For example, G1 cyclins are responsible for the transition from G1 into S phase, while mitotic cyclins are responsible for the transition from G2 into M phase. The abundance of cyclins is regulated at the levels of transcription, translation, and degradation. In addition, the CDK-cyclin complex may be rendered inactive by binding to a stoichiometric inhibitors such as Sic1 [8]. Ultimately, the cell cycle completes when CDK activity is reduced by the degradation of cyclins by the Anaphase Promoting Complex (APC) [9]. This allows the cell to progress through anaphase and cytokinesis. An illustration of this dynamic progression is shown in Fig 1C and 1D. As shown, the cell is born with low but increasing levels of G1 cyclins. When the level of G1 cyclin reaches a threshold, S-phase is initiated. Levels of G1 cyclins then decrease, with a complementary increase in mitotic cyclins maintaining Cdk activity. After sufficient time for progression through mitosis and the satisfaction of additional checkpoints, CDK inhibitors and components responsible for cyclin degradation (such as the APC) become active, along with the phosphatase Cdc14, which dephosphorylates Cdk substrates. This rapidly depletes CDK activity, allowing cytokinesis to occur and a new cell to be produced.
The distinct morphology of S. cerevisiae—in particular the correspondence between the initiation of S-phase and the appearance of the bud—means that it has been a useful model organism for the study of the cell cycle. A number of environmental cues have been found to regulate cell cycle progression in S. cerevisiae. For example, addition of glucose to cells growing in ethanol increases the average size of the cells at bud initiation and reduces the duration of the cell cycle [10]. The cell cycle is also responsive to changes in other nutrient signals [11–16], growth [1, 17, 18], osmotic stress [19, 20], and temperature [21]. In addition, under certain conditions the cell cycles of a population of cells can spontaneously exhibit partial synchronisation with an oxidative metabolic cycle [22, 23]. Finally, experiments in which cyclin expression is inducible by an external signal have demonstrated the possibility of mode locking the cell cycle to a periodic stimulus [24]. The responsiveness of the S. cerevisiae cell cycle to environmental conditions is a generic property of the eukaryotic cell cycle.
Despite the rapid accumulation of knowledge of the molecular details of the cell cycle mechanism and its regulation, such are the number of pathways and the complexity of the cell cycle itself that it is difficult to predict a priori how the system will respond to changes in conditions. As a result, it is also difficult to evaluate and interpret experimental observations and determine whether an observed phenomenon can be accounted for by known regulatory mechanisms. To this end, mathematical modelling approaches are useful to investigate hypotheses about cell cycle regulation. Models describing the dynamics of essential cell cycle components have existed for some time [25], and have reached high levels of molecular detail [26–31]. These models describe the interactions between key regulators of cell cycle progression, and formalise the understanding accumulated over decades of fundamental cell cycle research.
In this paper, a framework is developed for the investigation of the dynamic regulatory capabilities of cell cycle models, and by extension the cell cycle itself. This framework consists of exhaustive computational sensitivity analysis, allowing evaluation of how the cell cycle might respond to changes in conditions, both dynamically and after a sustained change in conditions. While the cell cycle is a highly nonlinear system, we note that similar approaches using sensitivity analysis of complex biological systems have been applied successfully before, e.g. in the study of circadian clocks [32, 33]. We apply this analysis to three models of the S. cerevisiae cell cycle [30, 34, 35]. This allows several key questions about cell cycle regulation to be addressed, focusing on understanding the interaction between the cell cycle and the key developmental transitions of S. cerevisaie (Fig 1). For example: to what extent can key cell cycle characteristics such as period and size at division be regulated independently? What qualitative behaviours can be observed in the response of the cell cycle to a sudden change in conditions? How flexible can this dynamic response be for a given eventual change in behaviour?
In this section, we describe the mathematical models under investigation and the parametric sensitivity analysis of these models. We begin with a basic phenomenological description of the budding yeast cell cycle, following [24]. This describes the phenomenology of cell cycle progression, rather than the biochemical details. Specifically, under some simple assumptions about the growth of the cell, it is possible to interrelate macroscopic cell cycle properties such as daughter cell size, cell cycle duration, and cell size at budding. This mathematical description then provides the orientation and basic framework for understanding the three detailed models that follow. These detailed models consist of ordinary differential equations (ODEs), and include both S. cerevisiae-specific models and a general model adapted here for use with budding yeast.
The underlying models of the regulatory network are coupled to this basic description of growth in two key ways, summarised in Fig 1. First, cell size affects cell cycle dynamics in each model by regulating the synthesis of molecular components. Second, the cell cycle dynamics determine the timing of budding and division through thresholds on the molecular components representing G1 and mitotic cyclins, respectively (Fig 1C). These interactions are described in detail for each model in S1 Text.
All of the models considered here share the same basic behaviour, with a continuously growing cell alternating between division and budding. The volume of the cell at budding and division, and the duration of cell cycle phases, constitute a simple description of the dynamics. Following [24], this model incorporates the assumptions that growth is exponential [3] (growing at an exponential rate μ), that all growth after budding is localised to the bud, and that the daughter cell receives all of the volume of the bud. The variables of interest are the cell cycle period of a daughter cell (i.e. the time from birth to division, denoted Tdiv), the time from birth to budding (i.e. the duration of the G1 phase, denoted TG1), the size of the cell at division (denoted Vdiv), at budding (denoted Vbud), and the initial size of the daughter cell (i.e. the size of the daughter cell at birth, denoted Vdau), and the fraction of the cell volume given to the daughter cell after division (denoted f). At constant growth rate, these variables are interrelated according to the following expressions:
V d i v = V d a u e μ T d i v V b u d = V d a u e μ T G 1 f = V d a u / V d i v (1)
Note that the underlying molecular models control the timing of budding and division events, with the result that the fraction f is an emergent property of the models rather than a parameter. Similarly, Tdiv is determined by the dynamics of the underlying models, and is in general be different from the mass doubling time (MDT), TMDT, which depends only on μ (TMDT = ln(2)/μ).
All models considered here give a pattern of behaviour that can be related directly to this simple description, after slight alteration to include a budding event where appropriate. The differences between the cell cycle models thus arise from the quantitative details of their structure and their parameter values. While more detailed models of the coupling between cell cycle to growth and metabolism have been suggested [36], the above description is an adequate minimal representation for the purposes of our investigation.
In this section, the models analysed are described, with model equations presented in S1 Text. The number of variables and parameters used in each model are also given. For more complete descriptions of these models, reference should be made to the corresponding papers. The three models are presented in order of increasing complexity, from a minimal model due to Pfeuty and Kaneko [34] (referred to here as the Pfeuty model), through a modified version of the Chen model [26, 35] (referred to here as the Chen model), and a more recent model incorporating detailed representations of multisite phosphorylation [30] (referred to here as the Barik model).
The simplified (Pfeuty) and detailed (Chen, Barik) molecular cell cycle models play complementary roles in the analysis. In particular, the simplified model demonstrates the range of behaviours possible with a minimal description of the molecular interactions. Thus, behaviours that are identified across all three models are unlikely to have arisen from a special combination of parameter choices and model structure. The detailed models, in turn, provide complementary insights as they contain explicit representations of important molecular regulators (e.g. cyclins). This allows specific hypotheses about regulation to be investigated (e.g. in the case of glucose signalling, below). The Chen and Barik models share several essential and well-established features with each other, for example the distinct roles of different cyclins in determining progression through different cell cycle checkpoints. In addition, the Barik model incorporates several additional mechanistic features that have been discovered more recently, such as the role played by Whi5 in progression through the G1/S transition [37–39].
All three models consist of systems of ODEs. This formalism is useful in this context as it provides a level of detail that allows investigation of how incremental changes in parameters change the system behaviour. Furthermore, a straightforward framework exists for the calculation and interpretation of sensitivity analysis of ODE models. It should further be noted that models that just consider particular phases of the cell cycle (e.g. models of the G1-S transition [20, 29] or mitosis [40, 41]) are not suitable for investigation here, since they cannot be run across multiple cell cycles.
Sensitivity analysis provides a straightforward way of understanding how combinations of parameter perturbations change cell cycle behaviour. In particular, we can approximate changes in behaviour (in the linear regime) by the linear combination of changes elicited by each perturbation, following [42]. For example, in the case of changes in Vdau in generation i following a perturbation in parameter k at time t, we have:
Δ V d a u , i ( t ) = Δ k k S k V d a u , i ( t ) (8)
Thus, for changes in multiple parameters k1, k2, …, kn, we have:
Δ V d a u , i = ∑ j = 1 n Δ k j k j S k j V d a u , i ( t ) (9)
An assessment of the accuracy of this approximation to changes in model behaviour away from the basal parameter set is shown for 8 parameters in the Barik model in S2 Fig. While the approximations are generally good, the highly non-linear nature of the model dynamics means that the range of parameter values for which this approximation is accurate is limited in some cases. However, even in these cases the qualitative changes in behaviour are matched across a wide range of parameter values. This demonstrates the utility of sensitivity analysis for understanding changes in model behaviour in a wide regime of parameter space.
We begin by evaluating the steady-state parameter sensitivities of the models, focussing on the macroscopic observable quantities such as the cell cycle duration (Tdiv) and cell volume at division (Vdiv). First, we note that, for a particular growth rate, the macroscopic cell cycle observables can be calculated in terms of only Tdiv and Vdiv. For example, for Vdau and TG1:
V d a u = V d i v e - μ T d i v T G 1 = - 1 μ ln ( V d a u ( V d i v - V d a u ) ) (10)
As a result, the sensitivity of the cell cycle to changes in parameters can be understood in terms of changes in Tdiv and Vdiv alone (or, equivalently TG1, Vdau). This makes it natural to visualise the distribution of sensitivities in 2-dimensional scatter plots for each model, with each parameter shown as a point with position ( C k T d i v , C k V d i v ) (or, similarly, ( C k T G 1 , C k V d a u )). This as shown in Fig 2A. This allows comparison across models of the properties of particular parameters, and identification of general trends across many parameters and models.
Some parameters of particular interest are those representing the regulation of cyclin synthesis and degradation. For example, the G1/S-specific cyclin Cln3 controls the timing of Start. Cln3 activity has been hypothesised to increase with cell size, and to therefore communicate cell size information to the cell cycle [47, 48]. Parameters representing the synthesis of Cln3 are present in both the Chen and Barik models, and an analogous parameter can be identified in the Pfeuty model (see S1 Text for details). As can be seen in Fig 2A, increasing the rate of synthesis of Cln3 acts to reduce the cell size in all three models, consistent with its role in cell size sensing. While changes in Vdiv are consistent across models, Tdiv is sensitive to changes in this parameter only in the Chen model.
Other species of interest are the mitotic cyclins. Mitotic cyclins increase through the G2-M transition, and are rapidly degraded by the APC upon exit from mitosis [49]. Parameters representing the synthesis of mitotic cyclins and the synthesis of APC subunits Cdc20 and Cdh1 are present in the Chen and Barik models (analogous components are not present in the Pfeuty model; see S1 Text for details). As can be seen in Fig 2, in both models these parameters act primarily to change Tdiv in opposing directions, with increased mitotic cyclin levels leading to a longer cell cycle period. While this is consistent across models, it should be noted that the changes in Vdiv predicted by the models are not.
Apart from the molecular species represented in the models, all three models also naturally include a parameter that specifies the growth rate (named μ by convention). In all three models, increasing the growth rate reduces the duration of the cell cycle, and increases the size of the daughter cell (S3 Fig), in agreement with experimental observations [10, 12, 50, 51]. This qualitative agreement has previously been noted for other cell cycle models [28]. In summary, it is clear that the models broadly agree on some, but not all, qualitative features of regulation by particular parameters.
Beyond specific parameters, it is also interesting to look at patterns observed across all parameters. It is clear from Fig 2A that in all three models most parameters act to modulate Vdau and TG1 in opposite directions (with a few clear exceptions in the case of the Chen model). This is quantified in Fig 2B. As a result, most combinations of parameter perturbations are expected to either increase Vdau and decrease TG1, or vice versa. This suggests that, for cells growing at the same rate under different conditions (i.e. with different environmental cues perturbing cell cycle components), Vdau and TG1 should be negatively correlated. A dataset that is useful for evaluating this model prediction was generated by Brauer et al. [12]. In their experiments, cells were grown in chemostats at 6 different growth rates (0.05, 0.1, 0.15, 0.2, 0.25, and 0.3 h−1) under 6 different nutrient limitations (glucose, nitrogen, phosphate, sulphur, leucine, and uracil). Average cell volume (denoted V ¯, proportional to Vdau) and the fraction of unbudded cells (denoted FG1, proportional to TG1 (see S1 Text for derivation)) were measured. Analysis of these data reveals a negative correlation between V ¯ and FG1 at all 6 growth rates, as shown in Fig 2C. Similarly, a recent study by Soma et al. measured Vdau, TG1, and μ for various strains under different conditions [46]. Selecting those experiments for which μ was within a 0.02 h−1 window, a clear negative correlation between Vdau and TG1 is again observed (Fig 2D). Finally, recently a high-throughput screen of cell cycle behaviour by Soifer et al. measured Vdau and TG1 in a range of mutants [52]. Considering only those mutants which were classified as having wild-type growth rates, this correlation was again observed (S4 Fig). The consistency of the qualitative behaviour of all three models with these experimental data suggests that they share essential dynamics that correctly describe cell cycle progression.
While the steady-state sensitivity analysis allows the characterisation of cell cycle models under constant conditions, it is also interesting to ask how the cell cycle responds to dynamic changes in parameters. Dynamic sensitivity analysis allows us to understand the complex dynamic behaviour which the cell cycle is capable of producing on its own. This provides a foundation for understanding how signalling networks with their own complex dynamics interface with the cell cycle.
As detailed above, dynamic sensitivity can be characterised by the change in cell cycle characteristics down generations to a sustained step change in a parameter, starting at a particular time t. By way of example, the sensitivity of daughter cell size and the combined duration of the S/G2/M phases (TS/G2/M) to changes in the rate of synthesis of mitotic cyclin in the Barik model (specified by the parameter ks,bM) are shown in Fig 3. In this example, the sensitivity functions S k V d a u ( t ) and S k T S / G 2 / M ( t ) are evaluated at two different times—one early (t = 30), and one late (t = 104) in the cell cycle (Fig 3B and 3C). This illustrates the changes in Vdau and TS/G2/M that follow step changes made at these times. Two characteristics are apparent in this example, and are seen frequently in many parameters across all models: the dependence of the response on the timing of the perturbation, and the non-monotonic dynamics of this response. This sensitivity can also be visualised as a continuous function of the time of perturbation, as shown in Fig 3D.
As before, it is also instructive to consider the biological significance of this particular example. First, the qualitative characteristics of the response change depending on the time at which the perturbation is applied. Increasing mitotic cyclin synthesis early in the cell cycle reduces TS/G2/M and Vdau in all subsequent generations, as compared to the initial state (Fig 3B). However, increasing mitotic cyclin synthesis at the end of the cell cycle increases TS/G2/M and Vdau in the short term (Fig 3C). This can be understood by the role played by mitotic cyclins: their level must first increase to initiate mitosis, but must then decrease to allow the cell cycle to restart. Increasing mitotic cyclin synthesis at a time when cyclin levels need to decrease might be expected to temporarily delay cell cycle progression, as demonstrated by this sensitivity analysis. While this sensitivity analysis is qualitatively consistent with known biology, we note that an assessment of how mitotic cyclins drive the cell cycle in S. cerevisiae found that the models mis-predicted the quantitative extent of this sensitivity [53].
In summary, dynamic sensitivity analysis provides a useful tool for understanding the range of behaviours which the cell cycle is capable of producing. In all three models considered here, nontrivial dynamic behaviours were identified, including nonmonotonic changes in cell size down generations.
It has been observed qualitatively in many studies that the duration of the G1 phase of the cell cycle is especially sensitive to changes in conditions. This manifests itself in a change in the fraction of unbudded cells in populations [10, 50]. It has also been observed that cells subjected to stresses transiently arrest the cell cycle at the G1/S-phase transition, without undergoing budding [54–58]. As a result, there has naturally been significant interest in understanding how signals determine progression through this transition. In this section, we investigate how the duration of the G1 phase changes under parameter perturbations of the models.
We begin by asking how changes in the duration of the pre-budded (TG1) and post-budded (TS/G2/M) phases of the cell cycle are related to one another in the phenomenological model (Eq 1). From this, we identify the relationship:
C k T S / G 2 / M = - f C k T G 1 (11)
Where f denotes the fraction of cell mass taken by the daughter cell upon division (see S1 Text for derivation). This demonstrates how parameter changes which alter the duration of the pre- and post-budded phases of the cell are fundamentally coupled to one another in the model. Furthermore, it shows that the magnitude of changes in the duration of the S/G2/M phases of the cell cycle are expected to be less than half the change in the duration of the G1 phase (since f ≤ 0.5, both in silico and in vivo [24]). This relationship is depicted for all three models in Fig 4A.
One counter-intuitive consequence of this is that changes in parameters affecting cell cycle progression during S/G2/M will modify TG1 more strongly than they modify TS/G2/M. Therefore, at steady-state, the duration of a particular phase of the cell cycle may be altered by perturbations that act during other phases. In particular, perturbations affecting processes during the G2/S/M phases will alter the duration of G1.
As discussed above, it is also commonly observed that moving cells into a stress condition can result in a transient accumulation of cells in G1 before the cell population eventually returns to its original state. At the single-cell level, this corresponds to a transient increase in TG1. One interpretation of this behaviour is that the cells take time adapt to the stress, during which cell cycle dynamics are perturbed, before the cells eventually return to their original state (and their original cell cycle behaviour). In the context of the analysis presented here, this would be analogous to changing model parameters for some time (while the cells are experiencing stress) before returning them to their original values (after the cells have adapted to the stress). However, we previously noted that a step-change in parameters can result in complex cell cycle dynamics, including transient changes away from the eventual behaviour. This was observed in the examples of S k T S / G 2 / M and S k V d a u given previously (Fig 3), and is also true of changes in TG1. Since growth rate is held constant in these simulations, this behaviour is not the result of temporary changes in growth rate that might also be expected to accompany some changes in conditions. The prevalence of this behaviour can be quantified by calculating fraction of time which the sensitivity functions S k T G 1 , i ( t ) display a nonmonotonic sensitivity down generations, averaged across all parameters, with all models displaying at least 80% nonmontonic responses (Fig 4D). This suggests that transient responses of the cell cycle to changes in conditions must be interpreted with some caution. There are cases in which transient signalling appears to give rise to transient changes in cell cycle dynamics (e.g. [20]). However, the models suggest that transient signalling or changes in growth rate are not required for this behaviour to be observed.
In conclusion, these results demonstrate two causes for caution in the interpretation of changes in cell cycle dynamics in different conditions. First, that in cases where cells are grown under constant conditions, it is difficult to identify the cause for a change in cell cycle timing. This is because the duration of one cell cycle phase might change significantly as a result of regulation occurring during a different phase. Second, that transient changes in the duration of the G1 phase are a generic property of these models, and do not imply that signalling to the cell cycle is itself transient.
The core yeast cell cycle oscillator interacts with other cellular oscillators, including the yeast metabolic cycle (YMC) [22], and is postulated to entrain slave oscillators such as oscillations in Cdc14 activity [59] and a transcriptional oscillator [60]. In addition, it is possible to partially mode-lock the cell cycle to an external periodic signal [24]. In other organisms, additional oscillator interactions have been identified, for example gating of cell cycle transitions by circadian clocks [61–63]. In this context, it is interesting to ask how dynamic perturbations alter the timing of cell cycle events. This has been investigated previously in the context of cell cycle responses to periodic forcing signals [64, 65]. Here, we are able to link control of cell cycle timing to the modulation of macroscopic cell cycle variables. The phase shift, Δϕ, resulting from a perturbation applied between the times t1 and t2 is given by its resultant effect on Tdiv down generations:
Δ ϕ = ∑ i = 1 ∞ ( T d i v , i - T d i v , 0 ) (12)
This can also be calculated according to:
Δ ϕ = Δ k k ( S k p h a s e ( t 1 ) - S k p h a s e ( t 2 ) ) (13)
This enables us to predict the mode-locking behaviour of the cell cycle to periodic forcing. For example, for a given periodic perturbation of the parameter sx,2 in the Pfeuty model, analogous to stimulating Cln3 synthesis, the phase response is predicted to mode-lock the cell cycle so that the stimulus occurs ∼39 minutes after cell birth (see S5 Fig). This prediction is borne out by simulations, with some error (∼7 minutes, S5 Fig).
The phase shift between two cells can be related to differences in the mass fraction donated to the daughter cell down generations. In particular, consider a perturbation which causes a temporary change in the fractions of mother cell volume donated to the daughter cell. Denote the initial fraction f0, and denote the deviation from this in the ith generation Δfi. Then the phase shift is given by:
Δ ϕ = 1 μ ln ( ∏ i = 1 ∞ f 0 + Δ f i f 0 ) (14)
(see S1 Text for derivation). In practice this limit converges rapidly (within a few generations). This establishes a link between how a parameter changes the mass of daughter cells, and how it changes the phase of the cell cycle. This correspondence is demonstrated in Fig 5A and 5B. We note that this is independent of any details of the models considered here, and applies to any asymmetrically dividing cell growing exponentially at a constant rate.
One notable qualitative feature of some of these phase response curves is that they are biphasic (i.e. both phase advances and delays are possible, depending on the timing of a perturbation). This property can be quantified for a parameter k by the metric Bk:
B k = 1 - | ∫ 0 T d i v Z k p h a s e ( t ) d t | ∫ 0 T d i v | Z k p h a s e ( t ) | d t (15)
This gives values of Bk ranging between 0 and 1. Bk is 0 for a completely monophasic pattern of sensitivity, as Z k p h a s e ( t ) is strictly positive or negative, so | ∫ 0 T d i v Z k p h a s e ( t ) d t | = ∫ 0 T d i v | Z k p h a s e ( t ) | d t. The distribution of Bk across the parameters of all models are shown in Fig 5E. From this, it is clear that many parameters in all models display this property. This is a property shared with other biological oscillators, for example circadian and neuronal oscillators [34, 66].
Another observation that can be made is that the phase shifts are most pronounced when perturbations are applied later in the cell cycle (from TG1 onwards). The distributions of the times of peak sensitivity of the parameters of all models are shown in Fig 5F. In all models there are two main groups of parameters—those peaking around TG1 and those peaking around Tdiv—with very few parameters displaying peak sensitivity before TG1. This is somewhat counter-intuitive given the noted sensitivity of TG1 to parameter changes (see above). The robustness of the cell cycle model behaviour to perturbations during G1 has been observed previously in the case of the Chen model [43]. In summary, these results show that the cell cycle models consistently predict a preponderance of biphasic phase response curves, and further illustrate the qualitative differences in sensitivity observed before and after the G1-S transition.
The analysis presented above provides a framework for understanding the effects of perturbations on the dynamics of cell cycle progression. In order to demonstrate how the analysis presented can be applied to understanding signalling to the cell cycle, it is useful to consider a specific example. Here, we investigate how glucose-sensing signalling pathways might affect cell cycle progression. Glucose sensing is particularly important in this context, as the extra- and intracellular glucose levels are key determinants of nutrient availability. As such, several pathways have been identified through which glucose affects cell cycle components, both through direct sensing [13, 14, 16, 67, 68], and indirect effects via metabolism and growth rate [15, 16]. Here, we consider the effects of direct signalling pathways, and note that their effects can be separated from indirect, growth-rate-mediated effects in conditions where growth rate does not change in response to glucose levels. An example of this was recently demonstrated in experiments by Soma et. al in which changing glucose concentrations in the range of 0.05% to 2% had no effect on growth rate but did perturb the cell cycle [46].
We consider three particular forms of cell cycle regulation by glucose (Fig 6A). The first mechanism of cell cycle regulation by glucose involves the control of translation of Cln3—a cyclin responsible for inducing G1-S transition. The regulation of Cln3 translation is mediated in part through the direct regulation of the translation initiation factor eIF4E [69], and can also be controlled through the relief of competition for translation initiation factors (e.g. due to rapid degradation of GAL1 transcripts in the transition from galactose- to glucose-driven growth [70]). The rate of translation of Cln3 is represented in the Barik model by the parameter ks,n3. The second mechanism we consider is the repression of Cln2 expression by glucose [71]. In the Barik model, Cln2 falls within the class of G1 cyclins, denoted by ClbS. The rate of ClbS transcription is represented by the parameter ks,mbS. Finally, it is known that signalling through the TOR kinase complex is capable of modulating the activity of the PP2A phosphatase complex [72, 73]. Upon phosphorylation by the TOR1C complex, this phosphatase dephosphorylates a wide range of targets, including Net1 [74]. Net1, in turn, is responsible for sequestering the cell-cycle phosphatase Cdc14, which is required for progression through mitosis. The dephosphorylation of Net1 in the Barik model is represented by the constitutive activity of a generic phosphatase, Ht1. The model parameters representing this activity are kd,t1 and kd,nt, regulating free Net1 and Net1 in the RENT complex, respectively. A natural assumption is that regulation of this pair of parameters is coupled, and therefore that they are modulated proportionally to one another.
The above summary of some regulatory mechanisms is by no means complete, partly as a result of some regulatory components not being present in this model (e.g. the regulation of Cdk1 phosphorylation by Cdc25 and Swe1 [75]). However, since it includes components involved in regulating different cell cycle phases, it provides a useful starting point for understanding the range of behaviours that might be achieved by glucose regulation of the cell cycle. The effects of these regulatory mechanisms can be summarised by the following constraints on the parameter perturbations applied through this pathway (where the “signal”, assumed to be proportional to the availability of glucose, is represented by G, and the sensitivity of a parameter k to changes in G is denoted R G k):
R G k s , n 3 = 1 k s , n 3 d k s , n 3 d G > 0 R G k s , m b S = 1 k s , m b S d k s , m b S d G < 0 R G k d , n t = R G k d , t 1 = 1 k d , t 1 d k d , t 1 d G = 1 k d , n t d k d , n t d G > 0 (16)
These constraints imply a certain attainable range of responses in Vdau and TG1, meaning that only particular changes in Vdau and TG1 are possible in response to increases in G. For each parameter k, we can calculate contribution of that parameter to the changes in ΔVdau and ΔTG1 that result from a change ΔG:
( Δ T G 1 Δ V d a u ) = ( C k T G 1 C k V d a u ) R G k Δ G (17)
The responses possible in response to increasing glucose (ΔG > 0) can then be plotted as vectors in the (ΔTG1, ΔVdau) space for each pathway, as shown in Fig 6B. The shaded region represents the space spanned by linear, positive sums of these vectors, which is the attainable range of responses. Here, the regulatory mechanisms that we consider are limited to speeding up the cell cycle with increasing glucose levels (i.e. ΔTG1/ΔG < 0). Additionally, while this form of regulation can freely decrease the cell size without having a significant impact on the cell cycle period, there must be a decrease in period in order to effect an increase in cell size.
The consistency of the attainable region with experimental observations can be assessed by evaluating measured Vdau and TG1 values under different glucose concentrations and constant growth rate, as reported in [46]. The linear correlation between these values at three glucose concentrations (0.05, 0.1, and 2%) suggest the following empirical relationship, as depicted in S6 Fig (note that Vdau, TG1 are in units of fL and minutes, respectively):
Δ V d a u = - 0 . 27 Δ T G 1 (18)
Note that this makes no assumption about the explicit relationship between glucose levels and the magnitude of parameter perturbations. The corresponding sensitivity to changes in glucose is then given by:
( Δ V d a u Δ T G 1 ) ∝ ( - 0 . 27 1 ) (19)
As shown in Fig 6B, this lies within the attainable region, confirming that this simple combination of regulations is consistent with the observed changes in behaviour.
An interesting aspect of the attainable region is that it is bounded by the opposing effects of stimulation of Cln3 translation and inhibition of ClbS transcription by glucose. This means that regulation of Net1 dephosphorylation does not broaden the range of behaviours that can be brought about through the pathway under constant conditions. Additionally, we observe that any change in behaviour (ΔTG1, ΔVdau) within the attainable region can be achieved in an infinite number of ways depending on the relative strengths of the three posited regulatory mechanisms (see S1 Text and S7 Fig). These different combinations of parameter perturbations will, by construction, have identical cell cycle behaviour under constant conditions, but may have distinct behaviours under dynamic changes in conditions.
In order to evaluate the potential for diverse dynamics in this system, we fix the change in behaviour achieved by parameter perturbations according to experimental observations (Eq 19), and consider three cases: no, weak, and strong up-regulation of Net1 dephosphorylation with increasing glucose levels. These changes are automatically balanced by changes in Cln3 and ClbS regulation by the constraint to achieve the specified (ΔTG1, ΔVdau). The resultant changes in the dynamic sensitivity (S k V d a u , 1 ( t )) shown in Fig 6D are the result of differences in the timing of sensitivity of the cell cycle to the different parameters (see Fig 6C for the individual sensitivity profiles). Regulation of Cln2 transcription and Cln3 translation alone is only capable of modulating cell cycle progression around the G1-S transition, while regulation of Net1 dephosphorylation modulates progression through mitosis. Therefore, regulation of Net1 in the model allows for a faster response to changes in glucose levels by extending the time window of responsiveness to glucose levels.
It has been noted previously that glucose levels act predominantly to modulate duration of the G1 phase of the cell cycle [76], as discussed above in the more general case. An important conclusion arising from the work presented here is that this form of regulation does not exclude active regulation of processes occurring during mitosis (or other phases of the cell cycle). Indeed, as long as counteracting pathways can be modulated in tandem, regulation of processes occurring in mitosis may be a useful strategy for dynamic adjustment of cell cycle characteristics after a change in conditions. In the particular example of strong Net1 regulation shown in Fig 6D, this is seen to lead to a more rapid modulation of Vdau than would be possible if only Cln2 and Cln3 were regulated. As discussed above, observations of cell populations under constant conditions (e.g. the chemostat experiments in [12, 76]) are not capable of distinguishing between these strategies of regulation.
In summary, this analysis demonstrates that control of the G1/S transition is insufficient for rapid adjustment of the cell cycle to changing conditions. In order for rapid response to changing conditions, it is necessary for the components that are active during the S/G2/M phases of the cell cycle to be regulated by environmental signals. Furthermore, the effects of such perturbations may only be observable in experiments in which response dynamics are observed. Cell cycle sensitivity during mitosis has been observed experimentally in response to sudden nutrient starvation or application of rapamycin [77, 78], suggesting that investigation of nutrient signalling under constant conditions can indeed mask important regulation.
Cell cycle progression is a highly regulated process. This is a result of the importance of the processes it coordinates, and of the fine-tuned response required in changing conditions. A number of environmental stimuli have been observed to regulate cell cycle progression [16], and in some cases regulatory components have been identified. While mathematical models have been able to provide insight into cell cycle responses to some particular environmental changes (e.g. in the case of osmotic stress [20]), a broader view of how the cell cycle regulatory network might respond to environmental changes, and how that might affect subsequent growth patterns, has been lacking. It is clear that this is a problem of importance in both basic and applied contexts and that its analysis requires a systematic approach.
Our analysis was focussed on three mathematical models spanning a range from the simple (the Pfeuty model [34]), to the complex (the Chen [26, 35] and Barik [30] models). This revealed that some patterns of sensitivity were common to all models. For example, an anticorrelation between changes in G1 phase duration and daughter cell size was observed in all three models, and matched experimental observations [12, 46, 52]. These are also reminiscent of correlations observed at the single-cell level within populations of cells [3]. In addition, the models were shown to exhibit other qualitative behaviours that are observed experimentally, such as a sensitivity of G1 phase duration to perturbations. The consistency of model behaviours with experimental observations demonstrates that the models capture essential properties of cell cycle behaviour beyond those typically considered (e.g. the behaviour of cell cycle mutants [26]). The fact that that these behaviours are observed even in the simplified Pfeuty model suggests that they are robust features of the cell cycle. This suggests that other aspects of model behaviours identified here, such as the prevalence of biphasic phase response curves, are good candidates for further investigation.
Sensitivity analysis characterises changes in model behaviour in response to small perturbations, providing a platform for understanding their behaviour under large perturbations that may elicit nonlinear responses. However, it is important to recognise that changes that result in bifurcations which transform the qualitative behaviour must be analysed with tools from bifurcation theory. This constitutes an important class of cell cycle behaviours, including cell cycle arrest, meiosis, or the transition to endoreplication. Bifurcation analysis has been applied to understand these behaviours in a variety of cases (e.g. [79]), and provides insights into system behaviour that are complementary to those obtained by sensitivity analysis. Bifurcation analysis has also been an vital tool for understanding the dramatic changes in cell cycle behaviour caused by loss of some cell cycle genes (e.g. [28]). Through a combination of sensitivity analysis in the linear regime, and bifurcation analysis and simulations in the nonlinear regime, a comprehensive analysis of cell cycle behaviour in dynamic environments can be undertaken.
While this study has focussed on S. cerevisiae as a model system in which to study the cell cycle, related questions arise in a range of contexts of both fundamental and applied interest. For example, the question of how environmental cues regulate cell cycle progression is a general one, and is of interest in other yeast species, as well as in plant and animal systems. Though the cell cycle mechanisms are somewhat different in these systems, a similar approach to that taken here can be used to address this class of questions. In an applied context, having a mechanistic basis for understanding the connections between extracellular conditions, growth, and cell cycle progression in yeast is an important practical tool, for example in maximising yield of a valuable product.
The relevance of parameter sensitivity analysis to experimental studies of cell cycle behaviour depends on technology for accurate observation of cellular behaviour, fine control of cellular environment, and manipulation of cellular network structures. Rapid advances in microfluidic and imaging technology are addressing these issues [39, 52, 80–82], with current methods capable of observing hundreds of cells over many generations under rapidly changing conditions [83]. The ability to measure coordinated changes in regulatory network and cell cycle dynamics in response to perturbations is allowing increasingly detailed understanding of molecular mechanisms (e.g. [78]). This then makes it feasible to perform controlled changes in the environment and observe the resulting macroscopic changes in growth patterns. In addition, the burgeoning possibilities of synthetic approaches allow hypotheses about molecular mechanisms to be explored at unprecedented levels of detail [84]. The combination of quantitative modelling methodologies such as those employed here with these high-throughput, quantitative experimental approaches will allow for a significant improvement in our understanding of cell cycle and growth progression in varying environments.
Overall, by performing systematic steady state and dynamic sensitivity analysis to a range of detailed and simplified models, we have established a methodological platform to investigate the effects of dynamically varying environments on the cell cycle. Future extensions may incorporate bifurcation analysis to understand qualitative transformations in behaviour in the nonlinear regime. While we have applied our analysis to understand characteristics such as cell size and cell cycle duration in this particular study, it can also be applied to understand other characteristics of cell cycle behaviour. In conjunction with experiments, this approach provides a sound basis for beginning to understand the roles of the different parts of the cell cycle machinery in generating these responses. Furthermore, it also provides a basis for developing simplified descriptions which combine biological realism and mathematical soundness. This may be important in application domains. Finally, this approach provides a new window into the cell cycle as a complex system, and a route into understanding how dynamic information processing is undertaken by the cell cycle control system.
Mathematical models of the cell cycle have been useful in describing how known molecular interactions give rise to the observed complex dynamics [85], and to predict the behaviour of cell cycle mutants [86]. Despite the fact that these models may contain many parameters, they exhibit a fairly limited range of behaviours. This arises from the fact that these models encode similar regulatory logic. As a result, while our understanding of the biochemical details may change substantially, if the regulatory logic is broadly the same, we expect future mathematical models to exhibit similar behaviours. We further note that, if we focus on particular macroscopic, experimentally observable features (as is the case here), the range of behaviours for these features is especially restricted. A reduced effective dimensionality has been noted in a range of biological models, including models of the cell cycle [42, 87, 88]. Overall we find that the models are capable of reproducing a range of experimental observations. This consistency in the face of considerable molecular and dynamic complexity suggests that these models will be valuable tools for understanding how the cell cycle responds to changing environments and for utilizing this in multiple applications.
|
10.1371/journal.pcbi.1005339 | Estimating the Respective Contributions of Human and Viral Genetic Variation to HIV Control | We evaluated the fraction of variation in HIV-1 set point viral load attributable to viral or human genetic factors by using joint host/pathogen genetic data from 541 HIV infected individuals. We show that viral genetic diversity explains 29% of the variation in viral load while host factors explain 8.4%. Using a joint model including both host and viral effects, we estimate a total of 30% heritability, indicating that most of the host effects are reflected in viral sequence variation.
| Viral loads of Human Immunodeficiency Virus infections are correlated between the donor and the recipient of the transmission pair. Similarly, human genetic factors may modulate viral load. In this study we estimate the extents to which viral load is heritable either via the viral genotype (from donor to recipient) or via the host’s Human Leukocyte Antigen (HLA) genotype. We find that a major fraction of inter individual variability is explained by the similarity of the viral genotypes, and that human genetic variation in the HLA region provide little additional explanatory power.
| There are differences in the rate of disease progression among individuals infected with HIV. An easy to measure and reliable correlate of disease progression is the mean log viral load (HIV RNA copies per ml of plasma). The viral load measured during the chronic phase of infection (referred to as setpoint viral load, spVL) exhibits large variation in a population. Several studies have been carried out to elucidate whether this variation is primarily driven by host genetics [1–4], viral genetics [5–9], or environmental effects [7]. Genome-wide association studies consistently show that amino acid polymorphisms in the peptide binding groove of the HLA-A and HLA–B proteins are associated with the viral load of an individual. Furthermore, variants in the HLA-C and CCR5 genes have also been shown to impact spVL. However, those host factors explain less than 15% of the observed phenotypic variance [4]. In contrast, viral genetic studies and studies of donor-recipient transmission pairs established that 33% of the phenotypic variance is attributable to the transmitted virus itself [5, 10–13].
HIV is an extremely variable and adaptive organism with a rapid replication time, and high rates of mutation. Within-host evolution of the viral population occurs during the chronic phase of infection in which the pathogen adapts to its host environment. Several studies showed that a major proportion of the viral sequence is under selective pressure in the host environment, and several viral amino acid changes are associated with host genetic variants in the Human Leukocyte Antigen (HLA) genes [14, 15].
Viral strains harbor epitope sequences that can be presented by HLA class I proteins of the infected host, which allows the detection and killing of infected cells. The viral population evades detection through escape mutations that modify the epitope sequence but may incur a fitness cost. Compensatory mutations may follow until the viral population reaches its optimal place in a sequence space constrained by the host immune system [16].
There are two main different approaches to viral heritability estimation in the literature. The first one is based on the regression of phenotypic values in donor-recipient transmission pairs, while the other quantifies the difference between the observed phenotypic variance-covariance structure and the phylogenetic variance-covariance structure. Because our study population did not include donor-recipient data, we used the latter strategy. In particular we used linear mixed models (LMMs) to explain inter-patient differences in spVL while taking into account host and viral genetic relatedness. LMMs use the pairwise relatedness of individuals with respect to a large set of features (rather than the individual data points) to estimate the fraction of phenotypic variance attributable to those features. Such models have been successfully applied to estimate narrow-sense heritability from genome-wide genotype data [17]. Concurrently, LMMs were proposed to incorporate phylogenetic relatedness between samples in comparative analyses [18], a technique that was further developed to estimate the viral genetic contribution to spVL [6, 8].
To estimate the respective contribution of host and viral genetics to the variation in spontaneous HIV control, we collected paired viral/host genotypes along with spVL measurements from 541 chronically infected individuals enrolled in two prospective cohort studies in Switzerland and in Canada. We estimated the respective contributions of host and viral genetics to spVL by defining two relatedness measures, one with respect to the host genotypes, the other with respect to the viral genotypes, and used these jointly in a linear mixed model.
On the host side, we focused on amino acid variations in the HLA-A, B and C genes due to their established associations with HIV control [1]. In particular, we used 33 amino acid polymorphisms selected by L1 regularized regression [19] to represent the genetic relatedness of the host (S1 Table). Principal component analysis based on host genome-wide genotype data confirmed the lack of major population stratification in the host sample.
We built three LMMs, one containing human variants, one derived from phylogenetic trees, and one including both host and virus information (Fig 1). The genetic relatedness matrix created from 33 amino acid polymorphisms of the human class I HLA genes explained 8.4% (SD = 4%) of the observed variance in spVL. In contrast, 28.8% (SD = 11%) of phenotypic variation was attributable to the viral phylogenetic tree. Combining the two relatedness matrices in one model yielded a total variance explained of 29.9% (SD = 12%), less than the sum of the latter two models. Thus, we show that HLA polymorphisms do not explain additional phenotypic variance beyond viral sequence variation.
We next assessed the contribution of viral variants most likely to have an impact on spVL. These included amino acids in known CTL epitopes [20] and those positions whose variation is associated with host polymorphisms [14] (82%, 60% and 84% of gag, pol, nef codons respectively, S2 Table). We used phylogenetic trees built from those codons to show that viral variation in epitopes or other HLA-associated positions explain 23.6% (SD = 11%) of phenotypic variance. However, this explained fraction might be overestimated due to linkage disequilibrium on the viral haplotype. We therefore repeated the analysis after randomly picking 70% of variable viral positions, and obtained very similar results. We thus cannot conclude that viral variants in known epitopes contribute disproportionately to variance in spVL. Additional evidence for the existence of substantial linkage disequilibrium on the viral haplotype comes from the analysis of the smaller, complementary set of variable viral positions (located in non-epitope regions), which explained 18.5% (SD = 10%) of the phenotypic variance. This leads to lower bounds of 11.4% and 6.3% of variance in spVL explained by variation in epitope and non-epitope regions, respectively, leaving 12.2% of variance unresolved due to linkage disequilibrium.
By jointly analyzing host and viral genetic relatedness, we here provide estimates of the total and respective contributions of human and viral genetic variation to HIV control. Our results do not challenge the current consensus estimates of the host or viral contributions to spVL. Nevertheless, our combined analysis demonstrates that human HLA polymorphisms do not explain additional variance in spVL once viral genetic diversity is taken into account.
The difference between the variance explained by viral phylogeny and the variance explained by HLA polymorphisms may be attributed to two effects. First, selected viral variants might provide a better surrogate of the impact of the host genotype than the imputed host amino acid variants we used. Rare host genetic factors outside of the major histocompatibility complex region (e.g. the CCR5 deletion), as well as environmental interactions may influence viral fitness, and these effects are not accounted for in our estimate of host heritability. Thus some host effects might be missed from the host partition, while their footprint in the virus is still detected in the viral partition. Second, the difference could partly be due to the effect of viral variation independent of the current host, including transmitted escape mutations, i.e. viral sequence variation carried over from the previous host, rather than induced by the current host. Indeed, a recent study showed that spVL is dependent on the degree of pre-adaptation of the viral strain to the HLA class I genotype of the current host [21]. In particular, an increase in the frequency of pre-existing escape mutations, at the population level, led to higher viral heritability estimates. This indicates that both host and viral estimates of heritability depend on the amount of pre-adaptation in the sample population, which varies based on the level of HLA diversity. It has also been shown that reversion of some fitness reducing escape variants is very slow, potentially allowing for a transitory but measurable effect on viral load at the population level [15, 22].
A limitation of our study is the fact that study participants were collected from two cohorts. To reduce batch effect, we included a cohort-specific variable in all our models. Still, differences in inclusion criteria, health system, geographical exposure and other factors are very likely to increase environmental variance, thus negatively impacting our heritability estimates.
Another potential shortcoming is our implicit assumption of the absence of selection on spVL, which might be incorrect, as suggested by recent studies [23, 24], and might thus lead to over- or under-estimation of heritability due to model misspecification. Still, because our estimates are comparable to results obtained in donor-recipient transmission studies and in host-genetic studies, we conclude that they are useful for the purpose of delineating the respective amounts of host and viral contributions to phenotypic variation of HIV spVL.
In conclusion, our results suggest that host genetic association studies not taking the virus into account underestimate the population level effect of host genetic variation. Combining host and pathogen data provides additional insight into the genetic determinants of the clinical outcome of HIV infection, which can serve as a model for other chronic infectious diseases.
All participants were HIV-1-infected adults, and written informed consent for genetic testing was obtained from all individuals as part of the original study in which they were enrolled. Ethical approval was obtained from institutional review boards for each of the respective contributing centers.
Bulk sequences of the HIV-1 gag, pol and nef genes, human genome-wide genotyping data and viral load measurements were obtained for 541 individuals of Western European ancestry infected with HIV-1 Subtype B, and followed in the Swiss HIV Cohort Study (SHCS, www.shcs.ch) or in the HAART Observational Medical Evaluation and Research study in Canada (HOMER, www.cfenet.ubc.ca/our-work/initiatives/homer) [14].
Viral sequences data were generated from samples collected two to five years after infection (for SHCS) or during chronic infection (for HOMER) but prior to the initiation of antiretroviral therapy. Thus, the viral genotypes reflect the result of natural adaptation of the pathogen to the host environment. The viral sequences for 1262, 2187 and 548 nucleotides of the gag, pol and nef genes were available for at least 80% of samples studied. The analysis was limited to these three genes because sequences of the rest of the retroviral genome were not available for the majority of study samples. Overlapping viral genomic regions were excluded from gag, to avoid duplicated sequences in the analysis.
Human DNA samples were genotyped in the context of previous genome-wide association studies. High-resolution HLA class I typing (4 digits; HLA-A, HLA-B, and HLA-C) was imputed from the genome-wide genotyping data as described previously [14].
Set point viral load (spVL) was defined as the average of the log10-transformed numbers of HIV-1 RNA copies per ml of plasma obtained in the absence of antiretroviral therapy, excluding VL measured in the first 6 months after seroconversion and during periods of advanced immunosuppression (i.e., with <100 CD4+ T cells per ul of blood). The distributions of spVL in the two cohorts are shown in S1 Fig.
The pairwise genetic relatedness of the dominant viral strains observed in the samples was calculated from phylogenetic trees similarly to [6]. Nucleotide sequences were translated to amino acid sequences, which were in turn aligned with MUSCLE [25] and used to derive the correct codon-aware nucleotide alignment. The phylogenetic tree was built from the aligned nucleotide sequences using RAxML [26] with the following command line: “raxml -w {PATH} -s {PATH} -m GTRCAT -f a -N 30 -k -n {NAME} -T {NUMBER} -x 1234 -p 1234”. Individual sequences were then rooted to the HIV-1 group M ancestral sequence, downloaded from the Los Alamos sequence database. Using an HIV-1 subtype C sequence as outgroup led to similar results. The whole tree was scaled with the inverse of the median height of the branches. We followed the method of Hodcroft et al, to create a relatedness matrix from a phylogenetic tree [6]. The genetic relatedness of two samples in a given phylogenetic tree is the amount of shared ancestry, i.e. the distance from the root of the tree (excluding the outgroup) to their most recent common ancestor [27].
We selected 33 amino acid variants with L1-regularized regression (LASSO) out of all polymorphisms in the HLA-A, B and C genes and used them to generate a genetic relatedness matrix as described in [17]. Our relatively small sample size made it necessary to use a small subset of selected markers rather than genome-wide variant information to create the genetic relatedness matrix. Doing otherwise would have resulted in very large errors of the estimates.
To estimate heritability, we used the gcta software as a generic implementation of the linear mixed model [17]. In such a framework, a multivariate Gaussian distribution models HIV viral load with a variance-covariance matrix consisting of the linear combination of the sample-sample genetic relatedness matrices (one for the host and one for the virus) and the identity matrix (representing sample-specific noise). The total heritability estimate is the fraction of variance explained by the genetic relatedness matrices over the total variance. All models included a binary variable indicating cohort as a fixed effect. Variance components were estimated by restricted maximum likelihood.
|
10.1371/journal.pgen.1000407 | Computationally Driven, Quantitative Experiments Discover Genes Required for Mitochondrial Biogenesis | Mitochondria are central to many cellular processes including respiration, ion homeostasis, and apoptosis. Using computational predictions combined with traditional quantitative experiments, we have identified 100 proteins whose deficiency alters mitochondrial biogenesis and inheritance in Saccharomyces cerevisiae. In addition, we used computational predictions to perform targeted double-mutant analysis detecting another nine genes with synthetic defects in mitochondrial biogenesis. This represents an increase of about 25% over previously known participants. Nearly half of these newly characterized proteins are conserved in mammals, including several orthologs known to be involved in human disease. Mutations in many of these genes demonstrate statistically significant mitochondrial transmission phenotypes more subtle than could be detected by traditional genetic screens or high-throughput techniques, and 47 have not been previously localized to mitochondria. We further characterized a subset of these genes using growth profiling and dual immunofluorescence, which identified genes specifically required for aerobic respiration and an uncharacterized cytoplasmic protein required for normal mitochondrial motility. Our results demonstrate that by leveraging computational analysis to direct quantitative experimental assays, we have characterized mutants with subtle mitochondrial defects whose phenotypes were undetected by high-throughput methods.
| Mitochondria are the proverbial powerhouses of the cell, running the fundamental biochemical processes that produce energy from nutrients using oxygen. These processes are conserved in all eukaryotes, from humans to model organisms such as baker's yeast. In humans, mitochondrial dysfunction plays a role in a variety of diseases, including diabetes, neuromuscular disorders, and aging. In order to better understand fundamental mitochondrial biology, we studied genes involved in mitochondrial biogenesis in the yeast S. cerevisiae, discovering over 100 proteins with novel roles in this process. These experiments assigned function to 5% of the genes whose function was not known. In order to achieve this rapid rate of discovery, we developed a system incorporating highly quantitative experimental assays and an integrated, iterative process of computational protein function prediction. Beginning from relatively little prior knowledge, we found that computational predictions achieved about 60% accuracy and rapidly guided our laboratory work towards hundreds of promising candidate genes. Thus, in addition to providing a more thorough understanding of mitochondrial biology, this study establishes a framework for successfully integrating computation and experimentation to drive biological discovery. A companion manuscript, published in PLoS Computational Biology (doi:10.1371/journal.pcbi.1000322), discusses observations and conclusions important for the computational community.
| In order to understand molecular biology at a systems level, it is first necessary to learn the functions of genes by identifying their participation in specific cellular pathways and processes. While protein sequence and structural analyses can provide valuable insights into the biochemical roles of proteins, it has proven much more difficult to associate proteins with the pathways where they perform these roles. Recently, high-throughput and whole-genome screens have been used to form basic hypotheses of protein participation in biological processes. However, the results of these studies are not individually reliable enough to functionally associate proteins with pathways. Many computational approaches have been developed to integrate data from such high-throughput assays and to generate more reliable predictions [1], but protein function cannot be confidently assigned without rigorous experimental validation targeted specifically to the predicted pathway or process. Surprisingly few follow-up laboratory efforts have been performed on the basis of computational predictions of protein function, and as such, these computational approaches remain largely unproven, and consequently underutilized by the scientific community [2],[3]. Here, we demonstrate that computational predictions can successfully drive the characterization of protein roles using traditional experiments. To test the approach, we systematically measured the mitochondrial transmission rates of a tractable set of S. cerevisiae strains carrying deletions of genes predicted to be necessary for this biological process.
The mitochondrion is an organelle central to several key cellular processes including respiration, ion homeostasis, and apoptosis. Proper biogenesis and inheritance of mitochondria is critical for eukaryotes as 1 in 5,000 humans suffers from a mitochondrial disease [4]. Saccharomyces has proven to be an invaluable system for studying a variety of human diseases [5],[6], including cancer [7], neurologic disorders [8], and mitochondrial diseases [9]–[11]. Yeast is a particularly attractive model system for studying mitochondrial biology due to its ability to survive without respiration, permitting the characterization of mutants that impair mitochondrial function. The process of mitochondrial biogenesis and inheritance [12] (hereafter, mitochondrial biogenesis) comprises a number of sub-processes that together ensure that new mitochondria are generated and segregated to a daughter cell. Mitochondrial biogenesis begins with the nuclear genes encoding mitochondrial proteins being transcribed, translated, and targeted to the mitochondria for import [13],[14]. The mitochondria must also replicate its own genome [15] and assemble the numerous membrane-bound complexes necessary for proper function [16]. During mitochondrial transmission, the mitochondria are actively transported along actin cables to the bud neck, where they are then segregated between the mother and daughter cells [17]. In addition to the experimental utility of yeast, it is well suited for the application of computational prediction approaches due to the availability of manually-curated annotations of yeast biology and the available wealth of genome-scale data.
Previous efforts have focused on identifying mitochondria-localized proteins through laboratory techniques such as mass spectrometry and 2D-PAGE [18],[19] and through computational predictions of cellular localization [20],[21]. These approaches have resulted in the identification of over 1,000 mitochondria-localized proteins in S. cerevisiae [22]. However, despite yeast's convenience as a model system, mitochondrial phenotypes of ∼370 of these 1,000 localized proteins have not been characterized, so the mitochondrial role of these predictions is unknown (over half of these 370 have no known function in any cellular process). Previous computational efforts have attempted to address this problem by predicting putative mitochondrial protein modules [20] and examining expression neighborhoods around mitochondrial proteins [23]. While valuable, these predictions of protein function have not been confirmed through laboratory efforts. Rather, these studies have performed assays for protein localization to the mitochondria, which is not sufficient to convert these predictions to concrete knowledge of protein roles [24].
Here, we describe a strategy that combines computational prediction methods with quantitative experimental validation in an iterative framework. Using this approach, we identify new genes with roles in the specific process of mitochondrial biogenesis by directly measuring the ability of cells carrying deletions of candidate genes to propagate functioning mitochondria to daughter cells. We assayed our 193 strongest predictions with no previous experimental literature evidence of phenotypes and interactions establishing a function in mitochondrial biogenesis. By these assays we experimentally discovered an additional 109 proteins required for proper mitochondrial biogenesis at a level of rigor acceptable for function annotation. Further, we identified more specific roles in mitochondrial biogenesis for several predicted genes through mitochondrial motility assays and measurements of respiratory growth rates. We also discovered genes with redundant mitochondrial biogenesis roles through targeted examination of double knockout phenotypes. This demonstrates that using an ensemble of computational function prediction methods to target definitive, time-consuming experiments to a tractably sized set of candidate proteins can result in the rapid discovery of new functional roles for proteins. Our results also show that most mutants resulting in severe respiratory defects have already been discovered. This is likely to be the case for mutant screens in many fundamental biological processes, because saturating screens have discovered mutations with strong phenotypes. However, even in a well-studied eukaryote like S. cerevisiae, there are many processes that have not yet been fully characterized by identifying all proteins required for its normal function [25]. As such, most of the remaining undiscovered protein functions are only identifiable by rigorous, quantitative assays that can detect subtle phenotypes, such as those used by our study.
We utilized an ensemble of computational gene function prediction approaches to systematically identify candidates for involvement in mitochondrial biogenesis. These candidates were experimentally assayed, and the confirmed predictions were then utilized as inputs for a second round of prediction and experimentation. A schematic overview of this approach is shown in Figure 1.
We trained an ensemble of three computational prediction methods (bioPIXIE [26],[27], MEFIT [28], and SPELL [29]) using genomic data that we collected from many sources (full list in supplementary materials) and a set of 106 genes known to be involved in mitochondrial organization and biogenesis based on published experiments as curated by the Saccharomyces Genome Database (SGD) [30]. Genes are assigned by SGD to this biological process if published experiments have definitively demonstrated functions involved in the formation, assembly, or disassembly of a mitochondrion. The classification of mitochondrial organization and biogenesis includes genes that affect mitochondrial morphology and distribution, replication of the mitochondrial genome, and synthesis of new mitochondrial components.
An intuitive description of our computational methods is that each employs “guilt by association” to identify genes exhibiting similar data patterns to the genes used for training (further details in Methods). The ensemble was used to rank all genes in the genome from most likely to be involved in mitochondrial biogenesis to least likely. We selected the top 183 most confident genes that were not included in the training set for experimental validation. Of these, we found existing experimental literature evidence of involvement in mitochondrial biogenesis for 42 proteins, and as such we included these in our set of positive controls (along with 6 genes from the training set). The remaining 141 proteins comprised our set of first iteration predictions, as none of these proteins appeared in published experiments that demonstrated their requirement for proper mitochondrial biogenesis. We assayed these predicted genes experimentally as described below. We then augmented our training set of genes known to be involved in mitochondrial biogenesis with the experimentally verified predictions (using both our experiments and the uncurated published literature, see methods) and repeated this process to generate a second iteration of predictions. From this second iteration, we selected the 52 most confident predictions that were not previously tested and performed the same experimental assays.
In order to confirm the potential roles of our candidate genes in mitochondrial biogenesis, we employed an experimental assay that measures the rate of generation of cells lacking respiratory competent mitochondria (called “petite” cells [31]). This assay reliably detects defects in mitochondrial biogenesis, but it is too time consuming to perform on a whole genome scale. Wild-type yeast from the S288C genetic background produce petite daughter cells at a baseline frequency of ∼20% [32], but mutation of genes involved in mitochondrial biogenesis can significantly alter this rate.
We measured the frequency of petite formation for single gene deletion strains of all 193 candidate genes (141 from the first iteration, 52 from the second) and for 48 positive control genes. To reduce the effects of suppressor mutations and aneuploidy associated with the yeast deletion collection [33], we sporulated the heterozygous Magic Marker deletion collection [34] and isolated six independent haploid deletion mutants for every gene tested. Individual deletion strains were grown in media requiring aerobic respiration for growth (glycerol), and strains completely unable to grow were deemed respiratory deficient and did not continue in the assay. The remaining mutants were then assayed by measuring the ratio of petite cells to total cells in a colony founded from a single cell. At least twelve matched wild-type sporulation isolates were assayed on each day of experiments in order to establish baseline frequencies. For each gene tested, petite frequencies were measured for at least eight colonies and compared to the distribution of wild-type frequencies measured in parallel on each day of experiments, which allowed us to quantitatively detect subtle phenotypes with statistical rigor. A schematic of this assay is shown in Figure 2 and further details are available in Methods.
In our first iteration of prediction and experimental testing, 83 of our initial 141 predictions (59%) were confirmed to play a role in mitochondrial biogenesis as they exhibited a significantly altered petite frequency rate compared to the wild-type distribution (FDR corrected Mann Whitney U-test p-value <0.05; see Figure 3A). These 83 newly confirmed predictions were added to the training set, and we then performed another iteration of prediction and experimentation. In this second iteration, 17 of the 52 predictions (33%) were experimentally confirmed (Figure 3A). Based on the second iteration predictions, we also examined a targeted set of double knockout mutants and experimentally confirmed 9 more proteins that exhibit synthetic petite frequency defects (full details below). Further, the petite frequency assay demonstrated a high level of sensitivity as 44 of our 48 positive controls (92%) exhibited a significant phenotype (the remaining 4 are discussed further below). All together, after both iterations of our approach we discovered a role in mitochondrial biogenesis by demonstrating significant phenotypic alterations for 109 of our 193 (56%) total predictions (see Table 1 for breakdown).
These newly characterized functions include 42 genes with other previously known functions (not in mitochondrial biogenesis) and 68 genes with no previously characterized cellular role. For example, we observe that mutation in the functionally uncharacterized TOM71 causes a 44% increase in petite frequency. While Tom71 has been co-localized with the translocase complex responsible for protein import through the mitochondrial outer membrane, previous work (largely in vitro) has not identified a strong functional defect associated with Tom71 in translocase activity [35]. Our confirmation that TOM71 significantly affects mitochondrial transmission rates strongly suggests that it does indeed play a role in mitochondrial import, at least for some subset of proteins required for mitochondrial inheritance or biogenesis. The identification of a functional role for 68 previously uncharacterized proteins is particularly striking as this covers roughly 1 in 18 of the remaining ∼1200 proteins in yeast that still have no known functional role [25].
We observed a striking difference in the severity of petite frequency phenotypes in single gene knockouts between the confirmed gene predictions and the positive controls (Figure 3B). Of the 44 positive controls demonstrating a significant phenotype, the majority exhibited a complete loss of respiratory function (28 of 44, 64%) as opposed to the more subtle phenotype of altered mitochondrial transmission (16 of 44, 36%). The proportions of subtle and severe phenotypes were reversed in our predictions experimentally confirmed by single gene knockouts, in which 79 of 100 mutants (79%) showed altered mitochondrial transmission while only 21 of 100 mutants (21%) were respiratory deficient. The quantitative nature of these phenotypes among our novel discoveries may indicate why they have not been previously associated with mitochondrial biogenesis by either classical genetic screens or high-throughput techniques [24],[36], which generally assay extreme rather than subtle phenotypes. In further support of this observation, since undertaking this study, 8 of our 100 confirmed candidates have been associated by other groups to mitochondrial biogenesis (COA1 [37], IBA57 [38], GUF1 [39], ATP25 [40], QRI5 [41], GRX5 [42], REX2 [43], RMD9 [44]), and 4 of these 8 exhibited the most extreme phenotype of respiratory deficiency in our study (Table S2).
The confirmation rate from our second iteration decreased from our first iteration (59% to 33%), which suggests we may be nearing the limit of predicted genes that can be verified using the single knockout petite frequency assay. In particular, examining single gene deletion strains prohibits characterization of the roles of redundant proteins or genes that only exhibit synthetic phenotypes. In fact, all four of our 48 positive controls that did not exhibit a significant petite frequency phenotype are known to synthetically interact with at least one other gene involved in mitochondrial biogenesis and inheritance [45]–[48]. Our second iteration prediction results indicate which of our unconfirmed predictions are worthy of further investigation with double mutant analysis or additional assays, particularly in light of additional localization evidence. Following the second round of computational prediction, 26 of the 58 initially unconfirmed predictions persisted as highly ranked candidate genes while the remainder decreased in confidence. Of these, 22 (85%, hypergeometric p-value <10−9) candidates are known to localize to the mitochondria, while only 1 of the remaining 32 unconfirmed candidates (3%) is similarly localized.
To test the hypothesis that these 26 high-confidence unconfirmed predictions represented genes that had redundant mitochondrial function, we performed targeted double mutant analysis looking for synthetic interactions. We chose 4 deletion mutants (aim17Δ, rvs167Δ, tom6Δ and ehd3Δ) confirmed to be involved in mitochondrial biogenesis with modest petite frequency phenotypes to cross with these 26 candidates. Choosing mutants with modest phenotypes was necessary to allow for a strong synthetic interaction to be observed. We tested 99 double mutant strains and observed 11 significant synthetic phenotypes (FDR corrected Wilcoxon rank-sum p-value <0.05) spanning 9 of 26 mutants that did not display a single mutant phenotype (Figure 4). While some of our double mutants exhibit suppression, we did not focus on these interactions because of the modest nature of the single mutant phenotypes. Instead we focused on synthetic defects which we could rigorously define as the double mutant petite frequency being significantly different from both single mutants and the wild-type petite frequency. Of the genes exhibiting significant double mutant phenotypes, 1 was synthetic respiratory deficient and 8 demonstrated altered petite frequency. The 9 genes showed a specific pattern of synthetic phenotypes, as 7 interacted with only 1 of the 4 known mitochondrial biogenesis genes used to generate double mutants. These specific synthetic interactions suggest the functions these genes may perform in mitochondrial biogenesis. For example, the four genes (AIP1, MPM1, YDL027C, and YDR286C) that specifically interact with rvs167Δ are potentially involved in the actin-based transmission of mitochondria to the daughter cell as Rvs167 is a regulator of actin polymerization [49]. In fact, the only known actin-localized protein among our 26 candidates, Aip1, had a genetic interaction only with the rvs167Δ(Figure 4).
The high rate of synthetic phenotype recovery (9 out of 26 candidates tested) was made possible by the use of computation to limit the number of double mutants queried. There were 58 unconfirmed predictions from the first round of our analysis, and 95 genes tested in this study have the quantitative petite frequency phenotypes necessary for double mutant analysis. Combining these 95 confirmed genes with the 58 unconfirmed genes yields 5,510 possible double mutants to assay, which is far too large to reasonably test with the quantitative petite frequency assay. However, we used computation in two ways to reduce the number of double mutants screened to ∼100. First, we used computational iteration to identify the subset of unconfirmed predictions most likely to be involved in mitochondrial biogenesis. Second, we used the functional networks generated by the bioPIXIE algorithm [26] to select four genes from different sub-functions in mitochondrial biogenesis. This allowed us to test less than 2% of the possible double mutants, but still identify phenotypes for 9 of 26 candidates (35%) due to the efficiency of our computational approach.
While we expect high correlation between localization to the mitochondria and involvement in mitochondrial biogenesis, many proteins not localized to mitochondria are vital for regulating mitochondrial function and biogenesis [17]. Thus, a candidate gene approach based solely on protein localization would neglect many important participants in normal mitochondrial biogenesis. Our use of computational predictions to drive experimental discovery is unbiased with respect to any one genomic feature or assay. In this study, 47 (43%) of our 109 newly confirmed discoveries are not known to localize to the mitochondria [30],[50] and would have been overlooked in a screen of mitochondria-localized proteins lacking known functions. Further, the accuracy of our predictions for non-mitochondria-localized proteins is comparable to that for mitochondria-localized proteins (44% vs. 59%, respectively). Thus, computational predictions can broaden the scope of potential discoveries beyond a more restricted candidate gene approach based on a single experimental technique or data source.
Specific examples of non-mitochondria-localized proteins critical for mitochondrial biogenesis include proteins linking mitochondria to the actin cytoskeleton. Several of our novel discoveries have literature evidence associating them to the actin cytoskeleton but no evidence suggesting a role in mitochondrial transmission [51]–[53]. One of these genes, the uncharacterized ORF YIR003W (AIM21), has been shown to co-localize with actin in high-throughput studies [51] and was predicted as an interactor with the actin cytoskeleton with high confidence by our system bioPIXIE [27]. We found that strains carrying a deletion of YIR003W grow normally on glycerol but form petites at a frequency of 166% of wild type cells, one of the highest petite frequencies observed in our experiments.
To better understand the mitochondrial transmission defect in this mutant, we used our computational predictions to direct experiments targeting the role of the actin cytoskeleton in mitochondrial transmission. The morphology of the actin cytoskeleton and of the mitochondria in this mutant was visualized by dual immunofluorescence (Figure 5A,B). In the yir003wΔ mutants, the actin skeleton appears relatively normal, with typical polarization of actin patches toward the daughter (Figure S1), and the mitochondria show no gross structural perturbation in these mutants. However, by observing sustained mitochondrial movement events, we assessed mitochondrial motility for this mutant and found severe defects comparable to a puf3Δ strain (Figure 5C), a gene known to be involved in mitochondrial motility [54]. Even though this mutant displayed no overt morphological phenotypes, detailed analysis of YIR003W uncovered a more subtle, specific defect in mitochondrial motility.
To further characterize our predictions, we assayed single gene knockout mutants for respiratory growth defects, as assembly of the complexes required for respiration is a critical step in mitochondrial biogenesis. We quantitatively measured growth profiles of most of our single gene deletion mutants under respiratory growth conditions (glycerol) comparing them to growth in fermentative conditions (glucose) as a control. A 96-well plate incubator and optical density reader was used to determine growth profiles for six independent replicates of each deletion strain tested and for two matched wild-type isolates of each strain (24 control wells per plate, see Methods for details). Exponential growth rates and saturation densities were calculated for each strain (Figure 6A), and both of these parameters were assessed for statistical significance relative to the distribution of all wild-type controls. Significant phenotypes were only reported if the defect was unique to the glycerol growth condition (i.e. was not present in the glucose growth curve) in order to ensure that the growth defect is respiration specific. By combining the growth rates and saturation densities (Figure 6C), we arrived at a respiratory growth phenotype that classifies each mutant as severe, moderate, weak, or unaffected. An example growth curve of each class is shown in Figure 6B.
As expected, nearly all mutants classified as respiratory deficient in the petite frequency assay were classified as severely defective in the respiratory growth assay. However, we also observed significant respiratory growth phenotypes for 29 mutants without previously reported respiratory impairments in the literature. Of these, 22 exhibited a weak or moderate defect that may have been difficult to observe in whole-genome screens assaying respiratory growth [55],[56]; the remaining 7 severe phenotypes might have been previously overlooked due to suppressor mutations in the systematic deletion collection. While employing multiple replicates in such assays lowers overall throughput, these results suggest that testing many replicates enables more complete discovery of subtle respiratory growth phenotypes.
We employed thorough assays performed in replicate in order to detect important but subtle phenotypic variations. As such, it is impractical to scale these assays to the entire genome at the same level of rigor. In fact, given our rate of experimental efforts, it would require nearly 7 years for us to apply the petite frequency assay to all viable single gene deletion strains. However, by using computational predictions of protein function as a form of initial genetic screen, we were able to target our efforts towards the most promising candidates first. This is important for testing single gene deletions, but it is imperative for assaying potential synthetic defects. There are 18 million possible double gene knockouts in S. cerevisiae, a number far too large to comprehensively test for a broad range of phenotypes. However, we were able to discover 11 synthetic mitochondrial biogenesis defects by assaying a small, computationally chosen fraction of this available space. In all, by utilizing computational predictions of proteins involved in mitochondrial biogenesis, we have rapidly characterized new functional roles for 109 genes.
We have used computational predictions of gene function to direct focused, non-high-throughput laboratory experiments, confirming 109 proteins required for normal mitochondrial biogenesis in S. cerevisiae. These discoveries include 68 genes with no previously known function (5% of the remaining ∼1,200 uncharacterized S. cerevisiae genes) and 47 proteins not currently known to localize to the mitochondria. For several genes, our results provide evidence of involvement in specific sub-processes of mitochondrial biogenesis (e.g. AIM21/YIR003W in mitochondrial motility). No previous study has systematically tested computational predictions of protein functions using non-high-throughput laboratory techniques; the 56% accuracy established by our study demonstrates the potential of such computationally driven genetic investigations for direct future biological discoveries. In addition to the biological discussion presented here, this study resulted in several observations and conclusions important for the computational community, which are discussed in a companion manuscript [57]. Of our newly characterized mitochondrial genes, 53 have strictly defined human orthologs, 5 of which are associated with known diseases (see Methods).
Computational function prediction and non-high-throughput laboratory experiments complement each other in another important way highlighted by these results: the combination of these two techniques can rapidly identify subtle, quantitative phenotypes that are difficult to detect with high-throughput assays. When investigating well-studied processes (such as mitochondrial biology), most genes for which loss of function completely disrupts the process have already been discovered, since such extreme phenotypes are relatively easy to detect. This is evidenced by the strong enrichment for severe phenotypes among our positive control set. Many important biological functions also tend to be redundant, such that disruption of a single gene results in only a mild (but quantifiable) perturbation of the process rather than loss of function. This is likely to be even more prevalent in higher organisms, which employ far more redundancy than does S. cerevisiae, and it is also key to understanding the molecular mechanisms of many diseases. Deletion of yeast orthologs of human mitochondrial disease genes is significantly more likely to cause a modest respiratory growth defect than a severe defect [36]; similarly, since aerobic respiration is essential for mammalian viability, many disease-related mutations are unlikely to completely disrupt human mitochondrial function. Rather, these mutations tend to cause diseases by partially compromising the mitochondria [24]. Recently, Fan et al. [58] compared several mouse models of mitochondrial disease, and found that subtle mutations caused disease in adult animals, while more severe mutations were suppressed at a high frequency. Subtle mitochondrial defects accrued over time have also been of increasing recent interest as related to aging in human beings [59]. As the field continues to investigate the molecular basis of human disease and aging, the relationship between diseases and mutations incurring subtle functional perturbations is likely to extend far beyond mitochondrial biology.
Using computational techniques to generate candidate gene lists for further investigation has several advantages relative to individual high-throughput experimental screens, with comparable accuracy. First, computational data integration has the capacity to take advantage of large collections of existing publicly available experimental data; this can reveal information on a process of interest (e.g. mitochondrial function) by simultaneously examining many previous results. Additionally, computational predictions can often be generated in days or weeks, in contrast to the months or years required to conduct many traditional experimental assays. Computational integration of multiple data sources can also be less biased to any one biological feature of the candidate genes. For example, high-throughput localization studies have identified hundreds of mitochondrial genes without known functions [22],[50], but this approach would have missed the 51 genes (∼50%) discovered in this study that do not have known mitochondrial localization. This lack of bias assisted us in discovering functions for 68 of the uncharacterized genes in S. cerevisiae, all of which represent healthy and viable mutants in the yeast deletion collection with no extreme single mutant phenotype detected by previous screens. Thus, while genetic screens are important and valuable for candidate selection, computational prediction approaches integrating existing data are a viable, accurate alternative, particularly in areas with prior knowledge.
The 51 genes we confirm to be necessary for mitochondrial biogenesis that have no known mitochondrial localization raise the possibility that these mutants are somehow indirectly affecting biogenesis. Several lines of evidence argue against this possibility. First, we expect that many of these 51 proteins will localize to specific cellular structures controlling biogenesis outside of the mitochondria. For example, 13 of the 51 are known to localize to actin cytoskeleton and/or the bud neck, both structures that play intimate roles in mitochondrial transmission. Of the remaining 38 proteins, 3 were computationally predicted to localize to the mitochondria by another study [50], 11 have no known localization, and 7 have only been localized to the cytoplasm by high-throughput microscopy (which does not exclude mitochondrial localization). Further study of these 38 proteins may identify as-yet-undiscovered mitochondrial localization or highlight the importance of other cellular processes necessary for mitochondrial biogenesis (e.g. transcriptional regulation of nuclear-encoded mitochondrial genes).
Among our deletion strains exhibiting the subtle phenotype of altered petite frequency, we observed mutants with both statistically significant increases and decreases in frequency. Increased petite formation clearly indicates a failure in normal mitochondrial biogenesis or transmission. One possible explanation for a decreased petite frequency is a distinct phenotype referred to as “petite negative” [60]. Petite negative mutants display synthetic lethality or sickness with respiratory deficiency, which impairs the survival of petite cells and thus decreases their frequency. Known petite negative mutations occur in mitochondria-localized proteins that normally support the maintenance of the mitochondrial membrane potential in the absence of respiration [60]. Decreased petite frequency was observed in nine (19%) of our positive controls, two of which (FMC1 and PHB1) are known petite negative mutants [61]. Previously, traditional genetics and genome-wide screens have identified 21 petite negative mutations that result in synthetic lethality [61]. Among our 100 discoveries in mitochondrial biogenesis from single gene knockouts, we found 32 additional mutants exhibiting a decreased petite frequency indicative of non-lethal synthetic interactions. Many of the characterized petite negative genes have roles in the assembly and turnover of ATP synthase complexes, and so these genes may be a rich target for further study [61].
While additional work will be necessary to associate all of the proteins discovered in this study with specific sub-processes (such as mitochondrial genome maintenance, mitochondrial protein import and mitochondrial complex assembly), we have already identified two groups with interesting potential responsibilities in mitochondrial biogenesis. The first group is identified by comparing our glycerol growth rate data with our petite frequency results. Mitochondrial biogenesis and respiratory growth are partially overlapping processes that intersect in the translation and assembly of respiration complexes. As such, 55 of the 67 assayed mutants (82%) that exhibited an altered petite formation frequency had only weak or unaffected phenotypes in the respiratory growth assay (Table 2). The remaining 12 mutants exhibiting altered transmission rates were classified as either severe or moderate in the respiratory growth assay, thus, these mutants demonstrate both an transmission defect and a strong defect in respiration. These include four positive controls (CIT1, COX14, FMC1, and MRP49) known to be directly involved in the translation and assembly of respiratory complexes [62]–[65]. Additionally, since the beginning of this study, two of the eight additional genes in this class (MAM33 and COA1) have been shown to function in aerobic respiration [37],[66],[67]. This suggests that the remaining six genes newly characterized by this study (AIM8, AIM23, AIM24, AIM34, CTK3, and UBX4) are also functioning in the assembly of respiration complexes. Though the components of the mitochondrial complexes that generate ATP have been identified for some time in yeast, extensive chaperone, assembly, and turnover machinery for these complexes remains to be fully elucidated. The assembly and maintenance of these respiratory complexes is thus a likely role for these 8 proteins.
The second group consists of 11 genes known to be associated with the actin cytoskeleton, including AIM21 as described in Results. The biochemical functions of the other 10 proteins with respect to actin have been previously described [68]–[71], but they had no previously known mitochondrial roles. For example, Cap2p has been characterized in vitro to bind the barbed ends of actin filaments and prevent further polymerization [72], but it has not been previously implicated in mitochondrial transmission. Interestingly, many of this specific subgroup of actin-associated proteins have also been implicated in actin/membrane interactions for endocytic trafficking [73],[74]. This raises the intriguing possibility that these proteins have specialized in interactions between actin and intracellular membranes.
Our general approach can be successfully extended to other processes beyond mitochondrial biogenesis in yeast and to other organisms. We have applied our computational ensemble [26],[28],[29] to 388 other processes in Saccharomyces with promising results (Figure S2), and we report functional predictions for these processes (Dataset S2). Computational methods have also been successfully applied in other organisms with readily available genomic data collections [1],[2],[75], and the iterative nature of our approach may be particularly useful in higher eukaryotes where current functional knowledge is relatively sparse. Directing assays with computational predictions is especially attractive in higher organisms where time and resource commitments are prohibitive.
These results demonstrate the utility of employing computation to direct quantitative, functionally definitive assays. Here, we have used this technique to newly confirm the involvement of 109 proteins in the process of mitochondrial biogenesis in S. cerevisiae by assaying the frequency of petite colony formation. A subset of these proteins was also characterized using growth profiling and immunofluorescence microscopy, revealing participation in specific sub-processes of mitochondrial biogenesis. In particular, AIM21 was shown to be required for proper mitochondrial motility, a discovery which would have been difficult to make without specifically targeted computational predictions. As these techniques can be naturally extended to additional organisms and processes, close integration of computational function prediction with experimental work in other biological systems promises to quickly direct experimenters to novel facets of their areas of interest.
This protocol is adapted from the original petite frequency [31] and tetrazolium overlay [76] assays. For each mutant strain tested, we grew several replicates of the strain for 48 hours in liquid YP Gycerol at 30°C [77]. Strains able to grow on glycerol were diluted and plated for single colonies on YPD plates, which releases the requirement for functional mitochondrial DNA. Thus, as colonies formed, cells without functional mitochondrial DNA were generated. When the colony is fully formed it is a mixture cells with functional mitochondrial DNA and cells without functional mitochondrial DNA. We measured this ratio by re-suspending a colony and plating a dilution of this re-suspension such that 100–300 colonies are formed on a YPD plate. By overlaying with soft agar containing tetrazolium, colonies with functional mitochondria were stained red, while colonies without functional mitochondria remained white. The final mixture for agar overlay contains: 0.2% 2,3,5 -triphenyltetrazolium chloride (Sigma T8877), 0.067 M sodium phosphate buffer pH 7.0 and 1.5% bacto agar. The ratio of white colonies to total colonies gives the petite frequency. Eight independent petite frequencies (biological replicates) were measured for each strain tested. The distribution of these frequencies was compared to the frequency of petite generation in wild-type yeast. Strains identified as having the altered mitochondrial transmission phenotype in this assay exhibit at least a 20% change in petite frequency from wild type, and have a p-value of less than 0.05 when comparing the petite frequency distributions of that strain to the wild-type petite frequency distribution, using a Mann-Whitney U test.
The three computational systems employed in our study were bioPIXIE [26],[27], MEFIT [28], and SPELL [29]. Each was used to analyze genes involved in the GO biological process ‘mitochondrion organization and biogenesis’ (GO:0007005). All methods were initially trained and/or evaluated using the 106 annotations to this process as of April 15th, 2007. Detailed descriptions of these methods can be found in their respective publications.
42 of our initial computational predictions had strong literature evidence for involvement in mitochondrial biogenesis and inheritance and were determined to be “under-annotated” – meaning that they already had strong literature evidence for their involvement in mitochondrial organization and biogenesis, but were not yet annotated to the corresponding GO term. These 42 genes, along with 6 genes already annotated, were included as our positive control set of 48 genes. In most of these 42 cases the information was already curated by SGD in the form of annotations to other GO terms, such as ‘integral to the mitochondrial membrane’ or ‘mitochondrial protein import.’ In addition to these 42 genes, we identified an additional 95 genes that we believe have enough literature evidence to warrant their inclusion in this process without further laboratory testing, for a total of 137 “under-annotated” genes. All 137 of these genes were included in the training set for our second iteration of computational predictions.
Novel candidates for laboratory evaluation were chosen on the basis of both the three individual computational approaches as well as the ensemble of their predictions. We limited ourselves to consider only those genes with viable knockouts available in the heterozygous deletion collection [55]. Furthermore, we chose to evaluate predictions to both genes with no previously known function as well as genes known to be involved in a biological process other than mitochondrial inheritance and biogenesis. We chose the 20 most confident genes of unknown function and the 20 most confident genes with existing annotations to other biological processes from each of the three individual methods for validation. Due to overlaps between the predictions of each method, there were 87 genes in this group, however, 20 of these genes we determined to be “under-annotated” and were tested as positive controls, leaving 67 genes used as novel candidates without any prior literature evidence. We then chose an additional 74 genes from the ensemble list of predictions with no previous literature evidence to arrive at our total of 141 test candidates in our first round of laboratory evaluation.
After our first round of testing, 82 of the 141 novel predictions were discovered to have involvement in mitochondrial inheritance and biogenesis. Combined with the original 106 annotated genes and the 137 genes identified as “under-annotated,” this results in a total of 325 genes. Each of the three computational methods was re-applied using this updated training set of 325 genes and the same procedure was used to form an updated ensemble list of predictions. We selected for laboratory investigation the 52 genes with the highest confidence from the updated results that were not previously tested. The petite frequency assay was used, and an additional 17 genes demonstrated a significant phenotype.
Deletion alleles marked with the ClonNAT resistance gene (rather than the G418 KanMX resistant marker) were prepared for the four tested strains (aim17Δ, rvs167Δ, tom6Δ, and ehd3Δ). A ClonNAT marked ura3Δ allele was prepared as a control (all other strains contained a ura3Δ allele. These ClonNAT resistant strains contained the Magic Marker reporter [34] as well as can1Δ and lyp1Δ mutations to reinforce haploid selection. These five strains were crossed to a set of deletion strains marked with the G418 resistant KanMX marker, and diploids were selected on YPD-G418-ClonNAT. The diploids were then sporulated as described for our single mutant assays, except that double mutants were selected on media containing G418 and ClonNAT, and three controls were isolated for each sporulation: G418 resistant mutant, ClonNAT resistant mutants, and wild-type strains. The petite frequency assay was applied to these double mutant strains as described above. Phenotypic calls were determined for the double mutants based on the significance of the difference between the distributions of petite frequencies of the double strain versus both of the corresponding single strains. If the FDR corrected joint Wilcoxon rank sum p-value of both of these comparisons was <0.05, and the distribution of the double mutant strain was significantly different from wild type, then we scored the double mutant as significantly altered.
All S. cerevisiae strains used in this study are descended from the S288C derivative used for the deletion consortium project [55]. Methods for individual mutant manipulation are described below. Standard methods for media preparation were used as previously described [77].
The Magic Marker heterozygous yeast deletion set [34] was pinned from glycerol stocks onto enriched sporulation agar as described [78]. Single colonies developing on these random spore plates were re-struck for single colonies on the same medium and tested for presence of the G418 resistant KanMX marker [34] to identify the spore as wild-type or a deletion mutant. Single colonies that grew from this re-streaking process were picked and arranged in 96 well plates containing YPD. Each set of strains for a given candidate gene of interest were placed in a single column (1–12 of a 96 well plate); mutant isolates were placed in the first six wells (A–F) and sister wild type isolates were placed in wells G and H. These 96 well plates were glycerol stocked.
Strains were measured for their ability to grow in both respiratory (2% glycerol as carbon source) and fermentative (2% glucose as carbon source) conditions in minimal media supplemented for auxotrophies. Cultures were grown at 30°C. Growth curves were generated in a 96-well plate format (described above in “Deletion Set Manipulation”) that tests 12 mutants per run. For each mutation tested, 6 independent deletion mutants of that gene were grown in separate wells. Twenty-four replicate wild-type strains were also present in each 96-well plate format. Plates were grown and measured using a Tecan GENios plate incubator and reader, which recorded densities every 15 minutes for 24 hours for glucose cultures and 48 hours glycerol cultures. The raw growth data are available in Dataset S1.
Growth rates were derived from these curves by using Matlab to fit an exponential model:For each well, this model was fit over the entire curve, the first 2/3, and the first half; whichever yielded the best fit was used in downstream analysis (to avoid plateau effects and to model only exponential growth). Wells with an adjusted R2<0.9 were marked as non-growing, and growth rates for the remaining wells were determined by subtracting the row, column, and plate means for each well from the exponential parameter b, yielding a rate b' for each well. These b' parameters for each mutant strain were tested for significance against the total wild type population (excluding non-growing wells) using a Mann-Whitney U test. Significance was only considered for b' parameters indicating a slower growth rate than wild-type.
To detect colonies growing exponentially but with significant differences in fitness, smoothed maximum densities d were also calculated for all wells, consisting of the average of the optical density readings for the last five time points in each growth curve. From these, plate, row, and column averages were subtracted from each well, generating adjusted maxima d'. Mutants which did not double in optical density at least once (i.e. where d' was less than twice the baseline optical density) were considered to be non-growing. The remaining d' values for each mutant were again compared with the wild type values (excluding non-growing wells) using a Mann-Whitney U test. Significance was only considered for d' parameters indicating a lower saturation density than wild-type. Combined with the exponential rate tests, this assigned each mutant phenotypes in rich media and in glycerol of no effect, no growth, or significant sickness.
In either assay, mutants with inconsistent results (disagreement among more than one of the six replicates) were deemed inconclusive and marked as “mixed”. Phenotypes were never assigned based on such mixed phenotypes. For a mutant to be classified as having a respiratory growth defect, that defect was required to be specific to the glycerol media (i.e. no phenotype in glucose). If the mutant grew slowly in both glycerol and rich media, then it was not considered to have a defect in respiration.
Yeast immunofluorescence was carried out using standard methods [77]. Briefly, strains were grown to exponential phase in synthetic complete medium, and fixed in freshly prepared formaldehyde for 1 hour at 30 degrees. (Mutant strains were isolated from the Magic Marker deletion set as described above; FY4 was used as a wild type strain for comparison.) Cells were washed, digested with Zymolyase and attached to polyethyleneimine-coated coverslips. Cells were blocked with BSA, and exposed to an anti-porin antibody (Invitrogen, A-6449) and a guinea pig anti-yeast actin antibody [79]. Secondary antibodies were Alexa 488-conjugated goat anti-guinea pig (Invitrogen, A-11073) and Alexa 555-conjugated goat anti-mouse (Invitrogen A-31621). Coverslips were mounted in PBS/glycerol/phenlyenediamene. Microscopy was performed on a Perkin Elmer RS3 spinning disk confocal microscope with a 100× objective. Exposures were 1 ms per slice, and Z-stacks were taken with a 0.15 um spacing, and images were deconvoluted and assembled into 3D volumes using Volocity (Improvision).
Phalloidin staining was performed according to Methods in Yeast Genetics [77]. Briefly, strains were growth to exponential phase in synthetic complete medium, and fixed in formaldehyde (Electron Microscopy Sciences 15712-5) for one hour. F-actin was stained using Alexa 488 conjugated phalloidin (Invitrogen, A12379). Cells were deposited on polyethyleneimine-coated coverslips and mounted in PBS/glycerol/phenlyenediamene. Slides were imaged and processed as for immunofluorescence.
The NatMX cassette was cut from pAG25 [80] using NotI and ligated into the EagI site of pYX122-mtGFP, which expresses a mitochondrially targeted GFP (directed by the Su9 peptide) under the control of the triose phosphate isomerase promoter [81]. This construct was used as a template to PCR amplify the NatMX-mtGFP cassette using primers with 40 bp homology to target the cassette for integration at the dubious ORF, YDL242W. This integration was performed in the strain Y5563 to create ACY50 (strain list, Table S1). ACY50 was then mated to the Magic Marker yeast deletion set [34] and selected for haploid deletion mutants carrying the cassette as described [78].
Exponential phase cultures of S. cerevisiae in Yeast synthetic complete media (YSC) were plated onto glass slides with an agarose bed growth chamber made of low melt agarose and YSC media. The slides were covered with a cover slip and sealed using VALAP [82]. Cells were then imaged using a Perkin Elmer RS3 spinning disk confocal microscope with a 100× objective. Images of mitochondrial GFP fusions were taken using a laser emitting at 488 nm at 100% power with an exposure of 1 sec. Phase contrast images were taken using an exposure of 3 ms. For all images, 2×2 binning was used and gain was set to 255. For each field of view, both an initial z-stack of images and a time course were taken. Each z-stack was taken at intervals of 0.2 µm through the entire depth of the cells. The time course was taken in a single focal plane for two minutes at 1 frame per second. (Raw imaging files are available upon request.)
To determine the frequency of sustained mitochondrial movement resulting from Brownian motion or other passive processes [83], sustained mitochondrial movement was also measured in the presence of the metabolic inhibitors sodium azide and sodium fluoride. These inhibitors were added to the YSC agarose used for imaging; 10 mM concentrations of these inhibitors were compared to a control of 10 mM NaCl.
To measure mitochondrial motility in vivo, individual mitochondrial tips were tracked through each frame of a 2 minute time course using the Manual_Tracking plugin for ImageJ. Image files were randomly coded with numbers so that the identity of each imaged strain was not known to the investigator performing image tracking. Mitochondria tips for tracking were identified using the Z-stacks (which avoids selection of tubules that appear to be tips because other sections are out of plane). These Z-stacks were assembled into a Z-projection and merged with the phase contrast image using ImageJ to permit identification of budded cells. The selection criteria for mitochondrial tips were that the tip is initially present in the mother cell of a budded cell. In cases where both termini of a mitochondrion were available for tracking, the tip closer to the daughter cell was selected. The position of the bud neck was set as a reference and the position of a mitochondrial tip in the mother cell was plotted for each of 120 frames. The distance from the mitochondrial tip to the bud neck was calculated at each frame (in our imaging hardware, each pixel corresponded to 0.15 µm). Mitochondrial movement in each frame was then calculated by subtracting the distances to the bud neck in two consecutive frames and dividing by the time interval of 1 second. The tracking data are available in Table S8. Sustained mitochondrial movement events consisting of 3 consecutive frames of motion towards (anterograde) or 3 consecutive frames of motion away (retrograde) from the bud were identified using a custom Perl script [54]. The number of these sustained movement events per minute was calculated.
In addition to the computational predictions used to study mitochondrial biogenesis and inheritance in this study, we have applied the same prediction techniques to many additional biological processes in S. cerevisiae. We used each of our three computational methods to predict gene functions for 387 additional biological processes in the same manner as described above. The full prediction lists for all of these processes are available in Dataset S2. In order to demonstrate that these methods are able to capture information about these processes, we have also calculated the average precision (AP) of the cross-validated results as:where G is a group of genes known to be involved in a process, and ranki is the rank order of gene i in the prediction results. A graph of the average precisions for all 388 processes (including mitochondrial organization and biogenesis, highlighted in red) is shown in Figure S2.
Orthology between yeast and human genes was based on orthologous clusters in the Homologene [84], Inparanoid [85], and OrthoMCL [86] databases as of June 2007. Each of these uses a published algorithm for determining clusters of orthologous genes, i.e. groups of genes thought to be conserved and perform near-identical functions in different organisms. We first took the union of these databases as applied to a core set of diverse organisms (yeast, human, mouse, fly, and worm), considering a gene pair to be orthologous if declared so by any of the three databases. This resulted in a set of unified orthologous clusters, from which we eliminated any cluster containing more than 50 genes. This resulted in 14,528 clusters spanning 61,702 genes in the five organisms, and from this set we report here the human orthologs of S. cerevisiae genes in our study.
Disease related human orthologs were determined based on the manual curation of the Online Mendelian Inheritance in Man (OMIM) resource [87], and the automated text mining available through GeneCards [88]. We considered all of the OMIM curations valid, while we required at least 2 independent publication citations in GeneCards for a disease relation to be valid.
Both mitochondrial and actin localization was based on the Gene Ontology cellular component curation. For mitochondrial localization curation to the term GO:0005739: mitochondrion was used. In addition six genes were marked as computationally predicted to the mitochondrion based on the study by Prokisch et al. [50]. For actin localization curation to the term GO:0015629: actin cytoskeleton was used.
|
10.1371/journal.pbio.1000287 | A Mitotic Phosphorylation Feedback Network Connects Cdk1, Plk1, 53BP1, and Chk2 to Inactivate the G2/M DNA Damage Checkpoint | DNA damage checkpoints arrest cell cycle progression to facilitate DNA repair. The ability to survive genotoxic insults depends not only on the initiation of cell cycle checkpoints but also on checkpoint maintenance. While activation of DNA damage checkpoints has been studied extensively, molecular mechanisms involved in sustaining and ultimately inactivating cell cycle checkpoints are largely unknown. Here, we explored feedback mechanisms that control the maintenance and termination of checkpoint function by computationally identifying an evolutionary conserved mitotic phosphorylation network within the DNA damage response. We demonstrate that the non-enzymatic checkpoint adaptor protein 53BP1 is an in vivo target of the cell cycle kinases Cyclin-dependent kinase-1 and Polo-like kinase-1 (Plk1). We show that Plk1 binds 53BP1 during mitosis and that this interaction is required for proper inactivation of the DNA damage checkpoint. 53BP1 mutants that are unable to bind Plk1 fail to restart the cell cycle after ionizing radiation-mediated cell cycle arrest. Importantly, we show that Plk1 also phosphorylates the 53BP1-binding checkpoint kinase Chk2 to inactivate its FHA domain and inhibit its kinase activity in mammalian cells. Thus, a mitotic kinase-mediated negative feedback loop regulates the ATM-Chk2 branch of the DNA damage signaling network by phosphorylating conserved sites in 53BP1 and Chk2 to inactivate checkpoint signaling and control checkpoint duration.
| DNA is constantly damaged both by factors outside our bodies (such as ultraviolet rays from sunlight) and by factors from within (such as reactive oxygen species produced during metabolism). DNA damage can lead to malfunctioning of genes, and persistent DNA damage can result in developmental disorders or the development of cancer. To ensure proper DNA repair, cells are equipped with an evolutionarily conserved DNA damage checkpoint, which stops proliferation and activates DNA repair mechanisms. Intriguingly, this DNA damage checkpoint responds to DNA damage throughout the cell cycle, except during mitosis. In this work, we have addressed how cells dismantle their DNA damage checkpoint during mitosis to allow cell division to proceed even if there is damaged DNA present. Using the observation that kinases phosphorylate their substrates on evolutionarily conserved, kinase-specific sequence motifs, we have used a combined computational and experimental approach to predict and verify key proteins involved in mitotic checkpoint inactivation. We show that the checkpoint scaffold protein 53BP1 is phosphorylated by the mitotic kinases Cdk1 and Polo-like kinase-1 (Plk1). Furthermore, we find that Plk1 can inactivate the checkpoint kinase Chk2, which is downstream of 53BP1. Plk1 is shown to be a key mediator of mitotic checkpoint inactivation, as cells that cannot activate Plk1 fail to properly dismantle the DNA damage checkpoint during mitosis and instead show DNA damage-induced Chk2 kinase activation. Two related papers, published in PLoS Biology (Vidanes et al., doi:10.1371/journal.pbio.1000286) and PLoS Genetics (Donnianni et al., doi:10.1371/journal.pgen.1000763), similarly investigate the phenomenon of DNA damage checkpoint silencing.
| Throughout the life of an organism, cellular DNA constantly encounters chemical and radiation-induced damage. Solar and terrestrial sources of radiation, along with the oxidative by-products of normal metabolism, result in chemical modifications of DNA bases and disruption of the sugar phosphate backbone. Additional DNA lesions, including mismatched bases, and single- or double-stranded DNA breaks, also arise during the process of replication, which is not an error-free process [1]. To cope with these types of genotoxic damage, cells activate powerful DNA damage-induced cell cycle checkpoints that coordinate cell cycle arrest with recruitment and activation of the DNA repair machinery [2]–[6]. Depending on the amount of damage and the specific cell type, cross-talk between the checkpoint and repair pathways with pathways involved in programmed cell death leads to the elimination of irreparably damaged cells by apoptosis [7]. The global importance of these cell cycle checkpoint pathways in maintaining genomic integrity is highlighted by the observation that loss, mutation, or epigenetic silencing of checkpoint genes is frequently observed in cancer [1],[4]. Conversely, deletion of checkpoint genes in non-neoplastic cells has been shown to cause genomic instability and predisposition to transformation [1],[4].
Loss of DNA damage checkpoints during early stages of tumorigenesis not only facilitates the acquisition of additional mutations over time [8],[9] but can also be exploited in various forms of human cancer treatment. Radiotherapy as well as many types of anti-tumor chemotherapy are believed to preferentially kill tumor cells by generating extensive amounts of DNA damage that promotes cell death in checkpoint-compromised tumors, but not in the surrounding non-neoplastic tissue where the checkpoint and repair pathways are intact [10]. The primary cytotoxic lesion created by therapeutic radiotherapy and most other genotoxic treatments are DNA double-strand breaks (DSBs). It has been estimated that a single unrepaired DSB is sufficient for cell lethality [11].
Early events following DSB generation include local alterations in chromatin structure, recruitment of the Mre11-Rad50-Nbs1 mediator complex to the DNA, and phosphorylation of the variant Histone H2AX by an initial wave of activation of the checkpoint kinase ATM [2],[12]–[14]. Subsequent recruitment of the protein MDC1 dramatically enhances further local activation of ATM as part of a positive feedback loop, which in turn recruits molecules like 53BP1 and BRCA1 [15]–[17]. 53BP1 facilitates DNA repair by the error-prone non-homologous end joining (NHEJ) pathway [18],[19], while BRCA1 is important for DNA repair by the error-free homologous recombination pathway during the S and G2 phases of the cell [20]. A major target of ATM is the effector kinase Chk2, a critical effector kinase that functions downstream of ATM to arrest the cell cycle after DSBs by inactivating phosphatases of the Cdc25 family through catalytic inactivation, nuclear exclusion, and/or proteasomal degradation [21],[22]. This, in turn, prevents Cdc25 family members from dephosphorylating and activating Cyclin-Cdk complexes, thereby initiating G1/S and G2/M cell cycle checkpoints.
In order for cells to survive DNA damage, it is important that cell cycle arrest is not only initiated but also maintained for the duration of time necessary for DNA repair. Mechanisms governing checkpoint initiation versus maintenance appear to be molecularly distinct. This was initially demonstrated by the observation that interference with specific checkpoint components can leave checkpoint initiation intact but disrupt checkpoint maintenance, leading to premature cell cycle reentry accompanied by death by mitotic catastrophe [7],[15],[23]–[25]. Although the process of checkpoint termination and cell cycle reentry has not been studied extensively, the existing data suggest that inactivation of a checkpoint response is an active process that requires dedicated signaling pathways, such as the Plk1 pathway [2],[26],[27]. Intriguingly, a number of proteins involved in terminating the maintenance phase of a DNA damage checkpoint also play critical roles during later mitotic events, suggesting the existence of a positive feedback loop in which the earliest events of mitosis involve the active silencing of the DNA damage checkpoint through one or more mechanisms that remain unclear.
Checkpoint silencing has been best studied in the budding yeast S. cerevisiae and has revealed several essential genes in this process, for example the phosphatases Ptc2 and Ptc3, Casein kinase-I, and Srs1 [28]–[30]. In addition, the Polo-like kinase Cdc5 is required for silencing checkpoint signaling, and this requirement appears to be widely conserved, since S. cerevisiae, X. Leavis, and human cells all depend on Plks for silencing of the S-phase or G2 checkpoints, respectively [29],[31]–[33]. The activity of Polo-like kinases has been shown to be required for inactivation of the ATR-Chk1 pathway and the Wee1 axis of checkpoint signaling. Specifically, Plk1 was shown to create β-TrCP-binding sites on both Wee1 and the Chk1 adaptor protein Claspin, resulting in efficient ubiquitin-mediated degradation of these target proteins [32]–[36]. Thus far, only inactivation of checkpoint components of the ATR-Chk1-Wee1 signaling axes has been identified in relation to maintenance and termination of cell cycle checkpoints. DSBs, however, primarily trigger a checkpoint arrest through the ATM-Chk2 signaling pathway. How, and if, the ATM-Chk2 signaling axis is actively silenced during release of the G2 DNA damage checkpoint is currently unclear. Here, we analyzed potential feedback mechanisms responsible for terminating this process. We reasoned that inactivation of cell cycle checkpoints after DSBs should involve at least two arms of the ATM-Chk2 checkpoint response—both the upstream sensor arm that maintains activation of ATM and the downstream effector arm that functions at and below the level of Chk2 must be silenced in order to facilitate cell cycle reentry. By using a combination of evolutionarily constrained bioinformatics analysis together with cell cycle–specific modifications of the highly conserved DNA damage checkpoint signaling network, we identified the Cdk- and Plk1-dependent phosphorylation of 53BP1 and Chk2 as critical checkpoint-inactivating events in the sensor and effector arms of the G2/M checkpoint pathway, respectively, that are important for checkpoint termination and cell cycle reentry.
To identify potential feedback and control mechanisms that extinguish the ATM-Chk2 signaling axis of the G2/M DNA damage checkpoint, we initially investigated whether we could observe silencing of this network under particular cell states or conditions. Molecular targets that are known to be inactivated in other G2/M cell cycle checkpoint control pathways, i.e. the ATR/Chk1 pathway, include Wee1 and Claspin [32]–[36] and inactivation of these components results in a shutdown of this checkpoint signaling pathway following mitotic entry. If the ATM-Chk2 pathway was also inactivated upon mitotic entry, clear differences would be expected when interphase cells are compared to mitotic cells following irradiation. To examine this, U2OS cells were exposed to 10 Gy of ionizing radiation (IR), and activation of the upstream checkpoint kinase ATM and the downstream effector kinase Chk2 were examined by immunoblotting (Figure 1A, B). In order to investigate whether mitotic cells remained in mitosis upon irradiation in our experimental set-up, we used two mitotic markers, MPM-2 immunoreactivity and the presence of Plk1 (Figure 1B). The monoclonal MPM-2 antibody was originally cloned on the basis of its ability to specifically recognize mitotic but not interphase cells [37]. MPM-2 recognizes multiple mitosis-specific phospho-proteins, and its reactivity thus indicates the abundance of mitotic cells. Plk1, on the other hand, is highly expressed in G2 and M-phase of the cell cycle and is degraded during mitotic exit [38]. Importantly, we observed that irradiation of mitotic cells did not lead to mitotic exit, as judged by the persistently elevated levels of Plk1 and MPM-2 immunoreactivity (Figure 1B). As shown in Figure 1A, in response to IR, ATM was efficiently activated regardless of cell cycle phase. We observed both rapid phosphorylation of Chk2 on Thr-68, a known ATM phosphorylation site, and enhanced Chk2 kinase activity (Figure 1B,C), after irradiation of interphase cells. However, irradiation of mitotic cells failed to result in phosphorylation of Chk2 on Thr-68, and the DNA damage-induced increase in Chk2 kinase activity was severely impaired (Figure 1B,C). This suggests that ATM may not efficiently form complexes with some of its critical downstream substrates such as Chk2 in response to DNA damage during mitosis, resulting in a failure to activate Chk2 and Chk2-dependent effector pathways required for cell cycle arrest. This hypothesis is in line with previous reports in which γ-irradiation or treatment with topoisomerase inhibitors were shown not to interfere with progression of cells already in mitosis [39],[40], indicating that DNA damage checkpoint pathways are functionally inactivated during mitosis.
To elucidate potential molecular mechanisms responsible for checkpoint silencing of the ATM-Chk2 axis in mitosis, we used a supervised computational network/bioinformatics approach. First, we identified a set of core proteins involved in the human G2/M checkpoint and mapped known in vivo phosphorylation sites [41]–[44] onto them (Figure 2A,B and Table S1). Next, this set of phospho-proteins was used to query for conservation of the phosphorylation sites, defined by five residues N-terminal and five residues C-terminal that flank the mapped phospho-residue, in protein orthologs across eleven vertebrate genomes. We computed the conservation as the mean percentage of conserved residues within this eleven-mer site window across these vertebrate genomes. The kinases responsible for generating these phosphorylation sites were identified using data from PhosphoELM [42] or predicted using the NetworKIN algorithm [45]–[47]. In addition, we used Scansite [48] to identify potential docking sites for the Plk1 Polo-Box Domain (PBD) [44],[49],[50] within the network. As would be expected, we observed that many of the checkpoint proteins contained highly conserved ATM/ATR sites (Figure 2A,B and Table S1). Importantly, we also identified highly conserved phosphorylation sites for Cdk1/2 and Plk1 kinases distributed relatively equally on proteins throughout the network, independently of whether the proteins were classified into “checkpoint” or “cell cycle” modules. No potential molecular targets could be uniquely pinpointed by looking only at the putative kinase-substrate level; thus the mitotic/DNA damage phosphorylation network seems to be robust in the sense that they are highly connected via relatively few but pleotropic kinases. However, when we searched for PBD binding sites, only a few network components appeared (Figure 2B) including the previously validated Plk1 binding target Cyclin B [51]. In addition, several components of the checkpoint signaling pathway appeared as putative Plk1 PBD-binding targets, notably MDC1 and 53BP1. Surprisingly, these two proteins belong to the non-enzymatic checkpoint adaptor family of proteins that function in the ATM-Chk2 pathway [16],[52]–[57].
We focused on 53BP1, since our analysis predicted eight highly conserved Cdk1/2 phosphorylation sites as well as three sites with lower conservation. Importantly, five of the highly conserved Cdk1/2 phosphorylation sites constitute putative PBD binding sites. We have previously shown that 53BP1 is a target of Cdk1-Cyclin B during mitosis [45]. Here, we aimed to investigate the functional implications of these phosphorylation events and again employed the MPM-2 antibody, which recognizes proteins that are phosphorylated on Cdk1/2 consensus motifs [37],[58],[59]. By immunoprecipitating 53BP1 from mitotic cell extracts, we observed clear immunoreactivity with the MPM-2 antibody, in stark contrast to 53BP1 immunoprecipitated from interphase cells (Figure 3A). These results were further strengthened by in vitro kinase assays, in which recombinant Cdk1-Cyclin B, but not Cdk2-CyclinA, efficiently phosphorylated 53BP1 (Figure 3B).
If 53BP1 is a critical target for checkpoint silencing by mitotic kinases, then the function of 53BP1 should be altered during mitosis. We therefore investigated the co-localization of 53BP1 and DNA damage–induced foci at different cell cycle phases. Few γ-H2AX foci were observed in untreated cells, while their number increased dramatically after 3Gy of IR (Figure 3C, Figure S1B left panel). Similar behavior was observed for 53BP1 (Figure 3C, Figure S1B left panel). In interphase cells, approximately 70% of γ-H2AX foci contained 53BP1 (Figure 3C, Figure S1B middle panel). However, when nuclear foci of 53BP1 were present, they always overlapped with γ-H2AX in both untreated as well as in IR-treated cells during interphase (Figure 3C, Figure S1B right panel). In contrast, in mitotic cells there were essentially no distinct 53BP1 foci that were observed regardless of the presence or absence of irradiation, and instead 53BP1 appeared to be largely excluded from chromatin. Hence, in mitosis no overlap was detected between the localization of γ-H2AX foci and 53BP1, showing that the function of 53BP1 in the DNA damage response (DDR) is indeed modified during mitosis, either directly or indirectly (Figure 3C).
In addition to changes in IR-induced localization of 53BP1, we also observed that the protein levels of 53BP1 rapidly declined as cells passed synchronously through the cell cycle (Figure 3D). However, the decrease in 53BP1 protein levels occurred only at late stages of mitosis or cell cycle reentry, after the destruction of Cyclin B, and thus may not have a prominent role in G2 checkpoint inactivation.
The NetworKIN algorithm, in addition to predicting 53BP1 as a substrate for Cdk1, also predicted putative Plk1 phosphorylation and PBD binding-site(s) in 53BP1 (Figure 2B). To investigate the functional roles of these sites, we immunoprecipitated endogenous 53BP1 from interphase or mitotic U2OS cells and examined the immunoprecipitates for co-association of Plk1 (Figure 4C). Whereas 53BP1 and Plk1 did not co-immunoprecipitate during interphase, a significant amount of Plk1 interacted with 53BP1 during mitosis (Figure 4C and Figure S1A). In addition to binding to 53BP1, Plk1 was able to efficiently phosphorylate 53BP1 in vitro (Figure 4D). To further identify the site(s) in 53BP1 that interact with Plk1, mutational analysis was performed (a selection of phosphorylation site mutants is indicated in Figure 4A,B). Lysates of interphase or mitotic U2OS cells stably expressing wt or mutant forms of GFP-tagged murine-53BP1 were incubated with the recombinant GST-tagged PBD from Plk1 (residues 363–562). Mitotic forms of 53BP1 show reduced migration on low percentage SDS-page gels, resulting in multiple bands in mitotic lysates from cells expressing endogenous as well as GFP-tagged 53BP1 (Figure 4E). As expected, wt-GFP-m53BP1 was efficiently pulled down by GST-PBD from mitotic lysates, but not from interphase lysates (Figure 4E). Like wt-GFP-m53BP1, both the GFP-m53BP1-1103A and mGFP-m53BP1-1620A mutants (corresponding to residues 1114 and 1635 in human 53BP1) efficiently bound to the Plk1 PBD (Figure 4F and unpublished data). A third predicted PBD binding site within 53BP1 (S380) resides in a cluster of potential PBD binding sites (some of which have not yet been shown to be phosphorylated in vivo). A 53BP1 mutant lacking this cluster of potential PBD binding sites (GFP-m53BP1 Δ196–439) did not interact with the PBD of Plk1 (Figure 4F). Importantly, the single highly conserved and predicted PBD-binding site that is found phosphorylated in vivo, S380 (corresponding to the S376 in murine 53BP1), appeared to be essential for the mitotic interaction between 53BP1 and Plk1, as the GFP-m53BP1-376A mutant could not be precipitated from mitotic lysates with recombinant PBD (Figure 4F). Furthermore, re-analysis of in vivo phosphorylation sites from mitotic Plk1 PBD pull-downs revealed the presence of phospho-S380 peptides from endogenous 53BP1 [44]. Combined, these results indicate that S380 is a critical site amongst the predicted CDK1/2 sites that is required for stable binding to Plk1. Although mass-spectrometry-based phospho-proteomics previously identified S380 as an in vivo phosphorylation site [41],[44], the dynamics of S380 phosphorylation during different phases of the cell cycle are unclear. In agreement with a model in which S380 is phosphorylated during mitosis, we could only observe S380 phosphorylation of 53BP1 in mitotically arrested cells, but not in interphase cells using a phospho-specific antibody raised against this site (Figure 4G). Furthermore, detailed analysis of phosphorylation during the cell cycle revealed intense phosphorylation of S380 when synchronized cells entered mitosis (Figure 3D), consistent with this site being a Cdk1 target. Finally, treatment of mitotic cells with the Cdk1 inhibitor roscovitine eliminated S380 phospho-reactivity (Figure 4G).
Although the identification of mitotic phosphorylation sites in DNA damage checkpoint proteins can elucidate potential feedback targets within the checkpoint networks, it is conceivable that mitotically phosphorylated checkpoint proteins could also possess alternative cellular functions. Mitotic phosphorylation of such proteins could, for example, be important for the regulation of normal mitotic progression, rather than facilitating feedback control during an exogenous G2 DNA damage checkpoint response. To investigate a possible role for 53BP1 during an unperturbed mitosis, we stably infected U2OS or MCF7 cell lines with 53BP1 RNAi hairpins and examined these cells for possible defects in mitotic progression (Figure 5). We used two independent hairpins that significantly decreased 53BP1 levels in both U2OS and MCF7 cell lines (Figure 5A). To select for a functional 53BP1 knockdown, MCF7 cell lines were treated with the MDM2 inhibitor Nutlin-3 [60]. Nutlin-3 treatment leads to a cell cycle arrest that depends on p53 as well as 53BP1 [60],[61]. As expected and reported previously, knockdown of 53BP1 significantly increased mitotic indices, the number of cells in S-phase, as well as the size and number of proliferating colonies following Nutlin-3 treatment (Figure 5B,C,D and unpublished data) in the 53BP1 knockdown cells. Although the increases in M- and S-phase content after Nutlin treatment in 53BP1 knockdown cells is minor, the increase in colony formation suggests that this effect is meaningful. A functional 53BP1 knockdown was also evidenced by a small but highly reproducible increase in mitotic content after low dose (2 and 3 Gy) ionizing radiation (unpublished data). In contrast, no differences in mitotic indices were observed in the untreated cell population, indicating that loss of 53BP1 does not interfere with normal mitotic progression. Likewise, paclitaxel treatment resulted in similar increases in the percentages of mitotic cells in control and 53BP1-depleted lines, suggesting that 53BP1 is not required for normal functioning of the spindle assembly checkpoint (Figure 5E).
Our finding that 53BP1 is not involved in spindle checkpoint functioning allowed us to use microtubule poisons to trap cells in mitosis after checkpoint escape, even in cells with modulated 53BP1 expression levels. In these experiments, if the observed mitotic phosphorylation of 53BP1 is important for attenuating the DNA damage checkpoint, one would expect to observe altered kinetics of G2-M transition when phosphorylation site mutants of GFP-m53BP1 are expressed, especially after cells are treated with genotoxic compounds. To first assess how phosphorylation by mitotic kinases alters the function of checkpoint components such as 53BP1, we utilized genetic and chemical inhibition of Plk1. Previously, a role for Plk1 in checkpoint silencing was identified by using siRNA technology [32]–[36]. Although clear differences in cell cycle reentry were observed after silencing Plk1 expression, a limitation of these RNAi experiments is that they cannot distinguish between a requirement for the mere presence of Plk1 in checkpoint recovery or for the enzymatic activity of Plk1 during this process. We therefore wished to confirm these results using the temporally controlled chemical inhibition of Plk1 [62]. As previously reported, chemical inhibition of Plk1 using BI-2536 led to spindle checkpoint activation and a concomitant mitotic arrest [63] with kinetics similar to those seen in nocodazole- or paclitaxel-treated cells (Figure 6A and unpublished data). Moreover, when the G2 DNA damage checkpoint was activated in U2OS cells by γ-irradiation, and the checkpoint then abrogated by treatment of the damaged cells with the ATM/ATR inhibitor caffeine, the cells rapidly entered mitosis, where they could be trapped in the presence of paclitaxel (Figure 6B). In contrast, cells treated with the Plk1 inhibitor were unable to enter mitosis and remained in G2, clearly indicating that Plk1 kinase activity, rather than physical presence of Plk1 per se, is required for cell cycle reentry after a DNA damage checkpoint arrest when the upstream checkpoint signaling pathways are silenced with caffeine. This effect does not appear to result from DNA damage induced by Plk1 inhibition, as was previously suggested [64], since Plk1 inhibition did not initiate DNA damage-induced foci (Figure S1C).
In addition to caffeine-induced checkpoint abrogation, we could show that Plk1 activity was equally important for spontaneous checkpoint recovery (Figure 6C). In response to low dose IR (2 Gy), U2OS cells delay cell cycle progression for up to 8 h, during which time cumulative mitotic entry is significantly lower (Figure 6C). When cells were treated with the Plk1 inhibitor following low-dose DNA damage checkpoint activation, similarly low mitotic indices were observed. However, unlike control cells in which the mitotic index had recovered to approximately 80% of the levels seen in unirradiated cells by 16 h after 2 Gy ionizing radiation, cells that were irradiated and treated with the synthetic Plk1 inhibitor maintained persistently low mitotic indices (Figure 6C). These results confirm a specific role for the kinase activity of Plk1 in spontaneous cell cycle reentry after a G2 DNA damage checkpoint arrest, as well as the requirement for Plk1 for normal mitotic progression beyond metaphase [31],[32],[34],[35],[65],[66].
Next, to explore whether the interaction of 53BP1 with Plk1 was important for the DNA damage recovery phenotype, we irradiated U2OS cells, expressing GFP-tagged wt-m53BP1 or a GFP-53BP1 mutant that was unable to bind Plk1 (Figure 6D), and monitored persistence of DNA damage checkpoint activity 24 h later by quantitatively measuring levels of H2AX phosphorylation by flow cytometry. As shown in Figure 6D, both the control untransfected cells and the cells expressing wt-53BP1 showed only background levels of γ-H2AX staining by this time after irradiation. In contrast, 24 h after irradiation cells expressing the Plk1-binding mutant GFP-m53BP1-S376A showed persistently increased γ-H2AX-positivity (Figure 6D). To assess the effects of such altered checkpoint activation on cell cycle progression, a parallel set of studies was performed in the absence (Figure 6E) or presence of low-dose IR (Figure 6F), and mitotic entry quantified by measuring phospho-Histone H3 staining in the presence of paclitaxel to trap all cells exiting G2 in mitosis. As shown in Figure 6E, in the absence of DNA damage cells, expressing the S376A-m53BP1 mutant showed no reduction in mitotic entry—if anything, the percentage of pH3-positive cells was slightly increased in m53BP1 mutant-expressing cells. In contrast, cells expressing S376A-m53BP1 were delayed in mitotic entry after irradiation with low-dose IR compared to either untransfected cells (unpublished data) or cells expressing wt-m53BP1 (Figure 6F), in agreement with the observed increase in checkpoint activity. These results strongly suggest that mitotic regulation of 53BP1 by Plk1 modulates DNA damage checkpoint activity to control checkpoint recovery.
It was previously suggested that 53BP1 functions as a molecular platform/scaffold for the efficient recruitment, phosphorylation, and activation of several checkpoint components including p53, BRCA1, and Chk2 [57],[67]–[70]. Chk2 is a Ser/Thr kinase that possesses an SQ/TQ-rich N-terminus, an N-terminal phosphopeptide-binding Forkhead-Associated (FHA) domain that is crucial for Chk2 activation, and a C-terminal kinase domain. Specifically, 53BP1 was shown to be required for Chk2 activation in response to DNA damage, as Chk2 activation was shown to be significantly impaired in 53BP1 null cells and in cells where 53BP1 was depleted by RNAi [57],[69],[70], particularly when exposed to low doses of IR [70], or when signaling through the MDC1 branch of the DNA damage signaling pathway is suppressed [69],[71],[72]. Interestingly, the inability of Chk2 to be activated during mitosis (Figure 1B,C) strongly correlates with the absence of 53BP1 from DNA damage–induced foci in irradiated mitotic cells (Figure 3C) and with the mitotic phosphorylation of 53BP1 on Ser-376 to generate a Plk1 PBD binding site. These data suggest that 53BP1 may function as a docking platform where Plk1 and Chk2 can bind and possibly interact.
To test the hypothesis that Plk1 kinase activity could inhibit Chk2 as part of the mechanism of checkpoint inactivation, we first examined whether the activity of Plk1 could be responsible for the inability of DNA damage to activate Chk2 during mitosis (Figure 1B,C). In these experiments, U2OS cells were treated with nocodazole in the absence or presence of the Plk1 inhibitor BI 2536, and mitotic cells then isolated and irradiated with 5 Gy of ionizing radiation. Chk2 activity was measured 1 h after irradiation using an immunoprecipitation/in vitro kinase assay (Figure 7A). No increase in Chk2 kinase activity was observed in the irradiated mitotic cells compared to the unirradiated mitotic cells, as expected. If the mitotic cells were treated with the Plk1 inhibitor, however, a marked elevation of Chk2 kinase activity was seen after DNA damage, consistent with a model where Plk1 kinase activity suppresses Chk2 activity during mitosis. We next examined whether Chk2 could be a direct substrate of Plk1. As shown in Figure 7B, incubation of full-length Chk2 with Plk1 in the presence of [32P]-γ-ATP resulted in significant Chk2 phosphorylation, as visualized by 32P incorporation and a clear phosphorylation-induced mobility shift (Figure 7B). In order to examine whether these effects could be recapitulated in vivo during checkpoint recovery, we irradiated U2OS cells expressing FLAG-tagged Chk2 in the absence or presence of Plk1 inhibitor (Figure 7C). Following checkpoint inactivation using caffeine, FLAG-Chk2 was immunoprecipitated and analyzed by SDS-PAGE. Cells treated with the Plk1 inhibitor showed a markedly faster migrating form of Chk2, confirming that the Plk1-dependent modification that was observed in vitro also occurs in vivo. Surprisingly, in vitro phosphorylation of Chk2 by Plk1 had only a minor effect on the ability of the Chk2 kinase domain to phosphorylate an optimal peptide substrate (Figure 7D). In marked contrast, in vitro phosphorylation of the FHA domain of Chk2 by Plk1 completely abrogated the ability of the FHA domain to bind to its phosphopeptide ligands (Figure 7E). Since the FHA domain is known to be critical for DNA damage–induced phosphorylation, oligomerization, and activation of Chk2 in vivo [73]–[76], our results indicate that loss of Chk2 activation and function in cells during both mitosis and recovery from a DNA damage checkpoint likely involves contributions from both Plk1 binding to 53BP1 and direct phosphorylation-induced inactivation of the Chk2 FHA domain. To further examine this, the Plk1 phosphorylation sites within the FHA domain of Chk2 were mapped using nano-liquid chromatography and mass spectrometry (Figure 7F and Figure S2A–C), revealing three sites, Ser-164, Thr-205, and Ser-210, that are both evolutionarily conserved and match the optimal phosphorylation motif for Plk1 ([77]; Alexander and Yaffe, manuscript in preparation). Mapping of these sites onto the X-ray crystal structures of the Chk2 FHA∶phosphopeptide complex [78] and the recently solved structure of the near-full-length Chk2 dimer (Figure 7G) [79] reveals that one of these sites, Ser-164, is in close proximity to the phosphopeptide-binding site, with its phosphorylation likely to disrupt ligand binding through electrostatic repulsion of the ligand phosphothreonine residue (Figure 7G right panel). Both Thr-205 and Ser-210 lie at the interface between the two monomers in the dimeric Chk2 structure that is believed to represent the early stages in the Chk2 activation process [79]. Phosphorylation of these residues would be expected to disrupt both the dimeric FHA∶FHA domain interaction as well as the interaction between the FHA domain of one monomer with the kinase-FHA linker of the other (Figure 7G left panel). It is not technically possible to directly assay Plk1-dependent alterations in phosphopeptide-binding capacity of the Chk2 FHA domain within cells expressing wild-type or mutant 53BP1. Therefore, to determine if phosphorylation of the FHA domain by Plk1 contributes to the observed Plk1 dependence of checkpoint silencing, we tested whether mutation of the identified phosphorylation sites affected the ability of cells to recover from a DNA damage checkpoint arrest. In these experiments, cells were transfected with wild-type or mutant forms of Chk2 in which each of the phosphorylation sites was replaced by Ala, along with an IRES-driven GFP (Figure 7H). Expression of wild-type or mutant forms of Chk2 did not result in altered cell cycle distributions under untreated conditions (Figure 7H). In marked contrast, mutation of Ser-164, Thr-205, or Ser-210 to a non-phosphorylatable residue was found to clearly impair checkpoint recovery, as judged by a significant decrease in cumulative mitotic entry at 24 h after irradiation (Figure 7I), with mutation of Ser-164 showing the greatest effect. These results show that Chk2 phosphorylation by Plk1 inhibits the function of the FHA domain and that these phosphorylation events contribute to inactivation of the DNA damage checkpoint during mitosis and checkpoint recovery.
In response to genotoxic injury, cells activate a network of DNA damage signaling pathways involving the upstream serine/threonine kinases ATM and ATR and the downstream kinases Chk1, Chk2, and MK2 to induce G1, S, and G2 cell cycle arrest, recruit repair machinery to the sites of damage, and target irreversibly damaged cells for apoptosis [4],[80]. ATR and its downstream effector kinase Chk1 are essential genes that respond primarily to single-strand DNA lesions and bulky base modifications. In contrast, the ATM-Chk2 signaling pathway, which is activated by DSBs (considered to be the most lethal type of DNA damage), is composed of nonessential genes. Their importance, however, is highlighted by the observation that interference with ATM and Chk2 function severely impairs the checkpoint response to IR and other DSB-inducing lesions, and mutation of the genes encoding for ATM and Chk2 results in the cancer-prone Ataxia-Telangiectasia syndrome, and familial breast and prostate cancer, respectively [81]–[87]. Following DNA repair, cells must extinguish the DNA damage signal to allow the cells to reenter the cell cycle, but the mechanisms through which this occurs, particularly with respect to the ATM-Chk2 pathway, are poorly understood.
Since DNA damage checkpoints respond to as little as a single DNA DSB in model systems [25],[26], it has long been assumed that human cells also maintain the G2/M checkpoint until all of the breaks are repaired. Recent evidence, however, shows that the G2 checkpoint in immortalized human cells in culture displays a defined threshold of approximately 10–20 DSBs [23]. Limited checkpoint control was not only apparent in response to IR doses that cause very few DNA DSBs, cells that had also repaired more extensive amounts of DNA damage also showed checkpoint release when fewer than 10–20 DSBs were left unrepaired [23]. Although the fate of cells that continue proliferating in the presence of unrepaired DNA breaks is unclear, and the identity of the rate-limiting DNA damage checkpoint components has yet to be uncovered, a picture is emerging in which certain cues are capable of overriding the DNA damage checkpoint machinery. G2 checkpoint escape in the presence of unrepaired DNA damage may be particularly common during the evolution of cancer cells [2],[4],[5],[8],[9], reinforcing the need to better understand this process in molecular detail.
Recently, a pathway comprising Aurora A, Bora, and Plk1 was shown to control inactivation of the G2 DNA damage machinery [65],[88]. Although several targets of Plk1 within or downstream from the ATR-Chk1 pathway that are involved in DNA damage checkpoint silencing have been described, no target within the ATM-Chk2 pathway has been identified thus far [32]–[36]. Here, we have used a combined bioinformatics and biochemical approach to identify targets of mitotic kinases within the DNA damage checkpoint. We show that the 53BP1 checkpoint protein interacts with Plk1 and is phosphorylated by Cdk1/Cyclin B and Plk1. In addition, we show that expression of a 53BP1 mutant that is unable to interact with Plk1 prevents proper checkpoint release. 53BP1 was previously identified as a non-enzymatic DNA damage checkpoint mediator protein that is recruited to sites of DNA damage through protein-protein interactions, oligomerization, and binding to methylated histones [89]–[94]. Although the recruitment of 53BP1 to sites of DNA damage has been studied intensively, the exact functions of 53BP1 are only beginning to emerge. 53BP1 was recently shown to regulate DNA repair as a component of the NHEJ network [18],[19],[95]. In addition, 53BP1 regulates checkpoint responses by interacting with a range of downstream checkpoint components, including Chk2 and p53 [57],[61],[69],[70].
Our results strengthen a role for 53BP1 as a checkpoint regulator and indicate that 53BP1 functions as a binding platform for Plk1 during the checkpoint recovery process. This suggests a model in which 53BP1 might mediate a direct interaction between Plk1 and the 53BP1-binding protein Chk2. We suggest that mitotic Cdk1 phosphorylation of 53BP1 and subsequent interaction of Plk1 and 53BP1 may function to bring Plk1 and the 53BP1-interacting protein Chk2 in close proximity (Figure 8, step 1). Subsequent direct phosphorylation of Chk2 by Plk1 (Figure 8, step 2) leads to impaired Chk2 phosphopeptide-binding ability by its FHA domain, which is required for continued Chk2 activation and function in cell cycle arrest (Figure 8, step 3). Our results fit well with previous observations in fission yeast in which a prolonged DNA damage–induced checkpoint arrest was observed when Cdk phosphorylation site mutants of the 53BP1 homologue Crb2 were expressed [96]. The budding yeast Polo-like kinase homologue Cdc5 has also been shown to be required for DNA damage checkpoint silencing in the presence of persisting DSBs [29]. Moreover, S. cerevisiae cells lacking a wt-CDC5 allele were unable to silence the activity of the Chk2 homologue Rad53 [97], indicating that, directly or indirectly, Polo-like kinase may regulate Chk2 function in that organism. The budding yeast 53BP1/Mdc1 homologue Rad9 has been shown to regulate checkpoint responses to DNA damage. Similar to 53BP1, Rad9 is activated by the ATM/ATR homologues Tel1/Mec1 and associates with the Chk2 homologue Rad53 [98]–[100]. A role for Rad9 as a target for feedback control to silence checkpoint functioning, however, has not yet been shown.
In addition to constituting a mechanism for silencing an activated DNA damage checkpoint, the Cdk1-Plk1-53BP1 feedback loop may be a more general means to prevent activation of the DNA damage checkpoint during mitosis. If a DNA damage–induced G2-like checkpoint were to become fully functional during mitosis, DNA lesions encountered during mitotic progression could result in inactivation of Cdk1/Cyclin B and result in forced mitotic exit. Such an event would cause the accumulation of 4N-DNA containing interphase cells, which were recently shown to have increased tumorigenic potential [101],[102]. Hence it can be expected that cellular mechanisms exist to prevent inappropriate Cdk1 inactivation during mitosis. Indeed, DNA damage during mitosis had previously been shown to be unable to delay mitotic progression or alter Cdk1 activity during mitosis [39],[40]. Our observation that Chk2 cannot be catalytically activated during mitosis by IR further strengthens this notion.
Immortalized proliferating cells are believed to have increased replication stress and elevated levels of associated DNA damage. The DNA damage checkpoint, therefore, was shown to form a barrier against malignant transformation [8],[9]. Feedback mechanisms, in which mitotic kinases can silence DNA damage checkpoints, may thus explain why Plk1 and Aurora A are frequently overexpressed in cancers and may form a rationale for including inhibitors of such mitotic kinases during cancer treatment [103]–[105].
A total of 244 in vivo mapped phosphorylation sites on 33 human DDR-related proteins were manually collected from Phospho.ELM [42], Phosphosite [43], and a phospho-proteomic study of Polo Box substrates [44]. The conservation level of these phosphorylation sites was measured by aligning predicted orthologues of these proteins in 11 species from the high-coverage vertebrate branch of Ensembl (release 46; [106]) with the human seed sequences in which the sites had been mapped by mass spectrometry and other means. Genomes used in the analysis included Homo sapiens (human), Bos taurus (cow), Canis familiaris (dog), Danio rerio (zebrafish), Gallus gallus (chicken), Macaca mulatta (rhesus monkey), Mus musculus (house mouse), Ornithorhynchus anatinus (platypus), Rattus norvegicus (Norway rat), Tetraodon nigroviridis (fresh water pufferfish), and Xenopus tropicalis (western clawed frog).
Ortholog mapping was performed as follows: initially the human DDR sequences and phosphorylation sites were mapped to their corresponding Ensembl gene entries by sequence comparison. Next, genes orthologous to the DDR genes were retrieved from Ensembl. Orthologous relationships between human and other vertebrate genes in Ensembl were inferred from phylogenetic trees constructed from multiple sequence alignment of CDS sequences [106]. A detailed description of Ensembl ortholog detection pipeline in release 46 is available at http://aug2007.archive.ensembl.org/info/data/compara/homology_method.html. Finally, each DDR protein sequence and all spliced variants of its orthologous genes across the 11 species were aligned using multiple-sequence alignment program MAFFT (v6.240, E-INS-i option with default parameters) [107].
Cross-species conservation of the human phosphorylation sites was then computed by evaluating the average number of amino acid substitutions within a −5 to +5 residue window of the modified residue (S, T, or Y) across the 11 vertebrate genomes from the sequence alignments using Perl scripting. S→T and T→S transitions of the central phosphoresidue were permitted, but S/T→Y transitions were not. The conservation level for each phosphorylation site is reported as the average across the 11 genomes, as a percentage of conserved residues within the 11-mer window, if the corresponding S/T is conserved.
Information about which of the 244 in vivo mapped phosphorylation sites were phosphorylated by the specific kinases ATM/ATR, Cdk1/2, Chk1/2, and Plk1 was collected from Phospho.ELM [42] and Phosphosite [43], along with whether phosphorylation at that site was known to create a binding site for the PBD of Plk1 [44]. In cases where multiple kinases are known to phosphorylate a single site, all of this information was retained and displayed. For sites where the upstream kinase was not experimentally known, we predicted the likely kinase responsible for phosphorylation at that site by computational analysis using the programs NetworKIN [45],[47] and NetPhorest [46].
Rabbit anti-53BP1 (304-A1) was from Novus Biologicals. Mouse anti-γ-H2AX (pS139, #05-636), rabbit anti-HistoneH3 pS10 (#06570), rabbit anti-Chk2 (#2662), rabbit anti-Chk2-pT68 (#2661), rabbit anti-53BP1-pS1778 (#2675), mouse anti-MPM2 (#05-368), and rabbit anti-Plk1 (#06-831) were purchased from Upstate. An additional rabbit anti-Chk2 antibody (#BL1432) was purchased from Bethyl Laboratories. Rabbit anti-Plk1 for immunoprecipitation was a kind gift from Dr. René Medema. Mouse anti-β-actin (A5441) was from Sigma. Mouse anti-Cyclin B1 (GNS1, sc-245), rabbit anti-GFP (sc-8334), and rabbit non-specific IgG (sc-2025) were from Santa Cruz Biotechnology. Mouse anti-GFP (clones 7.1 and 13.1) was from Roche. Rabbit anti-p-S380-53BP1 phospho-specific antibody was raised against peptide Pro-Phe-Iso-Val-Pro-Ser-pSer-Pro-Thr-Glu-Gln-Glu-Gly-Arg-Tyr and purified by Cell Signaling Technologies. Radiolabelled [32P]-γ-ATP (3,000 Ci/mmol) was purchased from Amersham/GE Healthcare. Plk1 inhibitor (BI 2536) was synthesized following the procedure described by Munzert et al. [108]. All other reagents and chemicals were from Sigma unless otherwise indicated.
The pEGFP-m53BP1 expressing murine GFP-Tagged 53BP1 was kindly provided by Dr. Yasuhisa Adachi. The Nhe1-Apa1 fragment of pEGFP-m53BP1 was cloned in the retroviral plasmid pLNCX2 (Clontech) containing a synthetic linker to generate pLNCX2-GFP-m53BP1. PCR-based mutagenesis was used to create pLNCX2-GFP-m53BP1-317A, m53BP1-330A, m53BP1-376A, m53BP1-922A, m53BP1-1103A, and m53BP1-1620A. All plasmid constructs were verified by automated sequencing. pLNCX2-GFP-m53BP1Δ196–439 was created by a nested PCR on two m53BP1 PCR fragments surrounding the deletion. The resulting 53BP1 fragment containing the deletion was used to replace wt-m53BP1 in pLNCX2-GFP-m53BP1. Full-length human Chk2 was cloned from pGEX6P2-Chk2 and subcloned into the Nhe1-EcoR1 sites of pIRES2-GFP (Clontech). Serine/Threonine to Alanine mutations at positions 164, 168, 205, and 210 were obtained by side directed mutagenesis and validated using automated sequencing. Full-length FLAG-tagged Chk2 was a kind gift from Dr. Domenico Delia. VSV-G pseudotyped retroviruses were prepared according to standard techniques. In brief, HEK293T packaging cells were transfected with the pLNCX-2 and the packaging plasmids pMDg/p and pMDg in a 4∶3∶1 ratio. Virus-containing supernatant was harvested at 24 and 48 h after transfection, filtered through a 0.45 µM syringe filter, and used to infect U2OS osteosarcoma target cells. A plasmid encoding the PBD of Plk1 (aa. 326–603) fused to GST was described previously [50].
U2OS osteosarcoma cells were maintained in Dulbecco's Modified Eagle medium, supplemented with 10% fetal calf serum, 100 units/ml penicillin, and 100 µg/ml streptomycin. To obtain mitotic cell populations, cells were incubated with paclitaxel (1 µg/ml) or nocodazole (250 ng/ml, Sigma). Where indicated, cells were harvested by mitotic shake-off. Where indicated, DNA damage was induced using a gamma-cell 40 irradiator equipped with a 137Cesium source for indicated doses. Alternatively, cells were incubated with doxorubicin (0.5 µM) for 1 h.
Human breast cancer cell line MCF7 or human osteosarcoma U2OS cells were retrovirally infected with control pRetrosuper or pRetrosuper-53BP1 (53BP1-targeting sequence #1, 5′-GATACTGCCTCATCACAGT-3′; 53BP1-targeting sequence #2 5′-GAACGAGGAGACGGTAATA-3′) for three consecutive 12 h periods [61]. Infected cells were selected with 2 µg/ml puromycin. pRS-53BP1-infected MCF7 cells were subsequently treated with 4 µM nutlin-3 to select for cells with a functional 53BP1 knockdown [61]. The statistical analysis of colony numbers, S-phase content, and phospho-HistoneH3 content in control-infected or pRS-53BP1-infected MCF-7 cells was done using the unpaired t test. Two-tailed p values were calculated using GraphPath software.
The Plk1 kinase domain (residues 38–346) was made as a His6-tagged construct in Escherichia coli (E. coli) Rosetta cells (Novagen) and purified by Ni-NTA chromatography followed by gel filtration on a Superose-12 column. Recombinant full-length GST-Chk2 and a GST-Chk2 FHA domain (amino acids 1–219) fusion were expressed and purified from E. coli. In brief, full-length Chk2 was cloned into pGEX-6P1 (GE Healthcare) and transformed into BL21 (DE3) cells. Cells were grown at 37°C to an OD600 of 0.6 and the culture temperature was reduced to 18°C for 30 min before a final concentration of 0.3 mM IPTG was added for overnight expression. Cells were pelleted and washed with MTPBS (16 mM Na2HPO4, 4 mM NaH2PO4, 150 mM NaCl, pH 7.3) and lysed by sonication in the same buffer with the addition of benzonase. The lysate was clarified by centrifugation, and the GST-Chk2 fusion protein was captured on glutathione 4B resin. After washing with 30 column volumes of phosphate-buffered saline (PBS), Chk2 was cleaved off the GST tag on the resin with 3C protease at 4°C overnight. The eluted full-length Chk2 was further purified by anion exchange on a Resource Q column (GE Healthcare) equilibrated with 20 mM Tris pH 8.0, 50 mM NaCl, 0.5 mM TCEP, and developed with 20 mM Tris pH8, 1 M NaCl, and 0.5 mM TCEP. Peak fractions containing full-length Chk2 were pooled and further purified with a Superdex S200 gel filtration column (GE Healthcare).
The GST-Chk2 FHA domain cloned into pGex-4T1 was transfected into BL21(DE3) cells, grown to an OD600 of 0.8, and induced with 1 mM IPTG at 37°C for 6 h. Cells were lysed in PBS containing 1 mM DTT and a mixture of protease inhibitors and disrupted by sonication. Benzonase (Novagen) was added at room temperature for 30 min and the lysate cleared via centrifugation. Roughly 500 µL of PBS-equilibrated GSH beads were added to the lysate and incubated at 4°C with rocking overnight. Non-bound material was aspirated off followed by 4×10 mL washes with PBS containing 0.2% NP-40 and 1 mM DTT, and the GST-Chk2 FHA domain eluted off the beads by incubation in 2.5 mL of elution buffer (20 mM HEPES pH 7.2, 40 mM glutathione, and 1 mM DTT; EB+G) at 4°C overnight. The purified GST-Chk2 FHA domain was dialyzed against elution buffer lacking 40 mM glutathione (EB) using a Slide-A-Lyzer (Pierce) dialysis cassette with a molecular weight cut-off of 6–8 kDa at 4°C with three buffer exchanges. Purity was assessed by SDS-PAGE and the protein concentration determined by absorbance at A280 using an extinction coefficient of 71,780 M−1 cm−1.
Streptavidin beads (Pierce, 75 pmol/µL gel) were incubated with a 5-fold molar excess (relative to binding capacity) of a biotinylated phosphothreonine-oriented peptide library (B-4pT5; biotin-Gly-AHA-Gly-AHA-Met-Ala-X-X-X-X-pThr-X-X-X-X-X-Ala-Tyr-Lys-Lys-Lys-NH2, where X indicates a equimolar degenerate mixture of all amino acids except Cys, and pThr denotes phosphothreonine) in 50 mM Tris pH 7.5, 150 mM NaCl, 0.5% NP-40, and 1 mM EDTA for 30 min at 4°C. Beads were washed five times with the same buffer to remove unbound peptides and then 20 µL of the bead-immobilized library was incubated with 10 µg of GST-Chk2 FHA domain prior to or following in vitro phosphorylation of the FHA domain by Plk1 kinase in the presence of [32P]-γATP. After a 60 min incubation, the beads were washed five times with 50 mM Tris pH 7.5, 150 mM NaCl, 0.5% NP-40, and 1 mM EDTA. Bead-bound protein was released by addition of SDS sample buffer with heating to 95°C and resolved by SDS-PAGE on 10% denaturing gels. Gels were analyzed by autoradiography using a phosphorscreen and a Typhoon variable mode imager (GE Healthcare, or transferred to PVDF membrane and immunoblotted using an HRP-conjugated anti-GST antibody to visualize binding of the GST-FHA domain).
U2OS cells were seeded on glass cover slips and treated as indicated. After treatment, cells were fixed in 3.8% formaldehyde in PBS for 15 min at room temperature. Subsequently, cells were permeabilized with 0.1% TritonX100 in PBS for 5 min. After extensive washing, cells were blocked and stained in PBS-0.05% Tween20 and mounted on slides. Images were acquired on a Zeiss Axioplan-2 inverted microscope, equipped with a Hamamatsu Orca-ER digital camera using OpenLab software.
After the indicated treatments, U2OS cells were lysed in lysis buffer (1% TritonX-100, 50 mM Tris-HCl (pH 7.5), 150 mM NaCl, 50 mM beta-glycerophosphate, 10 mM sodium pyrophosphate, 30 mM NaF, 1 mM benzamidine, 2 mM EGTA, 100 µM NaVO4, 1 mM dithiothreitol (DTT), 1 mM phenylmethylsulfonyl fluoride, 10 µg/ml aprotinin, 10 µg/ml leupeptin, 1 µg/ml pepstatin, and 1 µg/ml microcystin-LR) for 15 min at 4°C and cleared by high speed centrifugation. Protein concentrations were measured using the bicinchoninic acid assay (Pierce). 53BP1 was immunoprecipitated from 500 µg of clarified cell lysate using 3 µg of anti-53BP1 antibody and 50 µls of Protein-A-conjugated agarose beads (50% slurry) for 16 h. Immunoprecipitations were extensively washed and analyzed by SDS-Page and Western blotting. Alternatively, immunoprecipitations were subjected to in vitro phosphorylation by resuspension in kinase buffer (50 mM Tris-HCL pH7.5, 10 mM MgCl2, 1 mM EGTA, 2 mM DTT, 2 mM dithiothreitol, 0.01% BRIJ35, and 150 mM NaCl2), followed by addition of 25 µM unlabelled ATP, 10 µCi of [32P]-γ-ATP, and recombinant Cyclin A-Cdk2, Cyclin B-Cdk1, or Plk1 for 30 min. Kinase reactions were analyzed by SDS-page and autoradiography.
IP/kinase assays for Chk2 activity were performed as generally described [109] using lysates from either interphase cells or from mitotic cells generated by treating U2OS cells with 0.25 µg/ml nocodazole for 16 h followed by harvesting of the mitotitc non-adherent cells by gentle shaking. In brief, Protein A microtiter strips (Pierce) were coated overnight with 1.0 µg of anti-Chk2 antibody (Bethyl) or non-specific rabbit IgG per well and washed three times with blocking buffer (1% bovine serum albumin in 50 mM Tris-HCl (pH 7.5), 150 mM NaCl, 0.05% Triton X-100). Cell lysates (100 µgs) were placed in each antibody-coated well, incubated for 3 h, then washed twice with wash buffer (50 mM Tris-HCl (pH 7.5), 150 mM NaCl) and twice with kinase wash buffer (20 mM Tris-HCl (pH 7.5), 15 mM MgCl2, 5 mM beta-glycerophosphate, 1 mM EGTA, 0.2 mM Na3VO4, 0.2 mM DTT). Kinase reactions were performed in a total volume of 60 µl containing 20 mM Tris-HCl (pH 7.5), 15 mM MgCl2, 5 mM β-glycerophosphate, 1 mM EGTA, 0.2 mM Na3VO4, 0.2 mM DTT, 0.4 µM protein kinase A inhibitor, 4 µM protein kinase C inhibitor, 4 µM calmidazolium, 25 µM ATP, 10 µCi [32P]-γ-ATP, and 10 µM of Chk2tide substrate. Reactions were incubated for 60 min at 37°C, then terminated by addition of 60 µl of 20 mM EDTA. Forty µl of the terminated reaction mixture was transferred to a phosphocellulose filter plate (Millipore, Bedford, MA) containing 100 µl, 75 mM H3PO4, and mixed. The reaction contents were vacuum-filtered and washed five times with 75 mM H3PO4 and three times with 70% ethanol. Scintillation counting was performed using a Microbeta TRILUX luminescence counter.
In vitro phosphorylation of recombinant Chk2 or the Chk2 FHA domain by Plk1 was performed by incubating 3–10 µgs of the substrate proteins with Plk1 kinase domain in 50 mM Tris pH 7.5 containing 150 mM NaCl, 10 mM MgCl2, 100 µg/ml bovine serum albumen, 5 mM DTT, and 100–500 µM unlabelled ATP, in the presence or absence of 10–20 µCi [32P]-γ-ATP, for 60–120 min at 30°C. Kinase assays of recombinant Chk2 before or after Plk1 phosphorylation were performed in the above buffer containing 1 mM DTT, 20 µCi [32P]-γ-ATP, and 50 µM Chk2tide in a final reaction volume of 50 µl at 30°C for 60 min. Samples were quenched with an equal volume of 0.05% H3PO4, and 5 µl of the reaction spotted onto P81 paper, air dried, washed extensively with 0.05% H3PO4, and analyzed by scintillation counting.
Identification of Plk1 phosphorylation sites in the Chk2 FHA domain following in vitro phosphorylation was performed by separating the reaction products by SDS-PAGE. Gel slices containing Chk2 were excised, alkylated with iodoacetamide, and digested with trypsin. Peptides were resolved by nano-flow reversed phase liquid chromatography (Agilent 1100, Palo Alto, CA) and analyzed with a LTQ-Orbitrap equipped with a nanoelectrospray ionization source (Thermo, Bremen, Germany). Peptide and protein identification was analyzed using the Spectrum Mill MS Proteomics Workbench software (Agilent).
For the in vivo mobility shift analysis of Chk2, 293T cells were transfected with FLAG-tagged full-length hChk2. Twenty-four h after transfection, cells were treated with paclitaxel in combination with DMSO or in combination with Plk1 inhibitor for 8 h. Cell lysates were cleared by centrifugation and mixed with M2 FLAG resin for overnight immunoprecipitation. After washing, samples were analyzed by SDS-PAGE.
Cells were harvested with Trypsin/EDTA, washed with PBS, and subsequently fixed in ice-cold 70% ethanol for 4–16 h. After washing, cells were stained with anti-phospho-Histone H3 (1∶200) or anti-phospho-γ-H2AX (1∶100) in PBS-0.05% Tween20 and counterstained with Alexa647-conjugated secondary antibodies in PBS-0.05% Tween20. Cells were treated with Propidium Iodide/RNAse and analyzed on a Becton Dickinson FACScalibur using Cellquest software. A minimum of 10,000 events were counted.
|
10.1371/journal.ppat.1003839 | Production, Fate and Pathogenicity of Plasma Microparticles in Murine Cerebral Malaria | In patients with cerebral malaria (CM), higher levels of cell-specific microparticles (MP) correlate with the presence of neurological symptoms. MP are submicron plasma membrane-derived vesicles that express antigens of their cell of origin and phosphatidylserine (PS) on their surface, facilitating their role in coagulation, inflammation and cell adhesion. In this study, the in vivo production, fate and pathogenicity of cell-specific MP during Plasmodium berghei infection of mice were evaluated. Using annexin V, a PS ligand, and flow cytometry, analysis of platelet-free plasma from infected mice with cerebral involvement showed a peak of MP levels at the time of the neurological onset. Phenotypic analyses showed that MP from infected mice were predominantly of platelet, endothelial and erythrocytic origins. To determine the in vivo fate of MP, we adoptively transferred fluorescently labelled MP from mice with CM into healthy or infected recipient mice. MP were quickly cleared following intravenous injection, but microscopic examination revealed arrested MP lining the endothelium of brain vessels of infected, but not healthy, recipient mice. To determine the pathogenicity of MP, we transferred MP from activated endothelial cells into healthy recipient mice and this induced CM-like brain and lung pathology. This study supports a pathogenic role for MP in the aggravation of the neurological lesion and suggests a causal relationship between MP and the development of CM.
| Cerebral malaria (CM) is a potentially fatal neurological syndrome characterised by unrousable coma. Since the detection of high levels of plasma microparticles (MP) in patients with CM, it has been demonstrated that inhibition of MP production confers protection from murine CM. However, the precise mechanisms of action of these MP during CM have not been completely deciphered. In this study, we used experimental models of CM to measure the production and origins of MP over the course of infection. We found low baseline circulating MP in healthy mice and these were subsequently raised at the time of the neurological syndrome. Phenotypic analyses showed that circulating MP were predominantly from activated host cells that have previously been established to participate in CM pathogenesis. We show for the first time transferred MP impairing endothelial integrity and inducing CM-like pathology in the brain and lung of healthy animals. Our study dissects what tissues these MP localise to exert their effects, as little is known about their fate following the initial release. These data suggest a causal relationship between MP and the development of CM and also warrant further investigation into the representation of MP as a marker of CM risk.
| Cell activation by various agonists and apoptosis trigger the vesiculation of microparticles (MP) from all cell types [1], [2], [3]. During vesiculation, phospholipids are reorganised through the translocation of inward and outward membrane lipids, whereby phosphatidylserine (PS) is exposed on the outer leaflet of the membrane [4], [5]. The budding progeny are small (0.2–1 µm) plasma membrane-derived vesicles that express antigens of their cell of origin and PS on their surface, facilitating their role in coagulation, inflammation and cell adhesion [6], [7].
Once described as inert biological bystanders MP have now emerged as novel therapeutic targets in the treatment of diseases [8], [9], [10]. Under normal physiological conditions, baseline levels of circulating MP can be detected in the blood and are thought to be involved in maintaining cellular homeostasis. However, elevated levels of MP have been implicated in several diseases [11], [12], [13], [14], [15], [16], [17], [18], including cerebral malaria (CM), in patients as well as in experimental models [19], [20], [21], [22], [23], [24], [25], [26].
CM is a multisystem multi-organ dysfunction that develops as a syndrome following Plasmodium falciparum infection [27]. It is characterised by the presence of sustained impaired consciousness and those surviving may develop residual neurological sequelae [28]. Despite better campaigns targeted at the eradication of malaria, the global burden persists [29], [30]. The underlying pathogenesis that drives the manifestation of CM remains incompletely understood. What is known is that the pathogenesis is multifactorial, involving the dynamic interaction between cellular sequestration, a dysregulated inflammatory response, MP production and homeostasis disruption [24], [31], [32].
Little is understood about the role of MP in CM pathogenesis, although markedly high plasma levels of circulating platelet, erythrocytic, leucocytic and in particular endothelial cell-derived MP (PMP, EryMP, LMP, EMP respectively) have been detected in patients with CM [23], [26], [33]. In murine experimental CM (eCM), the overproduction of MP is also observed, and ablation of MP vesiculation via knock-down of the ATP-binding cassette transporter A1 (ABCA1) involved in the distribution of PS, confers protection against the neurological syndrome without interfering with the infection itself [21]. Pharmacological inhibition of MP production by pantethine also confers protection from CM [8]. The above findings indicate that MP may have an active role in the development of the CM lesion and are not merely an epiphenomenon, although the precise mechanisms of action of these MP during CM have not been completely deciphered [20].
Using murine experimental models of CM [34], [35] and non-cerebral malaria (NCM) [36], [37] we characterised the production of MP over the course of Plasmodium infection in CM-susceptible mice, and compared their cellular origins. We adoptively transferred MP, isolated from mouse blood obtained at the time of the neurological syndrome, into the circulation of recipient mice and followed their blood clearance. Our study dissects in which tissues these MP localise to possibly exert their effects, as little is known about their fate following their initial release. Since the endothelium is an active component of the CM lesion, and EMP have been found to be elevated in human CM (hCM) [23], [38], we transferred in vitro generated EMP and studied their induction of pathology and clearance kinetics in healthy and infected mice. This study shows MP localised at the neurovascular lesion in vivo and MP transfer elicited CM-like histopathology in the brain and lung of healthy recipients, supporting a role for MP in CM pathogenesis.
This study investigates the levels and cellular origin of MP produced during murine malaria and for the first time describes the clearance and fate of CM+ MP in vivo as well as the pathogenicity of EMP. Using CM-susceptible mice, we showed that P. berghei infection induces a rise in total circulating MP, although distinct differences exist in the cellular origin and production profiles of MP in mice developing CM versus NCM. We found that MP from CM+, but not healthy donors, adoptively transferred in vivo can be detected at the site of the lesion, sequestered amongst other cells within the cerebral vessels of CM+ recipient mice. We also showed that transferred EMP,induce CM-like pathology in the brain and lung of recipient mice, supporting a role for MP in the exacerbation of CM.
Ablating the increase of MP numbers either genetically [21] or pharmacologically [8] confers protection against murine CM. Elevated levels of plasma EryMP [33] and EMP [23], [38] in malaria patients correlate with severity and are particularly restricted to those with cerebral involvement. This also has been shown in studies on intracerebral haemorrhage, whereby higher MP levels correlate with coma and poor clinical outcome in patients [39][40]. Although suggestive of a role in the pathogenesis of CM, the existing data on MP in CM does not completely support a role for MP in the worsening of CM nor do they substantiate it solely as a predictive marker for CM.
Elevated levels of circulating Annexin V+ MP were detected in the plasma of CM+ mice (i.e. infected with PbA or PbK1/2) at the time of neurological development. This finding confirms what has already been observed upon CM onset in experimental and hCM [21], [38]. Our study extends from this and follows the production of MP over the course of infection, not just at CM onset. Interestingly, a biphasic production of MP was detected in the plasma of PbA infected mice, peaking during the early stages of infection and also at CM onset. These waves of MP coincide with the perpetuated cycle of endothelial activation, the production of cytokines and chemokines, the upregulation of adhesion molecules and the binding of vascular cells to microvessels [41]. Although the first wave of MP was absent in PbK1/2 infection, mice with high MP levels during the neurological phase did develop CM. MP overproduction was absent in mice that did not develop cerebral signs during the time of CM onset in PbA mice (i.e. PbK infected or 30% of the PbK1/2 infected). This suggests that MP may play a role in the development of the neurological syndrome.
At the time of CM onset, cell-specific MP numbers were higher in PbA infected mice than in PbK- and PbK1/2 infected animals or healthy controls. The sum of positive MP for single staining of cell markers gave a closer approximation of total circulating MP, as Annexin V staining indicates only PS-positive MP. MP can be PS negative or have low undetectable PS on the surface and remain unbound to Annexin V [20], [42]. We found that plasma PMP, EMP, EryMP and MMP were most elevated during the neurological phase in PbA infected mice. Previous studies have shown a dramatic increase in EMP, EryMP and PMP in patients presenting with CM [23], [33], [38]. In CM, the neurovascular lesion is comprised of sequestered vascular cells such as platelets, erythrocytes and leucocytes within the endothelial lining of microvessels [32], [43], [44]. It is not surprising that the sequestered cells also produce the most predominant MP populations detected. The development of CM is attributed to the cascade of events preceding the sequestration of cells and, consequently, the mechanical occlusion [31], [43], [44], [45]. The 70% of PbK1/2 infected mice that developed CM displayed comparable levels of total Annexin V+ MP levels to PbA infected mice, although their origins were predominately from platelets and erythrocytes. No significant differences were detected between the levels or proportions of cell-specific MP detected in PbK infected and healthy mice between days 6 and 14.
Our data show that during the acute stage of PbA-infection, a peak of PMP can be detected, and this is absent in the PbK- and PbK1/2-infections. This finding is consistent with the substantial loss of platelets as MP in the acute phase of the PbA-infection [46] and may correlate with the depth and duration of coma [38]. Thrombocytopenia is associated with poor prognosis in both human and experimental CM [46], [47] and higher platelet accumulation has been observed in cerebral microvessels in both human [48] and eCM [49], [50]. In vitro, platelets enhance the binding of infected erythrocytes (IE) to the cerebral endothelium [51] and their MP progeny are able to adhere preferentially to IE and also to the cerebral endothelium [19], [52]. PMP pathogenic potential is attributed to their mobility and their access via blood flow to other vascular cells [53]. Although this finding supports the platelet adhesion hypothesis, the absence of this first wave of PMP in PbK1/2 infection is interesting. Nothing is known about thrombocytopenia or PMP in PbK1/2 infection. The majority of mice develop CM+ despite the absence of the first wave of PMP; nevertheless, at CM onset, higher titres of PMP are detected.
EryMP were elevated in CM+ mice, both in PbA- and PbK1/2 infected mice, at CM onset. This finding is supported by human studies, whereby EryMP numbers are increased in patients with P. falciparum infections, even after antimalarial drug treatment [33] and also in patients with severe malarial anaemia [38]. In contrast, EryMP numbers are lower in patients with P. vivax and P. malariae malaria, similar to what is observed with the PbK infected mice in our study [33]. Interestingly, the PbK infected group, with no evidence of cerebral complications, had an overproduction of MP from erythrocytes and leucocytes during the later stage of infection. The evolving hyperparasitaemia in these mice leads to the destruction of IE and the gradual rupture and fragmentation of fragile erythrocytes increases the level of EryMP and cellular debris [33]. Higher levels of LMP could account in that case for the activation of the cells involved in the destruction of erythrocytes and their removal from the circulation.
Adoptively transferred CM+ MP are detectable in recipient mice but quickly subsided, indicating clearance from circulation. Some of these MP were found to be arrested in cerebral microvessels of CM+ recipients and also in the spleen, kidney and, to a lesser extent, the lung and liver. Recent studies in humans have shown that parasites induce the loss of endothelial protein C receptors in the cerebral microvessels, leaving them vulnerable to enhanced local thrombin generation and coagulopathy [54]. We know from previous murine studies that MP are procoagulant and proinflammatory [21], [43], [55], thus it is not surprising that MP are found lodged in the cerebral microvessels of infected recipients. No MP were detected in the heart of recipients. MP from both healthy and CM+ donors were also detected within the spleen of recipient hosts, suggesting that this organ could be a site of MP trapping and clearance from the circulation independent of the infection. It is possible that the spleen filters the PS+ MP in a similar way to PS+ cells in malaria [56] and Kupffer cells in the liver could remove EryMP as shown by Willenkens et al., [57], although further studies are required to elucidate this in our system.
Little is known on the mechanisms underlying the clearance of plasma MP from circulation in vivo. In our study, the disappearance of PKH67-labelled CM+ MP occurred within 5 minutes. In a rabbit model, no PMP were detected in the circulation at 10, 30 or 60 minutes post injection [58]. In contrast, transferred PMP, isolated from platelet concentrates from the peripheral blood of single donor patients, circulated for markedly longer with a half-life of 5.8 hours (Annexin V+) and 5.3 hours (CD61+) [59]. The authors attribute the elevated levels and the longer half-life of circulating MP to the infused platelets producing more MP in circulation [59]. Another possible explanation is that MP may clear the circulation and reach their target sites quicker in rabbit models, and also in mice, due to their faster heart rate [58]. The authors also suggest that the half-life is an overestimation, and the observations in the rabbits miss the initial rapid clearance of MP [58]. Furthermore, in vitro assays support the idea that MP can be detected in blood when there is no mechanism to remove them [58]. In pathological states, the continuous production of MP overrides the mechanisms of rapid clearance, hence elevated levels can be detected.
All cells are able to produce and release MP, although the emerging progeny of MP are heterogeneous and do not share the same properties. EMP represent the most abundant MP detected in pathologies that arise due to vascular injury or endothelial dysfunction although not all their roles are noxious [60], [61]. EMP numbers are elevated in patients with conditions in which the endothelium is injured and/or the endothelial barrier is compromised, including sickle cell disease, Alzheimer's disease, metabolic syndrome, hypertension, atherosclerosis and chronic obstructive pulmonary disease [62], [63], [64], [65], [66], [67]. EMP were first described in vitro [6] and high numbers were detected in CM patients presenting with coma [23]. Lower numbers of EMP and concentrations of TNF were detected in mice protected against CM during PbA infection [21]. Studies in vitro and in the murine model of CM indicated that EMP have similar procoagulant and pro-inflammatory potentials to, and express the same repertoire of antigens as, their corresponding mother EC [6], [21]. In CM, the endothelium is both a target and an effector in disease pathogenesis [24]. The direct role of EMP in inducing brain and lung pathology in murine models of CM has not been described. We addressed the hypothesis that EMP may be pathogenic in the CM lesion by transferring TNF-generated EMP into healthy and infected mice. Our findings demonstrate that transferred TNF-EMP can induce histopathological signs that are compatible with endothelial leakage leading to cerebral and pulmonary oedema and haemorrhage in healthy mice [68]. Control inert microsphere beads did not induce any change and remained in circulation for 60 minutes, compared to EMP that were cleared within the first few minutes post transfer, supporting the idea that the clearance of the MP is a physiological phenomenon that is mediated by receptors present at the surface of MP.
Exactly how EMP alter endothelial integrity in CM is unclear. Flow cytometry revealed that our EMP express endoglin, CD54 and CD106. Soluble endoglin (sCD105) overexpression is linked to typical systemic and vascular inflammation states such as pre-eclampsia and HELLP syndrome [69]. It is possible that endoglin-bound MP may have a role in inducing vascular injury. Endoglin is an RGD membrane protein acting as transforming growth factor-β accessory receptor and has been implicated in leucocyte recruitment and extravasation [70] and more recently in septic shock-induced disseminated intravascular coagulopathy [71]. In CM, CD54 and CD106 are established markers of EC injury and enable tethering and sequestration of cells to the endothelium [31], [43], [72]. TNF increased the expression of CD54 and CD106 on the EMP, consistent with other studies [25], suggesting a possible mechanism by which MP interact with the endothelium to induce injury.
Transferred EMP induced atelectasis in lungs from healthy recipient mice and increased alveolar cellularity, resembling the pathology seen in CM+ lung. EMP were shown to induce endothelial dysfunction, promote vasodilation, pulmonary oedema and acute lung injury in pathophysiological concentrations [73]. EMP sequester in lung tissue and elicit an immune response in vivo by increasing cytokine production leading to neutrophil recruitment [73], [74]. It is plausible that EMP initiate a cascade of events, beginning with the production of cytokines, that prime the endothelium thereby ultimately impairing vessel functions and resulting in tissue damage.
Taken together, our findings offer new evidence for a causal relationship between MP and the pathogenesis of CM. To our knowledge, this is the first time that MP have been localised at the neurovascular lesion in vivo and that their transfer elicited histopathology in the brain and lung of healthy recipients. We confirm that elevated levels of total MP are present at CM onset, predominantly from activated host cells that are known to participate in CM pathogenesis. Specifically, in CM, the early peak of total MP and PMP differentiated the kinetic production profile from NCM and could potentially be an indicator of prognosis. We showed that MP are rapidly cleared from circulation, and that some remain sequestered in organs, such as the brain and spleen. The EMP in this study carry VCAM-1 and ICAM-1 on their surface, which are upregulated following cytokine stimulation, potentially mediating their role in the worsening of CM.
During CM, activation of cells by parasite moieties, toxins, cytokines and/or cell death, delivers MP into the circulation. We propose that the interactions between MP, endothelium, circulating host vascular cells and their released circulating soluble factors influence the course of infection leading to the development of CM. Since the first human studies on MP in CM [23], [38], there has been growing interest in exploring the potential of MP as biomarkers for both diagnosis and follow-up and therapeutic targets since they represent both a consequence of, and contributor to, CM. The plasma membrane acts as the primary sensor to its external environment, thus, discriminating the cellular origin of MP may indicate what tissues or cells are undergoing activation or damage. Besides detection of malarial retinopathy, which is yet to be included as standard assessment for severe malaria, a clinical challenge still exists to distinguish CM from other encephalopathies [75]. Diagnosis relies on P. falciparum parasitaemia and impaired consciousness with the exclusion of other potential causes of severe disease, which in malaria-endemic areas is difficult to achieve due to the prevalence of asymptomatic parasitaemia and the lack of high-level diagnostic testing. Misdiagnosis is common and there is a need for reliable CM specific diagnostic and prognostic biomarkers [76]. This would support the promise of MP as clinical probes for CM and help provide targeted care of malaria patients at imminent risk of organ damage or cerebral complications as indicated by their detected MP [11].
In addition, molecules such as Pantethine that inhibit MP release have conferred protection in vivo and may be suitable candidates for an adjunct neuroprotective treatment of CM [8]. If the delay of CM onset in vivo is observed in human patients, this could potentially increase the therapeutic window available for treatment decreasing CM-associated mortality and neurological sequelae. In combination with anti-parasite chemotherapy, molecules that stabilize plasma membranes and reduce overproduction of deleterious MP and shedding may be protective or minimise the cerebral complications associated with CM.
Infections were performed as previously described [34][37][38]. All protocols adhered to the Australian Code of Practice for the Care and Use of Animals for Scientific Purposes. All protocols were approved by the Animal Ethics Committee of the University of Sydney (K20/7- 2006/3/4434 and K00/10-2010/3/5317).
Data were analysed using GraphPad Prism version 5.00 for Windows, GraphPad Software, San Diego California USA. Survival curves were analysed using the Log-rank (Mantel-Cox) Test and the Gehan-Breslow-Wilcoxon Test. To compare several groups, we used non-parametric analysis of variance (ANOVA, Kruskall-Wallis) with a Dunn's post-test. To compare mean total and cell-specific MP levels between two groups the Wilcoxon test was used; *p<0.05, **p<0.001, ***p<0.0001.
|
10.1371/journal.pntd.0001612 | In Vitro and In Vivo Efficacy of Ether Lipid Edelfosine against Leishmania spp. and SbV-Resistant Parasites | The leishmaniases are a complex of neglected tropical diseases caused by more than 20 Leishmania parasite species, for which available therapeutic arsenal is scarce and unsatisfactory. Pentavalent antimonials (SbV) are currently the first-line pharmacologic therapy for leishmaniasis worldwide, but resistance to these compounds is increasingly reported. Alkyl-lysophospoholipid analogs (ALPs) constitute a family of compounds with antileishmanial activity, and one of its members, miltefosine, has been approved as the first oral treatment for visceral and cutaneous leishmaniasis. However, its clinical use can be challenged by less impressive efficiency in patients infected with some Leishmania species, including L. braziliensis and L. mexicana, and by proneness to develop drug resistance in vitro.
We found that ALPs ranked edelfosine>perifosine>miltefosine>erucylphosphocholine for their antileishmanial activity and capacity to promote apoptosis-like parasitic cell death in promastigote and amastigote forms of distinct Leishmania spp., as assessed by proliferation and flow cytometry assays. Effective antileishmanial ALP concentrations were dependent on both the parasite species and their development stage. Edelfosine accumulated in and killed intracellular Leishmania parasites within macrophages. In vivo antileishmanial activity was demonstrated following oral treatment with edelfosine of mice and hamsters infected with L. major, L. panamensis or L. braziliensis, without any significant side-effect. Edelfosine also killed SbV-resistant Leishmania parasites in in vitro and in vivo assays, and required longer incubation times than miltefosine to generate drug resistance.
Our data reveal that edelfosine is the most potent ALP in killing different Leishmania spp., and it is less prone to lead to drug resistance development than miltefosine. Edelfosine is effective in killing Leishmania in culture and within macrophages, as well as in animal models infected with different Leishmania spp. and SbV-resistant parasites. Our results indicate that edelfosine is a promising orally administered antileishmanial drug for clinical evaluation.
| Leishmaniasis represents a major international health problem, has a high morbidity and mortality rate, and is classified as an emerging and uncontrolled disease by the World Health Organization. The migration of population from endemic to nonendemic areas, and tourist activities in endemic regions are spreading the disease to new areas. Unfortunately, treatment of leishmaniasis is far from satisfactory, with only a few drugs available that show significant side-effects. Here, we show in vitro and in vivo evidence for the antileishmanial activity of the ether phospholipid edelfosine, being effective against a wide number of Leishmania spp. causing cutaneous, mucocutaneous and visceral leishmaniasis. Our experimental mouse and hamster models demonstrated not only a significant antileishmanial activity of edelfosine oral administration against different wild-type Leishmania spp., but also against parasites resistant to pentavalent antimonials, which constitute the first line of treatment worldwide. In addition, edelfosine exerted a higher antileishmanial activity and a lower proneness to generate drug resistance than miltefosine, the first drug against leishmaniasis that can be administered orally. These data, together with our previous findings, showing an anti-inflammatory action and a very low toxicity profile, suggest that edelfosine is a promising orally administered drug for leishmaniasis, thus warranting clinical evaluation.
| The impact of the leishmaniases on human health has been grossly underestimated for many years, and this complex of diseases has been classified by the World Health Organization (WHO) as one of the most neglected tropical diseases [1]. During the last decade, endemic areas have been spreading and a sharp increase in the number of leishmaniasis cases has been recorded. The WHO classifies leishmaniasis as a category 1 disease (“emerging and uncontrolled”), and there is an urgent need to develop new therapeutic drugs and approaches. Currently, about 350 million people in 98 countries around the world are at risk, and an estimated 12 million people are infected [1]. Despite progress in the diagnosis and treatment, leishmaniasis remains a major public health problem, particularly in tropical and sub-tropical developing countries. Published figures indicate an estimated incidence of two million new cases per year, with 1.5 million cases of self-healing, but disfiguring, cutaneous leishmaniasis, and 500,000 cases of life-threatening visceral leishmaniasis [1], [2]. Approximately 60,000 people die from visceral leishmaniasis each year, a rate surpassed among parasitic diseases only by malaria; and a loss of about 2.4 million disability-adjusted life years (DALYs) throughout the world has been calculated as the total disease burden of leishmaniasis [1]–[3]. Furthermore, a number of reports have emphasized the increasing importance of visceral leishmaniasis as an opportunistic infection among HIV-positive patients in areas where both infections are endemic [4].
The chemotherapy currently available for the leishmaniases is far from satisfactory and presents several problems, including toxicity, many adverse side-effects, high costs and development of drug resistance [2], [5]. Two pentavalent antimonial (SbV) compounds, sodium stibogluconate (Pentostam) and meglumine antimoniate (Glucantime), were first introduced in the 1940's and have since been used as first-line chemotherapeutic agents against all forms of leishmaniasis through parenteral administration. Although SbV, administered by intramuscular or intravenous route, remains the first-line drug for the treatment of leishmaniasis worldwide, its efficacy is becoming increasingly lower [6], and highly depends on Leishmania species and distinct endemic regional variations, even within the same country. Resistance is now common in India, and rates of resistance have been shown to be higher than 60% in parts of the state of Bihar, in north-east India [7], [8]. In addition, the incidence of adverse effects, including myalgia, arthralgias, pancreatitis, nephrotoxicity, hepatotoxicity, and cardiotoxicity [1], [2], [9], makes the search for new alternative medicines to SbV an urgent issue, and a number of drugs are now in clinical trials [10]. Intravenous infusion of liposomal amphotericin B (AmBisome) is at present the most effective anti-Leishmania drug [2], [11], but its relatively high cost makes it unaffordable in several poor areas of the world where the disease is more prevalent [2]. In addition, the requirement for long periods of parenteral administration, frequently requiring hospitalization, has also limited the clinical use of amphotericin B.
Miltefosine (Impavido) is a new oral agent that has shown high cure rates in visceral leishmaniasis in India (L. donovani; 94% cure) [12], and in cutaneous leishmaniasis in Colombia (L. panamensis; >90% cure) [13]. However, a recent therapeutical trial has revealed a limited potential of miltefosine for the treatment of American cutaneous leishmaniasis, with an unsatisfactory cure rate of 69.8% in Colombia [14]. Furthermore, this percentage fell to 49% when miltefosine was administered to patients with lesions caused by L. braziliensis, which comprise more than 60% of cutaneous leishmaniasis in Colombia [14]. Additional recent clinical trials in Brazil showed a cure rate of miltefosine for the treatment of cutaneous leishmaniasis caused by L. braziliensis of 75% [15], and for the treatment of cutaneous leishmaniasis caused by L. guyanensis of 71% [16]. Miltefosine treatment also led to approximately 70% cure rate for mucosal leishmaniasis due to L. braziliensis in Bolivia [17], [18]. Moreover, the miltefosine cure rate was approximately 53% for cutaneous leishmaniasis (33% for L. braziliensis infection, and 60% for L. mexicana infection) in Guatemala [13], [19], [20], and a cure rate of 63% was reported for L. tropica in Afghanistan [20]. These figures contrast with cure rates of more than 82% in the treatment of visceral leishmaniasis (kala-azar) in India [21], [22] and Bangladesh [23]. These data point out the great variability in the outcome depending on the geographical area for reasons that are not well understood. In addition, miltefosine commonly induces gastrointestinal side-effects, such as anorexia, nausea, vomiting and diarrhea, that sometimes lead to drop out from treatment [1], [2], [22]. Miltefosine is potentially teratogenic and should not be administered to pregnant women [1], [2], for whom adequate contraception should be guaranteed during treatment and for up to 3 months afterwards [1], given the teratogenic potential of miltefosine in animal models [24]. An additional concern is the rapid in vitro generation of resistance to miltefosine [25]–[27] that could limit its clinical use. Thus, these studies reinforce the need to search for new therapeutic alternatives in the treatment of leishmaniasis.
Edelfosine (1-O-octadecyl-2-O-methyl-rac-glycero-3-phosphocholine, ET-18-OCH3) is a promising antitumor ether lipid drug [28]–[30], which is not mutagenic and acts by activating apoptosis through its interaction with cell membranes [31]–[34]. In addition to its antitumor activity, edelfosine has been shown to exert in vitro antiparasitic activity against different species of Leishmania parasites [35]–[37]. Edelfosine has been considered the prototype molecule of a rather heterogeneous family of synthetic compounds collectively known as alkyl-lysophospholipid analogs (ALPs), that comprise the above clinically relevant miltefosine as well as perifosine, which also shows anti-Leishmania activity [38], [39]. Although the mechanism of action of miltefosine against Leishmania parasites remains to be fully elucidated, there are some reports showing that the ability of this compound to promote an apoptosis-like cell death is critical for its leishmanicidal activity [40], [41]. Because edelfosine has been shown to have a higher proapototic activity than both miltefosine and perifosine in human cancer cells [29], [30], [33], we have carried out here a comprehensive in vitro and in vivo study, investigating the putative anti-Leishmania traits of edelfosine, as compared to other ALPs, using different Leishmania species as well as mouse and hamster experimental models.
Animal procedures in this study complied with the Spanish (Real Decreto RD1201/05) and the European Union (European Directive 2010/63/EU) guidelines on animal experimentation for the protection and humane use of laboratory animals, and were conducted at the accredited Animal Experimentation Facility (Servicio de Experimentación Animal) of the University of Salamanca (Register number: PAE/SA/001). Procedures were approved by the Ethics Committee of the University of Salamanca (protocol approval number 48531).
Edelfosine (1-O-octadecyl-2-O-methyl-rac-glycero-3-phosphocholine) was from INKEYSA (Barcelona, Spain) and Apointech (Salamanca, Spain). Miltefosine (hexadecylphosphocholine) was from Calbiochem (Cambridge, MA). Perifosine (octadecyl-(1,1-dimethyl-piperidinio-4-yl)-phosphate) and erucylphosphocholine ((13Z)-docos-13-en-1-yl 2-(trimethylammonio)ethyl phosphate) were from Zentaris (Frankfurt, Germany). Stock sterile solutions of the distinct ALPs (2 mM) were prepared in RPMI-1640 culture medium (Invitrogen, Carlsbad, CA), supplemented with 10% (v/v) heat-inactivated fetal bovine serum (FBS), 2 mM L-glutamine, 100 units/ml penicillin, and 100 µg/ml streptomycin (GIBCO-BRL, Gaithersburg, MD) as previously described [28].
The following Leishmania strains were used in this study: L. amazonensis (MHOM/Br/73/LV78), L. braziliensis (MHOM/CO/88/UA301), L. donovani (MHOM/IN/80/DD8), L. infantum (MCAN/ES/96/BCN150), L. major LV39 (MRHO/SU/59/P), L. mexicana (MHOM/MX/95/NAN1), and L. panamensis (MHOM/CO/87/UA140).
Leishmania promastigotes were grown in RPMI-1640 culture medium, supplemented with 10% FBS, 2 mM glutamine, 100 units/ml penicillin, and 100 µg/ml streptomycin at 26°C. Promastigotes were treated with the indicated compounds during their logarithmic growth phase (1.5×106 parasites/ml) at 26°C. Late stationary promastigotes were obtained after incubation of the parasites for 5–6 days with starting inocula of 1×106 parasites/ml. Leishmania axenic amastigotes were obtained at pH 5.0 in Schneider's culture medium following a stepped temperature increase to 30, 31 and 32°C, except for L. infantum amastigotes, which were exposed to 34, 36 and 37°C, as previously described [42].
The antileishmanial activity in promastigotes and axenic amastigotes was determined by using the XTT (sodium 3,3′-[1-(phenylaminocarbonyl)-3,4-tetrazolium]-bis (4-methoxy-6-nitro) benzene sulfonic acid hydrate) cell proliferation kit (Roche Molecular Biochemicals, Mannheim, Germany) as previously described [42], [43]. Cells were resuspended in FBS-containing RPMI-1640 culture medium (1.5×106 cells/ml for promastigotes, and 2×106 cells/ml for axenic amastigotes), and plated (100 µl/well) in 96-well flat-bottomed microtiter plates at 26°C, in the absence and in the presence of different concentrations of the indicated ALPs. After 72-h incubation at 26°C, IC50 (half-maximal inhibitory concentration) values, defined as the drug concentration causing 50% inhibition in cell proliferation with respect to untreated controls, were determined for each compound. Measurements were done in triplicate, and each experiment was repeated four times.
One and a half million Leishmania spp. promastigotes or axenic amastigotes were treated in the absence and in the presence of the indicated concentrations of ALPs for different incubation times. Then, parasites were pelleted by centrifugation (1000× g) for 5 min, and analyzed for apoptosis-like DNA breakdown by flow cytometry following a protocol previously described [44]. Quantitation of apoptotic-like cells was monitored as the percentage of cells in the sub-G0/G1 region (hypodiploidy) in cell cycle analysis [44], [45], using a fluorescence-activated cell sorting (FACS) Calibur flow cytometer (Becton Dickinson, San Jose, CA) equipped with a 488 nm argon laser. WinMDI 2.8 software was used for data analysis.
The mouse macrophage-like cell line J774, grown in RPMI-1640 culture medium, supplemented with 10% FBS, 2 mM L-glutamine, 100 U/mL penicillin, and 100 µg/ml streptomycin, at 37°C in humidified 95% air and 5% CO2, was infected overnight at the exponential growth phase (3×105 cells/ml) with stationary-phase L. panamensis promastigotes, at a macropage/promastigote ratio of 1/10 in complete RPMI-1640 culture medium. Non-internalized promastigotes were removed by 2–3 successive washes with PBS. Then, uninfected and L. panamensis-infected J774 macrophages were incubated for 1 h with 10 µM of the fluorescent edelfosine analog all-(E)-1-O-(15′-phenylpentadeca-8′,10′,12′,14′-tetraenyl)-2-O-methyl-rac-glycero-3-phosphocholine (PTE-ET) [34], [46], [47] (kindly provided by F. Amat-Guerri and A.U. Acuña, Consejo Superior de Investigaciones Científicas, Madrid, Spain) in complete RPMI-1640 culture medium. In addition, J774 cells were also incubated first with 10 µM PTE-ET for 1 h, then washed with PBS and infected with L. panamensis in the darkness for 6 h. Samples were fixed with 1% formaldehyde, and analyzed with a Zeiss Axioplan 2 fluorescence microscope (Carl Zeiss GmbH, Oberkochen, Germany) (40× magnification).
J774 cells were infected with L. panamensis promastigotes as above. The number of intracellular viable parasites was assessed by incubating infected cells with RPMI-1640 medium containing 0.008% SDS to gently disrupt macrophage plasma membrane, followed by addition of RPMI-1640 culture medium containing 20% FBS to stop further lysis. Samples were then sequentially diluted in 96-well plates containing biphasic Novy-MacNeal-Nicolle (NNN) medium. Plates were incubated at 26°C for 20 days, and examined weekly under an inverted Nikon TS-100 microscope (Nikon, Kanagawa, Japan) to evaluate the presence of viable motile promastigotes. The reciprocal of the highest dilution found positive for parasite growth was considered to be the concentration of parasites.
Macrophage-like J774 cells were plated in complete RPMI-1640 culture medium at a concentration of 1×106 cells/well in 24-well culture plates (Costar, Cambridge, MA), and let them adhere for 2 h at 37°C in 5% CO2. Non-adhering cells were removed by gentle washing with complete RPMI-1640 culture medium. Adherent J774 cells were incubated in the absence (negative control), or in the presence of 10 µg/ml lipopolysaccharide (Sigma, St. Louis, MO) (LPS; positive control) or of different concentrations of edelfosine. After 18-h incubation at 37°C in 5% CO2, supernatants were collected, centrifuged at 500× g for 10 min, and stored at −80°C until analysis. NO release was indirectly measured using a colorimetric assay based on the Griess reaction. Triplicate 100-µl aliquots of cell culture supernatants were incubated with 50 µl of freshly prepared Griess reagent (1% sulfanilamide, 0.1% naphthylethylene diamide dihydrochloride, and 2.5% orthophosphoric acid) for 15 min at room temperature, and then absorbance of the azo-chromophore was measured at 550 nm. Nitrite concentration was determined by using sodium nitrite as a standard. All samples were assayed against a blank comprising complete RPMI-1640 culture medium incubated for 18 h on the same plates as the samples, but in the absence of cells. All reagents were purchased from Sigma. Results were expressed in nanomoles of nitrite per 106 macrophages.
Six-week-old female BALB/c mice (18–20 g) and four-week-old male Syrian golden hamsters (Mesocricetus auratus) (about 120 g) (Charles River Laboratories, Lyon, France), kept in a pathogen-free facility and handled according to institutional guidelines, complying with the Spanish legislation under a 12/12-h light/dark cycle at a temperature of 22°C, received a standard diet and water ad libitum. Mice were inoculated s.c. into their left hind footpad (in a total volume of 50 µl PBS) with 2×106 infective stationary-phase promastigotes, whereas hamsters, previously anesthetized with inhaled Forane, were inoculated intradermally in the nose with 1×106 stationary-phase promastigotes in a volume of 50 µl PBS. When inflamation was evident (about 1 week in mice, and 6 weeks in hamsters, after inoculation), animals were randomly assigned into cohorts of 7 animals each, receiving a daily oral administration (through a feeding needle) of edelfosine (15 mg/kg for mice, and 26 mg/kg for hamsters, in water), or an equal volume of vehicle (water). In mice, the footpad thickness was measured with calipers every week, and compared with the uninfected right hind footpad to obtain the net increase in footpad swelling. In hamsters, nose swelling was measured with calipers every week, and compared with the nose size before inoculation and treatment. Evolution index of the lesion was calculated as size of the lesion during treatment (mm)/size of the lesion before treatment. Animal body weight and any sign of morbidity were monitored. Drug treatment lasted for 28 days, and animals were killed following institutional guidelines, 24 h after the last drug administration.
After the killing of the animals, the parasite burden in the infected tissues was determined by limiting dilution assays as previously described [48]. Biopsies were washed 3 times with PBS supplemented with 100 units/ml penicillin and 100 µg/ml streptomycin (GIBCO-BRL), and then incubated overnight (12 h) at 4°C with PBS containing 100 units/ml of penicillin and 100 µg/ml streptomycin. Following overnight incubation, biopsies were washed 2–3 times with PBS supplemented with the above antibiotics, and then a weighed piece of the infected area was homogenized in 1 ml PBS containing antibiotics using a sterile glass Potter-Elvejhem type tissue grinder. Homogenate was diluted at a final concentration of 0.1 mg/ml in Schneider's culture medium, containing 100 units/ml penicillin and 100 µg/ml streptomycin; and then serial dilutions were made in triplicate in 96-well plates containing biphasic Novy-MacNeal-Nicolle (NNN) medium. Plates were incubated at 26°C for 20 days, and examined weekly under an inverted Nikon TS-100 microscope to evaluate the presence of viable promastigotes. The reciprocal of the highest dilution found positive for parasite growth was considered to be the concentration of parasites per mg of tissue. Total parasite load was calculated using the total weight of the respective infected organ.
Parasites cultured in Schneider's culture medium supplemented with 10% FBS, 100 units/ml penicillin, and 100 µg/ml streptomycin at 26°C for 5 days, were washed twice with PBS, and centrifuged at 1000× g for 10 min at room temperature. Parasites were then resuspended at 2×106 promastigotes/ml in Schneider's culture medium, and incubated at 26°C for 5 days with 4 mg/ml Glucantime (Aventis Pharma, Sao Paulo, Brazil), which corresponded to its IC50 value, previously assessed by the XTT technique. Drug-containing culture medium was changed every 4–6 days, depending on parasite growth, and parasites were washed with PBS, analyzed by XTT assay, and resuspended again at 2×106 parasites/ml. This procedure was repeated until parasite viability in the presence of the drug was over 80%. Then, after achieving this viability rate, this process was repeated three times, with increasing concentrations of SbV, up to reaching a final concentration of 37 mg/ml. The volume of drug solution used in each passage was controlled not to exceed 10% of the total volume of culture medium.
The level of SbV resistance was further assessed by infection of golden hamsters with the above in vitro-generated SbV-resistant (SbV-R) L. panamensis parasites, growing in the presence of 37 mg/ml SbV, as well as with wild-type susceptible L. panamensis, followed by treatment with Glucantime. Hamsters were divided into two groups, eight animals infected with the resistant strain and eight animals infected with the susceptible strain. Each group was inoculated intradermally on the nose with 1×106 stationary-phase promastigotes in a volume of 50 µl PBS. These animals were previously anesthetized with ketamine (50 mg/ml) and xylazine (5 mg/kg) intraperitoneally. About six weeks after infection, lesions were evident in both animal groups, and animals were treated daily with 40 mg/kg Glucantime, intramuscularly using a 27-gauge needle, for ten days. Evolution of the lesions and drug efficacy were monitored as above.
ALP-resistant Leishmania strains were generated as indicated above for SbV-resistant parasites. Drugs were initially incubated at their corresponding IC50 values, and then drug concentration was gradually increased. Parasites were considered resistant when they could grow at a drug concentration of 30 µM.
Data are shown as mean ± SD. Between-group statistical differences were assessed using the Mann-Whitney or the Student's t test. A P-value of <0.05 was considered statistically significant.
We analyzed the antileishmanial potential of the four most clinically relevant ALPs, namely edelfosine, miltefosine, perifosine and erucylphosphocholine (Figure 1). By using the XTT assay, we found that edelfosine and perifosine were the most active ALPs inhibiting proliferation of distinct Leishmania spp. promastigotes with IC50 values in the range of low micromolar concentration (∼2–9 µM) in most cases (L. donovani, L. panamensis, L. mexicana, L. major, L. amazonensis) (Table 1). L. braziliensis and L. infantum promastigotes were more resistant to the action of edelfosine, perifosine and miltefosine than the other Leishmania species tested (Table 1). Erucylphosphocholine was the least efficient ALP in inhibiting parasite proliferation regarding most Leishmania spp. promastigotes, but interestingly it showed the highest antiparasitic activity against L. infantum promastigotes (Table 1). In general, the antileishmanial activity of the distinct ALPs ranked edelfosine≥perifosine>miltefosine>erucylphosphocholine against Leishmania spp. promastigotes.
Next, we analyzed the antileishmanial activity of the distinct ALPs against distinct axenic Leishmania amastigotes. Following an axenic amastigote drug screening, we found that edelfosine and perifosine behaved as the most potent ALPs in the inhibition of proliferation of distinct Leishmania spp. amastigotes (Table 1). A wider range of IC50 values was detected for amastigote than for promastigote forms of Leishmania (Table 1). The IC50 values for the anti-Leishmania amastigote activity of edelfosine and perifosine ranged between ∼3–12 µM and ∼2–15 µM, respectively. Miltefosine showed a higher degree of variability (IC50, ∼4–39 µM), with L. panamensis amastigotes being rather resistant (IC50, 39.3 µM) (Table 1). Erucylphosphocholine showed the highest IC50 values (∼28–66 µM) for the inhibition of cell growth in all the Leishmania spp. amastigotes analyzed (Table 1). Surprisingly, L. infantum amastigotes were very sensitive to the action of perifosine, edelfosine and miltefosine, whereas their cognate promastigotes forms were rather resistant (Table 1), with double digit IC50 figures for promastigotes and low one-digit IC50 values for amastigotes. Interestingly, L. braziliensis amastigotes were far more sensitive to edelfosine and miltefosine than their promastigote counterparts (Table 1), whereas perifosine and erucylphosphocholine showed similar IC50 values for both L. braziliensis promastigote and amastigote forms with IC50 figures over 14 µM (Table 1). In general, the antileishmanial activity of the distinct ALPs ranked edelfosine≥perifosine>miltefosine>erucylphosphocholine against Leishmania spp. amastigotes. These results indicate that sensitivity of Leishmania parasites to ALPs is highly dependent on each species as well as on their stage form, namely promastigote or amastigote. Interestingly, because we have recently found that the level of edelfosine in plasma, after daily oral administration of 30 mg/kg, was about 10.3–25.2 µM in both BALB/c and immunodeficient mice [29], [30], [49], a dose that was effective in inhibiting cancer cell growth in vivo [29], [30], [50], our results indicate that edelfosine was active against all Leishmania spp. tested at pharmacologically relevant concentrations (Table 1).
The above results showed that ALPs were able to inhibit Leishmania spp. proliferation at distinct rates. We next analyzed whether these agents, used at the pharmacologically relevant concentration of 10 µM, were able to induce an apoptotic-like cell death in Leishmania spp. promastigotes by determining DNA fragmentation by flow cytometry. Parasites displaying a sub-G0/G1 hypodiploid DNA content represent cells that undergo DNA breakdown and an apoptotic-like cell death [51]. We found that edelfosine was the most active ALP in promoting a potent apoptotic-like response in all Leishmania spp. tested (Figure 2A). The well nigh absence of apoptotic response in L. infantum promastigotes (Figure 2A) was expected, as ALPs were used at 10 µM, below the IC50 value for the inhibition of L. infantum promastigote proliferation measured by XTT assays (Table 1). Interestingly, edelfosine showed a much higher proapoptotic-like activity against L. donovani and L. mexicana promastigotes than miltefosine and perifosine (Figure 2A), despite the similar IC50 values (∼2–3 µM) of the three ALPs, assessed by XTT assays (Table 1). These results suggest that the induction of cell death by edelfosine might differ somewhat from the way by which miltefosine and perifosine promote parasite killing. The ability of the distinct ALPs to induce apoptosis-like cell death in Leishmania spp. promastigotes ranked edelfosine>perifosine≅miltefosine>erucylphosphocholine. Results shown in Figure 2A also show that the ability of edelfosine to promote an apoptosis-like cell death is highly dependent on the Leishmania sub-genus. In this regard, edelfosine inhibited proliferation of L. amazonensis (sug-genus Leishmania) and L. braziliensis (sug-genus Viannia) promastigotes with XTT IC50 values of 6.4 and 18.3 µM, respectively (Table 1), but the percentage of parasites with a sub-G0/G1 hypodiploid DNA content was higher in L. braziliensis than in L. amazonensis promastigotes (Figure 2A).
L. infantum promastigotes behaved somewhat different from other Leishmania species, with regard to their sensitivity to undergo apoptosis-like cell death by ALPs, requiring high ALP concentrations. A dose-response analysis of the apoptotic-like response of L. infantum promastigotes to the four ALPs tested was in agreement with the above XTT IC50 values of the corresponding drugs (cf. Figure 2B and Table 1), with erucylphosphocholine as the most efficient ALP at 30 µM (Figure 2B). However, at higher concentrations, edelfosine became as efficient as erucylphosphocholine in prompting an apoptotosis-like cell death in L. infantum promastigotes (Figure 2B).
A comparative dose-response analysis showed that edelfosine was more potent than miltefosine in inducing apoptosis-like cell death in L. panamensis promastigotes (Figure 2, C and D), edelfosine being highly effective even at 5 µM. These results agree with our above data on XTT IC50 figures (Table 1). The cell cycle profiles from propidium iodide-stained L. panamensis promastigotes showed a high percentage of parasites with apoptosis-like DNA breakdown following edelfosine treatment at either 5 or 10 µM (Figure 2, C and D), whereas miltefosine induced only a significant DNA breakdown response at 10 µM (Figure 2, C and D). Interestingly, edelfosine also induced apoptosis-like cell death in L. panamensis axenic amastigotes (25.8±4.6 and 55.4±2.8% sub-G0/G1 cells (n = 3) after 24 h incubation with 10 and 20 µM edelfosine, respectively).
Because Leishmania parasites use macrophages as their main host cell in mammalian infection, we next analyzed the localization of edelfosine in Leishmania-infected macrophages. To this aim, we used the fluorescent edelfosine analog all-(E)-1-O-(15′-phenylpentadeca-8′,10′,12′,14′-tetraenyl)-2-O-methyl-rac-glycero-3-phosphocholine (PTE-ET), which has been previously used as a bona fide compound to analyze the subcellular localization of edelfosine in cancer cells [30], [34], [46], [52], [53], and it fully mimics the antitumor [30], [34], [46], [52], [53] and antileishmanial [54] (data not shown) actions of the parent drug edelfosine. The mouse macrophage-like cell line J774 was rather resistant to undergo apoptosis following treatment with edelfosine (IC50 = 40.7±7.1 µM, assessed by XTT assays), and therefore it was used as a host cell line for Leishmania infection. Edelfosine (10 µM) was unable to induce apoptosis in J774 cells following 24 h incubation (<2% apoptosis), and caused less than 15% apoptosis after 48 h incubation. This is in stark contrast to the high sensitivity of other monocyte-like cell lines to edelfosine, such as human U937 cells [28], [55], [56], which undergo rapid apoptosis and can therefore not be used as host cells to analyze the effect of ALPs on intracellular parasites residing in macrophages. Incubation of J774 macrophages with PTE-ET showed that the fluorescent edelfosine analog was taken up into the cell (Figure 3A). The blue fluorescence of PTE-ET was mainly located around the nucleus (Figure 3A, left panel) that could be related to a predominant accumulation of this ether lipid in the endoplasmic reticulum of J774 cells, as previously reported for solid tumor cells [50], [52]. When macrophages were infected with L. panamensis parasites, an intense blue fluorescence was detected in the intracellular parasites (Figure 3A, middle panel), indicating that a major location of the PTE-ET fluorescent compound turned out to be in the intracellular parasites inside the macrophage. The PTE-ET location in the parasites residing in the macrophage was clearly detected, irrespective of whether PTE-ET was incubated with macrophages previously infected with parasites (Figure 3A, middle panel), or with intact macrophages and then subsequently incubated with parasites (Figure 3A, right panel). Macrophages containing a low number of Leishmania amastigotes are shown in Figure 3 in order to facilitate visualization of the fluorescent drug location in the parasites (Figure 3A). Similar data were obtained with primary mouse bone marrow-derived macrophages, which were resistant to 10 µM edelfosine, following infection with L. major (data not shown). These data suggest that edelfosine accumulates in intracellular Leishmania parasites inside macrophages, in a similar way as PTE-ET, to exert its anti-parasite action regardless drug treatment is before or after infection.
We also found that edelfosine efficiently killed J774 macrophage-residing L. panamensis by examining the parasitic burden of macrophages through limiting dilution assays (Figure 3B). The cytotoxic action of edelfosine against intracellular L. panamensis amastigotes was further confirmed by a dramatic decrease in the number of intracellular parasites, using J774 macrophages infected with green fluorescent L. panamensis, previously transfected with p.6.5-egfp to express green fluorescent protein (GFP) [57] (data not shown).
Some anti-parasite drugs are suggested to promote their action through the generation of nitric oxide (NO) [58], as this molecule exerts an important anti-parasitic effect [59], [60]. Miltefosine has been reported to induce NO in U937 cells [61]. However, we were unable to detect NO production following incubation of 10 µM edelfosine with J774 macrophages (<2 nmol nitrites/106 J774 cells after 18 h incubation), unlike cell incubation with 10 µg/ml LPS (100 nmol nitrites/106 J774 cells after 18 h incubation). Likewise, edelfosine treatment failed to prompt NO synthesis in mouse bone marrow-derived macrophages and rat alveolar macrophages (data not shown). These data suggest that the killing effect of edelfosine on macrophage-residing Leishmania parasites does not depend on NO generation.
We next examined the in vivo antileishmanial activity of edelfosine in BALB/c mice infected subcutaneously in the footpad with 2×106 infective stationary-phase L. major promastigotes. In agreement with previous estimates [29], [30], [49], we found that a daily oral administration of 15 or 30 mg/kg edelfosine was well tolerated, 45 mg/kg being the maximum tolerated dose, following toxicity analyses, where animals were monitored for body weight loss or any appreciable side-effect, including changes in strength and general condition (data not shown). We found that a daily oral administration dose of 15 mg/kg body weight edelfosine achieved a remarkable inhibition of both footpad inflammation (Figure 4A) and parasitic load (Figure 4B), as assessed by caliper measures of footpad swelling and limiting dilution assays, respectively, at the end of the 28-day treatment period. In comparison experiments, we found that oral treatment of L. major-infected BALB/c mice with edelfosine was slightly more effective than with miltefosine, although differences were not statistically significant (data not shown). The dose of edelfosine used in our assays was similar to the dose used for miltefosine in mouse models, ranging from 2.5 to 25 mg/kg of body weight/day given orally, and being 20 mg/kg/day the most widely used dose for in vivo murine experiments [35], [39], [62]–[65]. In addition, because the molecular masses for edelfosine and miltefosine are 523.7 and 407.6, respectively, the edelfosine dose used in our assays (15 mg/kg, corresponding to 28.6 µmol/kg) was even lower than the usual miltefosine dose (20 mg/kg, corresponding to 49.1 µmol/kg) in these in vivo murine studies.
Next, we used golden hamsters as an additional experimental animal model of leishmaniasis. Hamsters have been reported to better reproduce the clinicopathological features of human leishmaniasis than mice [66]–[68]. One million promastigotes of L. panamensis and L. braziliensis were inoculated in the nose of golden hamsters, as animal models for cutaneous and mucocutaneous leishmaniasis, since the above Leishmania species can cause both cutaneous and mucocutaneous disease [69], [70]. Then, hamsters were randomized into drug-treated and drug-free control (water vehicle) groups of seven hamsters each, and the animal models for L. panamensis (Figure 5, A–C) and L. braziliensis (Figure 5, D–F) infections were monitored for the antileishmanial efficacy of edelfosine. Serial caliper measurements during the course of the assays were made to determine the rate of nose swelling (Figure 5, A and D). Progression of the disease led to a dramatic swelling and ulceration of the nose. Oral administration of edelfosine (26 mg/kg body weight) on a daily basis for 4 weeks (28 days) induced a remarkable decrease in both nasal swelling and parasitic load at the site of infection (Figure 5). This dose is lower than the miltefosine dose (40 mg/kg/day) used in a recent study with L. donovani-infected hamsters [71]. Here, we found no appreciable adverse effects on the general condition of the animals following a daily oral administration of 26 mg/kg edelfosine. The effect of edelfosine on nose swelling became evident in both L. panamensis and L. braziliensis infections after two weeks of treatment (Figure 5, A and D). The parasite loads, assessed by limiting dilution assays, were significantly diminished in both animal models following oral treatment with edelfosine (Figure 5, B and E). Untreated infected animals displayed intense swelling and ulceration in their noses, but edelfosine treatment greatly ameliorated the signs of leishmaniasis (Figure 5, E and F).
Cutaneous leishmaniasis is the most common form of leishmaniasis and is endemic in many tropical and subtropical countries [1], [2]. Common therapies for leishmaniasis for more than 60 years include the use of SbV drugs as meglumine antimoniate (Glucantime) or sodium stibogluconate (Pentostam) [2], [72]. However, extensive use of these compounds is leading to SbV resistance. Thus, parasites have become resistant to antimony in many parts of the world, and primary resistance to SbV exceeds 60% of cases of leishmaniasis in the state of Bihar in India [73]. Different Leishmania species have been shown to display distinct susceptibility to antimonials [74], [75]. In addition, susceptibility of L. donovani to SbV has been reported to follow stage transformation from promastigotes to axenic amastigotes, while resistance to SbV is acquired when amastigotes differentiate into promastigotes [76]. SbV has also been reported to be active, even though to different degrees, against a number of Leishmania spp. promastigotes and amastigotes in vitro, including L. panamensis [77]–[81]. On these grounds and because of possible clinical implications, we generated a SbV-resistant L. panamensis strain to be tested for the antiparasitic activity of the distinct ALPs. Induction of resistance to SbV in L. panamensis promastigotes was achieved by continuous in vitro exposure of these parasites to increasing Glucantime concentrations for 1 year. The SbV-resistant L. panamensis strain was able to resist concentrations of Glucantime as high as 36 mg/ml, as assessed by XTT assays, a concentration 9-fold higher than the IC50 (4 mg/ml) for wild type L. panamensis promastigotes. Because of the different susceptibility to SbV shown by certain Leishmania spp., depending on their promastigote or amastigote stage, SbV resistance of L. panamensis promastigotes was further evaluated by in vivo experiments in a hamster model. Two groups of eight hamsters each were inoculated in the nose with wild type and SbV-resistant L. panamensis promastigotes, and then, after a 6-week post-infection period, when nose swelling was clearly detected, hamsters were injected intramuscularly with 40 mg/kg body weight Glucantime (SbV), on a daily basis for 4 weeks. As shown in Figure 6 (A and B), swelling was decreased in animals infected with wild type L. panamensis, but increased in animals infected with SbV-resistant L. panamensis. In addition, macrophages infected with Leishmania amastigotes were readily observed in smears from the nose of SbV-resistant L. panamensis-infected hamsters, but not from wild type L. panamensis-infected animals, treated with SbV (Figure 6B). Moreover, the parasitic burden in the nose of the two groups of animals indicated that the amount of viable wild type L. panamensis was dramatically diminished following treatment with the pentavalent antimonial drug, but the SbV-resistant L. panamensis parasites remained viable in the in vivo assay (data not shown). These results indicate that the generated SbV-resistant L. panamensis strain was highly resistant to pentavalent antimonial treatment both in vitro and in vivo.
Next, we tested in vitro the activity of the four ALPs edelfosine, miltefosine, perifosine and erucylphosphocholine against both wild type and SbV-resistant L. panamensis promastigotes by XTT assays. We found that all ALPs were effective in inhibiting proliferation of SbV-resistant L. panamensis promastigotes showing similar IC50 values to those found against wild type L. panamensis (Figure 6C). Edelfosine was the most effective ALP against SbV-resistant L. panamensis promastigotes and no difference in edelfosine sensitivity was observed between wild type and SbV-resistant strains (Figure 6C).
Infection of hamsters with SbV-resistant L. panamensis parasites in the nose, showed that a daily oral treatment with edelfosine (26 mg/kg body weight) for 4 weeks led to a dramatic decrease in the evolution index, parasitic burden and local inflammation (Figure 6, D–F). The first signs of improvement were detected after about two weeks of treatment (Figure 6A). These data indicate that oral treatment with edelfosine was efficient against leishmaniasis caused by SbV-resistant L. panamensis parasites.
A major concern in anti-parasitic chemotherapy is the generation of drug resistance. Thus, we next analyzed the feasibility to generate drug resistance to miltefosine and edelfosine in different Leishmania species, by a gradual increase in drug concentration. We determined the time required to achieve resistance to 30 µM miltefosine or edelfosine. This drug concentration could be appropriate to distinguish between specific and unspecific effects, and thereby drug resistance was considered when parasites became resistant to a final drug concentration of 30 µM. We found that the continuous exposure of L. donovani, L. major and L. panamensis promastigotes to increasing amounts of miltefosine led to a rather rapid advent of drug resistance following 40–64 days of treatment (Table 2). However, relatively more protracted continuous treatments were required to generate edelfosine resistance in L. major and L. panamensis promastigotes (Table 2). Interestingly, whereas miltefosine treatment led to drug resistance in L. donovani after a relatively short period of time (Table 2), no drug resistance was detected after 100-day treatment of L. donovani with edelfosine (Table 2).
Our results show the in vitro and in vivo antileishmanial activity of edelfosine against different Leishmania species. The ability of edelfosine to kill distinct Leishmania spp. promastigotes and amastigotes is in general higher than other ALPs, and the antileishmanial activity of ALPs ranked edelfosine>perifosine>miltefosine>erucylphosphocholine. Edelfosine also shows a higher capacity to induce an apoptosis-like cell death in Leishmania than miltefosine (Impavido), which has been approved as the first oral drug active against visceral leishmaniasis [2]. However, recent studies have challenged the efficacy of miltefosine against some cutaneous leishmaniasis [13]–[15], [17]–[20], and relapse cases of miltefosine-treated parasites have been reported in visceral and diffuse cutaneous leishmaniasis [82]–[84] as well as in HIV-positive patients [85], [86].
Here, we have found that edelfosine shows an outstanding activity against a wide number of Leishmania spp. causing cutaneous, mucocutaneous and visceral leishmaniasis. Edelfosine was able to kill parasites in both promastigote and amastigote forms through an apoptosis-like process that involved DNA degradation, as assessed by an increase in the percentage of cells with a hypodiploid DNA content. Leishmania parasites infect macrophages wherein they reside and replicate in a fusion competent vacuole (parasitophorous vacuole). Interestingly, edelfosine efficiently killed intracellular parasite amastigotes inside macrophages, without affecting the host cells. This killing activity on intracellular parasites seems to be mainly due to a direct action of the drug on the parasite, as edelfosine was unable to induce NO generation in macrophages, while a fluorescent edelfosine analog accumulated in the intracellular parasites within macrophages.
Our data also reveal a remarkable antileishmanial activity of edelfosine in several in vivo assays using mouse and hamster animal models infected with L. major, L. panamensis or L. braziliensis. To our knowledge this is the first study using hamsters as animal models for the in vivo evaluation of ALPs against cutaneous leishmaniasis. In addition, both in vitro and in vivo evidence showed that edelfosine was very effective against SbV-resistant Leishmania parasites. This is of importance as pentavalent antimonials Glucantime and Pentostam are being used in the treatment of leishmaniasis for over more than six decades, and still they are the first line drugs of choice and the traditional treatment worldwide. However, resistance to pentavalent antimonials is emerging as a result of their widespread use. A stark example of SbV resistance is well documented in Bihar (India), which houses approximately 90% of Indias's cases of visceral leishmaniasis, representing about 50% of the world's cases, and where resistance ended the usefulness of SbV more than a decade ago [2].
A major potential drawback in the use of miltefosine could be the relatively rapid generation of drug resistance in vitro. We have found here that generation of drug resistance required longer incubation times of Leishmania spp. with edelfosine than with miltefosine. Furthermore, whereas miltefosine generated drug resistance in L. donovani following a 40-day treatment, no resistance to edelfosine was detected after 100-day incubation.
It is worthwhile to note that miltefosine treatment has been reported to be unsatisfactory against infections caused by L. braziliensis [13]–[15], [17]–[20], whereas here we have found a remarkable antiparasitic activity of edelfosine in L. braziliensis-infected hamsters. In addition, edelfosine offers a number of additional advantages as compared to miltefosine, such as the fact that edelfosine shows a potent anti-inflammatory action [87], and no apparent toxicity [87]. Leishmania parasites enter first neutrophils through the regulation of granule fusion processes that prevents any deleterious action on the parasite [88]. Leishmania parasites use polymorphonuclear neutrophils as intermediate hosts before their ultimate delivery to macrophages, following engulfment of parasite-infected neutrophils, and in this way Leishmania can escape the host immune system [89]. A significant part of the destruction caused by cutaneous leishmaniasis is due to severe inflammation at the site of infection in the skin, leading to ulceration [90]. Neutrophils are recruited into the site of infection during cutaneous leishmaniasis [91], [92], and accumulation of neutrophils have been linked to tissue damage [93]. Edelfosine induces L-selectin (CD62L) shedding, and thus prevents neutrophil extravasation to the inflammation or infection site [87]. On these grounds, leishmaniasis could be ameliorated by oral treatment of edelfosine, which could reduce the parasite burden, by direct parasite killing, as well as the ulcerative process and subsequent scar formation, by a reduction in the recruitment of neutrophils into the site of infection.
A serious drawback of miltefosine is its teratogenic effects [24], however no studies have been conducted so far for a putative teratogenic action of edelfosine.
The studies reported here provide compelling evidence for the potent antileishmanial activity of edelfosine, which together with the low toxicity profile displayed by this ether lipid and its anti-inflammatory activity, warrants further clinical evaluation as a possible alternative treatment against leishmaniasis.
|
10.1371/journal.ppat.1005368 | TNF Drives Monocyte Dysfunction with Age and Results in Impaired Anti-pneumococcal Immunity | Monocyte phenotype and output changes with age, but why this occurs and how it impacts anti-bacterial immunity are not clear. We found that, in both humans and mice, circulating monocyte phenotype and function was altered with age due to increasing levels of TNF in the circulation that occur as part of the aging process. Ly6C+ monocytes from old (18–22 mo) mice and CD14+CD16+ intermediate/inflammatory monocytes from older adults also contributed to this “age-associated inflammation” as they produced more of the inflammatory cytokines IL6 and TNF in the steady state and when stimulated with bacterial products. Using an aged mouse model of pneumococcal colonization we found that chronic exposure to TNF with age altered the maturity of circulating monocytes, as measured by F4/80 expression, and this decrease in monocyte maturation was directly linked to susceptibility to infection. Ly6C+ monocytes from old mice had higher levels of CCR2 expression, which promoted premature egress from the bone marrow when challenged with Streptococcus pneumoniae. Although Ly6C+ monocyte recruitment and TNF levels in the blood and nasopharnyx were higher in old mice during S. pneumoniae colonization, bacterial clearance was impaired. Counterintuitively, elevated TNF and excessive monocyte recruitment in old mice contributed to impaired anti-pneumococcal immunity since bacterial clearance was improved upon pharmacological reduction of TNF or Ly6C+ monocytes, which were the major producers of TNF. Thus, with age TNF impairs inflammatory monocyte development, function and promotes premature egress, which contribute to systemic inflammation and is ultimately detrimental to anti-pneumococcal immunity.
| As we age, levels of inflammatory cytokines in the blood and tissues increase. Although this appears to be an inevitable part of aging, it ultimately contributes to declining health. Epidemiological studies indicate that older adults with higher than age-average levels of inflammatory cytokines are at increased risk of acquiring, becoming hospitalized with and dying of Streptococcus pneumoniae pneumonia but how age-associated inflammation increased susceptibility to was not entirely clear. We demonstrate that the increase in the inflammatory cytokine TNF that occurs with age cause monocytes to leave the bone marrow prematurely and these immature monocytes produce more inflammatory cytokines when stimulated with bacterial products, thus further increasing levels of inflammatory cytokines in the blood. Furthermore, although old mice have higher levels of these inflammatory monocytes arriving at the site of S. pneumoniae, they are not able to clear the bacteria. By pharmacologically or genetically removing the inflammatory cytokine TNF or reducing the number of inflammatory monocytes we were able to restore antibacterial immunity in aged mice. Thus we demonstrate that monocytes are both influenced by and contributors to age-associated inflammation and that chronic exposure to age-associated inflammation increases susceptibility to S. pneumoniae due to altering monocyte maturity and function.
| Monocyte phenotype and function change with age but whether these changes contribute to susceptibility to infectious disease is unclear. In mice, monocytes can be subdivided based on their expression of the Ly6C antigen into Ly6Chigh (Ly6Chigh, CCR2high, CX3CR1low) and Ly6Clow (Ly6Clow, CCR2low, CX3CR1high) monocytes [1,2]. In humans, the functional equivalents are CD14++CD16-/+ and CD14+CD16++ monocytes, respectively [1,3]. Ly6Chigh monocyte output from the bone marrow to the blood increases in a CCR2-dependent manner early during infection [4,5], and they become the dominant monocyte subtype in the blood, preferentially homing to sites of inflammation[6]. Ly6Chigh monocytes produce high levels of inflammatory cytokines[4,5,7]; thus, they are often called “inflammatory monocytes”.
In the elderly, numbers of circulating CD14++CD16+ and CD14++CD16- monocytes, are significantly higher[8]. CD14++CD16+ monocytes derived from elderly individuals are more senescent (i.e. have shorter telomeres) than other monocyte subsets and produce more pro-inflammatory cytokines (IL6, TNF, IL1β, IL12p70) and have higher levels of some chemokine receptors (e.g. CCR2, CCR5, CCR7, CX3CR1) [9,10]. Due to their ability to produce large amounts of pro-inflammatory cytokines, Ly6Chigh monocytes contribute to the pathology of several models of chronic inflammation [11,12,13,14,15,16,17]. During chronic inflammatory conditions, the number of circulating Ly6Chigh monocytes increase progressively[18] and their ablation is an effective strategy for decreasing pathology [16,17,19,20]. Whether Ly6Chigh monocytes contribute to chronic age-associated inflammation and increased susceptibility to infection is not known and is the focus of this study.
Aging is accompanied by an increase in the levels of pro-inflammatory cytokines such as tumour necrosis factor (TNF) and interleukins 1β (IL1β) and 6 (IL6) in the serum and tissues, a phenomenon that has been termed “inflamm-aging”[reviewed in[21,22]]. This age-associated, systemic state of chronic, low-grade inflammation (defined as “para-inflammation” by Medzhitov[23])is well-documented although its cellular source has yet to be definitively identified. Adipose tissue[24], CD4+ T cells or macrophages[25,26] have all been proposed to contribute. Increases in serum cytokines (particularly IL6 and TNF) are generally thought to be a pathological consequence of aging, as they correlate with risk of classical “diseases of age” such as dementia[27], stroke[28], cardiovascular disease[29] as well as frailty[30,31] and all-cause mortality[32,33]. Conversely, lower than average levels of age-associated inflammation predict good health in age[34]. Furthermore, most chronic inflammatory conditions are characterized by increased numbers of CD14++CD16+ and/or CD14++CD16- monocytes [35,36,37,38,39,40,41]. Herein, we investigate the role of monocytes, which are known to be critical mediators of chronic inflammation, contribute to age-associated inflammation.
Inflamm-aging contributes to susceptibility to infectious disease, and particularly pneumonia, which is a major cause of death in the elderly[42]. Susceptibility to pneumonia correlates with increased levels of IL6 and TNF before an infection [43,44,45]. When young mice are infused with TNF, they become as susceptible to experimental infection with Streptococcus pneumoniae as old mice[46]. Using a mouse model of pneumococcal colonization, we investigated whether changes in monocyte phenotype adversely affect host defense towards S. pneumoniae. We show that with age that there is an in increase in circulating Ly6C+ monocytes during the steady state due to increased expression of CCR2. Using heterochronic bone marrow chimeras, we demonstrate that the aging microenvironment, rather than intrinsic changes in myeloid progenitors, drives changes in monocyte phenotype, including decreased expression of F4/80 (a marker of maturity), and increased expression of CCR2 (required for monocyte mobilization). We demonstrate that age-associated increases in TNF are the driving factor behind changes in monocyte phenotype, as TNF deficiency or treatment with anti-TNF antibodies normalizes expression of CCR2 on Ly6C+ monocytes. Decreased CCR2 expression results in decreased numbers of monocytes in the circulation and reduced production of TNF and IL6. Finally, we demonstrate that, although TNF levels and the recruitment of Ly6C+ monocytes are increased in old mice during nasopharyngeal S. pneumoniae colonization, this, counterintuitively, results in diminished bacterial clearance.
To our knowledge, this is the first mechanistic study that investigates the role of Ly6C+ monocytes as central mediators of inflamm-aging and demonstrates that TNF is a key contributor to age-associated defects in myeloid phenotype and anti-bacterial function. These data indicate that Ly6C+monocyte frequency and increased production of pro-inflammatory cytokines contributes to both age-associated inflammation and declining anti-bacterial immunity.
It has been reported that with age the proportion of myeloid cells and cytokines in the blood is increased. We quantitated circulating leukocyte populations in old (18–22 mo) mice and found that, consistent with previously published data[47,48], there was a decrease in the percentage of T cells and an increase in the number of myeloid cells when compared with young (10–14 wk) mice (Fig 1A & S1A Fig). Analysis of monocyte subsets indicated that the absolute number of both Ly6Chigh and Ly6Clow monocytes was increased with age (Fig 1A). An increase in Ly6Chigh monocyte frequency within the blood of old mice was paralleled by a similar increase in the bone marrow (Fig 1B), suggesting that increased myelopoiesis within the bone marrow may precede increased numbers of these cells in the blood. Consistent with this, we also found that the number of M-CSF responsive cells (myeloid precursors and monocytes capable of differentiating into bona fide macrophages ex vivo) in the bone marrow was significantly increased with age (S1C Fig).
The C-C chemokine receptor type 2 (CCR2) is expressed at high levels on Ly6Chigh monocytes and is essential for their entry into the blood in response to the production of CCL2[49]. Since CCR2 is required for monocytes, and especially Ly6Chigh monocytes, to leave the bone marrow and enter the blood, we hypothesized that enhanced CCR2 expression on Ly6Chigh monocytes could prompt their premature emigration from the bone marrow and could explain the increased number of Ly6Chighmonocytes seen with age. CCR2 expression was measured on Ly6Chigh monocytes in the blood and bone marrow of old mice and found to be dramatically increased (Fig 1C). Consistent with previous research[1], CCR2 expression was more pronounced on Ly6Chigh monocytes (S1E Fig). As Ly6Chigh monocytes represent an intermediate stage in monocyte-to-macrophage differentiation, we investigated their maturity using the monocyte/macrophage maturity marker, F4/80. Interestingly, we found that there was an inverse relationship between CCR2 expression and F4/80 expression on Ly6Chigh monocytes in the blood of old mice. With age, these cells showed significantly decreased levels of F4/80 (Fig 1D), suggesting that their increased CCR2 expression may prompt these cells to enter the circulation in an immature state. When CCR2 expression was measured on myeloid precursors undergoing M-CSF-stimulated differentiation into macrophages, increased CCR2 expression occurred during an intermediate stage of differentiation (day 5) on cells from old mice (S1D Fig).
To determine whether increased CCR2 expression was sufficient to increase Ly6Chigh monocyte egress, we intraperitoneally injected young and old mice with 100 nM of CCL2 and measured Ly6Chigh monocyte recruitment after 4 hours. We found that despite administering an equivalent dose of CCL2, Ly6Chigh monocyte recruitment to the peritoneum was increased ~5-fold in old mice relative to young mice (Fig 1E). A less dramatic increase in Ly6Clow monocytes was also observed (Fig 1E), consistent with previous studies.
Since we found that there was an expansion of monocytes with age and these cells are known to be potent producers of pro-inflammatory cytokines, we postulated that they might contribute significantly to age-associated inflammation. To determine whether the increased numbers of monocytes with age contributed to age-associated increases in IL6 production, we targeted this cell population using carboxylated polystyrene microparticles (PS-MPs), which have been shown by others to lead to a reduction of primarily Ly6Chigh monocytes in the blood[50]. We found that when circulating monocytes were decreased in old mice (Fig 2A), this reduced circulating levels of IL6 (Fig 2B). In humans, CD14++CD16+HLA-DR+/intermediate monocytes are the biggest producers of inflammatory cytokines under a variety of stimulation conditions [3]. Intracellular cytokine staining reveals that of the three human monocyte populations (classical, intermediate, non-classical) intermediate monocytes are the major producers of TNF (Fig 3A) and IL6 (Fig 3B) after stimulation with LPS or S. pneumoniae and older donors (63–70 yrs) produce more cytokines than younger donors (26–52 yrs). Additionally, CD14+ monocytes isolated from PBMCs from older donors produced more TNF (Fig 3C) and IL6(Fig 3D) in response to LPS than did younger donors. As in mice, the numbers of intermediate monocytes may be influenced by levels of age-associated inflammation since the frequency of intermediate monocytes, are positively correlated with plasma TNF (Fig 3E) as has been shown to occur in other chronic inflammatory conditions [51]. A weaker correlation (p<0.02) was observed between TNF levels and the numerically dominant classical monocytes and no correlation was found between non-classical monocytes and TNF (p = 0.2).
To determine whether age-related changes in Ly6Chigh monocyte numbers, phenotype and inflammatory capacity were caused by changes in the aging bone marrow microenvironment or due to intrinsic changes in the myeloid precursors themselves, we created heterochronic bone marrow chimeras. When young bone marrow was transferred to old recipient mice the number of Ly6Chigh and Ly6Clow monocytes was increased to levels comparable to old mice (Fig 1A) or old recipient mice who had received old donor marrow (Fig 4A). In contrast, young recipient mice that had received old donor marrow had Ly6Chigh and Ly6Clow monocyte numbers comparable to young mice (Fig 1A) or to young recipient mice that had received young donor bone marrow (Fig 4A). In addition, the increase in CCR2 expression observed on circulating monocytes from old mice (Fig 1C) was also observed in circulating monocytes from old recipient mice who had received young donor marrow but not on young recipient mice who received old donor marrow(Fig 4B). These data demonstrate that increases of Ly6C+ monocytes and increased CCR2 expression occur in a manner entirely dependent on the bone marrow microenvironment.
Since TNF is one of the central cytokines associated with inflamm-aging, we investigated whether TNF was sufficient to drive expansion of the Ly6Chigh monocytes. We aged TNF knockout (KO) mice (18–22 mo) and quantified Ly6Chigh monocytes in their blood. We found that, unlike their WT counterparts, old TNF KO mice did not have higher numbers of circulating Ly6Chigh monocytes (Fig 4C), but did have an increase in bone-marrow Ly6Chigh monocytes compared to their young counterparts (Fig 4D). Surface expression of CCR2 on Ly6Chigh monocytes in both the blood (Fig 4E) and the bone marrow (Fig 4F) of old TNF KO mice was comparable to the levels seen in young mice. Similarly there were no changes in Ly6Clow monocytes in aged TNF KO mice (S1D Fig).
These data suggest that increased production of Ly6Chigh monocytes in the bone marrow occur independent of TNF, but that increases in CCR2 expression on these cells in the bone marrow, and their subsequent mobilization to the blood is TNF-dependent. Consistent with our observation that Ly6C+monocytes contribute to elevated levels of circulating cytokines with age (Fig 2), old WT mice produced more IL6 than young mice following 24 hour stimulation of whole blood with either PBS or LPS (Fig 4G). In comparison, old TNF KO mice, which did not have an increase of Ly6C+monocytes in the blood did not have an age-associated increase in IL6 in whole blood in response to PBS or LPS (Fig 4G).
We investigated whether it was chronic or acute exposure to TNF that mediated age-related increases in serum IL6 and changes in monocyte phenotype and function. We first sought to determine whether increases in circulating Ly6C+ monocytes were inducible after administration of TNF. TNF (5ng/g) was administered intraperitoneally for 3 weeks, a time point chosen because it would allow for multiple cycles of monopoiesis and complete turnover of pre-formed monocytes [52]. Young mice showed a large increase in Ly6Chigh monocytes in the blood and a less dramatic increase of Ly6Clow monocytes (Fig 5A). This was accompanied by a significant increase in serum IL6 in TNF-treated, but not vehicle control mice (Fig 5B). We next asked whether blocking TNF could reduce numbers of Ly6C+ monocytes in old animals. Young and old WT mice were administered Adalimumab (HUMIRA), a human monoclonal antibody specific for TNF, or an IgG isotype control at a dose of 50 ng/g for a period of three weeks via intraperitoneal injection. Anti-TNF therapy reduced the levels of plasma TNF from an average of 1.5 pg/ml to below the level of detection (LOD = 0.25pg/ml) in old mice and decreased the number of circulating Ly6Chigh but not Ly6Clow monocytes in the blood to levels similar to young mice (Fig 5C). Anti-TNF therapy also reduced CCR2 expression on Ly6Chigh monocytes in the blood of old mice to levels that are equivalent to those seen in young mice (Fig 5D) and reduced the percentage of monocytes that stained positive for IL6 or TNF by ICS after LPS stimulation (Fig 5E). Anti-TNF treatment reduces IL6 levels in the circulation of old mice (Fig 5F) and when blood from young and old mice treated with anti-TNF or IgG controls was stimulated with LPS, IL6 levels were lower in old mice treated with anti-TNF compared to those that were treated with IgG(Fig 5G).
In order to determine if age-related changes in Ly6Chigh monocyte numbers or maturity might play a role in defective anti-bacterial immunity with age, we investigated the trafficking of these cells following nasopharyngeal colonization of young and old mice with the bacterial pathogen, S. pneumoniae. We selected this pathogen specifically because of the high burden of disease caused by S. pneumoniae in elderly individuals and because it has been previously demonstrated that its clearance from the nasopharynx is largely dependent on recruited monocytes/macrophages[53,54]. Following intranasal delivery of S. pneumoniae, we found that old mice had defects in clearance of the colonization. By Day 21 most of the young mice had cleared the bacteria, while old mice still harbored high bacterial loads (Fig 6A). Old mice were also more susceptible to bacterial invasion to the lungs at day 3 (Fig 6B) and mortality throughout the course of colonization (Fig 6C). Although serum production of CCL2 in old mice was comparable to that of young mice (Fig 6D), old mice had increased Ly6Chigh but not Ly6Clow monocyte numbers in the circulation during colonization (days 3, 7, 14, 21) (Fig 6E).
We next investigated whether the monocytes/macrophages recruited in the context of age had maturity defects (as measured by F4/80 expression). In old mice, circulating Ly6Chigh monocytes had decreased expression of F4/80 during colonization (Fig 6F), suggesting that the decreased F4/80 expression seen in the bone marrow during the steady state (Fig 1D) perpetuates following their egress during infectious challenge. Despite their inability to control bacterial loads in the nasopharynx, old mice also had a significant increase in the expression of CCL2 in the nasopharynx during colonization (Fig 6G), and had higher numbers of recruited Ly6Chigh monocytes (Fig 6H) and macrophages (Fig 6I) to the nasopharynx compared to young mice. Although resident macrophages from young and old mice present in the nasopharynx during the steady state expressed equal levels of F4/80, monocytes/macrophages recruited to the nasopharynx during S. pneumoniae colonization showed decreased expression F4/80(Fig 6J), similar to that seen in their counterparts in the blood(Fig 6F). In order to determine whether bacterial binding and internalization was different between monocytes derived from young and old mice we compared bacterial binding (measured at 4°C) and internalization/killing (measured at 37°C). Although there was a significant decrease in bacterial binding between young and old mice, this did not appear to affect internalization or bacterial killing (Fig 6K).
Although trafficking of blood monocytes was not impaired with age, old mice nonetheless displayed impaired clearance of S. pneumoniae. To explain this, we hypothesized that high levels of recruited but developmentally immature Ly6Chigh monocytes could, in fact, have negative consequences for clearance. Interestingly, TNF, which we showed caused increased numbers of Ly6Chigh monocytes in the blood (Fig 4A), was increased with age during S. pneumoniae colonization in the nasopharynx (Fig 7A) and blood (Fig 7B). We next compared nasopharyngeal bacterial loads in WT and TNF KO mice, to determine whether TNF production affected bacterial clearance. Although TNF had no effect on clearance of colonization in young mice we found that old TNF KOs had significantly fewer CFUs in the nasopharnyx compared to their old WT counterparts at day 3 (Fig 7C). Old TNF KO mice also had decreased recruitment of Ly6Chigh monocytes in the blood (Fig 7D), confirming that TNF can regulate mobilization of these cells during infection as well as in the steady state.
To determine whether the decreased recruitment of Ly6Chigh monocytes we observed was responsible for improved bacterial clearance in old TNF KO mice, we preferentially targeted this cell population using negatively-charged polystyrene microparticles (PS-MPs) (Fig 8A). We observed that there were also decreases in monocytes in the lungs, but not neutrophils with this treatment (S2 Fig). Old mice were given PS-MPs on day prior to and every 3 days during the course of S. pneumoniae colonization and bacterial loads were measured at day 7. PS-MP-treated old mice had increased survival (Fig 8B), less weight loss (Fig 8C)and lower bacterial loads in the nasopharynx (Fig 8D), lungs (Fig 8E) and spleen (Fig 8F) compared to old control mice. Similar results were observed with Gr-1 antibody, which reduces numbers of monocytes and neutrophils. These data confirm that increased trafficking of this cell type during S. pneumoniae colonization impairs host defense.
Epidemiological data strongly suggests that there is a reciprocal link between pneumonia and age-associated inflammation. Older adults who have higher than age-average levels of the cytokines TNF and IL6 in their circulation have a much higher risk of acquiring pneumonia than their peers who have lower than age-average levels[55]. Although a robust inflammatory response is generally thought to be protective against infection, in the elderly, high levels of circulating inflammatory cytokines during pneumonia are associated with more severe disease and higher mortality[56,57]. Similarly, having a chronic inflammatory disease such as dementia, diabetes, or cardiovascular disease is strongly associated with susceptibility to acquiring pneumonia [58,59,60]. Conversely, having a pneumonia in mid- to late-life can often exacerbate or accelerate sub-clinical or existing chronic inflammatory conditions and can be the harbinger of declining health and decreased quality of life[58,59]. Although descriptions of this reciprocal relationship between chronic, age-associated inflammation and pneumonia, especially that caused by S. pneumoniae, are strong, the mechanistic explanations are weak. Herein we demonstrate that monocytes, both contribute to age-associated inflammation and are impaired by chronic exposure to the inflammatory cytokine TNF, and this ultimately impairs their anti-pneumococcal function.
Our data using aged TNF KO mice or anti-TNF therapy indicate that the increased levels of TNF that occur with age impair monocyte development and ultimately anti-bacterial immunity. Although macrophages have previously been shown to promote inflamm-aging[61], our data suggest that this may begin earlier in myelopoesis since monocytes produce more inflammatory cytokines such as TNF and IL6 with age and ablation of monocytes reduces levels of serum cytokines. The increase in circulating monocytes did not occur in old TNF KO mice. Furthermore, by treating young WT mice with a low-dose regime of TNF delivered intraperitoneally, we found that Ly6C+ monocytes were increased in the blood in a manner similar to old mice, demonstrating that TNF is sufficient to increase numbers of circulating Ly6C+ monocytes. Monocytes appear to be both highly responsive to increased levels of TNF but also seem to be a major source of age-associated TNF.
Our observational studies in humans imply that the numbers of intermediate (CD14++CD16-) monocytes, which we have previously shown express higher levels of CCR2 with age [62], correlate with increased levels of TNF and contribute to hyper-inflammatory responses to bacterial infection. Studies in patients on anti-TNF therapy for rheumatoid arthritis validate our observations that TNF drives increases in inflammatory monocytes. In these patients anti-TNF therapy decreases the levels of circulating CD14++CD16- monocytes in the blood and synovial fluid as well as decreases CCR2 expression on peripheral blood mononuclear cells and thus is consistent with our data demonstrating that TNF-mediated changes in CCR2 expression are sufficient to alter the numbers of Ly6Chigh monocytes in the circulation [63,64]. Interestingly, decreases in CD14++CD16- monocytes correlate with a positive prognostic response for patients, but whether this is because they contribute directly to disease progression or the inflammatory tone of rheumatoid arthritis is not known [63].
Increases in Ly6Chigh monocytes are associated with defects in maturity. Interestingly, our chimera data demonstrate that phenotypic changes in monocytes (i.e. CCR2 and F4/80 expression) were not due to intrinsic defects in myeloid precursors but rather the influence of the bone marrow microenvironment, and, since these changes did not occur in TNF KO mice, TNF produced in the context of the microenvironment. Although F4/80 levels were equivalent on blood monocytes during the steady state, they were lower on Ly6Chigh monocytes/differentiating macrophages recruited during nasopharyngeal S. pneumoniae colonization in old mice. These changes had functional significance; despite robust Ly6Chigh monocyte recruitment and TNF production in old mice, bacterial clearance was significantly impaired. In fact, our data suggest that TNF is detrimental to clearance of S. pneumoniae from the nasopharynx with age, as old TNF KO mice had lower bacterial loads compared to their WT counterparts. Although TNF is often thought of as a key anti-bacterial cytokine, mouse studies have demonstrated that TNF is required for control for S. pneumoniae bacteremia but not for survival in lung infection[65]. In our study, old TNF KO mice recruited fewer circulating Ly6Chigh monocytes during S. pneumoniae colonization compared to old WT mice and counter-intuitively, this appeared to be protective against infection as when we depleted circulating Ly6Chigh monocytes using carboxylated polystyrene microparticles colonization, bacterial loads in the nasopharynx decreased. These data are consistent with the clinical observation that rheumatoid arthritis patients (who have high levels of circulating TNF) are at increased risk of pneumonia but that there is no increase in risk of pneumonia for patients on anti-TNF therapy [66]. Whether pneumonia risk is decreased with anti-TNF therapy is not known; however, patients on anti-TNF therapy do live slightly longer than their untreated counterparts, despite an increased risk in re-activation of chronic infections[67,68].
These observations have important therapeutic significance, since the belief that host responses to bacteria are impaired with age due to poor innate cell recruitment has been the foundation of two large clinical trials testing the use of cytokines (G-CSF) to mobilize myeloid cells as an adjunct to antibiotics and one clinical trial testing GM-CSF as an adjuvant for pneumococcal vaccination. Although mouse models (tested in young mice) showed promise, these strategies all failed when tested in populations where the median ages were 59, 61 and 68, respectively [reviewed in[69] and[70]]. Our data suggests that use of G-CSF, GM-CSF or other myeloid chemoattractant-based therapies in older adults would enhance recruitment of a population that is fundamentally immature and predisposed towards TNF and IL6 production that provides no functional benefit to the host for clearance and may even exacerbate infection.
In summary, our data suggest that monocytes are both contributors to age-associated inflammation and have altered anti-pneumococcal function as a result of age-associated inflammation. Lowering levels of TNF may be an effective strategy in improving host defence against S. pneumoniae in older adults. In fact, it has been shown that immunosuppressive steroid use in combination with antibiotics reduces pneumonia mortality in the elderly[71,72,73,74], although uptake for this therapy has been limited. Although it may be counterintuitive to limit inflammatory responses during a bacterial infection, the clinical observations and our animal model indicates that anti-bacterial strategies need to be tailored to the age of the host.
All experiments were performed in accordance with Institutional Animal Utilization protocols approved by McMaster University’s Animal Research Ethics Board (#13-05-13 and #13-05-14) as per the recommendations of the Canadian Council for Animal Care.
Participants or Power of Attorney for participants were approached to determine interest in the study. Informed written consent was obtained from the participant or their legally authorized representative approved by the Hamilton Integrated Research Ethics Board (#09–450).
Female C57BL/6J mice were purchased from Jackson Laboratories and aged in house. Colonization was performed as previously described[75]. To protect from age-related obesity aging mice are fed with a low protein diet Teklad Irradiated Global 14% protein Maintenance Diet and provided with an exercise wheel, as were young controls. The average weight of a young mouse is this study is 20g+/-1g and the old mice are on average, 27g+/-2.5g. TNF knockout mice (KO) mice (C57BL/6J background) were bred in the barrier unit at the McMaster University Central Animal Facility (Hamilton, ON, Canada) as previously described[76]. All mice were housed in specific pathogen-free conditions. Continual monitoring of the health status of mice was performed.
Monocyte frequency was measured in whole blood according to staining procedures described in [62]. Briefly, intermediate monocytes were positive for the expression of HLA-DR and CD16, stained brightly for CD14, and were negative for lymphoid and neutrophil markers (CD2, CD3, CD15, CD19, CD56, and NKp46). They are presented as cells per microlitre of whole blood, which was measured using CountBright Absolute Counting Beads (Life Technologies, CA, USA). Serum TNF was measured in elderly donors (61–100 yrs) using the Milliplex High Sensitivity ELISA kit (Millipore, ON, CA).
For intracellular cytokine staining, described in [62], the production of TNF and IL-6 was measured in classical (CD14++), intermediate (CD14++CD16+) and non-classical (CD14+CD16+) monocytes after a 6 hour incubation period in the presence of 50 ng/ml LPS and 5 x 106 CFU of heat-killed S. pneumoniae. For cytokine secretion, CD14+ monocytes were isolated from PBMCs of young(26–52 yrs) and older (63–70 yrs)by positive selection procedure (Stemcell, BC, CAN) and stimulated for 22 hours in the presence of 50 ng/ml LPS. TNF and IL-6 were measured by ELISA (eBioscience, CA, USA).
Monoclonal antibodies with the following specificities were used in this study: F4/80 (APC), Ly6C (FITC), CD45 (eFluor 450), CD11b (PE-Cy7 or PerCPCy5.5), MHC II (PerCP eFluor 710), CD3 (Alexa Fluor 700), CD4 (Alexa Fluor 605NC), Ly6G (PE), Ter119 (PE), B220 (PE), NK1.1 (PE), CCR2 (PE), IL6 (PE) or TNF (PECy7). Blood and single cell suspensions of lung were stained according to previously published procedures [75]. Total cell counts were determined using CountBright Absolute Counting Beads (Life Technologies). To attain a single-cell suspension of mouse lung tissue, half a lung was collected from each S. pneumoniae-colonized mouse and kept on ice. Immediately following, each lung was mechanically dissociated and enzymatically degraded using a Miltenyi Biotec Lung Dissociation Kit (Cat#: 130-095-927) along with the gentleMACS Octo-Dissociator with Heaters (Cat#: 130-096-427). Following dissociation as per protocol, cell suspensions were filtered (70 μM cell filter) and centrifuged at 300 x g for 10 min. Subsequently, single-cell suspensions were re-suspended in phosphate-buffered saline & processed for flow cytometry. A gating strategy for distinguishing Ly6Chigh and Ly6Clow monocytes is presented in S3 Fig.
100 nM of recombinant murine CCL2 (endotoxin-free, eBioscience) was diluted in sterile saline and administered intraperitoneally. Recruited cells were isolated via peritoneal lavage and quantitated using flow cytometry. Murine recombinant TNF (eBioscience) diluted in sterile saline was administered intraperitoneally every other day for 3 weeks at a dose of 5 ng per gram of body weight. Adalimumab (HUMIRA, Abbott Laboratories), a humanized anti-TNF antibody, or the human IgG isotype control diluted in sterile saline were administered intraperitoneally at a dose of 50 ng per gram of body weight for a period of 3 weeks.
FITC Fluoresbrite 500 nm carboxylated polsytrene microparticles (PS-MPs) were obtained from Polysciences. PS-MPs were injected via tail vein at 4 x 109 particles in 200 μl as previously described[50]. Monocyte depletion was confirmed by flow cytometry.
Serum TNF and CCL2 was measured using high-sensitivity ELISA as per manufacturer's instructions (Meso Scale Discovery). For quantitative PCR analysis, RNA Lysis Buffer (Qiagen) was used to collect nasopharyngeal RNA via nasal lavage. RNA was extracted using an RNAqueous Micro Kit (Ambion), reverse-transcribed to cDNA using M-MULV reverse transcriptase (New England Biolabs) and qPCR was performed using GoTaq qPCR Master Mix (Promega, WI, USA) and the ABI 7900HT Fast Real-time PCR System (Applied Biosystems, CA, USA) all to manufacturer’s instructions. Cycle threshold (Ct) values relative to the internal reference dye were transformed by standard curve, followed by normalization to the housekeeping gene GAPDH. Normalized results are presented as relative to an internal calibrator sample.
100μL samples of peripheral blood, were incubated with TRITC-labeled S. pneumoniae (MOI 20) resuspended in 100μL of complete RPMI at 4°C to allow binding, but not uptake. After 30 min of incubation, cells were stained for flow cytometry. Following RBC lysis (1x 1-step Fix/Lyse Solution eBioscience; ref: 00-5333-57) for 10min, cells were washed 2x with PBS to remove excess stain and non-adherent bacteria, and re-suspended in FACS wash (10% fetal bovine solution in PBS). Flow cytometry was performed and the amount of S. pneumoniae bound by Ly6Chigh monocytes was quantitated based on the mean fluorescent intensities of TRITC.
Adalimumab (HUMIRA, Abbott Laboratories), a humanized anti-TNF antibody, or the human IgG isotype control diluted in sterile saline were administered to mice. A dose of 50 ng per gram of body weight was given intraperitoneally in a volume of 200 μl every other day, for a period of 3 weeks to young and old WT mice.
Unless otherwise mentioned in the figure legend, statistical significance was determined by two-tailed Mann-Whitney-Wilcoxon tests, one-way analysis of variance or two-way analysis of variance with Fischer’s LSD post-tests where appropriate.
|
10.1371/journal.pcbi.1006001 | In silico analysis of antibiotic-induced Clostridium difficile infection: Remediation techniques and biological adaptations | In this paper we study antibiotic-induced C. difficile infection (CDI), caused by the toxin-producing C. difficile (CD), and implement clinically-inspired simulated treatments in a computational framework that synthesizes a generalized Lotka-Volterra (gLV) model with SIR modeling techniques. The gLV model uses parameters derived from an experimental mouse model, in which the mice are administered antibiotics and subsequently dosed with CD. We numerically identify which of the experimentally measured initial conditions are vulnerable to CD colonization, then formalize the notion of CD susceptibility analytically. We simulate fecal transplantation, a clinically successful treatment for CDI, and discover that both the transplant timing and transplant donor are relevant to the the efficacy of the treatment, a result which has clinical implications. We incorporate two nongeneric yet dangerous attributes of CD into the gLV model, sporulation and antibiotic-resistant mutation, and for each identify relevant SIR techniques that describe the desired attribute. Finally, we rely on the results of our framework to analyze an experimental study of fecal transplants in mice, and are able to explain observed experimental results, validate our simulated results, and suggest model-motivated experiments.
| The burgeoning integration of big data and medicine is a portent of personalized healthcare. There is a need for accurate, predictive, and mechanistic models that can be relied upon to forecast the course of a disease, test treatments in-silico, and ultimately inform the doctor’s prescription. These models, still nascent, are buoyed by rich datasets available due to recent advances in experimental methods (e.g. 16S rRNA high-throughput sequencing); one such model, which we build upon in this paper, was developed by Stein et al. to predict the growth of the infectious C. difficile (CD) and 10 other microbial genera. In this paper we extend the existing model to capture clinical treatments and biologically relevant phenomena. First, we incorporate fecal transplants and identify the mechanism by which they treat C. difficile infection (CDI). Then, we develop a methodology that endows a microbe with nongeneric attributes within the existing framework; specifically, we add CD sporulation and the developement of antibiotic-resistant strains of CD. By better reflecting the clinically relevant properties of CDI we can “personalize” a mathematical model to a given disease; this construction of generic yet customizable models will be relevant for personalized healthcare models in years to come.
| Microbiota are covertly instrumental in bodily functions including immune response [1] and colonization resistance [2, 3]. Some diseases are associated with an imbalanced microbiome, due to disproportionate regulatory action of the host in response to the microbiome composition [4]. Ironically, another pathway to disease is through antibiotic administration, which can dramatically alter microbial composition and diversity, hinder colonization resistance, and subsequently allow for pathogen infection. Specifically in this paper, we focus on antibiotic-induced C. difficile infection (CDI), a prevalent nosocomial disease [5, 6].
The advent of high-throughput sequencing provides cheap and accurate time-series abundance data of interacting microbial populations, which can then inform dynamic models that extrapolate system behavior [7, 8]. One idealization of interacting species is the generalized Lotka-Volterra (gLV) model, which assumes that the competitive dynamics of a system are entirely captured through pairwise (inter-species) and self (intra-species) interactions [9]. The gLV model ignores explicit external factors like availability of organic compounds, temperature, or location, but it is the most general possible second order differential equation that describes interacting populations, with some reasonable biological constraints.
Approximating microbiome dynamics as a gLV system is a first step towards quantifying the complex interactions between competing microbes. Inarguably this model misses many subtle, non-competition based, interactions: for example, a non-abundant type of bacteria (e.g. Escherichia) may produce proteins vital to general bacterial function (e.g. pili production) [10], but this contribution would not explicitly appear in the model.
In this paper we simulate the prevalence of C. difficile (CD) in the microbiome with a generalized Lotka-Volterra model. The work by Stein et al. [11] and Buffie et al. [12] serves as a point of departure, from which we develop a framework for evaluating the efficacy of different treatment protocols for CDI. This framework develops causal relationships between simulated therapies and microbiome compositions and also explores how bacterial adaptations such as sporulation and antibiotic-resistant mutation may be added to the gLV model. These clinically motivated approaches explain distinct qualitative aspects of CDI that are otherwise unexplored or inconsistent with previous models.
We begin by discussing the clinical background and existing models of CD infection, including the mathematical model we use in this paper, and by describing our in-silico implementations of CD treatments. Then we numerically construct phase diagrams that depict the available behaviors of the simulated system, implement in-silico clinical therapies for CDI, and quantitatively track the efficacies of these therapies. Lastly we describe how to include mechanisms for sporulation and mutation in our model, and evaluate their impacts on the efficacy of antibiotic treatment. Through these techniques, we reveal the importance of timing on the efficacy of fecal microbiota transplantation (FMT) and additionally recover the clinical recommendation for pulsed antibiotic administration when treating CD. Finally, we wield this framework to explain experimental FMT outcomes [13], validate simulated results, and propose future experiments.
The era of personalized medicine and prevalence of high-throughput sequencing will demand accurate microbiome models that can predict, diagnose, and recommend treatment for microbiome disease, and the framework developed in this paper builds upon existing models [14] to progress towards this goal.
CD is a spore-forming bacterium that can produce toxins which cause CD associated diarrhea, afflicting three million people each year [15]. CDI is especially common in the elderly and in patients who are prescribed antibiotics, since antibiotics deplete the microbiome so that ingested spores of CD— often acquired in healthcare facilities or nursing homes— may invade the vulnerable microbiome [16].
The link between antibiotic treatment, CDI, and microbiome composition was investigated by Buffie et al. [12] in a study that gathered mouse time-series phylogenetic data via high-throughput 16S rRNA sequencing. In the study three scenarios were considered, in which the mice were either left alone as a control, exposed to CD, or dosed with the antibiotic clindamycin and subsequently exposed to CD. Each scenario was performed in triplicate and consisted of around 10 time points spanning four weeks, and each time point consisted of thousands of phylogenetic 16S rRNA gene sequences which were mapped to taxonomic species and tallied. The study found that after antibiotic administration of clindamycin the mouse microbiome was less diverse (in terms of the Shannon diversity index) and vulnerable to CDI, which is consistent with clinical observations of humans who develop CDI [15, 16]. Because the anatomies of mice and humans are similar [17] and the microbiomes of both species react to changes in diet in a similar manner [18], it is common to treat the mouse model as a proxy for human CDI.
In a first attempt to model the relationship between CDI and antibiotic treatment, Stein et al. [11] proposed a generalized Lotka-Volterra (gLV) model to explain the interactions between different microbes. The parameters for this model were fit with the previously mentioned data from Buffie et al. [12]. To reduce dimensionality, Stein et al. assumed that bacteria within a given genus behave similarly, and consolidated the species-level data into genus-level data. The parameter fitting procedure was tested on in-silico data, and the fitted parameters satisfied biologically reasonable restrictions. This model— described in more detail in the text surrounding Eq (1)— produces microbiome composition trajectories which allow for simulated antibiotic treatment or exposure to CD. The Spearman rank correlation, a measure comparing the predicted microbe abundances with the experimentally measured abundances, was 0.62 (the largest achievable value is 1), and simulated trajectories for each microbe typically matched experimental trajectories within an order of magnitude. Especially, the model preserved the clinical and experimental conclusion that microbiomes treated with the antibiotic clindamycin were vulnerable to CDI.
In this paper, we start from a gLV model with previously fitted parameters [11], analyze the steady states, and then build upon this model to explore clinically motivated adaptations. In particular, we focus on simulated remedial treatments that can avoid or reverse C. difficile infected steady states, which we interpret as microbiomes suffering CDI.
The generalized Lotka-Volterra equations track the abundance of N populations xi through time; in our case, the populations are N − 1 genera plus the bacterial species CD. They read, for i ∈ 1, …, N,
d d t x i ( t ) = x i ( t ) ( μ i + ∑ j = 1 N M i j x j ( t ) + ε i u ( t ) ) . (1)
The dynamics of each population are of the same form, so the distinct individual trajectories are entirely determined by the choices of parameters and initial conditions. The parameters and initial conditions that are used to generate each figure are given in Table A of S1 Appendix. For a population xi, μi describes that population’s self-growth while Mij describes the pairwise effect of population j on population i, an interaction that can be interpreted as mutualistic, commensalistic, or parasitic. Lastly, εiu(t) is an external forcing term, which in our model represents the effect of an administered antibiotic u(t) operating with efficacy εi. In all, Eq (1) accounts for zeroth, first, and second-order terms, and approximates the competitive dynamics as a power series of the individual populations.
The procedure for parameter fitting is explained in detail and performed by Stein et al. [11]. Briefly, the fitted parameter values satisfy μi > 0 and Mii < 0 for each i, so that in isolation each population will grow and eventually self-limit. Most but not all microbial groups are inhibited by the antibiotic clindamycin. Since the interactions between populations have no clear hierarchy, we interpret the gLV model as microbes on the same trophic level competing for a shared resource— the pairwise interactions, then, effectively describe a food web which we visualize in Fig 1. While dynamical systems such as this one may in principle display an array of behaviors, with these fitted parameters we have only observed trajectories that approach biologically reasonable steady states (e.g. no periodic orbits have been observed); if we interpret the negative values in Mij as negative covariances between populations, then this stability is consistent with the covariance effect [19].
In clinical practice, CDI is defined by the presence of toxigenic CD or of CD toxins in a patient experiencing diarrhea— since there are asymptomatic carriers of CD the mere presence of CD is not sufficient for diagnosis [16]. However, since the model Eq (1) does not predict toxigenicity or toxin production, for the purposes of this paper we equate CDI to the prolonged presence of CD in a simulated microbiome.
Stein et al. [11] investigated the existence and stability of steady states for the system Eq (1). Additionally, they found that for some initial compositions, antibiotic administration can alter a microbial composition to the degree that the composition becomes susceptible to CD colonization. Building upon their work, we propose the following three clinically relevant interventions and their corresponding in-silico implementations:
We refer to a simulation which implements any combination of these external interventions as a treatment scenario.
Simulations are run in Python with the scipy package and the scipy.integrate.odeint function, which uses ordinary differential equation solver lsoda from odepack, written in FORTRAN. This solver adaptively switches between stiff and non-stiff solvers, and simulations are run with an absolute tolerance of 10−12. The code used to generate the figures in this paper is freely available at https://github.com/erijones/simulated_CDI_with_gLV.
Clinically, an external microbial transplant seeks to rejuvenate an unhealthy microbiome by infusing “healthy” microbes into the unhealthy patient. The infused samples typically consist of probiotics or a microbiome (often fecal) sample from a healthy subject [20]. Microbial transplants can confer attributes (e.g. obesity) from the donor to the donee [21], so in some sense a microbiome transplant is seeking to confer CD-colonization resistance from a CD-resilient donor to a CD-susceptible donee. Since antibiotics tend to be ineffective in treating CDI and additionally can facilitate the growth of drug-resistant mutant strains of CD by providing them with a selective advantage, fecal transplants are becoming an increasingly popular CDI treatment [15].
In our implementation we simulate transplants made of CD-resilient initial conditions, and demonstrate how these treatments can guide the system into a desired (i.e. noninfective) steady state. We model the administration of a transplant of some external microbial source v at time t* as
d x ( t ) d t = f ( x ) + v δ ( t - t * ) , (2)
where f(x) entirely encapsulates the right-hand sides of the gLV equations of Eq (1) in vector form and δ(t) is the Dirac delta function, which will serve to instantaneously add the transplant v to the microbial community x at time t*.
Under environmental pressures CD can sporulate, entering a defensive state of dormant spores that maintain the genetic information of CD while functioning at a fraction of the vegetative cell’s metabolism. These spores are resilient to antibiotics, and CD sporulation may be induced by environmental stressors such as heat [22] and alcohol [16]. While the entire gamut of environmental conditions that induce sporulation is not yet known [23], there is some evidence that in murine models antibiotics may induce sporulation [15]. The toxin-producing types of CD prevalent in nosocomial infections are notoriously difficult to kill, and their resilience has in part been attributed to sporulation [15].
Mathematically, sporulation can be modeled by creating a population of spores that, through conversion of active CD, grows when environmental conditions are harsh and declines when conditions are mild. This implementation is inspired by the treatment of latently infected T-cells in SIR models of HIV, in which the latently infected T-cells effectively hide from the immune response in the same way that the inert spore cells are uneffected by the presence of antibiotics and other microbes [24]. To capture sporulation, we augment the basic model Eq (1) by introducing a spore compartment s(t) so that the populations of the original gLV model become
d d t x i ( t ) = x i ( t ) ( μ i + ∑ j M i j x j ( t ) + ε i u ( t ) ) , d d t x c ( t ) = x c ( t ) ( μ c + ∑ j M c j x j ( t ) + ε c u ( t ) ) + β s ( t ) [ u ( t ) < u s p o r ] , and d d t s ( t ) = α x c ( t ) [ u ( t ) ≥ u s p o r ] - β s ( t ) [ u ( t ) < u s p o r ] , (3)
where the terms in square brackets should be interpreted as conditional statements that return 1 if true and 0 if false.
In Eq (3), we assume that the background microbes (which we define as the bacteria that are not CD) are uneffected by the presence of the inert spores. In the presence of antibiotics bacterial growth often acts as a step function, growing or not growing if the antibiotic concentration is lower or higher than the bacteria’s minimum inhibitory concentration (MIC) [25]. We similarly model the inflow and outflow of spores as a step function, where sporulation or germination occurs if the antibiotic concentration is larger or smaller than some threshold uspor. Since the spores are robust, we assume they have no death rate. We assume that some proportion α of the CD normally killed by antibiotics are converted to spores, so there is no explicit α term in the CD growth term, and as a consequence of this we require α < εcu(t). The experimental methods used to measure CD sporulation are not yet standardized, so there is no clear consensus on the rate of CD sporulation [22]; therefore, the sporulation parameters α, β, and uspor must be considered in a qualitative fashion.
The final augmentation we add to the gLV model is antibiotic-resistant mutation, which is culpable for many of the difficulties in treating CDI [26]. Existing antibiotic resistance models for both within-host [27] and between-host [28] versions of antibiotic-resistance typically only consider isolated bacterial systems which include only the native and mutant strains of a single bacterial species. Since we consider mutation in the gLV framework, in this paper we are able to probe the more realistic scenario of mutation occurring within a complex microbial community.
We modify the standard gLV model in Eq (1) to include terms that allow for mutation of CD into an antibiotic-resistant mutant strain of CD, denoted xm(t), so that the microbial dynamics are described by
d d t x i ( t ) = x i ( t ) ( μ i + ∑ j M i j x j ( t ) + ε i u ( t ) ) , d d t x c ( t ) = x c ( t ) ( μ c + ∑ j M c j x j ( t ) + ε c u ( t ) ) - k x c ( t ) , and d d t x m ( t ) = x m ( t ) ( μ m + ∑ j M m j x j ( t ) ) + k x c ( t ) . (4)
In addition to the standard gLV pairwise interactions, the background microbes xi of Eq (4) now interact with the CD mutant xm via the Mim term. Following existing mutation models [28], we (1) group all potential antibiotic-resistant mutations into the one mutant population xm and (2) neglect the possibility of mutation from a mutant strain xm back to the native strain xc. Furthermore, we assume that the mutations are fully resistant to antibiotics and so we omit the εm term in Eq (4). While other candidate models for antibiotic-resistant mutation exist and have been examined [29], here we focus on embedding this particular implementation of single-strain mutation into the gLV framework; other types of mutation models may be implemented in a similar way.
Since we are extrapolating beyond the mouse data collected in [12], it is not surprising that the mouse microbiome data does not distinguish between native and mutant strains of CD. Antibiotic resistant strains of CD are already rampant: one survey found that close to half of tested CD strains were resistant to at least one antibiotic, and about one quarter of tested strains were resistant to multiple antibiotics [30]. However, since the antibiotic susceptibility of CD εc is non-zero, we assume that the administered CD used to inoculate the mice is antibiotic-sensitive.
We first demonstrate the available behaviors of the system described by Eq (1). In Fig 2 we evolve our system from the nine distinct initial conditions experimentally measured by Stein et al. [11] for one particular treatment scenario, in which all initial conditions are initially treated with antibiotics and later inoculated with CD. All but one of these initial conditions are free of CD, and the remaining initial condition (IC 8) has a trace amount of CD. Despite the diverse composition of the initial conditions, under this treatment scenario the simulated trajectories evolve into only two steady states.
Then, in Fig 3 we apply four different treatment scenarios to one initial condition and identify three different reachable steady states, indicating that the initial conditions can be sensitive to which treatment scenario is applied. In this paper, within a single simulation microbe counts can vary by more than two orders of magnitude. For clarity, in our figures we plot the total microbe count on a log scale (where the total microbe count is the sum of all of the microbes in each microbial population), and then at each time we linearly color each microbial population according to its proportion at that time, so that at a given time regions of equal size correspond to equal microbe counts. The treatment scenarios that result in Fig 3 mirror the experimental mouse treatments [12] and include a control, high dosing with antibiotic (the inset of Fig 3b depicts the initial microbial response to antibiotics), low dosing with antibiotic followed by inoculation with CD, and high dosing with antibiotic followed by inoculation with CD. While the log scaling disguises changes in total microbe count between the different steady states, the steady state of Fig 3a contains seven times as many microbes as the depleted (in microbe count) steady state of Fig 3b, and contains more than twice as many microbes as the infected steady state of Fig 3d (for details on steady state compositions refer to Table B of S1 Appendix). This figure elucidates the mechanism for CDI: antibiotic-induced microbiome depletion followed by opportunistic CD colonization.
Taken together the complementary results of Figs 2 and 3 indicate that (1) for a given treatment scenario there are a limited number of achievable steady states across all initial conditions, and (2) for a given initial condition there are a variety of steady states that may be achieved across different treatment scenarios. Since the model was fit with data collected over a 30 day period but the obtained steady states are often slow to equilibrate (e.g. around 100 days in Fig 3), we should proceed with caution when extrapolating the model [31]. However, since the collected experimental data [12] roughly equilibrates by day 30, and because experimental validation on longer time scales is difficult to obtain, we follow convention [32] and study long-term system behavior through steady state analyses.
In the four weeks before the mouse experiment the mice were identically housed and fed, and during the experiment the microbial compositions of mice in the control group were approximately constant over time [12]. Hence, what we consider “initial conditions” may also be interpreted as steady states compositions of the mice before any external intervention. However, the gLV model Eq (1) does not capture these initial conditions as steady states. Over the course of the 13-day control group experiment the measured bacterial abundances maintained a relatively stable composition, with the 7 or 8 colonized bacteria varying by less than an order of magnitude over the course of the experiment. However, the model Eq (1) predicts that the same control group initial conditions (ICs 2, 5, and 8) will tend towards a simpler steady state that consists of only 3 bacteria.
This inconsistency demonstrates two limitations of the gLV model: the paucity of steady states, and the likelihood of their stability. For a generalized Lotka-Volterra system of N species there are 2N steady states, each corresponding to a different subset of bacteria— hence, there is just one steady state that consists exclusively of the 7 overlapping bacteria of the control group. Since there is variation between the control experiments, there can be no steady state that would simultaneously and precisely fit all three control trials. Furthermore, even if this steady state were relatively accurate for each trial it is unlikely that it would be stable: Stein et al. [11] found that 98% of the steady states of this system were unstable. Despite the fact that unperturbed initial conditions are not stable steady states, other qualitative features of the model (including antibiotic-induced depletion of the microbiome and opportunistic CDI) indicate the model’s utility in modeling CDI.
To summarize the available system dynamics, we construct the phase diagrams in Fig 4 by systematically sweeping through treatment scenarios for each initial condition; specifically, we vary the concentration of antibiotic treatment and whether the system is exposed to a small amount of CD.
Though we simulate nine initial conditions (ICs), the phase diagrams for some initial conditions are redundant. We classify the phase diagrams of Fig 4 as (a) CD-susceptible, ICs which become infected upon exposure by CD regardless of antibiotic usage; (b) CD-resilient, ICs which are not infected by CD regardless of antibiotic usage; and (c) CD-fragile, ICs which switch from CD-resilient to CD-susceptible upon sufficient administration of antibiotic (an antibiotic concentration of approximately 0.71). We label the five reachable steady states A through E, categorize them as CD-infected or CD-uninfected, and plot their compositions in S1 Fig. Each phase diagram is composed of a number of treatment scenarios; for each treatment scenario, a 1-day pulse of antibiotic with varying antibiotic concentration is administered on day 0, and then a small amount of CD may be administered on day 10 depending on whether the scenario is with or without CD. For reference, the experimental antibiotic dose was normalized in [11] to a 1-day pulse of antibiotic concentration 1.
With the phase diagrams of Fig 4, we may now identify the initial condition plotted in Fig 3 as CD-fragile. Furthermore, the steady states of Fig 3a–3d correspond, respectively, to steady states C, E, C, and D of Fig 4. Notably, IC 8 is CD-resilient despite the fact that the initial condition contains a small amount of CD; in fact, according to the fitted interactions the presence of CD promotes the growth of microbes that inhabit the uninfected steady state. Therefore, the isolated presence of CD inhibits colonization of an infected steady state.
One key takeaway from this survey of model behaviors is that there is no a priori obvious predictor for whether an initial condition will be CD-susceptible, CD-resilient, or CD-fragile, even with knowledge about the microbial food web. Often, the complex interplay of microbial interactions can lead to unexpected and even counterintuitive results.
In the numerical phase diagrams of Fig 4 we observe different regimes for different initial conditions, but we can substantiate this phenomenon analytically as well. We label steady state A in Fig 4 by x A *, and similarly label all other steady states. After all antibiotic has been administered, we perform a perturbative analysis of the uninfected steady states by introducing a small amount of CD (notated xc(t)) to the uninfected steady state x*. This CD will invade the steady state only if [ d d t x c ( t ) ] | x * > 0. Since the introduced xc(t) is positive, we may discern the invadability of an uninfected steady state x* by the sign of I(x*), defined to be
I(x*)≡1xc(t)[ ddtxc(t) ]|x*=(μ+Mx*(t))c. (5)
Here, we have rearranged Eq (1), removed the antibiotic dependence u(t), consolidated all the μi and Mij into their respective vector and matrix forms μ and M, and consolidated the individual populations xi(t) into their vector form x(t). Notationally, the subscript c denotes the value of a vector corresponding to the index of CD. While magnitude of the invadability |I(x*)| will correspond to the initial rate at which CD will grow or decay, only the sign of I(x*) is relevant in determining long-term susceptibility or resilience to CD.
In Table 1 we compute and compile this invadability for each of the three uninfected steady states x B *, x C *, and x E *. This table also provides the size of each steady state, where size is interpreted as the sum of all the bacterial populations (here written as the 1-norm). These conclusions provide analytic justification for why some initial conditions are susceptible to CD while others are not, and complement the phase diagrams in Fig 4.
CD is predominantly inhibited by the existence of other microbes (mostly, Mcj < 0) and so a larger |x*(t)| will tend to inhibit the growth of CD. Additionally, microbes tend to be inhibited by antibiotics (mostly, εi < 0). Together, these tendencies allude to a mechanism of CDI whereby antibiotic administration depletes the microbiome and induces CD susceptibility.
While Table 1 indicates that the reachable CD-susceptible steady states are smaller than CD-resilient steady states, the size of the initial condition had little effect on the overall steady state profile: growing or shrinking the initial condition sizes only marginally modified the resulting phase diagrams. Hence, the different steady states are robust to variations in initial condition size.
Having exhaustively explored the basic behaviorial regimes of Eq (1), we now implement in-silico two commonly administered real-world medical interventions: fecal microbiome transplantation and antibiotic administration.
Following Eq (2), we choose a microbial transplant v that is derived from a CD-resilient donor so that v is proportional to the composition of a CD-resilient initial condition, and we choose the donee microbiome to be the CD-fragile initial condition so that the effects of the transplant are more apparent. In the simulation we choose the timing of the treatment scenario to match the clinical counterpart of CDI, in which CD attempts to colonize a microbiome that has been recently depleted by antibiotics: we administer antibiotics on day 0, inoculate with CD on day 1, and insert a transplant on day d. By categorizing the resultant steady state as CD-infected or CD-uninfected and sweeping over antibiotic concentrations, relative transplant sizes, and transplant times, we realize the phase diagram in Fig 5.
Fig 5 demonstrates how a transplant can alter the steady state behavior of a system exposed to CD. We can bias the initial condition towards a CD-uninfected steady state with a proper fecal transplant via the mechanism of steady state conversion, wherein a transplant can convert a state from CD-susceptible to CD-resilient. This result, consistent with clinical practice, supplies a numerical framing for microbial transplants, narrowing the gap between real-world practice and simulation.
For transplants that are applied after antibiotic administration, this figure indicates that shorter transplant delays lead to more effective transplants. However, a transplant applied concurrently with antibiotic administration on day 0 (labeled d = 0 in Fig 5) is less effective than a transplant applied after antibiotics on day 1. This reflects that antibiotic administration depletes all microbes, so a transplant on day 1 will be unsullied by antibiotics whereas applying a transplant on day 0 will lead to the depletion of the aggregate composition.
In Fig 6 we examine the effect of transplant timing for a fixed antibiotic concentration and transplant size. Steady state conversion is most effective immediately after antibiotic administration, when the depleted microbiome has room to grow. During this time the malleable microbiome is especially responsive to transplants, and introduction of the right collection of microbes can direct the microbiome towards an infection-free steady state. However, as indicated in Fig 3, without any transplant the CD-fragile IC will naturally evolve towards a CD-susceptible steady state: hence, the timing of the transplant is critical, with more immediate transplants being more effective.
We found that out of the measured ICs, the collection of microbes that best deter CDI are derived from IC 8. This transplant replenishes the unclassified Lachnospiraceae (colored purple), which promote constituents of the uninfected steady state while inhibiting Blautia (colored yellow), a key member of the infected steady state. More surprisingly, the existence of CD in IC 8 amplifies the effect of the transplant— the same transplant but without CD was a functional but substantially less effective treatment, and similarly mediocre results were obtained with a transplant derived from the other CD-resilient initial condition (IC 2) as displayed in S2 Fig. This result is due to the deleterious and contradictory effect of CD on the CD-infected steady state. As an aside, note that since IC 8 contains CD, the transplant on day 0 effectively inoculates the system with CD on day 0 rather than on day 1.
While appropriately derived and implemented transplants are effective at reversing CDI, if we had mistakenly used a CD-susceptible donor instead, simulation confirms the intuitive expectation that these results would be flipped. Since these initial conditions are a priori unidentifiable as CD-resilient or CD-susceptible, this prompts a clinically relevant caution of whether some donor’s microbiome will be beneficial or detrimental to another’s microbiome.
In a recent experimental study by Buffie et al. [13] CD-vulnerable mice exposed to CD were given transplants consisting of a known microbial composition, and the transplant efficacy for each composition was measured. Our work on simulated transplants, which resembles the experimental study, provides context and explanation for the mechanism of the experimental transplants. In conjunction, simulated and experimental transplants could direct the development of model-guided and experimentally-validated “designer” transplants.
Antibiotic administration has traditionally been the standard approach to fight infection, but antibiotics have struggled to control CD infection: CDI has a recurrence rate of 30-65% following antibiotic treatment, while fecal transplantation has cure rates upwards of 90% [33]. Nonetheless, the Society for Healthcare Epidemiology of America (SHEA) and the Infectious Diseases Society of America (IDSA) jointly recommend treating CDI with antibiotics— often vancomycin— administered in one of three dosing regimens: a constant dosing regimen, a pulsed dosing regimen, or a tapered dosing regimen [16]. Other studies have found that vancomycin administered in tapered or pulsed doses reduced the likelihood of recurrent infections of CD, compared with treatment at a constant dosage [34]. Our model, which allows arbitrary control over the dosing schedule and concentration u(t), provides a computational framework on which we can compare the efficacy of different dosing schedules: our implementations of the three dosing regimens are plotted in S3 Fig.
Over short time scales of 1-2 days we found that given the same total amount of antibiotic, the rate at which antibiotics were administered (e.g. .5 doses for 2 days vs. 2 doses for.5 days) did not affect the eventual steady state. Over longer time scales of around 2 weeks, we observed similar behavior— the model does not capture long-term differences between different dosing regimens as long as the total amount of administered antibiotic is the same, reflecting that the time-scale for microbial growth is longer than the period over which antibiotics are typically administrated.
In modeling the different dosing regimens, we are faced with one main complication: only one antibiotic, clindamycin, was fit in [11], and furthermore clindamycin was acting to induce CDI rather than halt it. The antibiotic efficacy parameter ε therefore does not serve as a realistic proxy for vancomycin or metronidazole, antibiotics which are used to eliminate CD [34]. To simulate the effect of an antibiotic which eliminates CD, we introduce an artificial “targeted antibiotic” ε ˜, which by construction only inhibits CD; specifically, ε ˜ c = - 1 and ε ˜ i = 0 for i ≠ c.
Even with this targeted antibiotic our model does not capture significant differences between the treatment regimens, which is contrary to the clinical recommendation that pulsed or tapered dosing be preferred over constant dosing [16]. In Fig 7a and 7b we administer the same amount of targeted antibiotic via constant and pulsed dosing to the CD-infected steady state and find that the two dosing regimens produce near-identical microbe trajectories (a similar result, shown in S4 Fig, was found with tapered dosing). We propose sporulation (which acts on a much shorter time-scale) as a biologically relevant mechanism that could explain this inconsistency.
In considering the model for sporulation detailed in Eq (3), the steady state analysis we previously performed is still relevant since all steady states will eventually be spore-free— we assume that the antibiotics will eventually cease, so the sporulation term of Eq (3) will eventually decay exponentially. However, the naïve expectation that including sporulation would make CD-infected steady states more common is incorrect; once again, due to the interactions between CD and other background microbes (mediated by the interaction matrix M), the presence of CD encourages growth of the microbes that populate the infection-free steady state, and so increasing the prevalence of CD through sporulation only entrenches the non-infective steady state. Since the steady states and phase diagrams are mostly unchanged by the inclusion of sporulation, we concentrate on the dynamics of CD and CD spores on shorter time scales.
In Fig 7 we compare the effects of the standard gLV model Eq (1) (top panels) and the sporulation model Eq (3) (bottom panels) under constant and pulsed antibiotic dosing regimens. Here, we use the targeted antibiotic previously described and apply all treatments to the CD-infected steady state D. Sporulation causes spores to form as antibiotics are administered, and germinate once the antibiotics cease, which is on display in the pulsed dosing regimen scenario of Fig 7c. After targeted antibiotic administration CD recovers slightly more quickly with sporulation than without, and we interpret this expedited resurgence as a more robust CDI. For details about the parameters used in Fig 7, refer to Table A of S1 Appendix.
While many of the in-host dynamics and the biological mechanisms that underlie CD sporulation and germination remain under active investigation, studies have identified that both spores and vegetative CD colonize and persist in the gut [35], and other studies have discerned the role of bile acids in promoting spore germination [36]. Our model does not allow for the long-term establishment of spores because we assume that germination always occurs in the absence of antibiotics, and we include no mechanism for germination induced by bile acids. However, more detailed sporulation models (e.g. models that include bile acid-induced germination) may extend our basic model to build upon the qualitative features of CDI it possesses.
We emphasize that sporulation is simply a proposed biological mechanism that would modify the model’s predictions to better match clinical observations, and so these results should be interpreted in a qualitative manner. However, by including sporulation we regain (at least for short time scales) the clinically expected result [16] that pulsed dosing is more effective than constant dosing at eliminating CD— comparing the top panels with the bottom panels of Fig 7 indicates that a pulsed dosing regimen dramatically reduces the buildup of CD spores compared to constant dosing.
The mechanism of mutation, introduced in Eq (4), introduces new unconstrained parameters for the mutation rate k as well as for pairwise interactions Mim and Mmi. Here we identify an intuitive parameter choice that reflects the underlying biology, discuss the resultant steady states, and then demonstrate the effects of mutation on transient microbe dynamics.
Antibiotic-resistant mutations typically incur a fitness cost in the absence of antibiotics since resources are being allocated for defense against antibiotics rather than growth [37–39], so we choose μm = .9μc < μc. Our choice of M assumes that the background microbes interact with mutant and native types identically (i.e. Mmi = Mci and Mim = Mic for i ≠ c, m). In real systems, the mutation rate k is variable and depends on factors including the concentration of antibiotic, the type of antibiotic, the native strain type, and other environmental pressures. In our model we approximate the mutation rate as a constant k = 2 * 10−6 (in units of 1/day), a choice which is in the range of measured mutation rates of some bacteria [40], but for our purposes mostly serves to highlight the effects of mutation. For details about the parameters used in our simulations of the mutation model, refer to Table A of S1 Appendix.
Due to our parameter choices the steady states of the background microbes are largely unchanged between the mutation model and the basic model (the CD-infected steady states of the standard and mutation models are explicitly compared in Table B of S1 Appendix), but the transient dynamics shown in Fig 8 differ. In these plots the same amount of targeted antibiotic is applied to the same initial state, but Fig 8a uses the standard gLV model Eq (1) while Fig 8b uses the modified mutant gLV model Eq (4). The targeted antibiotic severely inhibits CD in the standard gLV model, but in the mutation model the antibiotic-resistant mutant compensates for the antibiotics and reinforces the colonization of CD despite the antibiotic pressures.
At the scale of a single bacterium experiments now track the growth and decline of individual lineages of bacteria when confronted with antibiotics [41], and at larger scales experiments track the spread and fixation of mutations across an entire bacterial community [42]. Since the gLV model considers populations of bacteria rather than individual cells, the individual lineages cannot be resolved. However, our model does capture the tendency of microbes with a selective advantage to outcompete microbes with lower fitness (in our case, CD mutants outcompete native CD in the presence of antibiotics), and these simulations resemble the selective sweeps found in experimental data [42].
Existing mutation models have studied native and mutant strains of bacteria in isolation, but by embedding mutation within a gLV framework we can probe the complex behaviors of mutant strains within a microbial consortia. Accordingly, the wealth of behaviors present in the simpler mutation models [28] may be observed within the gLV model with mutation Eq (4). This comprehensive and community-level view is essential in identifying, understanding, and resolving the role of antibiotic-resistant mutants in disease.
A study by Buffie et al. [13] follows the modeling method of Stein et al. [11] and fits a gLV model to both mouse and human experimental time-series data in order to predict the growth of CD following antibiotic administration. In this study they identify the microbes anticorrelated with CDI in experimental data as well as the microbes that most inhibit the growth of CD according to the interaction matrix M of the gLV model. They create and administer transplants made up of a subset of these identified microbes: four transplants consist of individual microbes in isolation while another consists of a combination of all four microbes.
Buffie et al. [13] find that of the four transplants made up of isolated microbes only one microbe is effective at curing CDI, despite the fact that the other three microbes were a priori supposed to inhibit CD. We provide two explanations for their findings, motivated by the results of our paper: (1) the ability of CD-resilient transplants to confer CD-colonization resistance is largely variable and depends on the transplant composition (e.g. the variation in transplant efficacies between Fig 5 and S2 Fig), and (2) inhibiting the growth of CD does not necessarily inhibit CD-infected steady states, since the presence of CD inhibits some of the microbes that populate the CD-infected steady state. By applying the results of our simulations to microbiome transplant experiments, we can offer a computational context for experimental results.
Many of the microbes identified by Buffie et al. [13] as potential transplants were of the genus Clostridium; in fact, the only isolated-microbe transplant that was effective in curing CDI was Clostridium scindens. If we resolve only to the genus level (as assumed when constructing the gLV model in Stein et al. [11]), this experimental result is consistent with our own transplant simulations in which a transplant made up of the CD-resilient IC 8 was significantly more effective with CD than without (Fig 5). Hence, the seemingly contradictory computational result— that the presence of CD inhibits CD-infected steady states— is validated by experiment.
Finally, we can formulate experimental questions that are couched in our computational framework. Our results point to the importance of timing when administering microbial transplants, an area that is mostly unexplored both experimentally and therapeutically, and experiments could elucidate how the the timing of transplants effects their efficacy. While Buffie et al. [13] inferred microbial interactions from a gLV model, when predicting the CD-inhibiting microbes their analysis did not include dynamic simulations; applying our method of simulated transplants to such experiments could inform the selection of “personalized” transplants, and the corresponding experiments could then be used to inform the model, the model’s limitations, and additional experiments.
In this paper, the principle driver of CDI was whether a given microbial composition was CD-resilient or CD-susceptible: for example, when administering a fecal transplant, the effectiveness of the treatment depended on the properties of the donor’s microbiome. In general these properties are unknown a priori, so picking the right donor is a gamble. In clinical practice, the screening process for potential fecal donors consists primarily of avoiding those with impaired microbiomes (e.g. due to recent antibiotic therapy) or poor health, with only about 10% of prospective donors being accepted [43, 44]. While fecal transplantation has been more successful at curing CDI than traditional antibiotic treatments [45], predictive models are not currently being implemented to quantitatively select an optimal donor. Eventually, predictive models could allow for “designer” fecal transplants that are engineered to optimally confer colonization resistance. Until donor selection methods consist of searching for optimal donors rather than excluding diseased donors, our model warns that donor selection— even of seemingly healthy donors— can have unexpected consequences.
In this paper, we follow Stein et al. [11] and model the pharmacokinetics (the in-host concentration of the antibiotic u(t)) as a pulse. In reality, clindamycin pharmacokinetics are characterized by an initial spike in the in-host antibiotic concentration, after which antibiotics are cleared from the system (driven by uptake and deterioration of the antibiotic) with a half-life of approximately 4 hours [46]. However, in our simulations we found that over short durations (1-14 days) it is the total amount of administered antibiotic that determines the long-term dynamics of Eq (1) rather than the shape of the dosing regimen u(t) (meaning that administering.5 doses for 2 days leads to the same outcome as administering 2 doses for.5 days). This insensitivity to the form of u(t) justifies our simplified pharmacokinetic form.
Additionally, we model the pharmacodynamics (the microbial response or killing rate due to antibiotics) as a linear response −εi u(t), while more realistic models use a saturating Hill function [47, 48]. However, we only use one antibiotic concentration for each simulation, corresponding to one killing rate for each simulation. For any killing rate in the range of the saturating Hill function, one may find an effective antibiotic concentration that achieves this killing rate via either the linear response or by the saturating Hill function. Since both the linear response and the saturating Hill function are monotonic, there is a nonlinear scaling for u(t) between the two response functions, meaning that our results— acquired with the linear response function— may be extended to a model that uses a saturating Hill function as long as the antibiotic scaling is observed (e.g. for the phase diagram of Fig 5, stretch the antibiotic axis). Since a linear antibiotic response qualitatively captures the same long-term dynamics that a saturating Hill function would, we are justified in using a simplified pharmacodynamic model.
The gLV model idealizes interspecies interaction, and this simplification imposes limitations on our framework. The gLV model does not explicitly model why populations grow or decay (due to the underlying resource excesses or limitations) [49], and populations are assumed to respond instantly to changes in other populations, failing to account for the time required to respond to change [50]. The number of parameters required for a gLV model scales as N2 for N species, and even with high-throughput sequencing, the number of data points per parameter is still low (e.g. roughly 5 data points per parameter in [11]). Despite these drawbacks, gLV models are commonly implemented to describe microbial growth [14] since they are predictive, manipulable, and often capture the qualitative characteristics of microbial consortia. Our framework attempts to resolve some of these limitations by treating the gLV model as a base model, then offering extensions to the model that incorporate nongeneric and mechanistic features in order to more accurately portray microbial growth.
There are many techniques that fit parameters to data [51, 52], but it is difficult to know that these fitted parameters are indeed the true parameters. Stein et al. [11] fit the parameters used in this paper with regularized linear regression with a Tikhonov regularization, but other fitting methods exist, such as LIMITs [53], a software specifically designed for fitting microbial time-series data to a gLV model. Analytically, there are sufficient conditions on the model parameters that ensure the Lyapunov stability of fixed points of generalized Lotka-Volterra systems [9], but the fitted parameter values in this paper do not satisfy all of these conditions. Leveraging fitting methods to simultaneously fit parameters to data while maintaining the analytic properties that ensure stability would alleviate the potential for non-biological divergences in microbe count (divergences which are not impossible in the given system since M has a single positive eigenvalue). Regardless of this possibility, no unstable behavior was observed in any of the simulations run for this paper, perhaps due to the relatively few symbiotic relationships.
In this paper we fuse standard SIR techniques with the gLV model, thereby introducing specific mechanisms for sporulation and mutation. In this way, our framework allows for non-generic attributes of populations to be captured and simulated, and the resulting analyses provide qualitative insights into different mechanisms. Effectively, this allows for the entire family of SIR methods to be used in conjunction with the gLV model.
As the era of personalized medicine approaches, there is a growing need for accurate computational models that reflect human biology and can predict the progress of disease. This pursuit will be aided by the availability of “big data” in medicine, but this data needs to be harnessed in a useful way. This paper addresses initial steps in developing these computation models by constructing a framework at the interface between computational models and clinical therapies. This modular framework allows for “plug-and-play” implementations of clinical techniques and observed phenomena: in this paper, we implement fecal transplant therapy, antibiotic treatment regimens, sporulation, and mutation.
Our in-silico implementations of clinical treatments were mostly congruent with the actual clinical realizations— there exist initial conditions that become susceptible to CD after exposure to antibiotics; administration of a fecal transplant can halt CDI; and (once sporulation is included) pulsed dosing is more effective at eliminating CD than constant dosing, though fecal transplants are more effective than antibiotic administration in the long run. Introducing mechanisms for antibiotic-resistant mutations and sporulation strengthens the resilience of CD to remedial treatments. In all, this framework captures the intention and qualitatively the results of real-world clinical techniques.
There are many avenues stemming from this framework that may be explored in the future, including research into “designer transplants” or of bile acid-mediated germination of CD spores. Eventually this framework could be used to suggest clinical practices, but first more experiments, better data, and novel modeling are needed. As we recognize the advancement of gene sequencing in the past few years, it is not inconceivable that user-specific personalized medicine programs, built upon mathematical models of human health, will be accessible in the future.
|
10.1371/journal.pbio.1001368 | The Evolutionary Consequences of Blood-Stage Vaccination on the Rodent Malaria Plasmodium chabaudi | Malaria vaccine developers are concerned that antigenic escape will erode vaccine efficacy. Evolutionary theorists have raised the possibility that some types of vaccine could also create conditions favoring the evolution of more virulent pathogens. Such evolution would put unvaccinated people at greater risk of severe disease. Here we test the impact of vaccination with a single highly purified antigen on the malaria parasite Plasmodium chabaudi evolving in laboratory mice. The antigen we used, AMA-1, is a component of several candidate malaria vaccines currently in various stages of trials in humans. We first found that a more virulent clone was less readily controlled by AMA-1-induced immunity than its less virulent progenitor. Replicated parasites were then serially passaged through control or AMA-1 vaccinated mice and evaluated after 10 and 21 rounds of selection. We found no evidence of evolution at the ama-1 locus. Instead, virulence evolved; AMA-1-selected parasites induced greater anemia in naïve mice than both control and ancestral parasites. Our data suggest that recombinant blood stage malaria vaccines can drive the evolution of more virulent malaria parasites.
| Vaccination can drive the evolution of pathogens. Most obviously, molecules targeted by vaccine-induced immunity can change. Such evolution makes vaccines less effective. A different possibility is that more virulent pathogens are favored in vaccinated hosts. In that case, vaccination would create pathogens that cause more harm to unvaccinated individuals. To test this idea, we studied a rodent malaria parasite in laboratory mice immunized with a component of malaria vaccines currently in human trials. We found that a more virulent parasite clone was less well controlled by vaccine-induced immunity than was its less virulent ancestor. We then passaged parasites through sham- or vaccinated mice to study how the parasites might evolve after multiple rounds of infection of mouse hosts. The parasite molecule targeted by the vaccine did not change during this process. Instead, the parasites became more virulent if they evolved in vaccinated hosts. Our data suggest that some vaccines can drive the evolution of more virulent parasites.
| Evolution is a significant challenge to malaria control. Malaria parasites have repeatedly evolved resistance to frontline drugs [1],[2], and mosquitoes have evolved resistance to all classes of approved insecticides [3],[4]. Here we report experimental studies investigating how malaria parasites might evolve in response to the “natural” selection imposed by a blood stage malaria vaccine. There is currently no licensed malaria vaccine, but a number of candidates are in human trials [5]–[9], and a vaccine targeting the pre-erythrocytic stages of Plasmodium falciparum has provided partial protection to young children in a large phase 3 trial in Africa [10].
There are two ways parasites could evolve in vaccinated populations. Vaccine developers have traditionally been concerned with epitope evolution (antigenic escape) [5],[8],[9],[11],[12]. This is where pre-existing or de novo variants of target antigens emerge and spread because they enable parasites to evade vaccine-induced immunity. Epitope evolution in response to vaccination occurs in a range of infectious agents, including hepatitis B virus [13],[14], Bordetella pertussis [15]–[18], and Streptococcus pneumoniae [19],[20]. Epitope evolution has been of particular concern for those developing blood stage malaria vaccines because target antigens are often highly polymorphic, presumably because of natural immune selection. Considerable ingenuity is currently going towards inducing variant-independent immunity against these targets [7],[21]–[27].
Epitope evolution is not the only type of evolution that can occur in response to vaccination. Immunization can also promote the emergence of variants at loci other than those targeted by vaccine-induced immunity [28]. Of particular interest are virulence determinants because, in theory, immunization can under some circumstances promote the emergence and spread of strains causing more severe disease (morbidity and mortality) [28]–[37]. The idea that vaccines could prompt the evolution of more virulent pathogens is controversial, but it has been described as one of the key unexpected insights to arise from the nascent field of evolutionary medicine [38]. Several veterinary vaccines have failed in the face of more virulent strains, apparently in the absence of epitope evolution [39]–[43].
Vaccination could favor virulent malaria parasites in two ways. First, if the primary force preventing the evolution of more virulent strains is that they kill their hosts and therefore truncate their infectious periods, keeping hosts alive with vaccination will allow more virulent strains to circulate [28]–[37],[44]. Second, immunity might be less effective against virulent strains [36]. For instance, a given antibody titer or a proliferating immune response might better control slower replicating strains than more aggressive strains [45]. Virulence factors that reduce the efficacy of primed immune responses might also have a selective advantage in vaccinated hosts [46].
Epitope evolution and virulence evolution are not necessarily mutually exclusive (some antigens can be virulence determinants), but they will have different consequences for public and animal health. Epitope evolution will erode vaccine efficacy but need not lead to more severe disease in unvaccinated individuals. Virulence evolution on the other hand would both erode vaccine efficacy and cause more severe disease outcomes in unvaccinated individuals [28],[35],[36]. Note that virulence evolution will not occur for vaccines that induce sterilizing immunity: evolution can proceed only where vaccines are leaky so that wild-type pathogens can transmit from vaccinated hosts. Because natural immunity against malaria is neither life-long nor sterilizing [47],[48], it seems likely that malaria vaccines will be leaky.
To investigate the consequences of blood stage malaria vaccination for epitope and virulence evolution, we performed serial passage experiments with the rodent malaria Plasmodium chabaudi in laboratory mice immunized with a candidate blood stage vaccine. In this system, virulence, which we measure as weight loss and particularly anemia, is positively related to transmission and competitive ability [35],[36]. Anaemia is due to direct red cell destruction by parasites and bystander killing by host responses [35],[36],[49]. As with many pathogens [50],[51], serial passage of P. chabaudi creates more virulent parasites [52]. Serial passage through mice immunized with live parasites augments this effect [30], consistent with the idea that parasites evolving in vaccinated populations could become more virulent. However, most probably, actual blood stage vaccines will consist of recombinant antigens [53]–[69]. Here we specifically test the evolutionary impact of vaccination with Apical Membrane Antigen-1 (AMA-1), a component of at least 10 vaccines in human trials [6],[66]–[68]. Antibodies elicited by this antigen are believed to confer protection by inhibiting the invasion of merozoites into red blood cells (RBCs) [55],[65],[69]. In nature, the ama-1 gene is highly polymorphic, and this antigenic diversity is thought likely to compromise vaccine efficacy in the long term [7],[70]–[72]. By immunizing with a highly defined single recombinant blood stage antigen, we could specifically determine whether antibodies raised against AMA-1 select for parasites with altered ama-1 sequence (epitope evolution) and/or for parasites that cause more severe disease (virulence evolution). We found no evidence of epitope evolution in response to vaccination, but virulence increased.
Our experimental evolution studies consisted of two serial passage experiments, denoted A and B, and four separate “evaluation” experiments to determine the virulence of the passaged lines, denoted experiments 1 to 4 (Table S1).
Before beginning experimental evolution in vaccinated animals, we wanted to test whether AMA-1 vaccine-induced immunity would be less effective against virulent parasites. In order to generate virulent parasites, we serially passaged a single clonal lineage of P. c. adami (clone DK) through 30 successive naïve mice (“serial passage A”). We then tested the performance and virulence of these virulent parasites and their less virulent ancestral precursors in sham- and AMA-1-vaccinated mice (“evaluation experiment 1”).
As expected, serial passage produced parasites that were more virulent in naïve mice than were the ancestral parasites (Figure 1A–B; anemia F1,6 = 6.5, p = 0.04). Vaccination with recombinant AMA-1 reduced anemia (Figure 1A–B). It also suppressed parasite densities (Figure 1C–D). Importantly, vaccine-induced immunity was disproportionately effective at containing the avirulent (ancestral) parasites, even though they shared complete sequence identity at ama-1 with the more virulent (derived) parasites (Figure 1C–D; total parasite density×vaccination: F1,12 = 5.4, p = 0.03). This suggests that AMA-1 vaccination has the potential to selectively favor more virulent P. chabaudi parasites. Serial passage did not affect the nucleotide sequence of ama-1 (Figure S1).
To test the evolutionary impact of vaccination with AMA-1, we contemporaneously passaged P. c. adami DK parasites every week for 20 wk through either sham-vaccinated mice or through mice vaccinated with recombinant AMA-1 (“serial passage B”). We refer to the parasite lines evolved under these contrasting conditions as C-lines and V-lines, respectively. We set out to evolve five independent replicate lines of each type, but particularly in vaccinated groups, lineage loss occurred when parasites failed to reach high enough densities to allow onward syringe passage. Failure to achieve transmissible densities in vaccinated hosts is likely to be an important evolutionary force. When lines were lost, sub-lines were derived from surviving lines. The full evolutionary history of the lines is shown in Figure S2.
Throughout the 20 passages, parasite densities on the day of passage were lower in AMA-1 vaccinated mice (Figure S3). However, the densities of those V-lines increased steadily over the successive passages, presumably because of parasite adaptation to vaccine-induced immunity.
To test whether parasite virulence had evolved during the passages, we evaluated the virulence of the parasite lines in naïve mice at two time points during the evolution of the lines: once after 10 rounds of serial passage (“evaluation experiment 2”) and again after 21 rounds (“evaluation experiment 3”). In that latter experiment, we also assayed the virulence of the ancestral parasites (passage 0). We used naïve mice in these experiments because the hypothesis under test is that evolution through AMA-1 vaccinated mice will produce parasites that do more harm to unvaccinated hosts.
Parasites passaged through AMA-1 vaccinated mice (V-lines) became more virulent than parasites passaged through sham-vaccinated mice (C-lines) (Figures 2 and 3). This difference had already arisen by the 10th passage and was still apparent after 21 passages. Thus, in naïve mice, V-line parasites from both the 10th and 21st passage “generations” caused more anemia than their comparator C-lines (Figure 2A–B; Figure 3A–B; F1,28 = 8.4, p = 0.007, and F1,27 = 6.2, p = 0.02, respectively). The V-lines also induced more anemia than the parasites from which they were derived (passage 21 versus passage 0: F1,22 = 8.2, p = 0.008). After 20 passages, no changes in ama-1 nucleotide sequence were detected in any of the lines (Figure S1). Thus, over the course of the experiment, parasites evolved in AMA-1 immunized mice became more virulent to naïve animals, and there was no evidence of nucleotide evolution at the ama-1 target sequence.
The virulence differences apparent at the 10th round of selection were associated with differences in parasite densities (Figure 2C–D). V-line parasites produced more parasites in total (Figure 2D; F1,28 = 11.5, p = 0.002), and had higher densities on the day of serial passage (F1,28 = 4.3, p = 0.04) than did C-line parasites. This is consistent with the hypothesis that selection by AMA-1 vaccination results in faster growing parasites, and that was why vaccine-evolved lines were more virulent. However, vaccine-adapted parasites from 21 passages, while still more virulent, did not achieve higher densities than C-line parasites (Figure 3D; V-lines versus C-lines: F1,27 = 1.6, p = 0.2), even though they did achieve higher densities than ancestral parasites (Figure 3D; passage 21 versus passage 0: F1,22 = 12.3, p = 0.002).
We performed another evaluation experiment, this time to compare the virulence and performance of V- and C-lines from passage 21 in AMA-1 vaccinated and sham-vaccinated mice (“evaluation experiment 4”). This allowed us to ask whether V-lines and C-lines were better adapted to the immune environment in which they evolved. Note that the half of this experiment conducted in sham-vaccinated mice closely replicates our previous evaluation of the virulence of the lines in naïve mice (“evaluation experiment 3”).
Again, we found that the V-lines were more virulent than the C-lines in control mice (Figure 4A–B; anemia F1,38 = 4.0, p = 0.05). This virulence difference was also apparent in vaccinated mice (Figure 4A–B; anemia F1,38 = 4.0, p = 0.05). The magnitude of the virulence difference was unaltered by the vaccine status of the host (Figure 4A–B; anemia, parasite×vaccination: F1,76 = 1.0, p = 0.3). Thus, vaccine-line parasites were more virulent in both sham- and AMA-1-vaccinated hosts.
If parasites had become adapted to the immune environment in which they evolved, we would expect V-lines to perform best in AMA-1-vaccinated hosts and C-lines to do better than V-lines in sham-vaccinated hosts. In fact, C- and V-lines did equally well in sham-vaccinated hosts (Figure 4C–D: F1,38 = 1.9, p = 0.1), just as they did in naïve mice in evaluation experiment 3 (Figure 3). The V-lines did achieve higher densities in AMA-1-vaccinated hosts (Figure 4C–D; F1,38 = 3.9, p = 0.05), as expected if indeed the V-lines were better adapted to vaccinated hosts, but this difference was itself not significantly different from that observed in sham-vaccinated hosts (Figure 4C–D; parasite×vaccination: F1,76 = 2.8, p = 0.09).
The main evolutionary concern of malaria vaccine developers is that antigenic escape will erode vaccine efficacy [7],[21]–[27]. Evolutionary biologists have raised a different concern, suggesting that some vaccines may drive the evolution of more virulent pathogen variants [28]–[37]. Virulence evolution would put unvaccinated individuals at risk of more severe disease should they become infected. In this study we used serial passage experiments in mice to test whether the candidate malaria blood-stage vaccine AMA-1 creates within-host conditions that selectively favor the emergence of more virulent parasite variants. In three separate phenotyping experiments, we found that parasites selected by passage through AMA-1-vaccinated mice caused more severe disease, removing 20% more RBCs in unvaccinated hosts than did the parasites evolved in unvaccinated mice (Figures 2–4, panels A and B). Importantly, vaccination did not select for antigenic escape at the ama-1 locus (Figure S1). Our data highlight the importance of considering the evolutionary repercussions of blood-stage vaccines. These vaccines evidently have the capacity to cause changes at pathogen loci other than target antigens, including those responsible for disease severity.
In our experiments, all parasites were from the same clonal lineage, so variants differing in virulence must have been generated either by mutational processes or by switching of expression among members of multigene families. Presumably the more virulent variants had a relative fitness advantage during the process of serial passage and this was disproportionately larger in vaccinated hosts. Consistent with this, AMA-1-induced immunity controlled ancestral avirulent parasites more effectively than it controlled virulent descendant parasites (Figure 1). Virulent clones out-compete less virulent clones in mixed infections [73],[74]. This competitive advantage could be associated with more aggressive extraction of resources (e.g., RBCs) during infection or better performance in immune-mediated competition [49],[75]–[78]. We expect that comparative expression or genomic analyses of our different parasite lines will open up research programs that could shed light on the virulence determinants favored by AMA-1-induced immunity.
Our experiments highlight the importance of considering all types of evolution during malaria vaccine studies. To date, reports on parasite evolution in response to candidate vaccines in both human and animal trials have focused on antigenic polymorphism. But in the few cases where virulence correlates are also available, it is impossible to disentangle the effects of antigenic polymorphism from virulence. For instance, in a human field trial in Papua New Guinea with the P. falciparum “Combination B” blood-stage vaccine, which contained recombinant 3D7 MSP-2, the vaccine was less effective against parasites of the FC27 MSP-2 genotype. This was interpreted as reflecting a strain-specific protective response [56],[79] but could have also been because the FC27 MSP-2 genotypes were more virulent [80].
Results from an AMA-1 vaccine trial in non-human primates are also consistent with the possibility that more virulent P. falciparum strains are harder to control [81]. Aotus monkeys were vaccinated with AMA-1 derived from the P. falciparum 3D7 strain and then challenged with one of two heterologous strains, FVO or FCH/4. AMA-1 vaccination afforded less protection against the FVO strain [82]. This could have been because of greater epitope dis-similarity between the AMA-1 of FV0 and 3D7 strain [81] or because of the greater virulence of FVO parasites [82].
Our data show that immunization with a recombinant malaria vaccine can create ecological conditions that favor parasites that cause greater disease severity in unvaccinated individuals. But we are a long way from being able to assess the likelihood of this occurring in human malaria populations, were a malaria vaccine to go into widespread use. Most obviously, generalizing from animal models is notoriously difficult in malaria (reviewed in this context by [76],[83]), so extreme caution is warranted. But in addition to this generic issue, many potentially important considerations remain to be evaluated. Some of these are the following.
First, in human populations there will be variation in levels of immunity due to prior infection. Whether existing natural immunity will act to enhance or suppress vaccine-imposed selection for more virulent parasite variants remains to be determined. In mice, live parasite-induced immunity [30] and AMA-1-induced immunity (this study) both promote the evolution of virulence. Further experiments are needed to determine whether both occurring together in the same host would further promote virulence or whether the effects might be less than additive. It could be argued that semi-immune individuals will already naturally be imposing selection for greater virulence in the field, and the effects of vaccination will be no worse. However, the aim of vaccination programs is to increase the number of immune people in a population, and if that is achieved, a greater proportion of the parasite population will be evolving in immune hosts.
Second, our data show that virulence rises with serial passage, as it does in many systems [51]. In nature, something must counter within-host selection for virulence (or all pathogens would be extremely virulent). It has been hypothesized that syringe passage, which by-passes natural transmission, eliminates this counter-selection against excessive virulence that arises through host death [51]. This must be true in the limit, but the virulence increases we observed here as a consequence of immunity are likely to be far from this limit because mouse death played no role in the selection process in our serial passages (Figure S2). In the P. chabaudi-mouse model, more virulent infections are more infectious to mosquitoes [35],[36], and serial passage enhances virulence and transmission stage production [30],[52]. Virulence differences generated by experimental evolution using protocols identical to ours, but using whole-parasite immunized mice rather than a recombinant antigen, were not eliminated by mosquito transmission [30],[84]. If within- and between-host selection on virulence are somehow antagonistic, an important question is how they play out in the field now, and how vaccination might affect that. Our data show that the within-host selection for virulence is strengthened by vaccine-induced immunity.
Third, our experimental design involved passaging parasites every 7 d. We chose that timing because that is after a period of rapid parasite population expansion (selection) but before naïve mice begin mounting a strong acquired response against malaria [49],[85]–[90]. This meant that, in contrast to parasites in our vaccinated mice, our control-selected lines were under only modest antibody-mediated immunity. Without further experimentation, it is unclear whether onward transmission on any other days would lead to more or less potent selection on virulent variants. Later passage could select for parasite variants that are even more resilient against the mounting immune response; earlier passage may relax selection against competitively less able variants. How that would play out in terms of transmission to mosquitoes summed over the whole infectious period remains to be determined.
Our data demonstrate that immunity induced by a recombinant antigen that is a candidate for human malaria vaccines can increase the potency of within-host selection for more virulent malaria parasites. In contrast, we found no evolution of the parasite locus controlling production of the target antigen. This does not exclude antigenic polymorphism as a challenge for vaccine efficacy, nor does it mean that virulence evolution is inevitable in populations immunized with a leaky (non-sterilizing) vaccine. But it does argue that a range of evolutionary trajectories are possible in response to vaccination [36],[44], and that epitope evolution is not the only evolution that can occur. We suggest that investigation of the impact on blood stage parasite densities and transmission should be a standard component of all Phase 3 malaria vaccine trials [10], and that whole genome analyses of parasites that survive and are transmitted from individuals in vaccinated and control arms in clinical trials should be a priority. Until there is a better understanding of the selection processes set up by imperfect vaccination, there is no reason to think that vaccine-driven evolution will occur only in genes encoding target antigens. Evaluating the medium term effects of widespread vaccination (evolutionary risk) is a substantial challenge, not least because evolutionary change is likely to occur long after clinical trials have concluded (Box 1). More generally, there is little reason to think the vaccine-driven virulence evolution we have seen will be limited to malaria parasites. Analysis of virulence evolution in range of infectious diseases for which leaky vaccines are in widespread use would be of substantial interest.
This study was carried out in strict accordance with the recommendations in the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health. The protocol was approved by the Animal Care and Use Committee of the Pennsylvania State University (Permit Number: 27452).
We used the DK clone of P. chabaudi adami, which was originally collected from thicket rats (Thamnomys rutilans) in the Congo Brazzaville [91]–[93], and subsequently cloned by limiting dilution. Laboratory genotypes are stored as stable isolates in liquid nitrogen with subscript codes used to identify their position in clonal history [52]. Mice in our experiments were female C57Bl/6, at least 6–8 wk old. Parasite densities were estimated from day 4 from samples of tail blood using Giemsa-stained thin smears and red blood cell density was estimated from day 0 by flow cytometry (Beckman Coulter), or by genotype-specific real-time quantitative real-time PCR (qPCR) assays as described previously [74]. For amplification of the DK genotype, we used the forward primer previously used to amplify AS/AJ genotypes [74] and the DK genotpe-specific reverse primer 5′ GATTGTAGAGAAGTAGAAAATACA GATACAACTAA 3′.
All mice were in one of the following three immune classes: naïve (never vaccinated with the adjuvant or the AMA-1 antigen), sham-vaccinated (which were immunized with adjuvant alone), or vaccinated (which were immunized with AMA-1 antigen plus adjuvant). We use that terminology consistently throughout.
Immunization protocols were similar to those described by Anders and others [53],[94],[95]. Briefly, vaccination was with the ectodomian of the AMA-1 protein derived from P. c. adami genotype DK [53]. AMA-1 was emulsified with Montanide ISA 720 adjuvant (Seppic). Each mouse was injected intra-peritoneally with a total of 10 µg of protein on two occasions with a 4-wk interval. Sham-vaccinated mice were injected with Montanide ISA720 plus PBS. During serial passage, and during the evaluation experiments, mice were infected with parasites 14 d after the second immunization.
We conducted two separate serial passage experiments (denoted A and B). All passages involved the syringe transfer of 0.1 ml of diluted blood containing 5×105 parasites between mice every 7 d.
We first used serial passage simply to derive a more virulent parasite lineage from the ancestral DK (“serial passage experiment A”). This allowed us to test whether AMA-1-induced immunity controlled the derived (virulent) line less successfully than the ancestral (less virulent) line. P. c. adami genotype DK294 was derived via serial passage of ancestral P. c. adami genotype DK122 after a total of 30 passages though immunologically naïve mice.
The second serial passage (“B”) was the experimental evolution phase of our study (Figure S2). This was aimed at comparing the evolutionary consequences of passaging parasites through two contrasting selection treatments: sham- and AMA-1-vaccinated mice. We used sham-vaccinated mice so as to ensure that any evolved differences could be attributed to AMA-1 antigen, and not the adjuvant. We initially aimed to derive five independent parasite lines per selection treatment. At the start (generation 1), five mice that had been previously immunized with the AMA-1 vaccine (V- lines) or a sham vaccine (C-lines) were infected with P. c. adami genotype DK247 (generation 0) (Figure S2). Parasites from each one of the five mice at generation 1 were then used to infect at least two mice at generation 2 (forming a total of 10 sublines per treatment). Duplicate infections helped reduce the possibility of losing lines during the selection phase. Thus, from generation 2 to 21, parasites from each mouse within a selection treatment were used to infect a fresh mouse in the next generation. Some lines were lost (notably where AMA-1 vaccination induced a strongly protective anti-parasitic response) (Figure S2). When lines were lost, blood from a mouse in another line within that treatment group was used to infect at least two other mice in the next generation. This protocol ensured that at each generation 10 mice were infected with parasites within each selection treatment. A total of 410 mice were used during this experimental evolution phase.
Virulence and clone performance were assessed in four separate “evaluation” experiments conducted after the serial passages. In all cases frozen lines (P. c. adami-infected erythrocytes (IRBC)) were first introduced into naïve donor mice and then into naïve or sham-immunized experimental mice. Naïve donors are used because exact doses to initiate experiment infections cannot be obtained from frozen stock. Note that this single passage in naïve mice would, if it does anything, act to narrow the virulence differences observed in our experiments. Experimental mice were intra-peritoneally injected with 1×106 IRBCs.
Evaluation experiment 1 compared the performance of parasites derived from serial passage A with their pre-passage progenitors in vaccinated and naïve hosts (Table S1). Two mice died (one control immunized and one AMA-1 immunized both infected with derived parasites). These were included in the calculation of daily densities until death as death always occurred after the peak of infection (days 17 and 15, respectively).
Three further evaluation experiments were used to compare the virulence and parasites dynamics of the C-lines and V-lines from serial passage B (Table S1): evaluation experiment 2, parasites from passage 10 in naïve mice; evaluation experiment 3, parasites from passage 21 in naïve mice; and evaluation experiment 4, parasites from passage 21 in sham- and AMA-1-vaccinated mice. In these three evaluation experiments, we compared five surviving C-lines with five surviving V-lines, with each line used to infect three mice. The lines used and their history are as shown in Figure S2. In evaluation experiment 3, nine naïve mice were also infected with the ancestral lineage (P. chabaudi genotype DK247). During evaluation experiment 2, one mouse infected with C-line parasites died on day five and was thus excluded from all analyses
To test selected parasites for epitope evolution, ama-1 nucleotide sequences of the ancestral and derived parasites from experiment one and the ancestral, C- and V-line parasites from experiments 3 and 4 (passage 21 parasites) were established using a series of overlapping oligonucleotide primers designed by reference to the published sequences of P. c. adami DK [94],[95]. Parasite DNA was extracted as previously described [74]. AMA-1 was amplified as two gene fragments: Outer Forward 5′ CTTGGGTAATTGTTCCGA 3′ and Inner Reverse 5′ GCACTTCTAACCCTTTGGT 3′; Inner Forward 5′ GGGTCCAAGATATTGTAG 3′ and Outer Reverse 5′ GGGTTTCGTCTTTTCTAC 3′. PCR was performed using Nova Taq (Novagen), with the thermocycle profile; 95°C for 12 min, then 95°C for 1 min, 57°C for 1 min, and 72°C for 1 min (×30 cycles) ending at 72°C for 10 min. Amplified DNA was visualized on a 1% agarose gel and positive amplifications were cleaned with QIAquick Gel extraction kit (Qiagen) and sequenced in both directions with the same primers that were used for amplification. Sequencing was performed by Penn State DNA sequencing core facility and sequences were aligned and analyzed using ClustalW.
All analyses were conducted in R 2.10.1 [96]. All parasite density data were log transformed to meet normality assumptions of the models. For the analysis of evaluation experiments 2–4, which determined the consequences of evolution through sham- and AMA-1-vaccinated hosts (serial passage B), differences among sub-line variances (C-lines and V-lines) were first analyzed using mixed effect linear models with sub-line as a random effect [97]. In all experiments there were no sub-line variances with selection treatments so we only report the between-selection effects. For completeness, we report the more conservative analysis, based only on line means, in Table S2.
|
10.1371/journal.pgen.1006110 | Genetic Dissection of Sexual Reproduction in a Primary Homothallic Basidiomycete | In fungi belonging to the phylum Basidiomycota, sexual compatibility is usually determined by two genetically unlinked MAT loci, one of which encodes one or more pheromone receptors (P/R) and pheromone precursors, and the other comprehends at least one pair of divergently transcribed genes encoding homeodomain (HD) transcription factors. Most species are heterothallic, meaning that sexual reproduction requires mating between two sexually compatible individuals harboring different alleles at both MAT loci. However, some species are known to be homothallic, one individual being capable of completing the sexual cycle without mating with a genetically distinct partner. While the molecular underpinnings of the heterothallic life cycles of several basidiomycete model species have been dissected in great detail, much less is known concerning the molecular basis for homothallism. Following the discovery in available draft genomes of the homothallic basidiomycetous yeast Phaffia rhodozyma of P/R and HD genes, we employed available genetic tools to determine their role in sexual development. Two P/R clusters, each harboring one pheromone receptor and one pheromone precursor gene were found in close vicinity of each other and were shown to form two redundant P/R pairs, each receptor being activated by the pheromone encoded by the most distal pheromone precursor gene. The HD locus is apparently genetically unlinked to the P/R locus and encodes a single pair of divergently transcribed HD1 and HD2 transcription factors, both required for normal completion of the sexual cycle. Given the genetic makeup of P. rhodozyma MAT loci, we postulate that it is a primarily homothallic organism and we propose a model for the interplay of molecular interactions required for sexual development in this species. Phaffia rhodozyma is considered one of the most promising microbial source of the carotenoid astaxanthin. Further development of this yeast as an industrial organism will benefit from new insights regarding its sexual reproduction system.
| Some fungi are capable of sexual reproduction without the need for a sexually compatible partner, a behavior called homothallism. For some of these fungi, it was observed that they carried in a single individual all the genes normally determining sexual identity in two distinct sexually compatible individuals, but in most cases the role of these genes is still unclear. Here we examined in detail the homothallic sexual cycle of the yeast Phaffia rhodozyma that belongs to the Basidiomycota, which is the fungal lineage that also includes the mushrooms. Phaffia rhodozyma produces astaxanthin, a pigment with antioxidant properties used in the food and cosmetic industries and is accessible to genetic modifications, so far aimed mainly at improving astaxanthin production. Here we harnessed these genetic tools to dissect the self-fertile life cycle of this yeast and found that all genes normally involved in two-partner sexual reproduction are also required for self-fertile sex in P. rhodozyma and propose a model describing molecular interactions required to trigger sexual development. We also generated preferably outcrossing strains, which are potentially useful for further improvement of P. rhodozyma as an industrial organism.
| The ability to reproduce sexually is a widespread trait in all eukaryotic lineages and it is pervasive in fungi, where key regulators of sexual behavior are encoded in specialized chromosomal regions named MAT loci [1]. Mating in fungi is governed by mechanisms of self/non-self recognition, and it usually involves two individuals that carry different alleles at the respective MAT loci and thus belong to distinct mating types. A feature common so far to all fungal mating systems seems to be the involvement of mating type specific transcription factors that contribute to post-mating sexual development [1]. In Ascomycota and in the early derived Zygomycota, MAT identity is defined at a single locus that encodes one of two possible, very dissimilar idiomorphic sequences. Hence, in these two lineages, mating systems have so far been found to be bipolar, which means that that they have a single MAT locus and spores belonging to two distinct mating types may arise after meiosis [1,2]. While much remains to be unraveled concerning the genetics of sex systems in the Zygomycota, in the Ascomycetes it is well known that two types of pheromone receptors (dubbed Ste2 and Ste3 in the model yeast Saccharomyces cerevisiae [3,4]), and cognate pheromones, are encoded in the genomes of both mating types outside the MAT locus. Each type of haploid cell expresses only one type of pheromone and one type of pheromone receptor, which define their sexual identity and promote mate recognition and cell fusion [5–7].
The Basidiomycota form together with the Ascomycota a lineage of derived fungi named Dikarya. Interestingly, early in the evolution the Basidiomycota, which include many fungi relevant for human activity like the mushrooms and the human pathogen Cryptococcus neoformans, important modifications occurred in the mechanisms of mating type determination that originated the so called tetrapolar mating system [8–10]. In the tetrapolar system, mating type is determined by two genetically unlinked MAT loci, so that after meiosis four different mating types resulting from different combinations of the two MAT loci may form. The tetrapolar system is currently thought to be the ancestral mating system in this phylum [8,9,11]. One distinctive feature of the tetrapolar system resides in the mating type defining role acquired by genes encoding pheromones and pheromone receptors. In Basidiomycota the Ste2 type of receptor was apparently lost and different Ste3 type alleles contribute to define mating type identity and are usually encoded in one of the two MAT loci [8,12]. Hence, contrarily to what happens for example in S. cerevisiae, the contribution of receptors to define the mating type is determined by which pheromone receptor and pheromone precursor alleles are encoded in one of the two MAT loci, rather than by their differential expression in opposite mating types. In many basidiomycetes, the pheromone/receptor compatibility between two potential mating partners is a pre-requisite for cell fusion [1,9,13].
The second important distinctive feature of basidiomycete mating systems is the involvement of a pair of divergently transcribed genes encoding homeodomain transcription factors, named HD1 and HD2 encoded in the second MAT locus, which constitutes another checkpoint for self/non-self recognition in addition to the pheromone/receptor (P/R) system [1,8,9]. In the species where this has been studied in detail, the HD1 and HD2 proteins potentially form a heterodimer which is a key regulator of post mating sexual development that usually involves morphological changes like the formation of dikaryotic mycelium, and culminates in the development of basidia and basidiospores [1,9,14]. However, dimerization only occurs between HD1 and HD2 proteins originating from individuals of different mating types [14,15]. Hence, haploid cells do not produce functional HD1/HD2 heterodimers, which can only form after fusion of compatible mating partners. The P/R and HD1/HD2 recognition systems function independently so that successful mating only occurs when both compatibility hurdles are overcome [1,8,9]. The tetrapolar system favors outcrossing because it decreases the chances of mating compatibility between sibling spores (25%), while the prevalent bipolar mating type system in the Ascomycetes facilitates inbreeding with a 50% compatibility chance [9,16]. Moreover, the tetrapolar system is often associated with multiallelism at both loci resulting in the generation of up to thousands of mating types as observed for some mushrooms [9,16,17].
The architecture of fungal sexual reproduction systems is obviously directed primarily at governing mating between genetically distinct individuals, usually referred to as heterothallism. However, very early on, mycologists noted that some fungi are capable of completing the sexual cycle without the need for a genetically distinct partner [18,19]. Homothallism is seen as possibly favorable in that it ensures important advantages of sex, like purging of deleterious mutations, while circumventing the necessity to find a compatible mating partner. It should be emphasized that homothallism admits outcrossing as well as inbreeding, and in one instance it was even proposed to favor outcrossing [20]. Homothallism is found across all fungal lineages and its morphological manifestations are in all cases the appearance of sexual structures in cultures derived from a single individual [1,9,18,19]. However, in spite of the apparent similitude of different instances of homothallism, the advent of molecular genetics, and more recently of comparative genomics, unraveled profound differences between the underlying molecular mechanisms. Conceptually, the simplest form is primary homothallism, consisting in the co-occurrence in the genome of a single individual of all the MAT alleles required to trigger sexual development [19]. In basidiomycetes, this would require in principle the presence in a single individual of at least one compatible pheromone/pheromone receptor pair and at least one HD1/HD2 pair capable of forming a functional heterodimer, while in Ascomycetes it would entail the presence in a single genome of all the transcription factor genes present at the MAT loci of the two opposite mating types. Following the occurrence of an endoreplication event [21] or fusion with another cell, the resulting cells are in principle genetically fully equipped to undergo meiosis and complete the sexual cycle alone. Although several ascomycete species have been described in which the presumed genetic prerequisites of primary homothallism are present (e. g. Aspergillus nidulans) [22,23], the role of MAT genes in the respective homothallic life cycle has not been fully elucidated so far [24–26]. Primary homothallism has been proposed for a handful of species in the Basidiomycota, most notably Sistotrema brinkmannii [27] and the C- biotype of the cacao pathogen Moniliophtora perniciosa [28]. Other forms of homothallism, often mechanistically very complex, have been uncovered over the years, the best studied of which is the pseudo-homothallic life cycle of S. cerevisiae that is capable of mating type switching [1,2,5]. More recently another particular form of homothallism dubbed unisexual mating (a.k.a. homokaryotic fruiting) was described in several fungi belonging to both lineages of Dikarya, most notably the human pathogens C. neoformans [29,30] and Candida albicans [30]. However, these mechanisms are clearly distinct from primary homothallism in many aspects, including the fact that species where pseudo-homothallic or unisexual mating have been identified harbor different mating types [29–31].
In the present work we undertook the genetic dissection of the homothallic life cycle of Phaffia rhodozyma, an orange pigmented astaxanthin producing yeast belonging to the Agaricomycotina, a major clade of Basidiomycota that also includes C. neoformans and the mushrooms. Astaxanthin is a carotenoid of significant economic value, used in the food and cosmetic industries because of its antioxidant activity and in aquaculture feed as a source of pigment [32]. Phaffia rhodozyma is readily transformable and its biotechnological potential fostered the development of molecular genetic tools that have been extensively used to improve astaxanthin production [33,34]. We recently uncovered a diversity hotspot of P. rhodozyma in the Southern hemisphere [35] and identified four populations within the species, which seem to correlate with different host trees [35] and two new, more distantly related lineages proposed to represent new Phaffia species [35]. The new species resemble P. rhodozyma in that all individuals are homothallic [35,36]. This means that all individual strains within each of the three Phaffia species presently considered are capable of originating typically slender basidia that usually develop four basidiospores at the apex when cultivated on medium containing polyols as sole carbon source [37]. Three types of structures (S1 Fig) were observed to originate the formation of basidia in P. rhodozyma, the most common being conjugation between a cell and its bud (a.k.a. pedogamy) [37]. However, basidia were also observed to originate from unconjugated cells where possibly endoreplication precedes meiosis as observed in other fungi [21] and from independent cells that conjugate prior to sexual development [36–38]. Similar observations were previously reported for strains belonging to the two proposed new Phaffia species [35]. Hence, as far as can be presently ascertained, the lineage corresponding to the entire genus Phaffia consists entirely of homothallic species, wherein no heterothallic individuals have been identified [35].
Recently, inspection of the genomes of three P. rhodozyma strains [39] (Nicolás Bellora and Diego Libkind, personal communication) revealed the presence of pheromone precursor and pheromone receptor (P/R) gene clusters and one HD1/HD2 pair similar to those commonly found in heterothallic basidiomycete MAT loci. The two pheromone receptor genes exhibit considerable sequence divergence and each is flanked by a unique pheromone precursor gene. The P/R locus is seemingly genetically unlinked to the HD locus, since in all P. rhodozyma draft genome assemblies examined the P/R and HD genes are located on different scaffolds. The identification of the above mentioned MAT related genes raised the question of whether they have a role in the homothallic life cycle of P. rhodozyma. In the present study we first examined, and were able to discard, the possibility that cryptic molecular mating types might exist among available P. rhodozyma strains, in spite of their homothallic behavior. We subsequently made use of genetic tools available for this species to undertake the dissection of the genetic underpinnings of sexual reproduction in P. rhodozyma. Although ploidy of P. rhodozyma strains is thought to be diploid or higher in most cases [40], all genetic studies performed so far in strain CBS 6838 indicate that it is haploid [41]. This strain was, therefore, used as genetic background for the construction of various deletion mutants that contributed to show that all six MAT genes identified seem to be functional and to play a role in sexual reproduction. Taken together, these results allowed us to propose for the first time a mechanistic model for primary homothallism in a basidiomycete.
The existence of a minimum of two MAT alleles or MAT idiomorphs among distinct individuals within a population, defining at least two compatible mating types, can be considered as indicative of the potential for heterothallism, which may co-exist with homothallism in some species [2]. In the tetrapolar (heterothallic) system, the HD genes are multiallelic (more than two alleles are found in the population of a given species) [1,9,16] while the pheromone precursor and pheromone receptor loci can be either multiallelic, as is the case in many mushroom species [9,12,42] or biallelic as in red yeasts [43] and the model smut fungus Ustilago maydis [8] among many others. In each of the three available genomes of P. rhodozyma [39] (Nicolás Bellora and Diego Libkind, personal communication) we identified two clusters encoding each a pheromone precursor and a Ste3 type receptor (henceforth referred to as P/R1 and P/R2) located at approximately 5 kb from each other (Fig 1A). The Ste3 type receptors encoded in the P/R1 and P/R2 clusters are clearly different from each other (S2 Fig). The occurrence in the same genome of multiple pheromone receptor genes is quite common in heterothallic species within the Agaricomycotina. In several well-studied cases, different combinations of P/R clusters define distinct mating type identities [9,42]. To assess whether additional pheromone receptors alleles potentially encoding different mating identities could be retrieved in P. rhodozyma, we obtained partial sequences for the pheromone receptor genes found in the P/R1 and P/R2 clusters of 14 individuals representing all four previously identified populations of P. rhodozyma (S3 Fig) [35]. This survey uncovered 10 variants for both the STE3-1 gene (from P/R cluster 1) and the STE3-2 gene (from P/R cluster 2) that are nevertheless very similar within the cohort formed by genes originating from the same cluster (S3 Fig). A comparison of the three full length amino acid sequences of Ste3-1 and Ste3-2 available (S2 Fig) shows a maximum of seven and three amino acid substitutions between variants of Ste3-1 and Ste3-2, respectively. These differences are found between strains belonging to different populations (CRUB 1149 from population A vs. CBS 7918 and CBS 6838 from population C), while the two strains belonging to the same population (CBS 7918 and CBS 6838) exhibit identical amino acid sequences. We conclude that these variants are very unlikely to encode functionally different receptors. Both the STE3-1 and STE3-2 phylogenies including all the identified variants reproduce well the previously reported relationships between the phylogenetic lineages (populations) within P. rhodozyma (S3 Fig). For the genes encoding the pheromone precursors we did not conduct a broad population survey but a comparison of the genes present in the three genomes available that encompass two distinct P. rhodozyma populations (A and C; S2 Fig) also suggests a high similarity between sequences belonging to the same cluster. Hence, taken together, these results strongly indicate that the P/R locus is not bi- or multiallelic in P. rhodozyma, as opposed to heterothallic basidiomycetes.
Many heterothallic basidiomycete species, both bipolar like Cryptococcus neoformans [44] and Ustilago hordei [45] or tetrapolar, like Leucosporidium scottii [11] or Ustilago maydis [46] harbor two functionally distinct, and thus mating type determining, Ste3 receptors. When phylogenies are constructed using the amino acid sequences of these two pheromone receptors identified in various basidiomycete lineages, trans-specific polymorphism is usually observed, meaning that each of the two Ste3 receptors found in one species is phylogenetically more closely related to the receptor in the corresponding mating type of a different species than to its counterpart in the opposite mating type of the same species [9,47,48]. To find out how the two genes of P. rhodozyma fit in this general picture, a phylogenetic analysis was conducted including the Ste3-1 and Ste3-2 receptors of P. rhodozyma CBS 6938 and their closest known relatives from heterothallic species harboring two mating type specific Ste3 receptors, from the Cryptococcus, Kwoniella and Tremella lineages. While reproducing the expected trans-specific polymorphism for all other species examined, the phylogeny in Fig 2 shows that the two P. rhodozyma receptors share a more recent common ancestor with each other than with the receptors found in the other species and consequently do not exhibit trans-specific polymorphism. In line with this, the protein sequences of Ste3-1 and Ste3-2 receptors have diverged considerably less than the two Ste3 alleles found in the other (heterothallic) species examined (50% amino acid identity, S1 Table and S2 Fig). On the other hand, they seem to be more divergent than some alleles identified in heterothallic species like Coprinopsis cinerea possessing multiple functional receptor paralogs per genome that can be up to 80% identical while exhibiting different pheromone binding properties [42].
We also identified in P. rhodozyma one pair of divergently transcribed HD-like genes (Fig 3), which are usually hyperpolymorphic in heterothallic, tetrapolar basidiomycetes [8]. The HD-like genes were located in a scaffold different from the one harboring the P/R clusters in the three available draft genome assemblies, the two loci being therefore probably genetically unlinked. A comparison of the complete protein sequences encoded by representatives of all four P. rhodozyma populations shows that the proteins are quite similar among different strains (S2 Fig). Similarly to what was observed for the P/R loci, we found that HD1 and HD2 gene sequences retrieved from strains representing all four P. rhodozyma populations recapitulate previously reported phylogenetic relationships between populations (S3 Fig). Interestingly, the homeodomain region of the HD2 protein is strictly conserved in all the variants identified, while variable amino acid positions seem to be evenly distributed throughout the remainder of the protein (S2 Fig). This is in contrast to what is usually observed for different alleles of HD2 in tetrapolar species, where the N-terminal dimerization domain is much more divergent than the C-terminal domain, as a result of negative frequency dependent selection imposed by the functional constraints on the N-terminus of both the HD1 and HD2 proteins [49]. Notably, HD2 variants representing different P. rhodozyma populations exhibited different C-terminal domain lengths, except for the shortest variant (S2 Fig) that was found in strains representing both populations B and C. A similar analysis conducted for HD1 protein sequences revealed one conservative amino acid substitution within the homeodomain, identical lengths for all the variants and an even distribution of polymorphic sites, as observed for HD2 (S2 Fig).
In summary, we conclude that no evidence could be found for the existence of polymorphisms at the MAT loci that might represent cryptic molecular mating types in a set of P. rhodozyma strains that captures the diversity found so far within the species.
In heterothallic systems, pheromones and pheromone receptors are usually involved in the process of cell-cell compatibility recognition preceding plasmogamy [1,13]. This process might conceivably be dispensable in self-fertile sexual reproduction but was, on the contrary, found to be relevant for homothallic systems in species of the genera Neurospora [50] and Sordaria [51] (Ascomycota). To address the question of whether pheromone receptor genes were required for sexual reproduction of P. rhodozyma, we produced deletion mutants of each of the two receptor genes STE3-1 and STE3-2 in turn, using homologous recombination to target chromosomal integration of antibiotic resistance markers to the P/R locus so as to delete each of the receptor genes, but leaving the pheromone precursor genes intact (Fig 1 and Table 1). Assessment of the phenotype of the individual pheromone receptor mutants, ste3-1Δ and ste3-2Δ, in sporulation medium showed that their sporulation capabilities were similar to the wild type suggesting that neither receptor is required per se for sporulation (Fig 1E). On the contrary, the double mutant, ste3-1Δ ste3-2Δ, lacking both receptors, failed completely to sporulate indicating that the two receptors are functional and redundant. We subsequently reintroduced the STE3-1 gene in the double mutant ste3-1Δ ste3-2Δ, through integration in the rDNA locus. As expected, functional complementation restored the capability of the resulting strain ste3-1Δ ste3-2Δ +STE3-1 to sporulate (Fig 1C).
Genes encoding distinct pheromone precursors (MFA1 and MFA2) flank each of the two pheromone receptor genes (STE3-1 and STE3-2) that are located at a short distance from each other in the genome (Fig 1A). This arrangement is reminiscent of the P/R locus model proposed for the mushroom Coprinopsis cinerea, a tetrapolar species also possessing multiple P/R clusters per individual [42]. Should the P. rhodozyma P/R clusters functionally resemble those of C. cinerea, the receptor would be expected to be insensitive to the pheromone encoded by its flanking gene. To find out how, if at all, the pheromones encoded in the P. rhodozyma P/R clusters interact with the receptors, we constructed two deletion mutants, ste3-1Δmfa1 and ste3-2Δmfa2. Neither of the resulting strains was able to sporulate, strongly suggesting that each receptor is activated by the pheromone encoded in the other cluster, i.e. Ste3-1 is for example very likely activated by Mfa2 (Fig 1C). In accordance with this, reintroduction of the MFA2 gene in the ste3-2Δ mfa2Δ mutant complemented its sporulation defect (Fig 1C). We subsequently constructed double deletion mutants ste3-1Δmfa2Δ and ste3-2Δmfa1Δ, each expressing a distinct, presumably interacting pheromone receptor/pheromone pair. Both double mutants were able to sporulate at normal levels (Fig 1C and 1F) showing that, as predicted, Mfa1 interacts with Ste3-2 while Mfa2 activates Ste3-1.
Finally, when the ste3-1Δ mfa1Δ and ste3-2Δ mfa2Δ mutants are co-cultured in suitable medium, a low level of sporulation is observed (Fig 1 and S2 Table), two orders of magnitude lower than wild type, which is consistent with the functional receptor remaining in each strain being activated by the pheromone produced and secreted by the other strain.
Divergently transcribed candidate HD1 and HD2 genes were also identified in the P. rhodozyma genomes, resembling the genomic arrangement of homologous genes found in most tetrapolar species (Fig 3). Hence, we proceeded to investigate whether the putative HD1 and HD2 genes uncovered in the genomes were involved in sexual development. To that end, single mutants, hd1Δ and hd2Δ, in which each of the two genes was deleted in turn, as well as a double mutant hd1Δhd2Δ, were constructed. Sporulation was abolished in the hd1Δ single mutant and in the hd1Δ hd2Δ double mutant, but vestigial sporulation was observed for the hd2Δ mutant (Fig 3 and S3 Table). Reintroduction of the HD1 gene in the rDNA of the hd1Δ strain restored the wild type sporulation phenotype, confirming that loss of sporulation was truly a consequence of HD1 deletion (Fig 3C and 3D). Similarly, ectopic expression of HD1 in the hd1Δ hd2Δ mutant restored a limited ability to sporulate to levels similar to those observed for the hd2Δ mutant (S3 Table). Taken together, these results suggest that sporulation requires both proteins but that the absence of HD2 does not completely block completion of the sexual cycle. The most likely explanation for these observations is that the divergently transcribed HD1 and HD2 genes in P. rhodozyma work together to regulate genes required for sexual development, unlike similarly arranged genes found in tetrapolar species across the Basidiomycota. To assess whether the HD1 and HD2 proteins might be capable of forming a heterodimer, we first used a bacterial two-hybrid assay [52] to try to detect an interaction between the HD1 and HD2 proteins of P. rhodozyma. The results, shown in S4 Fig, fail to consubstantiate the occurrence of an interaction between the two proteins sufficiently strong to be detected by this assay. To consubstantiate this apparent absence of interaction, we subsequently performed a second assay using the yeast two-hybrid system, which was previously used to detect interactions between HD1 and HD2 proteins of U. maydis [14] and C. neoformans [53]. To this end, fusions were constructed mimicking those successfully employed to detect interactions between the U. maydis proteins, including fusion proteins comprehending the complete HD proteins as well as shorter versions including solely the N-terminal domains normally involved in dimerization and the homeodomain region (S4 Table). In line with results obtained for the bacterial two-hybrid system, we did not detect a clear interaction using the two possible combinations of the short dimerization domains of HD1 and HD2 (S5 Fig) with only one of the four transformants expressing the HD1 and HD2 N-terminal domains showing some activation of the MEL1 reporter gene and no discernible activation of the remaining two reporter genes. However, a weak interaction signal denoted only by the MEL1 reporter gene was consistently detected in all combinations involving fusion proteins that comprehended one complete coding region of either HD1 or HD2 (S5 Fig). Taken together, these results lead to the tentative conclusion that the two P. rhodozyma proteins may interact, albeit weakly. Interestingly, in the yeast assay, we also observed interactions that might support the formation of homodimers (S5 Fig). Homodimers of homeodomain proteins were observed in S. cerevisiae where they have a well-defined function [54], but no biological roles have been ascribed so far to HD protein homodimers in the basidiomycetes.
Close inspection of the P. rhodozyma genomes available revealed the presence of all but one (MLH2) of the core genes required to complete meiosis [39] (Nicolás Bellora and Diego Libkind, personal communication). However, some available evidence also argued against the occurrence of a typical meiosis: the ploidy of different P. rhodozyma strains is uncertain and may vary [40], extensive intraspecies chromosome length polymorphisms were noted [36,55,56] and segregation of markers was observed to deviate from Mendelian distribution [36]. This prompted us to examine the dependence of P. rhodozyma sporulation on SPO11, a core meiosis gene encoding an endonuclease [57] shown to be required for meiotic recombination in C. neoformans [58]. To achieve that, we compared the ability of a spo11Δ mutant to complete the sexual cycle and the viability of the spores produced with that of the wild type strain. The results, shown in Fig 4, indicate that sporulation is less efficient in the spo11Δ mutant than in the wild type, although this difference did not reach statistical significance. In addition, the viability of F1 spores isolated from the spo11Δ mutant was significantly lower than that of the wild type. These results suggest that meiosis is indeed part of the sexual life cycle of P. rhodozyma.
Interestingly, the results described in the previous sections show that the genetic determinants and mechanisms involved in the homothallic life cycle of P. rhodozyma are similar to those of heterothallic, tetrapolar basidiomycetes. On the other hand, from the point of view of strain improvement of P. rhodozyma for biotechnological applications, it would be very useful to be able to generate strains in which outcrossing is strongly favored, because it facilitates the selection of strains harboring desirable combinations of characteristics using classical genetic approaches. This prompted us to use our recently acquired knowledge of the sexual reproduction system to try to generate obligate outcrossing P. rhodozyma strains. To do that, we generated artificial “mating types” by constructing two triple mutants harboring complementary components of the two mating recognition systems, ste3-1Δ mfa1Δ hd1Δ and ste3-2Δ mfa2Δ hd2Δ. Our model (Fig 5) predicts that a cross between these two triple mutants should almost exclusively yield spores resulting from conjugation between independent cells belonging to complementary “mating types”, while formation of basidia originating from single cells or from pedogamy is almost completely prevented by the absence of complete HD1/HD2 gene pair in each of the “mating types”. Hence, while in a cross between ste3-1Δ mfa1Δ and ste3-2Δ mfa2Δ mutants, extracellular diffusion of the pheromones would permit any of the three possible modes for formation of basidia (S1 Fig) because each mutant possessed a complete HD1/HD2 pair, in the triple mutant cross, sporulation is expected to occur almost exclusively after cell fusion brings together the HD1 and HD2 partners originating from different “mating types”. Indeed, in a cross between the artificially created “mating types”, basidia with basidiospores were formed, albeit at levels three orders of magnitude lower than wild type and two orders of magnitude lower than the cross between ste3-1Δ mfa1Δ and ste3-2Δ mfa2Δ (S2 Table). The occurrence of sporulation in the triple mutant cross, even if at very low levels, paves the way for strain improvement based on outcrossing.
Here we characterize in detail the MAT loci of P. rhodozyma, an attractive model in which to study homothallism in basidiomycetes due to its genetic amenability and to the availability of draft genome sequences [39] (Nicolás Bellora and Diego Libkind, personal communication). In line with previous observations, all strains we examined are homothallic, which agrees with the limited role previously proposed for sexual recombination and hybridization between natural P. rhodozyma populations [35]. Our survey of variants of MAT genes in strains representing the four natural populations identified so far in P. rhodozyma failed to uncover additional potentially functionally divergent homologs of MAT genes that could represent cryptic molecular mating types. Indeed, the presence in a single genome of two sets of genes encoding pheromone precursors and pheromone receptors could represent a heterothallic MAT locus resembling that of C. cinerea where each mating type may harbor several receptor and pheromone precursor genes [9,42]. However, analysis of MAT gene sequences identified in P. rhodozyma strains do not favor this possibility, because they were all very similar in the different strains examined (S3 Fig). Therefore, the present study suggests that the species may be considered exclusively homothallic in the sense that no molecular mating types were identifiable, although it admits both uniparental and biparental (outcrossing) modes, the first being apparently much more frequent across a set of strains representing diversity within the species, as supported also by population analysis [35].
Primary homothallism has been observed in a considerable number of species across the entire fungal kingdom [18,19], sometimes in a few strains of an otherwise heterothallic species [11,43] or, alternatively in lineages consisting mainly of species formed entirely of homothallic individuals with a few heterothallic species, as observed in Aspergilli [19]. In a number of cases, like for A. nidulans, it was possible to demonstrate that homothallism was associated to the presence in one individual of the entire complement of genetic information normally present in the two opposite mating types of heterothallic individuals of closely related species [22]. Also, the purposeful introduction of genes determining sexual identity in one mating type into the opposite mating type, resulted in the emergence of a homothallic phenotype [59]. Nevertheless, a thorough explanation of how these genes interact to produce the homothallic phenotype is still lacking in most cases [1,18,19,28]. The genetic makeup of MAT loci in P. rhodozyma, although suggestive of primary homothallism, departs in important aspects from what would be expected from a simple assemblage in a single genome of two mating types as typically found in the Agaricomycotina. Firstly, in sexually reproducing species most closely related to P. rhodozyma, two lineages of pheromone receptors can be clearly discerned exhibiting trans-specific polymorphism [9], while our phylogenetic analysis shows with a high degree of confidence that the two receptors in P. rhodozyma are more closely related with each other than with receptors in other species. Therefore, our results strongly suggest that P. rhodozyma receptors, possibly similarly to species in the Agaricales (mushrooms), may descend from only one of these lineages [9]. A consequence of this is that the two pheromone/receptor pairs seem to have diverged relatively recently, at least after separation of the Phaffia and Cryptococcus lineages. Both receptors are functional and are activated by the pheromone encoded in the other cluster. We posit that the most likely setting for these pheromone/receptor specificities to have evolved is a heterothallic system, which supports that an ancestor of the Phaffia lineage was likely heterothallic.
Primary homothallism is thought to be rather uncommon in Basidiomycetes, which is in line with the complex genetic underpinnings of the mating system in this phylum, involving two independent non-self recognition checkpoints. Hence, a transition from heterothallism to primary homothallism would require a genomic rearrangement gathering in the same genome two compatible versions of both the P/R and HD loci, or that one or both loci would become self-compatible. It seems more likely that homothallism in P. rhodozyma is derived from a tetrapolar ancestor, rather than bipolar, because the HD and P/R loci are apparently genetically unlinked [39] (Nicolás Bellora and Diego Libkind, personal communication) and are both required for sporulation. However, the extant MAT locus in P. rhodozyma is not simply a gathering of all MAT genes (P/R and HD) normally present in two compatible mating types in tetrapolar systems because only one HD gene pair is present, encoding proteins that are both required to promote sexual development, as shown by the absence of sporulation in the hd1Δ and by a dramatic drop in sporulation (approximately 0.2% of wild type sporulation levels remaining) observed for the hd2Δ mutant. In heterothallic systems, HD proteins form heterodimers in which the interacting partners are encoded by different mating types, thereby enforcing mating between genetically distinct individuals (outcrossing). The domain structures of HD proteins and the molecular interactions they are likely to undergo have been examined in detail in U. maydis [14,15,60,61], C. cinerea [62], Schizophylum commune [63] and C. neoformans [53]. These studies showed that the protein domains required for proper functioning of the heterodimeric transcription factor, such as high affinity DNA binding, nuclear localization signal (NLS) and transcriptional activation were not present in a single protein. For example, the C. cinerea HD1 protein harbors a transcriptional activation domain and a NLS but its homeodomain is dispensable for sexual development, while its HD2 counterpart possesses no NLS but its homeodomain is absolutely required for DNA binding by the heterodimer [64]. In this manner, undimerized partners are doomed to be unsuccessful as transcription factors: HD2 is unable to get transported into the nucleus on its own while HD1 can get transported into the nucleus due to its NLS, but will fail to effectively bind DNA [64]. Conversely, in the HD2 proteins of Heterobasidion, no bona fide homeodomain could be found, so that DNA binding probably relies entirely on the homeodomain of the HD1 protein [65]. Interestingly, in P. rhodozyma we detected at most a weak interaction between the two HD proteins. This is in contrast to the results obtained when U. maydis [14] and C. neoformans [53] HD proteins were examined in a yeast two-hybrid assay. However, phenotypes of the various deletion mutants showed clearly that HD1 and HD2 are involved in the regulation of an overlapping set of genes essential for sporulation. Taking into account the well-established mode of operation of model species within the Agaricomycotina, this leaves room for two possible interpretations: i) the two P. rhodozyma HD proteins do not form heterodimers, as normally observed for proteins encoded by the same locus, and hence, they have independent contributions to regulate a set of genes essential for sexual development; ii) the two proteins interact to promote transcriptional regulation of genes essential for sporulation, but the interaction is much weaker than those previously characterized in heterothallic species. We favor the latter possibility for a number of reasons. Firstly, it has been shown for mutant alleles in U. maydis that they can promote sexual development in vivo despite their failure to interact detectably in the yeast two-hybrid assay [14]. Secondly, we found that HD1 alone supports the ability to sporulate, albeit at low levels (Fig 3 and S3 Table). This is consistent with, for example, a scenario in which HD1 is dependent on HD2 for high affinity DNA-binding, but can bind independently sufficiently well to support a low level of sporulation in the absence of the HD2 partner. In line with this, a prominent DNA binding role has been ascribed to the HD2 proteins of other members of the Agaricomycotina, such as C. cinerea [64], S. commune [66] and C. neoformans [67].
The third reason in favor of a weak interaction between the two HD proteins pertains to how self-compatibility of the P. rhodozyma HD protein pair, if it exists, may have evolved. For this, two possibilities may be considered. Firstly, a HD gene pair encoding self-compatible proteins could conceivably form through recombination between two distinct heterothallic ancestor alleles, in which case the interaction between the two proteins would probably be expected to be sufficiently strong to be unequivocally detected in the two hybrid assays, as observed for the large majority of naturally occurring HD alleles in the heterothallic species examined [14,53]. However, in face of previous findings [14], a very likely second evolutionary path to generate a compatible HD1/HD2 pair would be the emergence of one or more mutations relieving the structural hindrance normally preventing interaction between HD proteins encoded by the same locus. It has been previously shown that a single amino acid mutation may be sufficient to remove the obstacle for self-dimerization [14]. In that case, a weak interaction permitting for example cooperative DNA binding would probably suffice for normal function provided the two proteins can reach the nucleus independently. In fact, HD1 is apparently able to reach the nucleus independently of HD2, because it is capable of promoting some sporulation in the absence of HD2, although inspection of the sequences of both proteins revealed that only HD2 possessed a candidate NLS (S2 Fig). However, and in the absence of an unequivocal experimental demonstration of an interaction, the establishment of a definite mode of action for the HD proteins will have to await a detailed dissection of the functional domains present in each of the two proteins and the identification of their DNA binding sites, as previously accomplished in other systems [14,15,49,53, 60, 61, 66, 67].
While contemplating the various possibilities to reconcile the lack of a strong interaction between the HD1 and HD2 proteins of P. rhodozyma with the phenotypes exhibited by the various mutants, we also considered the possibility that HD1 might engage an alternative dimerization partner in the absence of HD2. However, close inspection of all P. rhodozyma genomes available, failed to detect genes encoding homeodomain proteins with the appropriate domain architecture (S5 Table). Hence, we consider this possibility very unlikely.
The genetic arrangement of the two P/R clusters is strongly reminiscent of P/R clusters in many basidiomycetes, each functional receptor gene being in the vicinity of a gene encoding a functional pheromone that nevertheless fails to activate its receptor counterpart in the same cluster or locus [8,42]. It can therefore be easily envisaged that this arrangement resulted from the fusion of two heterothallic loci.
We propose a model (Fig 5) summarizing the most likely roles of the six MAT genes in the sexual cycle of P. rhodozyma given our present knowledge concerning the genetics of MAT in P. rhodozyma and taking into account available information for other species in the Agaricomycotina. With this model as starting point, we predicted that it would be possible to generate artificial heterothallic “mating types” in P. rhodozyma, by constructing strains containing one P/R cluster and one HD gene complementary to those present in the other “mating type”. This was confirmed in successful heterothallic crosses between the ste3-1Δ mfa1Δ hd1Δ and ste3-2Δ mfa2Δ hd2Δ. The possibility of outcrossing in P. rhodozyma had been previously put forward based on the observation of fusion between independent cells and on genetic evidence, including strains that seem to be hybrids between the known P. rhodozyma populations [35,36]. However, in line with previous observations [35], our results indicate that outcrossing is probably infrequent, since sporulation is greatly diminished when the possibility of selfing is genetically prevented (S2 Table). The triple mutant crosses open a new avenue to facilitate improvement of P. rhodozyma strains for astaxanthin production, for example by combining industrially attractive features like growth at higher temperatures with genetic alterations that improve astaxanthin production.
We also gathered evidence that P. rhodozyma probably undergoes meiosis during its life cycle since deletion of a core meiosis gene shown to be required for meiosis in C. neoformans [58] the model organism most closely related to P. rhodozyma, possibly affects sporulation efficiency and significantly decreased spore viability. We deem it therefore unlikely that in P. rhodozyma ploidy changes are achieved by distinct mechanisms as postulated to occur in the parasexual cycles of other fungi, like the yeast C. albicans [68].
P. rhodozyma belongs to a lineage that corresponds to the order Cystofilobasidiales, comprising species consisting entirely of homothallic individuals and others that are heterothallic [69]. Homothallism has been suggested to be an evolutionary dead end so that transitions are proposed to be unidirectional, from heterothallic to homothallic [70,71]. The genetic makeup of the P. rhodozyma mating system presently described is in line with this hypothesis, since it suggests a transition from a (tetrapolar) heterothallic ancestor to homothallism. It will be very interesting to find out whether transitions in the entire clade of the Cystofilobasidiales also conform to this unidirectionality. To investigate that and other aspects of the biology of sexual reproduction in this group of organisms, we are currently using comparative genomics to illuminate the molecular underpinnings of the various mating systems in the Cystofilobasidiales. Evolution of HD compatible gene pairs and evolutionary mechanisms associated with transitions between homothallism and heterothallism will be addressed.
Escherichia coli DH5α (S6 Table) strain was used for all cloning steps and was grown in LB medium (10 g/L Tryptone, 5 g/L Yeast Extract and 5 g/L NaCl) with 100 μg/ml ampicillin at 37°C. Wild type Phaffia rhodozyma strain CBS 6938 was grown in YPD medium (10 g/L Yeast Extract, 20 g/L Peptone and 20 g/L glucose) at 20°C, while mutants strains were grown in YPD medium supplemented with appropriate antifungal drugs (50 μg/ml geneticin, 50 μg/ml hygromycin or 100 μg/ml zeocin). All wild type P. rhodozyma and mutant strains used and generated in this work are listed in Table 1.
Different gene deletion fragments (GDF) were generated in order to construct P. rhodozyma deletion mutants, using a common strategy, consisting in cloning the upstream and downstream flanking regions of the selected gene upstream and downstream of an antifungal resistance cassette, to promote integration of the GDF in the targeted genomic locus [34]. Standard molecular biology methods were employed [72,73] and three distinct plasmids were used as backbone, namely pPR2TN [74], pBS-HYG [75] and pJET1.2+ZEO (S7 and S8 Tables), encoding geneticin, hygromycin and zeocin resistance genes, respectively (S6 Fig). Specific primers for the STE3-1, HD1, HD2, SPO11, regions STE3-1/MFA1 and STE3-2/MFA2 flanking regions were designed to include restriction sites that allowed cloning of the amplified regions onto the chosen plasmids (S7 Table). The GDF for the STE3-2 gene was constructed by overlap extension PCR and then cloned into the pJET1.2 vector using CloneJET PCR Cloning Kit (Thermo Scientific) (S8 Table). Nested primers were used to amplify the complete GDFs by PCR using Phusion High-Fidelity DNA Polymerase (Thermo Scientific) (S7 and S8 Tables). PCR products were purified using GeneJET Gel Extraction Kit (Thermo Scientific) or Illustra GFX PCR DNA and Gel Band Purification Kit (GE Healthcare) and finally used to transform P. rhodozyma. Complementation of selected deletion mutants was accomplished using plasmid pUC18+rDNA+ZEO that was constructed by inserting the constructed 1.8 Kb zeocin resistance cassette (S9 Table) and a 3 Kb fragment of the rDNA from plasmid pPR2TN in plasmid pUC18 at the SmaI and SacI restriction sites respectively. DNA fragments containing the ORFs’ of the genes pertaining to the complementation with approximately 300 bp flanking regions, were subsequently cloned into the PstI and BamHI restriction sites of pUC18+rDNA+ZEO. Each of the four complementation plasmids were then linearized with ClaI within the rDNA sequence to promote integration of the plasmid into the ribosomal DNA of the mutant strains to be complemented (S10 Table) [76].
Linearized plasmids and GDF were used to transform P. rhodozyma by electroporation as previously described (S7–S10 Tables) [76]. Transformants were selected in YPD medium with the appropriate antifungal drugs. Mutants with multiple deletions were obtained by transforming a confirmed deletion mutant with a second or third GDF. Correct integration of the disruption cassettes was verified by PCR (S7–S10 Tables) as previously described [34]. Briefly, a primer inside the resistance cassette and a primer outside of the flanking region present in the GDF at both the 5’ and 3’ extremities were used in diagnostic PCR reactions to identify the desired mutants. Absence of the gene targeted for deletion was also assessed by PCR for each mutant. Key deletion mutants were also confirmed by Southern blot in order to ensure that integration of the GDF occurred only once and in the correct locus. For Southern blot, 5μg of genomic DNA was digested with ClaI and run in a 0.8% agarose gel. Southern blot was performed using standard methods. Primers MP091-MP092 and MP062-MP063 were employed to amplify fragments of the resistance genes present in the geneticin and hygromycin cassettes to be used as probes. Labeling of the probes was performed with (α-32P) dATP using the Prime-a-Gene Labelling system (Promega). Signals were detected on X-ray films (Hyperfilm MP, GE Heathcare Life Sciences) (S7 Fig).
In order to test the ability of the deletion mutants to sporulate, the various mutants were inoculated on DWR (2.5% agar and 0.5% ribitol) solid medium, incubated at 18°C and observed regularly for up to two months. Sporulation efficiency assays were performed as described previously by Kucsera [36]. Briefly, cells were grown on YPD medium overnight (180 rpms, 20°C, in 10% of the volume of the flask), collected by centrifugation and washed thoroughly with sterile distilled water to remove culture medium. Cells were subsequently distributed in 10 μl drops over the surface of DWR plates that were subsequently incubated at 18°C for 10 days. The number of basidia on each plate was determined by direct observation of the perimeter of the colonies using an optical microscope. Three independent assays were performed with CBS 6938 wild type (WT) strain and with all deletion mutants (in triplicate). Student’s t-test was performed (with a significance of 99%) to ascertain the statistical significance of the differences observed between the WT and each of the sporulating deletion mutants. Additional assays were conducted to determine the viability of F1 progeny of the spo11Δ mutant and wild type strain CBS 6938. Basidiospores were recovered by micromanipulation and were transferred to YPD solid medium to determine viability. Chi-square statistic was performed to verify if the difference between WT and spo11 deletion mutant was statistically significant.
Mutant strains to be crossed were cultivated and washed as previously described, were mixed 1:1 and were finally distributed in 10 μl drops on the surface of DWR plates. Additionally, direct mixture of strains on DWR plates was also performed. Plates were incubated at 18°C and observed daily.
In order to ascertain the possible interaction between HD1 and HD2 protein bacterial two-hybrid assays were performed using the Bacterial adenylate cyclase two-hybrid system kit from Euromedex. The cDNA of each of the HD genes (S1 File) was synthesized at Eurofins and delivered as an insert in vector (pEX-K4). Synthetic HD genes were amplified by PCR (S11 Table) and sub-cloned into plasmids pKNT25 and pUT18 using the Hind III and Pst I restriction sites present in the MCS of both of plasmids, according to standard molecular biology techniques and the Euromedex manual [52]. Genes were cloned in frame at the N-terminus end of the T25 and T18 peptides. After transformation, integrity of the constructs was assessed by PCR amplification and sequencing of the cloned fragments (primers MP191/MP192 and MP193/MP194) (S11 Table). Different combinations of the recombinant plasmids (S4C Fig) were co-transformed into BTH101 E. coli cells (S6 Table). After successful transformation, the phenotype of 8 clones (identified as 1.1–10.8) of each of the different co-transformations was assessed in X-gal and MacConkey/maltose media upon incubation at 30°C for 48h. Results were scored after 24h and 48h incubation times. The presence of both fragments and correct plasmids in each of the clones was assessed by PCR (S11 Table).
Matchmaker Gold Yeast Two-Hybrid System from Clontech was also used to assess a possible interaction between the homeodomain proteins of P. rhodozyma. Synthetic HD genes (S1 File) were cloned into pGBKT7 and pGADT7 plasmids by transforming each of the PCR fragments and the digested plasmid into the pertinent yeast strain (S4 Table and S5C Fig). All S. cerevisiae transformations were performed according to Yeastmaker Yeast Transformation System 2 User Manual from Clontech. Primers used to amplify the HD genes carried 40 bp 5’ tails homologous to the ends of the linearized plasmids to promote recombination. Plasmids pGBKT7 and pGADT7 were linearized with Pst I and Cla I, respectively. Two distinct versions of the HD genes were used, a shorter one comprising the complete N-terminal and homeodomain region of each of the genes (the first 183 amino acids of HD1 protein and the first 196 amino acids of HD2 protein respectively) and another, comprising the complete proteins. Each of the haploid S. cerevisiae strains generated was tested for the ability of the fusion protein expressed to activate any of the reporter genes on their own (S5E Fig). Haploid S. cerevisiae strains (S5D Fig) were them mated and diploid strains selected as described in the Matchmaker Gold Yeast Two-Hybrid System user manual. Three diploid strains (named X, Y and Z), selected from each mating experiment were tested for their ability to activate the reporter genes. Plates were incubated at 30°C for 72h and photographed daily (S5 Fig).
Previously published pheromone receptor sequences were used to reconstruct the phylogenetic tree depicted in Fig 2. The accession numbers of sequences used are listed in S12 Table. Available draft genome sequences of Kwoniella mangrovensis CBS 10435, K. heveanensis BCC 8398 and Tremella fuciformis tr26 were searched for the presence of pheromone receptor homologues by TBLASTN using P. rhodozyma Ste3-1 as query. Genomic regions corresponding to positive hits were retrieved (GenBank accession numbers ASQD01000019.1, ASQB01000005 and LBGW01000351, respectively) and protein sequences were deduced after removal of likely introns, either manually or using AUGUSTUS [77]. The final protein dataset was aligned using an iterative refinement method (L-INS-i) in MAFFT v.7.221 [78]. Poorly aligned regions were removed with trimAl v.1.2 [79] using the "gappyout" option. The resulting alignment containing 338 positions was analyzed in ProtTest v.3.2 using the corrected Akaike information criterion (AICc) to determine the model of sequence evolution that best fitted our data. A maximum likelihood-based phylogenetic tree was built in RAxML v.8.1.24 using PROTGAMMAILGF model of amino acid substitutions and branch support was determined using 1000 rapid bootstraps. The Saccharomyces cerevisiae Ste3 pheromone receptor was used to root the tree.
MAT genes STE3-1, STE3-2, HD1 and HD2 were partially amplified and sequenced using primers listed on S7 and S8 Tables. All sequences obtained were deposited in Genbank and are listed in S12 Table. Nucleotide unrooted maximum likelihood phylogenies were inferred with General Time Reversible model and 1000 bootstrap replications on MEGA5.1 software [80]. The trees with the highest log likelihood are shown with branch lengths measured in the number of substitutions per site.
|
10.1371/journal.pcbi.1006224 | Mechanical evolution of DNA double-strand breaks in the nucleosome | Double strand breaks (DSB) in the DNA backbone are the most lethal type of defect induced in the cell nucleus by chemical and radiation treatments of cancer. However, little is known about the outcomes of damage in nucleosomal DNA, and on its effects on damage repair. We performed microsecond-long molecular dynamics computer simulations of nucleosomes including a DSB at various sites, to characterize the early stages of the evolution of this DNA lesion. The damaged structures are studied by the essential dynamics of DNA and histones, and compared to the intact nucleosome, thus exposing key features of the interactions. All DSB configurations tend to remain compact, with only the terminal bases interacting with histone proteins. Umbrella sampling calculations show that broken DNA ends at the DSB must overcome a free-energy barrier to detach from the nucleosome core. Finally, by calculating the covariant mechanical stress, we demonstrate that the coupled bending and torsional stress can force the DSB free ends to open up straight, thus making it accessible to damage signalling proteins.
| Cancer therapy involves generation of damage to DNA by various means, ranging from chemicals to ionizing radiation. The most lethal form of damage is the breaking of the DNA molecule, especially effective when the rupture occurs simultaneously on both sides of the DNA double helix, in what is called a double-strand break (DSB). However, the fracture of the DNA backbone cannot be complete until the whole chemical bonds holding together the two strands are cleaved. This can occur even much later than the action of the chemical or radiation damage. By using sophisticated computer simulations at the molecular scale, we studied the evolution of the initial breaking in the nucleosome—the constitutive building block of chromatin and chromosomes. We show that the initial breaking event follows a complex path, before eventually arriving at the complete fracture of DNA; the energy barrier for separating the DNA from the protein core of the nucleosome are not very high, in the range of a few kBT. The internal mechanical stress is a key parameter, often underappreciated, in determining whether the DNA in the nucleosome will open up, and thus become accessible to the action of nuclear proteins, or it will rather remain deeply hidden in the chromatin and difficult to repair.
| Double-strand breaks (DSB) in the double helix of the DNA molecule are defined as the cleavage of the phosphate-sugar backbone on both sides, the two cuts being comprised within 10 base pairs (bp) at most. Such an occurrence is only one among the many types of DNA lesions that a cell suffers at any time [1–3], aside of single-strand breaks (SSB), base loss (AP site, removal of one purine or pyrimidine), base cross-linking or dimerization, various oxidative defects by reactive oxygen radicals. However, although being much less probable than most other types of lesions, and with relatively fast repair kinetics [4], DSBs stand out as the most critical lesions to the DNA, since they ultimately lead to chromosome breakage and genome instability, cell mutation, or apoptosis [5]. Moreover, several simpler lesions of the type mentioned above, can evolve into SSBs and DSBs upon further chemical processing, both during the repair process, and because of the interaction with other major nuclear proteins.
Because of their cytotoxic effectiveness, inducing DSBs in the DNA of malignant cells is one of the major objectives of chemo- and radiotherapy of cancer. Many powerful antitumor antibiotics, such as the enediyne C-1027, abstract hydrogen atoms at several C′ sites in the backbone ribose, initiating an oxidative chain that leads firstly to SSBs (mainly at adenylate and thymidylate residues), and then to DSBs, typically cleaved at a distance of 1-2 bp [6]. High-energy radiation creates swarms of ionization products, both directly on the DNA structure and, most importantly, on the surrounding water molecules [7]. The free radicals produced in the process can attack the backbone and induce many different lesions, often clustered over short distances. Notably, DSBs are produced by ionizing radiation with a relatively high probability, and the terminations at 5′ and 3′ strand ends are typically more complex than for DSBs produced by enzymatic cutting, making their repair more complex and error prone [4, 8].
The DNA of eukaryotic genomes is packaged into arrays of nucleosomes, which appear as cylindroid particles that make up the chromatin structure. About three quarters of the total nuclear DNA are included in the nucleosomes, with the remaining DNA acting as “linker”, in a sort of beads-on-a-string assembly. The nucleosome core particle consists of 147 bp of DNA tightly wrapped around a histone protein octamer containing two copies of four histone proteins (H2A, H2B, H3, and H4). The lysine-rich N-terminal tails of the histones extend from the protein core, making various contacts with the DNA minor groove (see Fig 1 below).
How the complex structural environment of chromatin is altered in the presence of DNA lesions is a longstanding question in the study of the cellular response to DNA damage [9, 10]. Incorporating a lesion within a nucleosome core particle introduces several features not present in naked (linker) DNA, which affect its reactivity. Firstly, wrapping the DNA around the octameric core introduces mechanical heterogeneity into the duplex, resulting in regions that are bent and/or in which base stacking is deformed. Secondly, the large number of lysine (Lys) and arginine (Arg) residues present in histone proteins (more than 22% on average, Lys being especially abundant in H2B, where it represents 16% of the total) can directly interact with the lesion; in particular, Lys side chains are directly involved in AP cleavage within nucleosomes, via Schiff base formation [11]. Furthermore, the whole nucleosome structure becomes less compact: histones at DSBs are susceptible to extraction in low salt [12], implying a weaker interaction between DNA and histones at DSBs; and biophysical studies demonstrate that presence of DSBs lead to a localized chromatin expansion [9].
Computer simulations are increasingly demonstrating their utility, by allowing deeper analysis of experimental data, also in conditions where experiments are difficult to carry out. Molecular dynamics (MD) simulations of nucleic acids are today capable of following the system evolution over length and time scales approaching the real experimental set up [13, 14]. We recently completed a first MD study of the mechanical evolution of SSBs and DSBs in random-sequence DNA oligomers taken as representative of the exposed “linker” DNA between two nucleosomes in the chromatin fiber [15]. We studied the mechanical response under tensile force of SSBs and DSBs with different spacing between the two strand cuts (or “nicks”). The results indicated that the absolute values of force necessary to break up a DSB-damaged, free DNA fragment can be very large, of the order of 100 pN, at elongations of ∼20%. Such values of longitudinal stress and strain are unlikely to be observed in the normal dynamics of chromatin, nor during chromosome mitosis. Most importantly, however, that study demonstrated that thermal fluctuations are unable to provide the energy necessary to overcome the barrier to rupture, unless the two DSB cuts are separated by 2-3 base pairs at maximum.
In the present work we turn to investigating the mechanical evolution of DSBs in a nucleosome immediately after the backbone breaking event, by using very-large-scale MD simulations in the microsecond time scale. All-atom MD simulations of the nucleosome have started a few years ago, initially restricted to a 10-100 ns time scale [16, 17], and very recently extended to the microsecond time scale [18, 19]; these works provided already a substantial description of many special features of DNA wrapped in a nucleosome, such as the effects of added torsion and bending, histone-DNA contacts, and much other. However, the structure and dynamics of DNA defects of any kind are yet unexplored, in the much wider context of the nucleosome. Here, we search for specific mechanical signatures induced by the DSB, by simulating microsecond-long trajectories of the entire system, embedded in a large box of water and neutralized with point ions (Fig 1). We use different analysis methods to characterize the mechanistic aspects of DSB structural evolution, at various positions in the nucleosome. Firstly, we look at the long-wavelength thermal fluctuations of the system, extracting the essential dynamics from the covariance matrix of the atomic displacements around the DSB region. Secondly, we determine the lability of the broken-DNA adhesion to the histone octamer, by force-pulling with the “umbrella sampling” method. Finally, we characterize the role of mechanical bending and torsional stress, in determining the evolution of the broken DNA ends at longer times.
Altogether, these analyses allow to trace a mechanical path of the evolution of individual strand breaks into a fully-developed DSB, up to the final fracture event, as well as suggesting the most probable late-stage mechanical evolution of the damaged nucleosome. We conclude that even the weakest DSBs can be resistant up to milliseconds against spontaneous disassembly by thermal forces, thanks to the strong DNA-protein interaction within the nucleosome; mechanical forces of some importance are needed to open up the DNA structure at the break site, in a manner that appears to imply also the intervention of external agents; once the DNA structure starts to be opened, its fate depends on the amplitude of the displacement, the DNA being able to fold back to its original configuration, or to straighten out from the nucleosome core, for a large enough initial opening; also in this case, external forces must help internal stress relaxation, to bring the DNA ends to a sufficiently wide opening. These crucial findings should have profound implications for the early stages of DNA damage detection and repair, for example implying that damage marker proteins (such as Ku70/80, which interacts strongly with broken DNA ends), should also be capable of exploiting complex mechanical actions, for the damaged DNA to be accessible to the repair agents subsequently recruited in cascade.
We obtained the nucleosome molecular configuration from the RSCB Protein Database, entry 1kx5 [20]. This is an x-ray structure of the entire histone octamer with 147 DNA bp resolved at an average RMS of 1.94 Å, reconstituted from human nuclear extract expressed in E. coli; only 6 histone residues were unidentified in this experimental structure, with respect to the known histone sequences, therefore the model can be considered nearly complete. The 147 bp DNA is a palindromic sequence, chosen to maximize the degree of ordering and increase the x-ray spatial resolution. To obtain a model structure useful for our computer simulations, we removed all the crystallization water molecules and ions from the published structure, and added two DNA extensions of length 20 bp at each end of the nucleosomal DNA, with repeated sequence d(AGTC) [18]. DNA bases are numbered from 1 to 187 in each chain, one running clockwise and the other counter-clockwise, the dyad being located at basis 94 of each chain. This pristine nucleosome model without strand breaks is shown in Fig 1a, and will be labelled O in the foregoing.
DNA is wrapped left-handed about the histone core, making two nearly complete turns that join at the dyad symmetry point; the two DNA turns define two circles lying in two ideally parallel planes, with a superhelical symmetry axis perpendicular to the center of the circles (for a thorough discussion of nucleosome geometry and structure, see e.g. Ref. [21]). The relaxed DNA double helix makes a complete twist around its double-helical axis, about every 10.4 bp, defining a major and a minor groove; therefore, when turning around the histone core, the wrapped DNA makes 14 nearly full twists. Correspondingly, 14 contact points between DNA and proteins can be identified within the nucleosome structure, loosely situated at the minor groove locations facing inwards.
Based on these geometrical features, we defined 4 potentially interesting sites along this wrapped structure, where to place a DSB in a “mechanically significant” position, labelled 1 to 4 in Fig 1a. Correspondingly, we introduced a DSB at an inner contact site (model M1); at an outer non-contact site (M2); at the dyad (M3); and at the entry point of the nucleosome (M4). To create a DSB at each such locations, we introduced 5′-OH and 3′-phosphate terminations at each end of the break, respectively between: bases C69-T68⋯A120-G121 in M1; bases C73-A74⋯T114-T113 in M2; bases A94-T95⋯A94-T95 of both chains in M3; bases T22-A21⋯T167-G168 in M4. (The—symbol indicates the break site along each backbone, the ⋯ indicate the central interacting base pair.) In this way, the two backbone cuts of each DSB are spaced by 1 bp always comprising an A⋯T pair (Fig 1b), which remains initially bonded by only its two hydrogen bonds, plus the stacking interactions on each intact side of the chain, while the other half of stacking is readily reduced, as soon as the MD relaxation starts.
The CHARMM-27 force field database [22, 23] and its extension to treat nucleic acids [24, 25] were used for the molecular bonding and non-bonding force parameters. Strict comparisons between CHARMM-27 and AMBER force fields [26, 27] ensure that the results of long-time, finite-temperature MD trajectories of nucleic acid fragments with largely different conformations are consistent, and able to correctly reproduce the key structural quantities (bond angles, hydrogen-bond structure, base tilt, twist, shuffle, etc.) compared to experimental data. However, in all cases great care must be taken by performing sufficiently long preparatory annealing cycles of the water and ion background, while keeping the nucleosome still, to obtain the right water density and allow a realistic arrangement of the counter-ions around the phosphate backbone, prior to starting the microsecond production runs.
For the molecular dynamics (MD) simulations we used the GROMACS 5.1 computer code [28, 29]. Nucleosome models O and M1-M4 were solvated in water box of size 14.5 or 18×19×10 nm3 with periodic boundary conditions in the three directions, containing about 82,600 or 110,500 TIP3P water molecules, plus 480 Na+ and 250 Cl− ions to ensure neutralization of the phosphate backbone charge, and a physiological salt concentration around 0.15 M. All the MD simulations were carried out at the temperature of 310 K and pressure of 1 atm, or 350 K and 50 atm for the thermal stability study (at constant-{NVT}, hence the small overpressure within the typical numerical fluctuation for a system of this size). Because of the requirements of stress calculations (see below), we could not use standard Ewald-sum electrostatics but plain cut-off Coulomb forces. This is known to be at the origin of possible artifacts, therefore we adopted for both electrostatics and long-range non-bonding forces an unusually large cut off radius of 1.6 nm. We compared 20-ns segments of MD trajectories starting from the same initial configuration, without observing significant RMSD deviations (see S6 Fig). The DNA terminal ends (linker) were restrained by soft harmonic constraints, allowing a fluctuation of ±5 Å, to represent embedding in the chromatin structure. We used rigid bonds for the water molecules, which allowed to push the time step to 2 fs for the thermal equilibration runs, and to 1 fs for the force-pulling simulations. Typical preparatory constant-{NPT} MD runs lasted between 10 and 20 ns; force-pulling simulations were carried out for 10 ns, and the subsequent force-free relaxation lasted up to 400 ns; thermal stability simulations at constant-{NVT} extended to ∼1,000 ns for O and M2-M4, and up to 1,800 ns for the M1 model. Overall, the study used about 4.2 million hours of CPU time on 2048 IBM BlueGeneQ processors (IDRIS supercomputing center in Orsay), and about 800,000 hours on 896/1064 Broadwell Intel E5-2690 multi-core processors (CINES supercomputing center in Montpellier), with typical running times of 1.3 and 7 ns/hour on the IBM and Intel machine, respectively. About 1.5 Terabytes of raw data were accumulated over a period of 8 months, from March to October 2017, for subsequent post-processing (S1 Table).
Steered molecular dynamics (SMD) was performed on the fragments with the constant-force pull code available in GROMACS, only on the M1 model. In this case, we enlarged the water box to 18 nm in the x-direction, to allow possible outward extension of the broken DNA end, resulting in a system of 107,000 water molecules. Since the objective was to promote the detachment of one of the broken DSB ends from the nucleosome core, we applied a constant force parallel to the direction x and perpendicular to the superhelical axis, by means of a harmonic-spring fictitious potential attached to the C4′ and P atoms of the last two base pairs at one DSB end. After some tests, the spring constants were set at 100 and 75 kJ mol−1 nm−2, respectively for the two DNA strand ends farther and closer to the nucleosome surface. To provide a reaction force keeping the system in place, all the atoms of the H3 opposite to the DSB were retained by soft harmonic restraints, with a spring constant of 250 kJ mol−1 nm−2. Pulling speeds of 1 to 5 m/s were used for most SMD simulations. Forces and displacements were recorded at intervals of 5-10 time steps. Umbrella sampling was performed by extracting 100 configurations spaced by 50 ps during the first 5 ns of the force pulling simulation; force bias was progressively reduced from 100 down to 10 kJ mol−1 nm−2, to extract the zero-bias limit of the free-energy profile; the weighted-histogram analysis was used to interpolate and connect the data from discrete configurations.
We use the so-called covariant central-force decomposition scheme (CCFD, [30, 31]) for the intra- and intermolecular forces, which ensures conservation of linear and angular momentum of the molecular systems under very general conditions. The method is implemented in a special-purpose patch to GROMACS 4.6, which reads (all or part of) a MD trajectory for the selected subset of atoms for which stress is to be computed, and performs the entire analysis. Since the GROMACS-LS patch [30, 31] constrains the code to run in serial rather than in parallel, care must be taken to define properly the subset of interest in order to avoid prohibitive computing times. We prepared simple scripts to extract the principal components of the stress, compare stress fields from different simulations, and write the outputs in the portable Gaussian-cube format for visualization. Typically, the stress field is averaged over segments of 1 ns, with 100 frames spaced by 10 ps. This choice is a compromise between obtaining significant statistics while reducing the noise: in fact, averaging over a longer time window would progressively smear out the differences, while averaging with more frames separated by shorter interval would progressively increase the noise. Comparison between stress fields from different MD runs poses an extra care, since the structures need to share exactly the same box size and center, to avoid numerical artefacts from the cancellation between large positive and negative values. According to the CCFD scheme, stress fields are calculated by GROMACS-LS on a continuous grid superposed on the molecular structure; however, stress components and individual force contributions (pair, angle, dihedral, etc.) can also be projected back on the atom sites by defining a conventional (but non unique) atomic volume.
In our previous study on linker DNA fragments [15], it was found that DSBs can be very stable against thermal fluctuations, unless the two cuts on the backbone are very closely spaced. In particular, we obtained an average bond lifetime of the order of 50 ns at T = 350 K for the DSB with a single-bp A⋯T pair, and from these data we extrapolated room-temperature lifetimes of the order of hundreds of milliseconds for a DSB with 2-bp spacing, and up to several hours for a DSB with 3-bp spacing.
Based on such results, we decided to use the most favorable DSB configuration in the present study, in order to increase the probability of eventually observing DNA break up. Therefore, we introduced in all models M1-M4 one single-bp DSB with a central A⋯T, which is the weakest bonded bp. We ran the MD simulations at the temperature of T = 350 K, or about 77°C, in order to stimulate the thermal dynamics of the system, while remaining within a range of vibrational excitations that is still meaningful for the molecular force field used. MD trajectories were extended to ∼1 μs for the M2-M4 models, and up to 1.8 μs for the M1, which displayed some potentially more interesting dynamic features. The reference model O with the intact nucleosome was simulated over a shorter trajectory of 500 ns. Shorter MD trajectories were also run at T = 310 K for all models, for comparison.
We firstly present the results for the models M2-M4. For all three, we could not observe any substantial evolution of the DSB into a fully broken DNA, over the whole duration of the simulation, despite the relatively high temperature. While it cannot be excluded that such an event could be produced over longer times, this is an increase of more than a factor of 20 in lifetime compared to the free (linker) DNA [15]. Upon scaling by the same factor at 310 K, the corresponding dissociation time is in the 100-μs time scale or longer, even for the most favorable (i.e., least bound) DSB configuration; this represents therefore a lower bound for the spontaneous dissociation time. Representative snapshots from the trajectories at the DSB sites are shown in S1 Fig.
The bonding configuration of the central base pair remains on average rather close to that of the pristine nucleosome, with the H-bonds providing a large fraction of the cohesive energy, and the mildly deformed stacking ensuring a substantial structure stability. An example can be observed in Fig 2, in which the time evolution of the H-bond lengths for the central A⋯T bp of model M3 are shown. The three bonds formed by the N1 (adenine), O1 and O2 (thymine) donors are indicated in red, blue and black, respectively. The relative strength of individual H-bonds in the A⋯T bp can be theoretically estimated [32] to be about 10:4:1 for the N1:O1:O2. The last one is not usually accounted as a true H-bond, since it is very weak and with a length fluctuating around 2.8 Å. Indeed, the central N1 bond remains always in the range [1.8-2.1] Å RMS (note that the simulation is at high temperature); the side O1 is more dynamic than the corresponding bond in normal DNA, with an average length of 2.3 Å (∼2.05 in normal DNA), and quite large RMS fluctuations due to the larger rotational freedom of the DSB about the central axis; the O2 length remains well beyond the definition of H-bond, fluctuating about an average of 3.2 Å. Overall, these interactions provide enough bonding to keep the DSB in place, even in this M3-dyad position that is the farthest from the histone protein core, among all the DSB configurations studied.
To characterize the dynamic motion of the DNA and of the closest protein residues around the DSB region, we performed for each model a study of the essential dynamics [33, 34]. This method of analysis looks at a small subset of collective coordinates of the system, to extract the large-scale, anharmonic movements (bending, torsion, etc.) that dominate the global molecular dynamics.
We firstly perform the analysis for the regions surrounding each location M1-M4 in the pristine nucleosome, model O. Typically, the analysis is restricted to a length of about 7 DNA bp on each side of the DSB, plus the 15-20 histone residues in the closest neighborhood of the DSB. MD trajectories are sampled at a rate of 40 ps−1. Such analysis of the undamaged system provides a spectrum of eigenvalues, from which we extract the first few significant ones, and an average reference configuration for each M1-M4 site. Then, we repeat the same analysis on each of the independent trajectories including a DSB at the M1-M4 positions, by using as reference molecular structure the corresponding average from model O, so as to highlight deviations from the normal DNA dynamics.
A key quantity providing information about the large-scale (or “long-wavelength”) movements of the fragments implicated in the DSB comes from the study of the first few eigenvectors, and of their root-mean-squared fluctuation (RMSF) on a atom-by-atom basis. (Note that, like the RMSD, the RMSF is in principle measured in Ångstroms; however, being obtained from the eigenvector analysis, these are not actual atomic displacements, but components of a theoretical displacement projected out according to a particular deformation eigenvalue. Therefore we indicate the units as arbitrary, although they are numerically coincident with Ångstroms.)
These new atomic variables capture the contribution of each group of atoms to the principal collective movements, as filtered out by the most important eigenvectors. For all the M1-M4 models, the first 4 eigenvectors are found to cover 65% of the weight, the 5-15 ones are responsible for another 20%, and all the remaining 3N-15 for the last ∼15%. Such a distribution is less extreme for the O model, in which large-scale movements are quite more restricted, with the first 15 eigenvalues carrying about 55% of the total weight. The physical meaning of such principal eigenvectors can be appreciated, for example, with the plot of S3 Fig, in which the extreme configurations spanned by the large-scale motion of the first eigenvector, for the DNA fragments in models M1 and M3, are all simultaneously represented; the frames are colored from blue to red, the ordering showing how each atom’s motion spans between the extremes of the eigenvector. It can be seen that the principal eigenvector for M3 describes quite homogeneous, local fluctuations of all DNA bases, with just a more evident oscillation along the stacking direction concentrated about the DSB; on the contrary, for the M1 this principal eigenvector describes a dramatic large-scale displacement of the central atoms making up the DSB, which tend to span ample areas across orthogonal planes, by turning about the backbone. This largely different behavior between M1 and the other models M2-M4 is discussed further in the following.
In Fig 3 we plot the RMSF for the first 4 eigenvectors of each DSB model; each plot compares the RMSF for the fragment of 7+7 bp of DNA enclosing the DSB on either side (black lines), with the corresponding RMSF of the same fragment intact (red lines). For the M2-M4 models, it can be clearly seen that the RMSF of the DSB fragments is comparable to that of the same fragment in the reference model O; despite local quantitative variations, also of some importance between the various DNA bases, the black and red traces remain always close to each other, for each eigenvector, within a range of 0.1 in the arbitrary units of the RMSF. Moreover, the regions of the DSB and the base-pairs immediately adjacent (indicated by grey shaded areas) do not seem to display a peculiar or specific behavior, compared to the bp more distant from the DSB locations. Only the 1st and 3rd eigenvectors of M4 are somewhat outstanding compared to all the others, since they display an even distribution of displacements among all the bp. As it can be seen in the detailed eigenvector plots in S4 Fig, this coordinated motion correspond to an ample twisting about the main axis, which exists both for the O and M4 model, therefore independently on the presence of the DSB. It may look that the first eigenvector of M4 is more perturbed than the first of M1 (compare S3 Fig); however, the displacements are homogeneous throughout the configuration for M4, whereas for M1 the large motion is concentrated around the 2-3 bp that make up the DSB, which “suck up” the entire eigenvector. Such a difference underscores once more that the displacements defined by the eigenvectors are not true atomic displacements, but relative weights of the total displacement.
This same analysis for the RMSF of the groups of about 16-18 histone residues closer to the DSB in each model, is shown in Fig 4. Also in this case, for the M2-M4 models it is hard to see a qualitative difference between the data for the intact fragments (red lines), and for the fragments with the DSB inserted (black lines). The lysine and arginine residues are overall more mobile than the others, as far as the 4 principal eigenvectors are concerned, describing a dynamic interaction with the DNA. However, with minor variations, this behavior is the same also in the absence of the DSB, therefore it reflects the usual affinity of such residues for the DNA bases. The M1 model, instead, is definitely different, as it was the case for the DNA analysis in Fig 3 above, and it will be treated later in this Section.
The Schlitter entropy formula [35] can be used to estimate an upper limit for the contribution to the free energy from the excess entropy due the presence of the DSB, as:
T Δ S D S B = T ( ⟨ S M X ⟩ - ⟨ S O ⟩ ) (1)
with MX = M1, …M4, and 〈…〉 indicating the time average of the Schlitter entropy for each molecular fragment:
S = 1 2 k B ln { det [ I + k B T e 2 ℏ 2 MC ] } (2)
with C the covariance matrix of the atomic displacements, I the identity matrix and M the mass matrix, having respectively 1 and the atom masses on their diagonals, and 0 elsewhere. Table 1 reports the values for each DSB model, divided into DNA and histone contribution.
The absolute DNA entropy SO from Eq 2 fluctuates about 18±0.5 kcal/mol/K for each bp, very homogeneously all along the most part of nucleosome, but increasing to 20 kcal/mol/K in the few terminal bps attaching to the straight segments. If the values of excess entropy of DNA are distributed to the 4 bases (green and red in Fig 1b) comprising the DSB, these correspond to an excess of 35 to 60% for the M2-M4 models, the excess per base being larger in the M4, in agreement with the somewhat larger mobility demonstrated in Fig 3. On the other hand, the excess entropy for the histone residues selected for this analysis remains relatively small, for the three models M2-M4. Despite some difference in the total masses of the groups selected, even when expressed per unit mass instead of per-moles, the absolute entropy of the histones remains comparable, between the model O and the models including the DSB. This is a further confirmation of the relatively minor role played by histone dynamics in the M2-M4 models.
We now turn to describing the behavior of the DSB in the M1 model. Contrary to our expectations, this location in which the DSB is constrained between the histone core and the mobile H2B tail, and close to a DNA-protein contact, was the one to display the most interesting dynamics. The most evident change in the immediate environment of the DSB is the modification of the H2B tail, which can fold into very different interacting positions, starting from the outward extended conformation of the experimental crystallographic structure.
This behavior is shown in Fig 5, where the arrangement of the H2B tail is represented for three configurations, averaged over the respective MD trajectories: the reference O model (yellow), the M1 model at T = 310 K (cyan), and the M1 model at T = 350 K (blue). The low-temperature average configuration of the H2B tail resembles well that of the O model, with the terminal wrapping the minor groove of the DNA strand on the left of the DSB (in the figure); the high-temperature average configuration, instead, has the H2B tail flipped down by about 180 degrees (occurring very early in the trajectory, and irreversible over the whole 1.8 μs), with the fold of Lys24-25 and Arg26 keeping close contact with the DSB (see the black arrow). That such a configuration may be dynamically sampled over ∼1 μs time by only a 40 K temperature difference, means that the corresponding energy barrier (chemical plus deformation) must be relatively small.
In this M1 model, the DSB is constantly enclosed between the two β-sheets of H3 and H4, which fluctuate about their equilibrium structure and interact with one side of the DSB, while the H2B tail experiences strong oscillations, coupling with the cut bases of the opposite DSB side. The time evolution of the four bases comprising the DSB (green-red colored in Fig 5) gives a qualitative appraisal of this strong interaction (S2 Fig). Notably, the interacting portions of both the two β-sheets, and the H2B tail, include more than 60% of lysine and arginine residues, as expected given the strong electrostatic affinity of such amino acids for DNA (notably for G and T, [36]). The DNA ends at the DSB are clearly perturbed by such interactions, and it can no longer be said that the two broken backbones preserve a geometrical continuity, as it was instead observed for the M2-M4 models for the entire duration of the respective MD trajectories.
By looking at the RMS fluctuation of the eigenvalues for the M1 model in Fig 3, it can be seen that in this case the group of DNA bases adjacent to the strand breaks take up the majority of the weight, indicative of their participation in the ample fluctuations of the open DSB ends. It is important to note that the types of motion described by the first few eigenvectors are definitely different between the DNA with, or without the DSB (black vs. red plot). In S7 Fig we show the projection plot on the first two eigenvectors (EIG1,EIG2) of the MD trajectories with (M1) and without (O) the DSB, from which it clearly appears that the two types of motion have nearly zero superposition. As it is very evident in S3 Fig, the motion corresponding to EIG1 in M1 is essentially carried by the few base pairs making up the DSB: this type of motion (eigenvector) does not exist in the intact fragment. Even if the eigenvalues are not the same for the red and black plots, the component analysis in Fig 3 shows that, similarly to all models, the principal motions are evenly shared by all atoms in the absence of DSB, whereas the high-energy dynamics becomes fully localized around the structural defect when the DSB is present. Also for the histone eigenvalues, shown in Fig 4, it is readily apparent a more dramatic dynamics, especially carried by the lysine and arginine residues. Finally, the values of excess entropies from Table 1 provide a further confirmation of the peculiar large-scale dynamics of this DSB configuration. Because of these indications, we continued the MD trajectory up to 1.8 μs, but never observed a true mechanical destabilization of the DNA structure: the two DSB ends remain firmly in place, even if the two terminal bp on each end fluctuate quite wildly (see again S3 Fig, and the motion indicated by blue arrows in S2 Fig), while promoting a strong interaction with the protein surfaces.
As shown in the preceding Section, spontaneous dissociation of one or both DSB ends of a broken DNA from the nucleosome remains a difficult event, never observed in our simulations. DSB opening, and DNA detachment from the nucleosome are likely governed by a free energy barrier of adhesion, which even such a critical defect as a fully-cut DNA could not easily overcome simply by thermal fluctuations. The way to estimate the free-energy barrier in such a large and complex molecular system is to resort to controlled-force pulling, in order to impose the detachment, and then to use the intermediate structures along the reaction coordinate as starting points for the “umbrella” sampling of the potential of mean force. From the latter, the free energy barrier along the chosen reaction coordinate can be extracted.
As briefly described above, we used as reaction coordinate ζ the separation distance between the moving DSB end and the histone core surface. This was measured by taking the center of the DNA axis, at the average position of the C4′ and P atoms of the last two bps, and projecting it on the closest histone surface atom, along the perpendicular to the superhelical axis.
Fig 6a shows the variation of ζ as a function of simulation time, at constant pulling force. It can be seen that the DNA broken end detaches from the histone surface in large steps (red segments), during which the internal energy builds up until some barrier is overcome; the final stage, indicated by the blue segment, is the complete detachment of the DSB end after t = 1.1 ns, in which the free end is simply drifting at the constant speed of about 2 m/s (later on dropping to 1 m/s).
During the final stage of the pulling simulation, the DNA is forcefully unwrapped from the histone core, as it can be seen in Fig 6b. Here we show the distance from the core surface of three P atoms facing the histones, belonging to the bp 71-118 (contact site close to the DSB), 78-111 (middle site) and 82-107 (next contact site). The first contact site is detached in the interval t = 1.-1.5 ns, as indicated by the black trace that follows the distance from the surface of of the P71 backbone phosphor. Then, under the continued pulling of the DSB end, also the P111 comes off, at t>3 ns (red trace); however, it may be noticed that this event is “cooperative”, the P82 (blue trace) following the instantaneous opening of P111 at t = 3.-3.4 ns, and then falling back into position, after which P111 is definitely “peeled off” the histone surface.
From this force-pulling simulation we can calculate the free energy profile of the barriers, which characterize the binding of the DNA end to the histone core surface. The potential of mean force (PMF, [37]) is a method to extract the free energy difference ΔA from a sequence of configurations, biased along a reaction coordinate that brings the system from a state a to a state b. In our case, the reaction coordinate is just the distance ζ defined above; the states a, b respectively represent the initial configuration at ζ = 0, with the DSB end still attached to the histone surface, and the final configuration with the end detached, at ζ ∼5 nm. The “umbrella sampling” technique [38] is used to obtain the PMF at discrete values of ζ, and the discrete values of A(ζ) between a and b are connected by the weighted-histogram method [39, 40]. We extracted 100 configurations from the force-pulling simulation, spaced by 50 ps in the first 5 ns of the trajectory (corresponding to about 0.5 Å spacing along the reaction coordinate ζ = 0 to ζ ∼5 nm); each configuration was equilibrated for 2 ns at 310 K under constant-{NVT}, while biased with a harmonic “umbrella” potential of variable strength, progressively reduced to zero to obtain the unbiased limit. The force probability distribution of the fluctuating DSB free-end at each value of ζ was reconstructed with the weighted-histogram analysis, and the free energy profile thereby extracted is shown in Fig 7.
Despite the noisy profile, a few features can be identified. The red circle defines the first barriers to the detachment of the DSB ends, corresponding to the red steps in Fig 6a; such barriers are quite small (<1 kcal/mol), and strongly depend on the choice of the point of application of the pulling force. The blue circle identifies the free-energy barrier for the detachment of the first contact at P71, about ΔA = 1.8±0.2 kcal/mol or 3 kBT; this does not represent a very large value, and should correspond to a ∼5% Boltzmann probability of spontaneous detachment at T = 310 K. It is worth noticing that this value for the detachment barrier fits very well with the experimental estimates of nucleosome unfolding energy, which obtain a value of about 27 kcal/mol [41, 42]: this corresponds to the detachment of all the 14 contact sites, from which it can be estimated an average energy of 1.9 kcal/mol per contact. The green circle roughly identifies the cooperative events leading to the detachment of P111, between 2 and 4 ns, with a sequence of ΔA again not larger than 2-3 kBT. Further detachment events were not observed, with the above values of pulling force; in particular, P82 remained in place for >500 ns, even at larger deformations of the DSB free end, because of the H3 histone tail acting as a sort of brace that maintains the DNA firmly in place about that position. Much larger forces, or cooperative events of histone tail fluctuation, likely involving other nuclear proteins, seem to be necessary to pull the free DSB end further beyond the limits observed in the present simulations.
In the last part of our study, we turn our attention to the internal relaxation dynamics of the nucleosome including a broken DNA. To demonstrate what it is meant by “internal relaxation”, we take two configurations along the final trajectory of the force pulling simulation of the M1 model described in Fig 6, C180 and C290, respectively extracted at times t = 1.8 and 2.9 ns, well beyond the detachment stage that ends at 1.1 ns in the figure. Each of these two configurations is used as initial structure for an MD simulation, and is then equilibrated and relaxed at 310 K and constant-{NVT}, without any external forces applied. The results of these two MD simulations are displayed in S5 Fig: starting from the two different initial conditions, after 40 ns the C180 tends to fold back into the initial M1 configuration, while C290 straightens out and increases its distance from the histone core. Notably, the C180 remains in a slightly open state, because of the free energy barrier to detachment that now has to be overcome in reverse. However, the important observation in both cases is that the folding back, or the straightening out, are driven entirely by the competition between the residual attraction between DNA and proteins (a “chemical” force), and the relaxation of internal constraints (mainly bending and torsion, therefore an “elastic” force).
The role of internal forces can be clearly understood by looking at the distribution of mechanical stress, which is a measure of the elastic energy accumulated by the bending and torsion of DNA while wrapping around the histones, and that is ready to be released if the structural constraints are softened, as it could be the case of a DSB cutting the DNA sequence. Recent developments led to alternative geometric derivations of the microscopic stress [31, 43], based on the invariance of the free energy with respect to surface deformations [44, 45], instead of the classical formulation based on invariance of momentum [46–51]). The so-called CCFD scheme [30, 31], incorporated in a special-purpose version of the GROMACS code, ensures conservation of both linear and angular momentum under a generic stress-induced transformation.
The mechanical stress σ(x) (a 3×3 tensor defined at any point x in space) is a meaningful way of representing the distribution of internal forces with respect to a given local direction vector. Once a DSB breaks the DNA backbone around the nucleosome, internal forces are going to be relaxed, and compete with the chemical (Van der Waals, electrostatic) forces from the interaction with the histone proteins. Looking back at S5 Fig for model M1, such a competition is very evident upon comparing the bottom configurations: in C185 the chemical forces overwhelm the internal stress, whereas in C290 the opposite holds, and the DNA ends up straightened out from the DSB site.
A tensor, such as the stress, can be meaningfully projected onto any direction vector, the choice of a particular projection being just a matter of convenience. In the present case, the “bent tube” structure of nucleosomal DNA makes it interesting to consider the stress projected onto its “tubular” surface. An intuitive way of looking at the mechanical stress as a “projected force” is through the surface traction vector, T(x) = σ(x) ⊗ n. The symbol ′⊗′ indicates the tensor product between the stress and the vector n, in practice the matrix product between the 3×3 matrix of the stress at each point x, and the 3-component vector locally perpendicular to the surface at x (see the local reference frame {n, τ, b} in Fig 1c). The traction vector T(x) contains a great deal of information on the state of internal tension, compression, and torsion, of a complex structure like the DNA in the nucleosome. The portion of DNA wrapped around the histone core is forced to bend into nearly two full circles of diameter about 8 nm, a size much shorter than the persistence length of free DNA, ξp ≃50 nm. Therefore, the DNA “tube” is here constrained in a geometry from which it should rather escape into a more straight structure, whenever possible, under the relaxation of internal forces. The state of tension and compression of a bent tube can be described by a particular projection of the traction vector, t(x) = T(x) ⋅ τ, in this case along the unit vector τ locally tangent to the continuous line sweeping the center of the tube.
Notably, a bent tube would experience a stretching force (a tensile, negative t(x)) in the half that lies outside the centerline with respect to the center of curvature, and a compressive force (a positive t) in the half lying inside the centerline, as shown in blue/orange in Fig 1c. The internal force should be zero along the centerline itself, because of this called the “neutral axis” (also the helical axis of DNA).
We computed the line tension t(x) all along the curved DNA pathlength, by averaging over slices of width 0.5 nm (see for example the white slice in Fig 1c), and by integrating separately over the inner and outer regions (orange and blue in the figure). Each slice averages all the atoms included in the white slice, centered at the midpoint between the two P atoms of each bp, therefore adjacent slices have some overlap to provide a smoother profile of the signal. In a perfectly smooth tube, one should see just two constant values of positive and negative tension, respectively in the blue and orange volumes. However, the DNA is not simply a smooth tube, but it has a complex geometry in which minor and major grooves alternate, and it contacts the histone surface in about 14, evenly spaced sites. At these points, there is an excess or a defect of tension/compression, as well as some amount of under/over twisting of the already twisted tube.
The twist stress is that part of the internal forces involved in the torsion about the central (neutral) axis of the tube. The DNA double helix is naturally twisted already in its normal B configuration; however, when it is bent in the nucleosome, the twist is necessarily modified with respect to the normal configuration. The twist component is obtained as well from the traction vector, as: w(x) = T(x) ⋅ (τ × r), where r is the vector from the neutral axis to the point x, parallel to the local surface normal n (see again Fig 1c). The vector product between τ and r defines a third vector threading like a spiral screw about the DNA tube; positive and negative values of w(x) indicate a rotational force (a torque) tending to over- or under-twist the DNA about its helical axis.
In Fig 8a and 8b we show for both configurations the tension profile t(x) along the helical axis of the DNA fragment right after the DSB (bp T68⋯A121). The portion mostly affected by the pulling under force and subsequently relaxed is comprised between the DSB and ∼bp T84⋯A105; we neglect the first few bp immediately next to the DSB, too disordered for such a calculation. Two sets of data are shown in each panel, at the beginning of the relaxation (black lines), and after 40 ns (red lines); stress values are averaged over 100 frames with 10 ps spacing, in either case (see S8 Fig for the error due to averaging procedure). In general, the terminal part of the DNA next to the DSB (indicated by a grey-shaded area in the panels) tends to lower values of both line tension and compression, for both configurations, compared to the rest of DNA beyond the dyad (bp A94⋯T94), indicative of the stress release at the free ends. The extra tension/compression from the DNA-histone contact points can be clearly observed in the alternating minima and maxima along the compression and tension sides of the DNA tube.
Despite some noise in the data, it can be appreciated that for the C185 configuration (Fig 8a) the red lines are at the same values than the black ones: this is a signature of the chemical residual attraction winning over the internal stress, thus tending to fold back the DSB open end into place. On the other hand the C290 (Fig 8b), starting from almost twice-higher stress values in the grey area compared to the C185, has red lines approaching a state of nearly zero stress after the relaxation time; also several sites beyond the dyad (outside the grey region) display sizeable variations of tension and compression. This suggests the release of internal stress as being responsible for straightening out the DSB end, into a mechanically less-constrained structure.
The extra twist stress (positive or negative) also contributes to the internal forces that are going to be relaxed, when the DSB cuts open the DNA, albeit to a much lesser extent, given the smaller absolute values of w compared to t. In Fig 8c and 8d the w(x) stress profiles are shown, under the same conditions of the two panels above for the line-tension/compression. It can be noticed that, also for the twist stress, generally smaller values (≤1 MPa in modulus) are seen in the DSB tail. However, the large numerical noise does not allow in this case to draw a more firm conclusion, concerning the (minor) role of twist stress in the chemical vs. mechanical force competition in the two configurations.
In the present work, we studied by very-large-scale molecular dynamics (MD) simulations the evolution under external force and temperature of double-strand breaks (DSB) in nucleosomal DNA. We collected and analyzed a large amount of raw data (more than 1.5 TBytes, and 5 million CPU hours on two large supercomputers), by running microsecond-long trajectories for 5 different, all-atom models of the experimental 1kx5 nucleosome structure [20]. The basic model is made up of the canonical 8 histones, plus a 187-bp DNA comprising the 147 bp wrapped around the histone core and 20-bp terminations on each end, and embedded in large boxes of about 80-100,000 water molecules with Na+ and Cl− ions at 0.15 M physiological concentration. The pristine nucleosome configuration (model O) was modified, by inserting a DSB at four different positions in the DNA (models M1-M4), and the stability of the resulting structures was compared with model O nucleosome.
A general observation from the μs-long trajectories, is that damaged DNA remains well attached to the nucleosome body, without qualitative differences compared to the intact DNA. Only the model M1, in which the DSB is tightly sandwiched between the histone H3 and the tail of histone H2B, displayed a dynamics substantially different from the corresponding region in model O, due to the increased interaction of the broken DSB ends with close-by histone residues; however, also this DSB configuration was stable over the entire observation time scale, which in this case was extended to 1.8 μs. In order to identify the free-energy barriers which maintain the broken DNA attached to the histone core, we carried out steered MD with a pulling force to “peel off” the free DSB end from the nucleosome; relatively small free-energy barriers of the order of 3 kBT were identified, which could allow spontaneous DSB end detachment at physiological temperatures, likely over longer time scales of hundreds of microseconds to milliseconds. Spontaneous unwrapping of DNA from the nucleosome core has been studied experimentally [52–54] because of its relevance in gene regulation and DNA transcription; notably, such experiments were carried out on isolated nucleosomes, with a length of DNA just matching, or barely longer than needed to wrap the histone core (147 to ∼180 bp). In such conditions, spontaneous detachment of the ends was indeed observed over the timescale of hundreds of milliseconds; simulations by coarse-grained MD methods roughly confirm such trends [55–57], despite being strongly dependent on the empirical parametrization of each different force model. To such experiments it may be objected that the nucleosome constrained in the chromatin could have a rather different mechanics: our molecular stress calculations demonstrate that the circularly bent DNA has a strong internal driving force, from the relaxation of line tension and, to a lesser extent, of twist (torsional) stress. The reason may be found in the persistence length of the free DNA, which is much longer (∼50 nm) than the average radius of curvature in the nucleosome (∼8 nm), and pushes the DNA to regain the straight average conformation on that length scale; the fact that spontaneous fluctuations were observed [53] both in presence of, and without binding proteins seems to support this view. In fact, our μs-long MD simulations were carried out with a soft restraining of the DNA linker (20 bp on each end), to simulate the effect of the background chromatin structure, and no fluctuations larger than thermal vibrations were detected for the terminal phosphors; on the other hand, the DSB free end, once extended beyond a distance of about 2.5 nm away from the histone core, tended to regain a straight conformation and detach completely, confirming the importance of stress relaxation as a main driving force in DNA unwrapping.
In conclusion, the results and potential implications of this study can be summarized by the following findings:
|
10.1371/journal.pcbi.1003885 | Glycolysis Is Governed by Growth Regime and Simple Enzyme Regulation in Adherent MDCK Cells | Due to its vital importance in the supply of cellular pathways with energy and precursors, glycolysis has been studied for several decades regarding its capacity and regulation. For a systems-level understanding of the Madin-Darby canine kidney (MDCK) cell metabolism, we couple a segregated cell growth model published earlier with a structured model of glycolysis, which is based on relatively simple kinetics for enzymatic reactions of glycolysis, to explain the pathway dynamics under various cultivation conditions. The structured model takes into account in vitro enzyme activities, and links glycolysis with pentose phosphate pathway and glycogenesis. Using a single parameterization, metabolite pool dynamics during cell cultivation, glucose limitation and glucose pulse experiments can be consistently reproduced by considering the cultivation history of the cells. Growth phase-dependent glucose uptake together with cell-specific volume changes generate high intracellular metabolite pools and flux rates to satisfy the cellular demand during growth. Under glucose limitation, the coordinated control of glycolytic enzymes re-adjusts the glycolytic flux to prevent the depletion of glycolytic intermediates. Finally, the model's predictive power supports the design of more efficient bioprocesses.
| Glycolysis generates biomass precursors and energy from sugars and is therefore a key element in the metabolism of mammalian cells. Changes in its activity greatly affect cellular function which is often recognized as metabolic disease but also as opportunity for the design of efficient bioprocesses. Metabolic research discovered that continuously growing mammalian cells often exhibit a high glycolytic activity but also delivered seemingly endless facets in the pathway operation. The latter call for a systems-level understanding regarding capacity and regulation for a broad range of cultivation conditions. In this work, we couple a cell growth model to a simple kinetic description of glycolysis to consistently explain intracellular metabolite pool dynamics of the Madin-Darby canine kidney cell line over a variety of experiments and time scales while considering the growth status and cultivation history of the cells. We argue that the many different dynamics in glycolysis result from an interplay between a growth-dependent sugar uptake together with simple intrinsic enzyme regulation.
| The primary metabolism of cells is essential for cell growth and maintenance. Glycolysis is a central element of the primary metabolic activity and supplies anabolic pathways with precursors and cellular energy in form of ATP. The detailed in vitro characterization of glycolytic enzymes, such as hexokinase (HK), phosphofructokinase (PFK) and pyruvate kinase (PK), with respect to their catalytic properties in the presence of substrates, products and allosteric effectors represents an initial step towards a kinetic description of metabolic phenomena of cells [1]–[3]. Dynamic mathematical models of glycolysis have been developed for many different organisms such as Escherichia coli, yeast, or mammalian cells. Such models range from simple to full kinetic descriptions with the intention to study specific observations, e.g., metabolic steady states [4]–[6] perturbation of substrates [7]–[9] or enzymes [10], flux sensors [11], oscillations in glycolysis [12], the glucose uptake system [13], or the link of liver cell glycolysis with blood glucose homeostasis [14], [15]. Although in many cases the existing experimental data sets do not allow for a full validation of highly complex models in a broad physiological context, there is a clear benefit regarding the integration of complex regulatory mechanisms, which helps to explain general phenomenological aspects that are typically found in the respective organism. However, an apparently complex metabolic behavior must not result from complex regulatory mechanisms [16]. In case of glycolysis, it seems that few regulatory mechanisms dominate the dynamics of intracellular metabolite pools and readily explain salient features of experimental observations [17]. Furthermore, with an increasing number of powerful assays, e.g. to determine intracellular metabolite concentrations or to measure enzyme activities in yeast and animal cells (e.g. [18]–[21]), changes in glycolytic activity for cell growth or substrate perturbations can be monitored at an unprecedented level. Based on the additional quantification of extracellular metabolite changes and cell number measurements a systematic analysis of basic dynamics of glycolysis for various cultivation conditions is possible.
Recently, we reported that adherent MDCK cells cultivated in two different media not only show similar and reproducible dynamics of many intracellular metabolite pools but also that changes in their concentrations are growth phase-dependent [22]. With the aim to elucidate the interplay between enzyme and growth regime-mediated regulation of glycolysis, a segregated cell growth model has been developed, which captures experimental observations during cell growth phases regarding number increase, diameter change and uptake of substrates [23].
Here, we couple this segregated cell growth model to a structured model which incorporates a simple kinetic description of glycolysis and focusses on a few well-known enzymatic properties to elucidate the control of glycolysis. In addition, the linkage to the pentose phosphate pathway and the glycogenesis are taken into account. We evaluate the model's ability to reflect changes in intracellular metabolite pools for a variety of cultivation conditions using a single set of parameters. This includes the transition from the exponential to the stationary cell growth phase, the fast replacement of medium by PBS at different time points of cultivation, and a substrate pulse experiment. Afterwards we discuss the influence of growth regime, changes in extracellular metabolite concentrations, and activity of key enzymes on the control of glycolysis. In addition, aspects of hierarchical regulation are addressed which, taken together, improve our understanding of the metabolism of fast proliferating cells. Finally, options for the modulation of metabolic activity are evaluated regarding the design and optimization of cell culture processes as well as the study of metabolic diseases.
In three independent experiments, adherent MDCK cells were grown in 6-well plates with the serum-containing medium GMEM-Z, which provides sufficient amounts of extracellular substrates over the chosen cultivation time. Therefore, cell growth occurs with maximum rate until the available surface becomes limiting [23]. The experimental data of intracellular metabolite pools is taken from Rehberg et al. [22], and analyzed in the following using the model described in the Materials & Methods section (see section “Model and simulation”). The model focuses on intermediates that were measured experimentally and is composed of a concise set of enzyme kinetics with few regulatory mechanisms. A schematic overview of the considered enzyme reactions, the measured metabolite pools and maximum in vitro enzyme activities, and the coupling to the previously developed segregated model of cell growth [23] is given in Fig. 1.
The implemented G6PDH and UT mediated conversion of G6P are entry points into the PPP and the glycogenesis, respectively. They eventually fuel the pools of ribose 5-phosphate (R5P) and uridyl diphosphate glucose (UGLC) and implementation of simple degradation reactions (Eq. (11), (12)) allows assessing the consistency between the flux through G6PDH and the R5P pool as well as between the flux through UT and the UGLC pool. The expense of an additional model parameter for the ribose 1,5-bisphosphate phosphokinase (RDPK) and glycogen synthase (GLYS), which both represent only one of the possible degradation reactions, enables the model to reflect the dynamics of R5P and UGLC during cell cultivation (Fig. 5). Note that in contrast to other intracellular metabolites, UGLC is diluted by cell volume growth to a visible extent, which reduces the typical peak-like behavior compared to other metabolites (Fig. 5D–F). During the limitation experiment, the pool of R5P decreases later than suggested by the model yet with similar dynamics. During the pulse experiment, the level of R5P is lower than suggested by the model (Fig. 6A,B,C). In both cases, the differences between experimental data and simulation results might be due to network properties of the PPP, which are not considered by the model (for instance, the high number of reversible reactions, and the linkage of its intermediates to the biosynthesis machinery). The data for UGLC shows only a minor decrease and a minor increase during the limitation and pulse experiments, respectively, which is described by the model (Fig. 6D,E,F) and clearly attributed to the low pathway activity (Fig. 1).
We developed a kinetic description of glycolysis that, coupled to a segregated cell growth model, enabled describing and analyzing the experimental data of this study comprising roughly 600 data points by using a single set of parameters for the enzyme kinetics. To describe the dynamics of enzyme activities different types of kinetics with arbitrary complexity can be found in literature. Here, we focused on the establishment of a relatively simple model, which incorporates only basic regulatory mechanisms of glycolytic enzymes and a minimum of reactions. Nevertheless, the model reflects the basic dynamics of metabolite pools for a variety of experimental data sets and time scales. In the model, the kinetics of TATK as well as the ENO represent lumped reactions and were realized with reversible mass-action kinetics (see supporting information 3 for further details on enzyme kinetics). The enzymes HK, GPI, G6PDH, UT, and aldolase (ALD) as well as the GLUT were defined as Michaelis-Menten kinetics, as they provide an upper activity bound that was measured in vitro by Janke et al. [19] (except GLUT), and appear either as reversible or irreversible reaction. So similarly to mass action kinetics, only one or two parameters of the Michaelis-Menten kinetics required estimation. Only the PFK, which is a strongly regulated enzyme in glycolysis, as well as the PK were considered to be influenced by allosteric effectors. A Hill-Kinetic with four subunits [10], [27] was sufficient for the PFK to fit all data and takes a direct activation by F6P [17] and an indirect activation via fructose 2,6-bisphosphate (F26BP) into account [28], [29]. The PK is influenced by the well-known F16BP-mediated activation. The chosen simplifications in enzyme kinetics renders the used parameters to be more abstract, such that, for example, the affinity of an enzyme for its substrates or products rather represents a constant sum of influential factors such as availability of cofactors and concentration of ions. As a result, a comparatively simple model is obtained that describes the experimental data with enzyme kinetics comprising only 19 parameters. In addition, two experiment-specific parameters were determined for each cultivation, which yields a total of 21 degrees of freedom not considering the parameters used in the segregated cell growth model. In principle, however, any model of glycolysis that takes into account the metabolites and enzyme reactions used here (even though with higher complexity) may equally well describe the dynamics of the intracellular metabolite pools of this study. Nevertheless, our relatively simple model features the identification of mechanisms that are involved in certain dynamics and has the advantage of efficient parameter estimation and model analyses. Furthermore, extension by additional reaction mechanisms is relatively easy in case further experimental data is available or other cellular functions are of interest, e.g. the response of primary metabolism to osmotic stress [30], and hypoxia [31] or its influence on the glycosylation of proteins [32].
The derived kinetic description of glycolysis simultaneously integrates data of three independent cell cultivation experiments, two limitation experiments and one pulse experiment and therefore required coupling to a model that takes explicitly into account the progress of the cell through different growth phases during the cultivation experiments Cult1–3 [23]. Because of the many different experimental settings, simulations would normally require a large set of initial conditions that comprise not only starting concentrations of intracellular metabolites (8 degrees of freedom) but also cultivation conditions (the actual medium volume, glucose concentration), and the growth status of the cells (cell number, cell-specific volume, enzyme level and glucose uptake rate). Considering that the perturbation experiments were performed at a certain time point of cultivation and that cultivations in turn were inoculated with cells from a defined preculture introduces a dependency of the cell status on the cultivation history. Accordingly, we transfer information regarding the cell status, which comprises information of growth and metabolism, as well as culture conditions (Table 1) from one simulation to another (Fig. S4, further explained in the supporting information 5). Estimating a certain cell cultivation history not only eliminates the estimation of initial conditions for glycolysis and the growth status of the cell but also supports consistent data simulation and can be used to evaluate biological variations [33]. However, inconsistent data sets or an unknown cell status (e.g. cell status different to those of Cult1–3) may pose a serious challenge for model fitting. For such scenarios the individual selection of initial conditions might be a better option. In this work, however, the estimation of two experiment-specific parameters, which are the Elevel for the respective cultivation and t* as starting point for the perturbation experiments, as well as a consistent consideration of all data sets outweighed a perfect data fitting and greatly supported our systems-level analysis of glycolysis.
The simulation of the limitation experiments was started with initial conditions of cells (growth status and a metabolic status) that corresponded to a time point t* of the Cult1 experiment (Table 1). The selection of different time points t* readily explains variations in the initial concentration of intracellular metabolite pools that were found between the Lim1 and Lim2 experiment. The actual limitation was induced by reducing the medium volume to 310−7 L, which is estimated as liquid volume that remains on the cellular surface or in the intercellular space. In comparison, the volume of all cells is roughly 610−6 L. In principle, a dilution of the remaining medium with PBS can be realized by choosing lower GLCx concentrations and a higher medium volume (VM). The affinity of GLUT for GLCx () was found to have a large confidence interval and, hence, lower concentrations of GLCx under a higher VM are likewise possible (Table 2).
With the limitation of glycolysis in substrates, the feed-forward regulation of PFK and PK stops the metabolite pool degradation while the TATK reactions partially reverse and fuel glycolysis with 0.03 mmol L−1 min−1 leading to a new steady state within minutes. Thus, the control of the glycolytic activity shifts from the growth regime that regulates the GLUT activity (see section “Tuning the ATP and biomass precursor generation”) towards an inherent regulation of enzymes by substrates and products in the glycolytic pathway (see also supporting information 1). Without the implementation of the TATK reactions, the remaining glycolytic activity eventually depletes the metabolite pools unless fueled from sources other than GLC. As the limitation applies to all possible extracellular substrates, the use of intracellular carbon sources that might be related to the PPP, glycogenolysis or glyconeogenesis from pyruvate seems likely. The PPP shares already three metabolites with glycolysis (G6P, F6P, and glycerine-aldehyde phosphate linked to 3PG) which are not depleted during the limitation experiments and may thus pose the most promising and simplest option among the aforementioned intracellular carbon sources. Also, the late decrease in R5P during the limitation experiment and its lower level during the pulse experiment may support a scenario in which the PPP fuels glycolysis under limiting GLC levels and, thus, can have a large influence on glycolytic intermediates, which is similarly found for hepatoma cells [34]. In turn, after addition of fresh medium, the PPP metabolite pools may be replenished by glycolysis and we hypothesize a certain buffering capacity of the PPP as it is composed of many reversible reactions and intermediates that participate in the biosynthesis machinery. In the model, the implemented reversible mass action kinetics allow for such a switch from metabolite consumption to metabolite production by the PPP under the lack of alternative sources for glycolysis. However, the flux rates as well as the parameters of the PPP cannot be uniquely identified on the basis of our experimental data (Table 2). Therefore, we have used the additional constrain that the flux from the PPP into glycolysis is low (supporting information 2). Although the implemented mechanisms may not definitely be attributed to the PPP, all parameterizations of Table 2 support the finding that metabolite pools can be maintained (or increased) under limited substrate availability. To this end, the model suggests that the allosteric regulation of PFK and PK as well as the reversibility of GPI and TATK modulate the glycolytic activity in scenarios characterized by limited substrate availability. This is consistent with findings that flux control in glycolysis can rely on a combination of many enzymatic steps [34] and can vary depending on experimental conditions [35]. Counter-intuitively, adenosine-based nucleotides, which are also considered to control the metabolic activity in general [36], are constant during our limitation and pulse experiments (Fig. S5). Similar observations were made for yeast and HeLa cells [17], [37]. Therefore, regulation of glycolytic enzymes of MDCK cells by adenosine-based nucleotides seems unlikely under the conditions investigated, which is also hypothesized by Renner et al. [38] for rat hepatoma cells. Furthermore, an activation of glycolysis by a possibly decreasing ATP/ADP ratio stands in contrast to the metabolite pool preservation and renders its influence to be limited. However, the general purpose of an enzyme-mediated control of the glycolytic activity through PFK, PK, TATK and GPI might lie in the prevention of unnecessary dissipation of valuable biomass precursors and may also guarantee a metabolic status that enables a fast reactivation of glycolysis and other cellular functions when new substrates become available after starvation conditions (Fig. 4C,F,I,L,O).
Over the full course of cultivation cells pass through several growth-phases with varying cell-specific volumes and with glucose uptake rates that both strongly influence the metabolite pool dynamics (Fig. 2,3). In addition, abundance of enzymes, their covalent modifications as well as the level of allosteric regulators may change over time which can additionally affect metabolite fluxes and pools [39], [40]. However, to our surprise most of the experimental observations were captured by the model under a parameterization that simultaneously explained the perturbation experiments. Obviously, other hierarchical control mechanisms besides the growth regime (for example on the genome or proteome level) were not essential for describing the observed metabolite pool dynamics. This may be attributed to the fact that initial culture conditions were tightly controlled and that the media composition provided adequate substrate and by-product concentrations in the time span analyzed. Nevertheless, the inclusion of other levels of hierarchical control, in addition to the growth regime of this work, may contribute to simulated aspects of the observed dynamics. The enzyme kinetics and the direct influence by the growth regime are in the following considered as the sole source of regulatory principles that control glycolysis during MDCK cell cultivation.
First, the peak in the metabolite pools can be explained with a high GLUT-mediated flux rate in combination with low cell volume-specific enzyme activities (based on higher cell-specific volumes during the growth phase). The implemented enzyme kinetics realize a relatively higher net flux into the PPP during cell growth, which is attributed to the higher metabolite levels in glycolysis and similarly described by Wu et al. [41] for bovine venular endothelial cells after addition of citrate in order to inhibit the PFK activity. Also, the activation of GLUT in rat thymus lymphocytes with concanavalin A resulted in higher fluxes in glycolysis and into the PPP [42]. Higher fluxes into the PPP possibly enables enhanced nucleotide, macromolecule, and lipid synthesis rates, as reviewed by Mazurek et al. [43]. According to our simulations the fluxes are in the range of 13—15% of the glycolytic flux, which is reasonable for continuously growing mammalian cells in the exponential growth phase [37], [44]. However, a much lower contribution e.g. 5.8% and 3.6% can be found in the late intermediate growth phase, which corresponds to findings for other transformed mammalian cells [45]–[47]. So, the regulation of enzymes by substrates, products and allosteric effectors can change concentrations of intracellular metabolite pools, and reorganize the pathway fluxes, especially under limiting conditions (see section “Glycolytic activity during substrate perturbation”). However, during MDCK cell cultivation the control over the glycolytic activity is exerted by the growth regime through modulation of the GLUT activity. For many microorganisms, the GLUT is described as the rate limiting step that can control the glycolytic flux [38], [48]–[51]. But also adenosine-based nucleotides are reported to play a major role in the control of the glycolytic activity [36], [52]. For MDCK cells, the influence of adenosine-based nucleotides on glycolysis seems to be negligible during cultivation conditions with excess of substrates [22]. So, neither during cell growth nor during substrate perturbation the adenosine-based nucleotides played a crucial role in describing the dynamics of the measured metabolite pools. Therefore, we assumed for the model that enzymes are insensitive against changes in the adenosine-based nucleotide levels, which is also reported by Soboll et al. [53] for rat liver cells.
Snoep and co-workers hypothesized that GLUT controls cell growth [54]. This, however, raises the question, whether metabolism regulates cell growth or vice versa [55]. In case of adherent MDCK cell growth with sufficient substrate supply, the growth status is exclusively defined by the availability of free space on the well surface. Eventually, space becomes limiting and cells reduce the glycolytic activity although high extracellular glucose concentrations are present. Therefore, we hypothesize that the growth regime of exponentially growing MDCK cells controls the GLUT activity to realize a higher metabolic activity yielding in turn higher metabolite pools that meet the energy and precursor demands of the biosynthesis machinery. On a lower level of regulation, the properties of the involved enzymes shape metabolism by influencing flux distributions. Under substrate limitation, however, regulation of enzymes has full control over the glycolytic activity (see section “Glycolytic activity during cell cultivation”). Thus, the model considers that the regulation of the glycolytic activity changes with the physiological status of the cell [55] and sheds light on the regulatory principles that are essential to simultaneously explain various experimental scenarios. Although regulation of glycolysis can change with the microorganism [52], we are convinced that the derived principles can be applied to other metabolic pathways, such as the citric acid cycle [56], and also support the study of other mammalian cell lines relevant for production of biologicals [57].
Within a GLUT activity of 0–4 mmol L−1 min−1, the model for glycolysis is validated with cultivation, limitation and pulse experiments. It already shows a good predictive power for an experiment were MDCK cells were grown in DMEM medium with low GLCx levels (Fig. 8), which further strengthens the confidence in the model structure and its parameterization. Although the model prediction for the DMEM cultivation would benefit from a lower Elevel to describe all maximum peak-heights, it still confirms the close linkage of GLUT activity and intracellular metabolite dynamics. Based on the finding that the GLUT modulates the glycolytic activity during cell cultivation (under sufficient substrate availability) it seemed desirable to explore the maximum capacity of glycolysis and the corresponding ATP and PPP metabolite production. However, such a maximum capacity clearly depends on the enzyme content (Elevel) and the cell-specific volume (). Therefore, we exemplary analyzed cells from the Cult1 experiment at 24 h of cultivation with an actual uptake of 3.3 mmol L−1. For these cells, in silico modulation of the GLUT activity revealed that an uptake of up to 3.8 mmol L−1 min−1 can be realized until the glycolytic flux saturates the PFK capacity, which slightly enhances the ATP production on average to 105%, and the PPP metabolite and NADPH production to surprising 153% for cells of Cult1 at 24 h. According to the model, a further increase in ATP production would require the simultaneous overexpression of the PFK, which illustrates the difficulty in fast up-regulation of metabolic activity while keeping a certain balance between ATP and PPP metabolite production. However, Janke et al. [19] measured higher maximum in vitro PFK activities than estimated in this study and glycolysis of MDCK cells may have higher capacities than estimated by the model. Higher biomass precursor and ATP production rates can support higher growth rates as shown for tumor and yeast cells with up-regulation of the GLUT activity [58], [59]. Furthermore, Schmidt et al. [60] described a correlation between the growth of tumor cells and the ATP production rate. Potentially, an increase in the ATP production to 105% may not or only slightly support higher growth rates for MDCK cells especially as they are described to have a large overproduction in ATP [47], [61]. But due to the importance of PPP metabolite production to pyrimidine [43], [62] and purine production [63] and NAPDH to lipid synthesis we believe that an increase to 153% positively affects the growth of cells (Fig. 6). A glycolytic activity above 5 mmol L−1 min−1 drastically enhances the production of PPP metabolites (433%) at the expense of the ATP production (77%) and seems to be an interesting scenario for future experiments. However, also the reduction in the glucose uptake, as done by Liebl et al. [64], poses an interesting strategy to design a more economic breakdown of glucose in biotechnological processes [65]. Currently, the reduction of the glucose uptake by interference with the glucose transporter is also studied as a potential target for cancer treatment [59], [66] which may benefit from the acquisition of mathematical models to evaluate corresponding dynamics in metabolism. Taken together, the model can greatly support the development of strategies that aim either at a faster or a more efficient cell growth, and is also an aid in the design of new experiments.
The differential algebraic equations of the glycolytic model were composed of first order rate laws, Michaelis-Menten and Hill kinetics which describe enzyme activity in dependence of metabolite concentrations and allosteric influences.
|
10.1371/journal.pntd.0001947 | IFN-γ Production to Leishmania Antigen Supplements the Leishmania Skin Test in Identifying Exposure to L. braziliensis Infection | Cutaneous leishmaniasis due to L. braziliensis (CL) is characterized by a positive delayed type hypersensitivity test (DTH) leishmania skin test (LST) and high IFN-γ production to soluble leishmania antigen (SLA). The LST is used for diagnosis of CL and for identification of individuals exposed to leishmania infection but without disease. The main aim of the present study was to identify markers of exposure to L. braziliensis infection.
This cohort study enrolled 308 household contacts (HC) of 76 CL index cases. HC had no active or past history of leishmaniasis. For the present cross-sectional study cytokines and chemokines were determined in supernatants of whole blood culture stimulated with SLA. Of the 308 HC, 36 (11.7%) had a positive LST but in these IFN-γ was only detected in 22 (61.1%). Moreover of the 40 HC with evidence of IFN-γ production only 22 (55%) had a positive LST. A total of 54 (17.5%) of 308 HC had specific immune response to SLA. Only a moderate agreement (Kappa = 0.52; 95% CI: 0.36–0.66) was found between LST and IFN-γ production. Moreover while enhancement of CXCL10 in cultures stimulated with SLA was observed in HC with DTH+ and IFN-γ+ and in patients with IFN-γ+ and DTH−, no enhancement of this chemokine was observed in supernatants of cells of HC with DTH+ and IFN-γ−.
This study shows that in addition of LST, the evaluation of antigen specific IFN-γ production should be performed to determine evidence of exposure to leishmania infection. Moreover it suggests that in some HC production of IFN-γ and CXCL10 are performed by cells not involved with DTH reaction.
| Both control of L. braziliensis infection and development of cutaneous leishmaniasis (CL) are dependent on the host immunological response. Due to the difficulty of finding parasites in leishmanial lesions, a delayed type hypersensitivity reaction - leishmania skin test (LST), is widely used to diagnose CL. In areas of L. braziliensis transmission a positive LST is also documented in up to 18% of individuals without disease, who are considered to be putatively resistant to leishmania infection. However the mechanisms involved in the control of parasite grow is not known. The aim of this study is to identify tests that could determine in house contact of CL (HC) without past or current evidence of leishmaniasis exposure to leishmania infection. We found that of the 308 HC, 36 (11.7%) had a positive LST but in these IFN-γ was only detected in 22 (61.1%). Moreover of the 40 HC with evidence of IFN-γ production only 22 (55%) had a positive LST. Therefore at least the two tests, the LST and IFN-γ production, should be used to determine exposure to L. braziliensis. Identification of subjects exposed to leishmania infection that may or may not develop CL is highly relevant to understand pathogenesis of L. braziliensis infection.
| American tegumentary leishmaniasis (ATL) is caused predominantly by Leishmania braziliensis, L. guaynensis, L. mexicana and L. amazonensis [1], [2]. It is endemic in South and Central America and cutaneous leishmaniasis (CL), characterized by well delimited ulcers with raised borders, is the most common clinical picture of ATL. The main characteristics of the immunological response in CL are a strong Th1 type immune response to soluble leishmania antigen (SLA), demonstrated by positive delayed type hypersensitivity (DTH) reaction to the leishmania skin test (LST), as well as lymphocyte proliferation and production of high levels of IFN-γ and TNF-α [3], [4]. Only a few parasites are found in the lesions due to L.braziliensis and because of this the leishmania skin test (LST) is widely used for diagnosis of ATL. A positive test in a patient with a typical cutaneous lesion has a high predictive value [5], [6]. The LST has been also used to measure exposure to leishmania infection, and a positive LST in the absence of clinical manifestations of ATL has been documented in up to 17% of healthy individuals living in endemic areas of L. braziliensis [6]. Individuals with a positive LST who do not develop leishmaniasis are considered as having a subclinical L. braziliensis infection [7]. Although a concordance between DTH and in vitro tests of cell mediated immune response is expected, discordant results between IFN-γ production and tuberculin skin test (TST) have been shown in individuals with latent tuberculosis [8]–[10]. Therefore it is important to determine whether this discordance also occurs in subclinical L. braziliensis infection as well as to evaluate if other tests may be indicative of exposure to leishmania infection.
As the ratio from infection to disease based on LST is 3.7 to 1, about 25% of individuals who are exposed to L. braziliensis will develop cutaneous leishmaniasis [7]. It is known that early events after penetration of leishmania in the skin are important to determine the outcome of leishmaniasis. Therefore characterization of immune response early after the infection is highly relevant. Moreover, an early detection of individuals exposed to L. braziliensis will allow a comparative analysis between individuals who will develop or not develop disease. The aim of this study was to establish and follow a prospective cohort of household contacts of CL patients, to evaluate initially markers of exposure to leishmania infection and ultimately to identify markers that are associated with resistance or susceptibility to develop disease. Our initial evaluation of this cohort indicates based in a cross-sectional study that more than one test is needed to determine exposure to L. braziliensis. In addition to LST, IFN-γ production in SLA stimulated cultures should be determined. Moreover, the production of CXCL10, a chemokine associated with recruitment and activation of T cells, gives support that in some cases CXCL10 and IFN-γ are produced by cells not involved with DTH.
This study was approved by the Ethical Committee of the Federal University of Bahia. Written informed consent was obtained from all enrolled subjects.
This study was conducted in Corte de Pedra, a rural region in Northeastern Brazil endemic for ATL, where we have performed clinical and immunological studies for over 25 years [3], [11], [12]. The area was previously dominated by Atlantic rainforest and is now a mostly deforested agricultural community. Lutzomyia whitmani and Lu. intermedia sandflies that transmit L. braziliensis are endemic in the local fauna [2], [13]. The health post of Corte de Pedra was created in 1986 and is a reference center for diagnosis and treatment of CL and is staffed by medical personnel from the Federal University of Bahia.
This is a cohort study enrolling household contacts (HC) of patients with history of CL. Index cases (IC, N = 76) were recruited at the Corte de Pedra health post. An index case was defined as a patient with confirmed CL diagnosed at the Corte de Pedra health post within two years prior to enrollment in the study, living within a 10 km radius of the health post. Patients with evidence of mucosal or disseminated leishmaniasis were not considered for enrollment as index cases. Researchers visited index cases in their homes to recruit HC from January to April 2010. Household contacts (N = 533) were defined as individuals without history of any type of leishmania infection who were living in the same home as the index case at the time of enrollment in the study and at the time of diagnosis of CL by the index case. Cutaneous leishmaniasis (CL) patients were diagnosed based on a typical clinical leishmaniasis lesion, associated with a positive leishmania skin test (LST) and documentation of parasites in culture or by histopathology. After obtaining informed consent, a negative history of CL in HC was established by a medical interview, assessing for symptoms consistent with previous CL or ML infection, and negative physical exam looking for scars consistent with past CL or ML on the skin, nose and soft palate. Exclusion criteria for HC were age less than 2 or more than 65 years, or frequent stays outside of the endemic area. After the initial immunologic studies a cross-sectional study was performed first comparing epidemiological and chemokine data among HC with evidence or without evidence of immune response to soluble leishmania antigen (SLA). Second, comparing cytokine data among groups according evidence of IFN-γ production and response to LST.
SLA for the skin test was prepared with an isolate of L. braziliensis as previously described [14]. Briefly, promastigotes of L.braziliensis were grown in Schneider's medium supplemented with 10% fetal bovine serum and 2% human urine. The promastigotes were washed in sterile phosphate buffered saline (PBS), resuspended in lysis solution (Tris HCl, EDTA and leupeptin), immersed in liquid nitrogen, and thawed at 37°C. After freeze-thaw procedure, they were sonicated. The disrupted parasites were centrifuged at 14,000G and assayed for total protein (BCA Protein Assay Kit, Thermo Scientific). For in vitro testing the filtrate was adjusted to a concentration of 500 µg/mL with sterile PBS. For the LST, the filtrate was adjusted to a concentration of 250 µg/mL with sterile PBS containing Tween 80 and phenol at final concentrations of 0.0005%(w/v) and 0.28% (w/v) respectively.
Once a negative history of CL was established, heparinized peripheral blood (10 mL) was collected immediately before LST was performed. About 6 h after collection, 1 mL of whole blood aliquots was dispensed into a 24-well tissue plate. SLA at 20 µg/mL, 50 µl of phytohemaglutinin (GIBCO, Grand Island- NY) as positive control or medium (negative control) were added to each well and incubated at 37°C 5% CO2 for 72 hours. Plasma supernatants (300–400 uL) were collected, and samples were stored at −20°C. The levels of IFN-γ released were quantified by ELISA, using commercially available reagent (BD OpTEIA, San Diego-CA). A standard curve was used to express the results in pg/ml. A positive test was defined as any detectable IFN-γ level after subtracting the SLA stimulated IFN-γ levels by unstimulated IFN-γ levels.
The levels of CXCL9, CXCL10, and CCL2 were measured by ELISA using commercially available reagents (BD OpTEIA, San Diego-CA). A standard curve was used to express the results in pg/ml. Chemokine data are summarized separately as spontaneous production (medium) and SLA-induced production (Ag).
The LST was performed after collection of blood for the in vitro test to avoid the possible influence of the skin test reaction on the immune response determined in vitro. SLA for the skin test was prepared with an isolate of L. braziliensis as previously described [14]. For LST, 0.1 mL (25 µg/mL) of the SLA was injected intracutaneously on the volar surface of the forearm, and the greater diameter of induration was measured 48–72 h later. Induration of ≥5 mm was defined as a positive reaction.
HC were considered to have been exposed to leishmania if they demonstrated a positive immune response to L. braziliensis antigen in either the in vitro and in vivo test.
The analysis of the concordance between LST and IFN-γ production was performed by calculating a κ statistic for agreement with 95% confidence interval. Demographic characteristics were compared across subject groups as follows: For continuous variables, one-way ANOVA or Kruskal-Wallis tests were used and if the overall P-value was <0.05, pair-wise Bonferonni or Dunn's post-hoc tests were performed. For categorical variables, chi-square or Fisher's exact test were performed. Student T test was used to compare the means of variables following normal distribution. For analysis of IFN-γ and chemokines production in unstimulated and stimulated culture, the Wilcoxon rank-sum test was used. Correlations were performed by the method of Spearman. STATA version 11 (College Station, TX) and GraphPad InStat3 (La Jolla, CA) statistical software were used for all the analyses.
Figure 1 depicts the design of the cohort study and distribution of study subjects. A positive LST was observed in 36 (11.7%) of the 308 HC tested and IFN-γ production was detected in 40 HC (12.9%). However in the 36 individuals with positive LST, production of IFN-γ was only observed in 22 (61.1%), and in the 40 HC with evidence of production of IFN-γ, only 22 (55%) had a positive LST. Considering evidence of immunological response to leishmania antigen as a positive LST and/or IFN-γ production in cultures stimulated with SLA, 54 (17.5%) of 308 HC had specific immunological responses to SLA. IFN-γ production was documented in supernatants from cells of all HC stimulated with phitohemaglutinin.
The demographic and epidemiological aspects of index cases and HC with and without evidence of immune response to SLA are shown in Table 1. There was no difference among the index cases and HC with evidence of immune response in all variables analyzed (age, gender, occupation, years living in the area as well as in the same house). There was also no difference among the three groups regarding the time of arriving at home. However, differences were found between HC with and without evidence of exposure to leishmania infection regarding age, occupation, time living in the endemic area and time living in the same house. Household contacts without evidence of immune response were younger and the majority was students; consequently, they had less time living in the same house and in the same endemic area of index cases than HC with evidence of immune response.
In order to determine the best test to detect exposure to leishmania infection, we compared the ability of the LST and IFN-γ production to identify exposure to L. braziliensis. When assessing the concordance between LST and IFN-γ production, among those who had at least one test positive, 61.1% of the HC were positive for both tests and there was moderate agreement beyond that expected by chance (Kappa = 0.49; 95% CI: 0.34–0.64). When the two tests were analyzed in the whole population a moderate concordance between IFN-γ and LST (Kappa = 0.52; 95% CI: 0.36–0.66) was confirmed.
As chemokines are produced early in leishmania infection and they participate in the immune response by activation and recruitment of T cells, expression of chemokines related to lymphocyte and monocyte recruitment were determined in all individuals with evidence of immune response (N = 54) and in a sub-group (N = 51) of HC without evidence of immune response. These 51 HC without evidence of immune response were selected among individuals living in the same house of HC with evidence of immune response. The IFN- γ, CXCL9, CXCL10 and CCL2 levels in IC, HC with evidence of immune response and HC without evidence of immune response are shown in Table 2. As expected IFN-γ levels were higher in IC than in the other groups. High levels of CXCL9 and CCL2 were found in unstimulated cultures of both IC and HC with and without evidence of immune response. While there was no difference regarding the production of CCL2 in unstimulated culture, the levels of this chemokine were higher in HC with evidence of immune response in cultures stimulated with SLA when compared with the control group (p<0.001). With regard to CXCL9 and CXCL10, levels of SLA induced chemokines were higher among the IC and HC with evidence of immune response than the levels observed in the control group (P<0.05).
To evaluate if the chemokine production was associated with a positive LST or ability to produce IFN-γ in vitro upon SLA stimulation, the 105 HC who had chemokines determined were divided into 4 sub-groups: 1) LST positive and evidence of IFN-γ production (LST+ IFN-γ+); 2) LST+ and absence of IFN-γ production (LST+ IFN-γ−); 3) LST negative and IFN-γ positive (LST− IFN-γ+); 4) Both LST and IFN-γ negative (DTH− LST−). The production of CXCL9 and CXCL10 in both spontaneous and in SLA stimulated cultures is shown in Figure 2. The levels of CCL2 were increased in cultures stimulated with SLA in all individuals who had a positive LST or IFN-γ production as well as in HC without evidence of immune response (data not shown). However, the high levels of CCL2 in cultures of individuals without evidence of immune response prevent the use of CCL2 as an indicator of evidence of immune response to SLA. An increase in the production of CXCL9 in cultures stimulated with SLA was documented in IC and HC with evidence of IFN-γ production and with positive LST (Figure 2A). In the majority of HC, positivity in both tests (LST and production of IFN-γ) was required for a high production of CXCL9 in SLA stimulated cultures. Production of CXCL10 in unstimulated and SLA stimulated cultures is shown on Figure 2B. IC and subjects with positive LST and IFN-γ production, as well as those who were LST negative but produced IFN-γ, had increased CXCL10 in SLA stimulated cultures. However, one striking observation was that while CXCL10 production was detected in SLA stimulated cultures of HC with a negative LST but producers of IFN-γ, absence or very low production of CXCL10 was observed in SLA stimulated cultures of individuals who, despite having a positive DTH, had an absence of IFN-γ production to SLA. Moreover, there was a positive correlation between production of IFN-γ and CXCL10 in SLA stimulated culture (Figure 3).
Epidemiologic and clinical studies of ATL have focused on determining the influence of host, parasite and environmental factors on the development of the different clinical forms of leishmaniasis [15], [7], [11], [16]. However, there is a lack of studies in individuals who have evidence of exposure to leishmania but may or may not develop disease. The documentation of a positive LST in the absence of current or past history of CL is the only test that has been used to identify individuals who have a subclinical form of L. braziliensis infection [6], [7]. Herein we showed that the use of in vitro immunologic tests such as production of IFN-γ not only increases the number of individuals with evidence of exposure to leishmania infection, but we also demonstrate discordant results between LST and in vitro IFN-γ production by cells of HC. Moreover, the increase in CXCL10 production in SLA stimulated cultures was predominantly associated with IFN-γ production rather with a positive LST.
The primary objective of this study was to identify individuals recently exposed to leishmania infection. A limitation of this type of study is that we cannot be sure when exposures to leishmania have occurred. However, the comparative analysis of demographic characteristics of the index cases and of the HC with and without evidence of immune response suggested that HC with evidence of immune responses more closely resembled index cases, supporting the contention that exposure to leishmania infection in this population likely occurred close to the period that the CL index cases acquired the parasite.
The DTH and in vitro immunologic response determined by lymphocyte proliferation or cytokine production to recall antigens have been widely used to determine evidence of cell mediated immune response [17]–[20], [9]. In patients with CL as well as in patients with mucosal leishmaniasis due to L. braziliensis infection there is a strong association between the positivity of LST and production of IFN-γ as well as evidence of lymphocyte proliferative response to SLA [3], [17], [15]. Because of the high predictive value of LST in the diagnosis of CL, this test has also been used to identify exposure to leishmania infection among healthy individuals living in areas of Leishmania sp transmission [21], [6], [7]. Individuals with a positive LST and absence of current or past history of leishmaniasis in areas of L. braziliensis transmission are considered as having subclinical L. braziliensis infection. We have previously shown that these individuals with a positive LST have a lower production of IFN-γ in SLA stimulated cultures than patients with CL, and in some of them even no detectable IFN-γ levels were found in supernatants of lymphocytes stimulated with SLA [7], [16]. In the present study in addition to showing that HC that have a positive LST may not produce IFN-γ upon SLA stimulation, there were also HC who produced IFN-γ but had a negative LST. Discordance between DTH and IFN-γ production has been observed in subjects with latent tuberculosis [22], [8], [9], [10], and several factors may explain the discordance between LST and in vitro IFN-γ production: 1) Presence of suppressor factors in vivo that prevent the documentation of DTH; 2) Production of IFN-γ by non T cells; 3) Lack of effector or effector memory T cells but presence of memory T cells. Discordance between DTH and in vitro tests have been documented in patients with active tuberculosis as well as in individuals with subclinical L. chagasi infection [9], [23]. In such cases malnutrition as well as the presence of soluble suppressor factors may explain the absence of response in vivo but the occurrence of response in vitro. Usually in this case restoration of the DTH test occurs after specific therapy as well as with improvement in nutritional status [24], [25]. Regarding memory, studies in an experimental model of leishmaniasis have shown that after control of leishmania infection memory effector cells may not be found but animals remain with central memory T cells [26]. IFN-γ in patient with CL is predominantly secreted by effector T cells or memory effector T cells [27]. However, the disappearance of these cells from the peripheral blood may make the in vitro test become negative. Since in DTH tests antigens are inoculated intradermally and immune response is evaluated 48 to 72 hours after the test, there is time for memory effector T cells that remain in lymph nodes or in other tissues to migrate to the site of the injection of SLA and react in vivo to leishmania antigen. As the participants of this study were healthy, well-nourished, and likely recently exposed to L. braziliensis based on the epidemiologic data, the more likely explanation for the discordance between the LST and IFN-γ production in the LST negative IFN-γ positive subjects is the production of IFN-γ by cells not involved with DTH. Giving support to this hypothesis is our data that CXCL10 was produced in HC with evidence of IFN-γ production but not in those with only a positive LST. Different cell types such as neutrophils, NK cells and NKT cells may produce IFN-γ [28]–[31] and future studies will address this subject.
This study shows that in L. braziliensis infection in addition to the LST, the documentation of exposure to leishmania antigen should be also evaluated by antigen specific IFN-γ production as it increases the evidence of exposure to leishmania infection from the 11.7% (as documented by LST) to 17%, when LST or IFN-γ were positive. The discordance between IFN-γ and LST was highlighted by the observation that production of CXCL10 was associated with IFN-γ production but not with a positive LST.
|
10.1371/journal.pcbi.1000872 | A Simple Self-Maintaining Metabolic System: Robustness, Autocatalysis, Bistability | A living organism must not only organize itself from within; it must also maintain its organization in the face of changes in its environment and degradation of its components. We show here that a simple (M,R)-system consisting of three interlocking catalytic cycles, with every catalyst produced by the system itself, can both establish a non-trivial steady state and maintain this despite continuous loss of the catalysts by irreversible degradation. As long as at least one catalyst is present at a sufficient concentration in the initial state, the others can be produced and maintained. The system shows bistability, because if the amount of catalyst in the initial state is insufficient to reach the non-trivial steady state the system collapses to a trivial steady state in which all fluxes are zero. It is also robust, because if one catalyst is catastrophically lost when the system is in steady state it can recreate the same state. There are three elementary flux modes, but none of them is an enzyme-maintaining mode, the entire network being necessary to maintain the two catalysts.
| The question of whether a whole organism (as opposed to particular properties of an organism) can be modeled in the computer has been controversial. As a step towards resolving it, we have studied the feasibility of simulating the behavior of a simple theoretical model in which all the catalysts needed for the metabolism of a system are themselves products of the metabolism itself, and in which there is a continuous loss of catalysts in unavoidable degradation reactions. In addition to a trivial (“dead”) steady state in which all rates are zero, the model is capable of establishing a stable non-trivial steady state with finite and reproducible fluxes. This can be achieved by “seeding” it with a sufficient quantity of at least one of the catalysts needed for functioning. It is also robust, because it can recover from a catastrophic disappearance of a catalyst.
| Several theories of life [1]–[5] coincide in the importance that they give to metabolic closure, the necessity for all of the catalysts essential for survival of an organism to be produced internally, as an organism cannot rely on any external agent for maintaining it. The same considerations must apply to the self-maintaining systems at the origin of life [6], [7]. Rosen [1] expressed this idea that catalysts must be produced by the system itself by saying that it must be closed to efficient causation. These theories differ in their details, and each includes important points missing from the others. Among them the theory of (M,R)-system s, or metabolism–replacement systems, perhaps comes closest to a complete explanation of life, but it is usually presented in abstract terms [1] that make it difficult to relate it to any ordinary ideas of chemistry, metabolism and catalysis.
To give concrete expression to the idea of an (M,R)-system , and to evaluate its possible relevance to the origin of life, we proposed [8]–[10] a simple system of three interlocking cycles: a metabolic process produces a metabolite ST from external precursors S and T in a reaction catalyzed by a component STU that is itself the product of a replacement process , in which U is another external precursor. The replacement process is necessary because STU, as a biological molecule, cannot be assumed to have an infinite lifetime, and even if it did it would be diluted by growth of the system and by other processes. Moreover, replacement also needs a catalyst, which also needs to be replaced. To escape immediately from the implied infinite regress we proposed that replacement is catalyzed by a similar type of molecule, SU, that results from a secondary reaction catalyzed by STU, . This system, illustrated in Fig. 1, is closed to efficient causation, because each of the three reactions is catalyzed by a product of the system itself. In our original proposal we assumed that only STU and SU were subject to unavoidable degradation (see Fig. 1b of [10]), but there is no logical reason to suppose that the other product of the system, ST, is indefinitely stable, especially as it is assumed to be a molecule similar to SU. In Fig. 1, therefore, there is a third degradation reaction, reaction 11.
A controversial aspect of Rosen's analysis is his contention that a system closed to efficient causation cannot have computable models [11]–[14]. Many aspects of biological systems can, of course, be simulated in the computer, and many examples of metabolic simulation can be found in the literature, but typically these examples do not simulate systems that are closed to efficient causation. In the recent simulation of aspartate metabolism in Arabidopsis thaliana [15], for example, the enzymes were taken as given; their production was not simulated. We discuss this controversy elsewhere [16] and will not do so here, apart from noting that there is no obvious reason why the system illustrated in Fig. 1 should not be simulated. On the contrary, it can certainly be simulated, as we shall show, with results that shed light of the conditions that need to be fulfilled by a self-maintaining system.
We shall show that a simple (M,R)-system can be robust, capable of a recovering from the loss of most of its catalysts, and in addition has the interesting property of bistability. As Delbrück [17] emphasized many years ago, multistability is also an important property for all but the simplest living organisms because it is essential for differentiation, an idea that has subsequently been developed by other authors [18]. Bistability can arise in systems considerably smaller and simpler [19] than the one we discuss in this paper, but we are concerned here with (M,R)-system s, which must be closed to efficient causation.
For the system to be simulated it needs to be defined in precise numerical terms, and for doing this it is convenient to expand the catalytic processes shown in Fig. 1a into the cycles of chemical reactions shown in Fig. 1b [9]. There is no fundamental difference in this model between catalysts (“enzymes”) and metabolites, and elsewhere [10] we have argued that no fundamental difference exists: all enzymes are products of the system in which they participate, and are thus metabolites, and many conventional metabolites (for example, ornithine in the urea cycle) participate in cycles of reactions, and thus satisfy the definition of a catalyst.
All simulations and studies of the stability of the steady states found were done with Matlab and checked with COPASI [20], or vice versa, and stoichiometric network analysis was done with MetaTool [21]. In the present paper all simulation is deterministic.
As we shall be supposing that the system in Fig. 1 can continue in operation indefinitely, despite containing irreversible degradation steps, we need to consider the thermodynamic feasibility of what we propose. In effect, we assume that the overall chemical reactions degradation products, degradation products and degradation products are irreversible, that synthesis of ST in the reaction is thermodynamically favored, and that the concentrations of the external molecules S, T and U are constant, either because the quantities consumed by the system are too small to have any effect on their concentrations, or because they are buffered by external chemistry. External constraints on a system of chemical reactions can be applied in two main ways, either with constant external concentrations or with constant input fluxes. In this model we have chosen the former approach, primarily to facilitate comparison with earlier work [8]–[10].
In this context it is important to note that organizational closure does not imply thermodynamic closure, or vice versa. In the Aristotelean terminology favored by Rosen [1], closure to efficient causation is not the same as closure to material causation [10]. An organism must clearly be open to material causation — it “feeds on negative entropy”, in Schrödinger's words [22] — but it can still synthesize all of its catalysts, and thus be closed to efficient causation. In a third type of closure, independent from both of these, an individual organism must be structurally closed, separated from other individuals by a skin or other barrier. This aspect was given almost no attention by Rosen [1], and we shall not discuss it further here, but it is clearly necessary, and it forms an important element of other theories of life such as autopoiesis [3].
The concentration evolution of the different metabolites in Fig. 1b can be described by a series of ordinary differential equations:The simple non-linear terms in these equations arise from applying simple mass action kinetics to the bimolecular steps.
Stationary solutions of the system of Fig. 1b were obtained by two different methods, numerical integration of the previous set of differential equations, and analytical solution of the nonlinear algebraic equations. Both revealed the existence of a region with three distinct steady states, one trivial and two non-trivial. It is obvious that the system shown in Fig. 1 cannot undergo any reactions if no form of any catalyst is present. Less obvious is whether it can construct itself and maintain itself indefinitely if it is seeded with a sufficient quantity of one catalyst. This has been tested in the first instance with various values in the range 0–0.6 of the degradation rate constants , and , other rate constants as defined in Fig. 1, and various initial concentrations of one intermediate, STU, all other intermediate concentrations being set initially to zero.
For the system cannot construct itself or maintain itself despite seeding with large or small initial concentrations of STU, and it always ends in a trivial steady state with all concentrations and all rates zero. However, with smaller degradation rate constants it can reach either the trivial steady state or a non-trivial steady state with all concentrations and rates non-zero, i.e. a self-maintaining regime. The results are summarized in Fig. 2 for and different initial concentrations of STU. For the system reached a trivial steady state with all concentrations zero, but at any it reached a non-trivial steady state with and all other concentrations and all rates non-zero. Hence there is a none-to-all transition at this critical point, as indicated by the broken line in Fig. 2a.
STU is not of course the only catalytic intermediate that could be used for seeding the system, and results with each of the others, and for some pairs of intermediates, are shown in Table 1, for two values of . Two important points are evident in this table: first, any metabolite apart from ST or SU can be separately used to seed the system, and although the concentration of seed metabolite necessary to drive it to a non-trivial steady state varies with the seed, the steady state reached depends only on the degradation rate constants, and is independent of the identity of the seed. We have also made simulations with various mixtures of metabolites used as seeds and these generalizations remain true.
The reason why ST and SU cannot act as seed can be seen by inspection of Fig. 1b: neither of these metabolites reacts directly with any of the external reactants S, T and U, and so no reaction can take place if none of the other metabolites are present. However, ST and SU can react with one another to give a product SUST capable of participating in additional reactions and closing all the loops. Not surprisingly therefore, the system can be seeded with a mixture of ST and SU even though neither of them is effective alone.
To verify the stability of the steady states, the Jacobian matrices were evaluated at the steady states obtained, and the eigenvectors and eigenvalues calculated. For those conditions in which three steady states were obtained, , the trivial and one of the non-trivial solutions always have all eigenvalues with negative real parts, and thus are asymptotically stable. Obviously, they correspond to those reached by numerical integration experiments. The additional non-trivial steady state calculated by the analytical solution of the non-linear algebraic equations has, however, one of the eigenvalues with positive real part, and is therefore an unstable steady state (a saddle point), so in this region the system exhibits bistability. Beyond the critical value, , only the trivial steady solution exists and is asymptotically stable, i.e. each of its eigenvalues has a negative real part. These results are summarized in the bifurcation diagram illustrated in Fig. 3.
The diagram of Fig. 3 predicts a sort of hysteretic behavior: if the system is in the stable non-trivial steady state with small values of the decay rate constants, it remains in the same state when these constants are increased, until it abruptly collapses to the trivial steady state when the critical point is reached. Once in the trivial steady state, it remains there even when the decay rate constants are decreased below the critical point. The hysteretic cycle cannot be completed unless we allow the possible appearance of trace quantities of any intermediate (such as might result from external chemistry) that could allow the system to recover the non-trivial steady state when close enough to the equilibrium condition .
The unstable steady state that appears in those conditions of bistability, , belongs to a separating barrier that constitutes a hypersurface limiting the attractor regions of both trivial and non-trivial stable steady states. A planar region of the phase diagram is illustrated in Fig. 4 for . Different initial conditions close enough to the separating barrier could drive the system either to one stable steady state or the other, as shown in Fig. 5.
It is clear that the system as described is capable of reaching a stable non-trivial steady state with finite fluxes and finite concentrations of all intermediates. However, before it can be regarded as a useful model of a self-maintaining system, and thus relevant to the early stages of metabolic evolution, it needs to be shown to be capable of recovering from catastrophic loss of one or more catalysts. To test this, it was allowed to reach the non-trivial stable steady state characteristic of , and the concentrations of all forms of STU (not only STU itself but also STUS, STUST and STUSU) were then abruptly set to zero, the others being left at their steady-state values. As seen in Fig. 6, both intermediate concentrations return to the previous steady-state values.
As STU catalyzes two different processes (synthesis both of SU and of ST), loss of STU is clearly the most stringent loss of catalyst one could consider, but for completeness we also tested the effect of loss of all forms of SU, with similar results. All of this shows that the system is highly robust, not only for infinitesimal perturbations, as tested by analysis of the Jacobian matrix, but also for large perturbations. Unless it is perturbed to such a large extent that the separating barrier mentioned is crossed, e.g. below the threshold requirements listed in Table 1 (for individual metabolites, but generalizable to combinations of metabolites), it always returns to the same non-trivial steady state. It can equally resist very large increases in metabolite concentrations, for example, when ST was abruptly raised to 100 times its steady-state value the system returned rapidly to the same steady state.
With the use of MetaTool [21] we have analyzed the structure of the model by means of an approximation to a stoichiometric analysis in the steady state. In this analysis. S, T and U are considered as external metabolites, the others being considered internal. With the rates numbered as in Fig. 1b the reaction subsets are as follows:As seen in this equation, subsets of reactions operate at the same rate in the steady state, i.e. , and , as illustrated in Fig. 7a. Notice that the degradation rates and for the two catalysts STU and SU are in the same subsets as the corresponding replacement reactions: with , and ; but with and . This explains how the replacement can efficiently balance the decay of each catalyst in the steady state.
The resulting convex basis can be expressed in the following way:All three basis elements are shown schematically in Fig. 7b. The first, , includes the reactions involved in the metabolism process, corresponds to both the metabolic and replacement cycles, and the third, , is the pathway that replaces the replacement catalyst SU. As previously shown in the subsets of reactions, the rate of decay of ST does not share the same rate with any other reaction. However, it also is compensated as a consequence of the performance of the metabolic reactions , and , as deduced from the inspection of the first element of the convex basis.
To study the relative contributions of the basis elements to the steady-state flux distribution, , and were evaluated from the numerical integration results for different values of the degradation rate constants (Fig. 8). The optimum operating rate value is obtained when is in the range 0.2–0.3, rather closer to the conditions for bifurcation than those for equilibrium (Fig. 8a). The contribution of turns out to be around double that of over most of the range. Nevertheless, as shown in Fig. 8b, as the degradation rate constants increase, the relative contributions of and decrease steeply until the bifurcation point is reached for . The rates of the replacement reactions, executed by these basis elements, then become incapable of withstanding the huge degradation rates, and the system collapses.
In the present model, the elementary flux modes coincide with the elements of the convex basis. Nevertheless, none of them is an enzyme-maintaining mode [23] because none of , and could indefinitely function alone, i.e. STU acting in and needs to be replaced but at the same time ST and SU in need the replacement function in and , respectively (Fig. 7b). Thus, , and should all be greater than 0. This is the reason why the entire system constitutes an indivisible enzyme-maintaining mode.
A simple model of an (M,R)-system consisting of three catalytic cycles organized so that all catalysts are products of reactions within the system is able to establish and maintain a non-trivial steady state capable of resisting degradation of all the catalysts, provided that this degradation is not so fast that the catalysts are eliminated faster than they are regenerated. This model was originally proposed as a way of giving concrete expression to the abstract view of life embodied in Rosen's (M,R)-system s [1]. It does of course oversimplify some aspects of his analysis, but we consider that it is helpful for understanding the nature of his concept of closure to efficient causation. It shows that a small system in which all catalysts are produced internally can not only organize itself into a non-trivial steady state, but it can also recover from large perturbations, such as complete loss of a catalyst. In favorable conditions and with a large amount of time available, the system in stable steady state can create itself from essentially nothing — a few suitable reactants present in vanishingly small amounts. As mentioned above, no elementary flux mode in this model is independently capable of maintaining itself. We are conscious that this does not constitute a proof of the simplicity of this model. In fact the model in the form originally proposed [8] did not allow for degradation of ST (reaction 11 in Fig. 1), and in a sense, therefore, represented a simpler system. However, the inclusion of this decay process is more realistic when considering the capacity of ST to be driven to new processes of increasing complexity and thus the evolutionary potential of the model.
As our original model [8] was designed to be self-maintaining the demonstration here that it is indeed capable of self-maintenance confirms our prediction. The bistability that it also shows was not consciously designed. This leads to more complex dynamics, and the advantages of multistability for a living organism have been discussed previously [17], [18]. The model is composed of various interconnected reactions, and it can be decomposed into individual circuits according to either logical or stoichiometric criteria; it was, in fact, constructed logically, with interplay of three basic building blocks, as described in the Introduction. These three cycles have both structural and dynamic roles in the self-maintenance of the entire system, and they exert constraints on the conditions for a “living”, non-trivial steady state, as discussed already and illustrated by Fig. 8. We have checked that none of them exhibits bistability by itself, and the occurrence of multistationarity is a consequence of the combined action of all of them: no “living” steady state is achieved in the system if any reaction (other than a degradation step) of the model is eliminated.
As mentioned in the Introduction, a smaller system [19] than the one in Fig. 1 can show bistability: this was presented as the smallest chemical reaction system with bistability, but it is not a model of an organism because it includes no mechanism for regenerating the catalyst, and if this is lost, for whatever reason, no recovery is possible. We do not claim to have demonstrated that the model studied here is the simplest system capable of self-maintenance.
The simplicity of this robust self-maintaining system and its capacity to be easily seeded may allow us to regard it as a plausible prebiotic system. Specifically, the establishment of a reflexive autocatalysis, i.e. autocatalysis that results from the structure of the whole network rather than from specifically autocatalytic components, is a typical common feature of models that illustrate recent theories of the origin of life; for example, the “lipid-world” scenario [24] and the theory of autocatalytic sets of proteins [25] share this property. Although the chemical nature of the components in the system analyzed in this paper is not specified, its autocatalytic organization is sufficient to satisfy the definition of an autocatalytic set: STU catalyzes synthesis of SU and vice versa. Of course, the difficulty of spontaneously developing a realistic {STU, SU} dual set of molecules performing such a special task of autocatalysis is arguable, but no other simple model of organizational closure escapes this criticism either. In any case, the essential postulate is that acquisition of some kind of recursive autocatalytic network should have been a necessary step at the very beginning of prebiotic evolution, before the development of more complex infrabiological systems [26].
In this analysis we have effectively assumed that a primitive self-maintaining system has metabolism but does not have information processing, in other words a metabolism-first scenario for the origin of life. All of the principal current theories of life [1]–[5] incorporate metabolism, but only a minority [2], [5] explicitly incorporates storage of information; even the autocatalytic sets [4] treat RNA molecules only as catalysts, not as information stores. Particularly interesting is that this simple (M,R)-system shows functions that depend on the arrangement of elements in its intermediates: multiple components have the same composition but different functions, depending on the arrangement of their elements, e.g. SUSTU and STUSU are isomers with different activities, and the same is true of STUS and SUST. As the model is drawn, the structural differences are differences in sequence, suggesting sequence-dependent information storage even in a metabolism-first model of the origin of life: thus the borderline between replication-first and metabolism-first approaches to the origin of life may not be absolute. Indeed, this typical dichotomy may be blurred when considering simple organizational recursive systems in which the different chemical intermediates necessarily have parts of their structures in common.
|
10.1371/journal.pmed.1002312 | Estimation of the cost-effectiveness of HIV prevention portfolios for people who inject drugs in the United States: A model-based analysis | The risks of HIV transmission associated with the opioid epidemic make cost-effective programs for people who inject drugs (PWID) a public health priority. Some of these programs have benefits beyond prevention of HIV—a critical consideration given that injection drug use is increasing across most United States demographic groups. To identify high-value HIV prevention program portfolios for US PWID, we consider combinations of four interventions with demonstrated efficacy: opioid agonist therapy (OAT), needle and syringe programs (NSPs), HIV testing and treatment (Test & Treat), and oral HIV pre-exposure prophylaxis (PrEP).
We adapted an empirically calibrated dynamic compartmental model and used it to assess the discounted costs (in 2015 US dollars), health outcomes (HIV infections averted, change in HIV prevalence, and discounted quality-adjusted life years [QALYs]), and incremental cost-effectiveness ratios (ICERs) of the four prevention programs, considered singly and in combination over a 20-y time horizon. We obtained epidemiologic, economic, and health utility parameter estimates from the literature, previously published models, and expert opinion. We estimate that expansions of OAT, NSPs, and Test & Treat implemented singly up to 50% coverage levels can be cost-effective relative to the next highest coverage level (low, medium, and high at 40%, 45%, and 50%, respectively) and that OAT, which we assume to have immediate and direct health benefits for the individual, has the potential to be the highest value investment, even under scenarios where it prevents fewer infections than other programs. Although a model-based analysis can provide only estimates of health outcomes, we project that, over 20 y, 50% coverage with OAT could avert up to 22,000 (95% CI: 5,200, 46,000) infections and cost US$18,000 (95% CI: US$14,000, US$24,000) per QALY gained, 50% NSP coverage could avert up to 35,000 (95% CI: 8,900, 43,000) infections and cost US$25,000 (95% CI: US$7,000, US$76,000) per QALY gained, 50% Test & Treat coverage could avert up to 6,700 (95% CI: 1,200, 16,000) infections and cost US$27,000 (95% CI: US$15,000, US$48,000) per QALY gained, and 50% PrEP coverage could avert up to 37,000 (22,000, 58,000) infections and cost US$300,000 (95% CI: US$162,000, US$667,000) per QALY gained. When coverage expansions are allowed to include combined investment with other programs and are compared to the next best intervention, the model projects that scaling OAT coverage up to 50%, then scaling NSP coverage to 50%, then scaling Test & Treat coverage to 50% can be cost-effective, with each coverage expansion having the potential to cost less than US$50,000 per QALY gained relative to the next best portfolio. In probabilistic sensitivity analyses, 59% of portfolios prioritized the addition of OAT and 41% prioritized the addition of NSPs, while PrEP was not likely to be a priority nor a cost-effective addition. Our findings are intended to be illustrative, as data on achievable coverage are limited and, in practice, the expansion scenarios considered may exceed feasible levels. We assumed independence of interventions and constant returns to scale. Extensive sensitivity analyses allowed us to assess parameter sensitivity, but the use of a dynamic compartmental model limited the exploration of structural sensitivities.
We estimate that OAT, NSPs, and Test & Treat, implemented singly or in combination, have the potential to effectively and cost-effectively prevent HIV in US PWID. PrEP is not likely to be cost-effective in this population, based on the scenarios we evaluated. While local budgets or policy may constrain feasible coverage levels for the various interventions, our findings suggest that investments in combined prevention programs can substantially reduce HIV transmission and improve health outcomes among PWID.
| The US drug-injecting population is growing, with consequent deaths from opioid overdose (30,000 deaths in 2014 alone) as well as increased HIV transmission.
Cost-effective HIV prevention programs for people who inject drugs are essential to the long-term health outcomes for this population and other high-risk groups in the US.
Our model-based analysis was designed to identify high-value portfolios of HIV prevention programs targeted to people who inject drugs in the US.
We adapted and extended a dynamic compartmental model that tracks multiple HIV risk groups in the US.
We projected the lifetime costs and benefits resulting from combinations of opioid agonist therapy, needle and syringe exchange programs, enhanced HIV screening and antiretroviral therapy programs, and oral HIV pre-exposure prophylaxis (PrEP) for people who inject drugs.
We estimated that opioid agonist therapy, which reduces injecting frequency and results in multiple, immediate quality-of-life improvements, can be cost-effective, while less expensive programs like needle and syringe exchange can provide additional value when used in combination. PrEP was not likely to be cost-effective in this population in most of the scenarios we evaluated.
Several prevention interventions—opioid agonist therapy, needle and syringe exchange programs, and enhanced HIV testing and treatment services—have the potential to prevent HIV transmission in PWID and improve health outcomes while meeting commonly accepted thresholds for cost-effectiveness. Opioid agonist therapy may reduce HIV transmission and also improve quality of life for people who inject drugs; both effects contribute to its favorable cost-effectiveness.
Some prevention programs for people who inject drugs are controversial, and access is limited. We demonstrate that some of these programs can be both effective and cost-effective investments that have the potential to reduce not only the spread of HIV but also the size of the injecting population.
| Over the past decade, injection drug use, particularly heroin injection, has increased across most US demographic groups, making substance-abuse-related mortality and morbidity a public health crisis [1]. In 2014, there were 47,055 deaths from drug overdose in the US, with almost 30,000 due to opioid overdose [2]. Because HIV spreads relatively efficiently through the transfer of blood in shared injecting equipment [3], people who inject drugs (PWID) account for a disproportionate share of HIV prevalence and incidence in the US [4,5]. Although HIV prevalence and incidence among US PWID have been falling over the past decade [4,6,7], recent growth in the size of the injecting population has raised concerns that HIV risks could rise [8]. Programs targeted to PWID, which have the additional benefit of preventing downstream sexual transmission of HIV to others in the population, are therefore a public health priority. Given the current epidemic of injection drug use in the US, benefits that extend beyond HIV prevention are also a critical consideration [1,8].
A recent empirical study demonstrated that a combined prevention strategy effectively halted HIV epidemics in PWID populations [9]. This strategy included opioid agonist therapy (OAT), which reduces injecting frequency [10,11], needle and syringe programs (NSPs), which reduce injection equipment sharing [12], and enhanced services for HIV testing and treatment (Test & Treat), which identify and virally suppress infected individuals by enrolling them in antiretroviral therapy (ART) [13,14]. Additionally, the US Centers for Disease Control and Prevention (CDC) now recommends daily oral pre-exposure prophylaxis (PrEP), which reduces uninfected individuals’ risk of acquiring HIV, for PWID [15]. Although all of these programs have demonstrated efficacy, they have diverse delivery methods and target populations, as well as different costs and anticipated benefits, and have not been modeled comparatively in a cost-effectiveness context.
The complexity of transmission dynamics and intervention scenarios makes it difficult to deduce a priori the highest value portfolio of prevention programs for PWID. To address this, we extended an empirically calibrated model of the US HIV epidemic [16] to assess the cost-effectiveness of alternative HIV prevention portfolios for US PWID. Each portfolio included some combination of OAT, NSPs, Test & Treat, and PrEP scaled to various coverage levels. Our model integrated clinical, epidemiologic, and economic data and captured the dynamic accrual of total population costs and benefits by tracking the spread of HIV through injection-based and sexual transmission routes.
Our analysis builds on a previously published dynamic compartmental model of the US HIV epidemic [16]. A simplified schematic (Fig 1) illustrates how the model stratifies the adult population aged 18–64 y by HIV infection and awareness status, CD4 count, ART status, OAT status, and risk group. We instantiated the model with US data (Table 1) and calibrated it to match a range of targets, including CDC estimates of US HIV prevalence [5,17–21] and incidence [4,21] across all risk groups. The model consists of a system of differential equations programmed in Matlab R2015b (MathWorks) that track compartment populations monthly from 2015 to 2035. We used a societal perspective to calculate the costs, quality-adjusted life years (QALYs), and incremental cost-effectiveness ratios (ICERs) associated with each portfolio of interventions. Costs and QALYs were measured over the lifetimes of all individuals active in the model over the 20-y time horizon and were discounted at 3% annually [22,23]. We also measured health outcomes such as HIV infections averted and change in HIV prevalence.
The majority of the modeled population is considered low-risk heterosexual. Consistent with CDC estimates, PWID and men who have sex with men (MSM) are smaller populations with higher initial HIV prevalence and HIV-related risk behaviors [5,21]. Between 2015 and 2035, US birth cohorts age into the model when they turn 18 y, and individuals either age out of the actively injecting population at age 65 y or die (at age <65 y) at background mortality rates adjusted for risk behavior [31,48,55], HIV infection, and ART status [39,40,42,43,47]. Additional Markov models follow individuals who mature out and those alive at the end of the 20-y analytic time horizon to capture all lifetime costs and benefits (S1 Appendix, Section 1). The model tracks incident infections, disease progression, HIV screening, enrollment in ART, and transitions into and out of OAT.
Through calibration, the model reflects HIV infection risks given current OAT, NSP, Test & Treat, and PrEP coverage. We assume that PWID, when they do share injecting equipment, are equally likely to share with any other PWID [17,27,28,38,44]. Injection-based HIV transmission depends on the infected partner’s HIV stage [3,41] and ART status [13,14], along with the uninfected partner’s use of PrEP [13]. Sexual mixing patterns approximate partnerships among and between risk groups, with transmission between sero-discordant partners additionally depending on male condom use [18–20,27,29,38,56], condom effectiveness [57], and whether both partners are MSM [3].
Upon infection, an individual enters a brief but highly infectious acute stage [3,41,58], followed by asymptomatic HIV (CD4 count 500 to 1,200 cells/mm3), symptomatic HIV (CD4 count >200 to <500 cells/mm3), and AIDS (CD4 count ≤ 200 cells/mm3) [25,41,47]. As CD4 count falls, infectivity increases [3,41,58] and quality of life decreases [44,59]. ART moderates these effects, reducing injection-based transmission by 59% [13,44] and sexual transmission by 90% [14,41,44], and extends life expectancy by suppressing HIV viral load [40,42,43]. To be eligible for ART, an individual must first be diagnosed with HIV infection, at which time that person may also modify risk behaviors, such as injection equipment sharing or condom usage [18,27,29,38]. In the model, HIV detection rates depend on risk group and are higher in symptomatic HIV compartments [47].
Although the model captures multiple risk groups in order to calibrate to the US HIV epidemic, all interventions in this analysis are directed exclusively to PWID. Program scale-ups are incremental to the status quo, which assumes baseline 2015 levels of OAT, NSP, Test & Treat, and PrEP coverage. Scale-ups to low, medium, and high coverage levels were chosen to standardize comparisons and provide intuition for program costs, benefits, and interactions as coverage increases. In practice, feasible enrollment levels in terms of budget and participant retention are likely to vary by community.
Methadone and buprenorphine are the most common pharmacological therapies prescribed as OAT in the US [10]. We assume that OAT decreases the number of injections by 55% [10,11], thereby reducing overdose risk [27,31] and the chance of HIV transmission, and improving quality of life [27,28,32]. Additionally, we assume that individuals on OAT have higher HIV screening rates than the general PWID population and are more likely to connect to ART services if diagnosed [60,61]. OAT also provides the sole pathway by which individuals permanently cease drug use (3.6% annually) and move to a lower-risk population [28,30]. Previous analyses of the costs and benefits of OAT have consistently found it to be cost-effective [28,62,63].
We assume that 25% of PWID receive OAT under the status quo [17,18,27–29]. At low, medium, and high coverage levels, enrollment increases to 40%, 45%, and 50% of the PWID population, respectively. Such scale-up would involve both short-term investments (e.g., overhead costs for starting methadone clinics) and the long-term costs of delivering the therapies themselves [27,28,33]. In sensitivity analyses, we varied parameters affecting OAT’s effectiveness and cost.
In the model, NSP broadly refers to any of a range of local programs, such as those at pharmacies, hospitals, or designated facilities, through which PWID access sterile hypodermic injecting equipment [64]. We assume that NSPs reduce equipment sharing by 45% [12,34]. Despite being considered a cost-effective HIV prevention strategy [35,65], social and political barriers often prevent NSP expansion [66], and it remains the most controversial of PWID-targeted interventions.
We calibrated our analysis to the current effects of NSPs, which we assume to be minimal at a national level in the status quo [37,67]. Low, medium, and high coverage levels expand NSPs to reach 40%, 45%, and 50% of PWID, respectively. We assume a fixed annual operating budget for NSPs, and low scale-up costs [35–37]. We varied the estimated costs and effectiveness of NSPs in sensitivity analyses.
US guidelines recently eliminated a CD4 count threshold for ART initiation and now recommend immediate ART following diagnosis [68]. US cities adopting this policy have used aggressive HIV testing, same-day treatment initiation, and routine follow-up to significantly increase overall viral suppression [69,70]. Previous analyses of Test & Treat in the US have found favorable cost-effectiveness ratios [44,47,71], and while PWID remain a difficult demographic to reach and sustain in care [70], other community-based interventions targeted exclusively to high-risk populations have demonstrated the potential for sustained case management for PWID [49].
Our status quo reflects current ART engagement levels [27,39,40], with risk group determining the probability that a newly diagnosed individual becomes virally suppressed [40,72]. At low, medium, and high coverage levels of Test & Treat, 40%, 45%, and 50% of infected PWID, respectively, enroll in sustained ART care by 2035. (Because an infection must occur before enrollment in Test & Treat, this intervention, unlike the others, cannot scale immediately in the model.) Costs associated with Test & Treat include those of screening, diagnosis confirmation, counseling, and ART [44,47,48,50,51], as well as the associated costs of comprehensive community-based care programs [49]. We varied the parameters determining enrollment probability (calibrated to additionally reflect the phenomenon of loss to follow-up), program costs, and participant quality-of-life benefits in sensitivity analyses.
Our previous analysis [16] found that PrEP, a daily oral pill of 300 mg tenofovir disoproxil fumarate and 200 mg emtricitabine (Truvada), is most valuable for PWID when delivered per the CDC’s clinical guidelines (e.g., HIV screening every 3 mo, toxicity monitoring every 6 mo) [15] and with prompt and sustained provision of ART for those who do become infected. Nonetheless, PrEP for PWID is expensive in terms of total budget outlay and ICER, although it could be considered cost-effective in the highest prevalence communities [16].
Our status quo assumes negligible PrEP use among PWID in 2015. At low, medium, and high coverage levels, 40%, 45%, and 50% of uninfected PWID, respectively, receive PrEP. The direct costs of PrEP reflect the costs of Truvada [48,53,54] as well as ongoing monitoring costs [48,51]. We assume that the PrEP enrollment process modestly increases HIV screening for the entire PWID population and that individuals diagnosed with HIV discontinue PrEP immediately [15].
Each compartment is associated with an annual cost (adjusted for inflation to 2015 US dollars [73]) and QALY value depending on the characteristics of that subpopulation. A one-time scale-up cost is additionally associated with each intervention (S1 Appendix, Section 3). For every individual in the model, discounted costs and QALYs accrue at each time step, yielding a total cost and QALY estimate for each scenario. We also include lifetime costs and QALYs for individuals maturing out of the population and for individuals alive in the population at the end of the modeled time horizon (S1 Appendix, Section 1). We calculate ICERs by comparing the incremental discounted costs and QALYs for each scenario to the next best alternative [23].
We used a random search algorithm to repeatedly sample from estimated distributions for each model input and then empirically fit the model to US epidemiologic data, resulting in 182 calibrated parameter sets (S1 Appendix, Section 2) [74,75]. Ranges on parameter values are presented in Table 1. We performed all analyses over the calibrated sets to incorporate parameter uncertainty [76]. We present the averages over these sets as our base case results, with 95% confidence intervals where appropriate.
Table 2 compares the four prevention programs implemented singly. We estimate that expansions of OAT, NSPs, and Test & Treat up to coverage levels of 50% can cost less than US$30,000 per QALY gained relative to the next highest coverage level (95% confidence intervals for these programs fall below commonly accepted thresholds of cost-effectiveness, with NSPs having the widest range), while PrEP is likely to cost more than US$300,000 per QALY gained (95% confidence intervals fall above commonly accepted thresholds of cost-effectiveness). Our model estimates that achieving 50% coverage for OAT may avert fewer infections than achieving 50% coverage for PrEP. Because we model nearly immediate direct decreases in mortality rates and increases in quality of life and HIV treatment for those enrolled in OAT, we estimate that expanded OAT coverage is likely to produce higher QALY gains than any other intervention, even when averting fewer infections. An underlying assumption in the model is that PWID must transition through OAT before being “eligible” to cease injection use. At low, medium, and high OAT coverage, respectively, we estimate that the actively injecting population could decrease by 23%, 31%, and 37% over 20 y, accounting for a substantial proportion of the quality-of-life gains. Relatively low delivery costs combined with these benefits have the potential to make OAT the most cost-effective choice among the singly implemented prevention programs.
Figs 2 and 3 and Table 3 illustrate how programs can be combined to construct the highest value prevention portfolio from all considered combinations of OAT, NSPs, Test & Treat, and PrEP. Our base case analysis indicates that scaling OAT coverage up to 50%, then scaling NSP coverage to 50%, then scaling Test & Treat coverage to 50% can be a cost-effective approach to maximizing health benefit, with each additional coverage expansion having the potential to cost less than US$50,000 per QALY gained relative to the next best portfolio. Over 20 y, the combination of high OAT, NSP, and Test & Treat coverage can avert up to 43,400 (95% CI: 23,000, 74,000) infections and decrease HIV prevalence among PWID by 27% (95% CI: 12%, 45%). This is 5,700 more infections and a 1% greater decrease in prevalence than model projections from a program of high PrEP alone. At the same time, our analysis estimates that combinations of OAT, NSPs, and Test & Treat could cost up to US$40 billion less than a high-coverage PrEP program over 20 y.
We conducted multiple sensitivity analyses to assess the importance of uncertainty around model parameters and to evaluate factors important to developing high-value portfolios. We present key insights below, with further details in S1 Appendix, Section 5.
Our findings presented in Figs 2 and 3 and Table 3 are largely insensitive to the majority of one-way sensitivity analyses conducted on model parameters relating to the implementation of each intervention. Several one-way sensitivity analyses conducted on program delivery parameters cause an intuitive interchange in the relative priority of interventions. For instance, at low NSP delivery costs or high efficacy of NSPs, or at high OAT delivery and start-up costs, our analysis suggests that the highest value portfolio would first increase NSP coverage before scaling up OAT. In probabilistic sensitivity analysis (PSA), 41% of sampled sets first add NSPs to the portfolio before adding OAT, while the rest add OAT first. We estimate that additions of OAT and NSPs to the portfolio cost less than US$50,000 per QALY gained in 100% and 74% of all samples considered in PSA, respectively. When the threshold is US$100,000 per QALY gained, NSPs are a cost-effective addition in 93% of PSA samples.
We estimate that increasing Test & Treat coverage could cost less than US$50,000 per QALY gained over a range of delivery costs. We did not identify a one-way sensitivity scenario under which Test & Treat replaces OAT as the most favorable investment, although there are several scenarios in which, unlike in the base case, Test & Treat is a higher priority investment than NSPs. Furthermore, in PSA there were no sampled sets for which the model projected Test & Treat as the first addition to a highest value portfolio. The addition of Test & Treat costs less than US$50,000 per QALY gained in 4.7% of PSA samples, less than US$100,000 in 33% of samples, and less than US$150,000 in 67% of samples. Thus, in terms of cost-effectiveness, our findings indicate that a portfolio of prevention programs could achieve highest value by first investing in OAT or NSPs before scaling up Test & Treat.
Over the majority of one-way sensitivity analyses, we estimate PrEP to cost more than US$500,000 per QALY gained when added to the portfolio. Only when we decrease PrEP’s drug cost by 90% do we project its ICER value to fall below US$100,000 relative to the next best alternative. In PSA, when we vary both the cost and the efficacy of PrEP within estimated, currently feasible ranges, 1% of sampled sets have a highest value portfolio for which the addition of PrEP costs less than US$150,000 per QALY gained.
To further explore the dependence of our results on underlying calibrated parameters, such as ART efficacy, we performed all analyses on a limited subset, the sets containing a parameter value in the bottom or top 5% of all sets, for each parameter. Limiting the analysis to extreme values does not substantially change our findings, nor do other sensitivity analyses on the duration or implementation of interventions (S1 Appendix, Section 5.2). To probe the effects of wide confidence intervals on several parameters, we performed multiple joint sensitivity analyses outside the calibrated context as well as over three calibrated sets for which both OAT’s effectiveness in reducing injection frequency and the decrease in injection equipment sharing following HIV diagnosis (with implications for the effectiveness of Test & Treat) were highly unfavorable (S1 Appendix, Section 5.3). We estimated that in such circumstances NSPs can replace OAT as the priority investment, but OAT, NSPs, and Test & Treat remain cost-effective additions to the portfolio, while PrEP is not likely to be, although its value can increase when other interventions are less favorable.
The opioid epidemic is a global public health burden that has become particularly acute in the US [1,2,8]. In addition to the substantial mortality associated with substance abuse [2], high rates of HIV transmission among PWID make the successful prevention of HIV in this population a public health priority. To that end, we consider portfolios of HIV prevention programs that include OAT, NSPs, Test & Treat, and PrEP scaled to various coverage levels. Although model projections can only provide estimates of health benefits and costs, such analyses can provide intuition around critical mechanisms and assumptions to inform decision making. Our main finding is that, over 20 y, high coverage (enrollment of 50% of the eligible population) of OAT, NSPs, and Test & Treat in combination could avert nearly 43,400 (95% CI: 23,000, 74,000) HIV infections among PWID and reduce HIV prevalence among PWID by 27% (95% CI: 12%, 45%). The construction of such a portfolio has the potential to be cost-effective at each incremental expansion, with projected ICERs below US$50,000 per QALY gained. Moreover, our analysis suggests that the estimated benefit obtainable by PrEP alone (measured in QALYs) could potentially be achieved and even surpassed at substantially lower cost by combining other prevention interventions into high-value portfolios.
Advocates for efficient investment in PWID-specific interventions have asked, “What good is preventing HIV if we do not first save that life at HIV risk?” [77]. Our analysis suggests that the high competing mortality risks of PWID can explain why interventions that immediately improve quality of life can have substantially higher estimated benefits than those that focus on HIV prevention alone. Our analysis estimates that OAT, in particular, which we assume has a direct impact on the length and quality of life of treated individuals [27,28,30–32,60,61], can provide substantially more benefit, measured in QALYs, than other interventions, even when it prevents fewer infections (Table 2).
Although our analysis did not identify a scenario in which OAT was not a cost-effective addition to a high-value portfolio, deterministic and probabilistic sensitivity analyses can provide intuition regarding scenarios in which NSPs could replace OAT as the priority investment. Because the assumed delivery cost of NSPs is so much lower than that of other programs, our findings suggest that it is reasonable to invest in NSPs concurrent with OAT scale-up. While Test & Treat is often estimated in our analysis to be a cost-effective addition to the portfolio, our model does not project it to be a priority investment. Our estimates for ART’s reduction of transmission risk via injection-based contact [13,44] are lower than those for sexual contact [14,41,44], which may explain our projection of smaller benefits in the PWID population. It should also be noted that HIV prevalence in US PWID is less than 10% [18], and the direct QALY increases from Test & Treat programs were therefore low relative to programs that served the entire PWID population.
Costs and cost-effectiveness are but one factor among several in the decision to provide prevention interventions. Policymakers and clinicians may decide that considerations of ethics and social justice outweigh economic considerations for this vulnerable population. PrEP, for instance, can provide benefits for PWID and should not be denied on the basis of injection drug use. Moreover, our findings are based on the current cost of PrEP in the US. If the cost of PrEP were substantially reduced, its cost-effectiveness could become more favorable. Nonetheless, as policymakers address the broader epidemic, our findings suggest that increasing the availability of a full spectrum of prevention interventions would have the highest health and economic benefits.
Our analysis assumes that each intervention is an available option, which is not true in many settings. Treatment and prevention programs for PWID remain controversial, and interventions may be infeasible for reasons beyond budgetary impact. Despite evidence of effectiveness and cost-effectiveness, a current ban in the US prevents federal funding for NSPs [67], and although a previous study found no significant correlation between neighborhood crime and treatment centers [78], proposals for new methadone clinics often face community opposition. Moreover, nearly 60% of individuals on methadone in the US receive insufficient dosing [79]. For this reason, the model’s calibrated sets reflected a wide range of possible effectiveness levels of methadone. If barriers to OAT access were lowered and treatment offered at international evidence-based standards, our analysis suggests that the value of this already cost-effective intervention could increase.
Our analysis has several limitations (Table E in S1 Appendix). First, although our model captures dynamic interactions between programs, we implement combinations of programs as independent. This means that the efficacy and cost of each program does not depend on the efficacy, cost, and impact of other programs at the individual level (for a PWID enrolled in multiple programs) or the population level; that is, declining HIV incidence from an already implemented intervention does not change the delivery cost of another intervention. Rather, our parameters reflect average, aggregate effects in the targeted population. As policymakers move toward a “one stop shop” approach [66], there may be synergies between programs that we do not account for. At the same time, our analysis likely overestimates cost in such a situation by counting the same overhead multiple times. PrEP’s cost, however, comes primarily from the drug itself and is unlikely to decrease if combined with other services. As our analysis suggests that OAT, NSPs, and Test & Treat can be cost-effective even without combining overhead costs or incorporating synergies on an individual level, modeling programs as interdependent would likely not change our general ranking of programs, although it might increase value in an absolute sense if, for instance, adherence to any one program could be improved with multiple enrollment.
Second, certain limitations are inherent to our choice of model. Although extensive sensitivity analysis allows us to investigate parameter sensitivity, the use of a dynamic compartmental model prevents the exploration of structural sensitivities [80,81]. Because compartmental models do not track individuals, we do not explicitly model such phenomena as loss to follow-up in HIV care, although we do calibrate linkage rates to account for long-term drop-offs in the care cascade [27,39,40]. Our model does not explicitly account for networks or distinguish risk on an individual basis. All of the interventions we consider would be more cost-effective if targeted to individuals central to injecting or sexual networks. Because an individual’s decrease in needle sharing must also affect the number of shared needles of his or her injecting partners, the effects of NSPs, in particular, may be underestimated by a compartmental model, where we can only estimate average effectiveness for the individual accessing the service. However, as we find NSPs to be cost-effective under most circumstances, and as we find the most substantial benefit to come from direct, individual health gains accruing from OAT, independent of network effects, this simplification is not likely to alter our model’s general findings.
Third, we assume constant returns to scale (i.e., per person costs do not increase as coverage expands) even though, in practice, enrolling more marginalized PWID, especially at higher coverage levels, is likely more costly. While diminishing returns as coverage rises may decrease the value of investments in absolute terms, the relative ranking of programs, assuming that marginalized PWID are equally difficult to enroll in any program, would not be affected. Moreover, our framing of low, medium, and high coverage levels is meant to be illustrative of possible investment patterns, but our prioritization rankings and cost-effectiveness conclusions do not change with arbitrary definitions of “high” coverage (S1 Appendix, Section 5.2.3). In practice, achievable coverage levels may vary extensively by region. However, our cost-efficient frontier (Fig 2) suggests that OAT investments along a range of feasible coverage levels can provide high value, even when such levels are below 50%.
Finally, many of our model parameters are uncertain. To address this uncertainty, we calibrated our model to empirical data and conducted extensive sensitivity and uncertainty analyses (S1 Appendix, Section 5). Although these analyses suggest that our main findings are consistent across many scenarios and analyses, there remains uncertainty about HIV transmission dynamics and the effectiveness of prevention programs when implemented jointly. Our findings should be interpreted in view of these limitations.
Our model-based analysis builds intuition for the mechanisms behind the conclusions of Des Jarlais et al. [9], who found that a combined prevention approach was effective in ending HIV epidemics among PWID in a number of settings. We also project a range of scenarios under which these combined portfolios can be cost-effective. Where budgets are limited, our analysis suggests that a reasonable approach is for resources to be allocated first towards expanding OAT coverage, which we assume can have additional quality-of-life benefits for PWID beyond HIV prevention, including cessation of drug use. Combining NSP scale-up with OAT expansion, and investing remaining budget resources in further NSP expansion and in Test & Treat, has the potential to be both cost-effective and beneficial to the entire population. While budgets may dictate the extent to which programs can be scaled, we project that the relative ranking among modeled programs remains consistent across varied delivery contexts. Investment in cost-effective programs is critical given the current epidemic of injection drug use in the US [1,8]. Although further empirical studies of combined prevention programs would be very useful, model-based projections can inform the development of high-value HIV prevention portfolios.
|
10.1371/journal.pntd.0006641 | A prospective cohort study comparing household contact and water Vibrio cholerae isolates in households of cholera patients in rural Bangladesh | Household contacts of cholera patients are at a 100 times higher risk of developing cholera than the general population. The objective of this study was to examine the incidence of V. cholerae infections among household contacts of cholera patients in a rural setting in Bangladesh, to identify risk factors for V. cholerae infections among this population, and to investigate transmission pathways of V. cholerae using multilocus variable-number tandem-repeat analysis (MLVA).
Stool from household contacts, source water and stored water samples were collected from cholera patient households on Day 1, 3, 5, and 7 after the presentation of the index patient at a health facility. Two hundred thirty clinical and water V. cholerae isolates were analyzed by MLVA. Thirty seven percent of households had at least one household contact with a V. cholerae infection. Thirteen percent of households had V. cholerae in their water source, and 27% had V. cholerae in stored household drinking water. Household contacts with V. cholerae in their water source had a significantly higher odds of symptomatic cholera (Odds Ratio (OR): 5.49, 95% Confidence Interval (CI): 1.07, 28.08). Contacts consuming street vended food had a significantly higher odds of a V. cholerae infection (OR: 9.45, 95% CI: 2.14, 41.72). Older age was significantly associated with a lower odds of a V. cholerae infection (OR: 0.96, 95% CI: 0.93, 0.99). Households with both water and clinical V. cholerae-positive samples all had isolates that were closely related by MLVA.
These findings emphasize the need for interventions targeting water treatment and food hygiene to reduce V. cholerae infections.
| Household members of cholera patients are at a 100 times higher risk of developing cholera infections than the general population. This risk is highest during the seven days after the cholera patient presents at a health facility. In this study we investigated the rate of cholera transmission within cholera patient households, identified risk factors for household cholera transmission, and performed genetic characterization of cholera strains collected. Stool was collected from patients, their household members, and from water sources and stored water during the seven days after the cholera patient presented at the health facility. A total of 230 human and water V. cholerae strains were collected and analyzed. Thirty seven percent of households had at least one household member with a V. cholerae infection. Thirteen percent of households had V. cholerae in their water source, and 27% had V. cholerae in stored drinking water. A water source with V. cholerae, consuming street vended food, and younger age were risk factors for cholera infections for household members of cholera patients. All strains from within households with water and human samples were closely related. These results demonstrate the importance of interventions focusing on water treatment and food hygiene for prevention of cholera.
| The World Health Organization estimates that there are 95,000 cholera deaths per year with 2.9 million cases worldwide [1]. Studies have identified risk factors for becoming infected with cholera such as age [2], drinking street-vended water [3], placing ones hands into stored household water [4], bathing in a river [4, 5], eating leftover food [6], eating food prepared by a recently ill food handler [7], not washing hands with soap before eating food [8], and being a first degree relative of a cholera case [9]. These findings indicate that water and food borne contamination are the main transmission routes for V. cholerae infection.
Previous studies in Bangladesh have demonstrated that household contacts of cholera patients are at a much higher risk of developing a V. cholerae infection than the general population [2, 10, 11]. The average rates of cholera in Bangladesh are 1.6 cases per 1000 individuals [1], while two studies in rural Matlab, Bangladesh found 240 V. cholerae infected individuals per 1000 household contacts of cholera cases [1, 5, 10]. Most recently, a study in urban Dhaka, Bangladesh, found 210 V. cholerae infected individuals per 1000 household contacts of cholera cases [2]. The highest risk for V. cholerae infections among household contacts is within 7 days of the onset of symptoms in the index case [2, 5, 12]. However despite this high risk, there has been little work done to determine the main transmission routes of V. cholerae infection for this population. This is likely because it is difficult to elucidate whether cholera transmission to household contacts is from an external source that is shared by household members such as a piped water supply, or by a cholera case infecting family members by contaminating household food or water.
Multilocus variable-number tandem-repeat analysis (MLVA) is a method to distinguish between different strains of V. cholerae that are typically indistinguishable by methods such as pulse field gel electrophoresis (PFGE) and multilocus sequence typing (MLST) [13–15]. MLVA records the number of repeating sequences found in short DNA fragments at five loci in the genome. Different isolates vary in the number of tandem repeats at each locus thus providing a fingerprint to differentiate between isolates [16]. Our recent study found that V. cholerae isolates with the same MLVA genotype had significantly fewer pairwise differences by whole genome sequencing (WGS) compared to isolates with different MLVA genotypes [17]. This is consistent with findings from Rashid et al. which found that isolates closely related by MLVA had significantly fewer nucleotides differences by WGS than isolates distantly related by MLVA [18].
In our recent study in urban Dhaka, Bangladesh which followed cholera patient households during the one week high risk period for V. cholerae infections after the presentation of the index case, we found that stored household drinking water with V. cholerae and a median free available chlorine concentration below 0.5 mg/L were associated with V. cholerae infections among household contacts of cholera patients [19]. MLVA and WGS were performed to investigate cholera transmission patterns. The findings showed a combination of person-to-person and water-to-person cholera transmission with the proportions of the two modes varying within and between outbreaks [17].
Building on this previous urban work, our current study focuses on cholera patient households in a rural setting in Bangladesh. Our objective was to examine the incidence of V. cholerae infections among household contacts of cholera patients in a rural setting, to identify risk factors and investigate transmission pathways for V. cholerae infections using MLVA. This approach allows for intervention strategies to be identified that can be used to reduce the incidence of cholera among household contacts of cholera patients.
Informed consent was obtained from all study participants, and study procedures were approved by the research Ethical Review Committee of the International Centre for Diarrhoeal Disease Research, Bangladesh (icddr,b) and the Johns Hopkins Bloomberg School of Public Health IRB.
This prospective cohort study was conducted in rural Bakerganj and Mathbaria upazilas in Barisal district of Bangladesh from April 2015 to June 2016. Suspected cholera patients were defined as patients presenting at the Bakerganj and Mathbaria upazila health complexes with acute watery diarrhea (3 or more loose stools over a 24 period). The stool samples from these patients were screened for the presence of V. cholerae using the Crystal VC Rapid Dipstick test (Span Diagnostics, Surat, India) [20, 21]. All positive findings by dipstick were confirmed by bacterial culture. Cholera patients were defined as diarrhea patients with a stool bacterial culture result positive for V. cholerae. Screening and study recruitment at Bakerganj and Mathbaria upazila health complexes occurred Saturday to Thursday each week during the study period. The final sample size was based on the number cholera patients that were recruited between April 2015 to June 2016. A cluster was defined as the index cholera patient and their corresponding household contacts. Household contacts were defined as individuals sharing the same cooking pot as the index cholera patient for the previous three days. To be eligible for the study household contacts had to plan to reside in the same household as the index cholera patient for the next week. Eligible household contacts present in the health facility at the time of patient enrollment were invited to participate, and a household visit was made to recruit household contacts within 36 hours of patient enrollment.
Cholera patient households were visited at Days 1, 3, 5, and 7 (Visits 1–4) after the presentation of the index cholera patient at the health facility for clinical and environmental surveillance. For clinical surveillance, household contacts were asked if they had diarrhea (3 or more loose stools over a 24 hour period) or vomiting in the past 48 hours, and a stool sample was collected from willing household contacts at each household visit to test for the presence of V. cholerae in stool by bacterial culture. For environmental surveillance, water samples were collected from the household’s water source and stored drinking water in the home at each visit to test for the presence of V. cholerae by bacterial culture. A spot checks was also conducted to observe if soap was present near the latrine and cooking areas of households (within ten steps) as a proxy measure of handwashing with soap behavior, and to assess if household stored drinking water was completely covered [22]. In addition, a structured questionnaire was administered to obtain information on household and individual characteristics.
Stool samples were collected in stool cups from cholera patients and household contacts and water samples were collected in 500 mL bottles. Fecal specimens were enriched in alkaline peptone water (APW) broth for six hours, streaked on Thiosulphate Citrate Bile Sucrose Agar (TCBS) and Taurocholate Tellurite Gelatin Agar (TTGA) plates, and incubated overnight. Serotyping was performed according to previously published methods [23]. Water samples were filtered through 0.22 micron polycarbonate membrane filters and then enriched in APW and cultured as previously described [24]. Two or more colonies were selected from each sample.
MLVA was performed on DNA from 230 V. cholerae water and clinical isolates from 27 cholera patient households (139 clinical isolates and 91 water isolates). For three households, there was no MLVA data available. DNA was isolated from 5 μl of culture using Prepman (ABI) according to the manufacturer’s instructions. To perform MLVA, the DNA from the V. cholerae O1 isolates was genotyped at each of five previously identified MLVA loci (VC0147, VC0437, VC1650, VCA0171 & VCA0283) using previously published methods [13]. An MLVA genotype was defined by the alleles at each locus. Genetic relatedness was defined by the similarity of alleles at MLVA loci. If an allele on the large chromosome (VC0147, VC0437, and VC1650) was found to be missing after MLVA analysis, we used the available information from the large chromosome to impute missing values using SAS (version 9.3). If the matching sequences were variable beyond 2%, we assumed that the missing allele could not be deduced and these alleles were not imputed. For locus VC0147, 3 (1% of 230) alleles were imputed, while 99 (43%) and 54 (23%) were imputed for VC0437 and VC1650, respectively. No alleles were imputed for the small chromosome.
Our primary outcomes were: (1) the incidence of cholera infected household contacts defined as an individual with a culture result positive for V. cholerae, and (2) the incidence of household contacts with symptomatic V. cholerae infections, defined as a V. cholerae infection with diarrhea or vomiting. Logistic regression models were performed to estimate the odds of developing a V. cholerae infection with household and individual level covariates using generalized estimating equations (GEE) to account for clustering within households and approximate the 95% confidence intervals (CI). If there were no V. cholerae infections in one of the categories, a chi square test was performed.
All analyses were performed using SAS, version 9.4 (SAS Institute Inc., Cary, NC, USA). Pairwise comparisons were made of the number of allele differences in the five locus genotype (e.g. this would be one if a single locus varied) [17]. Fisher’s exact, paired t-tests, and permutation tests were computed using SAS (version 9.3) to analyze MLVA data.
During April 2015 to June 2016, we screened 1081 diarrhea patients presenting at Mathbaria and Bakerganj health facilities using the Crystal VC rapid dipstick test for V. cholerae. Fifty-one diarrhea patients had positive results by dipstick, 5 of these individuals refused to participate in our study, and 16 were culture negative for V. cholerae. All 30 dipstick positive and culture confirmed cholera patients were enrolled in our cohort study. Seventy-six household contacts from these 30 cholera households patients were enrolled (S1 Dataset). Twenty-three households were from Bakerganj and 7 households were from Mathbaria. We observed 3 cholera outbreaks: Outbreak 1 (April -June 2015); Outbreak 2 (October 2015), and in Outbreak 3 (May-June 2016). All three outbreaks had households from both Mathbaria and Bakerganj. Forty-seven percent of index cholera patients (14) and 64% (49) of household contacts were female (Table 1). The mean age was 26 years for index cholera patients and 22 years for household contacts. The mean number of individuals in the household was 5. Eighty-seven percent of households (26) did not completely cover their stored drinking water during the surveillance period (assessed by spot checks), and only 10% (3) reported boiling their household drinking water during the surveillance period. Seventy-nine percent (23) of households had no soap present in the latrine area during the surveillance period and 97% (28) had no soap present in the kitchen area (assessed by spot checks). Twenty-seven percent of households (8) had unimproved latrines using the World Health Organization/ United Nations Children's Fund Joint Monitoring Programme definition [25]. Unimproved sanitation options include pit latrines without a slab or platform, hanging latrines, bucket latrines, and flying toilets. Improved sanitation options include ventilated improved pit latrines, pit latrines with slabs, composting toilets, and flush/pour flush latrines/toilets to piped sewer systems, or septic tanks. Ninety-seven percent of households reported that groundwater was their primary drinking water source, and one household reported pond water and was using a pond sand filter. Seventy-eight percent (59) of household contacts reported consuming water outside the household and 87% (66) reported consuming food outside the household during the surveillance period. Forty-seven percent (36) of household contacts reported eating street vended food during the surveillance period.
All V. cholerae strains belonged to serotype Ogawa; and all possessed the cholera toxin gene, ctxA. Thirty-seven percent of households (11) had at least one household contact with a V. cholerae infection, with 13% (4) of households having an infected household contact on the first household visit (Table 2). Twenty percent of households (6) had a household contact with a symptomatic V. cholerae infection, defined as a V. cholerae infection accompanied with diarrhea or vomiting in the past 48 hours. Eighteen percent (14) of household contacts had a V. cholerae infection during the surveillance period and 8% (6) had a symptomatic infection. Five household contacts had a V. cholerae infection on Visit 1, six on Visit 2, one on Visit 3, and two on Visit 4. Five household contacts had two visits with a stool specimen positive by bacterial culture for V. cholerae.
Among the 76 household contacts of cholera patients, 28% (21) were the mother of the patient, 24% (18) were a sibling, 12% (9) were the father, 9% (7) were a grandparent, 8% (6) were a spouse, and 20% (15) were another relative. Among those household contacts with a V. cholerae infection, 4 (29%) were mothers of the index patient, 4 were a sibling (29%), 2 were a spouse (14%), 2 were the father (14%), one was a grandparent (7%), and one was an uncle (7%). During the surveillance period, 15 household contacts reported taking antibiotics, only 5 of these individuals had a symptomatic V. cholerae infections. One household contact with a symptomatic V. cholerae infection was referred to a health facility and required IV fluids.
Thirteen percent of households (4) had V. cholerae in their source water during the surveillance period (all groundwater sources), and 27% (8) had V. cholerae in stored household drinking water. Seventeen percent (5) of households had V. cholerae in stored household drinking water on the first visit, and 10% (3) had detectable V. cholerae in their source water on the first visit. All households with detectable V. cholerae in their water source had unimproved latrines. All households with V. cholerae in their source water also had V. cholerae in stored drinking water. Source water with V. cholerae only occurred in Outbreak 2. Stored water isolates were from Outbreaks 1 and 2.
Household contacts consuming street vended food had a significantly higher odds of a V. cholerae infection (Odds Ratio (OR): 9.45, 95% confidence interval (CI): 2.14, 41.72) (V. cholerae infections: 33% (consumed street vended food) vs. 5% (did not consume street vended food)) (Table 3). Older age in years was significantly associated with a lower odds of a V. cholerae infection (OR: 0.96, 95% CI: 0.93, 0.99). No household contacts that reported boiling their drinking water during the surveillance period had a V. cholerae infection compared to 21% of household contacts that did not report boiling their drinking water (p = 0.18). Twenty-one percent of household contacts that reported consuming food outside of the home had a V. cholerae infection compared to no infections for those not consuming food outside the home (p = 0.11). V. cholerae infections among contacts were not associated with the presence of soap in the kitchen or latrine area.
Contacts with V. cholerae in their source water had a significantly higher odds of a symptomatic V. cholerae infection (OR: 5.49, 95% CI: 1.07, 28.08) (25% (V. cholerae in water source) vs. 6% (no V. cholerae in water source)) (Table 4). Nine percent of contacts residing in households with no soap in the kitchen area had a symptomatic V. cholerae infection compared to no V. cholerae infections among household contacts with soap present in the kitchen area (p = 0.49). All symptomatic V. cholerae infections occurred among contacts residing in households with unimproved sanitation options (p = 0.12).
A total of 230 clinical and water V. cholerae isolates were compared by MLVA: 139 clinical isolates from 40 stool samples from cholera patients and their household contacts; 31 source water isolates from 7 source water samples; and 60 stored water isolates from 11 stored water samples (S2 Dataset). These isolates were collected from 27 households across 3 outbreaks: 16 Households in Outbreak 1; 9 Households in Outbreak 2, and 2 Households in Outbreak 3. We identified 31 MLVA genotypes: 7 MLVA genotypes from both clinical and water isolates, 9 genotypes from only water isolates, and 15 genotypes from only clinical isolates. There were 3 alleles at VC0147, 3 at VC0437, 4 at VC1650, 6 at VCA0171, and 6 at VCA0283. There were multiple MLVA genotypes among isolates collected from a single sample. This was similar for clinical samples (mean: 1.9 MLVA genotypes, range: 1–3) and water samples (mean: 2.1 genotypes, range: 1–4), p = 0.53. Eighty one percent of clinical samples (25/31) had at least two isolates with different MLVA genotypes compared to 83% (15/18) of water samples.
Eight households had both clinical and water isolates. When the relatedness of water and clinical isolates from the same household was compared, all households had at least one clinical and water isolate with an identical MLVA genotype or a single locus variant of the same MLVA genotype. However only one household had all clinical and water isolates with identical MLVA genotypes or single locus variants of the same MLVA genotype. For the four households with a positive source water and stored water sample, all had source water and stored water samples with identical MLVA genotypes or single locus variants of the same MLVA genotype. Among the ten households with multiple infected household members, nine out of ten had isolates from different household members with identical MLVA genotypes or single locus variants of the same MLVA genotype. Five out of ten of these households had all clinical isolates with identical MLVA genotypes or single locus variants of the same MLVA genotype.
Isolates collected from the same household had significantly fewer pairwise differences in MLVA loci than those from different households (mean: 0.98 pairwise differences in MLVA loci (same household) vs. 1.79 (different household), p<0.0001). Isolates from the same outbreak also had significantly fewer pairwise differences than those collected from different outbreaks (mean pairwise differences: 1.33 (same outbreak) vs. 2.10 (different outbreak), p<0.0001). When comparing clinical and water isolates, the number of pairwise differences were significantly higher for water compared to clinical isolates (mean 1.79 (water isolates) vs. 1.69 (clinical isolates), p<0.0001).
Nearly 40% of cholera patients had a household member with a V. cholerae infection during the surveillance period, and 18% of household contacts overall were infected. V. cholerae was detected in both groundwater and stored water in patient households. Significant risk factors for V. cholerae infections among household contacts of cholera patients were the presence of V. cholerae in drinking water sources, consuming street vended food, and younger age. The genetic characterization of V. cholerae isolates from cholera patient households showed a high diversity of MLVA genotypes within and between clinical and water samples, and all water and clinical samples within the same household had V. cholerae isolates that were closely related. These findings emphasize the need for interventions targeting water treatment and food hygiene to reduce V. cholerae infections among contacts of cholera patients.
Eighteen percent of household contacts of cholera patients had a V. cholerae infection in our rural setting in Bangladesh. This is similar to previous studies conducted in rural Matlab, Bangladesh which found 23% and 24% of household contacts of cholera patients to be V. cholerae infected [5, 10]. Our findings are also similar to the 19% of household contacts infected in our recent urban cohort of cholera patient households in Dhaka, Bangladesh [26].
Household contacts using drinking water sources with V. cholerae were significantly more likely to have symptomatic V. cholerae infections, with 13% of tube wells having detectable V. cholerae. This finding is consistent with Hughes et al. and Spira et al. conducted in rural Bangladesh where households using a water source positive for V. cholerae were significantly more likely to have V. cholerae infections [5, 10]. In Hughes et al. 33% of tube well water samples were positive for V. cholerae, while in Spira et al. no tube wells had detectable V. cholerae by bacterial culture only surface water samples [5, 10]. In our current study twice as many stored household water samples were positive for V. cholerae compared to source water samples (27% vs. 13%). This finding suggests high rates of household contamination of stored water. This result is in contrast to Spira et al. which found similar V. cholerae concentrations in stored and water source samples (23% vs. 26%) [10]. Hughes et al. did not analyze stored water samples for V. cholerae. However, this previous study did find that water from cooking, bathing, and washing dishes was often contaminated with V. cholerae and that this contamination was a significant risk factor for V. cholerae infections in cholera patient households [5]. Future studies should test all water sources utilized for household tasks for V. cholerae, not only the household primary drinking water source and stored household drinking water.
Only households with unimproved sanitation options had detectable V. cholerae in their tube wells. This could be because cholera patients and their infected household members were using these unimproved sanitation options which were contaminating nearby tube wells, or alternatively there could be biofilm growing within tube well pipes from stored household water being used to prime wells. In rural Bangladesh tube wells are often located adjacent to household latrines, making these water sources more susceptible to fecal contamination. Consistent with this an intervention trial conducted in the Philippines found that communities with improved sanitation options had significantly fewer symptomatic V. cholerae infections compared to communities with no improved water or sanitation access [27]. Furthermore, when sanitation facilities were combined with improved drinking water sources reductions in cholera were doubled compared to sanitation alone.
Both our previous urban cohort study and current rural cohort study found drinking water to be a significant risk factor for V. cholerae infections among household contacts [19]. However, in our urban setting stored water was a significant risk factor for V. cholerae, while the source water was significant in our rural setting. This was unexpected given that there was a higher proportion of stored water samples with detectable V. cholerae in our rural compared to urban site (27% vs. 6%). One potential explanation for this is that households with V. cholerae in source water had a higher overall burden of fecal contamination in their households, likely from unimproved household latrines.
Street vended food was a sigificant risk factor for V. cholerae infections in our rural cohort. This is consistent with studies from South Africa, Guatemala, Nigeria, and India where street vended food and water were risk factors for V. cholerae infections. [3, 8, 28, 29] This is likely because of a recently ill food handler contaminating food or water as was found in rural Micronesia. [7] Younger age was also found to be associated with an increased risk of V. cholerae infections in our study as was previously shown in rural and urban cohort studies of cholera patient households in Bangladesh. [2, 5] In Weil et al. being less than 14 years of age was a risk factor for V. cholerae infections, while in Hughes et al. children 5–9 years of age were at highest risk. This association is likely because of young children lacking the naturally acquired immunity to cholera found in older individuals previously exposed.
We observed substantial diversity in MLVA genotypes in clinical and water samples. There were significantly more pairwise differences in water samples compared to clinical samples. This is consistent with the findings from our urban cohort study in Dhaka, Bangladesh. [17] While in Rashed et al. there was greater diversity in clinical isolates compared to water isolates in rural Bangladesh.[30] However this was likely attributed to the low number of water isolates collected.
Vibrio cholerae isolates from the same household were more closely related than isolates from different households. This finding is consistent with the source of V. cholerae infections within cholera patient households being either a shared environmental source in the household such as the drinking water source or street vended food, or person-to-person transmission through poor hygiene practices in the home. Consistent with water-to-person transmission, households with both water and clinical V. cholerae-positive samples all had isolates that were closely related by MLVA. Our findings are consistent with our urban cohort study where isolates from the same household were also more closely related than those from different households, and the majority of households had water and clinical isolates that were closely related [17]. In support of person-to-person transmission or a shared contaminated source in the household, we found that the vast majority (90%) of infected household members had closely related MLVA genotypes. This is consistent with our urban cohort study which found 82% of household member isolates with identical MLVA genotypes [17]. Future studies are needed that perform whole genome sequencing of water and clinical isolates from cholera patient households in this rural setting to further elucidate transmission pathways for V. cholerae infections, and these studies should include sampling of all water sources used for household tasks.
This study has several strengths. The first is the rural setting, since recent household contact studies in Bangladesh have all been conducted in urban settings [2, 26]. Second, we collected multiple isolates from all samples allowing us to investigate the diversity of MLVA genotypes within samples. Third, we performed intensive clinical and environmental surveillance that included collecting stool specimens from all enrolled household contacts, not only those presenting with diarrhea or vomiting, and included sampling of water sources and stored household drinking water. Fourth, we included both a risk factor analysis and genetic characterization of water and clinical isolates collected from cholera patient households.
Our study also had a few limitations. First, our sample size was small. We had fewer cholera patients than anticipated during our study period. Second, we did not perform whole genome sequencing on collected isolates which would have provided a higher level of resolution to distinguish the genetic relatedness of isolates collected. In our recent cohort study, however, we found that isolates with the same MLVA genotype were also closely related by whole genome sequencing, with significantly less pairwise differences in single nucleotide-variant counts than isolates with different MLVA genotypes [17]. Third, we did not include community control households. This would have allowed us to estimate the odds of V. cholerae infections for household contacts of cholera patients compared to community control contacts.
Vibrio cholerae in drinking water sources, consuming street vended food, and younger age were important risk factors for V. cholerae infections among household contacts of cholera patients in our rural setting in Bangladesh. Consistent with the findings from our risk factor analysis, the genetic characterization of strains from cholera patient households showed that the majority of water and clinical samples within the same household had isolates with closely related MLVA genotypes. These results highlight the urgent need for water treatment and food hygiene to reduce V. cholerae infections among highly susceptible household contacts of cholera patients.
|
10.1371/journal.pntd.0004066 | Factors Associated with Non-typhoidal Salmonella Bacteremia versus Typhoidal Salmonella Bacteremia in Patients Presenting for Care in an Urban Diarrheal Disease Hospital in Bangladesh | Non-typhoidal Salmonella (NTS) and Salmonella enterica serovar Typhi bacteremia are the causes of significant morbidity and mortality worldwide. There is a paucity of data regarding NTS bacteremia in South Asia, a region with a high incidence of typhoidal bacteremia. We sought to determine clinical predictors and outcomes associated with NTS bacteremia compared with typhoidal bacteremia.
We performed a retrospective age-matched case-control study of patients admitted to the Dhaka Hospital of the International Centre for Diarrhoeal Disease Research, Bangladesh, between February 2009 and March 2013. We compared demographic, clinical, microbiological, and outcome variables of NTS bacteremic patients with age-matched S. Typhi bacteremic patients, and a separate comparison of patients with NTS bacteremia and patients with NTS gastroenteritis.
Of 20 patients with NTS bacteremia, 5 died (25% case fatality), compared to none of 60 age-matched cases of S. Typhi bacteremia. In univariate analysis, we found that compared with S. Typhi bacteremia, cases of NTS bacteremia had more severe acute malnutrition (SAM) in children under five years of age, less often presented with a duration of fever ≥ 5 days, and were more likely to have co-morbidities on admission such as pneumonia and clinical signs of sepsis (p<0.05 in all cases). In multivariable logistic regression, SAM, clinical sepsis, and pneumonia were independent risk factors for NTS bacteremia compared with S. Typhi bacteremia (p<0.05 in all cases). Notably, we found marked differences in antibiotic susceptibilities, including NTS strains resistant to antibiotics commonly used for empiric therapy of patients suspected to have typhoid fever.
Diarrheal patients with NTS bacteremia more often presented with co-morbidities and had a higher case fatality rate compared to those with typhoidal bacteremia. Clinicians in regions where both typhoid and NTS bacteremia are prevalent need to be vigilant about the possibility of both entities, especially given notable differences in antibiotic susceptibility patterns.
| Salmonella are a group of bacteria that cause illnesses and death worldwide. There are two types of Salmonella–Typhi and non-typhoidal (NTS). In humans, the majority of illnesses caused by NTS are related to gastro-intestinal problems, though uncommonly, it also invades the bloodstream. On the other hand, typhoid fever caused by Salmonella typhi commonly invades the bloodstream. Since the treatment of the two types may differ, we wanted to compare the risk factors for each. We studied patients who had NTS or Typhi isolated from blood in a diarrheal hospital in Bangladesh. We observed that patients with NTS bloodstream infection frequently presented with severe malnutrition, clinically diagnosed sepsis and pneumonia compared to those with typhoid fever. We also found that NTS and Typhi differed in what antibiotics they were sensitive to. These observations may help our clinicians to initiate aggressive treatment from the very beginning of the illness in children with NTS bacteremia in order to attain better outcomes.
| Non-typhoidal Salmonella (NTS) are a group of Gram negative bacteria known to cause disease in both animals and humans worldwide. They include all Salmonella enterica spp. except for S. enterica serovar Typhi, Paratyphi A, Paratyphi B, and Paratyphi C. In humans, NTS are responsible for an estimated 94 million cases of gastroenteritis each year globally, causing upwards of 150,000 deaths [1]. NTS is also associated with systemically invasive disease and bacteremia, especially in immunocompromised hosts [2]. In sub-Saharan Africa, NTS is a common cause of bacteremia in both adults and children, especially in areas of high HIV and malaria prevalence [2]. In contrast, the burden of invasive NTS disease in many areas of Asia is thought to be much less than that of sub-Saharan Africa. One multicenter community-based fever surveillance study detected only 6 cases of invasive NTS from over 20,000 blood cultures [3]. Despite this, rates of NTS bacteremia are increasing in Asia, most notably in HIV positive populations [4–6]. Due to the low incidence of disease, there is a paucity of data regarding the clinical presentation, risk factors, resistance patterns, and outcomes for NTS bacteremia in Asia, particularly in areas of low HIV prevalence such as Bangladesh. Given the high incidence of Salmonella enterica serovar Typhi infection, a cause of typhoid fever, in this region, comparison and contrast of NTS bacteremia with S. Typhi bacteremia may be clinically important, especially with regards to empiric therapy. Also, comparing invasive with non-invasive NTS disease may provide important data for understanding disease pathogenesis.
The aim of this study was to characterize the demographics, clinical presentation, resistance patterns and clinical outcomes of patients admitted to a medical facility in a large urban center in Bangladesh with NTS bacteremia, and to compare them to those associated with S. Typhi bacteremia. A secondary objective was to compare cases of NTS bacteremia with those of NTS gastroenteritis.
This study was approved by the Research Review and the Ethical Review Committees of the icddr,b and the Institutional Review Board of Massachusetts General Hospital.
We retrospectively extracted data from the electronic charting system of the icddr,b Dhaka Hospital, which is a hospital providing care free of charge to approximately 120,000 patients per year, most of whom present with diarrhea. The country of Bangladesh has a HIV prevalence of <0.1% [7], and malaria is not known to occur in the city of Dhaka.
We identified patients admitted to the icddr,b from February 2009 to March 2013 who had Salmonella spp. isolated from either blood or stool cultures. We identified all patients with NTS bacteremia (blood culture positive for Salmonella spp. other than S. Typhi and S. Paratyphi), and as a control group, selected age-matched patients with S. Typhi bacteremia at a 3:1 ratio. Previous studies from Nepal, Indonesia, and Bangladesh have demonstrated that the clinical presentations of S. Typhi and S. Paratyphi infection are nearly indistinguishable [8–10]. Thus, to simplify the analysis, we have excluded patients with blood culture positive for S. Paratyphi. In a secondary analysis, we identified patients who had NTS isolated from stool and whose blood culture did not grow any organisms.
Microbiologic culturing of venous blood was performed in patients at the discretion of the attending physician. Blood was seeded directly into BacT/ALERT culture bottles and entered into the BacTAlert 3D system.
Microbiologic culturing of stool was performed as part of a systematic surveillance of every 50th admitted patient, and also for inpatients at the discretion of the attending physician[11]. Stool was cultured using standard isolation methodology from a single, fresh stool specimen collected from the patient [12].
Antibiotic susceptibility testing was performed using Disk Diffusion Method. The detailed procedure has been described elsewhere. [13] Susceptibility pattern was interpreted by using CLSI guideline.[14]
Pneumonia was diagnosed by following the World Health Organization (WHO) criteria for children under 5 years of age [15]; in patients older than 5 years of age, diagnosis of pneumonia was based on the attending physicians’ interpretation of clinical signs and symptoms and admission chest x-ray, if performed.
Clinical sepsis was defined as presence or suspected presence of infection, plus any two of the following: 1) hypo- (≤35.0°C) or hyperthermia (≥38.5°C), 2) abnormal age-adjusted white blood cell (WBC) count or >10% bands, 3) tachycardia (defined as heart rate above the upper normal limit according to age), 4) tachypnea (defined as respiratory rate above the upper normal limit according to age), and 5) abnormal cognition [16].
Acute kidney injury was defined as a creatinine level greater than the upper limit of normal according to age.
Fever was defined as axillary temperature more than 37.8°C.
Diarrhoea is defined as having loose or watery stools at least three times per day, or more frequently than normal for an individual. [17]
Acute diarrhea was defined as passage of three or greater number of abnormally loose or watery stools in the preceding 24 hours [17].
Acute diarrhea was defined as new onset of diarrhea in a person without a history of diarrhea in the past 14 days
Invasive diarrhea was defined as presence of WBC >20/HPF and any number of red blood cells which was detected by routine and microscopic examination of stool.[18]
Persistent diarrhea was defined as an episode of diarrhea, with or without blood, lasting at least 14 days. [17]
Severe acute malnutrition (SAM) was defined following WHO anthropometry as previously described [19].
Abnormal WBC count was defined as a WBC count outside of the reference values according to the age (0 to 1 month: 6000-36000/cmm; 6 month to 3 years: 6000-17500/cmm; 4 to 10 yrs: 5500-14500/cmm).
Hypokalaemia was defined as a serum potassium level below the reference value (3.5–5.3 mmol/L).
Hyponatraemia was defined as a serum sodium level below the reference value (135–146 mmol/L).
Hypocalcemia was defined as a serum calcium level below the reference value (2.12–2.16 mmol/L).
De-identified clinical and laboratory data were collected and analyzed using Statistical Package for Social Sciences (SPSS), Windows (Version 17.0; Chicago, IL) and Epi Info (Version 1.0.3, USD, Stone Mountain, GA).
In the period February 2009 to March 2013, a total of 12,940 blood cultures were collected from patients admitted to the icddr,b Dhaka hospital. Of these, 20 were positive for NTS, 567 for S. Typhi, and 64 were for S. Paratyphi (all were Paratyphi A). We compared the clinical and microbiological data from the 20 NTS patients to age-matched controls with S. Typhi bacteremia in a 1:3 ratio (60 S. Typhi patients). We also identified 27 patients from whom NTS was isolated from stool and not blood (24 blood culture negative, 3 without blood cultures checked). We observed that patients with NTS bacteremia had a significantly higher case fatality rate compared to those with S. Typhi bacteremia (25% vs. 0%, p<0.001). The age and causes of death in the five who died with NTS bacteremia are displayed in Table 1.
We compared the clinical and demographic characteristics of patients with NTS bacteremia with age-matched controls with S. Typhi bacteremia. We found that cases of NTS bacteremia were more likely to have SAM, concurrent pneumonia, and clinical sepsis, and less likely to have fever of 5 days or more than those with S. Typhi (Table 2). In logistic regression analysis, after adjusting for potential confounders, SAM, pneumonia, and clinical sepsis were independent risk factors for NTS bacteremia (Table 3).
When examining laboratory characteristics on admission, we found that age-matched patients with S. Typhi bacteremia were more likely to have hypokalemia, hypocalcemia, and a lower hematocrit in comparison to patients with NTS bacteremia. NTS bacteremic patients were more likely to present with increased creatinine level, and an abnormal (high or low count in age-specific range) white blood count (WBC)
We observed that the susceptibility patterns to conventional antibiotics differed between NTS and S. Typhi isolates (Table 4). We found that some strains of NTS had full or intermediate resistance to antibiotics used in the empiric treatment of typhoid fever, such as ceftriaxone and azithromycin.
We also examined the differences in clinical characteristics between patients with NTS bacteremia and those who had diarrhea and NTS isolated from stool but not blood. We found that patients with NTS bacteremia were more likely to have a history of drinking unsafe water, to present with fever, to have concurrent pneumonia, and clinical sepsis. However, the significant statistical differences were lost after Holm-Bonferroni correction (Table 5). The serogroups of NTS isolated are shown in Table 6. We found no significant difference between invasive and non-invasive strains with regard to serogroup.
Invasive NTS infections are a cause of significant morbidity and mortality worldwide, most notably in sub-Saharan Africa where it is associated with HIV, malnutrition, and malaria. Here we report the characteristics of patients presenting with invasive NTS to a diarrheal hospital in Dhaka, Bangladesh, where high rates of malnutrition, but not HIV or malaria, are present. We found that NTS bacteremia is a rare occurrence, and that compared with patients with S. Typhi bacteremia, patients with NTS bacteremia had a higher in-hospital fatality rate, were more likely to be malnourished, and to have concurrent pneumonia and clinical sepsis.
NTS bacteremia is an extremely uncommon finding at our diarrheal hospital in urban Bangladesh. NTS was isolated in less than 0.2% of all blood cultures, and less than 1% of positive blood cultures. On the other hand, S. Typhi accounted for 22% (567/2,573) of all positive blood cultures. We observed that 5 of 20 patients with NTS bacteremia died during hospitalization. This high fatality rate is consistent with studies from Sub-Saharan Africa [20], and underscores the importance of this pathogen despite the lower incidence seen in Asia. When we examined the age distribution of patients with NTS bacteremia, including those who died, we found that the many of the cases were infants and the elderly, age groups known to be of highest risk for invasive NTS. However, we also found cases of NTS bacteremia in older children and middle-aged adults, including in two of the five deaths.
A recent systematic review has identified S. Typhi as the most common community-acquired blood stream infection in South and Southeast Asia among both adults and children [21]. Previous studies have reported differences in age distribution between individuals with NTS and S. Typhi bacteremia [22]. Thus, we examined the factors differentiating NTS and S. Typhi bacteremia through an age-matched case control study using both univariate and multivariable regression analysis. We showed that in children under 5 years of age, SAM was more prevalent in patients with NTS bacteremia than those with S. Typhi bacteremia, and in multivariate regression analysis, SAM remained an independent predictor for NTS bacteremia. Studies from Africa have also suggested malnutrition to be a risk factor for NTS infection, compared with non-bacteremic hospitalized children [23,24], and compared with non-Salmonella bacteremia [25]. We confirm that this risk factor remains when compared to age-matched patients with S. Typhi. The reasons for this is unknown, but we suspect that the immunocompromised state of children with SAM may be a greater predisposing factor for NTS than for S. Typhi, which has not been associated with any immunodeficient state. Studies from both developed and developing countries have shown NTS to be more often invasive in immunocompromised patients than the otherwise healthy [26], and the role of the IL-17 axis has been attributed to differences between NTS and S. Typhi infection [27]. Furthermore, studies of humans with IL-12p40 deficiency have revealed an increased susceptibility for invasive NTS infection [28], suggesting that IL-12 may play a critical role in protection against NTS. Indeed, IL-12 expression is diminished in malnourished children compared to well-nourished controls [29], and may contribute to their susceptibility to invasive NTS infection.
We showed in our logistic regression analysis that NTS bacteremic patients were more likely to present with concurrent pneumonia and clinical signs of sepsis. Previous studies have shown that both children and adults with NTS bacteremia have co-infection of the lower respiratory tract [23,30]. Given the similarity in clinical signs and symptoms of bacterial pneumonia and NTS bacteremia [31], differences in microbial etiology and choice of empiric antibiotics, and the high mortality rate of NTS, care must be taken to consider the presence of NTS bacteremia in cases of lower respiratory tract infection where risk factors of NTS are present.
Previous reports from our institution have shown that drinking unboiled water is a risk factor for typhoid fever [32], but that the type of water source is not a risk factor for NTS infection [33]. In this study, we show that patients with S. Typhi bacteremia are more likely to drink unboiled water than those with NTS bacteremia. This may be a reflection of the large role that contaminated food and water plays in the acquisition of typhoid (a human-restricted pathogen), and does not rule out the possibility that NTS is acquired through contaminated food and water.
The higher proportion of NTS patients presenting with clinical signs of sepsis, acute kidney injury, and abnormal WBC count, together with associated mortalities due to septic shock, are all likely a reflection of the higher severity of illness on presentation in NTS bacteremia compared to S. Typhi bacteremia. Despite the higher severity of illness, we found that patients with NTS bacteremia are more likely to present with a shorter duration of fever on admission, similar to findings from a study of Tanzanian children in a malaria-endemic region [34]. The reasons for this are unclear, but may include factors such as higher rates of malnutrition among NTS patients and respiratory co-infections as described above. We hypothesize that patients with NTS blood stream infection may progress to severe disease more quickly than those with S. Typhi. On the other hand, duration of fever was not an independent predictor based on multivariable regression, and previous studies suggest that NTS and typhoid fever may have similar duration of fever.[35,36] Thus, studies with larger sample sizes are needed to clarify whether severity of illness on presentation is clearly different between the two entities.
Patients presenting with S. Typhi bacteremia also had a lower hematocrit than those with NTS. This is in contrast to findings from the Tanzanian study, which did not account for the contribution of age (NTS patients younger) nor malaria co-infection. Hematological changes are common in typhoid fever and bone marrow suppression and hemophagocytosis are considered to be an potential mechanisms for such effects [37]. We hypothesize that the higher proportion of anemia and electrolyte imbalances seen in S. Typhi-infected patients may also be indicative of the more insidious nature of typhoid fever and lengthier time prior to presentation to hospital.
The resistance patterns of blood NTS isolates in this report are comparable to previous reports of stool NTS isolates from Bangladesh [33], including high rates of intermediate resistance to ciprofloxacin, and lower rates of resistance to ampicillin, chloramphenicol, and TMP/SMX. However, there were several differences between the antimicrobial sensitivity patterns of isolated NTS and S. Typhi strains. Most notably, we found strains of NTS isolated from bacteremic patients that are resistant to ceftriaxone, and two with intermediate resistance to azithromycin. All strains of S. Typhi isolated were fully susceptible to the above agents, consistent with previous reports [36,38,39], and ceftriaxone (or oral equivalent) and azithromycin are commonly used for empiric therapy of patients suspected to have typhoid fever. While we did not find significant differences in time-to-appropriate antibiotic between the two groups, the differing susceptibility patterns of NTS and S. Typhi may have implications for empiric treatment of children presenting with fever.
Our study has several notable limitations. First, this is a hospital-based study and thus our findings can only be generalized to individuals in the population whose symptoms were severe enough to seek medical care. Also, since the majority of patients presenting to our diarrheal hospital have diarrhea, our findings may be biased towards those who prominently presented with diarrhea. Secondly, this was a retrospective study of a population where not all admitted patients had blood cultures drawn, and thus there is a selection bias towards those who had clinical suspicion for bacteremia. Thirdly, we did not test patients for HIV, though the prevalence of HIV in Bangladesh is low at < 0.1% (7). Fourthly, we only used conventional culture methods to detect Salmonella infections, as assays using other methodologies were not routinely available. Fifthly, the numbers of NTS bacteremia cases are low (n = 20), our confidence intervals are quite wide, and larger studies are needed to confirm our findings. We matched by age so that we could compare other features; however, our approach could also have introduced limitations for overall analysis by infecting pathogen. We also focused our analysis of enteric fever to that caused by S. Typhi, and did not include enteric fever caused by S. Paratyphi A in our analysis, given previous reports demonstrating their similarities in clinical presentation. Lastly, as we had extracted the data from the hospital EMR, we were not able to ascertain pre-admission use of antimicrobials, which may influence the severity and duration of presenting symptoms. Despite these limitations, our study is the largest analysis of NTS bacteremia in South Asia and largely matches findings from sub-Saharan Africa,[22,40] Given the influence of HIV and malaria co-infection on NTS infection in sub-Saharan Africa, we believe that our report is significant and of particular import to those working in Asia.
In conclusion, we have compared clinical and microbiological characteristics of patients with NTS and S. Typhi bacteremia at a diarrheal hospital in Bangladesh, and found a higher incidence of comorbidities on presentation and a higher fatality rate among those with NTS bacteremia. These results, combined with a different antibiotic susceptibility profile than S. Typhoid isolates, warrant further investigation into the epidemiology of invasive NTS infections in South Asia, and surveillance of resistance patterns of both non-typhoidal and typhoidal Salmonella infections in this region.
|
10.1371/journal.pntd.0003343 | Comparative Effectiveness of Different Strategies of Oral Cholera Vaccination in Bangladesh: A Modeling Study | Killed, oral cholera vaccines have proven safe and effective, and several large-scale mass cholera vaccination efforts have demonstrated the feasibility of widespread deployment. This study uses a mathematical model of cholera transmission in Bangladesh to examine the effectiveness of potential vaccination strategies.
We developed an age-structured mathematical model of cholera transmission and calibrated it to reproduce the dynamics of cholera in Matlab, Bangladesh. We used the model to predict the effectiveness of different cholera vaccination strategies over a period of 20 years. We explored vaccination programs that targeted one of three increasingly focused age groups (the entire vaccine-eligible population of age one year and older, children of ages 1 to 14 years, or preschoolers of ages 1 to 4 years) and that could occur either as campaigns recurring every five years or as continuous ongoing vaccination efforts. Our modeling results suggest that vaccinating 70% of the population would avert 90% of cholera cases in the first year but that campaign and continuous vaccination strategies differ in effectiveness over 20 years. Maintaining 70% coverage of the population would be sufficient to prevent sustained transmission of endemic cholera in Matlab, while vaccinating periodically every five years is less effective. Selectively vaccinating children 1–14 years old would prevent the most cholera cases per vaccine administered in both campaign and continuous strategies.
We conclude that continuous mass vaccination would be more effective against endemic cholera than periodic campaigns. Vaccinating children averts more cases per dose than vaccinating all age groups, although vaccinating only children is unlikely to control endemic cholera in Bangladesh. Careful consideration must be made before generalizing these results to other regions.
| Bangladesh has a high burden of cholera and may become the first country to use cholera vaccine on a large scale. Mass cholera vaccination may be hard to justify to international funding agencies because of the modest efficacy of existing vaccines and their limited duration of protection. However, mass cholera vaccination can induce high levels of indirect protection in a population, i.e., protecting even unvaccinated individuals by lowering cholera incidence, and a case for cost-effective cholera vaccination could be made. Mathematical modeling is one way to predict the magnitude of indirect protection conferred by a proposed vaccination program. Here, we predict the effectiveness of various mass cholera vaccination strategies in Bangladesh using a mathematical model. We found that maintaining high levels of vaccination coverage in children could be very effective in reducing the burden of cholera, and secondary transmission of cholera would virtually stop when 70% of the population is vaccinated. Mathematical modeling may play a key role in planning widespread cholera vaccination efforts in Bangladesh and other countries.
| Vibrio cholerae, the bacterium responsible for clinical cholera, has long been associated with the Bay of Bengal where it exists as an autochthonous member of the estuarine ecosystem [1], [2]. This area of south Asia has been the origin for six of the seven cholera pandemics, and the burden of disease remains high. Today cholera is endemic in much of the Ganges River Delta with an estimated 350,000 treated cases per year in Bangladesh alone [3]. Improvements in water, sanitation, and hygiene are the long-term solutions for cholera, but oral cholera vaccines (OCVs) may constitute a shorter-term option to reduce morbidity and mortality from the disease.
Oral cholera vaccines are safe and effective [4]–[6]. A recent large field trial in Kolkata, India, has shown that Shanchol, one of two World Health Organization prequalified OCVs, provides 65% protection over 5 years [5]. Further, successful demonstration campaigns conducted by the International Centre for Diarrhoeal Disease Research, Bangladesh (icddr,b) in urban and rural communities show promise for expanding vaccination coverage in Bangladesh [4], [7]. The expanded use of OCV as a component of cholera control is supported by the recent decision by the World Health Organization to establish a global stockpile of 2 million doses of OCV, and the GAVI Alliance has committed to finance and leverage support for the global stockpile through 2018 [8], [9]. Although OCV will likely be more widely used in the coming years, its effectiveness at a population level is not well understood. This information is crucial for planning vaccination programs on a large scale.
It can be difficult to justify the widespread use of OCVs on economic grounds because: OCVs confer only moderate protection for a few years [5], [10], the incidence of cholera in most settings is relatively low, and the number of deaths attributed to cholera is relatively small because of the availability of inexpensive and effective treatment. However, large vaccine trials have shown that as OCV coverage increases, indirect protection from vaccination, also known as herd protection, increases [11], [12]. When indirect protection is considered, the effectiveness of mass cholera vaccination can be high, and accounting for the effects of indirect protection appears to be necessary to make OCVs cost effective in the developing world [3], [13]–[15]. Mathematical models can be used to predict the effectiveness of mass cholera vaccination, including indirect effects [16]–[20], so modeling may be an essential component of any economic case for cholera vaccination. An earlier modeling study found that vaccinating 50 to 70% of the population of Bangladesh would virtually eliminate transmission [16]. Here, we expand upon that work, using a mathematical model to predict the effectiveness of targeting different age cohorts for vaccination at various coverage levels and schedules over a 20-year horizon.
Matlab is a rural community of approximately 220,000 people 30 kilometers southeast of Dhaka [21]. Data on cholera cases in Matlab were collected during long-term passive surveillance as described in detail elsewhere [22]. Briefly, between 1997 and 2001, twice a month for three days a study physician attended to all patients presenting with acute watery diarrhea at the icddr,b clinic. Following orally obtained informed consent, these patients were tested for V. cholerae by rectal swab, and cultured on the same day in the Matlab laboratory [22], [23]. Patient data, including age and sex, were obtained with identifying information removed. The Committee on Human Research of the Johns Hopkins University Bloomberg School of Public Health approved the research, and its guidelines were followed in the conduct of the clinical research.
We developed an age-structured mathematical model of cholera transmission. Compartments in the model are unvaccinated susceptible (S), vaccinated susceptible (V), symptomatically infected (I), asymptomatically infected (A), or recovered and immune (R) from cholera (Figure 1). The concentration of V. cholerae in the environment (water) is tracked in an additional compartment (W). Susceptible individuals may become infected by direct contact with infected individuals (direct transmission) or by exposure to V. cholerae in the environment (indirect transmission). A complete description of the model is given in Text S1.
The model aggregates the population in compartments by disease status and age. Age cohorts represent children under 2 years old, pre-school aged children (2 to 4 years old), school aged children (5 to 14 years old), and adults (15 years old and older). Younger age groups are assumed to be more susceptible to infection [24]. Births are modeled by adding unvaccinated susceptibles to the youngest age cohort each year, and deaths are modeled by removing individuals from all age cohorts. Birth and age-specific mortality rates were based on data from the Matlab Health and Demographic Surveillance System [21]. Cohorts are aged by moving individuals into the next older age compartment at the appropriate rates.
Frequency-dependent transmission rates are assumed for infections acquired through short cycle transmission (person-to-person) while a saturation (Holling type II [25]) function in terms of cholera concentration in water (W) is used to model the force of infection from long cycle transmission (environmental exposure). A fraction p of the infections are symptomatic, a fraction r of which seek treatment. We refer to r as the reporting rate. The asymptomatically infected individuals (proportion 1-p of all infections) are less infectious and shed bacteria into the environment at a lower rate than cholera cases. Throughout this paper, “cholera cases” refers to the number of symptomatically infected individuals. The reporting rate of cholera cases is set to 10% in the main scenario and to 25% in alternative scenarios to test the sensitivity of results to this parameter [26].
Infected individuals recover after five days on average and are immune to infection until they transition back to the susceptible state after an average of 3 years [27], [28]. An alternative scenario, assuming different duration of natural immunity protection across age groups, is also investigated.
The model is calibrated to fit the dynamics of cholera cases recorded between 1997 and 2001 in Matlab. The proportion of recovered individuals at the beginning of 1997 is estimated based on time-series data of cholera incidence in Matlab [29], the assumed reporting rate, and the duration of natural immunity. Two periods of increased environmental transmission occur annually with the first peak occurring in spring (approximately April to May) followed by a larger peak in autumn (approximately September to November) [30], [31]. An iterative fitting procedure is implemented in which one cholera season (1997–1998) is simulated to estimate: i) the initial distribution of the recovered individuals by age groups by running the model for 5 years and rescaling back to the estimated overall recovered proportion; ii) the transmission rates for both short and long cycle transmission by fitting the number of monthly symptomatic infections based on data for reported cases; and iii) the relative susceptibility of each age group by fitting the observed age distribution of cholera cases. Next, the estimated transmission rates are used to estimate the magnitude of elevated environmental risk during spring and autumn periods as well as the start time of the autumn period by fitting the monthly cases reported between 1998 and 2001. The resulting best fit of the dynamics, minimizing the residual sum of squares for the number of reported cases per month, are presented in Figure 2. Parameters used in the model are presented in Table S1 in Text S2 and the values of the estimated parameters are presented in Tables S2 & S3 in Text S2. The magnitude of the elevated environmental risk during peak periods is sampled from the aggregated ranges in Table S3 in Text S2, which represent the variation fitted over 5 consecutive annual cycles.
Vaccinated susceptibles in the model are protected for an average of five years. Adults and children 5 years and older are 65% less likely to become infected upon exposure to cholera than unvaccinated individuals and children 1–4 years are 40% less likely to become infected [5]. OCVs may decrease the probability of developing symptoms upon infection [32], but insufficient trial data is available to include this effect in the model. Therefore, we took a conservative approach and assumed that upon infection, vaccinated individuals have the same probability of becoming symptomatic as and are as infectious as non-vaccinated individuals. The model was implemented in Matlab R2012a (The MathWorks, Inc.).
We modeled vaccination programs that target one of three age groups: the entire vaccine-eligible population (those one year old and older), all children (ages one to fourteen years), and preschoolers (ages one to four years). We did not model the vaccination of those under one year old, since no vaccine is currently licensed for that age group [33]. The age structure of the model does not precisely match the age cohorts targeted for vaccination. The age cohorts in the model were chosen to match the age groups in the available epidemiological data for calibration. The age cohorts for vaccination were chosen to match current vaccine licensing and logistical considerations. Vaccination of one-year-olds in all scenarios is modeled by targeting half of the population younger than 2 years old. If cholera vaccines are later licensed for use in infants (i.e., under one year old), one could vaccinate a larger fraction of the youngest age cohort in the model.
We modeled three distinct schedules for vaccinating these target populations: one-time campaign, periodic campaigns, and continuous vaccination. For the one-time campaign, a proportion of the targeted population is vaccinated at the start of the first year only. For the periodic campaigns, every five years a proportion of all susceptible and recovered individuals are vaccinated. The period between campaigns was chosen to match the duration of vaccine protection. The continuous vaccination strategy is an approximation of an annual vaccination program. In this strategy, a proportion of the targeted population is vaccinated at the beginning of the first year, then starting in the second year the unvaccinated susceptible and recovered individuals are vaccinated at a fixed rate for the duration of the simulation. A detailed description of the implementations of all vaccination strategies in the model is in Text S1.
We define the overall effectiveness of mass vaccination to be the number of cholera cases prevented (i.e., the difference in the number of cases in a simulation without vaccination and the number of cases in a simulation with mass vaccination) divided by the number of cases when there is no vaccination [34]. We measure the efficiency of a mass vaccination strategy by the number of vaccinations per case averted (VPC), calculated as the number of people who are vaccinated divided by the number of cholera cases prevented.
Seasonal cholera transmission was simulated in a rural population in Bangladesh using a mathematical model calibrated to reproduce the two annual peaks (Figures 2A and S1) and the age distribution of cases observed in surveillance from the community of Matlab (Figure 2B). We found that infants and children younger than two years old, preschoolers (ages 2 to 4 years) and school children (ages 5 to 14 years) were 6.3, 5.2, and 1.8 times more susceptible than adults, respectively, to best fit the data assuming the same duration of immunity after infection across ages (Table S2 and Figure S2 in Text S2). We also estimated that asymptomatically infected individuals were 15% as infectious as the symptomatic cholera cases assuming that 20% of all infected individuals become symptomatic and 10% of cholera cases are reported.
We compare the effectiveness over 20 years of one-time mass vaccination, recurring campaigns every five years, and continuous vaccination targeting 70% of all individuals one year old and older (Figure 3). All three vaccination strategies avert about 94% of the cholera cases in the first year (Figures 3B and 3C). Vaccination of 50% of the population would reduce the incidence of cholera by 88% in the first year following vaccination (Figure S3 in Text S2). This is consistent with projections from a previous modeling study that found vaccination coverage of 50% would be sufficient to avert 93% of cholera cases in one season in Matlab [16]. With a one-time mass vaccination campaign, cholera incidence rebounds as protection from vaccine wanes and new susceptible individuals are born, and the overall effectiveness of the campaign is only about 20% after 20 years (Figure 3C).
Because protection conferred by vaccination lasts five years, one might choose to conduct campaigns once every five years for logistical reasons. However, susceptibility in the population accumulates between campaigns and the proportion of the population protected by vaccine drops to 20%–25% due to the waning of vaccine efficacy and the birth of new susceptible individuals (Figure 3A). Vaccination campaigns every five years could result in 70% overall effectiveness over 20 years but cholera incidence oscillates and peaks in the years preceding each campaign (Figure 3B).
To avoid the fluctuations in vaccination coverage associated with 5-year campaigns, we modeled continuous vaccination in which people are vaccinated at a constant rate throughout the year every year after year 1. When calibrated to use nearly the same amount of vaccine as the 5-year campaigns (Figure 3D), 58% of the population is always protected by vaccine (Figure 3A). When the population is continuously vaccinated, cholera incidence remains low over the 20 years with overall effectiveness above 95% (Figure 3C), and onward cholera transmission is essentially interrupted after ten years. The continuous strategy achieves 25% higher overall effectiveness than the 5-year campaigns (Figure 3C) while using slightly less vaccine over 20 years (Figure 3D).
We compared the effectiveness of targeting different age groups with campaigns every five years. Our modeling results suggest that vaccinating everyone (100% of) one year old and older at 5-year intervals would prevent 89% of cholera cases over 20 years (Figure 4A, red boxes). The efficiency of the 5-year campaigns decreases with higher coverage, with the number of vaccinations per case averted (VPC) rising from 11 to 14 (Figure 4B). Mass vaccination of all children 1 to 14 years old at 5-year intervals would prevent approximately 33% of cholera cases (Figure 4A, blue boxes) while vaccinating all preschoolers would prevent only 6% of cholera cases over 20 years (Figure 4A, green boxes). Because the proportion of the population protected by vaccine drops between campaigns, this vaccination strategy is not able to suppress cholera activity over 20 years, even at 100% coverage (Figure 4A). Targeting children (1 to 14 years old) is most efficient; requiring about 11 VPC over a wide range of vaccination coverage levels (Figure 4B). Targeting those 1 to 4 years old is less efficient, primarily because of the lower vaccine efficacy in this group.
Continuous vaccination is associated with higher overall effectiveness than the 5-year campaigns. Coverage above 70% of the general population is sufficient to virtually interrupt onward transmission of cholera (Figure 4C). VPC associated with the continuous vaccination declines with increasing coverage when children (1 to 14 years old) are targeted. However, VPC increases, thus efficiency decreases, for coverage over 70% when the general population is vaccinated because effectively all cholera transmission is prevented above this level (Figure 4D).
If vaccine efficacy in young children were as high as that in adults, then children ages 1 to 4 years old would be the most efficient age group to target, and vaccinating 70% of them every 5 years would have a VPC of 7, and maintaining 70% coverage with continuous vaccination would have a VPC of 6.5 (Figure S4 in Text S2). If the entire vaccine-eligible population were targeted, then this vaccine would be associated with only a modest increase in overall effectiveness compared to the vaccine with lower efficacy in young children (Figure S4 in Text S2). If the vaccine confers protection for only 3 years instead of 5 and has 65% efficacy among all age groups, one could achieve effectiveness similar to the 5-year campaigns described above by vaccinating every three years, but more vaccine would be required (Figure S5 in Text S2).
Simulated cholera epidemics are sensitive to the assumed proportion of cases that seek treatment. An alternative scenario was calibrated assuming 25% of cholera cases seek treatment in Matlab, resulting in a lower underlying disease burden than the main analysis, which assumed a 10% reporting rate. This alternative scenario projects substantially smaller epidemics and consequently stronger impact of all vaccination programs (Figures S6 and S7 in Text S2). Approximately 30% coverage was enough for the 5-year campaigns and continuous vaccination to eliminate 90% of cholera cases. The same reduction is achieved by campaign vaccination of 80% or continuous vaccination of 60% of children (1 to 14 years old). However, the projected recovered fractions (Figure S6B in Text S2) for all age groups are substantially lower compared to the extrapolations based on the Matlab data (Figure S2 in Text S2), which argues against the plausibility of this high reporting rate. We also modeled an alternative scenario in which children are protected against cholera for a shorter time than adults after infection. In this model, the fraction of susceptible individuals in each age group differed from those seen when all individuals become susceptible an average of three years after infection, and mass vaccination was somewhat more effective (Figures S8–S9 in Text S2).
We used a mathematical model to explore the potential effectiveness of mass cholera vaccination in rural Bangladesh and believe that the results apply more broadly to cholera endemic areas in Bangladesh. With the model, we were able to predict the overall effectiveness, which includes indirect effects, of different mass vaccination strategies. Our results indicate that maintaining 60% or higher vaccine coverage in the population would stop cholera transmission, which is consistent with an earlier modeling study [16]. However, a continuous vaccination schedule might be difficult to implement, as it requires a constant effort to keep a substantial proportion of the population protected by vaccine by identifying unvaccinated individuals and vaccinating them and revaccinating individuals as protection from vaccines wane. The continuous vaccination strategy as described is a mathematical idealization of vaccination efforts that occur throughout the year rather than vaccination campaigns that occur every few years. Vaccination campaigns that occur only once a year would maintain approximately the same level of vaccine protection in the population while being more logistically practical. We also model mass vaccination campaigns that occur once every five years, the average duration of protection from vaccination [5]. This strategy might be easier to implement, but as vaccine protection wanes and birth (and possibly immigration) introduces new susceptible individuals to the population between campaign years, large cholera outbreaks can occur.
We found that vaccinating all vaccine-eligible children, ages 1–14 years, requires the fewest number of vaccinations per case averted compared to vaccinating preschool-aged children (1–4 years) or the general population (ages 1 year and older). Although preschool-aged children have the greatest burden of cholera, represented by both disease incidence and mortality [24], [35], selectively vaccinating this group is the least efficient strategy, primarily due to the lower modeled vaccine efficacy in this age group (40%) [5].
Delivering OCV to children could build upon existing delivery mechanisms like the Expanded Programme on Immunization (EPI) or National Immunization Days [3]. However, our analysis suggests that endemic cholera is unlikely to be eliminated by vaccinating only children. A major consideration of immunization plans could be the most efficient use of the limited supply of doses, currently around two million globally but anticipated to expand over the next five years [8], [9]. Vaccinating populations with the highest risk of disease is efficient and also supported by evidence from cost-effectiveness analyses, the priorities of decision makers, and health equity considerations [3], [15], [19], [36].
Previous cost-effectiveness studies have found that untargeted mass cholera vaccination in Bangladesh may not be effective unless one accounts for indirect protection [3], [13]. However, the magnitude of indirect protection is difficult to estimate without proper studies and/or mathematical modeling. The results from this study suggest significant indirect protection from OCV that may improve the economic case for expanding its use. Although we do not evaluate the cost-effectiveness of the modeled vaccination strategies, the number of vaccinations per case averted can be used to estimate cost-effectiveness. If it requires between 10–25 vaccinations per case averted (Figures 4B, 4D) and costs $5.33 to fully vaccinate an individual (two doses at a public sector cost of $1.85 per dose [3] and a delivery cost of $1.63 per individual [4]), the vaccination programs considered here would cost between $53–$133 per cholera case averted, and assuming a 1.5% case fatality rate [3], $3,500–$8,900 per death averted.
There are several limitations to this study. The model was calibrated to match the demographic and epidemiologic characteristics of cholera in Matlab, Bangladesh; so extrapolating the results from this study to other settings requires careful consideration. We modeled transmission of cholera in an endemic setting where the incidence is much higher in children than adults. However, data from cholera outbreaks in non-endemic settings suggest a more even distribution of cholera incidence by age [24], [28], [37]. Therefore, the results described apply to regions that experience annual cholera outbreaks at a scale similar to Matlab's, but the model should be recalibrated for settings with substantially different epidemiology or demography. The actual cause of the relatively high cholera incidence among children in Bangladesh is not known, but it has been hypothesized to be due to higher levels of previous exposure in adults and differences in the immune system in children and adults [38]–[40]. The model assumptions required to create this differential susceptibility could affect the effectiveness of mass vaccination. We assumed that the differences in cholera incidence between age groups were due largely to differential intrinsic susceptibility. We had tested an alternative hypothesis that the duration of immunity conferred by infection differed between age groups and could explain the differences in incidence, but this model also required differential susceptibility to fit the data from Matlab.
The model was not intended for the prediction of cholera activity for a particular year. We calibrated the model using five years of data from Matlab, assuming that the epidemiology of cholera will not change substantially, so the 20-year projections described here should be considered average outcomes over this time horizon. We assumed that current demographic trends, sanitation levels, and climate would remain constant over the next 20 years, but changes in population movement, development, rainfall, the frequency of severe flooding events, sea level, and ocean temperature could change the epidemiology of cholera [41]–[44].
There is growing momentum toward incorporating oral cholera vaccine into cholera control and outbreak response planning. Field and feasibility trials have been conducted in urban and rural Bangladesh and there appears to be interest to include targeted OCV use as part of comprehensive cholera control strategies [3], [4], [33], [45], [46]. As Bangladesh and other countries begin to consider the role of OCV in comprehensive cholera control plans, this work provides insight into how OCV may diminish cholera transmission dynamics. This analysis demonstrates that mass immunization with oral cholera vaccines may greatly reduce the burden of disease, and mathematical modeling can provide guidance on the targeting of populations and scheduling of campaigns to maximize impact.
|
10.1371/journal.pntd.0003838 | Acceptability and Willingness-to-Pay for a Hypothetical Ebola Virus Vaccine in Nigeria | Ebola virus disease is a highly virulent and transmissible disease. The largest recorded fatality from Ebola virus disease epidemic is ongoing in a few countries in West Africa, and this poses a health risk to the entire population of the world because arresting the transmission has been challenging. Vaccination is considered a key intervention that is capable of arresting further spread of the disease and preventing future outbreak. However, no vaccine has yet been approved for public use, although various recombinant vaccines are undergoing trials and approval for public use is imminent. Therefore, this study aimed to determine the acceptability of and willingness-to-pay for Ebola virus vaccine by the public.
The study was a community-based cross-sectional qualitative and quantitative interventional study conducted in two communities, each in two states in Nigeria. An interviewer-administered questionnaire was used to collect information on respondents’ knowledge of the Ebola virus, the ways to prevent the disease, and their preventive practices, as well as their acceptability of and willingness-to-pay for a hypothetical vaccine against Ebola virus disease. The association between acceptability of the vaccine and other independent variables were evaluated using multivariate regression analysis.
Ebola virus disease was considered to be a very serious disease by 38.5% of the 582 respondents (224/582), prior to receiving health education on Ebola virus and its vaccine. Eighty percent (80%) accepted to be vaccinated with Ebola vaccine. However, among those that accepted to be vaccinated, most would only accept after observing the outcome on others who have received the vaccine. More than 87.5% was willing to pay for the vaccine, although 55.2% was of the opinion that the vaccine should be provided free of charge.
The level of acceptability of Ebola virus vaccine among respondents was impressive (though conditional), as well as their willingness to pay for it if the vaccine is not publicly funded. In order to achieve a high uptake of the vaccine, information and education on the vaccine should be extensively shared with the public prior to the introduction of the vaccine, and the vaccine should be provided free of charge by government.
| Ebola virus disease (EVD) is highly virulent and transmissible. The transmission is mostly by direct contact with an infected person or indirectly through contact with material contaminated with the secretions or body fluids of an infected person. Currently there is no vaccine or drug for EVD. Maintaining good personal and environmental hygiene remains the only control strategy, and its implementation was a challenge in West Africa countries. Ebola virus vaccine (EVV) is being developed and may soon be deployed; thus a need to evaluate factors that can improve or discourage the uptake of the vaccine when it becomes approved for public administration. This study highlights the acceptability and willingness-to-pay for EVV. Majority of the respondents were willing to accept the vaccine and pay for it if it is not publicly funded. Of interest was that among those that accepted to be vaccinated, most would only accept to do so after they had observed the outcome on others that had received the vaccine. There is need for early dissemination of correct information and education on EVV to the populace so as to prevent any misinformation and misperception about the vaccine. This will improve universal coverage with the vaccine when deployed.
| Ebola virus disease (EVD) is caused by Ebola virus (EBV), a highly virulent and infectious virus that infects humans and non-human primates. EVD is transmitted through human-to-human contact [1,2] and has up to 70% case fatality rate [3]. The current outbreak of EVD in six West African countries; Sierra Leone, Liberia, Guinea, Senegal, Mali, Nigeria [4] and reported cases in developed countries [5] have infected about 20,416 persons and caused 8,483 deaths [6,7] as at January 13, 2015.
The only available control strategy is strict personal and environmental hygiene, since no drug [8] has been approved for the treatment and no vaccine has been approved for its prophylaxis. The implementation of adequate hygiene; (avoiding contact with body fluids from an infected person or contact with items handled by an Ebola-infected patient, regular hand washing with soap and water and use of sanitizer) in West Africa is a challenge [9,10] principally due to poverty with existing low standard of living; lack of access to clean water, inadequate sanitation and overcrowded housing. Also inadequate health system in these countries lead to lack of or delayed case identification, inadequate supportive case management, contact tracing and surveillance which aid the spread of the disease. In view of the above limitations, effective prophylaxis through the introduction of Ebola virus vaccine (EVV) is urgently needed.
On August 28 2014, the National Institute of Health (NIH), USA announced that the first testing of EVV on humans by the National Institute of Allergy and Infectious Disease (NIAID) and GlaxoSmithKline (GSK) was imminent. The EVV, is a viral-vector-based recombinant vaccine in which genes encoding protein of Ebola virus is inserted into the genome of another virus (not Ebola virus), recombinant replication-deficient Chimpanzee-derived adenovirus 3 or cAd 3) [11] which when injected will generate both cellular and humoral immunity in the recipients. If approved for usage, the countries that have reported EBV cases may be among the first to benefit.
Whenever a new vaccine is introduced, it has to contend with public acceptability. Although, previous studies have reported favorable attitudes towards newly introduced vaccines [12–14] it would be an over assumption to conclude that introduction of the EVV will be welcomed with the same attitude and uncritical acceptance [15]. The existing poor knowledge on the vaccine by the uninformed masses [12], and the misconception of the possible risk of contracting an illness through a vaccine: as was attributed to oral polio vaccine [16,17], may dissuade majority from accepting the EVV with resultant low uptake of the vaccine, through propaganda [18,19].
Therefore, health program managers have to be proactive in identify early factors that can either facilitate or militate against the effective implementation of EVV program. Issues such as: the acceptance of the vaccine, making the decision to be vaccinated [20,21] and the willingness-to-pay (WTP) for EVV if not publicly funded need to be sorted out. The issue of willingness-to-pay for a newly introduced vaccine is paramount in Nigeria, since some vaccines that have previously been approved for public use are yet to be introduced in the National Programme on Immunization which is funded by the government. Therefore the aim of this study in Nigeria (West Africa) is to determine the public acceptability and willingness-to-pay for EVV. The outcome of this study will contribute to the strategic plan for a successful EVV implementation.
The study was conducted in two sites: an EBV low risk community (Umuahia), Abia State, Southeast and EBV high-risk community (Ajah) in Lagos State, Southwest, of Nigeria. The study took place from August to September 2014 during the period of Ebola outbreak in Nigeria, and data collection was completed before the Monday 20th October 2014 when the World Health Organization (WHO) certified Nigeria free of EBV. The distance between the two communities is about 600 kilometers [22] (Fig 1). The first case of EVD in Nigeria was reported in Lagos in Ikoyi-Obalende Local Council Development Area (LCDA) about 22 kilometers from Ilaje community in Eti-Osa East LCDA, both administrative areas within Eti-Osa Local Government Area (LGA). By the time the WHO certified Nigeria free of EBV a total of 20 cases were reported out of which 8 deaths occurred. Among the reported cases and deaths, 19 cases occurred in Lagos State, out of which 7 deaths were recorded. Throughout the period, there was no report of EVD in Abia State.
Umuahia-North and Eti-Osa LGAs have populations of 359,230 [23] and 983,515 [24] respectively. The population density of Umuahia and Eti-Osa LGAs were 450 persons/km2 and 20,000 persons /km2 respectively. Ugba ward is one of the 12 wards in Umuahia-North and Ilaje is a community within Ward A of Eti-Osa East LCDA, one of the four council’s areas controlled by Eti-Osa LGA. The two communities are mixtures of both urban and rural areas.
It was a community-based cross-sectional qualitative and quantitative interventional study conducted in two communities, each in two states in Nigeria. A stratified random sampling was used to select Ugba and Ilaje wards from a sample frame of 12 and 20 wards from Umuahia-North and Eti-Osa LGA respectively. A systematic random sampling was used to select the households from the house numbering done by the National Primary Health Care Development Agency. Households were selected, beginning with a house randomly selected and subsequent sampling was in alternate of four houses until the stipulated number was obtained. The minimum sample size of 260 for each study site was calculated using a power of 80%, 95% confidence level and based on the vaccine acceptability rate of 81.3% as reported by Williams et al [25]. The household heads participated in the study; if the head of the household was not around during the visit, the spouse was interviewed.
The Health Research and Ethics Committee of the University of Nigeria Teaching Hospital (UNTH) Enugu, gave ethical approval for this study. The committee approved the use of only verbal consent from each respondent, the reason was to reduce contact between the researchers and their multiple respondents which would occur through exchange of writing materials. This was precautionary due to the EBV threat during the period of the study. Although a prior information sheet was given to the identified households seeking their consent to be part of the study, there was no signing of the counterpart consent sheet attached to the questionnaire.
The questionnaire was pre-tested in a community that was not involved in the final study. Few questions were modified to clear ambiguity and some translated words were changed to convey appropriate meaning. Also provisions were made to record comments made by the respondents, since it was realized that there were a lot of valuable information that were not originally included in the initial questionnaire design. The pre-tested interviewer-administered questionnaire was used to collect information on socio-demographic characteristics, respondent’s knowledge on EVD, preventive practices, attitude towards EVV, their knowledge and acceptability of EVV, and their WTP for EVV from the head of the family (S1 Text; sample of the questionnaire). Also information on the medium through which they first heard about Ebola virus disease outbreak was collected. One respondent per household, was interviewed. The preferred respondent was the head of the family, and in a situation where the father/husband was not available, his spouse was interviewed. If neither the head of the house nor the spouse was available to be interviewed, a second visit was rescheduled. The respondents’ acceptability and WTP were assessed pre and post health education on Ebola virus and its potential vaccine. They were informed that EVD is caused by Ebola virus (EBV) which is highly infectious and is transmitted through human-to-human contact [1,2]. That the virus has a very short incubation period and a victim manifests the disease within a very short time of exposure to the virus. The illness has 70% case fatality rate [3] and for a person to be protected by the vaccine, he/she has to receive the vaccine before exposure or not later than five days from the time of exposure. The EVV will be neither an inactivated vaccine which has been found to be unsuccessful with Ebola virus, nor live-attenuated vaccine which is generally considered too dangerous in the case of Ebola virus. The EVV will be a viral-vector-based recombinant vaccine in which genes encoding protein of Ebola virus will be inserted into the genome of another virus (not Ebola virus), a recombinant replication-deficient adenovirus (Ads) or attenuated vesicular stomatitis viruses (VSVs), which are known to cause no serious side effects or disease in human [26] The Ebola virus genes encoded proteins are recognized by the immune system and stimulate immune response against the disease but do not cause Ebola virus disease.
The inequality in WTP for EVV was done using the SES of the households [27,28] which was generated based on functional household asset they owned [29,30]. The household expenditure consumables were used to estimate the household income. The Principal Component Analysis (PCA) was used to create a continuous SES quartiles based on household asset owned and expenditure on food [31]. It is easier to elicit monetary information, like income [32,33] with this approach. The estimation was in Nigerian naira (N) and converted at the rate of 1 United States dollars (USD) to N170.00.
Respondents were asked to recall the first medium through which they got to know about EVD, and multiple options were not allowed. This was to determine the effective channel of disseminating information to the wider population. They were also asked to state their perception of the severity of the disease when they first heard about of EVD and this was to indirectly assess the content and impact of the first information on the public awareness on the disease.
Respondents were asked questions to elicit their knowledge on ways EBV could be contracted, ways to prevent Ebola virus infection, and their preventive practices. They were also tested on their awareness of any vaccine against EVD. The respondents were allowed to give their reply to the questions without pre-empting them with answer options. There were five possible correct preventive measures, as well as preventive practices [34]. A normative value of 1 or 0 was given for correct and wrong responses respectively for the questions. For their knowledge on how EBV could be prevented, a score of 1 was given for any correct response. The lowest and highest possible scores for preventive knowledge were 0 and 5 respectively. The respective total scores were grouped into “very adequate” if 5, “adequate” if 3–4 and “inadequate” if 0–2. Adequacy means that the respondent scored at least three on knowledge of prevention of Ebola virus infection.
On the evaluation of practice, personal hygiene which has two major components was assessed. Nigeria had only 20 cases of EBV and 8 deaths. Therefore most of the respondents had neither seen a person suffering from EVD nor seen someone die from it.
The respondents were asked whether they knew of any EVV. Those that responded stated that: a) there was a vaccine but not approved for use or any related response, were categorized as “correct” and those that stated that: b) there was a vaccine available to combat EBV, or any related response were categorized as “wrong”. To assess the respondents’ acceptability of EVV, a hypothetical cAd3 [17] was described to them. They were informed that the vaccine would be safe and may cause little or no adverse events and lacks the potential of causing the disease. Their willingness to be vaccinated was elicited. Those that accepted to be vaccinated were asked their preferred time to receive the vaccine. A 5-point Likert scale scoring system ranging from 1 = “very unwilling”, 2 = “unwilling”, 3 = “not sure’, 4 = “willing” and 5 = “very willing” was used to rate their level of acceptability of EVV. The respondents who replied either “1’,”2” or “3” were categorized as unwilling, while those whose responses were either “4” or “5” were grouped as willing.
The WTP for the hypothetical EVV was evaluated only among those who accepted to be vaccinated. Their WTP for the vaccine was determined using the contingent valuation method. The highest amount that they were willing to pay for the vaccine was sought after. Since the vaccine is yet to be deployed to the market, no market price is yet available. The contingent valuation method (CVM), is a survey-based approach to elicit monetary valuation of products of healthcare [35,36] by individuals’ using bidding game approach (BGM). This is best suited for exploring individual preferences for goods and services with no known market price, as in this case where there is no market price for EVV. The respondents were presented with a scenario describing the hypothetical EVV, as an effective injectable vaccine, with no risk of getting infected with Ebola virus through the vaccine and should be given preferably before exposure to Ebola virus or at most within five days of exposure. Since no market price was available for the vaccine, the respondents were asked an open ended question: “How much will you be willing to pay for the Ebola virus vaccine?” After the initial response, they were allowed one more option to either increase or reduce the initially stated amount by a second question: “If due to inflation or other uncertainties, the cost for the vaccine is higher than what you have just stated, what is the maximum amount you are very certain to pay bearing in mind that your entire household (both adults and children) may have to receive the vaccine at about the same period?” The effect of cost of vaccine on acceptance rate was elicited by comparing the willingness to vaccinate before knowledge on cost and thereafter. Any other important comments made were documented and analyzed as direct speech.
Certain variables were dichotomized into binary variables: educational status into primary education or less and secondary education and above, acceptability into before and after knowledge on WTP. The households were categorized into those with household size of 1–3 and those of 4 and above. The two major components of practice care hygiene, were each scored 0.5 and a total score of 1 = adequacy and 0.5 = inadequacy. Association of other variables with acceptances of EVV was tested using univariate and multivariate analysis. Two-by-two table was created to test for statistical significance and p-value of <0.05 was taken to be statistically significant. The continuous socio-economic status (SES) index was generated using Principal components analysis (PCA) based on combination of household assets owned and mean weekly expenditure on food items. The SES index was categorized into four equal quartiles (Q) and these four groups were: the poorest (Q1), very poor (Q2), poor (Q3), and least poor (Q4). The correlation between SES and WTP was measured using SES groups.
A total of 645 households were identified for the study and 25 households had neither the father nor the mother to be interviewed while 20 households, the identified respondents declined to participate in the study. Therefore, 600 households were studied giving a cooperation rate of 93%. Of the 600 respondents that participated, 18 questionnaires were not analyzed, the reason being that 11 respondents did not give any responses to most of the questions in the questionnaire and 7 withdrew their consent to participate at some point during the interview (Fig 2).
The Ilaje community had more male respondents (73.3%, 215/293) while (54.1%, 133/289) of the respondents in Ugba were females. The mean ages of the respondents were 37.2 years and 38.32 years in Ilaje and Ugba respectively. In Ilaje and Ugba communities 93.3% (271/293) and 86.2% (249/289) respectively have at least secondary education. The occupation of most of the respondents; 34% (100/293) and 35.6% (103/289) in Ilaje and Ugba, respectively, was small scale business (see S1 Table).
The media through which most of the respondents first learnt about EVD were television (32.4%, 178/582) and radio (27.1%, 150/582). No one heard it from either church/mosque (0.0%) or hospital (0.0%) (Fig 3).
Ninety five percent of respondents stated that EBV can be transmitted by contact, Table 1. Majority (73%, 425/582) mentioned washing of hands as a preventive measure. Only 38.5% (224/582) appreciated the seriousness of the disease when they heard of the EVD for the first time. Hand washing (66.7%, 388/582) was the commonly adopted preventive measure while 12.3% (72/582) took no precautions.
The respondents in Ilaje and Ugba that acknowledged that there was an EVV were 2.7% (8/293) and 41.0% (119/289) respectively (p = 0.0001) (Table 2). Prior to health information on EVV, 80.0% (234/293) and 79.5% (230/289) of the respondents in Ilaje and Ugba respectively would accept EVV (p = 0.93). After information on EVV, 86.3% (253/293) (Ilaje) and 82.1% (237/289) (Ugba) were willing to be vaccinated. Among those that accepted to vaccinate once EVV is available in Ilaje and Ugba were 79.8% (202/253) and 47.7% (113/237) respectively, while 12.2% (31/253) and 51.5% (122/237) in Ilaje and Ugba would like to receive the vaccine later, after observing the effect on those that received it (p = 0.0001). The respondents that were unwilling to vaccinate in Ilaje and Ugba were 6.9% (16/233) and 37.8% (90/237) respectively (p = 0.0001). Among those willing to vaccinate, 91.2% (212/233) and 87.5% (207/237) in Ilaje and Ugba respectively were willing to pay for EVV (p = 0.2). The very poor and the least poor were willing to pay the least amount of money for the vaccine, while the poorest and the poor were willing to pay a higher amount for the EVV. Households with household size of 4–5 in number were willing to pay the highest amount of money for the EVV (S2 Table).
Some of the respondents who were not willing to pay for the EVV as well as some of those who were willing were of the opinion that government should provide the vaccine at no cost to the recipients (commented by 55.2%). Respondents stated that “Government should pay for it and make it free.” “It is among the duties of the government to protect the citizens and providing this vaccine should be one way to do that”. “I don’t think it is right to expect people to pay for a vaccine that will protect them from a disease they do not have any control on how it can be contracted.” Others suggested that government should coerce people to receive the vaccine. “EBV disease is a public health problem” and “Government should persuade everybody to receive the vaccine, and the only way they can do that is to provide it free of charge”. “It should be made compulsory and enforced. If it would be enforced, you cannot ask people to pay for it.” Other common suggestions on how to avoid or minimize out-of pocket payment by the people (reported by 21.0%) were: “National Health Insurance Scheme (NHIS) should cover the cost.” “The vaccine bill should be incorporated into the costs of GSM phone bills….”
Univariate and multivariate analyses demonstrated that educational status strongly correlated with acceptance of EVV (Table 3). The lower the educational status, the more likely they are to accept the vaccine (95% CI: 0.20–0.74, p-value = 0.001). It was also shown that giving health education on EVV improved its acceptability and the difference found to be statistically significant (95% CI: 0.54–1.01, p-value = 0.046)
The study showed that majority of the respondents learnt about EVD from the local media (television and radio) and none heard it from health care worker. This is similar to what have been reported elsewhere [37,38]. This could be partially attributed to the prevailing industrial action which closed the services in many public medical facilities during the early period of EBV threat in Nigeria. Although, private sector healthcare providers were functional, it has been revealed that healthcare providers in developing countries rarely educate their clients/patients on health related issues [39]. The lack of healthcare providers role in providing initial information on EVD, may account for the high proportion of the respondents that did not understand and appreciate the seriousness of the disease. Healthcare providers should assist the public at every contact in obtaining health information to optimize their health outcome.
Majority of respondents identified personal hygiene as an effective preventive measure, but practiced avoidance of casual hand shake as a preventive action, thus, revealing a gap between knowledge and practice. It has been established that EBV can spread from person to person through contact with body fluids from an infected individuals’ blood, feces, or other body fluids, not by avoidance of handshakes, hugging or being in a gathering like most respondents stated [40].
The level of acceptability of EVV among all socioeconomic classes was high. This high acceptability may not translate to prompt EVV uptake since majority would like to observe the effect of the vaccine on others before vaccination. The delay by some people in receiving the vaccine may segregate the population and provide an opportunity to spread inaccurate vaccine information which may lead to stigmatization of those that have received the vaccine [41,42] Healthcare providers should strive to gain public trust on the new vaccine by providing information on safety and side effects [43]. Respondents’ comments such as “EBV disease is a public health problem and Government should compel everybody to receive the vaccine….” and “It should be made compulsory and enforced.” are calls for strong political will to ensure public acceptance of EVV. Government should play a pivotal role in sharing the correct information about EVV with the people, Also as stated by one of the respondents, “It is among the duties of the government to protect the citizens and providing this vaccine should be one way of doing that….,”. However the government should not stop at providing this vaccine free of charge to the community, but encourage and possibly, coerce people to receive it. The classification of the current EVD outbreak as a public health emergency of international concern, justifies the use of reasonable legal measure to control the outbreak, including the option of compelling people to get vaccinated.
Majority of the respondents were WTP for EVV. This is similar to what has been reported on other new vaccines [44]. This may be due to the anticipated benefit which influences decision to pay for a product [45] In this case, it is the protection from EBV, since the risk of infection was high during the outbreak. However, no information was available on the market price to compare the average amount they were willing to pay for EVV. Most new drugs and vaccines are not affordable for out-of-pocket payment especially for the poorer households. Therefore, government should make plans to either subsidize the cost of the vaccine or at best bear the full cost. The suggestion by some respondents for NHIS to cover the cost of EVV may not be an ideal option, in a country like Nigeria, where NHIS currently covers only employees in the Federal formal sector which accounts for less than 5% of the population [46]. In Nigeria, WTP studies have been conducted for several healthcare products and services [47–49]. However, we are not aware of any WTP study in both Nigeria and beyond with respect to pre-vaccine deployment for EVV. Thus there is little information on how this can affect the accessibility of the vaccine when introduced. Nonetheless, the envisaged assistance of Global Alliance for Vaccine Initiatives (GAVI) and UNICEF with funds for the vaccine would increase the possibility of no financial cost to recipients when EVV is deployed in many countries that reported outbreak of EVD. However, a major obstacle that these nations will encounter is establishing a long-term financial sustainability structure that will continue to maintain accessibility and affordability of EVV if in future there is donor fatigue and reduction in funds. This raises the pertinent issue on how much people will be willing to pay for EVV. Furthermore, government may in an effort to make the EVV easily accessible to the public in situation of lack of external sponsor, find it necessary to subsidize the cost of the vaccine. In such a situation, government may need information on willingness-to-pay for the vaccine.
One of the limitation of this study is that it was conducted during the outbreak of EBV disease in the West African sub-region including Nigeria. The reality and fear of potential EBV infection might have affected the responses obtained. The findings of our study could be a representative of the highest level of acceptability and WTP for EVV. A concurrent study in any of the western countries where the threat of EBV was virtually non-existent might reveal a lower acceptability. Secondly, the hypothetical nature of the study may differ from the real practice. However the findings give an insight to the possible challenges that may exist when the real vaccine is introduced. Another limitation is the use of ownership of household assets in the classification of SES, in community where people purchase either used cheap items or brand new expensive items. This has the potential of leading to wrong SES classification. The most objective assessment would have been classification based on income, but previous experiences have shown that obtaining such information is often shrouded with difficulty. The other socioeconomic classification approach by Oyedeji or Olusanya which relies on highest education and occupation of the parents was not used.
The level of acceptability of Ebola virus vaccine among respondents was high, although majority confirmed that they would not hastily receive the vaccine until they observe the effect on others. Nonetheless, they was willingness to pay for the vaccine whenever they are to receive it should the vaccine not be publicly funded. However, it is recommended that for high uptake to be achieved, the vaccine introduction should be preceded with wide public health education, counselling and persuasion, and government should endeavor to provide the vaccine free of charge.
|
10.1371/journal.pbio.2001586 | History of antibiotic adaptation influences microbial evolutionary dynamics during subsequent treatment | Antibiotic regimens often include the sequential changing of drugs to limit the development and evolution of resistance of bacterial pathogens. It remains unclear how history of adaptation to one antibiotic can influence the resistance profiles when bacteria subsequently adapt to a different antibiotic. Here, we experimentally evolved Pseudomonas aeruginosa to six 2-drug sequences. We observed drug order–specific effects, whereby adaptation to the first drug can limit the rate of subsequent adaptation to the second drug, adaptation to the second drug can restore susceptibility to the first drug, or final resistance levels depend on the order of the 2-drug sequence. These findings demonstrate how resistance not only depends on the current drug regimen but also the history of past regimens. These order-specific effects may allow for rational forecasting of the evolutionary dynamics of bacteria given knowledge of past adaptations and provide support for the need to consider the history of past drug exposure when designing strategies to mitigate resistance and combat bacterial infections.
| Bacteria readily adapt to their environments and can develop ways to survive and grow in the presence of antibiotics. While many studies have investigated how bacteria evolve to become resistant to single drugs, it is unclear how adaptation to other drugs and environments in the past affect the way bacteria adapt to new drugs and environments. In this study, we allowed bacteria in a laboratory setting to adapt to three different antibiotics. We first exposed wild-type susceptible bacteria to high concentrations of the three antibiotics individually and then exposed these populations to each of the other drugs. By tracking the levels of resistance to all three drugs in all of the treatments, we identified cases in which past adaptation to one treatment influenced subsequent evolutionary dynamics with regard to both phenotypes (levels of resistance) and genotypes (genes that became mutated). Additionally, by allowing bacterial isolates originating from human patients to adapt to the three drugs, we recapitulated a subset of the adaptation history-dependent evolutionary dynamics. Overall, this study sheds light on how adaptation history in bacteria can potentiate or constrain the rise of multidrug resistance, depending on the particular order of drugs used in a sequential therapy.
| Antibiotic resistance is a growing healthcare concern whereby bacterial infections are increasingly difficult to eradicate due to their ability to survive antibiotic treatments [1]. There have been reported cases of resistance for nearly every antibiotic we have available [2]. Coupled with the fact that the antibiotic discovery pipeline has slowed over the past few decades [3], there is a dire need to find better treatment strategies using existing antibiotics that can slow or even reverse the development of resistance.
Adaptive laboratory evolution is a technique that can be used to study and test evolutionary principles in a highly controlled laboratory setting [4]. Microorganisms with short generation times such as bacteria are especially amenable to adaptive laboratory evolution and can be adapted to an environment through repeated cycles of growth in a specific media environment, dilution of the culture, and subsequent passaging into fresh media [5]. Multiple replicates for each condition can be evolved in parallel to investigate the reproducibility of evolutionary dynamics [6]. The evolutionary trajectories of the bacteria can be measured as they adapt to different nutrients and stressors over time [7]. Whole-genome resequencing on the evolved strains can subsequently be used to determine the mutations that occurred that may be associated with the observed phenotypes [8,9].
Adaptive laboratory evolution can be used to study the development of antibiotic resistance in bacterial pathogens [10]. Resistance to antibiotics is an evolutionary response of bacteria to withstand and survive the effects of the stressor. Deliberately evolving bacteria to withstand antibiotics through experimental evolution can yield insights into the evolutionary dynamics and trajectories of this adaptive process [11,12]. These evolution experiments can provide a longer-term perspective, which can yield information for the design of novel treatment strategies that can reduce the rate of resistance evolution or potentially even reverse the effects of resistance [13–15].
Recent studies have explored how adaptation to an antibiotic can cause bacteria to concurrently become more susceptible or more resistant to other drugs, an effect termed collateral sensitivity or collateral resistance [14,16,17]. Collateral sensitivities between drugs have been used to design drug cycling strategies and to explain the decreased rate of adaptation to certain antibiotics [12,14,18–23]. Drug deployment strategies that exploit such collateral sensitivities between pairs of antibiotics to minimize resistance evolution have been tested in vitro. A recent study determined the collateral sensitivity drug interaction network in Escherichia coli and demonstrated how an alternating sequential treatment of 2 reciprocal, collaterally sensitive antibiotics can slow down the rate of resistance evolution [14]. In this drug cycling strategy, the development of resistance to 1 drug concurrently increased the sensitivity to the second drug, and this allowed wild-type cells to outcompete the resistant cells when exposed to the second drug. In a different study, evolution experiments of alternating sequential therapies of pairs of antibiotics were performed on Staphylococcus aureus, and the study showed that the alternating treatments slowed the rate of resistance development compared to single-drug treatments [12]. Consistent with the E. coli study, this study found that collateral sensitivities could explain the evolutionary constraints in the cases in which alternating treatment resulted in decreased resistance development compared to the single-drug treatment.
Most of the prior studies that test the use of alternating antibiotic therapies to reduce the rate of resistance development employ an adaptive laboratory evolution scheme in which the drugs are switched at daily or subdaily intervals with the purpose of testing if rapidly changing antibiotic environments can diminish the rate of drug-resistance adaptation [12,19,23,24]. In this study, we expand on these prior works, but we are not focused on studying the evolutionary dynamics of bacteria adapted to rapidly changing drug environments. Rather, we explore the evolutionary dynamics of sustained, longer treatments of drugs and how the development of high levels of resistance to 1 drug influences the subsequent dynamics of sustained adaptation to a second drug. In clinical settings, when antibiotic cycling strategies are employed, they are used typically at the level of the hospital ward, and the cycling of antibiotics are often done at monthly intervals [25,26]. The rationale here is that if resistance to 1 drug arises after frequent use in a ward, switching to an antibiotic of a different class may allow resistance rates to the withdrawn drug to stabilize or even fall, enabling the first drug to be efficiently reintroduced again at a later time [27]. This practice of cycling drugs of different classes over the course of monthly intervals is done empirically, and it remains unclear how these regimens constrain the evolutionary dynamics of antibiotic resistance development. Here, we explore the evolutionary trajectories of bacteria as they evolve high levels of resistance to 1 antibiotic and the subsequent trajectories as the selection pressure from the first drug is withdrawn and replaced with the sustained pressure of a different drug. It remains unexplored how prior adaptation to 1 drug environment affects the evolutionary dynamics of a bacterial population during subsequent adaptation to a second drug in terms of the amount of resistance it can potentially develop and the resistance profile of the first drug. Collateral sensitivities and collateral resistances between 2 drugs have been studied in the context of adaptation to single drugs [21], i.e., as bacterial populations evolve and become resistant to 1 drug, do the cultures concurrently become more resistant or sensitive to other drugs? In this study, we focus not on if bacteria become concurrently more resistant or sensitive to other drugs, but rather, if adaptation to 1 drug constrains or potentiates the evolutionary dynamics to sustained adaptation to a second drug. How does the history of adaptation to 1 drug influence the subsequent adaptation to a second drug? If there are such historical dependencies, can we use this knowledge to design sequential therapies that slow down the evolution of resistance to the drugs used? What happens to the previously developed resistance once the drug pressure is taken away or switched to a different drug? Do compensatory adaptations sustain the high resistance, or do the bacteria revert and become susceptible again [28]? The answers to these questions are important for understanding how bacteria adapt to different antibiotic environments. Bacterial pathogens have complex evolutionary histories and elucidation of any historical dependencies of antibiotic resistance evolution would allow for rational forecasting of future resistance development and would aid in the design of strategies for mitigating antibiotic resistance.
To test how different antibiotic-resistance backgrounds affect the subsequent adaptation dynamics when evolved to a new antibiotic, we used a laboratory evolution approach to evolve P. aeruginosa to all 2-drug sequences of the 3 clinically relevant drugs piperacillin (PIP), tobramycin (TOB), and ciprofloxacin (CIP). In each of the experimental sequences, P. aeruginosa was subjected to 20 days of adaptation to each drug by serially passaging parallel replicate cultures to increasing concentrations of the drugs followed subsequently by 20 more days of adaptation to each of the 3 drugs or to lysogeny broth (LB) media without a drug (Fig 1A). Additional parallel replicates were adapted to LB media without a drug for 40 days as a control. For each drug treatment, changes in the resistance to the other 2 drugs were concurrently measured (Fig 1B). Minimum inhibitory concentration (MIC) gradients in microtiter plates were used to simultaneously measure the drug resistance level and to propagate the bacteria daily. To adapt the bacteria to a drug, a sample is taken from the population from the well of the highest drug concentration that allowed for growth (i.e., MIC/2) and then used to inoculate a new MIC gradient. This serial dilution cycle is done daily. More explicitly, 20 μl of culture is sampled from the well of the highest concentration that allowed for growth, then diluted in 5 ml of fresh LB media, and then this diluted culture is used to inoculate a new MIC gradient. This dilution protocol results in a daily dilution factor of the bacterial population of 1/500 (Materials and methods, Fig 1B). S1 Data provides the estimated number of generations per day for the evolved lineages based on the daily measurements of the OD600. For each lineage, the OD600 values are fairly consistent from day to day, and so with a dilution factor of 1/500, the cultures undergo approximately 9 generations of growth per daily dilution cycle (S1 Data).
We observed differences in final resistance levels to the different drugs depending on the history of past treatments (or lack of treatments), an effect we call drug-order–specific effects of adaptation. Our results show that a history of past drug adaptation can affect the rate at which resistance can potentially arise when subsequently adapted to a new antibiotic. Furthermore, in some cases, adaptation to a second drug or to LB can partially or fully restore sensitivity to the first drug. These observations suggest that in order to limit the rate of development of antibiotic resistance, it is important to consider which drugs a bacterial population may have been exposed to in the past when choosing which drugs to subsequently deploy.
The 3 drugs tested have different mechanisms of action and are clinically used to treat P. aeruginosa infections [29]. Piperacillin is a beta-lactam that inhibits cell wall synthesis [30]; tobramycin is an aminoglycoside that binds to the prokaryote ribosome and inhibits protein synthesis [31]; and ciprofloxacin is a fluoroquinolone that binds DNA gyrase and inhibits DNA synthesis [32]. We chose to study these 3 antibiotics because of their common use in the clinical setting to treat P. aeruginosa infections [29], their diverse mechanisms of action, and their well-studied resistance mechanisms [33]. Adaptive evolution for 20 days to these drugs individually resulted in 1-drug–resistant mutants denoted as PIPR, TOBR, and CIPR. The Day 20 PIPR, TOBR, and CIPR populations had averages of 32-, 64-, and 64 times higher MICs to piperacillin, tobramycin, and ciprofloxacin, respectively, compared to their initial levels. To determine if the population MICs that were measured during the course of the adaptive laboratory evolution experiments were representative of individual colony MICs, we retrospectively measured the MICs of cultures grown from multiple revived colonies from the saved frozen stocks (S2 Data). Overall, 73% of the retrospectively measured MICs were within one 2-fold dilution step of the originally measured population MICs, which suggests that the reported MIC values for each of the evolved lineages are well representative of the bacterial populations (S1 Fig).
By following how the resistance to each of the 3 drugs changes for each of the drug sequences (Fig 2; S2 and S3 Figs and S3 Data), we observed 3 types of drug-order–specific effects in the MIC profiles (Fig 3). Note that for now, we focus on summarizing the different drug-order–specific effects (as seen by the changes in drug MICs), and later, we discuss several hypotheses for the underlying mechanisms of the drug-order–specific effects based on analysis of the genomic mutations of the adapted lineages. In the first type of drug-order–specific effects, adaptation to a second drug or to LB restores the susceptibility to the first drug (Fig 3A). In these experiments, we were first interested to see if the increases in MICs of the 1-drug–resistant lineages (Day 20 PIPR, TOBR, and CIPR) were permanent or transient. By evolving them to LB and hence removing the selection pressure of the drug for 20 days, we observed that the high MICPIP was maintained in Day 40 PIPRLB (Fig 3A [top], p = 0.80; Fig 2A), while MICTOB declines (leading to partial resensitization) in Day 40 TOBRLB (Fig 3A [middle], p < 0.0001; Fig 2E), and MICCIP declines (although not significantly) in Day 40 CIPRLB (Fig 3A [bottom], p = 0.18; Fig 2I). Thus, for these 3 treatments, removal of the antibiotic pressure can maintain the high resistance or lead to resensitization in a drug-specific manner. Similar trends were seen in a recent adaptive evolution study whereby P. aeruginosa was evolved to tobramycin, ciprofloxacin, piperacillin/tazobactam, meropenem, and ceftazidime, followed by subsequent adaptation in the absence of the drug (growth medium only) to determine the effects of removing the drug selection pressure [34]. Similar to the patterns seen in our study, they observed that the tobramycin-resistant cultures partially resensitized, the ciprofloxacin-resistant cultures had a modest resensitization, and the 3 beta-lactam–evolved cultures maintained high levels of resistance.
Next, we were interested to see if evolving the 1-drug–resistant lineages to the other 2 drugs would show the same patterns seen as when evolved to LB. Interestingly, we saw unique outcomes for each of the 3 lineages. When Day 20 PIPR was evolved to tobramycin, the MICPIP of Day 40 PIPRTOBR remained high (p = 0.90), similar to how the MICPIP of Day 40 PIPRLB remained high (Fig 3A [top]). This result suggests that subsequent tobramycin adaptation has no role in altering the high piperacillin resistance. This specific order of drug treatments can then result in multidrug-resistant P. aeruginosa cultures that are resistant to both piperacillin and tobramycin. On the other hand, when Day 20 PIPR was evolved to ciprofloxacin, the resulting cultures became resensitized to piperacillin (Fig 3A [top], p < 0.05), and the MICPIP declined to levels comparable to those of the initially susceptible cultures (MICPIP of Day 1 PIPR versus Day 40 PIPRCIPR, p = 0.80), indicative of a full resensitization. Since resensitization did not occur after subsequent adaptation to tobramycin or LB, we suspect that the subsequent ciprofloxacin adaptation had an active role in the resensitization to piperacillin in such a way that tobramycin and LB did not. These results show that if a piperacillin-resistant culture (that is also sensitive to tobramycin and ciprofloxacin) is evolved to tobramycin, multidrug resistance can occur. However, if it is evolved to ciprofloxacin, despite the fact that ciprofloxacin resistance increases, the culture becomes susceptible to piperacillin again, making piperacillin a potentially rational choice for further treatment.
When Day 20 TOBR was evolved to ciprofloxacin, partial resensitization occurred (MICTOB of Day 20 TOBR versus Day 40 TOBRCIPR, p < 10−5), and the MICTOB of Day 40 TOBR-CIPR fell to a comparable level as that of Day 40 TOBRLB (p = 0.98) (Fig 3A [middle]). This result suggests that the resensitization seen during the subsequent ciprofloxacin adaptation is not caused by the selection pressure of ciprofloxacin, but rather by the absence of the selection pressure of tobramycin. On the other hand, evolving Day 20 TOBR to piperacillin also led to a partial resensitization (MICTOB of Day 20 TOBR versus Day 40 TOBRPIPR, p < 0.05) but not as much as it did when Day 20 TOBR was evolved to ciprofloxacin (MICTOB of Day 40 TOBRPIPR versus Day 40 TOBRCIPR, p < 0.01) and LB (MICTOB of Day 40 TOBRPIPR versus Day 40 TOBRLB, p < 0.05). Because of this difference, we suspect that the maintenance of the comparably high tobramycin resistance is a consequence of the piperacillin selection pressure, since we observed that adaptation in the absence of the drug pressure in LB led to substantially greater resensitization. This case highlights how the removal of all drug pressures may lead to the resensitization of the culture more than with the treatment of the culture to a new drug. In conjunction with the result that subsequent tobramycin adaptation of Day 20 PIPR still maintained a high MICPIP, this case then also shows how regardless of the order, sequential adaptation to piperacillin and tobramycin leads to multidrug resistance of the 2 drugs.
Lastly, when Day 20 CIPR was evolved to piperacillin and tobramycin, both treatments lead to a partial resensitization to ciprofloxacin (Fig 3A [bottom]). During subsequent tobramycin adaptation, the decrease in the MICCIP from Day 20 CIPR to Day 40 CIPRTOBR (p < 0.01) was marginally more than the decrease in the MICCIP from Day 20 CIPR to Day 40 CIPRPIPR (p < 0.05) during subsequent piperacillin adaptation. As mentioned above, subsequent adaptation of Day 20 CIPR to LB led to a decrease in MICCIP that was not statistically significant; however, we argue that the decrease is comparable to that seen when adapted to piperacillin and tobramycin as the final MICCIP of Day 40 CIPRLB was not significantly different than that of Day 40 CIPRPIPR (p = 0.93) and that of Day 40 CIPRTOBR (p = 0.53). Hence, in this case, evolution of a ciprofloxacin-resistant culture to either a different drug or to a no-drug condition led to a partial resensitization of ciprofloxacin. Interestingly, we also observed that the resensitization that occurred during subsequent piperacillin adaptation happened more quickly than the resensitization that occurred during subsequent tobramycin and LB adaptation (Fig 2I). After 5 days of subsequent piperacillin adaptation (Day 25 CIPRPIPR), the MICCIP was significantly different than that of Day 20 CIPR (p < 0.001), while this was not the case after 5 days of subsequent tobramycin (p = 1.00) or LB (p = 0.57) adaptation. These cases in which partial or full resensitization to the first drug occurs after adaptation to a second drug or LB highlight opportunities in which resistance to 1 drug can potentially be reversed by treating with a second drug or by removing the drug pressure completely.
In the second type of drug-order–specific effects, prior adaptation to a first drug reduces the rate of subsequent adaptation to a second drug (such that the endpoint level of resistance to that second drug is lower compared to the amount of resistance developed when the Day 0 Ancestor is directly evolved to that second drug). We observed that evolution first to piperacillin reduces the rate of subsequent evolution to tobramycin (Fig 2D and 2E). That is, the MICTOB of Day 40 PIPRTOBR was less than that of Day 20 TOBR (Fig 3B, p < 0.05). This observation suggests that, in some cases, different bacterial populations may evolve resistance to a given antibiotic at different rates depending on the history of prior adaptations that the populations have experienced. Having knowledge of prior adaptations may then potentially be used to slow down the development of resistance to a drug if that drug is selected rationally. Interestingly, we observed no cases in which prior drug adaptation led to an enhancement in the rate of adaptation to a second drug.
The last type of drug-order–specific effects is when the final MIC of a drug is different after adaptation to a 2-drug sequence compared to after adaptation to the opposite order of the 2 drugs (Fig 3C). This third type of drug-order–specific effect exists as a consequence of a combination of the first type of effect (resensitization of the 1-drug–resistant lineages during subsequent adaptations to other drugs) and specific cases of collateral sensitivities during the adaptation of certain lineages. First, the MICPIP was higher when piperacillin was used after ciprofloxacin (Day 40 CIPRPIPR) compared to when piperacillin was used before ciprofloxacin (Day 40 PIPRCIPR) (Fig 3C [top], p < 0.05). In this case, adaptation to piperacillin first led to high levels of piperacillin resistance, and subsequent adaptation to ciprofloxacin led to the resensitization to piperacillin as discussed before (Fig 2A). On the other hand, even though adaptation to ciprofloxacin first led to a collateral sensitivity to piperacillin (S4A Fig [right], p < 0.01), subsequent adaptation to piperacillin resulted in a final MICPIP comparable to that of Day 20 PIPR (Fig 2C).
Next, we observed that during the adaptation to tobramycin followed by ciprofloxacin and vice versa, the final MIC values of piperacillin and ciprofloxacin were different depending on the order of adaptation to the 2 drugs (Fig 3C [bottom and middle]). With regards to the difference seen in the final MICCIP (Fig 3C [bottom], p < 0.05), the partial resensitization to ciprofloxacin starting from Day 20 CIPR during subsequent tobramycin adaptation (Fig 2I) resulted in the MICCIP to be less than adaptation to tobramycin first, followed by ciprofloxacin (Fig 2H). Finally, it was interesting that even though piperacillin was not the direct selection pressure, there was a difference in the final MICPIP level whether ciprofloxacin adaptation occurred after tobramycin adaptation or vice versa (Fig 3C [middle], p < 0.01). In this case, initial adaptation to tobramycin first did not affect the MICPIP (Fig 2B), but subsequent adaptation to ciprofloxacin resulted in a collateral sensitivity to piperacillin (S4C Fig, p < 0.01). On the other hand, as previously mentioned, adaptation to ciprofloxacin first initially resulted in the collateral sensitivity to piperacillin (S4A Fig [right], p < 0.01); however, the MICPIP returned to baseline values during subsequent adaptation to tobramycin (Fig 2C). Thus, regardless if ciprofloxacin adaptation occurred before or after tobramycin adaptation, ciprofloxacin adaptation led to piperacillin collateral sensitivity. However, in order to take advantage of this collateral sensitivity, ciprofloxacin adaptation should be used after tobramycin adaptation, rather than vice versa. In a contrasting example, we also found it interesting that while ciprofloxacin adaptation also led to collateral sensitivity of tobramycin, subsequent piperacillin adaptation did not cause the MICTOB to return to baseline levels (Fig 2F) in the manner in which subsequent tobramycin adaptation returned the MICPIP to baseline values (Fig 2C). Altogether, these cases highlight how treating an infection with a sequence of 2 drugs can result in different resistance profiles depending on the order used.
All the cases of collateral sensitivity that were observed occurred during ciprofloxacin treatment whereby ciprofloxacin adaptation resulted in a lower MIC of piperacillin or tobramycin compared to baseline levels (S4 Fig). First, adaptation to ciprofloxacin starting from the Day 0 Ancestor resulted in collateral sensitivity to both piperacillin (Fig 2C; S4A Fig [right], p < 0.01) and tobramycin (Fig 2F; S4A Fig [left], p < 0.0001). Next, adaptation to ciprofloxacin starting from both the 1-drug–evolved lineages Day 20 PIPR (Fig 2D) and Day 20 TOBR (Fig 2B) resulted in collateral sensitivity to tobramycin (S4B Fig, p < 0.01) and piperacillin (S4C Fig, p < 0.01), respectively. These results suggest that regardless of historical background, ciprofloxacin adaptation results in collateral sensitivity to the other 2 drugs. While we observed that collateral sensitivity of other drugs occurs only during ciprofloxacin adaptation, a recent study in which P. aeruginosa ATCC 27853 was evolved to different antibiotics reported that evolution to tobramycin resulted in collateral sensitivity to piperacillin-tazobactam and ciprofloxacin, whereas we did not observe this effect [34]. Also, this study did not observe that adaptation to ciprofloxacin resulted in collateral sensitivity to piperacillin and tobramycin, as we reported here. We suspect that these inconsistences may be due to strain-specific differences in the different P. aeruginosa strains used (strain PA14 was used in this study).
We were interested in measuring the fitnesses of the evolved lineages to see if the adaptations to the different treatments altered their growth dynamics. We measured the growth curves (OD600) of the 68 evolved replicate lineages as well as the Day 0 Ancestor in quadruplicate grown in LB for 24 hours (S5 Fig). The exponential growth rates were subsequently calculated from the growth curves (S6A Fig, S1 Text and S4 Data). While we observed many different growth rates across the different lineages, we did not observe any correlation between the growth rate and the change in MIC between the Day 20 1-drug–evolved lineages and the subsequent Day 40 lineages (i.e., altered growth rates could not explain the cases in which subsequent adaptation led to the maintenance of high resistance or resensitization to the first drug) (S6B Fig and S1 Text).
We hypothesized that genomic mutations acquired during adaptive evolution contributed to the drug-order–specific effects observed in the MIC profiles. We sequenced genomes of the Day 0 Ancestor, Day 20 PIPR, TOBR, CIPR, and LB Control lineages and the Day 40 1-drug–and 2-drug–evolved lineages, as well as the LB Control lineages. Genome sequencing of the Day 20 and Day 40 mutants revealed a total of 201 unique mutations across the 56 samples consisting of 77 SNPs, 31 insertions, and 93 deletions (Fig 4; S7 Fig, S1 and S2 Tables). The 77 SNPs were found within 49 genes. Two SNPs were synonymous, and 6 were intergenic. To test how representative the sequencing results were of the mutant populations, we used PCR and Sanger sequencing to test for the presence of specific mutations in multiple colonies of different lineages after reviving the lineages from the saved frozen samples. We used the primers from S3 Table to test for the presence of 1 mutation from 1 replicate of each lineage, with 4 colonies of each lineage. Overall, while there may be limited heterogeneity in the populations with respect to a few of the mutations, the large majority of the mutations were homogeneous in the populations and fixed within the lineages (S1 Text).
While some genes were mutated during evolution to all drugs, other mutations were drug-specific and were related to the drugs’ primary mechanisms of action as would be expected (S4 Table). Genes encoding transcriptional regulators for multidrug efflux pumps were commonly mutated during evolution to all 3 drugs (mexC, mexR, mexS, nalC, nalD, nfxB, parS) [35]. Ribosomal proteins (rplJ, rplL, rpsL, rplF) [36] and NADH dehydrogenase subunits (nuoB, nuoG, nuoL, and nuoM) [37,38] were frequently mutated during tobramycin evolution. The most commonly mutated gene was fusA1, which encodes elongation factor G and was mutated in 11 different replicate lineages adapted to tobramycin. fusA1 has been observed to be mutated in clinical isolates of P. aeruginosa [39–41], as well as in adaptive evolution studies to aminoglycosides in P. aeruginosa [34] and E. coli [12,16,18]. Mutations in fusA1 may also contribute to altered intracellular (p)ppGpp levels, which may modulate virulence in P. aeruginosa [41]. Mutations in gyrA and gyrB were observed during ciprofloxacin evolution, but none were observed in parC and parE (the other genes of the quinolone resistance-determining region [29]). Lastly, genes encoding peptidoglycan synthesis enzymes (dacC, mpl) and beta-lactamase regulators (ampR) were mutated during piperacillin treatment. Many of these genes have also been observed to be mutated during human host adaptation of P. aeruginosa [42], highlighting the importance of several of these clinical resistance determinants (S1 Text).
We next analyzed the genomic mutations to see how the historical context affects which mutations occur during adaptation to a drug. For example, how do the mutations that occur during adaptation to piperacillin only (Day 20 PIPR and Day 40 PIPR) compare to the mutations that occur during piperacillin adaptation when there is a prior history of adaptation first to tobramycin (Day 40 TOBRPIPR) or ciprofloxacin (Day 40 CIPRPIPR)? To this end, we first categorized the genes in which mutations occurred into 23 broad categories based on the available literature and on the PseudoCAP functional classifications from the Pseudomonas Genome Database [43] (Table 1). Next, for each lineage, we tallied the number of times a gene in a functional category was mutated across the 4 biological replicates for each of the lineages (Fig 5). For a complete list of genes in each functional classification and descriptions of the genes, see S2 Table.
We observed several general trends in the genes mutated during adaptation to the 3 drugs, depending on their historical context. In the lineages adapted to piperacillin, we saw history-dependent trends in the mutated genes that were related to multidrug efflux pumps (Fig 5, dashed-black box). While all the piperacillin-adapted lineages had mutations in genes related to the MexAB-OprM efflux pump (which is the primary efflux pump of piperacillin [44]) such as nalD and mexR (whose products repress the expression of mexAB-oprM [45]), the Day 40 CIPRPIPR lineage had additional mutations in the structural subunit genes of the other efflux pumps MexCD-OprJ (mexC) and MexEF-OprN (mexF). Lastly, no mutations in genes related to the MexXY-OprM pump were observed in any of the piperacillin-adapted lineages. With regard to adaptation to piperacillin only, most of the mutations that occurred in genes related to MexAB-OprM occurred within the first 20 days, with only a few additional mutations occurring between Day 21 and 40. Regardless of historical context, metabolic and cell wall genes tended to be frequently mutated in piperacillin-adapted lineages, whereas metabolic and cell wall genes did not seem to be consistently mutated across the tobramycin-adapted and ciprofloxacin-adapted lineages. This result is perhaps due to the fact that the primary target of piperacillin is cell wall (peptidoglycan) synthesis, which is largely a metabolic process. Interestingly, we also observed that the lineages adapted only to piperacillin (Day 20 PIPR) sustained large chromosomal deletions that were not seen in the lineages in which there was prior tobramycin or ciprofloxacin adaptation (Day 40 TOBRPIPR and Day 40 CIPRPIPR). We discuss and explore the potential implications of these large deletions below.
The tobramycin-adapted lineages consistently had mutations occur in ribosomal subunit genes and other ribosomal machinery genes, regardless of historical context. In the lineages adapted only to tobramycin, mutations in genes related to the ribosome, membrane, energy, and NADH dehydrogenase tended to occur by Day 20, followed by mutations in efflux pump–related genes by Day 40. The mutations in genes related to membrane, NADH dehydrogenase, and energy likely reflect the unique requirement of the proton-motive force for the uptake of aminoglycoside antibiotics [46], and the mutations occurring during tobramycin adaptation may contribute to the resistance by reducing the proton-motive force [16]. While we observed mutations in the NADH dehydrogenase genes in the lineages adapted only to tobramycin, we saw no such mutations in the lineages in which prior piperacillin or ciprofloxacin adaptation occurred (Day 40 PIPRTOBR and Day 40 CIPRTOBR). Also, while efflux pump–related genes were mutated in the Day 40 TOBR and Day 40 CIPRTOBR lineages, no such mutations were seen in the Day 40 PIPRTOBR lineages in which prior adaptation to piperacillin occurred (Fig 5, dashed-purple boxes).
The mutations in the ciprofloxacin-adapted lineages were fairly consistently distributed regardless of historical context. For all ciprofloxacin-adapted lineages, mutations were seen in genes related to DNA/RNA synthesis as expected, as well as in genes related to membrane, flagella, efflux pumps, metabolism, and transcriptional regulators. Mutations related to the MexAB-OprM, MexCD-OprJ, and MexEF-OprN efflux pumps (mostly in genes encoding negative regulators of the pumps) are seen in the ciprofloxacin-adapted lineages, reflecting the ability of these different pumps to extrude ciprofloxacin; however, no mutations were seen in genes related to MexXY-OprM, even though this pump is also known to contribute to fluoroquinolone resistance [44]. Further experiments in measuring the gene expression of the different efflux pumps may help elucidate the roles that these pumps play in contributing to the different drug-order–specific effects.
Next, we sought to determine if the patterns in mutated genes could explain the mechanisms of some of the drug-order–specific effects that were observed in the MIC time courses. We first discuss the cases of resensitization or maintenance of high resistance in which the 1-drug–evolved lineages were subsequently adapted to the other 2 drugs or to LB (Fig 3A). While subsequent adaptation of Day 20 PIPR to LB and tobramycin maintained high piperacillin resistance, subsequent adaptation to ciprofloxacin led to full resensitization to piperacillin (Fig 3A [top]). We hypothesize that these differences stem from the different efflux pump-related genes that were mutated in these lineages (Fig 5, dashed-purple boxes). Evolution of the Day 0 Ancestor to piperacillin resulted in 2 different SNPs in nalD, and 1 SNP in mexR across the 4 biological replicates of Day 20 PIPR, likely leading to the overexpression of the MexAB-OprM efflux pump [45]. We suspect that MICPIP remained high during subsequent adaptation to LB and tobramycin due to continued overexpression of MexAB-OprM.
However, when Day 20 PIPR was adapted to ciprofloxacin, several mutations occurred in genes related to other efflux pumps, including 1 in mexA, 2 in nfxB, and 2 in mexS (Fig 5, dashed-purple boxes). In particular, mexS encodes a negative regulator of the expression of MexEF-OprN, and mutations in this gene likely lead to the overexpression of the efflux pump [47]. Interestingly, expression of MexEF-OprN has been observed to correlate inversely with the expression of MexAB-OprM [47,48]. Hence, we suspect that the resensitization to piperacillin when Day 20 PIPR was subsequently adapted to ciprofloxacin may be have been due to a concurrent decrease in MexAB-OprM expression (leading to reduced piperacillin efflux) as MexEF-OprN expression increased. That is, it is possible that the mutations that occurred during ciprofloxacin adaptation that led to the overexpression of MexEF-OprN negated the effects of the mutations that occurred during prior piperacillin adaptation that led to overexpression of MexAB-OprM. Furthermore, we observed no mutations in efflux pump–related genes in Day 40 PIPRTOBR (Fig 5, dashed-purple boxes), which supports the notion that because no mutations occurred, which would have negatively correlated with the expression of MexAB-OprM, expression of this efflux pump was maintained throughout the subsequent adaptation to tobramycin, and hence the MICPIP remained high.
We observed that subsequent adaptation of Day 20 TOBR to LB and ciprofloxacin resulted in a partial resensitization to tobramycin, and that while subsequent adaptation to piperacillin also led to a significantly lower MICTOB, it was not as low as that of Day 40 TOBRLB and TOBRCIPR (Fig 3A [middle]). In this case, the partial resensitization during subsequent adaptation to LB may be attributable to adaptive resistance of aminoglycosides in P. aeruginosa. Adaptive resistance is a phenomenon in which resistance to a drug is transiently induced in the presence of the drug, and resistance recedes upon the removal of the drug [49]. In contrast to acquired resistance, which is mediated through genetic mutations, adaptive resistance is explained by phenotypic alterations that allow for temporary increases in resistance. P. aeruginosa is known to exhibit adaptive resistance to aminoglycosides [50,51], and it is primarily mediated through up-regulation of MexXY-OprM during drug exposure and subsequent down-regulation after the removal of the drug [52]. We suspect that the partial resensitization during subsequent ciprofloxacin adaptation is also a consequence of adaptive resistance once the tobramycin selection pressure is removed. We further speculate that during the initial adaptation to tobramycin, the increase in tobramycin resistance was a combination of adaptive resistance and acquired resistance from accumulation of the mutations as seen in Day 20 TOBR. Thus, the resensitization during subsequent LB and ciprofloxacin adaptation was not a full resensitization but rather a partial one, perhaps reflecting the remaining contribution of the acquired resistance. Lastly, with regards to Day 40 TOBRPIPR, it is unclear how subsequent piperacillin adaptation seemingly resulted in the maintenance of high MICTOB compared to that of Day 40 TOBRLB and TOBRCIPR. We hypothesize that the subsequent piperacillin adaptation somehow counteracted the resensitization effects of adaptive resistance, even when the tobramycin selection pressure was removed.
The mechanism of ciprofloxacin resensitization when Day 20 CIPR was subsequently adapted to LB, piperacillin, and tobramycin remains unclear (Fig 3A [bottom]). While reversion of aminoglycoside sensitivity has been the most characterized case of adaptive resistance in P. aeruginosa, other studies have suggested that adaptive resistance may be prevalent in other classes of antibiotic classes as well, and that it may be mediated by epigenetic processes such as methylation and stochastic gene expression [53], particularly affecting the expression of efflux pumps [54]. It could be possible that adaptive resistance partially explains the resensitization to ciprofloxacin. We also note that qualitatively, there was much more variability in the MIC time courses between the individual replicates of the CIPR lineages, as seen by the larger error bars in Fig 2I, compared to that of the PIPR (Fig 2A) and TOBR (Fig 2E) lineages. Taken together, further investigation of the partial ciprofloxacin resensitization is needed.
While we observed clear cases of collateral sensitivity develop to piperacillin and tobramycin during the course of ciprofloxacin adaptation (S4 Fig), other adaptive evolution studies of P. aeruginosa evolved to ciprofloxacin showed mixed results. In one study, the adaptation of P. aeruginosa ATCC 27853 to ciprofloxacin showed no change in the MIC of 3 different beta-lactams (including piperacillin-tazobactam), nor of tobramycin [34]. In another study, while no statistical significances were assigned, adaptation of P. aeruginosa PAO1 to ciprofloxacin appeared to result in slight collateral sensitivities to piperacillin-tazobactam and tobramycin in some of their replicates. Nevertheless, in our study, we hypothesize that the collateral sensitivity to piperacillin and tobramycin during ciprofloxacin adaptation is attributable to the mutations seen in nfxB (which encodes a transcriptional repressor that regulates MexCD-OprJ [55]) in the Day 20 CIPR lineages. Three of the Day 20 CIPR replicates had deletions in nfxB (15, 13, and 16 base pairs), likely resulting in the inactivation of NfxB and concomitant up-regulation of MexCD-OprJ and increased ciprofloxacin resistance [56]. In fact, nfxB mutants have been reported to be hypersusceptible to certain beta-lactams and aminoglycosides [57,58].
Lastly, with regards to the decreased rate of tobramycin adaptation given a history of prior piperacillin adaptation (Fig 3B), we attribute this effect to the large chromosomal deletions that were sustained in 3 of the 4 Day 20 PIPR replicates. The consequences of these deletions are discussed in more detail in the subsequent sections of the manuscript. In summary, based on the genomic mutations, we have presented our interpretations of potential mechanisms that contribute to the drug-order–specific effects. These include how historical context can influence the frequency of mutations in certain genes, the varying contributions of adaptive and acquired resistance to total resistance, and specific cases of inverse correlation of the expression of different efflux pumps. While mutations are likely not the sole determinants of the differences [34,59], many of the observed genomic mutations can partially explain the drug-order–specific effects.
One striking mutation we observed was that 3 of the 4 replicates of Day 20 PIPR (Day 20 PIPR-1, -2, and -3) had large, approximately 400 kbp deletions (corresponding to approximately 6% of the genome) in a conserved region of the chromosome (Fig 4 [large red rectangles]; S5 Data), suggestive of selective genome reduction [60–63], and have been associated with directed repeats [64] and inverted repeats [65] at the boundaries of the deletions. These large deletions were also fixed in the corresponding Day 40 PIPRTOBR, Day 40 PIPRCIPR, and Day 40 PIPRLB lineages. Interestingly, the 3 PIPR lineages with these large deletions hyperproduced the brown pigment pyomelanin during piperacillin evolution, and this visually observable phenotype also persisted when evolved to tobramycin (PIPRTOBR), ciprofloxacin (PIPRCIPR), and LB (PIPRLB). The loss of hmgA as part of the large chromosomal deletions correlates exactly with the pyomelanin phenotype of these lineages. Indeed, hmgA mutants of P. aeruginosa hyperproduce pyomelanin [66]. This observation shows that evolving to piperacillin results in a high probability of sustaining large deletions spanning hmgA, which results in the pyomelanogenic phenotype. However, when we evolved the Day 20 TOBR and CIPR lineages to piperacillin to yield the Day 40 TOBRPIPR and Day 40 CIPRPIPR lineages (4 replicates each), none of them became pyomelanogenic, suggesting that prior history of tobramycin or ciprofloxacin adaptation leads to a lower propensity of becoming pyomelanogenic when subsequently evolved to piperacillin. Interestingly, 1 of the Day 20 TOBR replicates became pyomelanogenic when subsequently evolved to ciprofloxacin (Day 40 TOBRCIPR-2). Hence in this study, pyomelanin hyperproduction is a consequence of piperacillin and ciprofloxacin evolution, yet the likelihood to evolve this visually striking and observable phenotype depends on the history of prior drug adaptation.
While the 3 PIPR lineages that produced pyomelanin were not significantly more resistant to piperacillin than the nonpyomelanogenic PIPR lineage, pyomelanin-producing strains have been observed clinically [60] and have been shown to be more persistent in chronic lung infection models [66]. We tested the reproducibility of this example of a phenotypic dependence on the history of drug adaptation with a higher throughput approach. Starting with clonal populations of Day 0 Ancestor, Day 20 TOBR, and Day 20 CIPR, we seeded 92 replicate populations of each lineage into microplates, and we used a 96-pin replicating tool to serially propagate these populations and evolve them to increasing concentrations of piperacillin daily. The lineages that started from Day 0 Ancestor had the highest propensity to become pyomelanogenic (Fig 6A) compared to lineages starting from Day 20 TOBR (Fig 6B) or Day 20 CIPR (Fig 6C). Still, certain lineages starting from Day 20 TOBR and Day 20 CIPR did also produce pyomelanin, albeit with less propensity than starting from Day 0 Ancestor (Fig 6D; S8–S10 Figs).
To explore the relevance of our laboratory evolution results clinically, we tested for the drug-order–specific MIC evolutionary dynamics in clinical isolates. We first tested the evolutionary dynamics of clinical isolates that were resistant to piperacillin but susceptible to tobramycin and ciprofloxacin. We evolved 3 piperacillin-resistant clinical isolates of P. aeruginosa to piperacillin, tobramycin, and ciprofloxacin for 10 days and tracked how the piperacillin resistance changed in these lineages. If the results from the adaptive evolution experiment applied to these piperacillin-resistant clinical isolates, then we would expect that evolving to tobramycin would not affect the high piperacillin resistance, but evolving to ciprofloxacin would restore susceptibility to piperacillin. As mentioned previously, evolving Day 20 PIPR to LB did not result in a reduction of MICPIP, which suggests that the resensitization to piperacillin when Day 20 PIPR was evolved to ciprofloxacin was a consequence of the switch to the ciprofloxacin drug pressure. Of the 3 isolates we tested, the evolutionary dynamics of 2 of these isolates matched these expectations (Fig 7; S11 Fig and S3 Data). After normalizing to Day 1 MIC values, the MICPIP after 10 days of ciprofloxacin adaptation was significantly less than the MICPIP after 10 days of LB adaptation in isolate #2 (Fig 7B, p < 0.05) and in isolate #3 (Fig 7C, p < 0.001), indicating resensitization to piperacillin during ciprofloxacin adaptation. This observation suggests that the MIC evolutionary dynamics we observed are not limited to laboratory strains of P. aeruginosa and may be observed in diverse strains of P. aeruginosa, including those originating from human patients. Note that these 3 clinical isolates were isolated from different patients, and their phylogenetic relatedness between each other and to the laboratory PA14 strain used in our study is untested. In isolate #1, there was no significant difference in the normalized MICPIP values after 10 days of adaptation to tobramycin, ciprofloxacin, and LB (Fig 7A, p = 0.237, one-way ANOVA). Interestingly, this isolate evolved higher levels of piperacillin and ciprofloxacin resistance than the other 2 isolates (S11 Fig and S3 Data), which suggests the possibility that adaptation to ciprofloxacin in these higher piperacillin-resistant cultures could still result in a restoration of piperacillin susceptibility.
In the next set of evolution experiments, we investigated the role that the large chromosomal deletions play in a drug-order–specific effect. We had observed that compared to the Day 20 PIPR replicate that did not have a large deletion, the 3 Day 20 PIPR replicates with the large deletions, when subsequently evolved to tobramycin, developed less tobramycin resistance (S3 Data and S12 Fig). This observation suggests that the large deletions were involved in reducing the subsequent rate of tobramycin resistance evolution given a prior history of piperacillin adaptation. A recent study isolated 4 pairs of clinical isolates of P. aeruginosa, in which each pair consisted of a pyomelanogenic (PM) isolate and a “parental wild-type (WT)” nonpyomelanogenic isolate [64]. In each of the 4 pairs, the only genomic difference between the pyomelanogenic (denoted APM, BPM, CPM, and DPM) and its corresponding parental wild-type isolate (denoted AWT, BWT, CWT, and DWT) was the presence of large chromosomal deletions that overlap with parts of the deletions seen in Day 20 PIPR-1, -2, and -3 (Fig 8E; S5 Data). Indeed, all of the large deletions encompass hmgA, whose loss accounts for the pyomelanin phenotype. We used these 4 pairs of clinical isolates to test the hypothesis that the large deletions play a role in lowering the rate of tobramycin resistance evolution. We evolved the 4 pairs of isolates to tobramycin using the same daily serial passaging technique used throughout this study and tracked the MICs of tobramycin, piperacillin, and ciprofloxacin over the course of 15 days (Fig 8; S3 Data and S13 Fig). At the end of the 15 days, we saw that APM, BPM, and CPM had lower relative increases in MICTOB, compared to AWT (p < 0.01), BWT (p < 0.05), and CWT (p < 0.05), respectively (Fig 8A–8C). These data provide support for the idea that the large chromosomal deletions do indeed play a role in reducing the rate of tobramycin adaptation, and potentially even in limiting the maximum level of tobramycin resistance that can be developed comparatively. In the case of the fourth pair, we saw that DWT and DPM had comparable increases in MICTOB over the course of the tobramycin adaptation (Fig 8D, p = 1.00). It can be speculated that some combination of the presence or loss of specific genes in DPM led to this evolutionary trajectory that is different from the other 3 pyomelanogenic isolates. We would also like to point out that within each pair, the “WT” and “PM” isolates vary in initial Day 1 MICTOB. The BPM and BWT pair was the most disparate pair, as BPM had a much lower MICTOB than BWT (S13 Fig).
Interestingly, a recent study also observed large genomic deletions spanning hmgA when P. aeruginosa PAO1 was evolved to meropenem, which is another beta-lactam antibiotic [65]. These mutants were also pyomelanogenic. The large deletions in both our study as well as this recent study also span mexX and mexY, which encode portions of the efflux pump that is a significant determinant of aminoglycoside resistance [67]. The loss of these genes in the 3 PIPR replicates may partially explain why subsequent tobramycin adaptation is limited compared to the replicate that did not sustain the large deletion.
This study presents evidence of how the evolutionary history of bacterial adaptation to antibiotics can complicate strategies for treating infections and for limiting the further development of multidrug resistance. Exposing bacteria to fluctuating environments has been shown to be a potentially good strategy for slowing down the development of resistance [12,19,68]. More broadly, mechanisms of memory and history dependence in bacterial systems are being uncovered to better understand the dynamics of bacterial survival and adaptation in changing environments [69–71]. For example, a recent study showed that the survival of Caulobacter crescentus in response to a high concentration of sodium chloride stress depended on the duration and timing of an earlier treatment of a moderate concentration of sodium chloride and that this effect was linked to delays in cell division, which led to cell-cycle synchronization [72]. Another study described what they call response memory, which is when a gene regulatory network continues to persist after the removal of its external inducer. The study showed that in E. coli, lac induction transiently continued when the environment was switched from lactose to glucose, which may be beneficial when the environment fluctuates over short timescales [73]. The results of those studies as well as the results from this study challenge the notion that bacteria respond solely to their present environment. Bacteria can encounter different stressors over time such as osmotic, oxidative, and acidic stress, and other studies have looked at how adaptation to 1 stressor protects the bacteria against other stressors [7,74]. Another example of bacteria adapting to changing environments is how P. aeruginosa, which can be found in the natural environment in the soil and water, can readily adapt to a human host under the right conditions and consequently become pathogenic [75].
There are several factors involved in the emergence of antibiotic resistance that are clinically important that are not considered in this study. We have not taken into account any pathogen/host interactions, such as the role of the immune system. We also do not take into consideration the pharmacokinetics of the drug and the time-dependent fluctuation of drug concentration as experienced by the bacteria in a human-host environment. Furthermore, the dosages of clinical regimens are typically much higher than the wild-type MIC, and the evolutionary dynamics of the bacteria under these conditions may be different from those seen in our study, in which the drug pressure is slowly increased over time. We neglect to consider the role of horizontal gene transfer, which is a common mechanism of antibiotic resistance transfer, and focus rather on the role of de novo mutations acquired during adaptation. Because of the nature of the serial passaging method, we may be selecting for fast growers that may not necessarily have mutations that confer the most amount of resistance in terms of the MIC. We used a strong selection pressure in this study by propagating from the highest concentration of drug that showed growth, but it has been shown that weak antibiotic selection pressure can greatly affect the adaptive landscape [76,77]. Lastly, these bacteria were evolved to 1 antibiotic at a time, and we do not know how different mutant lineages would adapt if competed against each other. It would be interesting in the future to conduct competition experiments to measure the fitness of the different lineages with respect to each other.
While adaptive evolution of clinical isolates suggests that the drug-order–specific effects are clinically relevant, actual clinical studies must be performed to test the true clinical applicability of these effects. A major challenge that still needs to be addressed is how to translate the results of in vitro adaptive evolution experiments to effective therapies that can be used in a clinical setting [78]. For example, while we observed that piperacillin adaptation often led to the large chromosomal deletions and concomitant pyomelanin hyperproduction, the clinical isolates that had the large deletions (Fig 8) were not necessarily resistant to piperacillin. On the other hand, the other set of clinical isolates, which did have resistance to piperacillin, did not have the large deletions (Fig 7). Disparities between the phenotypic and genotypic adaptations such as this will need to be studied further in terms of strain-specific differences, actual history of antibiotic exposure, and other factors that are beyond the scope of this study.
Despite these caveats, there are several key factors of this study that provide confidence in the claims made. We saw consistency in the parallel replicates for the treatment lineages. An interesting exception is Day 40 PIPRTOBR-4, which had a higher final tobramycin resistance compared to Day 40 PIPRTOBR-1, -2 and -3, which we believe is attributed to the large genomic deletions seen in the first 3 replicates but not in the fourth replicate. We observed parallel evolution in which several genes were mutated independently across multiple lineages, and overall, there were less than 15 mutations per 20 days of evolution, which suggests positive selection. Furthermore, many of the mutated genes are also observed in clinical isolates of P. aeruginosa, further giving credence to the clinical relevance of these mutations.
As mentioned previously, studies that have looked at alternating treatments of antibiotics have primarily looked at the effects of rapid switching, typically at daily or subdaily intervals. One of such recent studies evaluated how E. coli responded to 136 different sequential treatments of subinhibitory concentrations of doxycycline and erythromycin, with each treatment consisting of 8 “seasons” of 12-hour-long adaptation periods to 1 of the drugs [19]. Using final optical density as an endpoint metric, the study found that 5 of the sequential treatments could clear the bacteria at the end of the eighth season. Interestingly, one of those 5 successful treatments consisted of 4 seasons of erythromycin, followed subsequently by 4 seasons of doxycycline. On the other hand, the treatment consisting of 4 seasons of doxycycline followed by 4 seasons of erythromycin did not manage to clear the bacteria at the end of 8 seasons. While the experimental setup is much different compared to that of this present study in terms of organism, antibiotics used, duration of treatment, and endpoint metric, these 2 treatments (4 seasons of erythromycin then 4 seasons of doxycycline and vice versa) are quite analogous to the types of treatments tested in our present study. The fact that these authors found a difference in the outcomes of this pair of opposite sequential treatments may suggest that drug-order–specific effects similar to those presented in our study may play a role in the evolutionary dynamics of their experiments.
Cycling between 2 drugs that exhibit collateral sensitivity to one another has been proposed and tested as a strategy to slow down the rate of resistance development [14]. Studies that have systematically tested for collateral sensitivities across a variety of antibiotics in E. coli have consistently discovered that when E. coli is adapted to drugs of the aminoglycoside class, it develops collateral sensitivity to several other drugs of different classes including beta-lactams, DNA synthesis inhibitors, and protein synthesis inhibitors [14,16,77]. In our present study, we tested 1 aminoglycoside (tobramycin), and we did not observe any collateral sensitivity arise to piperacillin or ciprofloxacin during adaptation to tobramycin. Instead, we saw collateral sensitivity to piperacillin and tobramycin arise as P. aeruginosa was adapted to ciprofloxacin, which is a DNA synthesis inhibitor. While we only tested 1 drug in each of 3 drug classes, the dissimilarity of collateral sensitivity profiles between those studies and this present study may highlight how collateral sensitivity profiles may be organism-specific and drug-specific. Further supporting this idea, these prior studies also showed that while adaptation to drugs of the aminoglycoside class as a whole tended to lead to collateral sensitivity to other drug classes, not every aminoglycoside drug that was tested induced the same collateral sensitivity profiles.
While we did observe cases of collateral sensitivity, the main focus of our study was not to look at how resistance profiles to other drugs concurrently change during adaptation to 1 drug, but rather to see how adaptation to 1 drug influences the future evolutionary dynamics as the resistant population adapts to a new drug. Additionally, we wanted to see how adaptation to the second drug affected the resistance profile of the drug that the bacteria originally developed resistance to. Our sustained drug adaptation scheme can be thought of as being more akin to month-long antibiotic cycling at the level of the hospital ward or the environments that bacteria in persistent chronic infections are exposed to. The history-dependent evolutionary dynamics presented in this study highlight the complexity of bacterial adaptation to multidrug therapies, serve as a framework for forecasting evolutionary trajectories based on genetic and phenotypic signatures of past adaptation, and ultimately help elucidate our fundamental understandings of the evolutionary forces that drive resistance adaptation.
Asymmetrical evolutionary responses in changing environments have been studied in terms of collateral sensitivity/resistance [14,16], temperature [79], other abiotic stresses [7], and in cancer treatments [80]. Here, we present the concept of drug-history–specific effects in multidrug resistance adaptation, whereby the history of adaptation to 1 antibiotic environment can influence the evolutionary dynamics during subsequent adaptation to another antibiotic environment. These history-specific effects have direct clinical implications on optimizing antibiotic treatment strategies to slow and prevent the emergence of dangerous multidrug-resistant bacterial pathogens.
The set of P. aeruginosa clinical isolates collected from the University of Virginia Health System (presented in Fig 7) were deidentified, did not require Institutional Review Board approval for their use, and were anonymized. The Hocquet P. aeruginosa clinical isolates (which were originally collected by the authors of the Hocquet study [64]; presented in Fig 8) also did not require Institutional Review Board approval for their use and were anonymized.
We evolved, in parallel, 4 independent replicates for each lineage in the primary adaptive evolution experiment and 3 independent replicates for each of the clinical isolates to balance the statistical power of the conclusions with the technical feasibilities of the daily serial propagations. In the primary adaptive evolution experiment, we concluded the 1-drug evolution at the end of 20 days because the resistance levels of the evolved lineages to their respective drugs were saturated or close to saturated at that point. The clinical isolates from Fig 7 and from Fig 8 were evolved for 10 and 15 days, respectively, because the similarities and differences of the drug-specific effects to those of the primary adaptive evolution experiment were readily apparent at that point.
MIC plates were made daily using the broth microdilution method with the standard 2-fold dilution series [81]. LB was used as the growth medium for all experiments (1% tryptone, 0.5% yeast extract, 1% NaCl). Antibiotics tested include piperacillin sodium, tobramycin, and ciprofloxacin HCl (Sigma). Aliquots of 1 mg/ml and 10 mg/ml antibiotic stocks were made by diluting the antibiotic powders in LB and were stored at −20°C. New frozen drug aliquots were used on a daily basis.
A frozen stock of P. aeruginosa PA14 was streaked on an LB agar plate, and a single colony was inoculated into 4 ml of LB, which was then grown overnight at 37°C, shaking at 125 RPM. This antibiotic-susceptible culture, denoted as the Day 0 Ancestor, was diluted to an OD600 of 0.001 (approximately 106 CFU/ml) and then inoculated into 3 identical MIC plates consisting of concentration gradients of piperacillin and tobramycin. A sample of the ancestor was saved in 25% glycerol and stored at −80°C. The 3 MIC plates were used to serially propagate cultures evolved to LB media, piperacillin, and tobramycin, with 4 biological replicates per condition. Wells for growth control (media + culture) and sterility control (media) were included in each MIC plate. For adaptation to LB media, bacteria were sampled from the growth control well. MIC plates were placed in a plastic container (to prevent evaporation) and incubated at 37°C with shaking at 125 RPM (Thermo Scientific MaxQ 4000). MIC plates were incubated daily for approximately 23 hours.
At the end of incubation, growth in the MIC plates was determined using a plate reader (Tecan Infinite M200 Pro). Growth was defined as OD600 > 0.1 after background subtraction. We recorded the MIC of each lineage for each drug, which was defined as the lowest antibiotic concentration tested that did not show growth (S1 and S3 Data). To propagate, cultures were passaged from the highest concentration that showed growth (i.e., MIC/2) from the corresponding MIC drug gradient. For adaptation to LB, cultures were passaged from the growth control well that contained only LB without any drug. For each culture to be passaged, the culture was first diluted by a factor of 1/250 in fresh LB (e.g., 20 μl of the culture was diluted in 5 ml of LB), which was then inoculated in fresh piperacillin and tobramycin drug gradients in the new day’s MIC plate. Wells of the MIC plate thus contained 100 μl of double the final concentration of the antibiotic and 100 μl of the diluted culture. Hence, the cultures were diluted by a total factor of 1/500 daily. Daily samples were saved in 25% glycerol and stored at −80°C. For Day 21, the piperacillin and tobramycin evolved cultures were subcultured in additional MIC plates such that they could subsequently be evolved to tobramycin and piperacillin, respectively.
A similar protocol was used to establish the ciprofloxacin-evolved lineages (CIPR). Starting with a clonal population of the Day 0 Ancestor, 4 replicates were established and propagated daily under ciprofloxacin treatment for 20 days. CIPR was then subpassaged to piperacillin and tobramycin to establish the CIPRPIPR and CIPRTOBR lineages in addition to continued ciprofloxacin evolution.
To establish the PIPRCIPR and TOBRCIPR lineages, bacteria from the frozen stocks of Day 20 PIPR and TOBR were revived on LB agar plates, and clonal populations were evolved to ciprofloxacin to establish these lineages. Similarly, to establish the PIPRLB, TOBRLB, and CIPRLB lineages, bacteria from the frozen stocks of Day 20 PIPR, TOBR, and CIPR were revived on LB agar plates, and clonal populations were evolved to LB.
Lastly, the MIC to ciprofloxacin was retrospectively measured for the Control, PIPR, TOBR, PIPRTOBR, and TOBRPIPR lineages. Frozen stocks were revived and plated on LB agar plates. The notation for the day numbering is such that Day X PIPR means X days exposure to piperacillin. For consistency, stocks were revived from Days 0 (Ancestor), 5, 10, 15, 19, 20, 25, 30, 35, and 39 for Control, PIPR, and TOBR. One day of exposure to ciprofloxacin would yield Days 1, 6, 11, 16, 20, 21, 26, 31, 36, and 40 MICs to ciprofloxacin. For PIPRTOBR and TOBRPIPR, stocks were similarly revived from Days 20, 25, 30, 35, and 39 and exposed to ciprofloxacin to measure Days 21, 26, 31, 36, and 40 MICs to ciprofloxacin. S3 Data shows the MICs to piperacillin, tobramycin, and ciprofloxacin, respectively, for all the lineages. Note that not all drug MICs were measured on a daily basis for all lineages.
During analysis of the mutations, we deduced that there were some cross-contaminations between replicates in a few lineages. Namely, we saw sets of mutations that were identical in 2 replicates. We believed that the most likely explanation was that the following 7 lines were cross-contaminated sometime between Day 21 and Day 40: CIPRPIPR-3, CIPRPIPR-4, TOBR-1 CIPRTOBR-1, CIPRTOBR-2, CIPRTOBR-4, and CIPR-3, where the number denotes the replicate. To redo these lineages, the corresponding Day 20 replicate frozen stocks were revived on LB agar plates. Then clonal populations were used to redo the propagation as described before. For example, CIPR-3 was evolved to piperacillin for 20 days to redo CIPRPIPR-3. We performed Sanger sequencing of replicate-specific mutations (S3 Table) on the Day 40 mutants to confirm successful propagation of the cultures.
Frozen samples of Day 0 Ancestor, Day 20 Control, PIPR, TOBR, CIPR, Day 40 Control, PIPR, TOBR, CIPR, PIPRTOBR, PIPRCIPR, TOBRPIPR, TOBRCIPR, CIPRPIPR, and CIPRTOBR were streaked on LB agar plates and incubated at 37°C. Agar plates were submitted to Genewiz Incorporation for sequencing service. A single colony from each plate was chosen for DNA extraction, library preparation, multiplexing, and sequencing using 101-bp paired-end reads with the Illumina HiSeq 2500 platform. Reads were aligned to the reference P. aeruginosa PA14 genome (NC_008463.1) with coverage ranging from 113X to 759X. This large range is due to the fact that we submitted samples for sequencing in 3 batches and had different numbers of samples for each batch but had relatively the same number of reads per batch. Nevertheless, the coverage was more than sufficient to identify the SNPs, insertions, and deletions in the genomes. The sequencing reads for Day 0 Ancestor and the 56 drug-evolved lineages are available via the NCBI SRA database (www.ncbi.nlm.nih.gov/sra), accession number SRP100674, BioProject number PRJNA376615.
Reads were aligned and mutations were called using the breseq pipeline [82] using default settings. All reported mutations were visually inspected by viewing the read alignments in IGV and the breseq output files, and mutations with less than 80% frequencies were not counted. The full list of mutations is presented in S1 and S2 Tables. The circos software package [83] was used to plot the mutations by genomic position for Fig 4 and the positions of the large chromosomal deletions in Fig 8.
We confirmed some of the mutations using Sanger sequencing. For each of the Day 20 PIPR, TOBR, and CIPR replicates, we chose 1 mutation each to confirm (S3 Table). We also used these to confirm that replicates were not contaminated before submitting them for whole-genome sequencing. These mutations were also confirmed in each of the Day 40 lineages that were derived from the Day 20 PIPR, TOBR, and CIPR replicates.
Clonal populations of Day 0 Ancestor, Day 20 TOBR-1, -2, -3 and -4, and Day 20 CIPR-1, -2, -3, and -4 were grown in LB starting from the frozen samples. These cultures were diluted in LB to OD600 of 0.001. On Day 1, in 96-well plates, 100 μl of the diluted cultures were inoculated with 100 μl of 4 μg/ml piperacillin (to yield a final concentration of 2 μg/ml piperacillin). Ninety-two wells were used to establish independent replicate populations exposed to piperacillin. Cultures were incubated at 37°C with shaking at 125 RPM. On Day 2, replicate populations were passaged using a 96-pin replicator tool (V&P Scientific, VP246, 100–150 μl per pin) into 200 μl of 4 μg/ml piperacillin. This protocol was continued until Day 10 with a final concentration of 20 μg/ml piperacillin. For each plate, 2 wells were used as sterility controls (only LB), and 2 wells were used as growth controls (LB with bacteria, without drug). Photographs were taken daily (S8–S10 Figs), and the number of visibly brown wells was recorded.
Three clinical isolates of P. aeruginosa with high piperacillin resistance and low tobramycin and ciprofloxacin resistance were obtained from the University of Virginia Health System and were evolved to the 3 drugs in the same manner as the main adaptive evolution experiment starting from frozen samples. These isolates were first confirmed to actually be P. aeruginosa with PCR by using primers that specifically amplify the 16S rRNA region of P. aeruginosa [84]. Three replicates of each isolate were evolved to each of the 3 drugs for 10 days, and their MICs to the 3 drugs were measured as before. In separate subsequent experiments, the 3 clinical isolates were evolved to LB with 3 replicates each. The MICPIP was measured for 10 days (S3 Data). This measurement was done by inoculating into piperacillin concentration gradients to measure the MICPIP but sampling and passaging from the “growth control” well (LB with bacteria, without drug) to adapt to LB.
The 4 pairs of clinical isolates of P. aeruginosa from the Hocquet study [64] were evolved to tobramycin for 15 days with 3 parallel replicates each, with the exception of BPM, which had 2 replicates due to cross-contamination in the third replicate. The MICs for piperacillin and ciprofloxacin were also measured every 5 days (S3 Data). At the end of the 15 days of evolution, primers amplifying part of the hmgA gene were used to check for the presence of the gene in the “WT” isolates and the absence of the gene in the “PM” isolates (S3 Table).
All statistical comparisons of MIC values were performed on the log2 transformed values. Unless noted otherwise, one-way ANOVAs were performed on the MICs of the relevant lineages. If the p-value from the ANOVA was less than 0.05, a post-hoc Tukey’s honest significant difference (HSD) multiple comparisons test was then performed to determine which pairs of treatments were significantly different from each other. The Tukey’s HSD tests report 95% confidence intervals for the true mean difference for each pairwise comparison. If the confidence interval does not contain 0, then the 2 groups being compared have significantly different means at the p = 0.05 level. To also assess the comparisons using nonparametric statistic tests, Kruskwal-Wallis tests followed by post-hoc Dunn’s multiple comparisons tests were also performed. All of the Kruskal-Wallis tests yielded comparable results to the one-way ANOVA at the alpha = 0.05 significance level, and the conclusions are the same for the key comparisons that drive the results highlighted in the manuscript. For a complete set of calculations, see S2 Text.
For the comparisons presented in Fig 3, treatments being compared consist of those listed on the x-axis of each graph in the figure. For the comparisons presented in Fig 7, the raw MIC values for each lineage were first normalized by subtracting the average Day 1 MIC of each of their respective lineages. For each of the 3 clinical isolates, a one-way ANOVA and a Kruskal-Wallis test were performed on the Day 10 MICPIP values of the lineages evolved to LB, tobramycin, and ciprofloxacin (piperacillin-adapted lineages were excluded in the comparisons). The Tukey’s HSD test and Dunn’s test were then performed to see if the Day 10 MICPIP values of the lineages evolved to tobramycin and ciprofloxacin were significantly different from the lineages evolved to LB. For the comparisons presented in Fig 8, the raw MIC values for each lineage were first normalized by subtracting the average Day 1 MIC of each of their respective lineages. A two-sample t test and a Wilcoxon rank sum test were performed for the Day 15 MICTOB values of the “WT” and “PM” lineages evolved to tobramycin in each of the 4 pairs of isolates. Calculations were done in MATLAB R2016b, using the functions “anova1” for one-way ANOVA, “multcompare” for Tukey’s HSD test, “ttest2” for two-sample t test, and “ranksum” for the Wilcoxon rank sum test. The Kruskal-Wallis test was done with the “kruskal.test” command in R, and the Dunn’s test was done with the “posthoc.kruskal.dunn.test” command with the PMCMR R package [85].
|
10.1371/journal.pgen.1006959 | Loss of the chromatin modifier Kdm2aa causes BrafV600E-independent spontaneous melanoma in zebrafish | KDM2A is a histone demethylase associated with transcriptional silencing, however very little is known about its in vivo role in development and disease. Here we demonstrate that loss of the orthologue kdm2aa in zebrafish causes widespread transcriptional disruption and leads to spontaneous melanomas at a high frequency. Fish homozygous for two independent premature stop codon alleles show reduced growth and survival, a strong male sex bias, and homozygous females exhibit a progressive oogenesis defect. kdm2aa mutant fish also develop melanomas from early adulthood onwards which are independent from mutations in braf and other common oncogenes and tumour suppressors as revealed by deep whole exome sequencing. In addition to effects on translation and DNA replication gene expression, high-replicate RNA-seq in morphologically normal individuals demonstrates a stable regulatory response of epigenetic modifiers and the specific de-repression of a group of zinc finger genes residing in constitutive heterochromatin. Together our data reveal a complex role for Kdm2aa in regulating normal mRNA levels and carcinogenesis. These findings establish kdm2aa mutants as the first single gene knockout model of melanoma biology.
| Epigenetic modifications of DNA and histones, the major components of chromatin, play a central role in transcriptional regulation. KDM2A is a histone demethylase that integrates DNA and histone modification signatures and is involved in transcriptional silencing through heterochromatin maintenance. Here we show that adult zebrafish homozygous for the orthologue kdm2aa develop melanomas, a malignant form of skin cancer, independently from oncogenes known to drive melanoma formation. We observe that transcript abundance is widely affected in kdm2aa mutants and find that gene expression of several DNA- and histone- modifying enzymes is stably altered. We furthermore demonstrate a specific de-repression of a group of genes encoding zinc finger-containing proteins that has the potential to be involved in transcriptional regulation. We suggest that these molecular disruptions underlie the melanoma formation, as well as the other observed phenotypes such as reduced growth and survival, a male sex bias and an oogenesis defect. This work demonstrates in vivo a role for Kdm2aa as a tumour suppressor and establishes, to our knowledge, kdm2aa-deficient fish as the first single gene knockout vertebrate model of melanoma.
| The World Health Organisation (WHO) reports that 132,000 melanoma skin cancers occur each year across the globe, with increasing incidence rates. Melanomas are cancers of the melanocytes, which are neural crest-derived pigment-producing cells in vertebrates. Accumulation of mutations, often due to UV damage, leads to the transformation of melanocytes to become a melanoma (reviewed in [1]).
Zebrafish models of melanoma provide a tractable resource to study melanoma biology, however current models require lineage-specific overexpression of an activated oncogene such as BRAFV600E, often in a tp53 or mitfa mutant background, to induce melanoma [2–6]. These models have enabled the identification of additional genes implicated in melanoma development by assessing a candidate gene’s ability to accelerate or delay onset of tumour formation. For example, both the histone H3 lysine 9 methyltransferase SETDB1 [7] and the transcription factor SOX10 [8] accelerate melanoma onset when coexpressed with BRAFV600E in a tp53 mutant line, whereas overexpression of HEXIM1 in this system suppresses tumour formation [9].
Setdb1 belongs to the class of chromatin-modifying enzymes that enable the same DNA sequence in every cell to produce distinct transcriptional outputs in different tissues. Chromatin-modifying enzymes function through the chemical modification of DNA or histone proteins to promote transcriptional activation or repression, either through direct alteration of overall chromatin structure, or by altering the ability of effector molecules to bind [10]. Whereas the primary modification found on DNA is cytosine methylation, histones can have a wide variety of post-translational modifications on various amino acid residues [11]. Due to their profound involvement in transcriptional regulation it is not surprising that mutations in chromatin modifiers have been implicated in cancers and developmental defects [12–14]. The general importance of chromatin-modifying enzymes also limits the in vivo study of their function in mammalian models since mice homozygous mutant for a number of different chromatin modifiers are embryonic lethal [15–19].
In order to gain insight into the in vivo function of chromatin regulators we have studied a zebrafish knockout model of the lysine de-methylase KDM2A. KDM2A specifically removes mono- and di-methyl marks on H3K36 [20]. KDM2A has been implicated in the regulation of CpG island promoters [21] and in the silencing of heterochromatin and rDNA repeats [22, 23]. KDM2A is recruited to H3K9me3-modified chromatin in cooperation with HP1 [24] and this interaction is blocked by DNA methylation [21, 25]. KDM2A knockout mice are embryonic lethal at E10.5–12.5 and exhibit severe growth defects [16] pointing to a role for KDM2A during development. Furthermore, cell culture studies suggest a role for KDM2A in cancer development, but there is conflicting evidence as to whether it acts to promote or inhibit tumourigenesis [26–31].
Here we highlight the complexity of the function of KDM2A by demonstrating that the zebrafish orthologue kdm2aa is required at multiple stages throughout the life of the zebrafish. Zygotic homozygous zebrafish carrying mutations in one of the KDM2A orthologues, kdm2aa, escape early embryonic defects and thus enable the interrogation of both embryonic and adult phenotypes. We show that kdm2aa-deficient fish have reduced growth and survival, a strong male sex bias and that females exhibit a progressive oogenesis defect. Furthermore, kdm2aa-deficient fish develop braf-independent, spontaneous melanoma, providing, to our knowledge, the first single gene knockout model of melanoma. Transcriptome analysis of individual kdm2aa mutant embryos reveals widespread effects on transcript abundance as well as stable regulatory responses of epigenetic modifiers of both histones and DNA, and a specific upregulation of a group of previously uncharacterised zinc finger (ZnF) genes located in constitutive heterochromatin. Our results provide insights into the in vivo function of KDM2A throughout the complete life span of a vertebrate model organism and establish kdm2aa-deficient zebrafish as a new model to study the aetiology of triple wild-type melanoma.
We assessed the in vivo function of KDM2A using zebrafish mutants generated by the Zebrafish Mutation Project [32]. KDM2A has two paralogues in zebrafish, kdm2aa (ENSDARG00000059653) on chromosome 1 (chr1) and kdm2ab (ENSDARG00000078133) on chr14 (Fig 1A). Embryonic expression of kdm2ab peaks during blastula stages, whereas kdm2aa expression is highest later in embryogenesis, during gastrula and early segmentation stages (Fig 1I). We raised two premature stop codon alleles affecting kdm2aa and one premature stop codon allele affecting kdm2ab (Fig 1A). kdm2aasa898 and kdm2absa1479 are assumed to produce non-functional protein. kdm2aasa9360 may produce a partially functional protein lacking the F-box and LRRs.
Fish homozygous for kdm2absa1479 showed no phenotype by 5 days post fertilisation (d.p.f.), grew to adulthood in the expected Mendelian ratio and had healthy offspring. We therefore concluded that kdm2ab loss of function (LOF) does not produce an overt embryonic or adult phenotype. Equally, both kdm2aasa898 and kdm2aasa9360 homozygous embryos did not display morphological defects at 5 d.p.f. (S1A and S1B Fig). We also generated double mutants between kdm2aasa898 and kdm2absa1479 to test whether there was compensation between the paralogues, but embryos homozygous mutant for both genes also showed no phenotypic difference to their siblings at 5 d.p.f. (S1 Table).
However, by 30 d.p.f. juvenile fish homozygous for either kdm2aa allele were thinner and shorter compared to their siblings (Fig 1B and 1C, S2 Table and S1 File). The size difference persisted into adulthood at 180 d.p.f. (Fig 1C). We confirmed that this phenotype was due to the loss of kdm2aa function by raising two clutches containing compound heterozygous kdm2aasa898/sa9360 fish (Fig 1B and S1 File). In addition, survival of homozygotes was reduced at 30 d.p.f. from the expected 25% to below 20% and fell further by 90 d.p.f. (Fig 1D, S2 Table and S1 File). Furthermore, incrosses for either kdm2aa allele produced at most two or three females out of a maximum of 20 homozygotes.
We next assessed whether homozygous mutant kdm2aa adults were fertile. Initial intercrosses of kdm2aasa9360/sa9360 or compound heterozygous kdm2aasa898/sa9360 adults produced phenotypically diverse clutches in which some embryos successfully inflated their swimbladders and either developed phenotypically normally (Fig 1E bottom panel), or with only mild defects (Fig 1E middle panel). Later crosses of the same females produced clutches in which over half of the eggs either failed to fertilise or did not divide beyond four cells (S1C Fig). The remaining eggs showed severe cleavage defects with asymmetric division, detaching cells, and slower division rate (Fig 1F). By 24 hours post fertilisation (h.p.f.) about a third of the maternal-zygotic mutant (MZ) kdm2aa-/- embryos had died and those that survived displayed degrees of generalised developmental defects (Fig 1G). This indicated that subsequent intercrosses from the same females displayed a progressive worsening of egg quality, with fewer eggs being fertilised and fewer embryos surviving beyond 24 h.p.f. Double labelling with DAPI and TRITC-conjugated phalloidin of 8–32 cell wild-type and MZkdm2aa-/- embryos confirmed asymmetric cells and unsynchronised division (Fig 1H).
To confirm that this phenotype was caused by kdm2aa LOF in the female, we outcrossed male and female kdm2aasa9360/sa9360 fish to wild-type fish of the same genetic background. Offspring from three kdm2aasa9360/sa9360 males were normal (Fig 1F and S1C Fig). By contrast, the majority of embryos from initial outcrosses of two kdm2aasa9360/sa9360 females died before 5 d.p.f., however some (12/64) embryos survived to 5 d.p.f. with 6 out of 12 showing no obvious phenotype (S1D Fig top panel) and the remaining 6 displaying only localised malformations (S1D Fig). Subsequent homozygous female outcrosses produced clutches with low fertilisation rates and embryos with severe defects very similar to MZkdm2aa-/- embryos (S1C and S1E Fig).
This demonstrates that embryos from oocytes devoid of functional kdm2aa mRNA or protein can develop normally and that the maternally deposited mRNA (Fig 1I) [33] does not explain the lack of phenotype in zygotic homozygous mutants. Instead the increase in unfertilized eggs and severity of the phenotype in the remaining embryos point to a role for Kdm2aa in maintaining the production of healthy oocytes.
From the age of 7 months, we observed that kdm2aa-deficient fish (homozygotes for either allele and also compound heterozygous fish) began to develop suspected cancers. We observed aberrant melanocytic pigmentation at the base of the tail extending into the tail fin (Fig 2Ai), masses behind one eye causing it to protrude (Fig 2Aii) and masses on the body (S2B Fig). 23/92 (25%) of kdm2aasa898/sa898 (Fig 2B) and 20/204 (10%) of kdm2aasa9360/sa9360 (S2A Fig) fish developed these suspected cancers within the first 28 months, whereas none of the heterozygous or wild-type siblings did. Of the 43 fish with potentially cancerous phenotypes, 10 fish had excessive melanocytic pigmentation on their tail, 12 fish had a tumour behind one eye causing it to protrude, 18 fish had a mass on their body, and 3 fish were found to have both excessive melanocytic pigmentation on their tail and a mass on their body.
To confirm that these growths were cancerous, tissue sections from 10 affected fish and 2 control siblings were haematoxylin and eosin (H&E) stained and analysed by two independent clinical histopathologists. Seven of the fish had excessive melanocytic pigmentation on their tails, and all of these fish were diagnosed with spindle cell malignant melanoma on the tail, invading the surrounding skeletal muscle and bone to varying degrees (Fig 2C). Furthermore pigmented melanophages were present in half of the tumours and these cells have previously been reported in zebrafish melanomas [34]. Two of the fish analysed had eye tumours, which confirmed as either spindle cell, or epithelioid and spindle cell melanoma and pigmented melanophages were present in one of the two tumours. A single fish was analysed with a suspected abdominal tumour and this was found to have a nodular lesion around the ultimobranchial body, vena cava and pancreas, composed of epithelioid and spindle cells (S2C Fig). No pigmented melanophages were present. Additionally in internal sections from one of the fish with excessive melanocytic pigmentation on the tail an abnormal spindle-cell proliferation within the proximal intestinal epithelium and the pancreas was found. Given the pigmentation, spindle cell morphology and malignant proliferation these two abdominal tumours are consistent with melanoma, but further analysis would be required for a firm diagnosis.
Three additional affected fish, two with excessive melanocytic pigmentation on the tail and one with an eye tumour, were analysed further using immunohistochemistry. H&E staining of both tail tumours revealed a biphasic appearance, with pseudoglandular or rosette-like structures alternating with areas of spindle cell growth (Fig 2C) and both the pseudoglandular and spindle cell elements stained positive for the melanoma marker Melan-A (Fig 2D) but negative for two alternative melanoma markers S100 and HMB-45 (Fig 2E and 2F). These tumours were also diagnosed as melanoma showing divergent differentiation. Both tumours stained positive for phospho-histone H3 (Fig 2G) indicating that they were mitotically active. To further characterise the pseudoglandular differentiation the tail tumours were stained for the neuroendocrine marker Synaptophysin (Fig 2H) which was negative and for the epithelial marker Cytokeratin (Fig 2I) which was positive. The eye tumour shared many characteristics with the tail tumours and was diagnosed as invasive melanoma; H&E staining revealed spindle cell morphology (Fig 2M), Melan-A and phospho-histone H3 were positive (Fig 2N and 2Q) and S100 and HMB-45 were negative (Fig 2O and 2P). Interestingly, both tail tumours stained positively for phospho-ERK (Fig 2J) indicating activation of the MAPK signalling pathway, whereas the eye tumour was phospho-ERK negative (Fig 2R). One tail tumour and the eye tumour stained positive for phospho-AKT (Fig 3K and 3S) indicating that PI3K signalling was activated, whereas the second tail tumour was phospho-AKT negative (Fig 2L).
To assess the mutational landscape in kdm2aa-deficient fish melanomas, we performed whole exome sequencing on four dissected tumours, adjacent non-tumour control tissue and sibling tissue, and called the single nucleotide variants (SNVs) and small insertions/deletions (indels) present. Across the 11 samples we obtained exome coverage of 50x (S3 Table). Laboratory zebrafish are not inbred and consequently there is a high level of natural variation [35]. We therefore used exome data from 3,811 individual fish generated in the Zebrafish Mutation Project [32] to define a common variant catalogue of 61,276,211 SNVs and filtered the SNVs found in sibling, control and tumour tissues using this variant set. This revealed on average 951 SNVs between sibling fish and control tissues (Table 1). Tumour tissues harboured on average 517 SNVs compared to non-tumour tissue from the same fish demonstrating that the tumours had accumulated mutations and increased their SNV burden by the equivalent of 50% of the normal sibling variation. By contrast the tumours had not increased their indel frequency with each tumour only showing one additional indel when compared to control tissue (Table 1).
Next, we sought to identify potential protein-disrupting mutations. Filtering the mutations for those which are predicted to disrupt the protein (for details see Methods) revealed between 9 and 21 disruptive mutations per tumour compared to control tissue, and there was no overlap between these acquired disruptive mutations in the different tumours (S3 Table). Whole genome sequencing would be needed in order to assess whether mutations were present in non-coding regions.
Although we did not detect a common mutational signature in the tumours analysed, the exome data showed that none of the oncogenes or tumour suppressors most commonly mutated in human melanomas had accumulated exonic mutations (S3 Table). We looked specifically at the 13 significantly mutated genes in human cutaneous melanomas identified by The Cancer Genome Atlas, as well as HRAS and KRAS which were used to categorise the samples and also GNAQ, GNA11, KIT, CTNNB1 and EZH2 which were found to carry mutations at low frequencies in the triple wild-type set of melanomas [36]. We used the Ensembl Compara database [37] to detect orthologues for these genes in our whole exome sequencing data, but none were found to carry potentially disruptive exonic mutations (S3 Table). Analysis of four tumours is not enough to prove conclusively that kdm2aa-deficient melanomas never harbour mutations in common oncogenes. However if mutations were present at the same frequency as in human melanomas there is, for example, a 94.7% probability that we would have detected a braf mutation (present in 52% of human melanomas [36]) and a 73.1% probability that we would have found nras mutations which are present in 28% of human melanomas [36]. Since BRAF and NRAS hotspot mutations are almost mutually exclusive [36] there is close to a 97.6% probability that we would have found either a BRAF or an NRAS hotspot mutation in at least one of the tumours. Therefore the absence of potentially disruptive mutations in any of the 20 genes assessed across the four melanomas supports a role for Kdm2aa in melanoma development independent of common oncogenes and tumour suppressors.
Given the role of Kdm2aa in chromatin regulation [20, 21, 26] we investigated whether loss of Kdm2aa resulted in an altered transcriptional profile which might explain the observed phenotypes. We performed a comparative transcriptome analysis (polyA RNA-seq) on four individual homozygous mutant, heterozygous and homozygous wild-type siblings each from both alleles at 5 d.p.f. and 12 d.p.f. giving us four mRNA expression profiles (Fig 3A). We chose these time points and individual embryos from two different alleles to capture changes at the mRNA level in morphologically normal individuals rather than secondary transcriptional deviations due to size differences, developmental delay and genetic background. Using DESeq2 [38] we determined differential transcript abundance as significant at an adjusted p-value <0.05. This revealed that kdm2aa transcripts were present at lower levels in fish homozygous for either allele and at both stages (log2 fold change between -0.34 and -0.73), indicating nonsense-mediated decay [39] had occurred. We tested for haploinsufficiency in heterozygous animals by running differential analysis between heterozygous and wild-type embryos in both alleles at 5 d.p.f. and 12 d.p.f. This yielded 0 and 1 differentially expressed (DE) genes for kdm2aasa9360/+ at 5 d.p.f. and 12 d.p.f., respectively. By contrast, kdm2aasa898/+ heterozygosity led to 29 (5 d.p.f.) and 80 (12 d.p.f.) DE genes, suggesting a mild haploinsufficiency effect of that allele on mRNA levels (S4 Table). When comparing homozygous mutants with siblings (Fig 3B–3D and S4 Table) between 539 and 1433 genes out of 32,261 detected genes were significantly differentially abundant in the four clutches. The four DE gene lists only had 19 genes in common (S3A and S3B Fig) with an additional 76 DE genes being significant in at least three of the four clutches (S3A Fig). The discrepancy between the clutches could either be due to clutch-specific and/or stochastic effects on transcript abundance or the fact that four biological replicates per condition do not provide sufficient power to detect differential gene expression above individual embryo variability. To test this we combined the samples for each stage and ran the differential analysis of homozygous mutants against siblings while controlling for clutch in the DESeq2 model (Fig 3A and Methods). The combined stage-specific analyses showed large overlap with their respective individual experiments (77% and 72% at 5 d.p.f., 57% and 75% at 12 d.p.f.), confirming that the majority of discrepancy was due to detection power rather than clutch difference (Fig 3B and 3C). The increase in power due to the larger sample size also enabled us to detect over 1100 additional DE genes for each stage (Fig 3B and 3C). Combining all four experiments in the analysis while controlling for stage and clutch identified 3,752 DE genes (Fig 3D). These results are consistent with previous findings that the number of biological replicates is the main factor in the ability to identify differentially expressed genes [40].
Gene ontology (GO) analysis of DE genes using topGO [41] revealed enrichment of a large number of terms relating to translation, DNA replication, energy metabolism, and chromosome organisation in the biological process (BP) domain in the four separate and the combined 5 d.p.f. RNA-seq analyses (S5 Table). The translation enrichment was driven mostly by upregulation of genes encoding ribosomal proteins (19/72 contributing genes in the 5 d.p.f. analysis), translation elongation or initiation factors (14/72) and mitochondrial ribosomal proteins (9/72), which is consistent with KDM2A’s described function in repressing ribosomal RNA genes [23]. This upregulation of ribosomal genes and energy generation processes together with differential expression of DNA replication genes suggests cellular stress.
We also found stage-specific differences. While different terms relating to translation, chromosome organisation and metabolism appeared in all individual analyses, the GO enrichment at 12 d.p.f. also included a large number of terms related to development of different tissues. This is very likely to reflect the emerging growth retardation observed morphologically from 30 d.p.f. onwards. To visualise this stage difference we filtered the lists for terms that are present in the stage analysis as well as their individual experiments (Fig 3G and 3H and S3C and S3D Fig). This showed a dominance of translation, DNA replication and chromosome segregation at 5 d.p.f., whereas the list at 12 d.p.f. contains mostly translation- and development-related terms.
In accordance with the role of Kdm2aa in chromatin regulation the theme of DNA replication and chromatin remodelling represents the core gene expression profile even in the comparatively small set of 95 DE genes that overlapped between at least three individual clutches across both stages and alleles (S3E Fig). Included in this core set are chromatin modifiers such as nsd1b, a methyltransferase for the KDM2A target H3 lysine 36, and the de novo DNA methyltransferase dnmt3bb.2 which is recruited to DNA by H3K36me3 [42], both of which were downregulated (Fig 3E and 3F). By contrast, the gene encoding the Snf2-related CREBBP activator protein Srcap, the catalytic subunit of a protein complex that incorporates the histone variant H2A.Z at promoters and eu- and heterochromatin boundaries was upregulated (S3F Fig) (S2 File for all count plots). This gene expression signature suggests a compensation for loss of H3K36-demethylase activity and a wider concerted response to chromatin disruption.
When plotting up- and downregulated genes separately onto their chromosomes, we noticed enrichment of upregulated genes on the long arm of chr4 (Fig 4A). This region is repeat rich (Fig 4A), and contains extensive constitutive heterochromatin [43, 44]. We therefore speculated that kdm2aa LOF causes generalised de-repression of genes located within this heterochromatin stretch. However, of the 208 genes that were upregulated on the long arm of chr4 in the combined analysis, 183 genes were annotated as containing a zinc finger (ZnF) domain. ZnF domain-containing genes represented 49.2% of detected genes on the long arm of chr4, but rose to 85.5% in the DE gene set and thus demonstrated specific enrichment (Fig 4B). Furthermore, none of the 183 DE ZnF genes on the long arm were downregulated whereas this was the case for 6 of the other 31 DE genes (Fig 4C). Kdm2aa therefore seems to have a function in repressing heterochromatic ZnF genes on the long arm of chr4 in a gene-specific manner. We have shown previously that these genes are normally expressed in a sharp peak at zygotic genome activation [33], pointing to a role for these genes in regulating zygotic transcription.
In this study we have used two non-complementing point mutations to identify a complex set of phenotypes caused by kdm2aa LOF, which affect different stages of development and adulthood: oogenesis is impaired, juveniles display reduced survival and grow to smaller adults with a strong male sex bias. We also demonstrate that Kdm2aa is not required for early embryonic development as a proportion of embryos from early clutches devoid of maternal wild-type transcript or protein develop normally. Furthermore, while oogenesis is abnormal, Kdm2aa is not required for meiosis per se, since embryos from homozygous male outcrosses are phenotypically wild type. Importantly, a significant proportion of mutants develop cancerous growths. All of the tumours analysed were diagnosed as melanomas, however they are atypical given their unusual histologic and immunologic characteristics and the absence of a mutational signature common to human melanomas. Cell culture studies have pointed to a role for KDM2A and other histone demethylases [45] in the development of human cancers, but it is unclear whether KDM2A acts to promote or suppress carcinogenesis [26–31]. Here we demonstrate that in vivo kdm2aa acts as a tumour suppressor. This is consistent with previous studies identifying chromatin modifiers as key players in cancer development [14, 46–49] and makes the kdm2aa mutant the first single gene knockout animal model of melanoma.
It has been shown previously that fish homozygous mutant for genes known to be involved in DNA damage repair, such as brca2, develop as all males [50]. The female-to-male sex reversal is caused by oocyte death, presumably due to an inability to repair the damage caused by recombination during meiosis [50, 51]. The strong male sex bias that we observe in homozygous mutant Kdm2aa adults raises the possibility that the DNA damage response might also be impaired in Kdm2aa-deficient fish.
A defect in DNA damage repair would also fit with the incidence of melanoma, since patients with Xeroderma Pigmentosum (XP) have a vastly increased risk of skin cancer [52]. XP is caused by mutations in genes involved in the nucleotide excision repair pathway which functions to repair bulky DNA helix distorting lesions such as those produced as a result of UV irradiation or endogenous reactive oxygen species [53, 54]. The effects of kdm2aa loss of function on the DNA damage repair pathway thus warrants further investigation.
Our RNA-seq analysis was carried out at 5 d.p.f. and 12 d.p.f. time points where the mutants do not display any discernible morphological phenotype. Nevertheless, we discovered significant effects on mRNA levels, indicating that we were able to identify the transcriptional profile underlying the later observed morphological phenotypes. We were able to confirm the core DNA replication and chromatin remodelling gene signature by examining the DE genes common to either all 4 or at least 3 of the 4 experiments. Out of the 19 DE genes significant in all four sets six genes are known to be involved in chromatin structure and function (rbbp5, smg9, chd3, rad23aa, kdm2aa and nsd1b). This reproducible gene signature suggests that kdm2aa LOF generally affects chromatin structure and function which is a main factor in transcriptional control. The de-repression of ZnF genes in heterochromatin on chromosome 4, which are normally expressed in a sharp peak at zygotic genome activation [33] and therefore likely to be involved in regulation of transcription at that stage, could also contribute to impaired control of gene expression. Consistent with our observations, disruption of transcriptional control is emerging as a key feature of cancer development and is proposed to favour malignancy [49, 55–58].
Disruption to chromatin has been shown to play a role in melanoma development. For example reduced acetylation and H3K4me2/3 marks at specific regions have been observed in a tumourigenic melanocyte cell model system [59]. Furthermore altered expression of chromatin modifiers has been associated with melanoma development. The histone demethylase KDM5B is highly expressed in many cancers [60] including melanoma cell lines and patient tumours and causes a slowing of the cell cycle which promotes resistance to chemotherapeutic drugs [61]. In a zebrafish melanoma model, overexpression of the histone methylase SETDB1 accelerates the onset of melanoma development [7]. Our Kdm2aa-deficient zebrafish model identifies kdm2aasa898 and kdm2aasa9360 as driver mutations in melanoma and therefore fits with current models demonstrating an important involvement of chromatin modifiers in melanoma. In further support of this, a recent study analysing whole genome sequences from cutaneous, acral and mucosal melanomas identified a number of chromatin modifiers as candidate driver genes harbouring protein-disrupting aberrations [62]. KDM2A is not among the commonly mutated chromatin modifiers in melanoma, but code-disrupting mutations have been identified in melanomas and other cancers [63, 64]. We also cannot exclude the possibility that kdm2aa-deficient fish additionally develop other types of cancers which were not assessed in this study.
Immunohistochemistry of Kdm2aa-deficient tumours with antibodies routinely used for clinical melanoma diagnoses revealed that they stained positive for Melan-A, but negative for two other melanoma markers S100 and HMB-45. Whilst this is unusual, a number of human melanomas do not stain positively for all three markers [65–67]. Additionally H&E staining revealed pseudoglandular or rosette-like features alternating with areas of spindle cell growth, and both tail tumours stained focally positive for the epithelial marker Cytokeratin, suggesting divergent epithelial differentiation within a melanoma. Divergent differentiation towards a range of cell types is a well-recognised although rare phenomenon in human melanoma [68] but the significance of this finding in several of our tumours is uncertain. At this time, with the diagnosis of two independent pathologists, these tumours are best classified as melanoma with divergent differentiation, although the atypical nature of the tumours, and the lack of similarity with human and other zebrafish melanomas suggest that additional evidence is needed to confirm the cell of origin.
All three tumours assessed were mitotically active, shown by phospho-histone H3 antibody staining. The rate of mitoses within a tumour has been identified as the second most powerful predictor of patient survival; a mitotic rate of 1 or more per square millimetre is associated with reduced survival [69, 70]. Furthermore MAPK signalling is activated in over 90% of human melanomas [71] and our immunohistochemical analysis showed that despite an absence of exonic mutations in braf or nras, both tail tumours but not the eye tumour had activated MAPK signalling. The eye tumour and one tail tumour however showed activated PI3K signalling. This suggests that there is not a uniform pathway to melanoma development in kdm2aa-deficient fish, but instead activation of either of the two major pathways known to be involved in human melanomas [62] leads to melanoma development in these fish.
This mutant provides an alternative genetic system to study melanoma development to previous zebrafish and mouse models which require overexpression of an activated oncogene or use xenografts [2, 4, 5] (reviewed in [72]). Our RNA-seq data show that key genes in melanocyte development, including mitfa and sox10, are expressed at normal levels. This is in contrast to fish that overexpress activated BRAF in a tp53-deficient background which already show altered expression of neural crest genes by 80 h.p.f. [57]. We also do not find a significant overlap between our core set of 95 genes DE in at least 3 of the 4 clutches and the gene signatures of either MITF high expressing or AXL high expressing human melanoma cells determined by single cell RNA-seq [73]. Taken together this suggests that the emergence of melanoma at later stages is not due to a direct effect on genes involved in melanocyte development. The melanoma predisposition due to a single gene knockout is comparable to deleterious germline variants in a number of genes such as CDKN2A and POT1 that have been shown to underlie familial melanoma cases in human patients [74, 75].
Due to the disparity between common human melanomas and Kdm2aa-deficient tumours this melanoma model is different from classic BRAF mutation model systems. It does not mimic all hallmarks of common melanomas, but it provides a unique opportunity to interrogate the relationship between chromatin regulation and cancer development. Indeed, transcriptional fluctuations rather than acquired mutations have recently been identified to underlie drug resistance in melanoma cells [76] and chromatin regulators have been demonstrated to function not only in melanoma development but also specifically in the emergence of resistance to BRAF inhibitors ([77] and reviewed in [48]).
Taken together, our work interrogates for the first time in vivo and across the vertebrate life span the role of Kdm2aa in development and disease. We uncover a function for Kdm2aa in oogenesis as opposed to embryogenesis and identify its role as a tumour suppressor. This loss of function model will be invaluable to further dissect the interplay of chromatin structure and transcription, and its impact on cancer.
Zebrafish were maintained in accordance with UK Home Office regulations, UK Animals (Scientific Procedures) Act 1986, under project licence 70/7606, which was reviewed by the Wellcome Trust Sanger Institute Ethical Review Committee. Embryos were obtained either through natural matings or in vitro fertilisation and maintained in an incubator at 28.5°C up to 5 days post fertilisation (d.p.f.). The mutant alleles kdm2aasa898, kdm2aasa9360 and kdm2absa1479 were obtained from the Zebrafish Mutation Project [32].
Standard length (SL) and height at the anterior margin of the anal fin (HAA) of anaesthetised offspring from heterozygous intercrosses were measured at 30, 90 and 180 d.p.f. Measurements were taken as previously described [78]. Tissue samples were taken from each measured fish for genotyping either by sacrificing whole individuals at 30 d.p.f. or by caudal fin biopsies at 90 and 180 d.p.f. To test whether there is a difference in SL or HAA as a function of genotype, we performed ANOVA on each clutch to check for significant differences between the three genotype groups of homozygous mutant, heterozygous and homozygous wild-type fish. Post-hoc testing (Tukey HSD) was used to assess which groups differed significantly.
DNA from embryos or fin biopsies was extracted and DNA samples were genotyped for kdm2aasa898, kdm2aasa9360 or kdm2absa1479 using KASP genotyping as previously described [79].
Fish samples were either collected in formalin and sent to Advance Histopathology Laboratory Ltd, 75 Harley Street, London, UK, for H&E staining and analysis, or fixed, processed and stained as described in [80]. Briefly, fish tissue was fixed in 4% PFA at 4°C for 3 days, decalcified in 0.5M EDTA (pH 8) at 4°C for 5 days and transferred to 70% ethanol. It was then processed in 95% ethanol, absolute alcohol, xylene and paraffin wax, embedded in wax blocks, cut into 5 μm thick sections and placed onto glass slides.
Hematoxylin and eosin staining and immunohistochemistry were performed as described in [80]. The slides were de-waxed by xylene and ethanol washes, stained, dehydrated and mounted with DPX. Antigen retrieval for IHC was performed in 0.01 M citrate buffer (1.8 mM citric acid, 8.2 mM sodium citrate, distilled water—pH 6) in a microwave pressure cooker. The samples were stained with the primary antibody (monoclonal mouse anti-human Melan-A clone A103, DAKO, Cat. No. M7196 concentration 1:75, anti-phospho-Histone 3, Cell Signalling Technology, rabbit, 1:200, anti-phospho-p44/42 MAPK (Erk1/2), Cell Signalling Technology, rabbit, 1:400 and anti-phospho-Akt, Cell Signalling Technology, rabbit, 1:50) overnight at 4°C and secondary antibody (HRP rabbit/mouse, DAKO) for 30 min at room temperature. DAKO Real EnVision Detection System (Peroxidase/DAB+, Rabbit/Mouse, Cat. No. K5007) was used to visualise the IHC staining. S100, HMB-45, Cytokeratin AE1/3 and Synaptophysin antibody stainings were performed under standard laboratory conditions at the Immunohistochemistry Laboratory in the Department of Pathology, Royal Infirmary of Edinburgh.
The stained slides were imaged using Pathology Nanozoomer SlideScanner and the images were processed using NDP.2 software.
For DAPI and TRITC-Phalloidin staining, embryos at the 8–32 cell stage were fixed in 4%PFA/PBS overnight at 4°C, washed in PBST (0.1% Tween-20 in PBS) and dechorionated. After 4 x 30 minute washes in 2% Triton/PBS they were incubated with 4',6-diamidino-2-phenylindole (DAPI) (1:300) in PBST and TRITC-Phalloidin (1:200) in PBST in the dark at 4°C overnight. Embryos were washed 3-4x in PBST, mounted in Vectashield Antifade Mounting Medium and imaged using a Leica SP5 confocal microscope.
Using Sera Mag beads, total nucleic acid was isolated from 96 larvae from heterozygous sibling intercrosses for both kdm2aa alleles at 5 d.p.f. and 12 d.p.f. resulting in four experiments. KASP genotyping was performed on all samples to identify 4 individual homozygous mutant, heterozygous and wild-type sibling samples for each of the four experiments. From these 48 samples 300 ng total RNA were used to prepare sequencing libraries with Ambion ERCC spike-in mix 1 (Cat. No. 4456740) according to the manufacturer’s instructions using the Illumina TruSeq Stranded mRNA Sample Prep Kit Set A and B (RS-122-2101 and RS-122-2102). Paired end sequencing with a read length of 75 bp was performed on four lanes of Illumina HiSeq 2500 machines.
Quality control of sequenced samples was performed using QoRTs [81] and 7 libraries showing characteristics of RNA degradation were excluded from further analysis. Sequence was aligned to the GRCz10 reference genome with TopHat 2.0.13, using a known transcripts file from Ensembl v87 (ftp://ftp.ensembl.org/pub/release-87/gtf/danio_rerio/Danio_rerio.Zv9.87.gtf.gz) and the "fr-firststrand" library type option. Read counts were obtained with htseq-count and used as input for differential expression analysis with DESeq2. For the analyses of individual clutch experiments, the DESeq2 model was “~ condition” where the condition is either “hom” or “het_wt”. For the stage-specific analyses, the model was “~ group + condition” with the same conditions as previously and where the group is either “sa898” or “sa9360”, corresponding to the different alleles. For the combined analysis, the model was also “~ group + condition” with the same conditions as previously and where the groups are “sa898_day5”, “sa898_day12”, “sa9360_day5” or “sa9360_day12”, corresponding to the different alleles and stages. Enrichment analysis for Gene Ontology terms from Ensembl v87 annotation was performed with topGO [41] using the Kolmogorov-Smirnov test and the "elim" algorithm with a nodeSize of 10. RNA-seq data were submitted to ENA under Study Accession Number: ERP007082 and to ArrayExpress under Accession Number: E-ERAD-326.
Biopsies were taken from tumours and adjacent non-tumour control tissue of homozygous mutants and from corresponding tissues of wild-type or heterozygous siblings. Dissected tissues were placed in 400 μl of 100 μg/ml proteinase K overnight at 55°C, followed by 30 min at 80°C to heat inactivate the proteinase K. DNA was precipitated by adding 400 μl of isopropanol and centrifuging for 40 min at 4100 rpm at room temperature. DNA pellets were washed twice with 400 μl of 70% ethanol followed by centrifugation at 4100 rpm for 25 min and 10 min, and resuspended in ddH20. The isolated DNA was whole exome enriched using Agilent SureSelect and used to generate standard Illumina sequencing libraries, which were paired end sequenced with a read length of 75 bp using two lanes of Illumina HiSeq 2500 machines. SNVs were called using MuTect [82] and indels were called using Strelka [83]. Known SNPs, obtained from the Zebrafish Mutation Project [32], were removed from the MuTect output. Potential protein-disrupting SNVs were identified using the Ensembl Variant Effect Predictor (VEP) [84] and filtering the output for stop_gained, missense_variant, transcript_ablation, splice_acceptor_variant, splice_donor_variant and frameshift_variant consequences. Whole exome sequencing data were submitted to ENA under Study Accession Number: ERP016095.
|
10.1371/journal.pntd.0001805 | Agrochemicals against Malaria, Sleeping Sickness, Leishmaniasis and Chagas Disease | In tropical regions, protozoan parasites can cause severe diseases with malaria, leishmaniasis, sleeping sickness, and Chagas disease standing in the forefront. Many of the drugs currently being used to treat these diseases have been developed more than 50 years ago and can cause severe adverse effects. Above all, resistance to existing drugs is widespread and has become a serious problem threatening the success of control measures. In order to identify new antiprotozoal agents, more than 600 commercial agrochemicals have been tested on the pathogens causing the above mentioned diseases. For all of the pathogens, compounds were identified with similar or even higher activities than the currently used drugs in applied in vitro assays. Furthermore, in vivo activity was observed for the fungicide/oomyceticide azoxystrobin, and the insecticide hydramethylnon in the Plasmodium berghei mouse model, and for the oomyceticide zoxamide in the Trypanosoma brucei rhodesiense STIB900 mouse model, respectively.
| Even though agrochemistry and infectious disease control have the same principle goal – the suppression of harmful organisms without harming human health and the environment – there have been only very limited activities to exploit this overlap for the development of new antiinfectious drugs so far. In this study and for the first time, over 600 commercial agrochemicals were systematically screened against the infectious pathogens causing malaria, sleeping sickness, Chagas disease and leishmaniasis. Many highly active compounds with known low mammalian toxicity were identified in cell based assays, and the activity of some of them could even be confirmed in first animal model studies. Further expansion of this concept to other pathogens and the examination of analogues of the identified hits, potentially available from agrochemical companies, would allow for a very efficient source of novel drug candidates.
| The Protozoan parasites of the genera Plasmodium spp., Leishmania spp., Trypanosoma brucei spp. and Trypanosoma cruzi, are the disease causative agents threatening entire populations in mainly resource poor countries around the world.
Malaria, due to infection with Plasmodium spp., is one of the most devastating diseases in developing countries, with 216 million cases in 2010, causing an estimated 655,000 deaths per year [1]. Other recent estimates assume up to 1.2 million deaths per year [2]. For the treatment of malaria several highly active drugs are available, like chloroquine, quinine, mefloquine, atovaquone, artesunate, and their analogs. Thus, malaria is often not included in the list of the neglected tropical diseases. Unfortunately, significant resistance to almost all of these drugs has developed; even to the “last resort” artemisinin-derivatives, first cases of delayed clinical efficacy have been reported [3]. Recently, large libraries from pharma companies have been screened against protozoan parasites and some interesting hits [4], [5], [6], [7] have been found, especially against malaria with the spiroindolones currently undergoing clinical evaluation [8], [9].
Most of the promising compounds in the development pipeline are in a rather early clinical stage, so that a high failure rate is expected [10]. Considering the rapid development of resistance, and the challenges seen with the development of malaria vaccines [11], a continuous refilling of research pipelines with compounds in preclinical/clinical evaluation will be necessary, for the long term perspective. Therefore new compounds for resistance management would be highly desirable, even if they might not show the same remarkably high activity levels as the recently promoted peroxide candidates like OZ439 [12]. In addition, the global malaria agenda has shifted from the mere control of clinical cases to malaria elimination and eventually eradication urgently requiring transmission blocking agents [13].
Human African trypanosomiasis (HAT), also known as sleeping sickness, is caused by infections of T. b. rhodesiense and T. b. gambiense. Populations living in remote rural areas of sub-Saharan Africa are at risk of acquiring HAT. The disease burden in 2000 was estimated at 1.3 Mio DALYs (Disability-Adjusted Life Years) and the estimated number of cases up to 70,000 in 2006 [14]. In recent years the public health situation has improved due to increased monitoring and chemotherapy, resulting in the decrease of reported HAT cases to approximately 10,000 [15]. Only 4 drugs are currently registered as HAT treatment. Pentamidine and suramin are used to treat the hemolymphatic stage (stage 1) of the disease, while melarsoprol and eflornithine (DFMO) are used in stage 2 of the disease when the parasites have invaded the central nervous system (CNS) and which is lethal if untreated. The available drugs are unsatisfactory due to cost, toxicity, poor oral bioavailability, long treatment and lack of efficacy. Melarsoprol is highly toxic, and up to 5% of the second stage patients treated with melarsoprol die of a reactive encephalopathy. Eflornithine treatment is expensive and logistically difficult; it requires four daily intravenous infusions over fourteen days. Recently the eflornithine-nifurtimox combination therapy (NECT) was introduced [16]. The requirement of intravenous administration although reduced to a quarter of injections as compared to monotherapy is still a limitation, with a need for new and more easily administrable drugs.
Trypanosoma cruzi infection elicts Chagas disease and is an important public health problem causing approximately 14,000 deaths and 0.7 Mio DALY annually [17]. Treatment options are limited due to toxicity of available drugs, parasite resistance, and poor drug activity during the chronic phase of the disease. Currently there are two medications being used to treat Chagas disease, nifurtimox and benznidazole [18]. Severe toxicity and long treatment requirements are associated with both drugs [19]. Therefore new medications are badly needed for treating this disease especially in its chronic phase.
Leishmaniasis causes approximately 50,000 deaths and 2.1 Mio DALY annually [20]. It threatens about 350 million people around the world and 12 million people are believed to be infected, with 1–2 million estimated new cases every year [21]. Widely used medications are still based on i.v. application of antimony compounds like stilbogluconate, resulting in severe side effects. More modern, but also more expensive medications are liposomal amphotericin B, miltefosine, and paromomycin [22].
Thus new affordable and effective therapies are urgently needed to combat these disastrous diseases. Registration requirements for agrochemicals are in some aspects even more stringent than for pharmaceuticals, as side effects that are tolerated for drugs against many life threatening diseases, are not acceptable for agrochemicals that potentially could enter the food chain [23], [24], [25]. As a consequence, all commercialized agrochemicals must go through broad toxicological profiles including e.g. chronic and reprotoxicological studies in different mammalian species, covering at least part of the preclinical studies required for drug development. Furthermore, agrochemicals are highly optimized on agrochemical pest targets with often good selectivities in mammals and excellent temperature and storage stability. Another interesting feature of commercial agrochemicals is the very low production cost of only a few cent/g, as the compounds are produced in highly optimized processes on the multi-ton scale. Surprisingly, these aspects have not led to a systematic evaluation of agrochemicals for pharmacological use so far [26].
Here we present data of over 600 commercial agrochemicals which have been systematically tested for the first time for their antiparasitic activity.
A library of over 600 compounds (for a list of CAS-numbers and common names of the tested agrochemicals see Supporting Information S1), that are or have been active ingredients in commercial agrochemical products, has been compiled from the BASF compound depository and was dissolved in DMSO stock solutions in a concentration of 10 mg/ml. These samples were then further diluted according to the requirements of the assays. The structural integrity of the dissolved samples has been confirmed subsequently by LCMS-analysis.
All work was conducted in accordance to relevant national and international guidelines. The in vivo efficacy studies were approved by the veterinary authorities of the Canton Basel-Stadt. The in vivo studies were carried out under license No. 1731 and license No. 739 of the Kantonales Veterinäramt, CH-4025 Basel, Switzerland adhering to the Tierschutzverordnung from 23.04.2008 (based on the Tierschutzgesetz from 26.12.2005).
Starting with the analysis of the phylogenetic relationship of the pests combated with agrochemicals, and the most important tropical infectious disease pathogens as defined by WHO [36], the close relationship of oomycetes, to which important agricultural pathogens like potato blight or downy mildew belong, with protozoan parasites was realized [37]. As a result, a first set of oomyceticidal agrochemicals was tested, resulting in a number of interesting hits. Based on this finding, over 600 commercially available agrochemicals were selected and their activity against the tropical disease pathogens Plasmodium falciparum, Leishmania donovani, Trypanosoma cruzi and Trypanosoma brucei rhodensiense tested in cell based screens.
38 agrochemicals with sub-µM activity on T. cruzi were identified, many of which being azoles with P450-inhibiting activity (Figure 2). P450-monoxygenases have been discussed before as targets against T. cruzi, especially the sterol 14α-demethylase [49].
The standard drug benznidazole (LD50 rat p.o. not available) [50], [51] has an IC50 of 1871 nM in this assay.
Ipconazole (LD50 rat p.o. 888 mg/kg), has an IC50 of 3.0 nM, the most active agrochemical against T. cruzi. It is a fungicide used predominantly in seed dressing. The tested material is, like the commercial material, racemic and a mixture of diastereomers, therefore an enantiopure isomer could potentially have even higher activity.
Difenoconazole (LD50 rat p.o. 1453 mg/kg), a broad spectrum and systemic fungicide, showed an IC50 value of 7.4 nM. This commercial agrochemical is again a racemic diasteromeric mixture and could therefore also have intrinsically higher activity as a pure isomer.
Clotrimazole (14 nM), and viniconazole [52] (26 nM), are two azole drugs used against fungal skin infections, that have also been discussed as agro fungicides and therefore have been tested in this screen. As they have a complete pharmacological dossier they might also be interesting drug candidates.
Zoxamide (LD50 rat p.o. >5000 mg/kg), a broadspectrum oomyceticide used in fruits and vegetables, showed 27 nM activity. It is sold and was tested as a racemate. Its mode of action against oomycetes is the inhibition of microtubule formation.
Pyridaben (30 nM), and tolfenpyrad (55 nM), are insecticides/acaricides inhibiting the complex 1 in the mitochondrial electron transport chain.
A number of further azole fungicides showed activities below 100 nM including metconazole 31 nM, tebuconazole 36 nM, bitertanol 35 nM, climbazole 55 nM, prochloraz 69 nM, hexaconazole 73 nM, and fenapanil 99 nM. Further agrochemicals with high activity in this assay were penconazole (130 nM), epoxyconazole (136 nM), imazalil (148 nM), propiconazole (160 nM), fenarimol (193 nM), fluquinconazole (199 nM), picoxystrobin (248 nM), cyproconazole (257 nM), myclobutanil (374 nM), tetraconazole (478 nM), and pyrifenox (491 nM).
In spite of the excellent in vitro activity initial experiments in a T. cruzi mouse model did so far not show in vivo efficacy for selected hits (personal communication Nazaré Soiro).
Against L. donovani only two agrochemicals showed sub-µM activity (Figure 3). The standard miltefosine (LD50 rat p.o. 246 mg/kg) showed in this assay an IC50 value of 250 nM.
Zoxamide (LD50 rat p.o. >5000 mg/kg) showed an IC50 of 250 nM. The oomyceticidal compound has been discussed in the T. cruzi section.
Tolylfluanid (LD50 rat p.o. >5000 mg/kg) resulted in an IC50 value of 861 nM. It is a protective fungicide and oomyceticide with presumed thiol conjugating activity.
Other agrochemicals with moderate activity against L. donovani were flocumafen (2451 nM), dimoxystrobin (3248 nM), bromofenoxin (3839 nM), cyhexatin (4517 nM), and cyazofamid (4988 nM).
Due to the split of most life science companies into their agro- and pharma branches in the 1990s, the companies active in agrochemistry have not been involved in the recent screening activities to identify new drugs against infectious tropical diseases, even though agrochemicals might have a high potential to yield interesting hits for these applications.
In this cooperation between industrial and public partners, it was shown for several commercial agrochemicals that they are highly active against some of the most important pathogens of infectious tropical diseases. Interestingly as anticipated, several of the oomyceticides (strobilurins against P. falciparum, zoxamide against T. b. rhodesiense and L. donovani) were active against these protozoans, but also other agrochemicals (e.g. hydramethylnon against P. falciparum; azoles like iproconazole against T. cruzi) showed very interesting activities. Exemplified by one of the major commercial agrochemicals, the fungicide azoxystrobin, as well as for the insecticide hydramethylnone, the reduction of parasitemia, and significant life extension for P. berghei infected mice was achieved. For zoxamide, an effect against T. brucei in the mouse model was also demonstrated. This successful in vitro– in vivo transfer without galenic optimization could not be taken for granted, as these agrochemicals have not been optimized for mammalian pharmacokinetics.
There is still a high probability that the identified hits in the end might not be suitable for human use, as there are still several hurdles to overcome. However, the results of this highly focussed and relatively low input approach are more promising than could have been hoped for. It is especially noteworthy, that the screen of less than 700 agrochemical resulted in e.g. 24 new sub-µM hits against P. falciparum, compared to 4 new sub-µM hit in over 2687 recently tested commercial drugs (excluding known antimicrobial and anti-cancer a.i.) [54], [55]. This clearly demonstrates that agrochemistry can be a very interesting and so far untapped source of new leads, and maybe even drug candidates, against protozoal diseases.
It would also be very interesting to screen commercial agrochemicals against the pathogens of other neglected diseases, like schistosomes, nematodes, food borne trematodes, diarrhoeal amoebas and also tropical bacterial pathogens, for which good antibiotic cures are missing. These studies are still to be done.
|
10.1371/journal.pcbi.1006525 | Steered molecular dynamics simulations reveal critical residues for (un)binding of substrates, inhibitors and a product to the malarial M1 aminopeptidase | Malaria is a life-threatening disease spread by mosquitoes. Plasmodium falciparum M1 alanyl aminopeptidase (PfM1-AAP) is a promising target for the treatment of malaria. The recently solved crystal structures of PfM1-AAP revealed that the buried active site can be accessed through two channel openings: a short N-terminal channel with the length of 8 Å and a long C-terminal channel with the length of 30 Å. It is unclear, however, how substrates and inhibitors migrate to the active site and a product of cleavage leaves. Here, we study the molecular mechanism of substrate and inhibitor migration to the active site and the product release using steered molecular dynamics simulations. We identified a stepwise passage of substrates and inhibitors in the C-terminal channel of PfM1-AAP, involving (I) ligand recognition at the opening of the channel, (II) ionic translation to the ‘water reservoir’, (III) ligand reorientation in the ‘water reservoir’ and (IV) passage in a suitable conformation into the active site. Endorsed by enzymatic analysis of functional recombinant PfM1-AAP and mutagenesis studies, our novel ligand-residue binding network analysis has identified the functional residues controlling ligand migration within the C-terminal channel of PfM1-AAP. Furthermore, from unbinding simulations of the Arg product we propose a charge repulsion as the driving force to expel the product out from the N-terminal channel of PfM1-AAP. Our work paves the way towards the design of a novel class of PfM1-AAP inhibitors based on preventing substrate entry to the active site.
| Malaria is a tropical disease caused by infections with protozoan parasites of the genius Plasmodium. Currently, the disease results in over 400,000 human deaths per year worldwide and future prevention and treatment is at risk due to the rise of drug-resistant parasites. Plasmodium falciparum M1 alanyl aminopeptidase is an enzyme involved in the terminal stages of haemoglobin digestion by the intra-erythrocytic stages of the parasite. Chemical inhibition of the enzyme activity prevents the supply of amino acids for parasite development within the blood cells and causes death of malaria parasites. Despite being a promising antimalarial drug target, this enzyme was challenging for inhibitor search using random screenings. One of the reasons is the poorly accessible active site. This project involves computer simulations of the enzyme crystal structures to understand ligand recognition and binding to the buried active site. We demonstrate that electrostatic interactions regulate ligand migration to the active site and offer opportunities for rational design of novel inhibitors.
| Malaria is a life-threatening tropical disease caused by parasites of the genus Plasmodium and spread by Anopheline mosquitoes. Malaria remains one of the top pernicious infectious diseases of humans with 212 million infections and 429,000 deaths each year, predominately in children under 5 years of age in sub-Saharan Africa [1]. The prevention and treatment of malaria is under threat due to the spread of parasites resistant to the current frontline antimalarial drugs [2,3]. As a result, there is a pressing need for new antimalarial therapies.
Plasmodium falciparum is the species of the malaria parasite that causes the majority of human disease and the highest mortality [4]. For the development of the intra-erythrocyte stages of parasite, the degradation of host cell hemoglobin is necessary to support protein synthesis and metabolism [5]. Hemoglobin is initially degraded to di- and tri-peptides by numerous parasite proteases within a specialized food or digestive vacuole [6]. Peptides generated by this process are then transported into the parasitic cytosol, where hydrolysis to free amino acids takes place with a help of the cytosolic M1 alanyl aminopeptidase (PfM1-AAP) as well as several other exo-metallo-aminopeptidases [5,7]. PfM1-AAP has a particularly broad substrate specificity cleaving peptide bonds involving hydrophobic, basic and aromatic amino acids [8–10]. Blocking of PfM1-AAP activity with inhibitor compounds such as Bestatin prevents the supply of amino acids for parasite development, therefore, and kill parasites both in vitro and in vivo making this an attractive strategy for design of novel anti-malarial therapies [11].
Information of the enzyme three-dimensional structure is crucial for an understanding of the molecular basis of substrate recognition and compound inhibition and essential for structure-based inhibitor design. In the past decade, 25 crystal structures of PfM1-AAP in the ligand-bound and unbound forms were solved, revealing a common N-fold architecture with bacterial aminopeptidases and binding interactions with inhibitors and products. Structurally, PfM1-AAP includes 26 α-helices and 7 β-sheets that form four domains (Fig 1). The active site with the catalytic Zn2+ ion coordinated by His496, His500, Glu519 and a catalytic water molecule is buried in the core of the protein. The crystal structure complexed with the inhibitor Bestatin [9], a natural analogue of dipeptide Phe-Leu, reveals 8 hydrogen bonds with the active site, where the Phe moiety is in the P1 hydrophobic pocket composed of Val459, Met462, Tyr575 and Met1034, and the Leu moiety sits in the P1’ pocket of Val493, Val523 and Thr492. From the Bestatin-PfM1-AAP complex substrate binding was readily predicted by molecular docking programs.
The crystal structure analysis shows that the active site is buried within the enzyme and can be accessed through two channel openings: a short N-terminal channel with the length of 8 Å and located at the interface between domains I and IV, and a long C-terminal channel with the length of 30 Å within domain IV [9] (Fig 1). Because the crystal structures show a static view of PfM1-AAP-ligand bound state, it remains unclear as to how substrates and inhibitors migrate to the buried active site and how products leave. Moreover, it is unknown whether the migration of substrates is via simple diffusion or regulated by PfM1-AAP and which channel allows the substrates to occupy a suitable configuration to fit into the active site for the catalytic reaction. Understanding the mechanisms of substrate entry and exit would set the groundwork for the development of a novel class of PfM1-AAP inhibitors aimed at blocking these events.
In the present study, we have selected two ligands as substrates, Met-Phe and Arg-Ala, two inhibitors, Bestatin and R5X (Fig 1) and the Arg product to examine substrate and product migration to and from the active site along both channels using multiple steered molecular dynamic (sMD) simulations [12]. Initially, the migration of the ligands in each channels was explored in short 30 ns multiple sMD simulations and then extended to a long 100 ns sMD simulation at a reduced speed. A table with the simulation systems and details of the sMD simulations is shown in the Supporting Information (S1 Table). Such an approach allowed enhanced sampling of ligand configurations along the binding pathways and highlighted ligand-protein interactions involved in the migration. We have also performed classical MD (cMD) simulations of PfM1-AAP and exploited the availability of >20 crystal structures to further extend the results of the sMD simulations. Our developed ligand-residue binding network together with other MD trajectory analyses highlight a stepwise passage of substrates and inhibitors from the external aqueous environment into the PfM1-AAP active site via the C-terminal channel, the mechanism of the product release via the N-terminal channel and the role of water molecules. Our computer simulations of the ligand migration were endorsed by enzymatic assays using functionally-active wild-type recombinant PfM1-AAP and a specific variant, PfM1-AAP Arg969Ala, and thus provide the groundwork for the design of new PfM1-AAP inhibitors for the treatment of malaria that prevent substrate entry and/or product exit from the active site.
The long C-terminal and short N-terminal channels are the main routes for ligands to reach the deeply buried active site of PfM1-AAP. To explore the dynamic properties of the two channels, we first performed cMD simulations of the ligand-unbound PfM1-AAP form and calculation of the radius of gyration (Rgyr) at several areas of the channels. In particular, the average Rgyr was calculated in four regions of the long C-terminal channel and in two regions of the short N-terminal channel. The position of residues selected for calculation of Rgyr is shown in Fig 1 and the residue number of Cα atoms is defined in the Methods section. The external opening and the start of the C-terminal channel cavity have the Rgyr of 8.9 and 8.6 Å, respectively. The channel then enlarges reaching the Rgyr of 12,7 Å and reduces again up to the Rgyr of 9.7 Å in the cMD simulations (Table 1). In the case of the N-terminal channel, the Rgyr is 6 and 7 Å at the internal and external openings, correspondently. Thus, the C-terminal channel is wider at all the length of the channel compared to the N-terminal channel. Fluctuations in the Rgyr are minor for both the channels, indicating that the channels are relatively rigid in the cMD simulations (Table 1).
We then compared the variation in the Rgyr values during the passage of the ligand in the sMD with the cMD simulations. Table 1 shows the results from the 100 ns sMD simulations and S2 Table in the Supporting Information from the 30 ns sMD simulations. The change in the Rgyr values for the C-terminal channel is smaller (±0.1–0.8 Å) compared to the N-terminal channel (±1.2–2.0 Å), indicating that small conformational changes are needed in the C-terminal channel to pass the ligands. By contrast, the N-terminal channel requires more structural changes, which increase with the size of the ligand (Arg versus Arg-Ala, Table 1). Furthermore, we have noted that the substrates typically adopt a bent conformation, resulting in the interaction between the C- and N-terminal ends, as they migrate through the N-terminal channel. This bent conformation has a strain energy of 5 kcal/mol compared to the extended conformation of the substrates observed through the passage of the C-terminal channel. Pulling of the Arg product requires the smallest change in the Rgyr of the N-terminal opening. The Rgyr data demonstrates that the migration of the substrates and inhibitors through the N-terminal channel is more hindered compared to the migration of the smaller Arg product.
To further gain insight into the preferable channel for ligand migration, we calculated the work that ligands undertake while passing through the N- and C-terminal channels using the 30ns and 100ns sMD trajectories (Fig 2). From the multiple 30ns simulations the results show that Arg-Ala, Bestatin and R5X require more work, 113±21, 141±6 and 68±9 kcal/mol to pass through the N-terminal channel compared to the C-terminal channel, 70±7, 82±9 and 47±4 kcal/mol. Less difference in the average work is observed for Met-Phe, 101±13 and 92±25 kcal/mol to unbind through the N- and C-terminal channels, respectively. The work for Arg unbinding is 142±19 kcal/mol in the C-terminal channel and 128±14 kcal/mol in the N-terminal channel, indicating the greater ease to pull Arg through the N-terminal channel. From the single 100 ns runs (Fig 2), the work required for the migration through the N- and C-terminal channels for the large and small ligands is even more distinct, suggesting that in longer simulations the observed pattern of work difference is more notable.
Although, more sMD simulation replicates are required to provide accurate calculation of the work, our data notably indicates that the migration of the substrates and inhibitors through the C-terminal channel is more favourable than through the N-terminal channel. In contrast, our simulations demonstrate that the product leaves the active site from the N-terminal channel. This scenario is in agreement with the proposed mechanism of substrate binding derived from the PfM1-AAP crystal structure analysis [9].
We next focused on the binding process of substrates and inhibitors as they migrate in the C-terminal channel to the active site. To examine a molecular mechanism of ligand migration through the channel entrances, we created the ligand occupancy map from the sMD trajectories. The ligand occupancy map represents a 2D population histogram, defined by the cross section through the channel centre (Fig 3A), which shows the relative position of the ligands in the channel (Fig 3B) during a simulation run. Using this approach, we have identified three main high-occupancy areas (the most frequently stayed regions) within the C-terminal channel, two areas at the entrance of the channel and one area in close proximity to the active site (coloured red). At the centre of the channel, the occupancy area is wider and less dense (coloured green). This reveals that the ligands sample a larger space in the wide part of the C-terminal channel (Rgyr = 12,7 Å). Overall, the ligand occupancy map indicates that the migration of the substrates and inhibitors along the C-terminal channel involves several transient states that are stabilized by interactions with specific residues of the channel.
To identify the amino acid residues that form interactions with the ligand and their importance in ligand migration we developed a ligand-residue binding network (LRBN) based on the calculation of the interaction energy between the ligand and a residue of the binding channel. LRBN aggregates information about ligand-residue interactions from the sMD simulated replicates of all ligand-protein systems in the form of weighted nodes and edges. The LRBN diagram of the migration through the C-terminal channel is presented in Fig 3C. A node represents a residue forming the interaction with the ligand with the node size corresponding to the strength of the average ligand-residue interaction energy calculated from the sMD simulations. The node colour ranks the residues into three groups based on the strength of the interaction energy; thus, red, yellow and grey nodes are the top 25% and 50% of strong ligand-residue interactions and the remaining frequently-occurring interactions, respectively. The nodes are connected to the average ligand binding pathway through the edge at the point of the highest interaction energy. The edge width is proportional to the timeframe of the ligand-residue interaction occurrence and, thus, the thick edge shows a prolonged interaction. Overall, this approach effectively creates a graph of a generic ligand passage through the channel to the active site highlighting residues that form strong and prolonged interactions with a ligand.
From LRBN (Fig 3C), it is evident that the ligand migration through the C-terminal channel is orchestrated by several charged residues, represented by the large nodes. In particular, five positively charged residues, Arg969, Arg489, Lys849, Lys907 and Lys849, and four negatively charged residues, Glu850, Asp830, Glu572 and Asp581 play an important role in the ligand migration and stabilization of the ligand transient stable states shown in the ligand occupancy map (Fig 3A). The specific role of these residues is explained in the following sections.
The crystal structure of PfM1-AAP bound to the Arg product was used as a starting point to study the product migration and release from the N-terminal channel of PfM1-AAP using the 30 ns and 100 ns sMD simulations.
From the proposed catalytic mechanism of the peptide bond cleavage in PfM1-AAP [13,16], the carboxyl group of the amino acid product is deprotonated. The COO- anion of the amino acid product is stabilized by the Zn+2 ion and the OH-group of Tyr381 in the transition state, while the protonated amine product is released from the active site. This mechanism is supported by the available crystal structure of PfM1-AAP bound to the Arg product [17]. The carboxyl anion of Arg is bound to the Zn+2 and the OH group of Tyr381, whereas the protonated amino group of Arg is coordinated by three glutamate residues. We, therefore, use the ionized from of Arg in our simulations.
The Arg occupancy plot shown in Fig 9A reveals an almost continuous narrow high-populated area throughout the entire unbinding pathway. The 2D occupancy map is plotted into the plane defined by the cross section through the N-terminal channel centre (S1 Fig). Although 184 residues of the N-terminal channel were considered in the construction of the LRBN diagram, only a couple of residues were identified as playing some role in the product egress (Fig 9B). Consequently, the high-populated area in the occupancy plot observed is due to sterically restrained migration rather than the formation of transient stable states by intermediate strong interactions with the enzyme. This is in correlation with a small Rgyr of the N-terminal channel and its notable deviation during the ligand migration (Table 1).
As mentioned above, the active site of PfM1-AAP has four negatively charged glutamate residues (Glu319, Glu463, Glu497 and Glu519) that interact with the ammonium of the substrate. Following the substrate cleavage these polar interactions are severed, placing the carboxylic group of the product within a highly repulsive negatively charged environment. Fig 10 represents the sum of the non-bonded interaction energy between the carboxyl group of the Arg product and the glutamic acid bundle along the representative sMD trajectory, quantifying the strong, albeit brief, repulsive force of +24 kcal/mol. This repulsion helps to quickly propel the product out of the active site by breaking the interactions between the Arg product and the active site, and inducing the conformational changes required for the successful migration through the N-terminal channel.
The opening of the N-terminal channel is encompassed by a protruding loop of domain II consisting of residues 570–575, referred to as the P1 pocket loop, and a loop of domain I, residue 317–325, referred to as the D1 loop (Fig 11A). Our sMD simulations indicate that the guanidine group of the product forms interactions with Glu572 of the P1 loop and the carboxyl group of the product forms temporal interactions with Arg325 of the D1 loop (Fig 11B and 11C). Interestingly, the 1μs cMD simulation of PfM1-AAP in the ligand-unbound form shows a salt bridge between Glu572 and Arg325, which is present in 90% of simulations. Calculation of the distance between the centre of mass of P1 and D1 loops shows that the loops are at the distance of 9 Å in the unbound form and 11.5 Å in the product-bound form of PfM1-AAP (Fig 11D). It is apparent that the Glu572-Arg325 interaction holds the loops together closing the access to the N-terminal channel, whereas in the absence of the salt bridge the loops are in an open position facilitating the egress of the product. The closed state of the N-terminal channel may prevent the substrates from prematurely exiting from the buried active site, while only when hydrolytic cleavage occurs the repulsive force separates the loops and allows the product to leave into the external environment (the parasitic cytosol).
To understand the physico-chemical properties between PfM1-AAP and its substrates/inhibitors, we carried out the multiple sMD simulation runs of five ligand-protein complexes using two simulation protocols totalling 2 μs simulation time and explored the atomistic detail of ligand migration. We then developed an efficient analysis of the MD trajectories to derive the common patterns of interactions along the binding channels from this large set of simulated data. Thus, by computing a ligand-residue interaction energy and ranking of the average interaction energy from the generated sMD trajectories, we have built a ligand-residue binding network that represents a general map of interactions during the ligand migration to and from the active site. This network was a key tool in our study to identify the critical residues in the ligand migration, which were then visually scrutinized to uncover their true function in a specific manner.
Our sMD simulation analysis has revealed a non-diffusive binding profile and identified several well-defined steps of ligand migration along the C-terminal channel, in particular: the ligand recognition at the channel entrance and the initial migration into the channel cavity (I), the ionic translation of the ligand to the water reservoir (II), the ligand reorientation in the water reservoir (III) and the final passage in a suitable conformation to the active site (IV). Several charged residues have been found subsequently interacting and pulling the ligands along the C-terminal channel pathway. In particular, we have identified two pairs of residues, Glu850/Lys907 and Asp830/Lys849, which coordinate the recognition and initial migration. Among them Glu850 forms the first strongest interaction with the enzyme, thus serving as a recognition point. Next, Arg969 is predicted to act as an ionic substrate translator, which facilitates the movement of the ligand to the water reservoir. The distinct conformations of this residue within the C-terminal channel that allow it to move the ligand a long distance have also been observed from the available crystal structures and the cMD simulations, validating our sMD findings. Finally, we identify Arg489 that controls the passage of the substrate in the suitable orientation to the active site. Overall, the extensive sMD simulations demonstrate that electrostatic interactions play a primary role in controlling the substrate and inhibitor migration in the C-terminal channel. To examine the conservation of these residues we have built the sequence alignment of PfM1-AAP from the 13 malaria species. We found that the residues are conserved or substituted with a residue of a similar charge (S2 Fig), suggesting a common mechanism of ligand migration in the M1-AAP of malaria species.
Measurement of PfM1-AAP activity in the presence of the haemoglobin-derived L-peptide and E-peptide provides a level of validation for the proposed ligand migration mechanism. In particular, there is a recognition and specificity spot at the entrance of the C-terminal channel, likely served by Glu850, which allows the L-peptide to enter and not the negatively-charged E-peptide. In addition, replacement of Arg969 with an Ala reduces the L-peptide binding, indicating its importance in migration of small ligands. Finally, although the L-peptide binds to the C-terminal channel, it is not cleaved in the active site because it either cannot reach the active site or is not in the suitable orientation to fit to the cleavage site. Thus, the ligand reorientation in the water reservoir is critical for binding to the active site and the large peptide cannot flip and turn in the small water cavity to form a suitable configuration for the active site. This also explains specificity of PfM1-AAP to cleave short (di and three)-peptides.
In all the simulations, the C-and N-terminal channels were fully hydrated. We consistently observe water-mediated hydrogen bonds (from one up to 5 contacts) between the ligand and enzyme during the ligand migration along the channels. The water molecules effectively substitute the contacts occurring between the ligands and the charged residues of the channel allowing a quick migration of the ligand toward the active site. In addition, the ability of water to shield long-range electrostatic interactions may prevent ligands from drifting back to the previous step. The water reservoir, which permits the formation of a water solvation shell around the ligand, assists the ligand to adapt a favourable conformation to best complement the active site.
From the sMD simulations, the sMD pulling of the Arg product along the N-terminal channel is likely accompanied by the repulsion force that pushes the product out from the active site and unlocks the external opening of the channel by breaking the Glu572-Arg325 interactions. These residues are also conserved among malaria species (S2 Fig), suggesting commonalities in the product release. The importance of Glu572 in the substrate binding and inhibitor specificity has been shown in the recent mutagenesis study [18].
Malaria PfM1-AAP has been a challenging target for the development of active site inhibitors; for example, high through-put screening of a >200,000 chemical library did not identify inhibitors with IC50s below 10 uM [19]. The discovered atomistic description of the complex binding process in PfM1-AAP here provides a possible clue to the failure of random screenings and suggests opportunities to design a novel class of antimalarial agents in a rational way. The significance of the identified charged residues in the C-terminal channel can now be further explored by mutagenesis studies. In addition, the intermediate states stabilized by the interactions with the channel residues identified in the C-terminal channel can be probed as binding sites for small molecule ligands that could block or allosterically modulate PfM1-AAP. Notably, examination of the available crystal structures shows the regular presence of a glycerol molecule around Arg489 and Arg969, further supporting our hypothesis of the existence of the alternative binding sites in the C-terminal channel. This work also provides a structural foundation for future optimization of inhibitor-enzyme binding and unbinding rates.
Our computer simulations further suggest that substrates and inhibitors enter the PfM1-AAP via the long C-terminal channel and for products to exit via the short N-terminal channel. We show that the migration of substrates through the C-terminal channel is predominantly controlled via long-range electrostatic interactions. Side chain rotations of key positively charged residues regulate the migration of ligands between the residues, endorsed by competitive enzymatic studies using peptides. The water reservoir in the core of PfM1-AAP is important for reorientation of substrates to fit to the cleavage site. The ligand-residue binding network tool developed here for an efficient analysis of the replicate sMD simulations can be employed in a wide variety of situations to identify and rank the critical residues of a protein interacting with other binding partners. The results presented in this study will facilitate design of efficient malarial M1 aminopeptidase inhibitors, helping in the race for new antimalarial drugs.
The initial coordinates for PfM1-AAP co-crystallized with the ligands Bestatin, R5X and Arg were acquired from the Protein Data Bank (PDB) repository (PDB ID: 3EBH, 4R5X and 4J3B) [9,17,20]. The Met-Phe and Arg-Ala di-peptides were created using the Maestro building tool [21] and docked into the active site (PDB ID: 3EBH) using the Induced fit protocol, version 6.6 [21]. The crystal structures were prepared with the Maestro protein preparation module [21]. The missing side chains were inserted using Prime version 3.8 ]21]. Overlapping hydrogen atoms were refined by hydrogen-only minimization and residues with alternate positions were defined. The engineered mutations (7 within 3EBH & 4R5X and 1 in 4J3B) were reversed to the wild type sequence. To alleviate the protein strain arising from these point mutations a two-stage minimization protocol was applied, i.e. hydrogen only minimization and all atom minimization with variable constraints set from the X-ray derived B factors in order to reduce the deviation from the original crystal structure coordinates. Minimization was performed using the MacroModel module [21]. The biosystems were solvated using TIP3P water molecules and 0.15 M Na+ and Cl- ions in an orthorhombic box of 12 Å size using the System Builder of Maestro GUI [21].
To fully equilibrate the five systems the following 6 step protocol was implemented: (1) simulation in the NVT ensemble with Brownian dynamics at 10 K for 120 ps with small time steps and solute non-hydrogen atoms restrained; (2) simulation using a Berendsen thermostat for 12 ps at the 10 K velocity re-sampling every 1 ps, a fast temperature relaxation constant and non-hydrogen solute atoms restrained; (3) simulation in the NPT ensemble using a Berendsen thermostat and barostat for 12 ps at 10 K and 1 atm, velocity resampling every 1ps a fast temperature relaxation constant, a slow pressure relaxation constant and non-hydrogen solute atoms restrained; (4) solvation of the protein cavity using the solvate pocket script; (5) simulation in the NPT ensemble using a Berendsen thermostat and barostat for 12 ps at 300 K and 1 atm with a fast temperature relaxation constant, a slow pressure relaxation constant velocity resampling every 1 ps and non-hydrogen solute atoms restrained; (6) simulation in the NPT ensemble using a Berendsen thermostat and barostat for 24 ps at 300 K and 1 atm with a fast temperature relaxation constant and a normal pressure relaxation constant. Each system was then subjected to a 70 ns MD simulation with no constraints applied.
Simulation conditions were maintained at 300 K constant temperature by Langevin dynamics and 1 atm constant pressure using the Nose Hoover Langevin piston method and the NPT ensemble. Long-range electrostatic interactions were calculated using the particle mesh Ewald method [22]. The catalytic Zn+2 ion with the pentahedral coordination was used in the simulations (13). A harmonic restraint was applied to the Zn+2 ion throughout all simulations with a force constant of 10 kcal/mol/Å2 to maintain the coordination with the active site residues. The RMSD of the Zn+2 ion from the crystal structure position was 1.5±0.3 Å during simulations, indicating the stability of the Zn+2 ion. The Zn+2 ion was at the distance of 2.3±0.1, 2.2±0.1 and 2.0±0.04 Å with the coordinating residues: His496, His500 and Glu519, respectively, during the simulations. The Zn+2 ion coordinates with a water molecule and the ligand. In the unbinding event, the ligand coordination was broken and substituted with a second water molecule.
sMD simulations were set up and carried out using the Desmond source distribution v3.6.1.1 [23] with the OPLS-2005 [24,25] force field and the biasing force plugin. A biasing potential was applied as a time-dependent harmonic spring. Here, atoms of the ligand are restrained with respect to the atoms of the protein. In this way, a force, which is distributed based on the centre of mass of the ligand and protein atoms, steers the ligand along a specified vector that is determined as a line through the centroid of the ligand and protein atoms. Steering was performed using a pulling velocity of 0.015 Å/ps and a biasing force constant of 10 kcal/mol/Å2 in the 30 ns simulation runs. The pulling velocity was reduced to 0.0045 Å/ps in the 100 ns simulation runs. This protocol represents the optimal pulling velocity and biasing force constant out of the following tested combinations: 0.030, 0.025, 0.020, 0.015, 0.010, 0.005 Å/ps and each with the force constant of 2.5, 5.0, 7.5 and 10.0 kcal/mol/Å2. The input velocities are determined by the Maxwell-Boltzmann distribution at 300 K.
The 1μs classical MD simulations of PfM1-AAP in the ligand-bound and unbound forms (PDB ID: 4J3B and 3EBG) [9,17] were performed. The buffer of water surrounding the protein was increased up to 15 Å. The distance for the short-range component of the electrostatic calculations was increased using a tapering function from 9 to 12 Å with the long-range electrostatic forces calculated every 4 fs (reduced from 6 fs). The harmonic restraint on the zinc ion was kept to preserve the pentahedral coordination.
Rgyr based on selection of the Cα atoms was calculated around the C-terminal channel opening using residues 972, 938, 907, 850, 846 and 1008 (1); around Arg969 selecting residues 969, 829, 900, 934, 1006 and 826 (2); around the water reservoir using 889, 484, 483, 965, 538, 550 (3) and 1042; and around Arg489 using 489, 581, 1038 and 997 (4) using VMD v1.9.2 (26). In the case of the N-terminal channel, Rgyr was calculated in the internal opening selecting residues 320, 305, 575 and 1034 (1); and in the external opening selecting residues 249, 570 and 1073 (2). In the sMD simulations, only frames, where the ligand sits around the selected residues were considered for Rgyr calculation to identify the maximum of Rgyr value variation.
The work required to pull ligands from the active sites was calculated by integrating the force over the ligand travelled distance using the VMD script [26] and the in house R script [27].
The ligand occupancy maps were computed using the in house R script [27]. Initially, all the trajectories were aligned based on the Cα atoms of the protein. The centre of mass (COM) of the ligand was used to derive a positional map for each ligand. The 3D coordinates of the COM of all the ligands were translated to the 2D coordinates using the plane defined by the cross section through the centre of either the C-terminal channel or the N-terminal channel. Next, the 2D coordinates of the ligand COM were used to create a heatmap. The generated graphs show the general binding path and the area sampled by the ligands. The VolMap plugin of VMD v.19.2 [26] was used to show the plane and 2D volume slice in Fig 3A and S1 Fig defined by the cross section through the centre of the channel.
For construction of the LRBN diagrams the ligand non-bonded interaction energy for each residue of the channels was calculated using the ‘analyze_trajectories.py’ script from the Desmond tools [23]. Long-range electrostatics was calculated using the Particle Mesh Ewald method with a cutoff value of 9 Å [22]. The OPLS-2005 force field [24,25] was used. The residues lining the N- and C-terminal channels were determined from a representative sMD simulation of each ligand, where those residues that fall within 4 Å of the ligand as it traverses the channel are selected. A visual inspection of the selection was made to ensure all residues encompassing the channel are included. We selected 214 and 184 residues of the C-terminal and N-terminal channels, respectivly. The LRBN diagrams were computed with the in house R script [27] and visualized with the igraph package [28]. The LRBN aggregates the results of ligand-residue interaction energies for all the simulations and visualizes as an ensemble average in the form of nodes (residues) and edges (the timeframe of residue-ligand interaction). A 2D representation of the ligand average path is derived from the COM of all the ligands. Nodes connect to this path at the point they have the strongest interaction energy. The ligand-residue interaction is counted at the distance of 9 Å. Nodes are ranked into three groups based on the percentile ranking, less than 0.5, between 0.5 and 0.75 and greater than 0.75.
The images were rendered in PyMol v2.0 [21] and VMD v1.9.2 [26]. The graphs of RMSF and ligand occupancies in water were calculated in VMD v1.9.2 and plotted using the R program [27]. The water-mediated occupancy was calculated in VMD using the Tk/Tcl interpreter. The sequence alignment of the M1 aminopeptidase from different malaria species was performed using ClustalX [29]. The videos 1S-2S and 4S were rendered in PyMol v2.0 [21] and video 3S in VMD v1.9.2 [26].
Functionally active recombinant wild-type PfM1-AAP and the variant PfM1-AAP Arg969Ala were expressed in Escherichia coli and purified by affinity chromatography on Ni-NTA columns as described by McGowan et al. [9]. Enzymatic assays were performed using the fluorogenic peptide substrate H-Arg-7-amido-4-methyl courmarin (H-Arg-NHMec) at a concentration of 5 μM in Tris HCl, pH 7.5 at 37°C. After initiating the enzymatic reaction, activity was recorded as relative fluorescent units in a fluorometre with excitation set at 370 nm and emission at 460 nm. The haemoglobin-derived peptides, L-SFPTTK (L-peptide) and E-EKSAVTA (E-peptide), were synthesised by GL Biochem, Shanghai, China, and dissolved in water before use.
|
10.1371/journal.pntd.0004485 | High Resolution Melting Analysis Targeting hsp70 as a Fast and Efficient Method for the Discrimination of Leishmania Species | Protozoan parasites of the genus Leishmania cause a large spectrum of clinical manifestations known as Leishmaniases. These diseases are increasingly important public health problems in many countries both within and outside endemic regions. Thus, an accurate differential diagnosis is extremely relevant for understanding epidemiological profiles and for the administration of the best therapeutic protocol.
Exploring the High Resolution Melting (HRM) dissociation profiles of two amplicons using real time polymerase chain reaction (real-time PCR) targeting heat-shock protein 70 coding gene (hsp70) revealed differences that allowed the discrimination of genomic DNA samples of eight Leishmania species found in the Americas, including Leishmania (Leishmania) infantum chagasi, L. (L.) amazonensis, L. (L.) mexicana, L. (Viannia) lainsoni, L. (V.) braziliensis, L. (V.) guyanensis, L. (V.) naiffi and L. (V.) shawi, and three species found in Eurasia and Africa, including L. (L.) tropica, L. (L.) donovani and L. (L.) major. In addition, we tested DNA samples obtained from standard promastigote culture, naturally infected phlebotomines, experimentally infected mice and clinical human samples to validate the proposed protocol.
HRM analysis of hsp70 amplicons is a fast and robust strategy that allowed for the detection and discrimination of all Leishmania species responsible for the Leishmaniases in Brazil and Eurasia/Africa with high sensitivity and accuracy. This method could detect less than one parasite per reaction, even in the presence of host DNA.
| The different clinical forms of the Leishmaniases range from cutaneous to visceral infections and are caused by organisms belonging to the genus Leishmania. Controversy over the validity of different molecular methods to correctly identify a species hinders the association of a given species with different clinical forms, complicating the prognosis and the development of suitable treatment protocols. A correct identification leads to a better understanding of the action and consequent development of new drugs and immunological reactions. It also provides important information about the relationship of each species with its hosts (humans, animal reservoirs and sandflies) in different geographical areas and ecological situations, helping to design control strategies. Today, PCR is the most commonly used method for Leishmania identification, but even though several targets have been described, no simple and direct protocol has emerged. In this paper, we coupled hsp70 real-time PCR with the determination of amplicon melting profiles in order to explore polymorphic regions by HRM analysis. This methodology yielded discriminatory melting temperature (Tm) values for Brazilian and Eurasian/African Leishmania species. The protocol has proven to be 100% reliable with both clinical and experimental samples. The major advantage of the presently described method is that it is simple, less expensive, highly sensitive and easily automated.
| Leishmaniases are a major worldwide public health problem and manifest themselves as a spectrum of diseases that may be exacerbated by other infections, such as human immunodeficiency virus. According to the World Health Organization, these diseases are endemic in 98 countries on 5 continents, with more than 350 million people at risk [1, 2]. Clinically, Leishmaniases can be broadly divided as either cutaneous or visceral, but neither form is exclusively linked to a particular species. Although cutaneous manifestations of the diseases are not life threatening, these manifestations can result in obstruction or destruction of the pharynx, larynx and nose in their final stages [2]. The visceral form is the most severe form, characterized by fever, loss of weight, splenomegaly, hepatomegaly, lymphadenopathies and anaemia, often with fatal outcomes if not timely treated [3].
The severity of the disease and its therapeutic responses are variable and depend on the patient’s immune response, the Leishmania species and even the parasite strain [4]. In this scenario, the development of optimized protocols for discriminating between the different Leishmania species is extremely useful and important in clinical management and treatment. The ability to evaluate the most appropriate species-specific treatments also supports the elucidation of the mechanisms of action of new drugs and the establishment of new species-specific treatment protocols. Furthermore, the identification of these parasites allows the generation of important data for clinical, epidemiological and ecological studies.
There are very few publications addressing a Leishmaniasis diagnosis using a High Resolution Melting (HRM) analysis, a methodology that detects differences in the nucleotide composition of a specific real-time PCR product. The method is based on thermodynamic differences in the dissociation curve profiles of amplicons generated from real-time PCR. The generated curves are specific signatures that identify polymorphisms due to small differences in nucleotide composition. In spite of the paucity of papers on the HRM method, some workers have already used it to discriminate Leishmania using targets against 7SL RNA [5, 6], haspb [7], the rRNA ITS sequence [8, 9], the rRNA ITS sequence coupled to hsp70 [10, 11] and a FRET-based assay using MPI and 6PGD [12].
Amongst several targets described for Leishmania identification, the heat-shock protein 70 coding gene (hsp70) has proven to be useful in identifying many species of different geographical origins [13–17].
In this work, we propose a more efficient protocol using HRM analyses targeting the hsp70 sequence for the discrimination of seven Brazilian Leishmania species, as well as three Eurasian and African species. This methodology was validated with DNA from reference strains, experimental infections in mice, human clinical samples and naturally infected phlebotomine sand flies.
Promastigotes of L. (L.) tropica (MHOM/SU/60/OD), L. (L.) donovani (MHOM/IN/80/DD8), L. (L.) infantum chagasi (MCER/BR/1981/M6445), L. (L.) major (MHOM/IL/81/Friedlin), L. (L.) amazonensis (MHOM/BR/1973/M2269), L. (L.) mexicana (MNYC/BZ/62/M379), L. (L.) lainsoni (MHOM/BR/81/M6426), L. (V.) braziliensis (MHOM/BR/1975/M2903), L. (V.) guyanensis (MHOM/BR/1975/M4147), L. (V.) naiffi (MDAS/BR/1979/M5533) and L. (V.) shawi (MCEB/BR/84/M8408) were grown at 25°C in M199 medium with 10% fetal bovine serum (Life Technologies, Carlsbad, CA, USA). Procyclic forms of Trypanosoma cruzi (Y strain) and T. brucei (427 strain) were grown at 28°C in liver-infusion-tryptose medium and SDM-79, respectively, with 10% fetal bovine serum (Life Technologies). Human DNA, FMUSP-IOF-2016, obtained from USP Medical School, was used in specificity tests.
DNA samples from reference strains were purified by a salting-out procedure using an adaptation of the protocol described by Miller et al. 1988 [18]. Approximately 2.5 x 109 promastigotes in stationary growth culture were centrifuged at 3000 x g for 10 min at 25°C. The cells were resuspended in 6 mL of lysis buffer (10 mM Tris-HCl, pH 7.4; 400 mM NaCl; 2 mM EDTA) and lysed by the addition of 600 μL of 10% SDS. After overnight digestion with 1 mg of proteinase K at 37°C, 2 mL of saturated NaCl solution was added to lysate, and then, the lysate was vigorously mixed for 15 seconds and centrifuged for 15 minutes at 25°C for the removal of precipitated proteins. Two volumes of cold absolute ethanol were added to the supernatant, and the precipitated DNA was washed with 70% ethanol and resuspended in 1 mL of TE buffer (10 mM Tris, pH 7.4; 1 mM EDTA).
DNA from samples obtained from fresh humans biopsies, collected by doctors at Clinical Hospital of Medical Faculty USP, or fixed and paraffin-embedded samples from the collection of Instituto Evandro Chagas, (Belem-Para) were used in accordance to the norms established by the National Committee of Ethics in Research (Comissão Nacional de Ética em Pesquisa, CONEP/CNS), resolution 196/96 with the approval of the Ethics in Research Committees of the Institutions of origin (CAPPesq no. 0804/07, IEC n°. 0029/2007).
Fresh experimentally infected BALB/c mice samples of L. (L.) amazonensis or L. (V.) braziliensis were obtained 6 weeks after infection when the animals were sacrificed and tissues were collected and DNA was obtained as described below; the procedures involving the use of BALB/c mice had the approval of the Ethical Committee for use of Animals of Biomedical Sciences Institute of University of São Paulo (CEUA-ICB-USP), under protocol #145 of October 20th, 2011, according to Brazilian Federal Law 11.794 of October 8th 2008.
DNA from infected phlebotomines captured in nature were purified using the commercial DNeasy Tissue & Blood kit (QIAGEN, Hilden, Germany), according to the manufacturer´s manual. Paraffin-embedded samples were prepared according to de Lima et al. 2011 [19]. The DNA concentration was measured by spectrophotometry.
Initially, we amplified the hsp70 234 bp fragments for all species analyzed in this study using the primers described by Graça et al. [17]. The alignment of the nucleotide sequence of those fragments was used to design primers for HRM analysis. Oligonucleotides used in the PCR assays to amplify a 144 bp fragment of hsp70 (amplicon 1) were hsp70C reverse, previously described by Graça et al. 2012 [17], and a new forward oligonucleotide designed and named hsp70F2 (5’–GGAGAACTACGCGTACTCGATGAAG–3’). For the amplification of a 104 bp fragment of hsp70 (amplicon 2) specific to the species from the L. (Viannia) subgenus, the oligonucleotides hsp70F1 (5’–AGCGCATGGTGAACGATGCGTC–3’) and hsp70R1 (5’–CTTCATCGAGTACGCGTAGTTCTCC–3’) were designed. The hsp70 amplicon sequences are shown in Fig 1 and indicate the position of the primers. Conventional PCR reactions were performed on a Mastercycler Gradient Thermocycler (Eppendorf, Hamburg, Germany) with TopTaq Master Mix (QIAGEN) in a final volume of 25 μL with 200 nM of each primer and 50 ng of genomic DNA as a template. The thermal cycling conditions were as follows: an initial denaturation step of 94°C for 5 min, followed by 40 cycles of denaturation at 94°C for 1 min, annealing at 60°C for 30 sec and extension at 72°C for 30 sec, with a final extension at 72°C for 10 min. Real-time PCR reactions were performed using MeltDoc Master Mix for HRM with the fluorophore SYTO9 (Life Technologies) in a final volume of 20 μL with 200 nM of each primer and 50 ng of genomic DNA. The real time amplification conditions were as follows: an initial denaturation step at 94°C for 5 min, followed by 40 cycles of denaturation at 94°C for 30 sec and annealing/extension at 60°C for 30 sec, with the acquisition of fluorescent signals at the end of each extension step, followed by the dissociation curve for HRM analysis in Thermocycler PikoReal96 (Thermo Fisher Scientific, Walthman, MA, USA).
The 234 bp hsp70 fragment produced by conventional PCR, as described by Graça et al. 2012 [17], from each Leishmania species used in this study was purified and cloned into a pGEM-T vector using the pGEM-T Easy Vector System (Promega, Madison, WI, USA) and E. coli SURE competent cells. The recombinant plasmids from at least three colonies were purified, and they were sequenced with T7 and SP6 primers and the BigDyeTerminator v3.1 Cycle Sequencing Kit (Applied Biosystems, Foster City, CA, USA). The sequencing was performed on an ABI 3130 XL Platform (Life Technologies).
Recombinant plasmids containing the hsp70 target were linearized with ScaI. The plasmid copy number was calculated considering the molar mass concentration, and a serial dilution on a tenth proportion was used to produce standard curves for each quantification test. The quality parameters for the standard curves were obtained by PikoReal Software (Thermo Fischer Scientific) analysis, including the correlation coefficient, linear dynamic range and PCR efficiency.
HRM assays were performed at the end of each real-time PCR. The amplicon dissociation analysis was performed by capturing fluorescence signals in 0.2°C intervals and holding for 10 seconds in each range of the melting curve (between 60°C to 95°C). The acquisition of fluorescence data and the construction of dissociation profiles were performed using PikoReal96 software. HRM software normalizes melting curves relatively to values from pre- and post-melting point assigned as 100% and 0%, respectively. Then the software determines the normalized difference that means the signal-to-noise ratio difference of each sample versus a user-defined sequence that can be any. The call efficiency is the benchmark measured in percentage of the similarity between two dissociation profiles using fluorescence and Tm values as parameters. The software performs a paired comparison between the profile of the sample of unknown identity and each standard and chooses the standard that has the closest value. The “call” identity refers to the designation allotted to the sample being identified based on that of the closest standard.
The graphs containing the means and standard deviations of the Tm values obtained by the HRM analyses were made in GraphPad PRISM v. 6.02 software.
The hsp70 sequences deposited in GenBank for L. (L.) tropica (FN395025.1), L. (L.) donovani (AY702003.1), L. (L.) infantum (HF586351.1), L. (L.) major (HF586346.1), L. (L.) amazonensis (EU599090.1), L. (L.) mexicana (EU599091.1), L. (L.) infantum chagasi (FN395036.1), L. (V.) braziliensis (GU071173.1), L. (V.) guyanensis (EU599093.1), L. (V.) lainsoni (GU071174.1), L. (V.) naiffi (GU071183.1) and L. (V.) shawi (GU071177.1) were used for oligonucleotide design. DNA from all Leishmania reference strains analyzed in this study was used as templates in conventional PCR, and the amplicons were cloned and sequenced to confirm the sequences to those deposited in GenBank. The obtained hsp70 amplicon sequences were then aligned, and we chose regions containing polymorphic sites to be used in HRM methodology (Fig 1).
The two pairs of oligonucleotides depicted in the alignment produced the two expected PCR fragments for all Leishmania reference strain DNA used as a template. The 144 bp amplicon 1 is the PCR product used in the amplification of all Leishmania species. The 104 bp amplicon 2 was produced by the oligonucleotide pair designed for species of the L. (Viannia) subgenus (Figs 1, S1 and S2).
The average and standard deviation of the melting temperature (Tm) of each amplicon was determined in duplicate from three independent experiments using 50 ng of DNA as a template from each reference species. The melting profiles and obtained Tm values of hsp70 amplicon 1 for all species studied are presented in Figs 2 and 3 and Table 1. For a reliable discrimination, we calculated the dispersion of Tm values and only considered differences in Tm values exceeding 0.3°C (Fig 2).
The standard curves for the quantification assays using the cloned target showed good linear correlations (0.99 for all curves) and efficiencies varying from 92,37 to 97.23% for all tested species, in the range of 101 to 107 copies (S3 Fig). Moreover, to evaluate the specificity/sensitivity of hsp70 amplicon 1 as a target, HRM assays were performed using genomic DNA from the seven references species of Leishmania in proportions of 1:1 or 1:100 in relation to a human reference DNA (FMUSP-IOF-2016), and the call identification agreed 100% with the reference samples, even in samples where the call efficiency was approximately 75% (Table 2).
To test if the initial amount of target DNA caused a variation in the Tm, serial dilutions containing 50 ng to 50 fg (DNA amount corresponding to 5.0 x 105 to 0.5 of parasite) of Leishmania DNA from reference strains were used as a template to produce both hsp70 amplicon 1 (Fig 4A) and hsp70 amplicon 2 (Fig 4B). The Tm variation obtained for both amplicons in each species showed that some species presented a fluctuation of Tm values that overlapped with other species.
In the case of overlapping Tm values for amplicon 1, a sequential discrimination can be performed by HRM analysis of amplicon 2. This amplicon is specific for the L. (Viannia) subgenus species, allowing the segregation of two patterns that group L. (V.) guyanensis, L. (V.) lainsoni and L. (V.) shawi with Tm = 83.92 ± 0.04°C or L. (V.) naiffi and L. (V.) braziliensis with Tm = 84.39 ± 0.04°C (Figs 4 and 5).
The Ct values obtained in the amplification curves of amplicon 2, using DNA of all Leishmania studied indicated that the reactions were at least 5 orders of magnitude more specific to Leishmania (Viannia) species than for the L. (Leishmania) species (Figs 5C and S2), confirming that amplicon 2 can be used to discriminate L. (Viannia) from the L. (Leishmania) species.
Moreover, using the information on the geographical origin of the samples associated with the HRM analysis of hsp70 amplicon 2 allowed for the discrimination between L. (L.) donovani and L. (L.) infantum chagasi; among L. (L.) major, L. (L.) amazonensis, L. (L.) mexicana and L. (V.) lainsoni.
DNA from uninfected mouse, human, or Trypanosoma cruzi and T. brucei were used as templates and compared to the standardized positive range of Tm values for the tested Leishmania species. No cross-reactivity was detected. For these controls, characteristic Tm values (T. cruzi: 83.08 ± 0.07°C and T. brucei: 83.91 ± 0.06°C) or no amplification was observed (mouse and human) (S4 Fig).
The HRM analysis of hsp70 amplicon 1 obtained with DNA from other Leishmania isolates also used as reference strains resulted in a 100% correlation with the Tm values of the reference species used in this study (Table 3). Some of those strains represent isolates obtained from different geographical regions in Brazil, and experimentally corroborated the identification through the HRM protocol for possible polymorphisms.
The intra-specific variability was further assessed by the in silico analysis of polymorphism of 186 hsp70 entries from L. (L.) tropica, L. (L.) donovani, L. (L.) infantum, L. (L.) major, L. (L.) amazonensis, L. (L.) mexicana, L. (V.) lainsoni, L. (V.) braziliensis, L. (V.) guyanensis, L. (V.) naiffi, L. (V.) shawi, L. (V.) peruviana, L. (V.) panamensis, L. (L.) aethiopica, L. (L.) martiniquensis and L. siamensis. All the sequences were aligned to include the regions of amplicons 1 and 2. The aligned sequences were then examined for polymorphisms among species as well as among strains of the same species. We then calculated the percentage of similarity and estimated the theoretical Tm value of both amplicons (S1 Table). If we assume that the nucleotide differences that we detected are real polymorphisms and not sequencing errors then we can see from S1 Table that the differences in the theoretical Tm values of each species results in the same discriminatory pattern. Of the 186 strains analyzed, only two strains of L. infantum, MCAN/IR/96/LON-49 and LEM75/zymodeme1, presented a theoretical Tm value whose difference was higher than 0.3°C. We cannot rule out the possibilities that this difference is in fact a real one, due to sequencing errors or reflects different taxa.
In the absence of bona fide samples we also determined the theoretical Tm of amplicons 1 and 2 (S1 Table) of two Leishmania species found in America, L. (V.) peruviana and L. (V.) panamensis, that occur outside Brazil. The obtained data indicated that these two species could be differentiated from the others L. (Viannia) species by the coupled HRM analysis of the two amplicons.
The theoretical Tm value of the African L. (L.) aethiopica, potentially allowed the discrimination from L. (L.) donovani, L. (L.) infantum and L. (L.) major, but not from L. (L.) tropica (S1 Table). The enriettii complex members L. (L.) martiniquensis and L. siamensis presented identical theoretical Tm values.
To validate the HRM protocol for different types of sample preparations, sixteen DNA obtained from real biological samples, like fresh tissue from hamster inoculated with infected sample from human or dog cases; cell culture of the human isolated strain; human fresh tissue; human paraffin-embedded tissue; tissues from experimentally infected BALB/c mice and naturally infected phlebotomines, that had been previously tested in our laboratory by sequencing of SSU rDNA [20] or by discriminatory PCR targeting g6pd [21], were submitted to HRM analysis. The results obtained presented a correlation with the results obtained with the other targets (Table 4).
The establishment of optimized protocols for the detection and identification of the aetiological agents of Leishmaniases are extremely useful tools in a clinical context. Identifying the species can lead to species-specific treatment protocols to promote a better efficacy of treatment, assessing the need for patient follow up as well as the development and understanding of the mode of action of potential new drugs.
Several methodologies targeting different genomic or mitochondrial DNA have been described in the past 20 years, and PCR is currently the preferred method in studies involving the detection and identification of Leishmania. These methodologies have been developed by designing primers that exploit species-specific sequence polymorphisms in different targets, such as kDNA [22], the SSU rDNA gene [20, 23], the glucose-6-phosphate dehydrogenase gene (g6pd) [21, 24], rDNA internal transcribed spacers (ITSs) [25], hsp70 [13–17] and cysteine proteinase B gene (cpb) [7, 26]. However, none of these methods represents a gold standard because the targeted polymorphisms were unsuitable for simple and direct identification protocols. These PCR analyses involved the use of multiple targets requiring a combination of several primers creating the need of running more than one reaction to identify a single sample. The multiplex PCR that uses several pair of primers in one reaction and restriction fragment length polymorphism analysis (RFLP) of PCR products both need of a subsequent DNA fractionation by gel electrophoresis. These procedures require experienced operators to interpret the results, besides the risk of laboratory contamination with amplicons, due to the manipulation of PCR product.
Another way to exploit DNA polymorphisms is the determination of the C+G composition of PCR products from conserved regions by calculating the Tm of the amplicon in a melting curve. HRM methodology has been successfully used for Leishmania identification using different targets, such as the 7SL RNA gene that discriminated L. tropica, L. major and species that cause visceral Leishmaniases in clinical samples [5, 6]. Additionally, using the same target, researchers determined that rodent Ctenodactylus gundi is a potential host of L. tropica in Tunisia [5]. Polymorphisms on haspb (Hydrophilic Acylated Surface Protein B gene) analyzed by HRM allowed the differentiation of strains of L. (L.) donovani from distinct regions of East Africa [7]. In Southeastern Iran, the rRNA ITS sequence incriminated Phebotomus sergenti as a natural vector of L. (L.) tropica [10], or the discrimination between L. (L.) tropica and L. (L.) infantum in Turkey [9]. HRM analysis of the ITS-1 rRNA region discriminated L. (L.) major, L. (.L) tropica, L. (L.) aethiopica and L. (L.) infantum in samples from Middle East, Asia, Africa and Europe [8]. The combination of two targets, hsp70 and the rRNA ITS1 sequence, using the absolute HRM values allowed for the discrimination of six American Leishmania species [11] and MPI/6PGD-FRET PCR distinguished L. (V.) braziliensis from L. (V.) peruviana [12].
Here, we described an algorithm using HRM methodology for the rapid detection and discrimination of Leishmania species circulating in Brazil and Eurasia/Africa (Fig 6). We used the sequence coding for hsp70, but in order to obtain a discriminatory PCR product, we designed the primers to encompass a region that was no larger than 144 bp and that had relevant polymorphisms for HRM analysis, that is, shifts of AT base pairs to CG or vice-versa. Moreover, to be effective, the total amount of polymorphisms was taken into account, and compensatory changes were avoided. Using these criteria, we obtained two PCR products: amplicon 1 and amplicon 2. Using the algorithm described in Fig 6, the analysis of the produced melting profiles of amplicon 1 for the Brazilian species allowed for the discrimination of L. (L.) i. chagasi, L. (L.) amazonensis/L. (L.) mexicana/L. (V.) lainsoni, L. (V.) braziliensis/L (V.) guyanensis, L. (V.) naiffi and L. (V.) shawi using differences in the Tm of at least 0.3°C. For Eurasian samples, amplicon 1 produced values with the same 0.3°C interval to discriminate L. (L.) tropica from L. (L.) major and from L. (L.) donovani/L (L.) infantum, but these two species cannot be discriminated from each other (Fig 2).
The occurrence of an overlap in the Tm value for the Brazilian species L. (L.) amazonensis and L. (L.) lainsoni after a positive reaction of amplicon 1 can be solved by a positive reaction of amplicon 2. This amplicon sequence is specific for Leishmania (Viannia) species, so L. (L.) amazonensis will not be amplified and L. (V.) lainsoni will present the corresponding Tm value (Fig 5). The occurrence of an overlap in the Tm value for the American species L. (L.) amazonensis and L. (L.) mexicana can be solved by amplicon 1 sequencing because this amplicon is not identical, but there are two mismatches (position 82 A to G and position 100 G to T in L. (L.) amazonensis and L. (L.) mexicana, respectively (Fig 1), that are compensatory in the melting profile. It is interesting to note that these two species are very closely related. Uliana et al. [23] distinguished L. (L.) amazonensis from L. (L.) mexicana by SSU rDNA, but Castilho et al. [21] also failed to distinguish these species by g6pd because the region of the g6pd sequence that was used is identical in the two species. It is also interesting that Hernandez et al. [11], using a larger amplicon (337 bp) of hsp70, succeeded in differentiating L. (L.) mexicana from L. (L.) amazonensis; however, Fraga et al. [13] failed to distinguish these two species using RFLP in another region of hsp70. However, when the complete nucleotide sequence of the hsp70 PCR fragment of 1268 bp is used, the discrimination between the two species can be achieved [27]. These problems once again emphasize that one gene or a particular sequence of a gene is not reliable to define a species or plot its phylogeny. Recently, Real et al. [28] showed that L. (L.) mexicana and L. (L.) major had, respectively, 5 and 7 species-specific orthologous gene families, while L. (L.) amazonensis had 23 different gene families. Moreover, the geographical parameter can also be used; Uliana et al. used SSU rDNA polymorphism to show that these species present a characteristic distribution in Latin America that correlates to monoclonal antibody profiles [29].
The Tm overlap for Eurasian species occurred for L. (L.) donovani and L. (L.) infantum, which presented identical sequences for amplicon 1. Again, the geographical origin of the sample can be used because L. (L.) donovani is more frequently found in India and East Africa and presents anthroponotic behavior. L. (L.) infantum is found in Africa, China and the Mediterranean and shows zoonotic behavior [30]. However, the two species can be discriminated by multilocus enzyme electrophoresis (MLEE) or multilocus microsatellite typing (MLMT) [30]. Recently, the haspb coding region was initially used in a classical PCR coupled to RFLP [31], while the gene coding for cpb was used as a target in conventional PCR [7]. We propose to use the latter in case of doubt between the two species (Fig 6).
The in silico analysis of amplicon 1 and 2 from other Leishmania species from America or from Eurasia/Africa, also indicated the potentiality of the hsp70 HRM protocol to discriminate L. (V.) peruviana, L. (V.) panamensis and L. (L.) aethiopica/ L. (L.) martiniquensis/L. siamensis from L. (L.) donovani and L. (L.) major but not from L. (L.) tropica. It is interesting to note that the ITS-HRM analysis applied to L. (L.) tropica L. (L.) aethiopica, L. (L.) infantum, L. (L.) major and L. (L.) donovani [8] presented exactly the same degree of resolution of the hsp70 HRM described here.
We also noticed that the initial amount of template DNA influenced the Tm determination (Fig 4). This Tm variation could be important in cases where the Tm values are in the same range and can lead to a misidentification if the reference sample is at a different concentration. This is the case for L. (L.) amazonensis and L. (L.) lainsoni. However, as has been previously explained, the use of hsp70 amplicon 2 allowed for the discrimination between these two species. The two other species that presented an overlapping Tm range depending on the initial amount of DNA were L. (V.) braziliensis and L. (V.) guyanensis, which could be discriminated by the use of an HRM analysis on the same amplicon 2.
In fact, when we applied the protocol described here to other Leishmania isolates, the obtained “call” (the identification of the problem sample in relation to the reference samples) presented a 100% correlation with the reference strains (Table 3).
The test of sixteen samples consisting of fresh hamster tissue from animals injected with human or dog biopsy macerates, fresh or paraffin embedded human biopsies, tissues of experimentally infected BALB/c mice or even naturally infected phebotominae, produced identification “calls” comparable to the identification results using SSU rDNA sequencing or g6pd PCR (Table 4), showing that the source of the sample as well as its conservation do not interfere in the HRM protocol. Moreover, the use of HRM protocol is easier than the use of SSU rDNA and/or g6pd PCR, since those methods require either sequencing of the product or three or more distinct PCRs followed by gel electrophoresis analysis.
Overall, the hsp70 HRM protocol described herein accurately and sensitively identified Leishmania species that are important in the majority of cases of Leishmaniases in the Brazil and Eurasia. The test is simple and rapid, and its use in the clinic or in research samples has many advantages, such as a lower total cost for the identification of a sample and other characteristics that facilitate its application. There is no need for sequencing or gel fractionation to analyze the product, thus avoiding laboratory contamination with PCR products because these products are discarded without being manipulated. It also reduces the need for trained personnel to analyze the fractionation profile of an electrophoretic gel or sequencing data to provide a result. Also the HRM assay presents a possibility of quantifying parasites present in samples because it is a real-time PCR-based technique. Moreover, the whole process can be automated because the analyzer software will produce the “call” result by comparing the tested samples to the reference sample identities, which must always be included in the reactions.
In conclusion, the protocol described herein is a low cost, reliable, easy to apply, potentially automated procedure that is a good alternative for the detection, quantification and identification of Leishmania species in biological and clinical samples.
|
10.1371/journal.ppat.1003704 | The Structure of Classical Swine Fever Virus Npro: A Novel Cysteine Autoprotease and Zinc-Binding Protein Involved in Subversion of Type I Interferon Induction | Pestiviruses express their genome as a single polypeptide that is subsequently cleaved into individual proteins by host- and virus-encoded proteases. The pestivirus N-terminal protease (Npro) is a cysteine autoprotease that cleaves between its own C-terminus and the N-terminus of the core protein. Due to its unique sequence and catalytic site, it forms its own cysteine protease family C53. After self-cleavage, Npro is no longer active as a protease. The released Npro suppresses the induction of the host's type-I interferon-α/β (IFN-α/β) response. Npro binds interferon regulatory factor-3 (IRF3), the key transcriptional activator of IFN-α/β genes, and promotes degradation of IRF3 by the proteasome, thus preventing induction of the IFN-α/β response to pestivirus infection. Here we report the crystal structures of pestivirus Npro. Npro is structurally distinct from other known cysteine proteases and has a novel “clam shell” fold consisting of a protease domain and a zinc-binding domain. The unique fold of Npro allows auto-catalysis at its C-terminus and subsequently conceals the cleavage site in the active site of the protease. Although many viruses interfere with type I IFN induction by targeting the IRF3 pathway, little information is available regarding structure or mechanism of action of viral proteins that interact with IRF3. The distribution of amino acids on the surface of Npro involved in targeting IRF3 for proteasomal degradation provides insight into the nature of Npro's interaction with IRF3. The structures thus establish the mechanism of auto-catalysis and subsequent auto-inhibition of trans-activity of Npro, and its role in subversion of host immune response.
| Mammalian cells respond to viral infection by inducing an innate immune response involving interferon-α/β that mediates cellular antiviral defenses. Viruses, in turn, have evolved mechanisms to counter the host's innate immune response by inhibiting the interferon response. Pestiviruses use the virally encoded N-terminal protease (Npro) to suppress interferon-α/β induction. Npro first cleaves itself off from the viral polyprotein using its own cysteine protease activity. Released Npro then interacts with interferon regulatory factor-3 (IRF3), a transcriptional activator of interferon-β, and induces proteasome-mediated degradation of IRF3. We have determined the crystal structure of Npro from classical swine fever virus. Npro has a unique protease fold consisting of two domains. The N-terminal domain carries the protease active site and has the C-terminus, the auto-cleavage site, bound in the active site. Thus, following auto-cleavage at the C-terminus, Npro obstructs the catalytic site preventing further activity, making the protease active only once in its lifetime. The C-terminal domain carries a zinc-binding site that is required for interaction with IRF3. Surface mapping of the Npro residues essential for subversion of interferon induction provides insight into the interaction with IRF3 and its subsequent degradation. To our knowledge, this is the first structure of a direct IRF3 antagonist.
| Cells sense RNA virus infections by pattern recognition receptors (PRR) such as Toll-like receptors and the cytosolic RIG-I and MDA5 that recognize different forms of single-stranded and double-stranded viral RNA [1], [2]. Engagement of these PRR triggers a signaling cascade leading to phosphorylation and subsequent activation of the interferon regulatory factor-3 (IRF3). Activated IRF3 then translocates into the nucleus where it induces transcription of the interferon-α/β (IFN-α/β) genes. This activation is essential for the host to mount innate and adaptive anti-viral responses [3]. Viruses have evolved a multitude of strategies to counter the initial steps of the host's innate immune activation [4], [5]. IRF3 is targeted by many different viruses that use virus-encoded proteins to counteract IRF3 functions. A few examples of viral proteins targeting IRF3 are NSP1 of rotavirus, ICP0 of herpes virus, the leader protein of mengovirus and of Theiler's murine encephalomyelitis virus, the ML protein of Thogoto virus, and the NS1 and NS2 proteins of bovine and human respiratory syncytial virus [6], [7], [8], [9], [10], [11], [12], [13]. The mechanisms of IRF3 antagonism employed by these viral proteins vary, and include inhibition of phosphorylation, nuclear translocation, and assembly of the transcription complex, as well as targeting IRF3 for proteasomal degradation.
Pestiviruses, such as bovine viral diarrhea virus (BVDV) and classical swine fever virus (CSFV) use a virally encoded protein, the N-terminal protease (Npro), to suppress the transcriptional activation of the IFN-α/β genes by interacting with IRF3 and inducing its ubiquitination and proteasome-dependent degradation [14], [15], [16], [17], [18]. Npro also interferes with the activity of IRF7 in plasmacytoid dendritic cells [19]. Unlike the other viral proteins encoded by the pestivirus genome, Npro has no counterpart in the other members of the Flaviviridae family. The Npro protein is a leader cysteine autoprotease that cleaves itself from the nascent polyprotein during translation of the viral mRNA, freeing itself for innate immune suppression activities. Self-cleavage of Npro releases the core protein that becomes a structural component of the virion, probably by associating with the viral RNA genome and forming a nucleocapsid [20], [21], [22]. Npro is not associated with virions and is dispensable for virus replication [23]. Interestingly, following the first self-cleavage reaction, Npro does not possess any proteolytic trans-activity [24]. Sequence comparison with other known protease families showed that Npro has a unique sequence and a novel arrangement of catalytic residues, and hence it is classified as its own C53 protease family [25]. The predicted catalytic triad of Npro is Glu22, His49 and Cys69, which differs from the known catalytic triads in serine and cysteine proteases. For example, the catalytic triad in papain-like cysteine proteases is Cys25-His159-Asn175, and that in subtilisin-like serine proteases is Asp32-His62-Ser245. Additionally, the catalytic activity of Npro is not inhibited by typical cysteine protease inhibitors such as antipain dihydrochloride and by the serine protease inhibitor aprotinin [21]. There is strong evidence that the function of Npro in IRF3 degradation and inhibition of IFN-α/β induction is independent of its autoproteolytic activity, since mutation of the catalytic Cys69 to Ala had a minimal effect on inhibition of IFN-α/β induction by Npro [26], [27]. However, mutation of His49 resulted in a loss of IFN-antagonistic activity, indicating that there is spatial overlap of regions that are involved in proteolysis and those responsible for IRF3 interactions [15], [16], [26]. We previously reported that Npro has two domains, a catalytic N-terminal domain and a zinc-coordinating C-terminal domain, and that a zinc-binding TRASH motif in the C-terminal domain is required for binding IRF3 and subverting IFN-α/β induction [28].
Here, we report the crystal structures of CSFV Npro and a C168A cleavage site Npro mutant to 1.6 Å resolution. To our knowledge, this is the first structure of an IRF3 antagonist. Npro has a unique ‘clam shell’-like protease fold that is distinct from other known proteins including all known proteases. As predicted, Npro consists of two domains, a cysteine protease domain and a zinc-binding domain. The active site of Npro's cysteine protease is formed by a catalytic dyad, Cys69 and His49. Contrary to previous reports, Glu22 is not in the active site. The C-terminus of the protein that constitutes the self-cleavage site is not only bound in the protease active site, but also contributes an integral β-strand to the central β-sheet that makes up the active site. Thus, the C-terminus of Npro occludes the catalytic site following cleavage, inhibiting any trans-activity of the protease and limiting the activity of the enzyme to a single catalytic turnover. The C-terminal domain contains the zinc-binding TRASH motif that is indispensable for binding IRF3 and targeting it for proteasomal degradation. Taken together, the Npro structures presented here establish the mechanism of auto-catalysis and subsequent auto-inhibition of Npro, and provide insight into its interaction with IRF3 in subversion of host innate immune responses.
Both crystal structures of Npro were determined using deletion mutants that lack the first 17 amino acids (Npro-Δ17N). The first 19 amino acids are not essential for proteolytic activity of Npro and can be deleted without affecting the in vitro protease activity [21], [26]. The 19 residue deletion also does not interfere with the ability of Npro to block the IFN-α/β induction in cell culture [26]. Thus the structures represent biologically relevant forms of Npro. The C168A protein was expressed as Npro-Δ17N protein that contains a Cys168 to Ala mutation along with an additional four residues at the C-terminus, 169Ser-Asp-Asp-Gly172, a sequence that corresponds to the first four amino acids of the core protein. This construct was intended to trap a substrate-bound form of Npro; by mutating the cleavage-site residue, Npro should not be able to catalyze self-cleavage at its C-terminus. Unexpectedly, the C168A mutant was active and the SDDG peptide at the C-terminus was cleaved, as confirmed by mass spectrometry (data not shown). Introduction of the Cys168 to Ala mutation resulted in different packing in the crystal lattice. The C168A Npro crystals belong to the spacegroup P212121 with one molecule per asymmetric unit and a solvent content of 54%. In contrast, wild-type Npro crystals were of spacegroup P21212 also with a monomer in the asymmetric unit, but with a solvent content of 35% (Table 1). Although it is not entirely clear why the C168A mutation resulted in such a dramatic change in packing, new crystal contacts at the active site contributed to the stabilization of the helix carrying the nucleophilic Cys69, which was disordered and not visible in the native structure (Figure 1a–b). In the C168A structure, the side chains of Cys69, His49 and the C-terminal carboxyl group of one molecule, along with the side chain of His74 from a neighboring molecule, coordinated a zinc atom (Figure 1c). Since His74 is not conserved among pestiviral Npro proteins, coordination of zinc near the active site is most likely a result of crystal packing interactions and may not be physiologically relevant. Presence of zinc was confirmed by determining the anomalous difference map from single wavelength anomalous dispersion (SAD) data collected at the zinc absorption edge. The distances between zinc and the coordinating groups are 1.9, 2.1, 2.1 and 2.3 Å for the carboxy-terminus, two histidines (His74' and His49), and cysteine (Cys69) residues, respectively, all of which are consistent with distances reported for tetrahedral zinc coordination [29]. The structure of the C168A mutant can be superimposed on the wild type structure with a root-mean square deviation (R.M.S.D.) of 0.38 Å for 136 Cα atoms. Thus, there is no gross structural difference between the wild-type and C168A mutant.
Npro is composed predominantly of β-sheets that adopt a unique ‘clam shell’-like fold. The protein can be divided into two distinct domains, the catalytic protease domain and the zinc-binding domain (Figure 1). The protease domain spans the N-terminus through residue 100 and also includes C-terminal residues 157 to 168. The domain harbors the protease active site along with the C-terminal protease cleavage site Cys168. The protease domain contains mostly coils without regular secondary structure and a single β-sheet formed by strands β1, β2, and β8. The first two β-strands of the sheet are contributed by the first 100 residues in the sequence. The last 6 residues at the C-terminus (163–168) form the final β-strand, and fold back into the protease active site, positioning the C-terminus Cys168 for cleavage (see the next section). The Npro protease domain was predicted to be disordered, perhaps due to the abundance of proline residues; the domain contains twelve prolines corresponding to an average of one proline for every 7 residues. Most prolines are located in the loops on the surface of the protein, contributing to the unique fold of the protein. A search for similar folds using the DALI server resulted in zero instances, indicating that the catalytic domain of Npro has a new fold. The zinc-binding domain of Npro spans residues 101 through 156 and forms an anti-parallel β-sheet consisting of five β-strands, β3, β4, β5, β6, and β7 (Figure 1). This domain carries a conserved metal binding TRASH motif consisting of Cys112-Cys134-Asp136-Cys138 that coordinates a single zinc atom [28]. The TRASH motif is located at one end of the β-sheet. The interface between the protease and zinc-binding domains is mostly hydrophobic, and the C-terminal domain partially covers the final β-strand in the protease domain (see below).
A cysteine protease triad in Npro was predicted to include Glu22, His49 and Cys69 by site-directed mutagenesis in a cell-free translation system [21]. The active site of Npro is solvent exposed on one end of the protease domain (Figure 2a). Cys69 is part of a single-turn helix formed by residues 68–71 in the C168A protein, while it is disordered in the native structure. Among the proposed protease triad, only Cys69, the nucleophile, and His49, the general base, are present to form a catalytic dyad flanking the C-terminal cleavage site. The sulfur atom of Cys69 is 3.8 Å away from the epsilon nitrogen of His49, and 3.5 Å away from the terminal carboxylate of Ala168. These distances are comparable to those seen in crystal structures of papain-like cysteine proteases, although the arrangement of the dyad (His49-Cys69) is different from that of papain (Cys25-His159) and other cysteine proteases. Since a zinc atom is present in the active site (figure 1C), this could lead to small artifacts in the observed active site geometry. Cysteine proteases often contain a stabilizing Asn/Asp residue in the vicinity of the catalytic His. No stabilizing anion group in the vicinity of His49 is apparent. However, the main chain carbonyl group of Asp50 is within a hydrogen-bonding distance of 2.8 Å of the delta nitrogen of His49, and thus could orient the imidazolium ring of His49 during catalysis (Figure 2b). Contrary to previous predictions, Glu22 does not complete a catalytic triad since its spatial location is approximately 23 Å from the nucleophilic Cys69 (Figure 2a). Instead, Glu22 forms a salt bridge with the conserved Arg100. Since mutation of Glu22, either a deletion or an Ala-substitution renders the protease inactive, breakdown of the salt bridge between Glu22 and Arg100 likely destabilizes the structure of the protease domain, resulting in the observed loss of proteolytic activity.
Both Npro structures represent a product-bound form of the enzyme, providing the geometry of the catalytic residues around the scissile bond of Cys168. The C-terminal β-strand (β8) provides substrate specificity by positioning the C-terminal cleavage site residue 168 for hydrolysis near the protease dyad. The C-terminal carboxylate of residue 168, either Cys in the wild-type structure or Ala in the C168A mutant structure, has a clear density with planar geometry (Figure 2b). The terminal carboxylate forms hydrogen bonds with the main chain amides of Gly67, Asp68 and Cys69. These hydrogen bonding interactions would also help stabilize the tetrahedral intermediate during catalysis and thus form the oxyanion hole.
Npro cleaves the peptide bond between Cys168 and Ser169 in the viral polyprotein such that Ser169 then becomes the N-terminus of the core protein. The residues at the C-terminus of Npro (the P sites) are reported to be essential for a functional Npro protease, while the residues following the cleavage site (P' sites) can tolerate many amino acid substitutions [20], [24], [30]. The cleavage site for substrates is defined as …P3-P2-P1- P1'-P2'-P3'…, where a cleavage occurs between the P1 and P1' residues. The last seven C-terminal residues Pro162-Leu-Trp-Val-Thr-Ser-Cys168 (P7 to P1) form the final β-strand (β8) which is an integral part of a β-sheet. The β-strand is partially occluded by the C-terminal domain, such that the P7 to P4 site is not solvent accessible and enclosed in a hydrophobic environment. In addition to the typical main chain hydrogen bonding with β1, the β8-strand is stabilized by side-chain interactions, most notably hydrophobic interactions, involving Leu163, Trp164, and Val165. For example, Trp164 (P4) shown to be critical for protease activity is located in a hydrophobic environment consisting of the conserved residues Leu45, Leu47, Arg51, Tyr82, Val97, and His99 (Figure 2c). This is consistent with the highly conserved nature of the C-terminal residues and the mutational studies in which Trp164 to Ala substitution renders the protease inactive and prevents the release of Npro from the BVDV-encoded polyprotein [20], [24]. The W164A mutation would likely destabilize the β-sheet and, by extension, the catalytic site of Npro such that the enzyme is no longer able to carry out catalysis.
Cys168 is absolutely conserved in pestiviruses, and thus Npro has been thought to be highly specific for Cys at the P1 site. Consistent with this prediction, Cys168 to Glu substitution abrogates the protease activity [30]. In our experiment, however, the recombinant C168A protein that contains the additional four residues (SDDG) from the core protein was proteolytically active, and the four C-terminal residues were cleaved by the protein. In the C168A structure, the side chain of Ala168 is located in a shallow hydrophobic pocket formed by Thr166, Pro64, Val78 and Gly80 (S1 subsite) (Figure-2c). The S1 subsite can only accommodate amino acids with small side chains. A long negatively charged Glu in the C168E protein thus would not fit in the subsite due to steric hindrance. Thus, although Npro does not require Cys at the cleavage site, Cys168 is conserved in pestiviruses, suggesting that the residue may have an additional function other than participating in the protease catalysis.
Following auto-proteolysis at the C-terminus, the catalytic activity of Npro is completely lost [24]. The structures presented here show that the C-terminal β-strand (one half of the product peptide) remains buried in the active site pocket, indicating that once cleaved, the C-terminus of Npro acts as an intramolecular inhibitor and thus prevents trans-activity, i.e, the enzyme is inactive toward additional substrates. This is consistent with limited proteolysis results showing that the C-terminus of Npro is protected from proteolytic degradation [28]. Additionally, the C-terminal β-strand (substrate) is also a part of the central β-sheet. Thus, no other peptide substrate can bind in the substrate binding site without disrupting the fold of the protease. In this way, Npro has evolved to carry out only a single catalytic event.
An analogous autoprotease mechanism, viz., intramolecular product inhibition, has been reported for several proteins. For example, pestivirus NS2 is an autoprotease that cleaves its own C-terminus from the NS2-3 protein. Although the full-length NS2 is limited to a cis-cleavage reaction, deletion of at least four amino acids from the C-terminus was sufficient to allow the protein to cleave a substrate in trans [31], indicating that the C-terminal residues (substrate) do not participate in either the fold or the activity of the protease. In contrast, Npro is unlikely to possess trans-cleavage activity even if the C-terminal residues are deleted from the protein because the C-terminal β-strand is critical for the fold and activity of the protease domain. Deletion of the C-terminal residues would result in an unstable protease rather than an active protease with an unoccupied substrate binding site. We have indeed observed that the deletion of the terminal 5 amino acids results in an insoluble protein, likely due to instability of the protein and the resultant formation of inclusion bodies upon expression in E. coli (unpublished data).
The C-terminal zinc-binding domain (residues 101 to 156) forms an anti-parallel β-sheet consisting of five β-strands (β3, β4, β5, β6, and β7). Though the domain is not directly involved in the proteolytic mechanism, it serves as a structural scaffold for the N-terminal protease domain and shields the C-terminal β-strand. The C-terminal domain likely maintains the structural integrity of the protein until the final β-strand carrying the cleavage site (Cys168) is translated, which then enables the catalytic domain to acquire its active conformation, thus allowing cleavage of the peptide bond at the C-terminus of Npro. We have shown that this domain carries a conserved metal binding TRASH motif with the consensus sequence C-X19–22-C-X3-C (X being any amino acid) [28]. Npro has a modified TRASH motif that consists of Cys112-Cys134-Asp136-Cys138, which coordinates a single zinc atom. Individual mutations of C112A/R, C134A, D136N, and C138A in the TRASH motif resulted in loss of zinc-binding, and also abolished IRF3 binding and subsequent inhibition of IFN-α/β induction when introduced into the virus. In the crystal structure, all four residues of the TRASH motif are located at one end of the β-sheet, consistent with previous biochemical data [28] (Figure 2d). The zinc-binding site consists of a loop that contributes the ligand Cys112 and a β-hairpin that contributes the other three ligands Cys134, Asp136, and Cys138, respectively. However, neither the wild-type nor the C168A structures contain a bound zinc atom at this site. Instead, a disulfide bridge was formed between Cys112 and Cys134. We surmise that the zinc atom escaped the binding site in the absence of a stable reducing agent in the crystallization conditions, which in turn allowed formation of a disulfide bridge. This displaced zinc atom could then be salvaged in the C168A protein, and coordinated by the secondary coordination site formed between two molecules of Npro as a result of crystal packing (see above). The distance between Cys112 and Cys138, the furthest two residues in the proposed zinc-binding site, is 7.7 Å, which is too long to form a zinc coordination site. Since both residues are located in flexible loop regions, they may come closer upon zinc binding without the need for major conformational changes. Alternatively, one of the Cys residues may be required to maintain the geometry of the other residues in the zinc-binding site, and may not be directly involved in zinc binding. A water molecule could then occupy the fourth coordination site.
The TRASH motif was first described as a novel sequence motif for genes involved in copper homeostasis, and was predicted to have a treble clef fold [32]. The treble clef fold consists of a β-hairpin at the N-terminus and an α-helix at the C-terminus that contribute two ligands each for zinc-binding [33]. However, the zinc-binding site in Npro does not resemble the treble clef fold or any other common zinc-finger motifs. It is close to a zinc ribbon in that the zinc-binding site contains a three-stranded anti-parallel β-sheet. Unlike a typical zinc ribbon that consists of two zinc knuckles (short β-strands connected by a turn) that each contribute two ligands, in Npro one ligand comes from the loop connecting β3 and β4, and the other three from the strands β5 and β6 and the loop connecting them (Figure 2d). Since the zinc-binding residues in Npro constitute a modified form of the TRASH motif, i.e., C-X21-C-X-D-X-C, it is not clear whether the zinc-binding motif in Npro forms a subset of the TRASH motif or a new zinc coordinating sequence motif.
An intact zinc-binding site in Npro is required for binding IRF3 and targeting it for proteasomal degradation in the host cell [28]. Similar to pestivirus Npro, rotavirus NSP1 and herpes virus ICP0 also inhibit IRF3 activation by binding to IRF3 and targeting it for proteasomal degradation [8], [34]. Both proteins also contain a conserved zinc-binding RING-finger motif (Cys3HisCys4) at their N-termini, and have been suggested to act as an E3 ubiquitin ligase. The E3 ligase transfers ubiquitin from the E2 conjugating enzyme to the substrate protein via direct interaction with the substrate protein. Although Npro contains a zinc-binding motif, the structure of the zinc-binding site is rather different from the classical zinc-fingers and does not resemble the RING-finger motif, the typical fold of E3 ubiquitin ligase. Thus, it seems unlikely that Npro functions as an E3 ubiquitin ligase and Npro may regulate the IRF3 degradation via a mechanism different from that of rotavirus NSP1 and herpes virus ICP0.
The role of Npro in the regulation of IRF3-dependent IFN-α/β induction is well-established in the pestiviral disease pathogenesis. Interaction of CSFV Npro with IRF3 has been shown in cell-based and in vitro binding assays, and interaction between BVDV Npro and IRF3 has been shown by immunoprecipitation, all of which were used to identify residues that affect its ability to interfere with IFN induction [15], [16], [26], [27], [28]. Mutations of the catalytic Cys69 had a minimal effect on IFN induction for both BVDV and CSFV Npro, indicating that the protease activity is not related to IRF3 binding [15], [26], [27]. However, mutations of His49 to Val or Leu resulted in a loss of IFN-antagonistic activity, suggesting at least partial structural overlap between the protease and anti-IFN functions of Npro [15], [16], [26]. Point mutations of Glu22 to Leu or Val also abolished the anti-IFN activity of Npro [15], [16], [26]. Cys112, Cys134, Asp136 and Cys138 in the zinc-binding domain were also required for the anti-IFN activity, as described previously [28]. N-terminal mutations of BVDV and CSFV Npro have different consequences on the suppression of IFN-α/β production in infected cells [15], [16], [26]. In BVDV Npro Leu8 to Pro substitution impaired its IFN-α/β antagonistic function. Although the mutant displayed binding to IRF3, it could no longer promote its ubiquitination and proteasomal degradation [16]. In contrast, CSFV Npro that contains a deletion of the N-terminal 19 amino acids maintained its ability to inhibit IFN-α/β response [26]. Deletion of 24 or more amino acids at the C-terminus of Npro also abolished the anti-IFN activity of Npro [16], [26], [27].
To determine if the residues involved in the anti-IFN response form a localized IRF3-binding surface, we mapped the above mentioned residues on the 3D structure of Npro, along with the conserved residues (Figure 3). Large deletions of the N-terminal 19–22 amino acids or the C-terminal 24 amino acids were not included in the surface mapping, since their loss of function may be caused by the disruption of protein folding. The residues form two spatial clusters on the opposite sides of the protein surface; one cluster is on a face of the protease domain, and the other on the zinc-binding domain (Figure 3). Whether IRF3 could simultaneously interact with both surfaces is not known. However, since the mutation on the N-terminus of BVDV Npro (L8P) only disrupted ubiquitination of IRF3 and not its binding, the two surface clusters of conserved residues in each domain may account for different functions in anti-IFN activity; one for direct interaction with IRF3 and the other for interaction with cellular proteins in a downstream response leading to ubiquitination and degradation of IRF3 [16]. Based on previous experimental data and the surface distribution of residues involved in IRF3 binding, we speculate that the C-terminal Zn binding domain interacts with IRF3, whereas the protease domain would bind to a cellular protein involved in the ubiquitination reaction.
Cysteine proteases fall into one of two major groups, or clans based on their structural homology and evolutionary relationship. Clan PA proteases are evolutionarily related to the chymotrypsin family and have the common double β-barrel fold with catalytic residues located between the two β-barrels. The catalytic nucleophile can be either serine or cysteine arranged similar to His57-Asp102-Ser195 in the chymotrypsin sequence. The other clan, CA, comprises all papain-like cysteine proteases which consist of an N-terminal α-helical domain and the C-terminal β-barrel domain with the active site located in the cleft between the two domains. The arrangement of the catalytic residues in papain is Cys25-His159-(Asn175); Asn helps to orient the imidazolium ring of the catalytic His [35]. Npro does not share sequence homology with any other known proteases, and thus was assigned to its own family of cysteine proteases, C53 [25]. The newly established catalytic His49-Cys69 dyad does not align in either sequence or structure with either type of cysteine protease, and the unique fold of the protein reported here supports this classification. One of the unique features of the protease is that the C-terminal residues, the substrate of its own protease activity, form a β-strand that contributes to the overall fold of the protein. This β-strand is further blocked by the zinc-binding domain, and the substrate binding site is partially enclosed in a hydrophobic environment. Such an arrangement of substrate peptide would prevent further access of any other endogenous substrates to the active site for cleavage, indicating that Npro evolved to catalyze a single cleavage event. After self-cleavage, the product peptide remains bound in the active site pocket, making the protease permanently inhibited by its own C-terminus. In fact, the β-strand would not be able to be released without distorting the structure, resulting in the loss of protease activity. Additionally, the short α-helix containing the catalytic Cys69 is disordered in the wild-type Npro structure, which may be an additional measure to deactivate the protease function after initial cleavage.
While our manuscript was under review, BVDV Npro structures that lack the first 21 amino acids have been published [36]. The overall fold is similar to the fold presented here, and BVDV Npro (PDB 3zfr) can be superimposed with an rmsd of 0.86 Å for the common 134 Cα atoms with CSFV Npro. The greatest deviations between CSFV and BVDV Npro and among BVDV Npros lie in the short helix containing the catalytic Cys69 and the loop preceding the helix (residues 63–69). This also contributes to two major differences in the structures and in their interpretation. First, the catalytic Cys69 and Cys168 forms a disulfide bond in BVDV Npro. Since this is an inactive state of the enzyme, an active form of the enzyme with reduced Cys69 was proposed to be in equilibrium with the non-productive form. Second, a bound hydroxide ion near Gly67 amide was proposed to deprotonate the catalytic Cys69 sulfhydryl for the nucleophilic attack. His49 would then function as oxyanion hole and polarize the scissile bond, instead of forming an imidazolium (His49)-thiolate (Cys69) ion pair. In comparison, CSFV Npro structures do not have a disulfide bond between Cys69 and Cys168 because the active site-containing helix is disordered in the wild-type Npro and Cys168 is replaced in the C168A Npro. Since C168A is as active as the wild-type Npro in our hands, disulfide bond formation is not a required step in catalysis. In addition, no water molecule was observed near the Gly67 amide. The Gly67 amide, along with Asp68 and Cys69 amides, is within hydrogen-bonding distance to the C-terminal carboxylate (Figure 2). Thus, the CSFV Npro structure supports the classical cysteine protease mechanism rather than the catalytic mechanism proposed for BVDV Npro. Although differences in crystallization conditions and protein preparations (native vs refolding) could have led to the observed differences in the catalytic site, it seems unlikely that BVDV and CSFV Npro utilize a different mechanism to catalyze the cleavage reaction. Since both proteins could have a distorted active site geometry either from the disulfide bond formation between Cys69 and Cys168 in BVDV Npro or zinc coordination in CSFV Npro, the measurement of pKa of the Cys69 and His49 may be required to distinguish between the two mechanisms.
The release of Npro from the polyprotein subsequently sets the stage for suppression of innate immune responses. Pestivirus Npro binds and degrades IRF3 via ubiquitination and the proteasomal degradation pathway, and thus subverts the IFN-α/β induction in host cells [16], [17], [18], [27]. However, following the initial binding to IRF3, the mechanism of IRF3 degradation is still unknown. In addition, although many residues have been indicated to be involved in IRF3 binding and subsequent IFN subversion, it is not known whether the mutations directly affect IRF3 degradation or simply decrease protein stability. For example, Glu22 was proposed to be important for both proteolytic activity and IFN subversion functions of Npro [15], [16], [21], [26]. In light of the crystal structures presented here, it is likely that the mutation would destabilize the protein folding leading to loss of function. Nonetheless, the Npro residues involved in the anti-IFN function cluster into two patches on opposite sides of the protein surface (Figure 3), suggesting there might be distinct functions for each of the patches. Since Npro is unlikely to ubiquitinate IRF3 directly, other cellular proteins probably need to bind to the Npro-IRF3 complex for ubiqutination and degradation to occur. This could also explain the observations that the N-terminal mutations of CSFV and BVDV Npro have different consequences in subversion of interferon induction. Although the same cellular proteins are likely recruited to CSFV and BVDV Npro, specific residues involved in the interaction between Npro and the protein partner may be different.
Several Npro-binding proteins including IRF-7, HAX-1, IκBα, and TRIM56 have been identified [19], [37], [38], [39]. IRF7 is a transcription factor for interferon-α genes and induced by type I interferon. Npro also interferes with the function of IRF7 in pDC and thus dampens interferon-α induction during viral infection. Similar to IRF3 and Npro interactions, the IRF7 interactions with Npro rely on the zinc-binding domain of Npro. In particular, individual mutations of TRASH motif residues (C112R, C112A, C134A, D136N, and C138A) in Npro abolished the IRF7 interaction in mammalian two-hybrid assays [19]. However, interaction between IRF7 and Npro does not induce proteosomal degradation of IRF7, suggesting a different mechanism of Npro-mediated IRF7 antagonism. The significance of Npro interactions with HAX-1, IκBα, and TRIM56 in viral pathogenesis is less clear. HAX-1 and IκBα are involved in controlling cell survival, while TRIM56 is involved in antiviral response. Interestingly, the consensus sequence for HAX-1 binding site was suggested to be present in Npro between residues 110 and 135. The peptide corresponding to Npro residues 106–143 interact with HAX-1 in co-precipitation assays [37]. This Npro peptide contains an intact TRASH motif, and it is likely that HAX-1 binds to the zinc-binding surface as with IRF3 and IRF7. Several additional proteins interacting with Npro were identified in random screens, but the functional relevance of these interactions have not yet been characterized [39], [40]. Identification of cellular proteins that interact with Npro and their interaction studies are essential next steps towards understanding how binding of Npro to IRF3 leads to the degradation of IRF3 and Npro-mediated viral pathogenesis.
The Npro gene of CSFV strain vA187-1 (Alfort/187, GenBank accession number X87939) [41] was amplified by PCR from pA187-1 and cloned into pCR4-TOPO (Invitrogen). The DNA fragment containing the Npro gene was then subcloned into the NdeI and XhoI restriction sites of pET-15b vector to obtain pET-6H-throm-Npro(Alf) [28]. The N-terminal deletion mutants were designed based on limited proteolysis results [28]. The Npro construct lacking the first 17 amino acids (Npro-Δ17N) was amplified from the full-length construct using the oligonucleotide primers 5′ ggcagccatatgggagtggaggaaccggtatac 3′ (forward) and 5′ cggatcctcgagttagcaactggtaacccacaatgg 3′ (reverse). The PCR product was again sub-cloned into the pET15b expression vector between the NdeI and XhoI restriction sites. The C168A mutant with the additional four residues of the core protein (NproΔ17N-C168A-SDDG) was cloned similarly using the reverse primer 5′ gtggtgctcgagttagccatcatcagaggcactggtaac 3′. All constructs were verified using DNA sequencing. The resulting proteins have a hexa-histidine tag (His-tag) and the thrombin cleavage sequence (LVPRGS) on the N-terminus of the protein.
The Npro mutants were expressed and purified as described in ref. 28. His-tagged Npro proteins, purified using Talon metal affinity chromatography resin (Clontech) were pooled and dialyzed in buffer A (20 mM Tris pH 8.0, 100 mM NaCl and 5 mM β-mercaptoethanol) overnight, and the N-terminal His-tag cleaved using thrombin protease immobilized on agarose beads (ThermoScientific). The cleavage reaction was performed with 1% thrombin (w/w) at room temperature in buffer A for 4 hrs. Cleaved Npro-Δ17N was separated by passing the mixture through the Talon resin equilibrated in buffer A. Npro-Δ17N-C168A was purified similarly except that 10% glycerol was added during thrombin cleavage of the His-tag. Glycerol was necessary to stabilize the protein during the cleavage reaction at room temperature. The protein was >95% pure judged by SDS-PAGE. Both proteins were monomers in solution as judged by size-exclusion chromatography.
Proteins were concentrated to ∼4.5 mg/ml. Initial crystallization trials were conducted using the sitting drop vapor diffusion method in 96-well plates using a Phoenix RE liquid handling robot (Rigaku). 200 nL protein solution was mixed with equal volumes of a range of precipitants obtained from commercially available crystal screens. Crystals appeared in several conditions within a week. Diffraction quality crystals of Npro-Δ17N grew in 25% PEG3350, 0.2 M MgCl2 (or 0.2 M (NH4)2SO4) and 0.1 M Hepes pH 7.4. Npro-Δ17N-C168A crystallized in 25% PEG3350 0.2 M (NH4)2SO4 (or 0.2 M Li2SO4) and 0.1 M Hepes pH 7.5.
The Npro-Δ17N crystals were cryo-cooled at 100 K using paratone as cryo-protectant. High redundancy data was collected using Bruker's Microstar microfocus X-ray Source equipped with a Platinum135 CCD detector. The data were indexed and merged using the Bruker AXS PROTEUM2 software suite for X-ray crystallography (Bruker AXS (2010). PROTEUM2, Version 2010.5, Bruker AXS Inc., Madison, Wisconsin, USA). The crystals diffracted to 1.6 Å resolution and belonged to the space group P21212 with a = 60.1, b = 62.6, c = 30.8 Å. The solvent content was 35% with a monomer in the asymmetric unit. The structure of Npro-Δ17N was solved via the single wavelength anomalous dispersion method (SAD) using the anomalous signal present in sulfur atoms illuminated by a copper K-α home X-ray source. Determination of the positions of sulfur atoms, phasing, and calculation of electron density maps were performed using AutoSol wizard in the Phenix package [42], [43]. The initial atomic model was obtained using the Autobuild wizard in Phenix [44]. The final model was achieved using manual model building with the program O [45] followed by iterative cycles of refinement with phenix.refine. All residues from Glu21 to Cys168 were visible in the electron density map except residues 65–71, which encompasses the catalytic Cys69.
Diffraction data for Npro-Δ17N-C168A crystals was collected to 1.6 Å using a Rigaku FRE++ X-ray source and an RAXIS-IV™ detector at UTMB. Data was indexed in the space group P212121 with unit cell dimensions of a = 41.5, b = 58.3, and c = 75.5 Å, different from the Npro-Δ17N protein. The solvent content of the crystals was 54%. The structure of Npro-Δ17N-C168A was determined using molecular replacement with the Npro-Δ17N structure as a model. An atomic model was built using Auto-Build, and phenix.refine was used for refinement of the final model. All residues from Met18 to Cys168 were visible except amino acids 145–149. New crystal packing interactions involving the active site contributed to the stabilization of the helix carrying the nucleophilic Cys69. Strong density (visible at >8σ) at the center of the coordination complex at the active site indicated the presence of a metal ion at the site. SAD data at the zinc absorption edge were collected at the Center for Advanced Microstructures and Devices (CAMD) synchrotron macromolecular crystallography beamline at Louisiana State University. An anomalous difference Fourier map was calculated to confirm the presence of zinc. Ramachandran plots for both structures were generated using the program PROCHECK [46] in CCP4. Data collection and refinement statistics are given in Table 1.
The coordinates and associated structure factors for this publication have been deposited into the Protein Data Bank and assigned the following accession codes: 4H9J and 4H9K for the Npro-Δ17N and Npro-Δ17N-C168A, respectively.
|
10.1371/journal.pntd.0005031 | Plasmodium vivax but Not Plasmodium falciparum Blood-Stage Infection in Humans Is Associated with the Expansion of a CD8+ T Cell Population with Cytotoxic Potential | P. vivax and P. falciparum parasites display different tropism for host cells and induce very different clinical symptoms and pathology, suggesting that the immune responses required for protection may differ between these two species. However, no study has qualitatively compared the immune responses to P. falciparum or P. vivax in humans following primary exposure and infection. Here, we show that the two species differ in terms of the cellular immune responses elicited following primary infection. Specifically, P. vivax induced the expansion of a subset of CD8+ T cells expressing the activation marker CD38, whereas P. falciparum induced the expansion of CD38+ CD4+ T cells. The CD38+ CD8+ T cell population that expanded following P. vivax infection displayed greater cytotoxic potential compared to CD38- CD8+ T cells, and compared to CD38+ CD8+ T cells circulating during P. falciparum infection. We hypothesize that P. vivax infection leads to a stronger CD38+ CD8+ T cell activation because of its preferred tropism for MHC-I-expressing reticulocytes that, unlike mature red blood cells, can present antigen directly to CD8+ T cells. This study provides the first line of evidence to suggest an effector role for CD8+ T cells in P. vivax blood-stage immunity. It is also the first report of species-specific differences in the subset of T cells that are expanded following primary Plasmodium infection, suggesting that malaria vaccine development may require optimization according to the target parasite.
anzctr.org.au ACTRN12612000814875; anzctr.org.au ACTRN12613000565741; anzctr.org.au ACTRN12613001040752; ClinicalTrials.gov NCT02281344; anzctr.org.au ACTRN12612001096842; anzctr.org.au ACTRN12613001008718
| The specific immune responses that contribute to protective immunity in humans following Plasmodium infection are yet to be fully characterized. The species P. vivax and P. falciparum account for most human infections, yet little is known about P. vivax specific immune responses and whether they are similar to or distinct from P. falciparum. Here, we establish that P. vivax and P. falciparum elicit distinct cellular immune responses following primary infection, with the expansion of a subset of CD38+ CD8+ T cells with a cytotoxic potential in P. vivax but not in P. falciparum infection. This study provides the first evidence for the activation of CD8+ T cells in P. vivax blood-stage infection and demonstrates the existence of species-dependent host immune responses to malaria. These findings have important implications for P. vivax vaccine development, and suggest that future malaria vaccine studies should be adapted according to the target Plasmodium spp.
| Malaria vaccine research efforts have been directed predominantly at P. falciparum, since globally it is the major cause of malaria-related mortality [1]. However, it is now recognized that P. vivax is poised to become the dominant species in areas where it is endemic [2] and can be associated with severe pathology [2,3]. Yet, compared to what is known about responses to P. falciparum, little is known about immune responses to P. vivax infection. This lack in knowledge is due in part to confounders that are present in samples from naturally-infected individuals living in malaria-endemic regions where parasitic co-infections and cross-species immunity are present; and technical difficulties associated with experimental infection of humans due to a lack of a method for the continuous in vitro culture of P. vivax [4]. It has been generally assumed that P. vivax would elicit similar immune responses compared to P. falciparum. However, the two parasites display very different features in terms of life cycle, invasion mechanism and immunopathology [2,3,5] and thus may generate distinct host specific immune responses. A few studies have compared global frequencies of circulating lymphocyte populations during P. falciparum or P. vivax infection in naturally infected humans [6,7], but have not investigated their activated or effector phenotype.
The recent establishment of different models of Controlled Human Malaria Infection (CHMI) provides the opportunity to obtain samples from malaria-naive healthy volunteers following first exposure to Plasmodium blood-stage parasites, thereby greatly enhancing our understanding of the host-parasite immune response [8,9]. Until recently, such experimental infection studies could be done only with P. falciparum due to the lack of a continuous in vitro culture system of P. vivax as a source of parasitized red blood cells [8]. Recently, however, a cell bank of cryopreserved P. vivax infected erythrocytes was successfully derived from a naturally-infected individual and used to experimentally infect malaria-naive healthy adult volunteers, establishing for the first time a CHMI model with P. vivax [10].
Here, we have taken advantage of this novel resource to compare cellular immune responses generated following experimental blood-stage infection of naive volunteers with P. vivax or P. falciparum. Overall, we found marked differences in the immune profiles generated following infection with the two species. Specifically, P. vivax but not P. falciparum infection led to the expansion of a specific subset of CD8+ T cells which were associated with an activated phenotype and cytotoxic potential. This study enhances our understanding of P. vivax associated immunity and Plasmodium species-specific immunity, identifying for the first time components of the immune response to blood-stage infection that are species-specific.
Experimental infection of malaria-naive healthy adult volunteers was undertaken at QPharm Pty Ltd (Brisbane, Australia); all clinical studies were registered on the Australian and New Zealand Clinical Trials Registry (ANZCTR): P. falciparum clinical trial ID numbers ACTRN12612000814875, ACTRN12613000565741, ACTRN12613001040752 and NCT02281344; and P. vivax clinical trial numbers ACTRN12612001096842 and ACTRN12613001008718, with written informed consent and approval of the QIMR Berghofer Medical Research Institute Human Research Ethics Committee (QIMRB-HREC) and the Western Institutional Review Board (ethics board for the trial sponsor, Program for Appropriate Technology in Health, PATH).
Inoculum preparation, volunteer recruitment, infection, monitoring and treatment were performed as described previously for P. falciparum [11] or P. vivax [10]. In brief, healthy malaria-naive individuals were intravenously inoculated with freshly thawed P. falciparum 3D7 or P. vivax parasitized erythrocytes and treated with anti-malarial drugs when the parasitemia exceeded the approximate threshold of 10,000 parasites/mL, at day 7–8 post-infection or day 14 post-infection for P. falciparum or P. vivax, respectively. The infecting dose for P. falciparum was 1,800 viable parasitized red blood cells. Parasite growth modeling using in silico analysis estimated that the infecting dose for P. vivax was 15 fold lower compared to P. falciparum (Khoury D & McCarthy JS, in preparation). Blood samples were collected prior to infection, at day 7 post-infection for P. falciparum infected volunteers, and day 14 post-infection for P. vivax infected volunteers. Peripheral blood collected in Lithium Heparin Vacutainers (Becton Dickinson) was either used directly for flow cytometry analysis, or peripheral blood mononuclear cells (PBMC) isolated using standard Ficoll density gradient centrifugation.
Parasitemia was determined using a consensus P. falciparum or P. vivax species-specific quantitative PCR assay as previously described [12]. Parasite levels were assessed once daily until day four post-infection and then twice daily until treatment. All samples were batch tested in triplicate together after each study completion. Limit of detection was 64 parasites/ml [12]. Exponential growth equation fitting parasitemia kinetics for P. falciparum or P. vivax infected volunteers was calculated with GraphPad Prism (version 6.0).
Staining buffer was PBS supplemented with 0.5% FCS and 4 mM EDTA. Whole blood collected in Lithium Heparin vacutainers was lysed and fixed with BD FACS lysing solution (Becton Dickinson) and lymphocytes permeabilized with BD FACS permeabilising solution 2 (Becton Dickinson) according to the manufacturer’s instructions. Cells were then resuspended in 50 μl of staining buffer containing anti-human CD4-BV510 (Becton Dickinson, 1:200 dilution), anti-human CD8-APC-H7 (Becton Dickinson, 1:400 dilution), anti-human CD19-PE-Cy7 (Biolegend, 1:200 dilution), anti-human CD38-APC (Biolegend, 1:400 dilution), anti-human Perforin-PE (Biolegend, 1:400 dilution), anti-human Granzyme B-Pacific Blue (Biolegend, 1:400 dilution) and 1 μl of human Fc receptor blocking solution (Human TruStain FcX, Biolegend) for 30 minutes at room temperature, washed and resuspended in staining buffer before acquisition on LSR Fortessa 4 (Becton Dickinson) with Diva software. FlowJo software version 6.0 was used for gating.
CD38+ CD8+ T cells, CD38- CD8+ T cells as well as CD8- cells were sorted from freshly isolated PBMC. Approximately 10x106 cells were resuspended in 50 μl of staining buffer containing anti-human CD4-BV510 (Biolegend, 1:200 dilution), anti-human CD8-APC-H7 (Becton Dickinson, 1:400 dilution) and anti-human CD38-PerCpCy5.5 (Biolegend, 1:400 dilution) for 20 minutes at 4°C, washed and resuspended in staining buffer. Just before the sorting, 1 μg/mL of propidium iodide (Sigma-Aldrich) was added to allow for assessment of viability. Pi-CD8+CD4-CD38+, Pi-CD8+CD4-CD38-, and Pi-CD8- cells were sorted using a BD Aria III cell sorter (Becton Dickinson) directly in staining buffer and kept on ice until further use for in vitro assays.
Sorted CD38+ and CD38- CD8+ T cells were plated at 50,000 cells/well in RPMI 1640 containing 25 mM Hepes, 2 mM L-glutamine (Invitrogen), and supplemented with 10 units/mL of Penicillin (Life Technologies), 10 μg/mL of Streptomycin (Life Technologies) and 10% fetal bovine serum (Life Technologies) in a 96-well plate pre-coated overnight with 10 μg/mL of anti-human CD3 OKT3 antibody (Biolegend) together with 0.75x106 cells/mL autologous CD8- cells, anti-human CD107a-FITC (Biolegend, 1:200 dilution) and 1 μg/mL of co-stimulatory antibodies anti-human CD28 and anti-human CD49d (Becton Dickinson) for 5 hours at 37°C in an atmosphere of 5% C02. Following stimulation, cells were resuspended in 20 μl of staining buffer containing anti-human CD4-BV510 (Biolegend, 1:200 dilution), anti-human CD8-APC-H7 (Becton Dickinson, 1:400 dilution) for 20 minutes at 4°C, washed and resuspended in staining buffer before acquisition on LSR Fortessa 4 (Becton Dickinson) with Diva software. FlowJo version 6.0 was used for gating.
P. falciparum and P. vivax experimental infection of malaria-naive volunteers was performed under similar procedures [10,11]. However, due to logistical reasons associated with parasite density in the inoculum stock, and a lack of a continuous in vitro culture system for P. vivax [4], the infecting dose for P. vivax was estimated to be 15-fold lower than that used in the P. falciparum studies (Khoury D & McCarthy JS, in preparation). Demographics of P. falciparum and P. vivax infected volunteers were comparable in term of age, gender, BMI and ethnicity (S1 Table).
Parasitemia growth curves determined by quantitative PCR (qPCR) from P. falciparum and P. vivax experimental infection studies were similar for both parasites (Fig 1 and Table 1) except for the delayed onset of detectable blood-stage parasitemia with P. vivax. Specifically, P. falciparum parasites were detected as early as day 4 of infection whereas P. vivax parasites were detected at day 8 of infection, consistent with the differences in the size of the starting inocula. Interestingly, all individuals infected with P. vivax developed symptoms of malaria before the time of treatment while more than 40% of P. falciparum infected individuals were asymptomatic until anti-malarial drug administration (S2 Table).
In order to compare cellular immune responses to infection between P. falciparum and P. vivax infected volunteers, we determined the phenotype of lymphocytes ex vivo from whole blood samples obtained prior to infection and during infection once the parasitemia exceeded 10,000 parasites/mL (which corresponded to day 7 and day 14 for P. falciparum and P. vivax infected volunteers, respectively). CD38 is a surface glycoprotein that modulates cell adhesion, signal transduction and intracellular Ca2+ levels, and is specifically upregulated on lymphocytes following activation [13]. We have recently shown that the frequency of CD38+ T cells and B cells circulating in the peripheral blood of test volunteers was dynamically regulated during experimental blood-stage infection with P. falciparum and that the expansion of CD38+ CD4+ T cells following infection was inversely correlated with parasite burden [14]. Thus, we compared the frequencies of CD38+ T cells and B cells circulating before and after infection in P. falciparum or P. vivax infected volunteers. There was a higher frequency of CD38+ CD4+ T cells circulating following infection in P. falciparum but not P. vivax infected volunteers (Fig 2A). Conversely, P. vivax infection but not P. falciparum elicited a higher frequency of circulating CD38+ CD8+ T cells (Fig 2B). No significant differences were observed in the frequency of CD38+ B cells circulating following infection with P. falciparum or P. vivax (Fig 2C). There was no correlation between the expansion of CD38+ CD8+ T cells following infection and parasite burden (S1 Fig). Overall, these data suggest that the quality of the immune response generated following primary blood-stage infection in humans is Plasmodium species-dependent.
Since very little information on P. vivax protective immune responses is available, we aimed to further understand the contribution of CD38+ CD8+ T cells to P. vivax blood-stage immunity. Effector CD8+ T cells perform classical cytotoxic functions by killing infected cells through perforin-mediated dependent mechanisms. To determine the cytotoxic potential of CD38+ CD8+ T cells generated during P. vivax infection, we measured their intracellular expression of granzyme B and perforin by flow cytometry, in an independent cohort (n = 2 because of logistics associated with vivax experimental infection). We found similar expression of granzyme B and perforin in CD38+ and CD38- CD8+ T cells before infection (Fig 3A and 3B). However, post-infection, CD38+ CD8+ T cells had a greater expression of granzyme B and perforin compared to CD38- CD8+ T cells (Fig 3A and 3B). Additionally, CD38+ CD8+ T cells circulating during infection contained a higher amount of granzyme B and perforin in P. vivax infected volunteers compared to P. falciparum infected volunteers (Fig 3C and 3D) whereas no differences were observed prior to infection.
In order to further investigate the cytotoxic potential of CD38+ CD8+ T cells and CD38- CD8+ T cells circulating during P. vivax infection, we tested their ability to degranulate in vitro following TCR stimulation. Prior to infection, CD38+ and CD38- CD8+ T cells had the same ability to degranulate, whereas post-infection CD38+ CD8+ T cells had a higher degranulation compared to CD38- CD8+ T cells (Fig 4).
Overall, these findings suggest that CD38+ CD8+ T cells have a greater cytotoxic potential compared to CD38- CD8+ T cells and that this cytotoxic function is specifically activated in the CD38+ CD8+ T cells circulating during P. vivax infection but not P. falciparum infection.
Herein we report on the first study to compare the quality of cellular immune responses elicited in humans by experimental blood-stage infection with P. vivax or P. falciparum. In this study we utilized a model of controlled infection of malaria-naive human volunteers, thereby avoiding potential confounders of pre-existing immunity and cross-species immunity. We found marked differences between responses to the two Plasmodium species in terms of the phenotype of T cells that expanded during infection. P. falciparum infection elicited a significant expansion of CD38+ CD4+ T cells whereas P. vivax infection led to the expansion of CD38+ CD8+ T cells. We have recently shown that the frequency of circulating CD38+ CD4+ T cells was significantly increased following experimental infection with P. falciparum and inversely correlated with parasite levels [14]. Here we show that P. vivax blood-stage infection elicited a substantially different type of immune response compared to P. falciparum, with significant changes in the CD8+ T cell compartment rather than in the CD4+ T cell compartment. There was no significant association between the expansion of CD38+ CD8+ T cells and parasite burden in P. vivax infected volunteers, suggesting that the expansion of CD38+ CD4+ T cells in P. falciparum infection and the expansion of CD38+ CD8+ T cells in P. vivax infection might have distinct contributions to the immune response to blood-stage infection.
A possible explanation for the qualitative differences observed between P. falciparum and P. vivax associated immune responses may relate to their distinct tropism during blood-stage infection. P. vivax merozoites preferentially invade reticulocytes [15] which, although they are enucleated, still express MHC-I molecules which remain from the nucleated reticuloblast stage. This is in contrast to mature erythrocytes that have completely lost expression of MHC molecules. Previous work using a genetically engineered mouse model of murine malaria has shown that CD8+ T cells can be activated by parasitized erythroblasts but not parasitized mature red blood cells through MHC-I dependent mechanisms [16]. Thus, we hypothesize that the higher proportion of infected reticulocytes in P. vivax infection leads to the activation of a higher proportion of CD38+ CD8+ T cells in an MHC-I dependent manner in comparison to P. falciparum infection.
While it is established that CD4+ T cells and parasite-specific antibodies are critical for protective immune responses to blood-stage malaria [17], the contribution of CD8+ T cells is less clear. Studies in mice have shown that CD8+ T cells were activated and associated with protective function in lethal [18] or chronic [19] blood-stage malaria. However, no association between CD8+ T cells and protective immunity to primary blood-stage infection in humans has been reported yet. Our data suggest that the CD38+ CD8+ T cells that specifically expand during P. vivax infection display increased cytotoxic function compared to other CD8+ T cells. Hence their function might be to kill parasitized reticulocytes through MHC-I dependent and perforin-mediated mechanisms. This proposal is further supported by the enhanced expression of cytotoxic molecules observed in CD38+ CD8+ T cells circulating during P. vivax compared to P. falciparum infection. In contrast, we could not identify an expansion of cytotoxic CD8+ T cell population following P. falciparum infection of malaria-naïve volunteers.
These findings have important implications in regard to P. vivax vaccine development. Indeed, most efforts so far have been directed toward the direct translation of the findings associated with P. falciparum vaccine development to P. vivax [20] (e.g. the use of ortholog antigens that were shown to be protective against P. falciparum). Here we show that P. falciparum and P. vivax elicit qualitatively different immune responses and are likely to require distinct vaccine-induced immune responses for protection. Thus, the immunization strategy may need to be adapted for each Plasmodium species to mount optimal protective immune responses, and the design of a universal vaccine conferring protection against multiple Plasmodium species might be conceptually more challenging than initially thought.
In conclusion, in this study we report that primary exposure of humans to different Plasmodium species elicited qualitatively distinct immune responses: P. falciparum infection generated changes in CD4+ T cells whereas P. vivax preferentially activated a subset of CD8+ T cells expressing the activation marker CD38 and associated with a cytotoxic function. The specific expansion of CD38+ CD8+ T cells following P. vivax infection might be due to the preference for P. vivax merozoites to infect reticulocytes which can activate CD8+ T cells through MHC-I dependent mechanisms. Overall our data are consistent with the proposal that protective immune responses to Plasmodium are species-dependent. These findings have important implications for malaria vaccine development strategies.
|
10.1371/journal.pgen.1004836 | Rrp12 and the Exportin Crm1 Participate in Late Assembly Events in the Nucleolus during 40S Ribosomal Subunit Biogenesis | During the biogenesis of small ribosomal subunits in eukaryotes, the pre-40S particles formed in the nucleolus are rapidly transported to the cytoplasm. The mechanisms underlying the nuclear export of these particles and its coordination with other biogenesis steps are mostly unknown. Here we show that yeast Rrp12 is required for the exit of pre-40S particles to the cytoplasm and for proper maturation dynamics of upstream 90S pre-ribosomes. Due to this, in vivo elimination of Rrp12 leads to an accumulation of nucleoplasmic 90S to pre-40S transitional particles, abnormal 35S pre-rRNA processing, delayed elimination of processing byproducts, and no export of intermediate pre-40S complexes. The exportin Crm1 is also required for the same pre-ribosome maturation events that involve Rrp12. Thus, in addition to their implication in nuclear export, Rrp12 and Crm1 participate in earlier biosynthetic steps that take place in the nucleolus. Our results indicate that, in the 40S subunit synthesis pathway, the completion of early pre-40S particle assembly, the initiation of byproduct degradation and the priming for nuclear export occur in an integrated manner in late 90S pre-ribosomes.
| During the synthesis of small ribosomal subunits in eukaryotes, the pre-40S particles formed in the nucleolus are rapidly transported to the cytoplasm. The mechanisms involved in the nuclear export of these particles and its coordination with other steps of the 40S synthesis pathway are mostly unknown. In this work we studied the function of Rrp12, the only major non-ribosomal factor of nuclear pre-40S particles that does not remain stably associated to them during maturation in the cytoplasm. We demonstrate that Rrp12 is required for the exit of pre-40S particles to the cytoplasm. Remarkably, we also found that Rrp12, together with the Crm1 exportin, participates in processes that occur in early pre-ribosomes in the nucleolus, including the processing of the pre-rRNA and the elimination of processing byproducts. Thus, Rrp12 and Crm1 participate in maturation steps that take place upstream of nuclear export. Our results indicate that, in the 40S subunit synthesis pathway, the completion of early pre-40S particle assembly, the initiation of byproduct degradation and the priming for nuclear export occur in an integrated manner in nucleolar pre-ribosomes.
| The formation of ribosomes in eukaryotic cells requires the production and subsequent assembly of four rRNAs and ≈80 ribosomal proteins into small (40S) and large (60S) ribosome subunits. In the yeast S. cerevisiae, three out of those four rRNAs (18S, 5.8S and 25S) are transcribed together in the nucleolus in the context of a common polycistronic 35S pre-rRNA (see scheme in Fig. 1A) [1], [2]. This primary rRNA precursor is bound by ribosomal proteins, as well as by the U3 small nucleolar ribonucleoprotein (snoRNP) and ≈70 non-ribosomal factors, to form the large 90S pre-ribosomal particle [3]. This process involves the recruitment of smaller multiprotein subunits that associate to the nascent transcript in a stepwise manner [4], [5], [6]. The 35S pre-RNA then undergoes serial cleavages at the A0, A1 and A2 sites to generate the 20S and 27SA2 pre-rRNAs (Fig. 1A) [1], [2]. These three cleavages can also occur co-transcriptionally within the so-called small subunit (SSU) processome, a complex very similar in composition to the 90S pre-ribosome [7], [8]. The disassembly of the 90S pre-ribosome leads to the formation of pre-40S and pre-60S particles containing the 20S and the 27SA2 pre-rRNAs, respectively [2], [9]. This process is accompanied by the release of the non-ribosomal components originally present in the 90S pre-ribosome and the rapid degradation of processing byproducts [2].
The early pre-60S particles contain >40 associated factors, and undergo multiple maturation steps that are accompanied by major changes in composition until exiting the nucleus [2], [10], [11], [12]. In contrast, the early pre-40S particles are thought to have a relatively low compositional complexity and are rapidly exported, consistent with the fact that the 20S pre-rRNA is not further processed inside the nucleus (Figure 1A) [2], [10], [11], [13]. Final maturation of 40S subunits, which includes a proofreading step through association to 60S subunits and the cleavage of the 20S pre-rRNA at site D, takes place in the cytoplasm [14], [15], [16], [17]. Due to the rapid kinetics of transit thorough the nucleoplasm, it is assumed that the major events of pre-40S particle assembly take place concurrently with the 35S pre-RNA cleavage steps in the nucleolus. Despite this, the pre-40S particles released from 90S pre-ribosomes have to undergo some transformations before leaving the nucleus. These include the recruitment of factors that will participate in cytoplasmic maturation processes as well as transport proteins involved in particle transit through nuclear pores [18], [19]. Pre-40S particles are also known to undergo a kinase-dependent conformational rearrangement that might be required for nuclear export [20].
Despite the significant progress in the understanding of the compositional changes that take place between 90S and pre-40S pre-ribosomes, there are still many questions about the nucleolar assembly and nuclear maturation of 40S subunits that remain unanswered. For example, it is still unclear how the early pre-40S particles are assembled within the 90S pre-ribosome and how similar they are, at the structural level, to the pre-40S particles that reach the cytoplasm. It is also unknown how and when pre-40S particles become competent for export, and how the export process itself takes place. The Ran GTPase and the Crm1 exportin are both essential for pre-40S particles to exit the nucleus [19], [21], but the factors or mechanisms that mediate their interaction with those particles are not known. Tackling these questions has been difficult so far due to the large number of components involved, the transient nature of nucleoplasmic transit and nuclear exit, and the lack of success in dissecting these activities in separable or mechanistically simple steps.
Rrp12 is a karyopherin-like protein previously described as essential for the export of pre-40S and pre-60S particles out of the nucleus [22]. In addition, Rrp12 has been found to facilitate ribosome-unrelated nuclear import processes [23]. In relation with the role of Rrp12 in pre-ribosome export, it is presently unclear whether such function is due to an implication in the assembly of pre-ribosomal complexes, their maturation in the nucleus, the actual transport event, or compound roles in some of the above processes. To address those issues, we studied in detail the consequences of a partial or total loss of function of Rrp12 in S. cerevisiae. Our results show that Rrp12 is required for nuclear export of pre-40S particles. However, and unlike previously-published observations, we found that Rrp12 is not essential for 60S subunit production. During the course of our experiments, we additionally uncovered that this protein, together with the Crm1 exportin, is important for the last processing events of the 35S pre-rRNA within the 90S pre-ribosome, and for the rapid elimination of the 5′-A0 byproduct. The characterization of these new roles indicates that the completion of assembly and the nuclear export of the pre-40S particle are intertwined processes.
A previous report described that Rrp12 was required for export of both pre-40S and 60S ribosomal subunits from the nucleus to the cytoplasm [22]. However, we observed using a yeast strain with the RRP12 gene under a galactose-inducible promoter (GAL::HA-rrp12) that this protein was specifically involved in the biosynthesis of 40S subunits. Evidence in favor of such conclusion included: (i) Polysome profile analyses showing that the loss of Rrp12 was associated with reductions in the content of free 40S subunits and polysomes, but not of free 60S subunits (Figure 1B and Figure 1C). In fact, the relative abundance of the large subunits was clearly increased in the absence of Rrp12 (Figure 1B and Figure 1C). (ii) Northern blot analyses demonstrating a decrease in the steady-state amount of the 18S rRNA (present in 40S subunits) but not in those of the 5.8S and 25S rRNAs (present in 60S subunits) in Rrp12-depleted cells (Figure 1D, left panels; and Figure 1E). Such alterations were found to be associated with an increase in the abundance of the 20S pre-rRNA, the immediate upstream precursor for the 18S rRNA (Figure 1D; see scheme in Figure 1A), indicating that the cleavage at site D is inhibited. Consistent with previously published data [22], we also observed some accumulation of the 35S and 32S pre-rRNAs, a reduction in the content of the 27SA2 pre-rRNA, and the generation of the aberrant 21S pre-rRNA (a species produced from direct cleavage of the 32S pre-rRNA at site A3) (Figure 1D; see scheme in Figure 1A). These results indicate that, in addition to the major defect in the cleavage at site D, the loss of Rrp12 causes partial defects in the early cleavages at sites A0 and A1 and, to a larger extent, at site A2. We also detected a delay in the processing events of 5.8S rRNA precursors manifested by the presence of both the 7S pre-rRNA and aberrant 3′-extended forms of the 5.8S rRNA (5.8S+30) in Rrp12-depleted cells (Figure 1D and Figure 1E). Curiously, we found that the absence of Rrp12 led to an increase in the abundance of the 5′-A0 fragment (Figure 1D), a byproduct produced when the rRNA precursor is cleaved at site A0 (Figure 1A). Similar defects, although milder in intensity, were observed in a constitutive manner when pre-rRNA analyses were performed in a yeast strain (rrp12-Δ198) expressing a hypomorphic version of Rrp12 (Figure 1D, right panels; Figure 1F). Taken together, these data indicate that Rrp12 is absolutely required for the generation of the 18S rRNA from 20S pre-rRNA and, in addition, important for both the rapid elimination of the 5′-A0 fragment and the normal processing of both 32S and 5.8S pre-rRNA precursors. Despite this latter function, Rrp12 does not seem to have any major influence on the overall production of 60S ribosomal subunits.
Our group and others have previously shown that Rrp12 copurifies with components of 90S and pre-40S particles [3], [18], [22], [23]. However, there is no detailed information about its relative content in different subsets of pre-40S complexes. Using coimmunoprecipitation experiments, we observed that endogenous Rrp12 interacted with green fluorescent protein (GFP)-tagged versions of factors present in nucleolar 90S (Pwp2, Enp1, Dim1, Pno1; Figure 2A and Figure 2B, lanes 1 to 4 and lanes 19 to 22) and nucleoplasmic pre-40S (Enp1, Dim1, Pno1, Tsr1; Figure 2A and Figure 2B, lanes 3 to 6 and lanes 19 to 22) particles. These interactions took place within the context of ribonucleproteic complexes, because they were eliminated by RNase treatment (Figure 2B, lanes 1 to 6 and lanes 19 to 22). By contrast, we did not detect any association of Rrp12 in these experiments with either Nob1 or Rio2, two proteins mostly present in cytoplasmic pre-40S particles (Figure 2A and Figure 2B, lanes 7,8 and lanes 23,24). Rrp12 did show an interaction with Ltv1, a protein that, like Nob1 and Rio2, is mainly detected in cytoplasmic pre-40S complexes (Figure 2A and Figure 2B, lanes 25,26). This interaction is the only one that cannot be disrupted by RNase treatment (Figure 2B, lanes 25,26), indicating that it survives pre-40S particle disassembly or, alternatively, that takes place outside those particles. In agreement with the results presented in Figure 1, we could not detect interactions of Rrp12 with proteins present in early (Ssf1, Nop7; Figure 2A and Figure 2B, lanes 9 to 12), intermediate nuclear (Rix1; Figure 2A and Figure 2B, lanes 13,14), late nuclear (Arx1; Figure 2A and Figure 2B, lanes 15,16) or cytoplasmic (Kre35; Figure 2A and Figure 2B, lanes 17,18) pre-60S complexes. These results suggest that Rrp12 is predominantly associated to both nucleolar and nuclear pre-40S pre-ribosomes while it is weakly associated, or not bound at all, to the cytoplasmic ones. Further analyses of Rrp12-MYC immunoprecipitates by Northern blot confirmed the predominant presence of this protein in the 40S synthesis pathway and, in addition, evidenced that its interactions with nucleolar and nucleoplasmic particles exhibit differential features. Thus, we observed that the association of Rrp12 to pre-40S particles had to be rather strong, as inferred by the stable coimmunoprecipiation of the 20S pre-rRNA with Rrp12-MYC (Figure 2C). Indeed, the amount of this pre-RNA in those complexes is even higher than that seen in the case of immunoprecipitations performed with Tsr1, a factor that stably associates with both nucleolar- and cytoplasmic-located pre-40S particles (Figure 2A and Figure 2C). By contrast, we could not detect any significant amount of 35S and 32S pre-RNAs in the Rrp12-MYC immunoprecipitates, suggesting that the association with the 90S particle is either labile or restricted to a minor pool of Rrp12-containing complexes (Figure 2C). As control, we found that these two pre-RNAs do coimmunoprecipitate with Pwp2 (Figure 2C), an integral component of the 90S pre-ribosome (Figure 2A). Consistent with the lack of Rrp12 in the purifications of pre-60S complexes (see above Figure 2B), we could not observe any interaction of Rrp12-MYC with the 27S or 7S pre-rRNAs. As expected, these two pre-rRNAs do coimmunoprecipitate with the early pre-60S particle component Nop7-MYC (Figure 2D, see scheme in Figure 2A). These results indicate that Rrp12 does not stably associate with pre-60S particles.
We next focused on the cause of the block in the maturation of 20S pre-rRNA to 18S rRNA found in Rrp12-depleted cells. Given the restricted presence of Rrp12 to nucleolar 90S and nucleoplasmic pre-40S complexes, this phenotype could be due to defects in the assembly of the pre-40S particle inside the nucleus. However, this does not seem to be the case because the depletion of Rrp12 does not affect the stability of both early and late nuclear pre-40S components (Enp1, Dim1, Tsr1, Rio2, Nob1; Figures 3A to 3C; top panels). Likewise, it does not block the interaction of those proteins with the 20S pre-rRNA (Figures 3A to 3C; bottom panels). However, the depletion of Rrp12, although not affecting the steady state levels of Ltv1 in cell lysates prepared by TCA precipitation (Figure 3D, compare lanes 7 and 9), does cause a destabilization of that protein under the conditions used for the pre-rRNA coimmunoprecipitation analyses (Figure 3C; top panel, compare lanes 4 and 6). Such behavior may reflect a functional relationship of Rrp12 and Ltv1 in vivo, because we observed using sucrose gradient fractionation experiments that the loss of Rrp12 leads to a substantial decrease in the amount of Ltv1 that is stably incorporated onto ∼40S complexes (Figure S1A). This effect is specific, because the depletion of Rrp12 does not affect the incorporation of both Enp1 and Rio2 onto those complexes (Figure S1B and Figure S1C).
Mass spectrometry experiments further confirmed that the absence of Rrp12 does not have a major effect in the composition of pre-40S complexes. Indeed, we found that both the pattern and strength of the associations exhibited by four pre-40S factors (Enp1, Tsr1, Nob1 and Rio2) with the rest of major pre-40S particle components are quite similar to those observed in wild-type cells (Figure 3E, compare columns 1 to 4 with columns 5 to 8). The only exception observed is the loss of the interaction of both Enp1 and Tsr1 with Ltv1 (Figure 3E, compare columns 1 and 2 with columns 5 and 6), a defect probably derived from the impaired recruitment of Ltv1 to the pre-40S particle seen in above experiments. Interestingly, we observed that the loss of Rrp12 promotes the formation of new interactions of both Enp1 and Tsr1 with the tRNA methyltransferase Ncl1 and the abundant hnRNP protein Npl3 (Figure 3E, compare columns 1 and 2 with columns 5 and 6). Likewise, Tsr1 and Nob1 interact with the 90S particle component Nop1 (Figure 3E, compare columns 2 and 3 with columns 6 and 7). These results indicate that Rrp12 is not required for the formation of the core structure of the pre-40S particle, although it may contribute to the release of specific nucleolar factors such as Nop1. In addition, they show that Rrp12 appears to be dispensable for the recruitment of some factors with hitherto unknown roles in the synthesis of 40S subunits (i.e., Ncl1, Npl3). Also consistent with a correct particle assembly in the absence of Rrp12, we found using Western blot analyses that Prp43 and Mex67 [24], [25], [26], [27], two factors that are not usually detected in this type of proteomics analysis due to their weak interaction with pre-40S particles, remain particle-associated in the absence of Rrp12 (Figure S2). Interestingly, the absence of Rrp12 does promote a reduction in the association of Enp1 with some, but not all, of its usual partners within the 90S particle (Figure 3E, compare columns 1 and 5). These data indicate that the lack of Rrp12 may affect either the composition or maturation dynamics of 90S pre-ribosomes.
The above findings indicated that the lack of Rrp12 blocks the 40S synthesis pathway at a step downstream the assembly of pre-40S particles. To investigate if this block occurred in the nucleolus, nucleoplasm or cytoplasm, we analyzed the subcellular localization of GFP-tagged versions of pre-40S particle (Enp1, Dim1, Pno1, Tsr1, Ltv1, Nob1, Rio2) and mature 40S subunit (Rps2) components in control and Rrp12-depleted cells. Consistent with previous reports [18], [28], [29], [30], [31], we found that these proteins exhibit nucleolar (Enp1, Figure 4A), nucleolar and nucleoplasmic (Dim1, Tsr1; Figure 4B and Figure 4C, respectively), and nucleoplasmic plus cytoplasmic (Pno1, Ltv1, Nob1, Rio2 and Rps2; Figures 4D to G, and Figure S3, respectively) localizations in both wild type cells and control GAL::HA-rrp12 cells. However, in Rrp12-depleted GAL::HA-rrp12 cells, we detected that most of those proteins undergo a major relocalization towards the nucleoplasm (Figures 4A to G; and Figure S3). The only exception was again Ltv1, since its subcellular distribution is fully Rrp12-independent (Figure 4E). The nuclear accumulation of Rio2, but not of the cytosolic Pgk1 control protein, in the absence of Rrp12 was demonstrated using independent subcellular fractionation experiments (Figure 4H). This effect is specific for the 40S subunit synthesis pathway, because the loss of Rrp12 does not alter the normal subcellular distribution of Rpl25 and Rpl11 (Figure S3), two 60S subunit components. These results show that pre-40S particles are blocked in the nucleoplasm when Rrp12 is absent. Collectively, our data indicate that Rrp12 does not participate in the major assembly events of pre-40S particles in the nucleus, and that it is essential for some event that immediately precedes or is concomitant to nuclear export.
In addition to the block in pre-40S particle export, the depletion of Rrp12 causes defects in the cleavage of the pre-rRNA at site A2 and in the elimination of the 5′-A0 fragment. The accumulation of this byproduct appears to be a rather specific feature, because it is not observed upon depletion of other factors, like Pno1, that do not affect the A0 cleavage but are essential for the A2–A3 cleavages (Figure S4A). We also found that the 5′-A0 fragment associates to Rrp12 in wild type cells (Figure S5), suggesting that Rrp12 might influence directly the elimination of this fragment. As a first approximation to obtain clues about the role of Rrp12 in this process, we decided to study the sedimentation behavior of the 5′-A0 fragment on sucrose gradients in the presence and absence of Rrp12. These experiments corroborated the increase in the abundance of the 5′-A0 fragment already seen by Northern blot analyses in Rrp12-depleted cells (see above, Figure 1D) and, in addition, revealed that this fragment was present in complexes that sediment broadly between the 60S and 90S regions of the gradient (Figure 5A; right panels, gradient fractions 12 to 15). A significant proportion of these entities cosedimented with the 32S pre-rRNA and U3 snoRNA (Figure 5A; right panels, gradient fractions 14,15), suggesting that they form part of a 90S transitional particle that has initiated, but not completed, the processing of the 35S pre-rRNA. This interpretation is consistent with the delay in the A2 cleavage evidenced by the formation of aberrant 21S pre-rRNA (see above, Figure 1D), and the increased coimmunoprecipitation of the 5′-A0 fragment with the 90S pre-ribosome-specific Pwp2 protein in Rrp12-depleted cells (Figure 5B, compare lanes 10 and 12). The interaction of Pwp2 with the 5′-A0 fragment appears to take place in the context of a 90S pre-ribosome-like particle, as inferred from the presence of Pwp2 in 80–90S complexes in Rrp12-depleted cells (Figure 5C). In agreement with an abnormal accumulation of a 90S transitional particle, we observed by microscopy experiments that Pwp2 shifted from an exclusively nucleolar localization to a more disperse distribution between the nucleolus and the nucleoplasm upon depletion of Rrp12 (Figure 5D). These results indicate that the loss of Rrp12 delays some event during pre-40S particle assembly in the nucleolus, leading to both the accumulation and delocalization of 90S transitional particles in the nucleoplasm.
We next characterized by mass spectrometry the complexes formed by Pwp2 in the absence of Rrp12 to investigate possible differences in the composition of 90S pre-ribosomes. Although highly similar to those formed in control cells, we observed the presence of new Pwp2 partners in the absence of Rrp12 (Figure 6A). Those included 90S pre-ribosome components involved in the cleavage of the 35S precursor at the A0–A1–A2 (Utp20, Rcl1) and A1–A2 (Dim1, Pno1) sites [29], [32], [33], [34], [35], [36]. Interestingly, we observed using RNA coimmunoprecipitation experiments that two of the above partners, Dim1 and Pno1, preferentially bind the 32S rather than the earliest 35S pre-rRNA (Figure 6B). This suggests that they become stably assembled onto the 90S pre-ribosome upon cleavage of the 35S precursor at the A0 and A1 sites (see above, Figure 1A). We also found among the new partners the nuclease Rrp44 (also known as Dis3), an exosome subunit shown to be involved in the direct physical interaction with the 5′-A0 fragment [37]. This finding was quite interesting for us, because previous results have shown that this interaction seems to be crucial for poising the 5′-A0 fragment for productive degradation [37], [38]. Thus, we surmised that the Rrp44-Pwp2 interaction detected in Rrp12-depleted cells could indicate that the exosome is normally recruited to 90S pre-ribosomes and that, in the absence of Rrp12, there is an enrichment or stabilization of some of those exosome-containing 90S pre-ribosomes. In agreement with this idea, we found using sucrose gradient sedimentation analyses that Rrp44 is indeed present in 80–90S complexes both in control and Rrp12-depleted cells (Fig. S1D). These data raised the possibility that the defect in the elimination of 5′-A0 fragment found in Rrp12-depleted cells could be due to an impairment of exosome function. Consistent with this idea, we found that the elimination of the exosome cofactor Mtr4 (also known as Dob1) elicited the expected accumulation of the 5′-A0 fragment (Figure 6C and Figure 6D) [39] and, most importantly, that such accumulation occurs in the context of 80–90S complexes, similarly to what is observed in Rrp12-depleted cells (Figure 6D; see above, Figure 5). Interestingly, Rrp12-depleted cells do not exhibit the sustained high levels of the 5′-A0 fragment seen in Mtr4-depleted cells (Figure 1D and Figure 6C), indicating that the exosome activity is affected but not fully compromised upon the loss of Rrp12. Consistent with this, we have seen that the loss of this protein does not trigger other terminal defects typically observed in exosome-deficient cells, such as the abnormal accumulation of the 35S pre-rRNA, the total block of 7S pre-rRNA maturation, and the balanced decrease in the contents of both ribosomal subunits (Figure 6C and Figure 6D) [39], [40]. Taken together, our data indicate that the loss of Rrp12 causes a 90S pre-ribosome maturation defect that precedes the A2 cleavage and the exosome-dependent 5′-A0 fragment degradation steps. As a result, it promotes either a delay or partial inhibition, but not a block, in the A2 cleavage of the pre-rRNA and the elimination of the 5′-A0 fragment.
Given the implication of Rrp12 in the export of pre-40S particles (see above, Figure 3 and Figure 4), we decided to investigate whether the pre-40S export step was associated to the Rrp12-dependent 90S pre-ribosome maturation step. If that were the case, we expected that the elimination of any other protein involved in pre-40S export would induce the same defects seen in Rrp12-depleted cells. To test this idea, we chose a yeast strain that constitutively expressed a mutant version of the Crm1 (Crm1T539C) exportin. This mutant protein, unlike its wild type counterpart, can be specifically inhibited by leptomycin B [41]. Using this strategy, we found that the inhibition of Crm1 recapitulates all the defects observed in Rrp12-depleted cells, including increased abundance of the 35S, 32S and 21S pre-RNA species (Figure 7A), abnormal levels of the 5′-A0 fragment (Figure 7A and Figure 7B), accumulation of this fragment in 80–90S complexes (Figure 7B) and, as expected [19], an increase in the content of the 20S pre-rRNA due to the halt in pre-40S particle nuclear export (Figure 7A). These results indicate that the 40S subunit export machinery facilitates a late 90S pre-ribosome maturation event that promotes the rapid cleavage of the pre-rRNA at site A2 and the efficient degradation of the 5′-A0 fragment. This function is quite specific for export regulators, because the elimination of factors specifically involved in the cytoplasmic maturation of pre-40S complexes (Rio2 and Ltv1) does not trigger any of the above defects [16], [28], [42] (Figure S4B).
The above results led us to investigate whether Crm1, like Rrp12, was present in 90S pre-ribosomes. We first assessed the potential interaction of Crm1 with two 90S pre-ribosome components, the 35S pre-RNA and Pwp2, using coimmunoprecipitation analyses similar to those that detect Rrp12 in 90S and pre-40S particles (see above Figure 2). This approach however did not reveal associations of Crm1 with any pre-ribosomal component, not even with pre-rRNAs or proteins present in the pre-40S and pre-60S complexes transported by this exportin. We therefore decided to change the experimental conditions of our coimmunoprecipitation assays. In particular, we changed the Triton X-100-containing lysis buffer by a NP-40-containing buffer that was similar to buffers used by others to detect interactors of Crm1 in vivo [43], [44], [45]. Notably, when we purified Crm1-GFP using the NP-40 buffer, we could readily observe that it interacts with the 35S pre-rRNA, the 20S pre-rRNA, 27S pre-rRNAs and the 25S rRNA (Figure 7C). The associations with these RNAs were specific because in the same Northern blots Pwp2-GFP coprecipitated the 35S and 23S pre-rRNAs, but not the 20S, 27S and 25S RNAs. These results indicate that Crm1 binds to pre-40S and pre-60S particles, as expected from its role in export, and also that it is already recruited to early 90S particles. Consistent with this, we found using sucrose gradient sedimentation analysis that Crm1 is indeed present in large 80–90S complexes that co-sediment with Pwp2 (Figure 7D). Furthermore, when 90S pre-ribosomes were purified from sucrose gradients using Pwp2 as bait it was confirmed that they do contain Crm1 (Figure 7D, right set of panels). Western blot analysis of Rrp12-containing complexes from total cell lysates evidenced that Crm1 interacts with Rrp12 (Figure 7E), a result consistent with the common presence of the two proteins in both 90S and pre-40S pre-ribosomes.
In our final set of experiments, we investigated whether the recruitment of Crm1 to 90S pre-ribosomes was Rrp12-dependent. For this purpose we analyzed the sedimentation behavior in sucrose gradients of a HA-tagged version of Crm1 that was coexpressed with the endogenous Crm1 either in wild type or in rrp12Δ198 cells. We found that in wild type cells the Crm1-HA protein is recruited to large assemblies, including 80–90S-like complexes (Figure 7F, left two panels). This sedimentation in large complexes is drastically reduced in rrp12Δ198 cells (Figure 7F, right two panels), suggesting that the incorporation of Crm1 onto large 80–90S pre-ribosomal particles is Rrp12-dependent. Altogether, our data indicate that Rrp12 and Crm1 act on 90S pre-ribosomes in a concerted manner.
The results presented here identify Rrp12 as a factor required for a number of intertwined steps of the 40S ribosomal subunit synthesis pathway (Figure 8). We have observed that Rrp12, together with Crm1, is first recruited to the pathway to facilitate the processing of the 35S pre-rRNA and the elimination of the 5′-A0 fragment in the context of a late 90S transitional particle (Figure 8). A lack of Rrp12 or Crm1 at this step delays but does not halt the assembly and release of early pre-40S particles. Interestingly, this early function of Rrp12 occurs immediately upstream and temporally close to the export of the pre-40S particles, a process that absolutely requires Rrp12 and Crm1. In addition to revealing a hitherto unknown role for export-related factors in a specific maturation step in the nucleolus, our results shed light onto the dynamics of 90S pre-ribosome factors upon cleavage of the 35S pre-rRNA at site A2. Indeed, some authors previously suggested that, after the A2 cleavage, the non-ribosomal components of the 90S particle are released en bloc in association with the 5′-A0 fragment [3], [18]. However, the formation of such disassembly complexes, and when and how was the exosome recruited, remained unclear. We find no evidence for the formation of a post-disassembly complex containing the 5′-A0 fragment upon which the exosome acts (Figure 6D). Rather, our results indicate that the exosome is present in transitional 90S pre-ribosomes to degrade the 5′-A0 fragment, either in the last step of pre-40S particle assembly or at the very time of pre-40S particle release (Figure 8). The implication of Crm1 in steps of ribosome synthesis, other than nuclear export, is also a new finding in yeast. In human cells, Crm1 has been implicated in the targeting of snoRNP complexes to the nucleolus [43], [46]. Whether Rrp12 and Crm1 utilize the same domains for the export-related and maturation-related functions, and whether the two proteins need to interact directly to exert their functions, remains to be determined. We have found that Rrp12 and Crm1 purified from bacteria do not stably interact in vitro (unpublished data). However, we cannot exclude the possibility that such interaction could require the participation of other proteins. Indeed, it has been shown before that the interaction of Crm1 with other molecules involves the participation of additional factors, including the Ran GTPase in its GTP-bound state. Ran can in fact be involved in these interactions, as suggested by the identification of allele-specific Ran mutants that elicit defects in the degradation of the 5′-A0 fragment [47]. Based on the present results, we hypothesize that such defects could be associated to the Rrp12- and Crm1-dependent mechanism reported here. An involvement of Ran on the association of Crm1 with pre-ribosomes could also explain the difficulties for detecting Crm1 in purified 90S and pre-40S pre-ribosomes, because these complexes are normally prepared under conditions that favor the conversion of Ran-GTP to Ran-GDP. Here we describe that, using a buffer that contains 0.2% NP-40, it is possible to detect the specific association of Crm1 to both pre-rRNAs and pre-ribosomal components by coimmunoprecipitation analysis. The reason for the efficiency of this buffer is unclear, but it must somehow favor the maintenance of some Ran-GTP levels and/or affect other currently unknown features that improve the stability or solubilization of Crm1-containing complexes.
In our model we propose that Rrp12 is an export factor rather than a nuclear maturation factor (Figure 8). Consistent with this, we have observed that the elimination of Rrp12 leads to the accumulation of pre-40S particles that, in addition to being dissociated from the 90S pre-ribosome machinery, are fully-assembled. This is evidenced by the recruitment to those particles of factors that are predominantly cytoplasmic in normal cells (Rio2, Nob1), and that therefore must join the pathway just before nuclear exit. One important inference of our results is that the major assembly events involved in the formation of pre-40S particles are separable and fully independent from the subsequent export step. A direct participation of Rrp12 in the export process is also supported by the previously-described interactions of this protein with some nucleoporins and with Ran [22]. Unexpectedly, we could not find any significant role for Rrp12 in the export of pre-60S subunits, as it had been previously published [22]. In addition to the phenotypic analysis of Rrp12-depleted cells, the prominent role of Rrp12 in the 40S rather than the 60S subunit pathway is supported by the RNA-protein interaction data showing the specific binding of Rrp12 to the 20S but not the 27S and 7S pre-rRNAs. The reason for these different results is not readily apparent. We have found that the loss of Rrp12 elicits the 40S subunit-specific phenotype both in the W303 and in BY4743 strains, indicating no influence of the genetic background. Still, it is worth noting that the depletion of Rrp12 causes delays in the processing of 5.8S rRNA precursors in the nucleus by a hitherto unknown mechanism. According to our results, such delays do not impact the overall production of 60S subunits, but it could be possible that, under some experimental conditions or in strains with genetic modifications that subtly affect ribosome biogenesis, the defect in 5.8S rRNA production became exacerbated and caused nuclear accumulation of pre-60S particles. It is also plausible that Rrp12 could interact either weakly or very transiently with some pre-60S particle subpools, as it would be expected if its influence on the processing of 5.8S precursors were direct. This possibility would be in agreement with the previously-reported detection of Rrp12 bound to 27SB pre-rRNAs using primer-extension analyses [22]. Despite the possibility of these alternative scenarios, we believe that our data clearly indicate that Rrp12 is not essential for 60S subunit synthesis. Consistent with this idea, it is also worth noting that mammalian Rrp12 has been shown to be required exclusively for 40S subunit synthesis [48], [49].
One distinctive feature of the intermediate particle formed in the absence of Rrp12 is the lack of Ltv1, a factor not essential for nuclear export. Previous studies indicate that this protein is recruited in the nucleus [31], [50], but some evidence suggests that its interaction with the nuclear pre-ribosomes that are about to be exported might be weak [20]. Thus, a possible explanation for the absence of Ltv1 in the pre-40S particles of Rrp12-depleted cells is that those particles are ready to be exported and have Ltv1 loosely associated. Alternatively, it is possible that Rrp12 could be actively required for the docking of Ltv1 to those particles during the export process. We currently favor the latter possibility, since we have observed that the interaction of these two proteins can occur in a pre-rRNA-independent manner. Based on the present data, we believe that Rrp12 probably promotes the recruitment of Ltv1 onto the pre-40S particle immediately prior to the step of transport (Figure 8). Upon this docking step, Rrp12 is carried along with the particle through the nuclear pores to be finally released when the particles reach the cytosol. Consistent with this hypothesis, our co-purification experiments and other proteomic analyses have shown that Rrp12 is not a major component of cytoplasmic pre-40S particles. Alternatively, it is also possible that Rrp12 could remain associated to cytoplasmic pre-40S particles and only becomes released upon completion of a specific maturation event that takes place right after the nuclear export step. This model would explain previous results indicating that Rrp12 can associate with di-methylated 20S pre-rRNA, a modified form of the 20S pre-rRNA that is generated in the cytoplasm [22]. Further work will be required to dissect the fate and specific roles of Rrp12 in these late maturation stages.
The reason for using the pre-40S export machinery to facilitate late 90S pre-ribosome-mediated processes is unknown. We propose that such mechanism could ensure a timely coordination of the recycling kinetics of 90S pre-ribosome components with pre-40S particle release and rapid export (Figure 8). An inter-relation between these three processes is indicated by our data, which shows that the impairment of nuclear export causes defects in the function, disassembly and subcellular localization of the 90S pre-ribosome. Future work will be needed to explain the precise mechanisms by which the export factors influence the activities of the exosome and A2 cleavage complexes within the 90S pre-ribosome.
The Saccharomyces cerevisiae strains and plasmids used in this study are listed in Tables S1 and S2, respectively. The conditional strain for RRP12 under the control of the GAL1 promoter (YPM7) was generated by one-step insertion of a KAN-MX6-GAL1 cassette upstream of the ATG of the RRP12 gene [51]. This strain (referred to in the text as GAL::HA-rrp12), and the other GAL1-driven strains used in this study, JDY144, WDG72, YGM168, YO470 and YGM174 (referred to in the text as GAL::HA-spb4, GAL::rsa4, GAL::HA-pno1, GAL::rio2-ProtA and GAL::HA-mtr4, respectively) were cultured at 30°C in media containing galactose (YPGal, 0.4% yeast extract, 0.8% peptone, 0.1 mM adenine, 2% galactose) or glucose (YPD, 0.4% yeast extract, 0.8% peptone, 0.1 mM adenine, 2% glucose). For protein depletion, the incubation times in YPD varied from 9 to 18 h, as indicated in figure labelings. The ltv1Δ strain was cultured at 25°C, the temperature at which the 40S subunit biogenesis defects of this strain are more exacerbated. For the experiments of inactivation of Crm1 we employed a strain with the CRM1 gene depleted that carried a plasmid for the expression of the crm1-T539C-HA allele (strain MNY8, plasmid pDC-crm1-T539C). As a control for those experiments, we employed the corresponding strain carrying a plasmid for the expression of crm1-HA (strain MNY7, plasmid pDC-CRM1). MNY7 and MNY8 cells were treated with 100 ng/ml of leptomycin B (LMB) for 5–15 min. All strains with MYC, hemagglutinin (HA) or green fluorescent protein (GFP) carboxy-terminal tagged alleles, except the crm1-HA and GFP-rrp12 ones, were generated by in-frame one-step integration of PCR cassettes in the corresponding locus of wild type cells. In these strains, the epitope-tagged versions are the only source of the proteins in the cell, and their expression is driven from the endogenous gene promoters. All epitope-tagged alleles were fully functional, as measured by normal growth rates and normal contents of rRNAs, pre-rRNAs and ribosomal subunits. The sedimentation analysis of Crm1-HA shown in Figure 7D was performed on the YGM193 strain (referred to in the figure as pwp2-GFP, crm1Δ, pRS315-crm1-HA). The coimmunoprecipitation experiment in Figure 7E was performed with the YMD6 strain carrying the pDC-CRM1 plasmid (referred to in the figure as pwp2-GFP, crm1-HA) and with the YPM7 strain carrying the pGM58 and pDC-CRM1 plasmids (referred to in the figure as GFP-rrp12, crm1-HA). The sedimentation analysis of Crm1-HA shown in Figure 7F was performed on the following strains maintained in glucose-containing media: YPM7 carrying the pBN18 and pDC-CRM1 plasmids (referred in the figure as RRP12, CRM1, pRS315-crm1-HA), and YPM7 carrying the pBN19 and pDC-CRM1 plasmids (referred in the figure as rrp12Δ198, CRM1, pRS315-crm1-HA). Preparation of media, yeast transformation and genetic manipulations were performed according to established procedures.
RNAs from total cellular lysates, gradient fractions and coimmunoprecipitations were prepared by the hot-phenol method [52]. Oligonucleotide labeling, RNA separation, Northern blotting and hybridization were performed as described previously [53]. The sequences of the oligonucleotides used as probes are shown in Table S3.
Preparation of total celular lysates for immunoblot, Western blot analysis, purification of GFP-tagged proteins and mass spectrometry analysis were performed as described previously [23], except for the Pwp2-GFP/Crm1-HA and GFP-Rrp12/Crm1-HA coimmunoprecipitation analysis in Figure 7E. In this case, instead of lysing cells in IP buffer (20 mM Tris-HCl, pH 7.5, 5 mM MgCl2, 150 mM potassium acetate, 1 mM dithithreitol, 0.2% Triton X-100, supplemented with Complete [Roche]), cells were lysed in IP-NP40 buffer (15 mM Na2HPO4, 10 mM NaH2PO4, pH 7.2, 150 mM NaCl, 2 mM EDTA, 50 mM NaF, 0.1 mM NaVO4, 0.5% NP-40 Alternative [Calbiochem], supplemented with Complete). Before purification of Pwp2-GFP and Rrp12-GFP with GFP-TRAP (Chromotek), the pre-cleared lysates were diluted to 0.2% NP-40. The anti-Rrp12 antibody used for Western blot in Figure 2B is a rabbit polyclonal antibody raised against a peptide mapping at the C-terminus of yeast Rrp12 (this study). Other antibodies used in Western blot analysis were: anti-MYC (Roche), anti-GFP (Clontech), anti-HA (Covance), anti-Nop1 (Pierce), anti-Mex67 (kind gift of C. Dargemont, Institut Jacques Monod), anti-Rps3 (kind gift of M. Seedorf, University of Heidelberg), anti-Rps8 (kind gift of G. Dieci, University of Parma), anti-Rpl1 (kind gift of F. Lacroute, Centre de Génétique Moléculaire, Gif-sur-Yvette), anti-Pgk1 (Abcam), and anti-Cdc11 (Santa Cruz). For the represention of the results of the proteomic analysis shown in Figure 3E, the four different dot sizes are indicative of the amount of the copurifying protein relative to the amount of bait: >80%, 60–80%, 40–60%, and <40%.
Cell cultures (200 ml) were grown to an optical density at 600 (OD600) between 0.8 and 0.1 and, before harvesting, cycloheximide was added to a final concentration of 0.1 mg/ml. After an incubation on ice for 5 min, cells were collected and lysed in 700 µl of HK buffer (20 mM HEPES, pH 7.5, 10 mM KCl, 2.5 mM MgCl2, 1 mM EGTA, 1 mM dithiothreitol (DTT) and 0.1 mg/ml cycloheximide) using a Fastprep apparatus. Cell lysates were pre-cleared by high-speed centrifugation, and extracts equivalent to 5–20 absorption units at 260 nm (A260) were loaded on 7–50% sucrose gradients (10 ml), which had been prepared in HK buffer without cycloheximide. Ultracentrifugation, subsequent fraction collection and polysome profile recording were performed as previously described [53]. For Western blot analysis, 40 µl samples of each fraction were mixed directly with 10 µl of SDS-PAGE loading buffer (SPLB) and loaded onto SDS polyacrylamide gels. For Northern blot analysis, total RNA was prepared by the hot-phenol procedure from 100 µl samples of each fraction and separated on 1.2% agarose-formadehyde gels. For the analysis of purified complexes shown in Figure 7D, two sets (pools 1 and 2) of four combined fractions were concentrated 6-fold by spinning on Microcon-10 (Millipore) filters. The recovery of proteins after the concentration step was ∼10 fold more efficient for pool 1 than for pool 2, probably due to the higher sucrose concentration in pool 2. Before performing the GFP-Trap purification, each concentrated pool was taken to 1 ml with NP-40 buffer (0,2% final concentration).
Cell cultures were grown to OD600 between 0.8 and 1.0, and polysome extracts were prepared as described above. Extract equivalents to 15 A260 units were taken to 250 µl with HK buffer and mixed with 0.5 ml of IP buffer containing Complete and 600 U/ml of RNasin (Promega). In the Crm1-RNA coimmunoprecipitations shown in Figure 7C, instead of using IP buffer it was used IP-NP40 (0.2% final concentration) buffer. For evaluation of protein content in total cell lysates, a 30 µl aliquot of the pre-cleared lysate was mixed with 30 µl of SPLB and kept frozen until analysis by Western blot. The rest of the extract was incubated with 2 µg of anti-MYC 9E10 (Roche) antibody or with 25 µl of GFP-TRAP beads at 4°C for 2 h. When using anti-MYC antibody, immunoprecipitates were immobilized with GammaBind sepharose beads (GE Healthcare). Immunoprecipitates were washed four times at 4°C with IP or IP-NP40 buffer. For protein analyses, one fifth of the immunoprecipitated material was resuspended in SPLB and analyzed, in paralel with the samples of total protein, by Western blot. For RNA analyses, the rest of the immunoprecipitated material was resuspended in 400 µl of 50 mM sodium acetate, 10 mM EDTA (pH 5.2), and processed for RNA extraction by the hot phenol method. After ethanol precipitation, the whole amount of recovered RNA was resuspended in formaldehyde loading buffer, separated on 1.2% agarose-formadehyde gels and analyzed by Northern blot. In parallel, in the same Northern blot experiments, it was evaluated the pre-rRNA content in cell lysates before immunoprecipitation, using 5 µg of total RNA prepared by the hot phenol method directly from extract equivalents to 10 A260 units of the corresponding polysome preparations.
Cells were visualized using a Zeiss Axioplan 2 microscope equiped with a 63× objective, a Hammamutsu ORCA-ER digital camera and Openlab (Improvision) cell imaging analysis software. The Rpl25-EGFP and Rps2-GFP reporter assays to monitor pre-40 and pre-60S nuclear accumulation were performed as previously described [54].
Cells were grown to OD600 between 0.8 and 0.1, harvested and spheroplasts prepared by incubation in S buffer (50 mM Tris-HCl, pH 7.5, 10 mM MgCl2, 1.2 M sorbitol, 1 mM dithiothreitol, 5 mg/ml Zymolyase T-100 (Seikagaku) at 30°C for 15 min. After two washes with the same buffer, the spheroplasts were lysed using a manual homogenizer in Ficoll buffer (10 mM Tris-HCl, pH 7.5, 20 mM KCl, 5 mM MgCl2, 3 mM dithiothreitol, 1 mM EDTA, 1 mM PMSF, 180 mg/ml Ficoll-400, supplemented with Complete). Pre-cleared lysates were ultracentrifuged in a TLA 100.3 rotor at 23.000 rpm for 15 min, and the supernatant cytosolic fraction collected. The nuclei pellet was resuspended in 50 mM Tris-HCl, pH 7.5, 100 mM NaCl, 30 mM MgCl2, 0.25% NP-40 supplemented with Complete. Aliquots of the precleared whole lysate (W), cytosolic fraction (C) and nuclei (N) were mixed with SPLB and loaded onto a SDS polyacrilamide gel for Western blot analysis.
|
10.1371/journal.pcbi.1003236 | Whisker Movements Reveal Spatial Attention: A Unified Computational Model of Active Sensing Control in the Rat | Spatial attention is most often investigated in the visual modality through measurement of eye movements, with primates, including humans, a widely-studied model. Its study in laboratory rodents, such as mice and rats, requires different techniques, owing to the lack of a visual fovea and the particular ethological relevance of orienting movements of the snout and the whiskers in these animals. In recent years, several reliable relationships have been observed between environmental and behavioural variables and movements of the whiskers, but the function of these responses, as well as how they integrate, remains unclear. Here, we propose a unifying abstract model of whisker movement control that has as its key variable the region of space that is the animal's current focus of attention, and demonstrate, using computer-simulated behavioral experiments, that the model is consistent with a broad range of experimental observations. A core hypothesis is that the rat explicitly decodes the location in space of whisker contacts and that this representation is used to regulate whisker drive signals. This proposition stands in contrast to earlier proposals that the modulation of whisker movement during exploration is mediated primarily by reflex loops. We go on to argue that the superior colliculus is a candidate neural substrate for the siting of a head-centred map guiding whisker movement, in analogy to current models of visual attention. The proposed model has the potential to offer a more complete understanding of whisker control as well as to highlight the potential of the rodent and its whiskers as a tool for the study of mammalian attention.
| The management of attention is central to animal behaviour and a central theme of study in both neuroscience and psychology. Attention is usually studied in the visual system (most often using cats or primates) owing to the ease of generating controlled visual stimuli and of measuring its expression through eye movement. In this study, we develop a model of the expression of attention in another sensory modality, that served by the tactile whiskers of small mammals (such as rats and mice). This sensory system has long been a popular model in neuroscience and is well characterised. It has become recognised in recent years that the modulations of whisker movements prevalent in the behaving animal represent “active sensing” (in the sense of moving the sensors to optimise sensing performance), yet a unified understanding of these modulations is still lacking. Our model proposes just such a unified understanding, suggesting that whisker movement modulations can be understood as an overt expression of the animal's changing focus of attention. This proposal, therefore, offers to provide both an enhanced understanding of the whisker sensory system and an insight into the management of attention in these animals.
| A succinct summary of contemporary models of primate visual spatial attention is that exogenous signals (those arising from external stimuli) from multiple sensory modalities and endogenous signals (those arising from internal processes) compete and combine to produce a spatial map of salience from which a single region of immediate spatial attention is chosen [1]–[3]. In the case of overt attention, this location is ‘foveated’ by the rapid re-positioning of the eyes with movements of the head and body following as necessary [4]. If multiple salient locations are present, they are visited sequentially. The degree and nature of integration between overt and covert attention (that expressed only internally), exogenous and endogenous influences, and inputs from different modalities are all matters of debate, as is a definition of attention itself [3]–[8]. One aspect, however, is uncontroversial: that overt attention is expressed by rapid orienting movements that centre the foveal region of the eye on the attentional target. Many small mammals, including laboratory rats and mice, possess in addition to vision a complementary and well-characterised sensory system driven by tactile stimulation of prominent arrays of sensitive whiskers, particularly those located around the snout [9]. Here, we will consider whether the movements of these whiskers might also represent an expression of overt attention, revealing areas of proximal space that are of high salience to the animal. Potentially, such a model would be useful to experimentalists interested in mammalian attentional processes and their neural substrates, not least owing to the growing ease with which observations of whisker movement and position can now be made and analysed in these animals even when they are freely behaving.
Whisker movements have been most studied in animals that express ‘whisking’, a periodic protraction and retraction of the whiskers, typically occurring at several cycles per second (each cycle being termed a ‘whisk’) and in bouts lasting several seconds, with a close coupling of the oscillatory motions of the left and right whisker fields. Most data have been gathered using rats [10]–[12], though analyses are also available for mice, shrews, opossums and hamsters [13]–[16]. Many studies have now described significant departures from spectrally pure, bilaterally symmetric and synchronous whisking, revealing that both spatial and temporal parameters of whisker movements are under active control and can change rapidly in response to environmental conditions as well as to the motivations of the animal [12], [17]–[24]. Furthermore, small changes in whisker position can lead to large changes in sensory signals [25]–[27]. Thus, the proposition that an understanding of whisker movement is a pre-requisite to an understanding of whisker sensory signals has become a key focus of research [28]–[33]. This shift has been facilitated by the increasing availability of experimental tools for measurement of whisker movements [18]–[20], [34] as well as for automated analysis of large high-speed video datasets [34]–[37]. Not only is whisker movement of interest to the researcher who wishes to understand whisker sensory processing (and sensory systems in general), but these movements may also provide data about the internal state of the animal [32], [38]. Since whisker motion can be modulated when the head is stationary some useful measures are available also in the head-restrained condition [17], [39].
The modulation of whisker motion parameters under different conditions has been previously explained as arising from reflex responses (e.g. [19], [20], [40]) or from task-specific sensing strategies (e.g. [12], [41]). Furthermore, computational models developed by the current authors and evaluated in biomimetic whiskered robots [42]–[45] have demonstrated that a mix of positive and negative feedbacks, such as could plausibly be mediated by brainstem loops [46], can produce some of the observed whisker modulations. However, a simple reflex model cannot explain all modulations—for instance, those driven by conditioning [47], [48] or anticipation [17], [18], [20], [23], suggesting the involvement of higher centres in motion modulation [49]. Below, therefore, we motivate and develop a new model of whisker movement control that has as its key variable the region of spatial attention. The explicit representation of this region, as a tactile ‘salience map’, represents a significant departure from current theories and our own earlier models of whisker control, and provides a theoretical bridge to the current paradigm for understanding visual attention in primates, in which salience maps are a core concept [6]. We go on to reprise three behavioural experiments in simulation using the new model and report comparable results to those obtained using animals [19], [20], [23] using analyses closely replicating those employed in the original studies. In our discussion, we summarize the key features of the model and of our results, compare it with competing models and discuss its limitations, suggest experiments that might invalidate it, and discuss its likely neural substrate. In addition, we highlight two architectural features common to any model of this form. Thus, this report both represents a step forward in our understanding of active sensing in rodents and highlights the potential of the rodent and its whiskers as a tool for the study of mammalian attention.
The upper panel of Figure 1 (and Video S1) shows the behaviour of a rat as it approaches, detects, and orients toward an object. This top-down view displays the most prominent degree of freedom of each whisker: rotation around the follicle (at the base of the shaft) resulting in ‘sweeping’ of the whisker rostro-caudally with the largest component of movement being in the horizontal plane [22], [50], [51]. Typical unperturbed periodic whisking can be seen in the first half of the trace of average bilateral whisker protraction angles shown in the lower panel of the figure. The current study focuses, however, on the modulations of periodic whisking that occur in response to environmental and internal conditions as illustrated, for instance, in the second half of the trace where whisking becomes strongly bilaterally asymmetric in response to contact with the object. Whisker positioning is, of course, dependent on head position, therefore our model also addresses the issue of moving the head and body in order to reposition the whiskers on larger spatial and temporal scales [52]. The model will not directly address variability in the periodic component of whisker motion, which can also be modulated (e.g. [21]), or the extension to three dimensions, although both of these topics are considered in the discussion. To explain the development of our model we next summarise some of the key observations of rat whisking behaviour that motivated its development together with some of the earlier functional explanations these observations gave rise to. We then operationalize the attentional hypothesis underlying the new model, and provide a detailed description first in conceptual form, then in terms of its implementation as a computer simulation, also explaining how the model will be evaluated in comparison to biological data.
We have previously shown that whisker motion in the horizontal plane can be well summarized by just two variables for each side of the snout [23]: mean (across whiskers) angular position (henceforth, ‘mean protraction angle’) and the angular position difference between caudal and rostral whiskers (henceforth, ‘angular spread’). Several distinct observations of correlations between these and other behavioural variables have been reported. An early result in rat, that whisker protraction angles increase as the animal approaches the location of an anticipated contact [11], [17], [18], [20], has been recently matched and quantified in mouse [24]. Two further observations first made in rat have also been extended to mouse and opossum [15]. The first, which we term Head-Turning Asymmetry (HTA), is that mean protraction angles are adjusted to be more caudal/rostral on the side of the animal into/away from a future turn of the head [15], [19]. The second, Contact-Induced Asymmetry (CIA), is the observation that mean protraction angles are adjusted to be more caudal/rostral on the side of the animal near/away from a nearby object [15], [20] (see also Figure 1). A further observation is the Rapid Cessation of Protraction (RCP) that interrupts the protraction phase of a whisk movement when whiskers on one side of the animal make contact with an obstruction [20], [23]. We use the term Spread Reduction (SR) for the observation that the angular spread on each side of the snout is reduced during contact with objects in the vertical plane versus non-contacting whisks [23]. Finally, recent work in our lab has shown that animals engaged in rapid ( m/s) goal-directed locomotion employ tonic protraction (increased mean protraction angles and a reduced amplitude of periodic whisker movement, [53]).
To account for the observation of HTA, Towal and colleagues proposed that the whiskers search in the space into which the head will shortly be moved, perhaps partly to avoid collisions [19]. To account for contact-driven observations (RCP, CIA, SR) we proposed the general control strategy of ‘Minimal Impingement, Maximal Contact’ (MIMC, [20], [23], [42]) whereby whiskers are controlled so as to maximize the number of contacts but avoid excessive whisker bending within each contact (minimizing impingement). In addition, we recently hypothesized that tonic protraction during rapid forward locomotion reflects a strategy for collision avoidance whereby the ‘look-ahead’ distance of the animal is maximized [53]. Here, we propose that a single mechanism may be sufficient to explain all of these observations, including responses to anticipated contact.
One clue to the nature of this mechanism is the observation that unilateral contact often elicits head-turning towards the contact point suggesting that CIA (Contact-Induced Asymmetry) and HTA (Head-Turning Assymetry), at least, may be related. The simplest possibility is that they are examples of the same response, to head movement or whisker-contact, expressed under different circumstances, but this is excluded by the following two cases. First, CIA is expressed regularly even where head-turning is precluded or absent, such as when the animal is following a wall ([20]; Video S1, S2, S3 all show examples of CIA in the absence of head-turning). Second, and conversely, HTA is expressed in the absence of any contact [19]. Nonetheless, these observations may be related through a hidden variable. In the case of HTA, whisker asymmetry precedes head-turning; therefore, unless whisker asymmetry drives head-turning directly—which seems unlikely—a hidden variable is implied.
Seeking this hidden variable, we ask: Why does unilateral contact often elicit head-turning? The intuitive answer is that contact will often elicit attention, and attention will typically elicit orienting. We hypothesize, accordingly, that the hidden variable relating these observations is the ‘attended region’—that region of the external world which is currently the subject of the animal's attention—which can be affected by both tactile signals and other influences. According to this hypothesis, then, the mechanism underlying CIA is that laterally-biased contact generates laterally-biased attention which, in turn, drives asymmetric whisking, whilst that underlying HTA is that laterally-biased attention (however generated) drives asymmetric whisking and also head-turning. This model, summarised in Figure 2, is also consistent with observations of increased whisker protraction when contact ahead of the animal is anticipated and during goal-directed locomotion, both of which are conditions in which we might expect the attention of the animal to be focussed to the fore. Furthermore, the model explains why CIA is not observed in response to contacts in cases where the animal does not subsequently indicate attentiveness by orienting towards the contacted object [20].
Thus, our central hypothesis is that a transformation from the attended region to whisker protraction angles is the primary driver of long-term modulations of whisker movement (that is, on timescales longer than that of a single whisk cycle). The second behavioural response seen in Figure 1, the orienting of the snout tip, also intuitively appears to be an expression of overt attention since this movement serves to reposition a generalised sensory ‘fovea’—a body region in which are located the microvibrissae, lips, teeth, tongue, and nose, [9], [54], [55]—as well as an important actuator for small mammals: the jaws. We have, therefore, previously argued that movement of the head driven by switches in spatial attention represents a very significant component of the exploratory behaviour of small mammals ([42], [44], [56], [57]; see also [55]). Therefore, in analogy with the literature on the behaviour of visual animals, we refer to discrete head movements delineated by attention switches as ‘foveations’. The current model ties together these two modes of expression of attention, using a single representation of the attended region—in the form of a ‘salience map’—to drive movements of both the whiskers and the head (and, consequently, of the body). The remainder of this section details our implementation of this model, starting with an overview, and continuing with sub-sections detailing each computation, the headings of which correspond to the labels on the boxes in Figure 3.
Below, we use computer simulation of our attentional model to reprise three earlier behavioural experiments. In each case, we position obstacles in the simulated environment, allow the model to control the whiskers and head for some period, and make the following measurements. First, we measure the location of the tip of the snout over time, , and the head bearing (that is, the angle of the head midline that runs from the neck joint to ). Second, we record the measured protraction angle of the th whisker, , according to the methodology we have used previously in the behavioural laboratory [23]. That is, we locate the base of the whisker, and a point two thirds of the way along its shaft, and derive the angle between the vector connecting these points and the head midline. Similar strategies were used in most of the other behavioural work with which we make comparison [15], [19]. We go on to obtain the instantaneous mean protraction angle of all the whiskers on each side of the snout, and , by simple arithmetic mean across the whiskers, again following precedent from analyses of behavioural data [15], [19], [23]. As a measure of whisker protraction angle that is unaffected by bending of the whiskers against obstacles, we also record the protraction angle at the base of the th whisker, , and compute the corresponding bilateral mean protraction angles, and . Presented examples of animal behaviour (stills and videos) were drawn from our archive of behavioural data to illustrate the text; recording methodology was described previously [20], [23]. Bilateral mean protraction angle presented in Figure 1 was recovered from the video data using the BIOTACT Whisker Tracking Tool (bwtt.sourceforge.net) and the ViSA tracking algorithm suite [37].
Above, we described an implementation of a new model of snout and whisker motor control as well as additional simulated components to permit observations of the model. In summary, this implementation (Figure 3) shares the basic form of models from the visual system literature (see [1] for a review)—that is, it includes a spatial map, bottom-up drive from the sensory periphery, non-specific top-down drive, inhibition-of-return (IOR), and outputs that drive overt attention. Experimental control over the model is exercised by choosing the location of any obstacles and the initial position of the head in a given trial. We have included only very simple models of motivation and IOR sufficient to generate patterns of exploratory behaviour, both around and away from obstacles, that can be compared to those seen in animals. In particular, in the absence of obstacles, foveation is driven only by a random signal, and the head model expresses stochastic exploratory-like behaviour (for instance, see Video S5). When obstacles are present, foveation is also driven by contact (for instance, see Video S6). The interaction between foveation to the points of contact with obstacles and inhibition of recently-visited locations leads to thigmotaxis—specifically, the fovea tends to follow obstacle contours and a form of ‘wall-following’ behaviour emerges. Maximum whisker protraction angles are controlled according to a transform driven by the current region of spatial attention and inspired by the ‘Minimal Impingement, Maximal Contact’ (MIMC) hypothesis [20]. In this section, we use this system to reprise three earlier behavioural experiments showing evidence for active touch sensing strategies in the rat—head-turning asymmetry (HTA), contact-induced asymmetry (CIA) and spread reduction (SR). For each study, data are extracted from the simulated model to emulate as closely as possible the original analyses of high-speed digital video recordings of behaving animals.
The results above can be summarised as follows. During exploration in free space, the simulation expresses HTA with a coefficient of linearity between those reported in two behavioural studies. During exploration near walls, the model expresses CIA with a strength comparable to that reported in two behavioural studies. During approach to a wall, the model expresses SR (some reduction in first contacting whisk, substantially more in second) with comparable strength to that reported in a behavioural study. To assess the sensitivity of these results to the ‘Reference’ parameter choices listed in Table 1, we realised the three experiments multiple additional times, making adjustments to one or a few parameters in each case, and assessing the results for the qualitative findings given above. We did not test adjustments to the parameters marked in Table 1 since these are fairly well-defined by previous reports ( is a temporal scale parameter which defines only the overall rate of behaviour; the other three are anatomical parameters). The effect of adjustment of the remaining parameters is reported below.
To begin with, we tried flipping the array along the left/right dimension after it had been built. The asymmetries of HTA and CIA had their senses reversed, as expected, whilst the SR result was somewhat weakened, also as expected. Next, we checked that integration error was not affecting our results by using higher spatial ( mm) and temporal ( s) resolution; the CIA result appeared a little strengthened, but otherwise there was no effect. Similarly, most other adjustments to the parameters (listed in Table 1, column ‘Adjusted’) had only minor effects and did not change the qualitative results; those that did impact the results are now listed. Increasing all three width parameters (, , ) had little impact; decreasing them somewhat weakened the CIA result (though the main lateral bias remained robust). Raising had little effect, but reducing it eliminated plausible gross behaviour in the CIA experiment so that the result could not be measured. Decreasing/increasing the excitation noise gain () strengthened/weakened the results, as expected (at the high noise level, the SR result was qualitatively degraded). Decreasing had little effect; increasing it had little effect on HTA or SR, and only slightly weakened the CIA result, apparently owing to changes in gross behaviour rather than any effect on whisker movement per se. Adjusting the nominal protraction angles up or down affected the scaling just of the SR result, but did not change it qualitatively. Increasing the protraction duty cycle, , to 80% had little effect; reducing it to 50% introduced some noise into the CIA result (though the main lateral bias remained robust). Adjusting the overall modulation strength, , had the strongest effect of any of the tested adjustments, unsurprisingly—however, whilst the strength of all three results was very directly affected, all the results were qualitatively unchanged for all non-zero tested values. As expected, with a modulation strength of zero, both HTA and CIA plots are flat, whilst the SR plot shows a small reduction in spread owing to the measurement of physical whisker deformation.
The central variable of the model is a representation of the immediate region of space attended by the animal which rapidly modulates, through a fixed transform, the maximum protraction angles of the whiskers and drives the movement of the snout (specifically, the positioning of a generalised sensory ‘fovea’ around the mouth) on a longer timescale. Thus, both whisker and head movements are modelled as overt expressions of attention. In the implementation presented, the attended region is represented in the activity of a salience map driven by contact and by an endogenous stochastic signal and inhibited by an IOR mechanism, maximum whisker protraction angles are set by an MIMC-like transform driven by activity in the map, and the fovea is driven towards the location of the peak in the map. This implementation expresses HTA, CIA and SR, when challenged by experimental paradigms equivalent to those used in the behavioural laboratory. Furthermore, these results were robust to parameter variation—this is unsurprising, given the intuitive development of the underlying model presented in Methods.
Attention is a prototypical example of what is generally considered to be a cognitive process. That is, compared to the simpler notion of a reflex arc, attention requires mechanisms that can implement bottom-up filtering of stimuli, working memory for recent events, competitive selection, and top-down modulation (e.g. by motivational systems) (see, e.g. [66] for a review of the nature of attentional processing). Components that implement each of these computations are required to create even a relatively simple model of spatial attention as demonstrated by the model system we describe above. Whilst it is reasonable to seek simpler mechanistic explanations of a phenomenon such as the sensory modulation of whisker movement, there is evidence in a wide-range of domains—time [67]–[69], number [67], [70], reward [71], [72], decision-making [73], [74], space [75]–[78], and working and long-term memory [79], [80]—that rodents process information in a manner that reflects the operation of cognitive mechanisms sometimes approaching, in terms of their sophistication, those identified in primates. We propose that in the case of spatial attention, rat cognition again shares interesting similarities to primate cognition that have been largely overlooked (though, see [81], [82]). Specifically, that models of visual attention using salience maps, that have proved effective in explaining primate eye movement data, could have a useful analogue in the attentional mechanisms underlying rat vibrissal touch.
Whilst not a minimal model in terms of the computations involved, we propose that our attentional hypothesis for rodent whisking modulation is parsimonious in the sense of being explanatorily powerful. That is, the model accounts for multiple observed phenomena (HTA, CIA, SR), and, moreover, does so in a way that is robust to parameter change (see Sensitivity Analysis, above). The model should also naturally reproduce phenomena described in the literature that cannot, even in principle, be explained by reflex mechanisms. Specifically, anticipatory ‘reaching’, in the form of increased whisker protraction, has now been reported in a range of experimental paradigms: Sachdev et al. (2003) [17] reported unilateral reaching in anticipation of contact with a sensor that triggered a reward; Berg & Kleinfeld (2003) [18] reported reaching (alongside changes in temporal parameters) when animals were challenged to contact a discriminandum on the other side of a gap; our own observations of a freely-exploring condition also suggest reaching [20] (see Figure 7B) as does evidence of rats increasing whisker protraction during running [53]; finally, SR also appears to be anticipatory at least in part [23]. All of these experiments used rats, but reaching has also recently been observed in mouse by Voigts et al. (2013) [24], who highlighted that “The precision in amplitude modulation is not due to current sensory input” but rather relies on historical sensory information (i.e. on working memory).
The validity of the attentional explanation of whisking modulation can be further tested in the behavioural laboratory. One key prediction is that non-attended objects will not elicit whisker modulation, as we have previously observed informally in a handful of trials but have not yet quantified [20]. A possible preparation to test this prediction might be, for instance, a motivated animal seeking particular objects preferentially over others positioned nearby. A second key prediction is that whisker movement is modulated by spatial attention, however generated. A preparation for testing this might be an examination of the whisker movements of a head-fixed animal, with spatial attention manipulated by olfactory, auditory, or visual cues rather than by tactile stimuli. If, for instance, a whiff of an attractive odor from a specific direction elicited whisker movement toward that direction this would be strong evidence in favour of an attentional model of whisking modulation, in this case showing cross-modal transfer of target salience.
Our model does not include modulation of whisk frequency, nor changes in whisker movement at very short time-scales. As a result, two notable observations not accounted for by the model are Rapid Cessation of Protraction (RCP) [20], [23] and the ‘touch-induced pump’ (TIP) [40] both of which occur within the time-scale of a single whisk. As previously discussed [20], [83], these observations may reflect the operation of a rapid negative feedback loop, though alternative plausible models for RCP and TIP include (i) that they represent contact-driven changes in the timing of an underlying motor pattern and (ii) that they follow from rapid switches in spatial attention through the attentional mechanism proposed here (given the rapidity of responses in midbrain to whisker contact, [84]). Further experiments will be required to establish whether brainstem mechanisms alone are sufficient to elicit these phenomena.
The model presented is abstract in form and also in substrate, however, neuroscientific evidence does point towards some likely substrates for different aspects of these attentional computations in the rat brain.
Most clearly, the superior colliculus (SC) would be a very plausible location for a spatial attention map to be sited. SC has the right inputs from somatosensory centres—rapid bottom-up inputs arrive directly from trigeminal sensory complex, whilst top-down inputs from somatosensory cortex are also present [84]–[86]—and the sensory organization is topographic [87]–[89]. More broadly, rodent SC receives inputs also from visual and auditory centres [90], reflecting that SC is an important centre for the integration of multi-sensory—specifically, spatial—information [91]. It also has the right outputs: it contains topographic motor maps for both orienting-like head movements [92] and apparently modulatory (non-periodic) whisker movements [93], [94] and has direct efferents to facial nucleus, the motor nucleus associated with the whisker musculature [85], [95]. Salience maps have been identified in SC [1] and it has been strongly implicated in the mediation of visual attention processing [96]–[99]. The proposal that SC plays a key role in rat orienting to whisker stimuli is consistent with its importance for rat prey capture [100]. Interestingly, adult-like HTA, CIA and SR emerge in the post-natal animal during overlapping periods in P12–16 [65], corresponding approximately to the time when SC is reported to be maturing anatomically (around the beginning of the third post-natal week, [89], [101]).
Aside from colliculus, other centres likely to be involved in attention management and/or whisker movement include motor cortex and the basal ganglia. Stimulation of vibrissal motor cortex (vMCx) can evoke whisking-like movements of the whiskers, and the parameters of stimulation affect the parameters of whisking [102]–[104]. In addition, motor cortex ablation significantly disrupts whisking parameters, particularly contralaterally [105]. These data suggest that vMCx is involved in initiating and modulating whisking even though whisking itself appears to rely on a CPG [32], [106], [107]. Activity recorded in vMCx during natural whisking reflects whisking onset as well as variations in amplitude and set-point, consistent with this hypothesis [108]–[110]. Interestingly, motor area M2 in rat has been analogised to the primate Frontal Eye Fields (FEF) [111], a key structure involved in primate oculomotor control and critical in relaying signals from frontal cortex related to voluntary control of visual attention [112]. In addition to projecting to the SC, the FEF, in primates, also project directly to the brainstem saccadic generator so that a primate with a SC lesion is still able to generate saccades. The M2 area in rat likewise has strong reciprocal connections with prefrontal cortex [113], projections to SC [114], and direct brainstem projections to areas involved in orienting [115]. Unilateral lesions in this area have been found to produce contralateral neglect in both primates and rats [111]. The basal ganglia (BG), in both rats and primates, are well-situated to gate switches of attention. SC, whisker somatosensory cortex S1, and whisker motor cortex, all project to similar regions of the dorsolateral striatum (DLS), the input region of the BG [116]. In the case of SC, the projection is via the thalamic intralaminar nuclei [117]. DLS then has an inhibitory projection to BG output structures including the substantia nigra pars reticulata which, in turn, tonically inhibits SC and, via the thalamus, areas of sensory and motor cortex related to the vibrissae, thus completing a double-disinhibitory loop that seems configured to select target representations that are of high salience to the animal [118], [119]. In primates, the role of BG in gating saccadic eye-movements to salient targets has been described in detail by Hikosaka et al. (2000) [120], and it seems plausible that the BG will play a similar role for whisker-guided orienting movement in rats.
The model has two interesting architectural features distinct to this system. First, whisker-centric data are mapped into a head-centric representation of space, implying dynamic routing of sensory data, in analogy to remappings of auditory and somatosensory data in other animals [91]. However, owing to the rhythmic exploration of space by the whiskers (along with inertial or contact-driven bending), the central representation of the periphery is constantly and rapidly on the move in such a model. In SC, rats have an approximately retino-centric topography in the superficial layers, whilst vibrissal data is represented in the deeper layers in spatial register with the overlying visual maps [87], [93]. At the same time, regions sensitive to stimulation of individual whiskers are large and overlapping under anaesthesia [86], [87], particularly in the rostral-caudal dimension, consistent with the large area of the visual field swept by individual whiskers as they move back and forth [50]. Whisker-sensitive cells in primary somatosensory cortex have been reported both to respond most strongly at particular whisker movement phases [121] and to encode whisker bending direction [122], and primary afferent cells that encode whisker phase have also been identified [27]. Thus, this highly dynamic model is consistent with existing data, whilst cells such as those identified could constitute part of a substrate for remapping, as has been previously discussed [27], [121], [122].
Second, whilst visual overt attention is primarily expressed through the azimuth and elevation angles of the eye [96], our model of tactile overt attention hinges upon the radial dimension since the generalised sensory fovea must be brought to an object rather than just pointed at it [9]. Accordingly, the current study could not have been performed without a representation of the radial dimension. In the current study, we did not represent the vertical dimension (primarily because behavioural data are lacking) but we routinely find it necessary to use three-dimensional representations of space as the substrate for spatial orienting in our work with robots (reviewed in [44]). The current proposal can be extended to three dimensions if a three-dimensional representation of the attended region is assumed, but whether extension in this way would respect the biological organisation remains an open and important question.
In summary, then, our findings support the general hypothesis that there exists in the rat a system somewhat homologous to the visual orienting system known from primate studies [96], with the primary outputs being re-location of a generalised sensory fovea around the mouth, supported by body movements as required [92], and adjustment of the protraction angles of the whiskers, perhaps to favour a ‘Minimal Impingement, Maximal Contact’-like control aim. Within this system, superior colliculus may well play a key role, along with areas of cortex and the basal ganglia [111], [123]. This system probably forms only part of a larger system that generates whisker movements but most or all non-periodic components of motion may be mediated therein. Thus, this sensorimotor model has the potential to substantially improve our understanding of the modulations of periodic whisker movements that are observed in behaving animals. As highlighted recently by Schwarz et al. (2010) [124], a particular disadvantage of the head-fixed rat preparation is that the behavioural repertoire of rodents includes many whole-body movements, whisker movements being an exception. In contrast to widely-studied rodent attentional measurement paradigms (such as the 5-choice serial reaction time task, [125]), whisker movements could reveal attention on relatively short timescales, in considerable spatial detail, optionally in head-fixed preparations, with measurement remaining highly automatable. Thus, if whisker movements can be confirmed to reveal the region of spatial attention, their observation might provide a novel and practical tool for its investigation in small mammals.
|
10.1371/journal.pgen.1001323 | Differential Genetic Associations for Systemic Lupus Erythematosus Based on Anti–dsDNA Autoantibody Production | Systemic lupus erythematosus (SLE) is a clinically heterogeneous, systemic autoimmune disease characterized by autoantibody formation. Previously published genome-wide association studies (GWAS) have investigated SLE as a single phenotype. Therefore, we conducted a GWAS to identify genetic factors associated with anti–dsDNA autoantibody production, a SLE–related autoantibody with diagnostic and clinical importance. Using two independent datasets, over 400,000 single nucleotide polymorphisms (SNPs) were studied in a total of 1,717 SLE cases and 4,813 healthy controls. Anti–dsDNA autoantibody positive (anti–dsDNA +, n = 811) and anti–dsDNA autoantibody negative (anti–dsDNA –, n = 906) SLE cases were compared to healthy controls and to each other to identify SNPs associated specifically with these SLE subtypes. SNPs in the previously identified SLE susceptibility loci STAT4, IRF5, ITGAM, and the major histocompatibility complex were strongly associated with anti–dsDNA + SLE. Far fewer and weaker associations were observed for anti–dsDNA – SLE. For example, rs7574865 in STAT4 had an OR for anti–dsDNA + SLE of 1.77 (95% CI 1.57–1.99, p = 2.0E-20) compared to an OR for anti–dsDNA – SLE of 1.26 (95% CI 1.12–1.41, p = 2.4E-04), with pheterogeneity<0.0005. SNPs in the SLE susceptibility loci BANK1, KIAA1542, and UBE2L3 showed evidence of association with anti–dsDNA + SLE and were not associated with anti–dsDNA – SLE. In conclusion, we identified differential genetic associations with SLE based on anti–dsDNA autoantibody production. Many previously identified SLE susceptibility loci may confer disease risk through their role in autoantibody production and be more accurately described as autoantibody propensity loci. Lack of strong SNP associations may suggest that other types of genetic variation or non-genetic factors such as environmental exposures have a greater impact on susceptibility to anti–dsDNA – SLE.
| Systemic lupus erythematosus (SLE) is a chronic autoimmune disease that can involve virtually any organ system. SLE patients produce antibodies that bind to their own cells and proteins (autoantibodies) which can cause irreversible organ damage. One particular SLE–related autoantibody directed at double-stranded DNA (anti–dsDNA) is associated with kidney involvement and more severe disease. Previous genome-wide association studies (GWAS) in SLE have studied SLE itself, not particular SLE manifestations. Therefore, we conducted this GWAS of anti–dsDNA autoantibody production to identify genetic associations with this clinically important autoantibody. We found that many previously identified SLE–associated genes are more strongly associated with anti–dsDNA autoantibody production than SLE itself, and they may be more accurately described as autoantibody propensity genes. No strong genetic associations were observed for SLE patients who do not produce anti–dsDNA autoantibodies, suggesting that other factors may have more influence in developing this type of SLE. Further investigation of these autoantibody propensity genes may lead to greater insight into the causes of autoantibody production and organ damage in SLE.
| Systemic lupus erythematosus (SLE) is the prototypic systemic autoimmune disease and can affect virtually any organ system. Manifestations of SLE are quite varied and include renal failure, hemolytic anemia, arterial and venous clots, and disfiguring skin rashes. Overall prevalence of SLE in the general population is 1 in 2000 individuals with a predilection for women (female to male ratio of 6-10∶1) [1]. Although the prevalence is relatively low, SLE creates tremendous health care and societal costs since affected individuals are typically young and can suffer significant morbidity and early mortality.
The pivotal immunologic disturbance in SLE is the formation of autoantibodies directed at cell membrane and nuclear components. Deposition of immune complexes containing these autoantibodies leads to inflammatory responses and end-organ damage. Autoantibodies directed against native double-stranded DNA (dsDNA) are observed in 40–60% of SLE patients. Anti–dsDNA autoantibodies can be present prior to clinical symptoms of SLE [2], and are implicated in the pathogenesis of lupus nephritis, a major cause of morbidity and mortality in SLE [3], [4]. Anti–dsDNA autoantibodies have also been associated with decreased survival [4]. Given its high specificity for SLE, anti–dsDNA autoantibody production is one of the 11 classification criteria for SLE developed by the American College of Rheumatology (ACR) [5], [6].
SLE susceptibility is strongly influenced by both genetic and environmental factors. Recent genetic association studies have successfully identified over 20 SLE susceptibility loci [7]. Odds ratios (OR) for these associations have been modest, with most OR <1.3. One potential factor influencing the magnitude of these associations may be the extensive clinical heterogeneity of SLE. Studying more specifically defined SLE manifestations may reveal stronger and novel genetic associations. Therefore, we conducted a genome-wide association study of anti–dsDNA autoantibody production in SLE to identify genetic associations with this clinically relevant autoantibody, and to determine if the genetic associations were different between those SLE subjects that do and do not produce this autoantibody.
For this genome-wide association study (GWAS), we utilized genotyping data from the GWAS of SLE published by Hom et al. [8] as the discovery dataset, and genotyping data from the GWAS of SLE published by The International Consortium on the Genetics of Systemic Lupus Erythematosus (SLEGEN) [9] as the replication dataset. Since both datasets utilized publicly available healthy controls from the same sources, we supplemented the controls in the replication dataset with 1142 healthy controls from the Cancer Genetic Markers of Susceptibility (CGEMS) study (http://cgems.cancer.gov/data/) [10]. After employing data quality measures, including removal of duplicate and related subjects (see Methods and Figure 1), a total of 1717 SLE cases and 4813 healthy controls of European descent were studied. The discovery dataset was comprised of 1278 SLE cases and 3334 healthy controls, while the replication dataset was comprised of 439 SLE cases and 1479 healthy controls. For both datasets, 47% of the SLE cases were anti–dsDNA +. In the joint dataset, 296,509 SNPs were typed in common between the discovery and replication datasets and passed data quality measures (see Methods). An additional 124,809 imputed SNPs (see Methods) passed data quality filters and were included for analysis in the replication and joint datasets. Figure 1 summarizes the autoantibody status and sample sizes of the datasets used in this study, as well as the number of individuals removed for each data quality measure. The clinical characteristics of the subjects in this study, provided in Table 1, are comparable to those in previously reported studies [1], [11].
We first compared anti–dsDNA + SLE cases to healthy controls using additive logistic regression models implemented in PLINK [12] (http://pngu.mgh.harvard.edu/purcell/plink/). The discovery and replication datasets were analyzed separately, and then combined into a “joint analysis” for maximal statistical power. All logistic regression models were adjusted for population stratification using principal components analysis. Table S1 presents the genomic control inflation factor (λGC) for each analysis prior to and after adjustment for population stratification. P-values for association were adjusted for the λGC observed after accounting for population stratification (see Methods for additional details). The quantile-quantile and Manhattan plots for the joint analysis are displayed in Figure S1.
Table 2 (and Table S2) displays each locus with significant (p<5E-07) or suggestive (p<1E-05) evidence of association in the joint analysis. Excluding the associations seen with the major histocompatilibity complex (MHC) on chromosome 6p21, 14 statistically significant associations were observed in the joint analysis of genotyped SNPs when none would have been expected under the null hypothesis. The most significant associations were observed in the MHC, with rs1150754 near TNXB (ORjoint 2.21, 95% CI 1.92–2.53, p = 6.4E-29) as the most significantly associated SNP. Outside of the MHC, the most significantly associated SNP was rs7574865 (ORjoint 1.77, 95% CI 1.57–1.99, p = 2.0E-20) located in STAT4. Strong evidence of association was observed with SNPs in/near IRF5 and ITGAM. Association results for these loci met the genome-wide significance threshold in both datasets, and thus are considered replicated findings. These 3 loci were previously shown to be associated with SLE [8], [13], [14], but only STAT4 has been previously associated with anti–dsDNA autoantibody production [15], [16]. Strong evidence of association was also observed for BLK in the joint analysis, but this association did not meet the threshold for genome-wide significance in both datasets. While SNPs in/near LAMC2 and COL25A1 met the genome-wide significance threshold in the discovery dataset, these associations were not observed in the replication dataset, possibly due to the limited statistical power of the second dataset.
Suggestive findings of association in the joint dataset (p<1E-05) were seen with SNPs near or in the PTTG1, UBE2L3, SLC1A7, and KIAA1542 loci, and with rs10737562 (no known gene within 100 kb). PTTG1, KIAA1542, and UBE2L3 have been shown to be associated with SLE [7], [9], and thus, are likely true associations that are specific for anti–dsDNA + SLE. The associations with SLC1A7 and rs10737562 have not been previously reported with SLE or anti–dsDNA autoantibody production, and should be replicated in another collection of anti–dsDNA + SLE cases.
Anti–dsDNA + SLE cases (n = 811) were also compared to the combined group of anti–dsDNA – SLE cases and healthy controls (n = 5719) to conduct an analysis maximally powered to identify SNPs only associated with anti–dsDNA + SLE. No new loci (i.e., loci not presented in Table 2) displayed significant or suggestive evidence of association. Also, analyses comparing anti–dsDNA + SLE cases to healthy controls excluding ANA negative subjects showed results similar to Table 2 (data not shown).
Next, we compared anti–dsDNA – SLE cases to healthy controls in the discovery, replication, and combined “joint analysis” datasets using additive logistic regression models adjusted for population stratification as described previously. Table S1 presents the genomic control inflation factor (λGC) for each analysis prior to and after adjustment for population stratification. The quantile-quantile and Manhattan plots for the joint analysis are displayed in Figure S2.
Far fewer statistically significant genetic associations were observed. Excluding the MHC, one statistically significant association was observed in the joint analysis of genotyped SNPs when none would have been expected under the null hypothesis (p<5E-07). The most significant associations were again seen in the MHC, with rs2301271 (∼9 kb downstream from HLA-DQA2) as the most significantly associated MHC SNP (ORjoint 1.47, 95% CI 1.32–1.63, p = 2.0E-12). In the discovery dataset, no SNPs outside of the MHC met our genome-wide significance threshold. In the joint analysis (Table 3), an additional association with rs10488631 near IRF5 met genome-wide significance (ORjoint 1.57, 95% CI 1.35–1.82, p = 6.2E-09). Three SNPs had suggestive evidence of association in the joint analysis: rs2669010 in RPL7AP50, rs918959 (no known gene within 100 KB), and the missense SNP rs7927370 in OR4A15. These novel findings need to be replicated in another collection of anti–dsDNA – SLE cases. Analyses excluding ANA negative subjects showed similar results (data not shown).
Using the combined dataset, we compared the anti–dsDNA + SLE cases (n = 811) to the anti–dsDNA – SLE cases (n = 906) using additive logistic regression models. Minimal population stratification was observed between these two groups (λGC = 1.01) without adjustment using principal components. However, we included the principal components in these models to decrease the possible influence of subtle stratification on our findings (λGC = 1.00 after adjustment for population stratification).
No SNP met our genome-wide significance threshold (p<5E-07) for anti–dsDNA autoantibody production in this analysis. Six SNPs showed suggestive evidence of association (p<1E-05), as shown in Table 4. Only three SNPs would be expected to have a p<1E-05 under the null hypothesis. Similar to the anti–dsDNA + analysis described above, rs7574865 in STAT4 was once again found to be associated with anti–dsDNA + SLE. rs1463525 in NAALADL2 is of interest, since this gene was recently identified as a susceptibility locus for Kawasaki disease [17], another autoimmune disease. However, the most significantly associated SNP for Kawasaki disease (rs17531088) is not in linkage disequilibrium with the SNP identified in our analysis (r2 = 0.002 in the CEU HapMap population). The statistical power for this analysis was limited by our relatively smaller sample size. Thus, additional studies are needed to fully explore this area and to replicate our findings.
Since two of the suggested SNPs are located within the MHC, our findings indicate that MHC associations may be heterogeneous between these two subgroups of SLE. This finding is further supported by a plot of the p-values for association among the MHC SNPs, as shown in Figure 2. The strongest MHC associations with anti–dsDNA autoantibody production among these SLE patients were within the class II region. Given the extensive linkage disequilibrium of the MHC, many of these associations may be driven by the MHC class II locus HLA-DRB1, a well established SLE susceptibility gene [18], [19].
Next, we examined the magnitude of association between 22 polymorphisms previously associated with SLE in Gateva et al. [7], stratified by anti–dsDNA autoantibody status using tests of heterogeneity. For each SNP, the association result in the anti–dsDNA + versus healthy control analysis was compared to the association result for the anti–dsDNA – versus healthy control analysis. A p-value of less than 0.05 was considered significant evidence of heterogeneity.
Table 5 presents the results of the tests of heterogeneity, along with the association results from the case-only analysis, for these 22 SLE susceptibility loci. Associations for HLA-DR3 (indicated by its tagSNP rs2187668) and SNPs in STAT4 and ITGAM differed substantially between the two anti–dsDNA subgroups (pheterogeneity<0.005). In addition, SNPs in the BANK1, KIAA1542, ITGAM, and UBE2L3 regions also showed differential associations in the two anti–dsDNA subgroups (p<0.05). For all of these SNPs, the associations with anti–dsDNA + SLE had stronger OR and smaller p-values when compared to anti–dsDNA – SLE or SLE itself. The differences are best demonstrated by rs7574865 in STAT4: OR for anti–dsDNA + SLE 1.77 (95% CI 1.57–1.99, p = 2.0E-20) compared to OR for anti–dsDNA – SLE 1.26 (95% 1.12–1.41, p = 2.4E-4), with p-value for the test of heterogeneity <0.0005. In contrast, ORs of association were quite similar between the 2 SLE subgroups for SNPs in/near the FCGR2A, OX40L, IL10, PXK, UHRF1BP1, PRDM1, BLK, and IRAK1 regions.
When examining these SNPs in the case-only analysis, rs2476601 (PTPN22), rs10488631 (IRF5), and rs2431099 (PTTG1) were more strongly associated with anti–dsDNA + SLE than anti–dsDNA – SLE (p<0.05). Sensitivity analysis of the 722 SLE cases with longitudinal anti–dsDNA autoantibody data (of which 46% were anti–dsDNA +, see Methods) showed good consistency in OR with the analyses performed using the full dataset (data not shown).
Among the SNPs studied in this comparison, we did not identify a single SNP that was more strongly associated with anti–dsDNA – SLE disease than anti–dsDNA + SLE or SLE itself, based on OR or p-values.
To study the relationship between cumulative genetic risk and anti–dsDNA autoantibody production, we calculated an SLE genetic risk score (GRS) by counting the total number of risk alleles an individual had for the 22 SLE-associated SNPs listed in Table 5. The mean SLE GRS was higher in anti–dsDNA + SLE cases (15.5, SD 3.1) compared to anti–dsDNA – SLE cases (14.5, SD 3.0) and healthy controls (13.1, SD 2.8), and the trend was highly statistically significant (ptrend = 1.0E-102). In logistic regression analyses adjusting for study source and population stratification, the odds of producing anti–dsDNA among SLE cases increased by 12% (OR 1.12, 95% CI 1.09–1.16) for each 1 unit increase in the SLE GRS. When comparing to healthy controls, the odds of having anti–dsDNA + SLE increased by 32% (OR 1.32, 95% CI 1.28–1.35) for each 1 unit increase in the SLE GRS versus 18% (OR 1.18, 95% CI 1.15–1.21) for anti–dsDNA – SLE. These findings indicate that SLE cases with higher genetic risk are more likely to be anti–dsDNA positive. A more thorough investigation of the association between SLE GRS and SLE manifestations is presented in Taylor et al. [20].
In this paper, we present the first GWAS of anti–dsDNA autoantibody production in SLE. We have shown that SNPs in the MHC, STAT4, IRF5, and ITGAM regions are associated with anti–dsDNA + SLE. Only SNPs in the MHC and IRF5 met genome-wide significance threshold levels in the analysis of anti–dsDNA – SLE, with lower OR and larger p-values compared to their associations with anti–dsDNA + SLE. Furthermore, many of the previously identified SLE susceptibility loci showed differential associations between the 2 anti–dsDNA subgroups. Using a genetic risk score analysis, we found that SLE cases with a greater number of risk alleles were more likely to be anti–dsDNA +. These results suggest that genetic factors may have a greater influence in the development of anti–dsDNA + SLE as compared to anti–dsDNA – SLE.
The strongest association signals for both the anti–dsDNA + and anti–dsDNA – analyses were observed with MHC SNPs. Previous studies have shown that the strongest, most consistent genetic signals with SLE have been with the HLA-DR2 and HLA-DR3 MHC serotypes [18], [19]. While we confirm these findings, we also show that the HLA-DR3 association with SLE (as suggested by its tagSNP, rs2187668) is far stronger in anti–dsDNA + SLE as compared to anti–dsDNA – SLE or SLE itself. Thus, the HLA-DR3 allele may have a greater impact on the propensity to produce autoantibodies compared to SLE susceptibility generally. Although a similar finding was observed with HLA-DR2 (tagSNP rs9271366), the test of heterogeneity was not statistically significant, possibly due in part to decreased statistical power since the DR2 tagSNP is less common than the DR3 tagSNP (DR2 tagSNP minor allele frequency 0.182 in anti–dsDNA + SLE, 0.167 in anti–dsDNA – SLE, and 0.143 in healthy controls). Examination of other MHC SNPs in the case-only analysis indicates that other (non-HLA-DRB1) loci may have associations with anti–dsDNA autoantibody production beyond the associations observed with SLE.
In addition to the HLA-DR3 tagSNP discussed above, the associations between the STAT4 and ITGAM SNPs and anti–dsDNA + SLE were stronger in magnitude than the associations with SLE per se in our datasets (Table 5). The smaller p-values seen in the associations for these loci with anti–dsDNA + SLE are especially striking given the substantially smaller sample size of this subgroup. Our results imply that STAT4, ITGAM, and HLA-DR3 may be more accurately considered “autoantibody propensity loci” rather than simply “SLE susceptibility loci” given their significant tests of heterogeneity (p<0.05). Using this criterion, three other previously identified SLE susceptibility loci may also be considered autoantibody propensity loci: KIAA1542, BANK1, and UBE2L3. In fact, these SNPs had no evidence of association with anti–dsDNA – SLE in this study (p>0.05). By characterizing these SNPs as autoantibody propensity loci, we identify a potential mechanistic role for these disease associations.
Are these autoantibody propensity loci associated with other autoantibodies? In rheumatoid arthritis (RA), other alleles of the HLA-DRB1 locus (collectively referred to as the “shared epitope”) are associated with anti-CCP autoantibody positivity [21]. While a study of STAT4 (rs7574865) in an early RA inception cohort suggested an association with the anti-CCP autoantibody [22], others have not a shown strong association between this SNP and seropositivity in RA [23]. PTPN22 (rs2476601) has been shown to be more strongly associated with autoantibody positive RA [22]. In our study, other SLE-related autoantibodies (anti-SSA, anti-SSB, anti-Sm, and anti-RNP) are more frequent in the anti–dsDNA + subgroup (Table 1), but correlations between anti–dsDNA and these other autoantibodies antibodies are modest, with Pearson correlation coefficients <0.2 (data not shown). Thus, additional studies are needed to further investigate whether these or other loci are associated with other autoantibodies.
Of note, not all of the previously identified SLE susceptibility SNPs showed differential associations between the anti–dsDNA subgroups. In fact, the OR for the SNPs in or near FCGR2A, OX40L, PXK, and UHRF1BP1 were strikingly similar between the anti–dsDNA + and anti–dsDNA – subgroups. These loci may represent more generalized SLE susceptibility loci, and their mode of conferring SLE disease risk is likely independent of anti–dsDNA autoantibody production. While PTPN22 (rs2476601), IRF5 (rs10488631), and PTTG1 (rs2431099) do not fulfill our criterion as autoantibody susceptibility loci, the results of the case-only analysis suggest that these loci may have a stronger effect in anti–dsDNA + SLE.
Interestingly, far fewer associations were seen in the anti–dsDNA – SLE analysis. Even in the joint analysis, which had the most statistical power, only 1 SNP outside of the MHC met our genome-wide significance threshold—rs10488631 in IRF5. This finding may be explained by a number of different reasons. SNP associations for anti–dsDNA – subgroup may be weaker, and thus would require a larger sample of anti–dsDNA – SLE cases in order to be identified. Other types of genetic variation or non-genetic factors, such as environmental exposures [24], [25], may have a stronger influence on susceptibility to anti–dsDNA – SLE. Lastly, the anti–dsDNA – subgroup may be more clinically heterogeneous or be comprised of individuals who develop SLE through different pathogenic (and genetic) mechanisms, thus decreasing our statistical power to identify genetic associations with this subgroup.
One limitation of this study is the potential misclassification of anti–dsDNA autoantibody status. This misclassification may have occurred because the anti–dsDNA autoantibody was assessed by different assays between the participating case collections, and a patient's anti–dsDNA status can vary over the disease course. However, this misclassification would bias our findings of differences between anti–dsDNA + and anti–dsDNA – SLE towards the null. Moreover, sensitivity analyses performed using the available longitudinal data showed consistent ORs, suggesting that the potential misclassification did not greatly influence our results. A second limitation is that all participants were of European descent. Limiting this study to those of European descent minimizes confounding due to genetic differences arising from differences in ethnicity. Future efforts should study non-European populations given their increased incidence of SLE [26], [27].
In summary, this GWAS of anti–dsDNA autoantibody production in SLE shows that there are more, and stronger, genetic associations in anti–dsDNA + SLE compared to anti–dsDNA – SLE. Previously identified SLE susceptibility loci such as STAT4, ITGAM, KIAA1542, BANK1, and UBE2L3 are more strongly associated with anti–dsDNA + SLE and may confer disease risk through their role in autoantibody production. Weaker associations in anti–dsDNA – SLE may suggest that other types of genetic variation or non-genetic factors have a greater impact on disease risk. Lastly, focusing genetic studies on clinical disease characteristics decreases the heterogeneity that could cloud association results and may provide greater insight into pathogenic disease mechanisms.
Written informed consent was obtained from all study participants and the institutional review board at each collaborating center approved the study.
For this study, all SLE cases and healthy controls were of European descent. All SLE cases fulfilled at least 4 ACR classification criteria for SLE [5], [6]. Figure 1 presents the final sample sizes in the discovery and replication datasets, and the final number of SNPs advanced to analysis.
Anti–dsDNA autoantibody status for all SLE cases was determined by medical record review and/or serologic testing of banked serum. Since anti–dsDNA autoantibody status can fluctuate with disease activity, a SLE case had to have at least one definitively positive laboratory result to be considered anti–dsDNA +. A SLE subject was considered anti–dsDNA – if all laboratory results in the medical record and serologic testing for this autoantibody were negative.
Longitudinal anti–dsDNA autoantibody status (i.e., at least 2 individually documented measurements) was available for a subgroup of SLE cases (n = 722). These data were used for a sensitivity analysis, where anti–dsDNA + was defined as having at least 2 positive anti–dsDNA laboratory results in the longitudinal data, and anti–dsDNA – was defined as having all negative laboratory results for this autoantibody in the longitudinal data.
Three GWAS were performed: anti–dsDNA + SLE subjects versus healthy controls, anti–dsDNA – SLE subjects versus healthy controls, and anti–dsDNA + SLE subjects versus anti–dsDNA – SLE subjects (referred to as the case-only analysis). To determine if genetic associations were significantly different between the 2 anti–dsDNA subgroups, tests of heterogeneity were performed for previously identified SLE susceptibility loci. For these loci, the case-only analysis was also repeated using only the longitudinal dataset as a sensitivity analysis. Lastly, a genetic risk score analysis was conducted using logistic regression.
For each GWAS, associations with anti–dsDNA autoantibody status were assessed using additive logistic regression models implemented in PLINK and included the first 5 principal components as co-variates to adjust for population stratification. Principal components were calculated using EIGENSTRAT [29] (http://genepath.med.harvard.edu/~reich/Software.htm). After removal of SNPs in regions with extensive linkage disequilibrium on chromosomes 5 (44–51.5 Mb), 6 (25–33.5 Mb), 8 (8–12 Mb), 11 (45–57 Mb), and 17 (40–43 Mb), the remaining SNPs common to all of the genotyping platforms were used to calculate the principal components. The first 5 principal components were selected based on review of the eigenvalues for the first 10 principal components. A plot of the first two principal components for each individual in the study is shown in Figure S3. P-values for association were adjusted for the genomic control inflation factor (λGC) observed for each analysis after accounting for population stratification (Table S1). Each GWAS also included calculation of the expected p-value distribution using PLINK to determine the expected number of statistically significant SNPs.
Analyses were conducted separately for the discovery and replication datasets. These datasets were then combined into a “joint dataset” for maximal statistical power. The study source (discovery versus replication dataset) was included as a co-variate in analyses of the joint dataset. For the discovery and joint datasets, a p-value of less than 5E-07 was considered statistically significant, and p-values between 5E-07 and 1E-05 were considered suggestive of association. Statistically significant SNPs in the discovery dataset were examined in the replication dataset, where a p-value of less than 0.005 was considered statistically significant. Analyses of the discovery, replication, and joint datasets first used only the assayed SNPs. These analyses were repeated for the replication and joint datasets to include the imputed SNPs that passed the data quality filters, since the λGC was expected to differ in analyses using imputed SNPs.
Based on the publication by Gateva et al. [7], 22 SNPs with previously established evidence of association with SLE were analyzed further. For these SNPs (or their proxy, if the listed SNP was not genotyped), the association results for the SNP in two of the GWAS conducted above (anti–dsDNA + or - versus healthy controls) were compared using tests of heterogeneity (STATA 9.0/SE, College Station, TX). A p-value of less than 0.05 was considered significant evidence of heterogeneity. The results for these SNPs were also examined for the case-only analysis (including the sensitivity analysis with longitudinal data), where anti–dsDNA + SLE cases were compared to anti–dsDNA – SLE cases. A p-value of less than 0.05 was considered significant evidence of a differential association between the 2 subgroups. For comparison, the association with SLE was assessed using the logistic regression methods described above.
Associations between the SLE GRS and anti–dsDNA status were calculated using logistic regression models (STATA 9.0/SE, College Station, TX). These models utilized the SLE GRS as a continuous predictor, and adjusted for population stratification (using the first 5 principal components) and study source.
|
10.1371/journal.pcbi.1003217 | Modeling the Effect of APC Truncation on Destruction Complex Function in Colorectal Cancer Cells | In colorectal cancer cells, APC, a tumor suppressor protein, is commonly expressed in truncated form. Truncation of APC is believed to disrupt degradation of β—catenin, which is regulated by a multiprotein complex called the destruction complex. The destruction complex comprises APC, Axin, β—catenin, serine/threonine kinases, and other proteins. The kinases and , which are recruited by Axin, mediate phosphorylation of β—catenin, which initiates its ubiquitination and proteosomal degradation. The mechanism of regulation of β—catenin degradation by the destruction complex and the role of truncation of APC in colorectal cancer are not entirely understood. Through formulation and analysis of a rule-based computational model, we investigated the regulation of β—catenin phosphorylation and degradation by APC and the effect of APC truncation on function of the destruction complex. The model integrates available mechanistic knowledge about site-specific interactions and phosphorylation of destruction complex components and is consistent with an array of published data. We find that the phosphorylated truncated form of APC can outcompete Axin for binding to β—catenin, provided that Axin is limiting, and thereby sequester β—catenin away from Axin and the Axin-recruited kinases and . Full-length APC also competes with Axin for binding to β—catenin; however, full-length APC is able, through its SAMP repeats, which bind Axin and which are missing in truncated oncogenic forms of APC, to bring β—catenin into indirect association with Axin and Axin-recruited kinases. Because our model indicates that the positive effects of truncated APC on β—catenin levels depend on phosphorylation of APC, at the first 20-amino acid repeat, and because phosphorylation of this site is mediated by , we suggest that is a potential target for therapeutic intervention in colorectal cancer. Specific inhibition of is predicted to limit binding of β—catenin to truncated APC and thereby to reverse the effect of APC truncation.
| We asked the question, how can the effects of APC truncation, a very common mutation in colorectal cancer, be understood and reversed? We addressed this question by formulating a computational model for destruction complex function that incorporates site-specific details about protein-protein interactions and protein phosphorylation and examined the differences in predicted behaviors when APC is full length, as in normal cells, and truncated, as in colorectal cancer cells. Our model offers an explanation for how and why destruction complex function is altered by APC truncation. The model indicates that phosphorylation of the first 20-amino acid repeat in APC (which is usually the only 20-amino acid repeat that remains in truncated forms of APC) together with the absence of SAMP repeats (missing entirely because of truncation) allows truncated APC to act as a diversion sink. In other words, phosphorylated APC can outcompete Axin for binding to , provided Axin is limiting, and thereby prevent from associating with Axin and the Axin-associated kinases and , which initiate phosphorylation-dependent degradation of . Thus, the model identifies inhibition of APC phosphorylation, which is mediated by , as a potential means by which the oncogenic effect of APC truncation could be reversed.
| (CTNNB1) is a key signaling protein in the pathway [1], [2], a regulator of cadherin cell adhesion molecules [3], and a regulator of the Tcf and Lef family of transcription factors [4]–[7]. In mesenchymal cells, levels increase when a Wnt ligand binds a cell-surface Frizzled (Fz)-family receptor. Activation of the Wnt/ pathway (transiently) inhibits proteosome-mediated degradation of . Wnt binding also has other important effects on , including regulation of phosphorylation state and redistribution of within subcellular compartments. In colorectal cancer cells, normal control of degradation is disrupted, resulting in elevated levels of .
Cellular degradation of is regulated by (in our view) oligomeric protein complexes, which have diverse compositions but common features; these complexes are often collectively referred to as the destruction complex [8]–[10]. The destruction complex, which characteristically contains and two scaffold proteins, Axin (axis inhibition protein, AXIN1) and APC (adenomatous polyposis coli protein), mediates phosphorylation of by recruiting (glycogen synthetase , GSK3B) and (casein kinase , CSNK1A1) [11]–[15]. These kinases, upon binding Axin, catalyze phosphorylation of on specific serine and threonine residues. Phosphorylation of Ser-45 by and subsequent phosphorylation of Ser-33, Ser-37, and Thr-41 by initiates ubiquitination and proteosome-mediated degradation of [12]–[15]. The destruction complex also recruits PP2A, a multimeric protein phosphatase, which opposes the action of kinases. It has been suggested that activation of Wnt/ signaling destabilizes the destruction complex by sequestering Axin in complexes with activated Fz receptors [16]–[18]. However, details about the early events in Wnt/ signaling are still emerging [19], [20]. In colorectal cancer cells, the destruction complex member APC is often truncated [21]. An important effect of APC truncation is believed to be a perturbation of the interactions amongst proteins comprising the destruction complex that alters regulation of degradation, perhaps by destabilizing the destruction complex in a way similar to the destabilization brought about by Wnt signaling.
The interactions responsible for assembly of the destruction complex are complex and are mediated by multiple functional sites within the member proteins of the destruction complex. The characteristic core of the destruction complex can be viewed as a ternary complex that forms through interactions of APC, Axin, and . contains twelve ARM (Armadillo) repeats, allowing it to bind both APC and Axin. In particular, ARM repeats 3 and 4 constitutively bind a central region of Axin [22], [23] as well as a phosphorylated 20-amino acid (20-aa) repeat region of APC [24], [25]. There are total of seven 20-aa repeats in this region. ARM repeats 5–9 constitutively bind three 15-amino acid (15-aa) repeats in the N-terminal region of APC [26]. APC contains three SAMP (serine-alanine-methionine-proline) repeats, which bind the RGS (regulator of G protein signaling) domain of Axin [27]. These interactions connect the three core proteins of the destruction complex (APC, Axin, and ) and enable each protein to bind the other two core proteins, possibly within a closed/cyclic ternary complex. A cyclic complex would presumably be highly stable, because dissociation of such a complex would require the sequential break up of two protein-protein interactions.
Stability of the destruction complex may be important for its function as a platform for phosphorylation of , and other proteins. The destruction complex mediates phosphorylation of by allowing Axin to colocalize the kinases and with their substrate . Axin contains binding sites for both [11], [28] and [12], [29]. Interestingly, the destruction complex is also thought to mediate phosphorylation of APC by colocalizing another kinase, (CSNK1E), with APC [30], although it is not known which protein in the destruction complex recruits . and together mediate phosphorylation at the 20-aa repeat region of APC [30]. Recall that this region in APC, when phosphorylated, mediates interaction with a site in that also interacts with Axin [22]–[25]. Thus, phosphorylated APC and Axin compete for binding to . The outcome of this competition is perhaps dependent on stability of the destruction complex.
Much of what we know about the functional effects of APC truncation has come from studies of a human colon adenocarcinoma cell line (SW480). SW480 cells express a truncated form of APC termed APC1338, which contains only the first 1,338 amino acids of the full-length protein [9], [31]. APC1338 contains all three 15-aa repeats and the first 20-aa repeat, but is devoid of the remaining six 20-aa repeats and the SAMP repeats, which bind Axin [9], [31]. Therefore, a model can be conceptualized wherein assembly of the functional destruction complex cannot be completed in the absence of interaction between APC1338 and Axin, leading to decreased phosphorylation, ubiquitination, and degradation of [2]. However, an absence of SAMP repeats in APC does not prevent direct binding of Axin to [22], [23], and there are some uncertainties about the validity of this model [19] because reports from different laboratories have shown that expression of recombinant APC can either promote degradation of or have no or little effect, depending on cell type and whether APC is expressed transiently or stably [31]–[34].
As discussed above, APC plays an important role in destruction complex function. However, APC is a multifunctional protein, subject to numerous post-translational modifications. It is believed to play a role in regulating not only phosphorylation and ubiquitination of but also localization of . There are several pools of : membrane-associated (e.g., complexed with E-cadherin), cytosolic (free, bound and Tcf bound), and nuclear. Other components of the destruction complex are also multifunctional proteins, which can be found in distinct subcellular locations and states. For example, Axin, through self-polymerization mediated by its DIX (dishevelled and axin) domain [35], localizes to cytoplasmic puncta. We will not consider these complexities, but they are mentioned at this point to caution the reader about the limitations of our study.
Here, our focus will be on APC regulation of phosphorylation within an idealized destruction complex, taken to comprise a ternary complex of APC, Axin, and with 1∶1∶1 stoichiometry. We will consider the site-specific details of the interactions amongst these proteins, because these details are relevant for understanding how the interactions of APC, Axin, and are perturbed by an absence of SAMP repeats in truncated APC (APC1338). We also consider, with less mechanistic resolution, proteins that mediate phosphorylation and dephosphorylation of APC and and degradation of . The set of proteins of interest are considered in isolation. Thus, for example, we do not consider interaction with E-cadherin, or the effects of Wnt. We also do not consider Axin puncta or the DIX domain in Axin. Axin puncta play a role in degradation but are not required for phosphorylation of [36].
To investigate the roles of APC and its oncogenic truncated forms in destruction complex function, we formulated a computational model for regulation of phosphorylation and degradation using local rules to represent the protein-protein interactions of interest [37]–[39]. This rule-based approach, ideal for modeling the chemical kinetics of biomolecular interaction networks, allowed us to consider the mechanistic details of protein-protein interactions at the resolution level of functional sites within the proteins of interest. These mechanistic details are complex, as summarized above, and arguably beyond our ability to comprehend without reasoning aids, such the model considered here. Using this model, we interrogated system behavior, which emerges from the states and state changes of protein sites, with the goal of elucidating the distinctive mechanisms by which APC and APC1338 regulate the rate of destruction in normal and SW480 colorectal cancer cells. We also used our model to investigate the functional significance of intracomplex interactions among APC, Axin, and , which have the potential to produce a highly stable cyclic ternary complex.
Although APC is a characteristic component of the destruction complex and thought to be important for degradation of [31]–[34], our analyses suggest that APC does not promote degradation of in a normal cell when overexpressed. However, we do predict that expression of recombinant full-length APC in SW480 cells promotes degradation, as seen in several studies [31]–[34]. These results are obtained because, according to our model, phosphorylated APC1338 in SW480 cells competes with Axin for . APC1338-mediated separation of from Axin reduces phosphorylation of by Axin-recruited kinases, and reduced phosphorylation of decreases its rate of degradation. In contrast, in normal cells, binding of phosphorylated full-length APC to , in competition with Axin, is not functionally equivalent because Axin can still colocalize with through indirect association via the SAMP repeats in APC, which are missing in APC1338. Because of these results and because is responsible for phosphorylation of APC (but not ), we identify as a potential target for therapeutic intervention in colorectal cancer. Inhibition of is predicted to limit sequestration of away from Axin and Axin-associated kinases and thereby to lower levels in cancer cells expressing truncated APC.
To investigate how the function of the destruction complex changes when APC is mutated, especially as a result of a typical C-terminal truncation that removes the SAMP repeats and all but the first of the 20-aa repeats, we formulated a model (as described below) for full-length APC interactions with other components of the destruction complex. We then used this model and variants corresponding to different mutated forms of APC to predict how levels and other readouts of system behavior depend on various parameters, such as the abundance of APC or truncated APC. Because APC contains multiple functional components or sites and we are interested in forms of APC containing different subsets of these sites, we formulated a model that tracks the chemical kinetics of the protein-protein interactions of interest with site-specific/structural resolution. This was accomplished by leveraging the rule-based modeling approach [37], [40], in which local rules are used to represent biomolecular interactions and their consequences. Modeling with site-specific resolution is difficult with traditional modeling approaches, such as that of ordinary differential equations (ODEs), because of combinatorial complexity [41], which arises from multisite phosphorylation, multivalent binding, and other common aspects of biomolecular interactions involved in cellular regulation. Combinatorial complexity is a motivating factor for the use of rule-based modeling here.
We developed a model for APC, Axin, and interactions and destruction complex function using the rule-based modeling framework of BioNetGen [37]–[39] (see Materials and Methods). We considered a base model, corresponding to a normal cell with full-length APC, and several variant forms of the base model. The base model is illustrated in Figs. 1 and 2. The model is annotated in Text S1 (Supporting Information). Executable BioNetGen input files are provided in the Supporting Information for the base model (Text S2) and eight variant forms of the base model (Text S3 through Text S10).
In the base model, both explicit and implicit interactions are considered. We explicitly consider the interactions of five signaling proteins (and their isoforms presumed to be functionally equivalent): APC, Axin, , , and . We implicitly consider the interactions of , PP2A, other phosphatases, and the proteins responsible for ubiquitination and proteosomal degradation of . In Fig. 2, proteins and their interactions are represented with site-specific/structural resolution using the conventions of Chylek et al. [42]. Briefly, proteins and their functional components are represented by nested boxes. Components excluded from consideration (e.g., the DIX domain of Axin) are not illustrated in Fig. 2. Arrows connecting boxes represent interactions. It should be noted that the visual elements of Fig. 2 correspond to the formal elements of our model [42]: boxes correspond to molecule types and arrows correspond to rules for interactions (Text S1). Each interaction included in the model is discussed below. The technical details of how these interactions are modeled/represented using rules are explained in Text S1. See also the Materials and Methods section.
Arrow 1 in Fig. 2 represents reversible binding of ARM repeats 5–9 to the 15-aa repeats of APC [22], [26]. In the model, ARM repeats 5–9 are considered to comprise a single binding site. Likewise, the three 15-aa repeats in APC are considered to comprise a single binding site.
Arrow 2 represents reversible binding of ARM repeats 3 and 4 to phosphorylated APC 20-aa repeats [22]. In the model, ARM repeats 3 and 4 are considered to comprise a single binding site. The seven 20-aa repeats of APC are taken to function as two distinct binding sites, with binding activity of one site considered to be mutually exclusive with binding activity of the other site. The first site (labeled 1) corresponds to the first 20-aa repeat and the second site (labeled 3) corresponds to the third 20-aa repeat. We consider binding of APC to to be mediated by the phosphorylated first repeat when the protein is APC1338 (or a comparable truncated form of APC), and predominantly (exclusively in the model as a simplification) by the phosphorylated third repeat if the protein is full-length APC. This distinction is made because APC1338 contains only the first 20-aa repeat, whereas full-length APC contains all seven 20-aa repeats. Binding of full-length APC to is mediated primarily by the phosphorylated third 20-aa repeat [25] because the phosphorylated third repeat binds with 100- to 1000-fold higher affinity than that of any of the other phosphorylated 20-aa repeats [25]. We take the stoichiometry of a -APC complex to be 1∶1.
Arrow 3 represents reversible binding of to Axin. ARM repeats 3 and 4 of bind a central region of Axin [22], [23]. As noted before, ARM repeats 3 and 4 also bind the phosphorylated 20-aa repeat region of APC (Arrow 2). Thus, ARM repeats 3 and 4 represent a binding site recognized by both APC and Axin.
Arrow 4 represents reversible binding of APC to Axin. The three SAMP repeats of APC bind the RGS domain of Axin [27], [43]. In the model, as a simplification, the SAMP repeats are considered to comprise a single binding site. Thus, we take the stoichiometry of an APC-Axin complex to be 1∶1.
Arrows 5 and 6 represent reversible binding of and to Axin, respectively. binds the GSK3 interaction domain (GID) of Axin [11], [28], [44]. binds a central region of Axin [29], which is distinct from the binding sites in Axin recognized by other binding partners. In the model, the binding of , and to Axin is taken to be non-competitive and non-cooperative.
Arrows 7 and 8 represent phosphorylation of by Axin-bound and , respectively. phosphorylation takes place in a processive manner [12], [13]. first phosphorylates Ser-45 (labeled as S45 in Fig. 1), and then phosphorylates Ser-33, Ser-37, and Thr-41. In the model, as a simplification, the latter three sites are lumped together (labeled as S33/S37 in Fig. 2). We model the phosphorylation reactions as processes with first-order kinetics that occur only when kinases and substrates are colocalized within a complex. In the model, phosphorylation at S45 occurs when is colocalized with Axin-associated . Phosphorylation at S33/S37 occurs when is phosphorylated at S45 and colocalized with Axin-associated [12], [13]. We do not consider phosphorylation of outside the context of the destruction complex.
Arrows 9 and 10 represent phosphorylation of APC 20-aa repeats by and [30], [45]. Both and are required for phosphorylation of APC [30]. In Fig. 2, is shown for illustration purposes only. In the model, we implicitly consider because it is not known which protein is responsible for colocalizing with APC. Phosphorylation of APC is taken to occur through a process with first-order kinetics when APC and are colocalized via Axin. Thus, we assume that is colocalized with APC in proportion to the extent to which is colocalized with APC via Axin. This assumption is equivalent to assuming that associates non-competitively with Axin (or directly with ).
We model dephosphorylation reactions as first-order processes (without explicit consideration of phosphatases). We allow dephosphorylation to occur if a site is exposed, i.e., not occupied and shielded by a binding partner. In the model, both phosphorylation sites of (i.e., S45 and S33/S37) are dephosphorylated according to the same rate law. In other words, the same first-order dephosphorylation rate constant is used for both sites. We allow the 20-aa repeats in APC to be dephosphorylated only if APC is in complex with Axin because Axin recruits PP2A, a phosphatase that mediates dephosphorylation of APC [46].
An important feature of the model is intracomplex binding of APC, Axin, and . In Fig. 2, Arrows 1–4 each represents two distinct types of binding reactions: intermolecular binding, and intracomplex binding. The former type of binding reaction occurs when the reacting sites are freely diffusing, i.e., not tethered. The latter type of binding reaction occurs when the reacting sites are already in a complex together, i.e., tethered and co-confined to a small subvolume of the cytoplasm. An intracomplex reaction can be marked by a high apparent affinity because of the high local concentrations of the tethered binding partners [47]. In the model, these reactions lead to complex stabilization. We account for the high local concentration effect on an intracomplex reaction by multiplying the corresponding forward rate constant by an enhancement factor . For instance, if and APC are already connected via Axin, then the effective forward rate constant for the reaction represented by Arrow 1 would be , where is the intrinsic forward rate constant when the proteins are not tethered together.
It should be noted that in the model and APC can form a binary complex held together by two-point attachment i.e., and APC can be held together through simultaneous interaction between ARM repeats 3 and 4 and APC 20-aa repeats (Arrow 1) and interaction between ARM repeats 5–9 and APC 15-aa repeats (Arrow 2). The intracomplex reactions between APC and are allowed to occur outside the context of a completely assembled destruction complex.
In the model, except for , the total concentrations of signaling proteins are taken to be conserved (i.e., constant). is produced in a process with zeroth-order kinetics and degraded in either a slow or fast process with first-order kinetics. When S33/S37 is not phosphorylated, is degraded at a slow rate, regardless of its bound state. When S33/S37 is phosphorylated, is degraded at a fast rate, again regardless of its bound state. Thus, we allow to be degraded, through a slow or fast process, independently of whether it is free or bound. We assume that releases any binding partner(s) upon degradation.
The model has 27 independent parameters, including five protein concentrations and 14 binding constants (Table 1). Parameter values were specified as described in Materials and Methods. A local sensitivity analysis indicates that model behavior is not particularly sensitive to any individual parameter value (Table S1).
Using the estimated parameter values summarized in Table 1 (see Materials and Methods), which were selected in part to allow the model to reproduce certain system behaviors (Figs. S1 and S2), we tested whether the model is able to predict the effects of transfection of SW480 cells with different truncated forms of APC. Munemitsu et al. [31] systematically transfected SW480 cells with various forms of APC. These experiments were designed to understand the effects of deletion of different functional components of APC on levels in SW480 cells, which almost exclusively express APC1338 instead of the full-length protein [31], [48].
Munemitsu et al. [31] transfected SW480 cells with full-length APC or one of 11 different truncated forms of APC (illustrated in Fig. 3). In our model, full-length APC and the 11 truncated forms of the protein can be grouped into six distinctive classes, Classes A–F (Fig. 3). The proteins in each class are functionally equivalent based on the components and interactions of APC included in the model (Fig. 2). For example, Munemitsu et al. [31] considered three forms of APC each containing the following components: 1) a partial or complete set of the 15-aa repeats, 2) all of the 20-aa repeats, and 3) the SAMP repeats of APC. These are the functional sites that we consider to be included in full-length APC (Fig. 2). Therefore, we will use APC-A to represent all three proteins, as we take these forms of APC to be functionally equivalent. Similarly, we will use APC-B to represent two other proteins, which both contain the 15-aa repeats and only the first 20-aa repeat. We take these two forms to be equivalent to APC1338, the truncated protein in SW480 cells. Henceforth, we will use APC-A, APC-B, APC-C, APC-D, APC-E and APC-F to refer to the proteins in Classes A (e.g., full-length APC), B (e.g., APC1338), C, D, E and F (Fig. 3).
Using the model, we investigated the effects of transfection of SW480 cells with APC-A through APC-F. In the model, the endogeneous concentration of APC1338 in an SW480 cell is set at 100 nM. Similarly, the endogeneous concentration of full-length APC in a normal cell is set at 100 nM (Table 1). Because APC1338 does not contain the third 20-aa repeat, nor SAMP repeats, Axin interactions associated with these sites (Fig. 2) are absent in an SW480 cell. In contrast, in a normal cell, all interactions considered in the model are active, except for the low-affinity interaction between APC and involving the phosphorylated first 20-aa repeat of APC and ARM repeats 3 and 4 of . This low-affinity interaction is omitted when considering a normal cell as a simplification (see Materials and Methods). In the model, when a representative of one of the six classes of APC is introduced into an SW480 cell, any novel interactions associated with the functional components of the transfected protein become active. For example, when APC-A is introduced, interactions associated with the third 20-aa repeat and the SAMP repeats (Fig. 2) become active. These interactions are normally missing in an SW480 cell. We assume that simulated transfections each introduce 100 nM of new protein into a cell. Thus, simulated transfection of SW480 with a particular form of APC implies that the cell contains 100 nM of a protein belonging to that form in addition to the 100 nM of the endogeneous form of APC (APC1338). (We systematically investigate how behavior depends on the amount of transfected protein below.)
In Fig. 4, we compare the model-predicted changes in levels in SW480 cells after simulated transfection of different forms of APC against the findings of Munemitsu et al. [31] (Fig. 4). The model is able to recapitulate the qualitative increase or decrease in level observed after transfection of each class of protein. Consistent with the findings of Munemitsu et al. [31], the model predicts that only transfection of APC-A and APC-E leads to a decrease in level, whereas the other four classes of APC have the opposite or no effect on level (Fig. 4) [31]. It should be noted that the results in Fig. 4 were obtained without adjustment or fitting of parameter values.
The results of Munemitsu et al. [31] suggest that exogeneous full-length APC downregulates by promoting degradation in SW480 cells. Similar results for SW480 cells have been obtained in other studies [33], [34]. However, transfection of different cell types have yielded different results [33]. Using our model, we investigated whether overexpression of APC can generally be expected to increase the rate of degradation in all cell types, or if the effect may be specific to SW480 cells only (Fig. 5).
Fig. 5A shows the model-predicted level in a normal cell as a function of APC level. A normal cell in the model is taken to have endogeneous full-length APC at a cytosolic concentration of 100 nM (Table 1). Fig. 5A illustrates the predicted effects of added APC. Fig. 5A shows that increased abundance of APC does not promote degradation, rather it has a concentration-dependent positive effect on level in normal cells, in contrast to the effect in SW480 cells (Fig. 4). The effects of exogenous full-length APC at different concentrations in SW480 cells are considered in Fig. 5B, which shows the model-predicted level in SW480 cells as a function of full-length APC level. An SW480 cell is taken to have endogeneous APC1338 at a cytosolic concentration of 100 nM (Table 1). The predicted effect of added full-length APC is a significant decrease in level in SW480 cells over a wide range of exogeneous full-length APC expression levels (Fig. 5B). This finding is consistent with the effects of transient expression of full-length APC in SW480 cells [31] and to some extent also with stable expression of full-length APC in SW480 cells [34].
In Fig. 4, we assumed a fixed amount (100 nM) of exogeneous expression for all six classes of APC. However, the results in Fig. 4 could depend on APC concentration, as seen for APC-A (Fig. 5). Therefore, we investigated the predicted concentration-dependent effects of APC-B, -C, -D and -E on levels in SW480 cells (Fig. 6). For APC-A, such effects have already been discussed (Fig. 5B). We do not consider APC-F because in our model it represents a non-functional form of APC with no binding sites.
The simulation results in Fig. 6 illustrate the concentration-dependent effects of APC-B, -C, -D and -E. As seen in Fig. 6A, added APC-B (e.g., APC1338) increases level over the entire concentration range considered. The level of doubles as the amount of exogeneous APC1338 approaches a 10-fold higher amount of endogeneous APC1338 (Fig. 6A). Unlike APC-B, the other three proteins do not increase level over the entire concentration range. APC-C reduces level at relatively high concentrations (Fig. 6B), APC-D does not alter level at any concentration (Fig. 6C), and APC-E reduces level over a range of intermediate concentrations in a manner similar to full-length APC (cf. Fig. 6D and Fig. 5B).
The only difference between APC-B and APC-C is that the former form of APC contains the first 20-aa repeat, whereas the latter form does not. This distinction leads to APC-B and APC-C having opposite effects on level in SW480 cells (cf. Figs. 6A and 6B). APC-D contains the first 20-aa repeat but no other functional components of APC that are able to interact with or Axin. Therefore, APC-D cannot interact with because of the consequent absence of phosphorylation of the 20-aa repeat. The 20-aa repeat in APC-D is never phosphorylated because the unphosphorylated protein is unable to interact with Axin. Thus, APC-D has no effect on level (Fig. 6C). APC-E entails all structural features of APC-B, but in addition it contains SAMP repeats, which mediate Axin binding (Fig. 2). Because of this distinctive feature, the model predicts that APC-E behaves differently from APC1338 and produces reduced levels at intermediate concentrations of APC-E similar to the predicted effects of full-length APC (Fig. 5B). These results indicate that the absence of SAMP repeats in APC1338 may have an important role in APC1338-mediated increases of levels in cancer cells.
The analysis of Fig. 6 indicated that APC-B (e.g., APC1338) and APC-C have opposite effects on level because APC-B contains a 20-aa repeat that APC-C does not. In Fig. 7, we analyze the effects of phosphorylation of the 20-aa repeat in APC-B on levels in SW480 cells. Recall that phosphorylation of the 20-aa repeat in APC is mediated by and [30], [45] and that phosphorylation of this site is necessary for direct interaction of APC with [22], [25].
The simulation results shown in Fig. 7A indicate that phosphorylation of the 20-aa repeat is needed for APC-B/APC1338-mediated stabilization of . In the figure, the solid line corresponds to default rate constants for phosphorylation and dephosphorylation of APC in the model (Table 1). For these parameter values, the 20-aa repeat is nearly always phosphorylated. This case can be viewed as the extreme opposite of the case where the 20-aa repeat is deleted and therefore never present in phosphorylated form. When the 20-aa repeat is deleted, APC-B becomes equivalent to APC-C and downregulates in a similar manner (cf. Fig. 7A and Fig. 6C).
Phosphorylated APC1338 binds to ARM repeats 3 and 4 in , which is also a binding site for Axin (Fig. 2). Thus, phosphorylated APC1338 competes with Axin for binding to and can inhibit phosphorylation of by sequestering away from Axin-associated kinases. Fig. 7B illustrates the predicted effect of APC1338 phosphorylation on association of and Axin. The simulation results of Fig. 7B show that phosphorylation of APC1338 inhibits interaction of with Axin.
The results of Fig. 7 suggest that competition between APC1338 and Axin for binding upregulates levels in SW480 cells. These results however do not explain how APC1338 and APC regulate differentially. Differential regulation is somewhat paradoxical because both proteins have phosphorylation sites in the 20-aa repeat region, which mediates binding. The distinction between APC and APC1338 can be attributed to the absence of SAMP repeats in APC1338, as explained fully below. In short, APC1338 sequesters away from Axin, whereas APC fails to do so (Fig. 8). The sequestration effect arises because APC1338, lacking SAMP repeats, cannot mediate indirect association of with Axin. We note that the bell-shaped curve in Fig. 8B represents a characteristic scaffold effect [49], [50]. Here, the scaffold is APC and the scaffold ligands are Axin and .
In normal cells, can associate with Axin in two ways: 1) direct binding via ARM repeats 3 and 4 in (Arrow 3; Fig. 2), and 2) indirect binding via APC, with APC acting as a linker between and Axin (Arrows 2 and 4; Fig. 2). As in an SW480 cell, the direct interaction in a normal cell is also inhibited by phosphorylation of the 20-aa repeat region in APC because of competition between phosphorylated APC and Axin for binding to ARM repeats 3 and 4 in . Nonetheless, in a normal cell, the indirect interaction still enables to colocalize with Axin via APC [30], thus allowing phosphorylation of to occur via Axin-associated kinases, which leads to degradation of . In contrast, in SW480 cells, can associate with Axin only through direct interaction. The indirect interaction does not occur because APC1338 lacks the SAMP repeats necessary for Axin binding. Thus, in SW480 cells, APC1338 phosphorylation effectively blocks association with Axin, leading to less degradation. A corollary of this finding is that increased expression of Axin would be expected to increase the degradation of , which has been observed [36], [51], [52].
The stability of the destruction complex can be perturbed (decreased) by preventing APC, Axin, and from forming a closed/cyclic ternary complex. The cyclic complex, which we assume can form in normal cells, cannot form in SW480 cells as a result of APC truncation. Formation of the cyclic ternary complex can be prevented not only by truncation of APC but also by other mutations. Any mutation affecting one of the three protein-protein interfaces of the ternary complex would prevent closure of the cyclic structure. Using our model, we simulated inhibition of formation of the cyclic structure by systematically blocking each of the three protein-protein interfaces of the closed/cyclic ternary complex, and we determined the resulting effect on level. As seen in Fig. 9, blocking the contact between APC and or and Axin did not change level, indicating that the cyclic structure is unimportant for regulation of level. Only blocking of the interface between APC and Axin (by removal of SAMP repeats) is predicted to upregulate level. However, as established above, this behavior arises for reasons other than destablilization of the cyclic ternary complex of APC, Axin, and . Thus, our model indicates that destabilization of this complex (through ablation of cyclization) is not an important effect of APC truncation.
In this study, we have modeled regulation by the destruction complex in normal and colorectal cancer cells, which express full-length and truncated APC, respectively. Our model, illustrated in Figs. 1 and 2, incorporates site-specific mechanistic details about the destruction complex, which comprises a number of signaling proteins. In colorectal cancer cells (e.g., SW480 cells), the interactions of these proteins are altered by truncation of APC. We have used our model to study the function of APC and the effects of its truncation on phosphorylation and phosphorylation-dependent degradation. We caution that our results pertain to only the function of APC within an idealized destruction complex and furthermore that we considered the interactions of particular (multifunctional) proteins in isolation from most of their binding partners. Thus, within the context of a cell, the functional effects of APC or truncated APC overexpression could potentially be very different from what our model predicts. Nevertheless, the model is qualitatively consistent with the observed effects of transient expression of various recombinant forms of APC in SW480 cells [31] (Fig. 4). Stronger, less ambiguous tests of model predictions in the future would ideally be performed using an in vitro reconstituted or cell-free system [53] to eliminate the uncertainties and complexities of the cellular milieu.
Our analyses indicate that whilst the expression of full-length APC in SW480 cells can be expected to increase degradation of , APC overexpression in normal cells may decrease degradation or have no effect. We show that phosphorylation of the first 20-aa repeat in truncated APC, together with the absence of the SAMP repeats, is crucial for the effect of APC1338 on levels in SW480 cells (Figs. 7 and 8). We suggest that phosphorylated APC1338 sequesters from Axin, thus blocking phosphorylation by Axin-bound kinases, viz. and . In contrast, phosphorylation of full-length APC, because of its SAMP repeats, which provide an indirect means for interaction between Axin and , does not block association with Axin and Axin-bound kinases, except at significantly higher levels of expression (cf. panels A and B in Fig. 8).
Several experimental studies have detected competition between phosphorylated full-length APC and Axin for binding [24], [30], [54], [55], although the effect of such competition on levels has not been previously characterized.
Our results suggest that APC1338, similar to full-length APC, can efficiently mediate competition with Axin for binding, even though it lacks the third 20-aa repeat (Fig. 3), the high-affinity binding site in full-length APC. In the model, APC1338 associates with with sufficient strength to displace Axin because of two-point attachment via its 15-aa and phosphorylated 20-aa repeat sites (i.e., because of the combined action of the interactions represented by Arrows 1 and 2 in Figs. 1 and 2). The single-site 's for the interactions mediated by these sites are and 80 nM, respectively [25], [43]. The affinities are comparable to the affinity of Axin for ARM repeats 3 and 4 ( nM [43]). However, if two-point attachment is possible, as we have postulated in our model, then there is an avidity effect. This effect has been studied in other systems [56]–[58] and may confer on phosphorylated APC1338 a competitive advantage, allowing it to outcompete Axin for .
A critically important feature of our model is a greater abundance of APC than Axin (Table 1). According to our model, as discussed above, truncated APC in SW480 cells acts as a diversion sink that sequesters away from Axin. This diversion-sink mechanism cannot be operative if Axin is more abundant than APC. Recent measurements of APC and Axin in SW480 cells indicate that the total amounts of Axin and APC are comparable [59]. At first, these results might seem to contradict the model presented here, which takes APC to be 10-fold more abundant than Axin. However, Axin is not homogeneously distributed in a cell. Much of the Axin in a cell is found in cytoplasmic puncta [35], [36]. Thus, only a fraction of total Axin may be available in a form capable of joining a destruction complex having the composition and structure considered here. A more complicated model than that presented here would be required to account for subcellular compartmentalization of APC, Axin and , which clearly play an important role in signaling [2], [3]. Such an effort is beyond the intended scope of our study.
The role of colocalization of signaling proteins within the destruction complex is not completely understood. In our model, we assumed that the core of the destruction complex, formed by mutual interactions of APC, Axin, and , has a closed/cyclic structure (as depicted in the cartoon diagram at the far left of Fig. 9). Within this cyclic structure, there are three protein-protein interfaces, and each of the three interfaces involves interaction between two adjacent proteins, which are connected indirectly via the third protein. Therefore, the binding sites at each interface are confined together in a volume that is small relative to the total volume of the cytoplasm and the local concentrations of tethered binding partners are high. Such high local concentrations can confer on a cyclic structure more stablity than a linear structure of the same composition [60]. It has been assumed that the destruction complex provides a stable platform for phosphorylation of by the Axin-recruited kinases and . However, according to our analyses, stability of the core destruction complex (i.e., the cyclic ternary complex of APC, Axin and ) is not important for efficient degradation of . By systematically simulating ablation of each possible contact between , APC, and Axin, we demonstrate that stability of the complex has little if any influence on degradation (Fig. 9). (Note that the bar at the far right of Fig. 9 is explained by the diversion-sink mechanism.) We caution that, in the model, stability of the cyclic structure is also determined by factors other than the local concentration effect. Degradation of in a core complex can terminate its cyclic structure, leaving behind a complex of APC and Axin only. In addition, phosphorylated APC can disrupt the cyclic structure by breaking the -Axin interface through competitive binding and sequestering of away from the complex.
Questions may arise as to what other roles APC plays besides destruction complex-mediated regulation of because our model indicates that elevated expression of APC in a normal cell does not have a positive effect on degradation (Fig. 5A), i.e., an increase in APC abundance is not predicted to cause a decrease in level. A variety of other potential functions of APC have been suggested. Phosphorylated APC has been implicated in subcellular localization and nuclear shuttling of [32], [61]–[63], and high-affinity binding of phosphorylated APC with has been suggested to disrupt interaction with other binding partners, such as E-cadherin and the Tcf and Lef family transcription factors [24]. It has been shown that APC competes with E-cadherin for binding to the ARM repeat region of [64]. Indeed, the main effect of stable expression of full-length APC in SW480 cells is not a reduction of level (although there is an approximate 2-fold reduction in the total amount of ), but rather a redistribution of from the nuclear and cytosolic compartments to the plasma membrane [34]. Transient expression of APC (at higher levels) causes a more dramatic reduction in the level of [31], [34]. In future work, it would be interesting to investigate how E-cadherin may regulate and vice versa [3].
Our study identifies as a potential target for therapeutic intervention in colorectal cancer. Inhibiting is expected to reverse the effect of truncation of APC in SW480 cells. According to the model, phosphorylation of APC1338 at the first 20-aa repeat plays a key role in upregulating in cancer cells (Fig. 7). Therefore, inhibition of phosphorylation of APC at this site might be an effective way to normalize levels in cancer cells. Phosphorylation of APC requires the combined action of two kinases, and [30]. Therefore, blocking of either kinase is predicted to reduce APC phosphorylation, as shown by Ha et al. [30]. Because is a common kinase for both and APC (Fig. 1) and its inhibition would stabilize , only is a potential target. We note that targeting of should be feasible in preclinical studies, as pharamacological kinase inhibitors specific to are available [65]–[67].
In this study, we used a detailed mechanistic modeling approach based on the principles of chemical kinetics to investigate regulation of phosphorylation and degradation by full-length and truncated APC. In a previous study, Lee et al. [53] developed a related model to investigate regulation of by Wnt stimulation. However, this model does not consider truncated APC. Another notable difference is that the model of Lee et al. [53] is an ordinary differential equation (ODE)-based model, wherein molecules and complexes of signaling proteins are treated as reactive chemical species, which must be enumerated along with all possible reactions to obtain an executable model. In contrast, because of the goals of our study, we developed our model using the rule-based modeling approach [37]–[40]. With this approach, local rules are used to represent protein-protein interactions, which are assumed to be modular. Assumptions of modularity can greatly reduce the complexity of a model for protein-protein interactions, and as a result, enable explicit consideration of multiple functional components within proteins (e.g., the multiple sites of phosphorylation in ). In our model, the components of the proteins considered are the basic reactive elements, i.e., we consider biochemical reactions, such as reversible binding and phosphorylation, to occur at the level of protein sites. This approach was critical for the goals of our study, which included a characterization of the effects of loss of sites in APC. Such effects, and biomolecular site dynamics in general, are difficult to capture in an ODE model [40]. The study presented here provides an example of how rule-based modeling, a fairly new approach in biology, can be used to study biomolecular site dynamics.
We focused on a part of the Wnt/ signaling pathway that controls degradation and expression level. Our primary goal was to understand the differential regulation of in normal and cancer cells at steady state and in the absence of Wnt signals, unlike in other modeling studies that have focused on the dynamics of regulation in response to a Wnt ligand [20], [53]. In future work, it would be interesting to extend our model by connecting it to other components of the Wnt/ signaling pathway and to further investigate the dynamics of regulation of .
Model parameter values are listed in Table 1. Most of the parameter values are based on previously reported estimates. However, some parameter values were set to allow the model to capture a set of observed system behaviors.
In the model, the concentration of depends in part on its rates of synthesis and degradation. As discussed below, we set parameters for these and other processes considered in the model such that the nominal, steady-state concentration of is 35 nM [53], which corresponds to 11,000 copies/cell assuming a cytoplasmic volume of L [59]. This concentration is consistent with the concentration of measured in Xenopus egg extract [53]. It is also consistent with the cytosolic (but not total) concentration of measured in HEK293T and MDCK cells, kidney epithelial cell lines, and in Caco-2 cells [59], an intestinal cell line. We take both APC and concentration to be 100 nM (31,540 copies/cell). Concentrations of APC and have been measured to be 100 nM and 50 nM, respectively, in Xenopus egg extract [53], and measured concentrations of these proteins in mammalian cells fall in the ranges of 4–34 nM and 10–120 nM, respectively [59]. We take concentration to be the same as , 100 nM. We assume Axin to be present at 10 nM (3,154 copy/cell), which is consistent with recent measurements of Axin abundance in mammalian cells; the Axin concentration measured in mammalian cells ranges from 20 to 150 nM [59]. Lee et al. [53] reported that Axin is present in Xenopus egg extract at a very low concentration, in the range of 10 to 20 pM [53]. We rejected this value, while accepting and using the qualitative observation of Lee et al. [53] that Axin is less abundant than APC, because a concentration of 20 pM corresponds to only six copies of Axin per cell for a human epithelial cell, which as stated above is taken to have a cytoplasmic volume of L [59].
For association of any two proteins that are not already in a complex together we assume the same forward rate constant () for all interactions: (Table 1). Each reverse rate constant () is determined from the relation , where is the equilibrium dissociation constant for the reaction of interest. Equilibrium dissociation constants are set at values reported earlier in the literature, as indicated below.
In the model, the region in containing ARM repeats 5–9 interacts with the 15-aa repeat region in APC (Arrow 1) with nM () [43]. The region in containing ARM repeats 3 and 4 interacts with the phosphorylated 20-aa region in APC (Arrow 2) with an affinity that depends on whether APC is full length or truncated (APC1338). Phosphorylated full-length APC binds the ARM repeats 3 and 4 with nM (), whereas phosphorylated APC1338 binds with nM () [25]. We assume that APC binds via only a single 20-aa repeat, which is consistent with binding of at one 20-aa repeat sterically hindering binding of additional copies of at other 20-aa repeats. When phosphorylated, the third 20-aa repeat mediates binding with nM, whereas the other six 20-aa repeats appear to bind with much lower affinites [25]. binds the central region of Axin via ARM repeats 3 and 4 (Arrow 3) with nM () [43]. binds the Axin GID domain (Arrow 5) with nM () [68]. We assume that APC and bind Axin (Arrow 4 and 6, respectively) with the same : nM () and nM ().
In the model, level is determined not only by synthesis and degradation rate constants, but also by other parameters that affect the phosphorylation of and APC. These parameters include phosphorylation and dephosphorylation rate constants and the enhancement factor . All of these parameters directly or indirectly determine the total level of . We selected values for these parameters, which are given in Table 1, such that the model captures a set of experimentally-observed system behaviors.
With the parameter values listed in Table 1, the model reproduces a steady-state concentration of (cytosolic) of 35 nM, which as discussed above is consistent with some measurements [53], [59]. Furthermore, the model predicts the effective half-life of to be min, and the half-life of mutated at S33/S37 to be h, which is consistent with the observations of Rubinfeld et al. [14] (Fig. S1). The model also reproduces the experimentally-determined kinetics of dephosphorylation at distinct phosphorylation sites in response to treatment with LiCl, an inhibitor of [12] (Fig. S2). The consistency of the model with observed system behaviors is further discussed in the Results section.
We note that the model is not entirely consistent with recent measurements of APC, Axin, (cytosolic) , and concentrations in SW480 cells [59]. The model cannot simultaneously reproduce these concentrations and observed system behaviors. This could be because the measured concentrations do not reflect the subcellular concentrations relevant for destruction complex function. Alternatively, our mechanistic knowledge of destruction complex function could be incomplete. We formulated our model to probe the limits of understanding of destruction complex function. In setting parameter values, we put an emphasis on selecting values that allow the model to reproduce observed system behaviors, rather than measured system parameters (viz., protein concentrations). This approach is guided by findings from computational analyses of model sloppiness, which indicate that fitting tends to improve the accuracy of model predictions more than parameter measurement [69].
We will refer to the model illustrated in Figs. 1 and 2 as the base model. The base model corresponds to the case of a normal cell with full-length APC. Variants of the base model correspond to SW480 cells with APC1338, SW480 cells transfected with different forms of APC, and other cases considered in this study (see below for further details). The base model and each of its variants was formulated using BNGL, a model-specification language [39]. Executable model-specification files are provided in the Supporting Information for the base model (Text S2) and its variants (Text S3, S4, S5, S6, S7, S8, S9, S10). BNGL is compatible with BioNetGen [39], [70], a software tool for rule-based modeling, and a number of other tools, such as RuleBender [71], an interface for BioNetGen. In addition to parameter values and initial data, each model-specification included molecule type definitions and rules. Molecule type definitions delimit the functional components of proteins and their possible phosphorylation states. The rules characterize protein-protein interactions and other processes (viz., synthesis and degradation of ). The rules that characterize protein-protein interactions can be subdivided into 10 sets, with one set for each arrow in Fig. 1 or 2. An arrow corresponds to multiple rules if the interaction represented by the rule can take place in multiple contexts and the context influences the rate of reaction. For example, Arrow 3 in Fig. 1 or 2 represents an interaction that can take place between either unconnected proteins or tethered proteins. Thus, there is a rule for each of these two cases.
The base model (Figs. 1 and 2) and each variant model was simulated by submitting the corresponding model-specficiation file to BioNetGen [39], [70] for processing. Model-specification files are provided in the Supporting Information (Text S2, S3, S4, S5, S6, S7, S8, S9, S10); the format of these files is plain text. Simulation protocols are included in each model-specification file. The captions of figures indicate which model-specification files were used in calculations. In all cases, the method used for simulation was an indirect method, meaning that the rules of the model being simulated were not used directly in the simulation procedure. Rather, the rules were expanded (i.e., used to exhaustively enumerate the distinct chemical species and individual chemical reactions implied by the rules) by invoking the generate_network function of BioNetGen to obtain a reaction network. The corresponding system of ODEs describing the mass-action kinetics of this network were then numerically integrated by invoking the simulate_ode function of BioNetGen. BioNetGen uses CVODE [72], [73] for numerical integration of ODEs. For the base model, the reaction network obtained by expansion of its rules comprises 410 distinct chemical species. The size of this network does not reflect the intrinsic complexity of the base model. Rather, the intrinsic complexity of this model is reflected by the number of its rules. The base model includes 29 rules. We used scripts to systematically vary default parameter values specified in BioNetGen input files to produce many of the figures shown in the Results section. Parameter scans are enabled by a function available within RuleBender [71].
|
10.1371/journal.pntd.0007299 | Midgut barriers prevent the replication and dissemination of the yellow fever vaccine in Aedes aegypti | To be transmitted to vertebrate hosts via the saliva of their vectors, arthropod-borne viruses have to cross several barriers in the mosquito body, including the midgut infection and escape barriers. Yellow fever virus (YFV) belongs to the genus Flavivirus, which includes human viruses transmitted by Aedes mosquitoes, such as dengue and Zika viruses. The live-attenuated YFV-17D vaccine has been used safely and efficiently on a large scale since the end of World War II. Early studies have shown, using viral titration from salivary glands of infected mosquitoes, that YFV-17D can infect Aedes aegypti midgut, but does not disseminate to other tissues.
Here, we re-visited this issue using a panel of techniques, such as RT-qPCR, Western blot, immunofluorescence and titration assays. We showed that YFV-17D replication was not efficient in Aedes aegypti midgut, as compared to the clinical isolate YFV-Dakar. Viruses that replicated in the midgut failed to disseminate to secondary organs. When injected into the thorax of mosquitoes, viruses succeeded in replicating into midgut-associated tissues, suggesting that, during natural infection, the block for YFV-17D replication occurs at the basal membrane of the midgut.
The two barriers associated with Ae. aegypti midgut prevent YFV-17D replication. Our study contributes to our basic understanding of vector–pathogen interactions and may also aid in the development of non-transmissible live virus vaccines.
| Most flaviviruses, including yellow fever virus (YFV), are transmitted between hosts by mosquito bites. The yellow fever vaccine (YFV-17D) is one of the safest and most effective live virus vaccine ever developed. It is also used as a platform for engineering vaccines against other health-threatening flaviviruses, such as Japanese encephalitis, West Nile, dengue and Zika viruses. We studied here the replication and dissemination of YFV-17D in mosquitoes. Our data showing that YFV-17D is unable to disseminate to secondary organs, as compared to a YFV clinical isolate, agree with previous studies. We have expanded on this knowledge by quantifying viral RNA production, viral protein expression, viral distribution and infectivity of YFV-17D in the vector midguts. We show that the midgut is a powerful barrier that inhibits YFV-17D dissemination in mosquitoes. Our study contributes to our basic understanding of the interactions between viruses and their vectors, which is key for conceiving new approaches in inhibiting virus transmission and designing non-transmissible live virus vaccines.
| Arboviruses, which are transmitted among vertebrate hosts by blood-feeding arthropod vectors, put billions of people at risk worldwide. Viral infection in arthropods is usually persistent. Following uptake of an infectious blood meal by a female mosquito, arbovirus must initiate a productive infection of the midgut epithelium, which consists of a single layer of cells [1]. To develop a disseminated infection, virus must then escape the midgut into the haemocoel and infect secondary tissues such as the fat body, trachea and the salivary glands [1]. Finaly, the virus needs to be released into salivary ducts for horizontal transmission to an uninfected vertebrate host [1]. Traditional means of controlling the spread of arbovirus infection include mosquito control and vaccination of susceptible vertebrates. However, in many cases, these measures are either unavailable or ineffective. To successfully implement the strategy of blocking the virus at the arthropod stage, further knowledge of the virus/vector interactions is required.
Flaviviruses constitute the most important and diverse group of arthropod-transmitted viruses causing diseases in humans. They are 50 nm-diameter enveloped viruses harboring a single positive-strand RNA genome of around 11 kb. The genome encodes a polyprotein that is cleaved into seven non-structural (NS) proteins (NS1, NS2A, NS2B, NS3, NS4A, NS4B, and NS5) and three structural proteins: capsid (C), pre-membrane/membrane (prM/M) and envelope (Env). The C, M, and Env proteins are incorporated into virions, while NS proteins are not [2,3]. NS proteins coordinate RNA replication, viral assembly and modulate innate immune responses.
Several members of the Flavivirus genus, such as dengue virus (DENV), yellow fever virus (YFV) and Zika virus (ZIKV) are highly pathogenic to humans and constitute major global health problems. YFV is responsible for viral hemorrhagic fever resulting in up to 50% fatality [4]. Despite the existence of the safe and effective live-attenuated vaccine YFV-17D, YFV regularly resurges in the African and South American continents, as illustrated by recent outbreaks in Brazil and equatorial Africa [5–7]. The YFV-17D vaccine has been used safely and efficiently on a large scale since the end of World War II [8]. It was developed in the 1930’s by passaging the blood of a human patient in rhesus macaques and later in mouse and chicken embryo tissues [9]. A single dose confers protective immunity for up to 35 years. During the attenuation process, the virus has lost its neurotropic and viscerotropic properties, which account for the major disease manifestations of yellow fever in primates [10,11]. The molecular determinants responsible for its virulence attenuation and immunogenicity are poorly understood. We have recently shown that YFV-17D binds and enters mammalian cells more efficiently than a non-attenuated strain, resulting in a higher uptake of viral RNA into the cytoplasm and consequently a greater cytokine-mediated antiviral response [12]. This differential entry process may contribute to attenuation in humans.
YFV-17D is also used as a platform for engineering vaccines against other health-threatening flaviviruses, such as vaccines against Japanese encephalitis virus (JEV), West Nile virus (WNV), the four serotypes of DENV, and, more recently, ZIKV [13–16]. These vaccines consist in a YFV-17D backbone in which sequences coding for prM/E proteins are replaced by those of the selected flavivirus. Some of these live-attenuated chimeric vaccines are commercially available [17,18], with variable success [19]. YFV-17D is thus a key component in controlling flaviviral disease and it must not disseminate in mosquitoes. Early studies have shown, using almost exclusively viral titration by plaque assays, that YFV-17D can infect Aedes aegypti midgut [20,21], but does not disseminate to other tissues and fails to be transmitted to a novel host. Here, we re-visited this question using a variety of techniques and showed that not only the midgut escape barrier, but also the midgut infection barrier, restrict YFV-17D replication in its vector.
The YFV-17D vaccine strain (YF-17D-204 STAMARIL, Sanofi Pasteur, Lyon) was provided by the Institut Pasteur Medical Center. The YFV-DAK strain (YFV-Dakar HD 1279) was provided by the World Reference Center for Emerging Viruses and Arboviruses (WRCEVA), through the University of Texas Medical Branch at Galveston, USA. Viral stocks were prepared on Vero cells, concentrated by polyethylene glycol 6000 (Sigma) precipitation and titrated on Vero cells by plaque assay as described previously [22].
The Aag2 mosquito cell lines (provided by the teams of M. Flamand and L. Lambrechts, Institut Pasteur, Paris) are derived from larvae of Aedes aegypti. They were cultured in a humid chamber at 28°C, with no CO2, in Leibovitz medium (Gibco Leibovitz's L-15 Medium, Life Technologies) supplemented with 10% fetal bovine serum (FBS), 2% tryptose phosphate buffer (Gibco Tryptose Phosphate Broth 1X, Life Technologies), 1/100 dilution of the penicillin-streptomycin (P/S) stock (final concentration of 100 units/mL and 100 μg/mL, respectively) (Sigma) and non-essential amino acids (GibcoTM NEAA 100X MEM, Life Technologies). Vero cells, which are African green monkey kidney epithelial cells, were purchased from the American Type Culture Collection (ATCC) and used to perform viral titration. They were maintained in Dulbecco's modified Eagle's medium (DMEM, Invitrogen), supplemented with 10% FBS and 1% P/S.
Env MAb 4G2 hybridoma cells were kindly provided from P. Desprès (La Réunion University, Sainte Clotilde). Anti-YFV-NS4B and anti-DENV NS1 17A12 (that recognize YFV-NS1) antibodies, were kind gifts from C.M. Rice (Rockefeller University, NY) [23] and M. Flamand (Institut Pasteur, Paris) [24], respectively. Anti-actin (A1978, Sigma) and anti-tubulin (T5168, Sigma) antibodies were used as loading controls for mosquito organs and Aag2 cells, respectively. Secondary antibodies were as followed: anti-mouse 680 (LI-COR Bioscience), anti-rabbit 800 (Thermo Fisher Scientific) and anti-rabbit Cy3 (Life Technologies).
The Paea strain of Ae. aegypti is a laboratory colony originated from mosquitoes collected in French Polynesia in 1960 and conserved in the laboratory since 400–450 generations. Adult mosquitoes were maintained at 25 ± 1°C and 80% relative humidity with a light/dark ratio of 12 h/12 h. The larvae were provided with brewer’s yeast tablets and adults were given continuous access to 10% sucrose solution. Sucrose was removed 24 h prior to the infectious blood meal. The infectious blood meal was comprised of half-human blood and half-viral suspension (4.107 PFU/mL in the mix). The blood donors were randomly selected from a population of healthy volunteers donating blood at the ‘Etablissement Français du Sang’ (EFS), within the framework of an agreement with Institut Pasteur. Experimental procedures with human blood have been approved by EFS Ethical Committees for human research. All samples were collected in accordance with EU standards and national laws. Informed consent was obtained from all donors. Seven day-old female mosquitoes were allowed to feed for 15 min through a collagen membrane covering electric feeders maintained at 37°C (Hemotek system). Blood-fed females were selected and transferred into cardboard boxes protected with mosquito nets. Alternatively, ice-chilled mosquitoes were injected intrathoracically with twice 69 nL of viral stock (2.5x104 PFU) with a micro-injector (Drummond, Nanoject II). Mosquitoes were anesthetized on ice at various time-points after infection. They were passed through a 70% ethanol bath and then in a PBS bath before being dissected in a drop of PBS under a magnifying glass using tweezers. The midguts, legs and salivary glands were removed and placed in individual tubes containing sterilized glass beads of a diameter of 0.5 mm (Dutscher) in a suitable lysis buffer. Experiments were reproduced in triplicate with 5–10 mosquitoes collected at each time-point for dissection.
The mosquito midguts, legs or salivary glands were crushed using a tissue homogenizer (Ozyme, Precellys Evolution) during twice 15 s at 1000 g. Total RNA was extracted from mosquito tissues with the NucleoSpin RNA II kit (Macherey-Nagel). YFV RNA was quantified using NS3-specific primers and TaqMan probe (NS3-For CACGGCATGGTTCCTTCCA; NS3-MFAM CAGAGCTGCAAATGTC; NS3-Rev ACTCTTTCCAGCCTTACGCAAA) with TaqMan RNA-to-CT 1-Step (Thermo Fisher Scientific) on a QuantStudio 6 Flex machine (Applied Biosystems). Genome equivalent (GE) concentrations were determined by extrapolation from a standard curve generated from serial dilutions of total YFV RNA of a known concentration.
Individual midguts and salivary glands were collected in RIPA buffer (Sigma) containing protease inhibitors (Roche Applied Science). Tissue lysates were normalized for protein content with Pierce 660nm Protein Assay (Thermo Scientific), boiled in NuPAGE LDS sample buffer (Thermo Fisher Scientific) in non-reducing conditions and 32 μg (midgut) or 14 μg (salivary glands) of proteins (corresponding to around 10 pooled organs) were separated by SDS-PAGE (NuPAGE 4–12% Bis-Tris Gel, Life Technologies). Separated proteins were transferred to a nitrocellulose membrane (Bio-Rad). After blocking with PBS-Tween-20 0.1% (PBST) containing 5% milk for 1 h at RT, the membrane was incubated overnight at 4°C with primary antibodies diluted in blocking buffer. Finally, the membranes were incubated for 1 h at RT with secondary antibodies diluted in blocking buffer, washed, and scanned using an Odyssey CLx infrared imaging system (LI-COR Bioscience).
After dissection, individual midgut were deposited on slides, fixed in cold acetone for 15 min and rehydrated in PBS for 15 min. The midguts were then incubated for 2 h in Triton X-100 (0.2%). After washing with PBS, they were incubated for 30 min with PBS + 0.1% Tween 20 + 1% BSA. The slides were then incubated overnight at 4°C with anti-YFV-NS4B antibodies diluted 1:1000 in PBS. After washing with PBS, they were incubated for 1 h with secondary antibodies and washed with PBS. The actin network was visualized with phalloidin Alexafluor 488 (Invitrogen). After washing, nuclei were stained using Prolong gold antifade containing 4′,6-diamidino-2-phenylindole (DAPI) (Invitrogen). All preparations were observed with a confocal microscope (ZEISS LSM 700 inverted) and images were acquired with the ZEN software.
Viral RNA was extracted from viral stocks (around 1.4x109 genomes for YFV-17D and 3.4x108 for YFV-DAK) using Trizol (Ambion, TRIzol Reagent), were re-suspended in RNAse-free water and treated with DNAse with the DNA-free kit (Ambion) before being stored at -80°C. Synthesis of cDNAs was carried out with the Maxima H Minus First Strand kit (Thermo Fisher Scientific) from 250 ng of viral RNA. Three fragments of the viral genome were amplified by 25 rounds of PCR using the Phusion High-Fidelity DNA Polymerase kit (NEB) using primers mainly described previously [25]. New primers targeting the 3'-UTR of the genome were designed for optimal amplification of YFV-17D and YFV-DAK (Table 1). The PCR products were purified with the NucleoSpin Gel kit and PCR Clean Up (Macherey-Nagel), resuspended in RNAse-free water and stored at -20°C. Libraries were prepared after pooling 400 ng of the three overlapping amplicons, which had a size between 3725 and 3891 pb. The PCR products were fragmented randomly with the NEBNext dsDNA fragmentase kit (NEB) and then purified with the AMPure XP Beads kit (Beckman Coulter, Inc.). The Illumina sequencing libraries were prepared with the NEBNext Ultra DNA Library Prep kit (NEB) by selecting 400 bp fragments. NEBNext Multiplex Oligos for Illumina primers (NEB) were used. Purification was performed with the AMPure XP Beads kit. The Qubit dsDNA BR Assay kit (Thermo Fisher Scientific) was used for quantification. Samples from the library, diluted to 4 nM, were sequenced on a NextSeq 500 sequencer (Illumina) machine with the NextSeq 500 Mid Output Kit v2 kit (150 cycles) (Illumina), to generate single-end reads of 150 nt. The PhiX control library served as a quality and calibration control in sequencing runs (Illumina, FC-110-3001).
Reads were trimmed for adapters and primer sequences. Low quality reads were filtered using Trim Galore! (www.bioinformatics.babraham.ac.uk/projects/trim_galore/) with the following parameters: quality 30, length 75 and stringency 4. Final reads quality was evaluated using FastQC (www.bioinformatics.babraham.ac.uk/projects/fastqc/). Reads were aligned on the YFV-Asibi reference genome AY640589.1 using BWA [26] and SAMtools [27]. Consensus sequences were obtained using SAMtools mpileup, VarScan mpileup2cns (min-var-freq 0.5) and BCFtools consensus [27]. When mapping YFV-DAK sequencing reads on the YFV-Asibi sequence, an uncomplete coverage was observed. We reiterated the mapping of YFV-DAK reads on this intermediate consensus sequence to obtain the final YFV-DAK consensus sequence. Consensus sequence of YFV-17D and YFV-DAK have been deposited on NCBI (ID numbers MN10624 and MN106242), and sequencing data have been uploaded on SRA (SRA accession number PRJNA548475). Variant determination was estimated using VarScan mpileup2snp (min-var-freq 0.01, strand-filter 0) with a cutoff of 3%.
Data were analyzed using GraphPad Prism 7. Statistical analyses were performed using two-tailed Fisher's exact test or Mann-Whitney test (* p < 0.05; ** p < 0.01; *** p < 0.001; **** p < 0.0001, ns, not significant.), as indicated.
The replication and dissemination of YFV-17D was studied in the Ae. aegypti strain Paea. The clinical isolate YFV-Dakar HD 1279 (YFV-DAK), whose replication in rhesus macaque is well characterised [28], was used as a positive control for these experiments. Virus produced on Vero cells were mixed with human blood to prepare a meal containing 4x107 PFU/mL of either YFV-17D or YFV-DAK. Five to ten mosquitoes were collected every 2–3 days until 14 days post-feeding (dpf). Mosquitoes were dissected to separate the midgut from legs and salivary glands. Virus production in these tissues was first assayed by calculating the viral titer by plaque assays on Vero cells. Several whole mosquitoes were also analyzed 20 minutes after feeding to ensure that the mosquitoes ingested a similar amount of viral particles of both viral strains (Fig 1A and 1B, black dots). Around 103 infectious particles of YFV-DAK were detected per midguts 3 dpf (Fig 1A). Viral titers remained high in midguts until 14 dpf. YFV-DAK infectious particles were present in legs as early as 5 dpf and in salivary glands as early as 7 dpf (Fig 1A). This replication pattern is comparable to that of South American and African YFV isolates in the strain Ae. aegypti AE-GOI [29]. Midgut of mosquitoes infection with YFV-17D produced 1 to 2 log less infectious particles than YFV-DAK at 3 dpf (Fig 1B). Infectious particles were detected in a unique leg sample at 14 dpf. No virus was detected in salivary glands of mosquitoes infected with YFV-17D. Thus, by contrast to YFV-DAK, and in agreement with previous studies performed with the Ae. aegypti strains Rexville or Rexville-D (Rex-D) [21,30–32], YFV-17D disseminated poorly in the strain Paea.
Viral replication was assessed in the midguts, legs and salivary glands by measuring viral RNA quantity over-time by RT-qPCR (Figs 1C, 1D and S1). Total RNA was also extracted from several whole mosquitoes the same day of the feeding to insure that they had ingested similar amount of infectious particles from both viral strains (Fig 1C and 1D, black dots). Around 107 copies of viral RNA were detected in midguts of mosquitoes infected with YFV-DAK since 3 days (Figs 1C, S1A and S1C). The viral RNA copy number per midgut remained high until 14 dpf, indicating that viral replication had already reached a plateau at an early stage of infection (Figs 1C, S1A and S1C). In agreement with titration assays (Fig 1A), YFV-DAK RNA was detected in legs and salivary glands of mosquitoes around 7 dpf. The quantity of viral RNA detected in these secondary organs increased over time to reach on average 107 copies RNA in legs and 106 copies in salivary glands at 14 dpf (Figs 1C, S1A and S1C). Around 5x105 copies of YFV-17D RNA was detected in 2 out of 5 midguts of blood-feed mosquitoes at 3 dpf (Fig 1D). At 12 dpf, around 107 copies of YFV-17D RNA was detected in 4 out of 8 mosquitoes, which is 10 time less than in YFV-DAK infected moquitoes. YFV-17D RNA was found in legs of 2 mosquitoes among the 46 blood-fed mosquitoes collected during 14 days (Figs 1D, S1B and S1D). No virus was dectected in the salivary glands of these 46 mosquitoes (Figs 1D, S1B and S1D). Therefore, in agreement with our titration assays (Fig 1B) and with previous studies performed with Rexville strains of Ae. aegypti [30–32], YFV-17D disseminated poorly in the strain Paea. The RT-qPCR analyses also revealed that the vaccine strain replicated less efficiently than YFV-DAK in its vector.
The percentage of mosquitoes that were positive for viral RNA among the mosquitoes that had taken blood was calculated based on RT-qPCR data obtained from 3 independent experiments (Fig 1E). Significantly less midguts were postivive for YFV-17D RNA than YFV-DAK RNA at days 7 and 14 post-feeding, suggesting that the midgut infection barrier restricts the replication of the vaccine strain. Viral dissemination to legs was defined by the presence of viral RNA in the legs of mosquitoes whose midguts were infected (Fig 1F). YFV-DAK had disseminated in around 40% of infected mosquitoes at 7 dpf and in around 90% of mosquitoes at 14 dpf. At this time, YFV-17D had disseminated in around 10% of them (Fig 1F). YFV-DAK dissemination rates are consistent with the ones reported for the YFV-Asibi strain [32] or clinical isolates from Peru [31] in Rexville mosquitoes. YFV-DAK RNA was detected in salivary glands of approximately 75% of mosquitoes whose midguts were infected, revealing that the virus had efficiently reach these secondary organs (Fig 1G). By constrast, no dissemination in salivary glands was observed in mosquitoes infected with YFV-17D.
To investigate the replication ability of the two viral strains further, the presence of viral antigens in pooled midguts and salivary glands of mosquitoes fed on blood containing 4.107 PFU/mL of either YFV-17D or YFV-DAK was analyzed by Western blots at days 7 and 14 post-feeding using antibodies against Env and NS1. The Env protein was detected at both time-points in the midgut of mosquitoes infected with the strain YFV-DAK, in a majority form of around 45 kDa and a minor form of around 35 kDa (Fig 2A). The Env protein was not detected in the salivary glands 7 days after the blood meal but was present as a 45 kDa form 14 days after the blood meal (Fig 2A). These data are in good agreement with the titration and RT-qPCR data presented in Figs 1 and S1. Like the Env protein, the NS1 protein was detected in the midguts of mosquitoes infected with the YFV-DAK strain at both 7 and 14 dpf (Fig 2A). In midguts, NS1 was detected at the expected size of 45 kDa, but also as heavier forms of around 80 kDa. These forms could represent NS1-2A, a polyprotein precursor consisting of NS1 and a portion of NS2A. This NS1-2A form was previously reported in human SW-13 cells infected with YFV-17D [23] and maybe generated by alternative cleavage sites in the NS2A region upstream from the cleavage site generating the N-terminus of NS2B. Alternatively, they could represent glycosylated versions of NS1 monomer or dimer. NS1 was also detected in the salivary glands of mosquitoes infected with YFV-DAK for 14 days (Fig 2A). No or very little signal was detected by the anti-NS1 or anti-Env antibodies in organs of mosquitoes infected with YFV-17D (Fig 2A). In order to ensure that the antibodies directed against the NS1 and Env proteins recognize YFV-17D proteins, control experiments were performed with the Ae. aegypti Aag2 cells infected for 24 or 48 hours at an MOI of 0.1 with both viral strains. Both proteins were well detected in cells infected for 48 hrs, independently of the viral strain used (Fig 2B). Thus, absence of detection of YFV-17D Env and NS1 proteins in the mosquito organs at 7 and 14 dpf is not due to poor recognition of the viral antigens, nor the antibodies used, but reflects a low-level replication. These data confirm our titration and RT-qPCR analyses (Fig 1). Of note, the YFV-DAK Env was detected as 2 forms in Aag2 cells infected for 48 hours while the YFV-17D Env was detected as a unique form. No YFV-17D proteins were detected at 24 hours post-infection, suggesting that the replication of the vaccine strain is slower in Aag2 cells than the one of YFV-DAK, as in Ae. aegypti (Fig 1).
Finally, to confirm RT-qPCR and Western blot data, immunofluorescence analyses were performed on midgut of mosquitoes fed since 7 days using antibodies againt the viral protein NS4B. YFV-DAK antigens were evenly distributed in foci over the entire epithelium at this time (Fig 3A). By contrast, YFV-17D antigens were found in one or two localized foci in infected midguts. In an attempt to investigate further this uneven distribution of YFV-17D replication sites, the midgut of mosquitoes infected with both viral strains for 3 or 7 days were cut longitudinally into two equal parts. The presence of viral RNA was determined by RT-qPCR analyses performed on individual half midguts (Fig 3B). Among 18 mosquitoes that ingested blood containing YFV-17D, 3 half midguts were positive for YFV-17D RNA at day 3 post-infection and only 2 at day 7 post-infection. This is in agreement for our previous results (Fig 1E). Among these five positive midguts, only one contained YFV-17D RNA in both sections (Fig 3B). As expected based on previous results (Fig 1E), it was easier to obtain midguts positive for YFV-DAK. Twelve out of the 15 midguts that were positive for YFV-DAK RNA contained viral RNA in both sections. These experiments revealed that YFV-17D replication in Ae. aegypti midgut is more confined than YFV-DAK replication.
Together, these data show that, by contrast the clinical isolate YFV-DAK, the vaccine strain replicated poorly in, and disseminated poorly from Ae. aegypti midgut.
To assess whether YFV-17D could infect Ae. aegypti when delivered via a non-oral route, mosquitoes were inoculated intra-thoracically with 2.5x104 PFU of YFV-17D or YFV-DAK, which corresponds to around 10 times less PFU than when mosquitoes are taking around 5 μL of a blood meal containing 4x107 PFU/mL. The presence of viral RNA was analyzed by RT-qPCR 10 days after injection. Mosquitoes infected via a blood meal served as controls. Several whole mosquitoes were also analyzed 20 minutes after feeding or injection to ensure that a similar amount of viral particles of both viral strains were delivered in mosquitoes (Fig 4, black boxes). In good agreement with our previous experiments (Fig 1E), around 35% of midguts (8 out 22) were positive for YFV-17D RNA, whereas 81% (18 out 22) were positive for YFV-DAK RNA at day 10 post feeding (Fig 4A). Moreover, significantly less viral RNA (around 10 times) was found in YFV-17D-infected midguts as compared to YFV-DAK-infected midguts (Fig 4A). YFV-DAK RNA was detected in legs and salivary glands of around 50% of these mosquitoes. By contrast, YFV-17D was detected in the legs of a unique mosquito out of 22 and was not detected in salivary glands (Fig 4A), confirming the inability of the vaccine strain to spread to secondary organs when orally delivered. When the midgut barriers were bypassed by injecting Ae. aegypti mosquitoes in the thorax, 100% of midguts were positive for both viral strains and similar amounts of YFV-17D and YFV-DAK RNA were detected in this organ, indicating that both viral strains successfully replicated in midgut-associated tissues when bypassing the lumen (Fig 4B). All legs and salivary glands were positive for YFV-17D and YFV-DAK RNA (Fig 4B), revealing that the two viruses were efficiently infecting secondary tissues once the midgut was bypassed. Of note, significantly more (around 10 times) YFV-DAK RNA was detected in salivary glands than YFV-17D RNA (Fig 4B), suggesting that YFV-17D is sensitive to the salivary gland infection barrier.
To ensure that viral RNA detected in secondary organs of injected mosquitoes represented replicative RNA and not input viral RNA, UV-treated viral RNA was also injected into the thorax of several mosquitoes. A signal, slightly above the detection threshold, was detected in two organs out of 39 tested (Fig 4B). These data confirm the ability of YFV-17D to replicate as efficiently as YFV-DAK in midgut and secondary organs when mosquitoes were inoculated intra-thoracically.
To determine the consensus sequence of the two viral strains, we performed next generation sequencing (NGS) analysis of the two viral stocks. Average coverage depths for these alignments were around 1000x (S1 Table) and homogeneous along their references. The comparison of the two consensus sequences identified 333 synonymous mutations (Fig 5A and 5B, blue bars) and 60 non-synonymous ones (Fig 5B, red bars). These differences were scattered along the genome. Single nucleotide variants (SNVs) and their frequency were identified all along the two genomes (Fig 5C). Only the SNVs representing a minimum of 3% of all observations were considered. The genome of YFV-17D contained more SNVs than the one of YFV-DAK (50 against 18). A SNV that lies in the NS2A gene of YFV-17D is represented in 44% of the population, but does not induce amino acid change.
Studies conducted shortly after the development of YFV-17D showed that Ae. aegypti fed on vaccinated volunteers or rhesus monkeys were unable to transmit YFV-17D to susceptible monkeys [33]. These results were confirmed five decades later by showing that suckling mice bitten by Ae. aegypti infected with YFV-17D did not exhibit sign of disease [31]. Poor dissemination of YFV-17D to mosquito heads was shown by examining head tissues by immunofluorescence or immunohistochemical studies [31,34]. Consistently, titration assays performed on organs of the Rex-D strain of Ae. aegypti revealed that YFV-17D infects the midgut, but does not spread to secondary organs [21,32]. Our RT-qPCR, immunofluorescence and titration analyses document the inability of YFV-17D to disseminate in the Paea strain of Ae. aegypti. Our analysis also revealed that YFV-17D replicates poorly in the midgut, as compared to the clinical isolate YFV-DAK. Among the mosquito with midgut positive for YFV-17D RNA, only 10% had viruses that disseminated to their legs and none had viral RNA in their salivary glands. Thus, our data suggest that the YFV-17D strain is not only sensitive to the midgut escape barrier, but also to the midgut infection barrier when orally delivered. When injected into the thorax of mosquitoes, YFV-17D replicated in midgut tissues as efficiently as YFV-DAK. These data suggest that the restriction of YFV-17D replication in the midgut occur at the level of epithelial cells. Our RT-qPCR analyses suggest that the major restriction occurs at a stage prior to viral RNA production. Several mechanisms, not mutually exclusive, could explain this restriction.
First, the restriction could occur during viral entry in midgut epithelial cells. The low number of loci revealed by immunofluorescence analysis of YFV-17D-infected midguts suggests that only few cells were initially infected by the vaccine strain and thus supports the hypothesis of an entry defect. Flavivirus entry mechanisms are poorly described in mosquito cells. Neither attachment factor(s) nor entry receptor(s) are identified yet. As in mammalian cells, the domain III of Env is involved in attachment and entry of flavivirus in mosquito cells [35]. Thus, it is conceivable that YFV-17D Env would have a lower affinity for cell entry factors than YFV-DAK Env. Our NGS analysis revealed that the consensus sequence of the two Env proteins differs from 75 mutations, including 14 non-synonymous mutations. Seven of these non-synonymous mutations lie within the domain III. Finally, in Aag2 cells infected for 48 hours, we detected two forms of YFV-DAK Env under non-reducing conditions and a single form of YFV-17D Env. These differences may reflect a different conformation and may explain a different affinity for a cell entry receptor. In agreement with this hypothesis, when domain III of the Env gene of a YFV able to disseminate was replaced by domain III of the Env gene of YFV-17D, the dissemination of the chimeric virus was strongly inhibited, suggesting an important role in Domain III in this process [21]. These results, however, may be the consequence of chimerization, as it is known that flavivirus chimeras replicate less efficiently than parental viruses [36]. We have recently shown that the YFV-Asibi enters a panel of human cells by canonical endocytosis mechanisms involving clathrin, while YFV-17D enters cells in a clathrin-independent manner [12]. We have shown that the 12 mutations differentiating YFV-Asibi Env from YFV-17D Env are responsible for the differential internalization process. Based on these data, we hypothesized that YFV-17D and YFV-Asibi use different cell receptors [12]. It is therefore possible that the YFV-17D and YFV-DAK strains also use different receptors in mosquito cells and that the receptor used by YFV-17D is poorly expressed at the apical surface of midgut epithelial cells, as compared to the one used by the clinical strain. This hypothesis is consistent with our data showing that YFV-17D succeeded in replicating into midgut-associated tissues when inoculated intra-thoracically. Alternatively, the glycosylation status of the Env protein could play a role in the differential entry abilities of the two strains. Flavivirus Env proteins possess a conserved N-glycosylation motif at amino acid 153/154. This modification is involved in important viral replication and pathogenesis functions [37]. Mutagenesis studies on many flaviviruses, including the DENV, WNV and ZIKV, indicate that the loss of this N153/154-glycosylation impairs viral replication in the midgut [38–40]. Unlike most flaviviruses, the YFV Env lacks the N153/154-glycosylation canonical site. A second non-canonical N-glycosylation site exists at position 470. However, it is unlikely that this site is functional because it is located in the hydrophobic carboxy-terminal domain and is therefore inserted into the endoplasmic reticulum membrane. The absence of an accessible N-glycosylation site in YFV-DAK Env therefore indicates that such motif is probably not necessary for replication and dissemination in mosquitoes. We therefore believe that mutations in Env, rather than its glycosylation status, are involved in vector competence.
Another mechanism that could explain the low replication of YFV-17D in the midgut of mosquitoes is its inability to escape the antiviral mechanisms in midgut epithelial cells. The RNA interference pathway (RNAi), is a major antiviral defense initiated by the recognition of viral replication intermediates by the Dicer-2 protein [41]. Its efficacy differs within organs, both in Anopheles gambiae [42] and Aedes aegypti [43]. One can envisage that the pathway is particularly efficient against YFV-17D in midgut epithelial cells. This pathway inhibits the replication of DENV and ZIKV viruses in the midgut and salivary glands of mosquitoes [44–46]. Interestingly, Myles and colleagues recently showed that the YFV C protein counteracts the RNA interference pathway in Ae. aegypti by protecting double-stranded viral RNA from Dicer-2-induced cleavage [47]. No amino acid sequence responsible for this effect has been identified. Our NGS analysis revealed that the consensus sequence of the C gene of our two strains of interest differ by 10 mutations, including a non-synonymous one. This unique mutation in C could modulate its RNA interference suppression activity.
The NS1 protein, which is a highly conserved glycoprotein secreted by flavivirus-infected cells, enhances DENV and JEV replication in their vectors [48]. It does so by allowing them to escape two important antiviral mechanisms: the production of reactive species of oxygen (ROS) and the JAK/STAT pathway [48]. One can envisage that, like the NS1 proteins of DENV and JEV, YFV-DAK NS1 protein could be a potent suppressor of these two antiviral strategies. The NS1 protein of YFV-DAK could also be more expressed and/or secreted than the one of YFV-17D.
Our NGS analysis detected more than 60 non-synonymous nucleotide differences along the genome of the two viral strains. These, together with the 8 nucleotide differences in the 3' untranslated region (UTR) between the 2 viral strains, could have functional consequences. The higher abundance of variants in YFV-17D genome as compared to YFV-DAK genome was unexpected since a recent study showed that the vaccine strains YFV-17D and YFV-FNV contained fewer variants than their respective parental strains [49]. However, our observations do not inform on the general variability of the genomes since they concern only a small number of nucleotides. Further analysis would be warranted to compare the genetic diversity of YFV-17D and YFV-DAK. Deep sequencing analysis of YFV-17D genome coupled to independent diversity measurements, such as the Simpson 1-D and Shannon entropy indexes, revealed that the vaccine strain lacks quasispecies diversity as compared to its parental strain Asibi [25]. This loss of genetic diversity has been proposed to contribute to YFV-17D attenuation in vaccinated patients [25]. Moreover, recent studies with Venezuelan equine encephalitis virus (VEEV), which belongs to the genus Alphavirus, have revealed that viruses able to disseminate in mosquitoes have an higher diversity than the ones that did not disseminate [50]. Thus, the poor genetic diversity of YFV-17D may contribute to its inability to infect and spread in Ae. aegypti.
Additional studies will be needed to identify the molecular mechanism(s) responsible for the low replication and dissemination of the YFV-17D vaccine strain in Aedes mosquito. These studies are essential to better understand the interactions between viruses and their vectors and can also contribute to the development of non-transmissible live-attenuated vaccines.
|
10.1371/journal.pntd.0004009 | Improved Quantification, Propagation, Purification and Storage of the Obligate Intracellular Human Pathogen Orientia tsutsugamushi | Scrub typhus is a leading cause of serious febrile illness in rural Southeast Asia. The causative agent, Orientia tsutsugamushi, is an obligate intracellular bacterium that is transmitted to humans by the bite of a Leptotrombidium mite. Research into the basic mechanisms of cell biology and pathogenicity of O. tsutsugamushi has lagged behind that of other important human pathogens. One reason for this is that O. tsutsugamushi is an obligate intracellular bacterium that can only be cultured in mammalian cells and that requires specific methodologies for propagation and analysis. Here, we have performed a body of work designed to improve methods for quantification, propagation, purification and long-term storage of this important but neglected human pathogen. These results will be useful to other researchers working on O. tsutsugamushi and also other obligate intracellular pathogens such as those in the Rickettsiales and Chlamydiales families.
A clinical isolate of O. tsutsugamushi was grown in cultured mouse embryonic fibroblast (L929) cells. Bacterial growth was measured using an O. tsutsugamushi-specific qPCR assay. Conditions leading to improvements in viability and growth were monitored in terms of the effect on bacterial cell number after growth in cultured mammalian cells.
Development of a standardised growth assay to quantify bacterial replication and viability in vitro.
Quantitative comparison of different DNA extraction methods.
Quantification of the effect on growth of FBS concentration, daunorubicin supplementation, media composition, host cell confluence at infection and frequency of media replacement.
Optimisation of bacterial purification including a comparison of host cell lysis methods, purification temperature, bacterial yield calculations and bacterial pelleting at different centrifugation speeds.
Quantification of bacterial viability loss after long term storage and freezing under a range of conditions including different freezing buffers and different rates of freezing.
Here we present a standardised method for comparing the viability of O. tsutsugamushi after purification, treatment and propagation under various conditions. Taken together, we present a body of data to support improved techniques for propagation, purification and storage of this organism. This data will be useful both for improving clinical isolation rates as well as performing in vitro cell biology experiments.
| Scrub typhus is a serious, neglected tropical disease that is endemic in large parts of Asia and northern Australia. It is caused by the bacterium O. tsutsugamushi, which is maintained in Leptotrombiculid mites, small arthropods that occasionally bite humans and transmit the disease. O. tsutsugamushi is an obligate intracellular bacterium, which means that it can only survive and grow when it is physically enclosed within a cell, both when it is living in its vector mite, and when it is living in the human or other mammalian host. This makes it difficult to work with in the laboratory, as it needs to be cultured together with host cells. This technical difficulty is one reason why our understanding of this human pathogen is less well-developed than for many other pathogens of equivalent incidence and severity. Here, we have performed a body of work that was designed to measure and improve methods for growing these bacteria in the laboratory, purifying the bacteria from their host cells without damaging them, and preserving bacteria for long periods of time by cryopreservation. This work will support future efforts to understand the basic science behind this and similar intracellular human pathogens.
| Scrub typhus is a serious febrile illness of broad geographical diversity, endemic in the majority of rural Asia and northern areas of Australia. Clinical symptoms resemble that of a number of other tropical diseases including malaria, dengue, leptospirosis, and other bacterial infections, and rapid, unambiguous diagnosis is often unavailable [1]. Consequently it is difficult to know the exact distribution and prevalence of scrub typhus, but best estimates suggest that one billion people per year are at risk and one million people per year are infected. Recent epidemiological studies showed that scrub typhus is a leading cause of serious, under-reported non-malarial fever in rural Thailand, Laos, China and Myanmar [2–6]. It has recently been shown to be a leading cause of CNS infections in a hospital in Laos [7], and is associated with high miscarriage and poor neonatal outcome rates in pregnancy [8].
Scrub typhus is caused by the obligate intracellular Gram-negative bacterium, Orientia tsutsugamushi. This is a member of the Rickettsiaceae family, but differs from bacteria of the genus Rickettsia in important aspects of genome structure, morphology and phenotypic properties [9,10]. O. tsutsugamushi has been shown to infect a wide range of cell types in vitro, but in vivo studies report that it is largely localised to endothelial cells, monocytes and dendritic cells in infected humans [11–17].
One reason why research into the fundamental mechanisms of cellular infection by Orientia tsutsugamushi is less well characterised than those of other equivalent human pathogens is because of the technical difficulties and uncertainties associated with culturing this bacterium in vitro. Clinical isolation of this organism from infected patients presents a major difficulty, where best practice results in isolation rates of only about 40% of PCR-confirmed patients [18], and where bacterial growth and isolation typically requires 4 weeks of growth. This is a serious concern because it means that bacterial culture cannot be used as a diagnostic tool for clinical purposes, and because it reduces the opportunity to generate bacterial isolates for research into strain diversity and potential antibiotic resistance.
In this study we set out to measure and improve methods for the quantification, propagation, purification and long-term storage of Orientia tsutsugamushi. We were inspired by a set of elegant experiments performed by the group of Barbara Hanson in the 1980s, which measured the growth of O tsutsugamushi (then called Rickettsia tsutsugamushi) under a range of conditions [19], and we aimed to update and expand on those observations. It is our hope our observations will benefit researchers engaged in both basic cell biology research and applied clinical diagnostic research.
L929, a mouse fibroblast line, was cultured in Dulbecco’s modified Eagle’s medium (DMEM; Gibco BRL). The media was supplemented with 10% fetal bovine serum (FBS, GIBCO BRL) without antibiotic, unless stated otherwise. Monolayers of L929 were cultured in T25 or T75 cell culture flasks at 37°C in a humidified atmosphere containing 5% CO2. When the cells reached 80–100% confluence, they were ready to be infected with O. tsutsugamushi or to be subcultured into a new flask by trypsinisation.
Subculture by trypsinisation was performed as follows. The culture medium was discarded and the cell layer washed one time with 1X PBS. To disaggregate the cells, 1 ml of 0.25% Trypsin/EDTA (10x Trypsin/EDTA 1:250, PAA, diluted with PBS) was added to the flask and incubated at 37°C. Flasks were tapped gently to facilitate detachment from the flask. Complete detachment was achieved at 3 mins, and was monitored by microscopy. Culture medium at 3x the volume of trypsin was added to neutralize the enzyme and to disperse the cells. The cell suspension was centrifuged at 1000 xg for 5 mins. The cell pellet was collected and resuspended into new medium, and the cell quantity was determined using a Trypan Blue dye exclusion assay. Cells were diluted into fresh growth medium at a ratio of 1:10, and transferred to a new culture flask or grown in appropriate vessels (12-well or 24-well cell culture plates) at 37°C with 5% CO2. After 1 day a confluent cell layer was formed.
O. tsutsugamushi strain UT-76 (a Karp-like strain from Thailand) was propagated in T25 cell culture flasks containing a confluent monolayer of L929 cells which were grown in DMEM supplemented with 10% FBS at 35°C with 5% CO2. After 7 days of infection the cells were sub-cultured onto fresh L-929 cell monolayers. The infected flasks were harvested by scraping using a sterile inoculating loop in 1ml spent medium and disrupted to release intracellular bacteria by lysing in a bullet blender (BBX24B, Bullet Blender Blue, Nextadvance, USA) used at power 8 for 1 min. The lysed bacteria were added to a T25 flask (100 μl/flask) containing an uninfected L-929 monolayer at 70–100% confluence. Infected flasks were passaged at a ratio of 1:5 (original culture flask: subculture flask). Bacteria were sub-cultured for a total of <20 passages, where passage 0 was the original clinical isolate.
Antibiotics were only used in the experiment testing the effect of antibiotic addition. Here antibiotics were used at the following concentrations: chloramphenicol 150 μg/ml, penicillin G 100 μg/ml, penicillin G + streptomycin (premix from Sigma-Aldrich) 125 μg/ml + 200 μg/ml.
Cell lines and bacteria were tested for the presence of mycoplasma and confirmed to be mycoplasma-free using the VenorGeM PCR detection kit (Minerva biolabs).
The DNA copy number of O. tsutsugamushi or L929 cells was measured using qPCR targeting the 47 kDa gene or cfd gene respectively. The primer and the TaqMan probe for the 47 kDa gene were as follows: 47 kDa FW, (5’-TCCAGAATTAAATGAGAATTTAGGAC-3’); 47 kDa RV (5’-TTAGTAATTACATCTCCAGGAGCAA-3’); and 47 kDa probe (5’-FAM-TTCCACATTGTGCTGCAGATCCTTC-TAMRA-3’). The primer and TaqMan probe for the cfd gene were as follows: cfd FW (5’- ACTGAGATCGCTTTTGGGTC-3’); cfd RV (5’- GGAGGGTAGGTGTATTGTAAGG-3’) and cfd probe (5’- 5HEX-CTGGGTTGGAGGTGTCTGTGGT-BHQ2–3’). The qPCR mixture was composed of 1X Platinum supermix (Platinum Quantitative PCR SuperMix-UDG, Invitrogen, USA) or 1X qPCRBIO Probe Mix (qPCR Probe Mix LO-ROX, PCR Biosystems, UK), 0.1 μM forward and reverse primers, 0.2 μM probe, sterile distilled water and 1 μl of extracted DNA. Real time PCR was performed in a 100 TM Thermal Cycle Instrument (Biorad, CFX96 real time system) using the following conditions: initial denaturation at 95°C for 2 min, followed by 45 cycles of denaturation at 95°C for 15 sec and combined annealing and extension at 60°C for 30 sec with the acquisition of fluorescence. The DNA copy number was calculated by comparison with a standard curve.
DNA extraction of cell pellets containing O. tsutsugamushi was performed using four different methods: Alkaline lysis, the commercial Qiagen DNeasy kit, boiling in water or boiling in DMEM.
The protocol for DNA extraction by alkaline lysis (hotshot) was as follows. The cell pellet was resuspended in 20–80 μl alkaline lysis buffer (25 mM NaOH, 0.2 mM EDTA) and boiled at 95°C for 15–60 min. The sample was cooled to 4°C and neutralization buffer (40 mM Tris-HCl, pH 7–8) was added at an equal volume (20–80 μl). Smaller volumes were used where lower bacterial numbers were expected.
DNA extraction using the Qiagen DNeasy kit was performed following the manufacturer’s instructions. DNA extraction by boiling in water or DMEM was performed by resuspending cell pellets in 40 μl water or DMEM and boiling at 95°C for 30 min.
Extracted DNA samples were stored at-20°C as required. Where DNA extraction methods were being directly compared, identical cell pellets were resuspended in an equal final volume.
Infected host cells grown in 12- or 24-well culture plates were harvested for DNA extraction using three different methods: trypsinisation, RIPA buffer and scraping.
For trypsinisation, cells were washed once with PBS and resuspended in 1 ml 0.25% Trypsin/EDTA (10x Trypsin/EDTA 1:250, PAA, diluted with PBS). Cell detachment was monitored by microscopy, and fully detached cells were transferred to 1.5 ml microcentrifuge tubes. Samples were centrifuged at 20,238 xg for 3 min, and the pellet processed for DNA extraction.
Cells released using RIPA buffer were treated as follows. First, the supernatant was removed and either discarded or collected by centrifugation. Then, 300 μl (24 well plate) or 500 μl (12 well plate) of RIPA buffer (Sigma) was added into each well and incubated at room temperature (RT) or 37°C for 3–5 min. Detached cells were monitored under the microscope, and fully detached samples were transferred to a 1.5 ml microcentrifuge tube and centrifuged at 20,238 xg for 3 mins. The pellet was collected for DNA extraction.
Samples prepared by scraping were processed as follows. A sterile plastic inoculating loop was used to remove attached infected cells from the bottom of a 12- or 24-well culture plate, and cells were resuspended directly in the original growth media. The entire volume of resuspended cells was transferred to a 1.5 ml microcentrifuge tube, spun at 20,238 xg for 3 min, and the supernatant removed. Wells were inspected after scraping to assess the efficiency of the process.
Bacteria were purified from host cells after 7 days growth. Host cells were released from the surface of T25 or T75 culture flasks by scraping with a sterile inoculating loop, then the total volume was transferred to a sterile 15 ml falcon tube and centrifuged at 4,000 rpm for 10 mins. The cell pellet was collected and resuspended in 1–2 ml spent medium or new medium. Host cells were lysed using either a vortex or a bullet blender (BBX24B, Bullet Blender Blue, Nextadvance, USA) either without beads or with 0.2 mm glass beads (Life Science AP, Thailand) or with 0.9–2.0 mm stainless steel beads (Next Advance, USA). Unless stated in the text, lysis was done using the bullet blender at power 8 for 1 min with no beads added. Lysis by bullet blender was performed using Eppendorf Safe Lock tubes (Eppendorf, USA) and lysis by vortex was done using regular 1.5 ml microcentrifuge tubes. Lysed cells were centrifuged once at 50 xg for 3 min to pellet residual host cells and large cell debris. The supernatant containing released bacteria was transferred to a clean 1.5 ml microcentrifuge tube, and this was centrifuged at 20,238 xg for 3 mins to pellet cell-free bacteria. Pellets were collected and resuspended in different buffers for further experiments. Purification was performed at room temperature unless stated otherwise.
O. tsutsugamushi purified as described above were treated in a number of different ways in order to assess viability after short- and long-term storage. For short-term storage, purified bacteria were resuspended in either sucrose-phosphate-glutamate (SPG) buffer (0.218 M sucrose, 3.76 mM KH2PO4, 7.1 mM KH2PO4, 4.9 mM monosodium L-glutamic acid), PBS, PBS + 10% BSA, DMEM + 10% FBS or water, and stored at room temperature for 30 mins or 120 mins before being grown on L929 cells in 24 well plates and assessed for growth after 7 days.
For long-term storage purified bacteria or bacteria in intact host cells were pelleted and resuspended in a range of different buffers: SPG, SPG + 15% glycerol, SPG + 10 mM MgCl2, SPG + 10 mM MgCl2 + 15% glycerol, RPMI, RPMI + 15% glycerol, or 90% FBS + 10% glycerol. All samples were transferred to a 1.5 ml cryovial and placed directly in the-80°C freezer and stored for 2 hours, then thawed at 37°C for 5 min. Samples were infected onto fresh monolayers of L929 in 24-well plates and assessed for growth after 7 days.
Experiments were performed to compare the effect of freezing and thawing at different speeds. Purified bacteria were pelleted and resuspended in SPG buffer. Samples were transferred to a cryovial and frozen and thawed under different conditions. For freezing, cells were either frozen fast (placed in a dry-ice gel for 10 min before being transferred to-80°C freezer) or slow (by placing samples in an CoolCell LX (Biocision) alcohol-free freezing chamber directly in the-80°C freezer, which freezes samples at a rate of ~ 1°C per minute). For thawing, cells were either thawed fast (placed in a 37°C water bath until thawed) or slow (placed one ice until thawed completely). Bacteria were then seeded onto L929 cells and assessed for growth after 7 days.
Infected L929 cells, which had been lysed in different ways, were assessed by fluorescence confocal microscopy in order to determine to effect of lysis on morphology. Two μl of lysed samples were fixed on the 12-well slide with 100% acetone for 10 mins at-20°C then left to dry completely. Slides were then incubated in a humidified slide chamber at room temperature with 1/100 anti-O. tsutsugamushi primary monoclonal antibody (made in house) for 30 mins, washed three times with PBS and incubated with 1/400 Alexafluor 488 anti-mouse secondary antibody (Invitrogen, USA) for 30 mins. Evans blue was included with the secondary antibody at a dilution of 1/200. The samples were washed three times with PBS, then mounted using Vectamount (Vecter Laboratories, Burlinggame, CA, USA) which contained the blue nuclear intercalating counterstain DAPI (4′,6-diaminidimo-2-phenylindole). Samples were examined under confocal laser scanning microscope (Zeiss LSM 700).
47 kDa antigen gene from O. tsutsugamushi (strain Ikeda): OTT_1319/ID 6337858 cfd gene from Mus musculus: ID 11537
First, we aimed to develop a standardised assay that would generate accurate and reliable quantitative information about the number of viable cells in a bacterial sample. This assay would be used to analyse two types of experiments. (1) We would treat purified bacterial samples in different ways and compare the effects of these treatments on bacterial viability and (2) we would quantify the growth of bacteria under different growth conditions. When deciding how to assess viability and growth for these experiments we rejected fluorescence microscopy-based assays, such as the syto9/propidium iodide live/dead reagents, because we wanted to specifically measure the ability of viable O. tsutsugamushi to invade and replicate within host cells, and we could not exclude the possibility that microscopy-based assays would score some bacteria as live, that were structurally intact but metabolically dormant or inactive. Therefore, we decided to develop an assay based on active uptake and replication of bacterial cells.
We opted for qPCR to measure bacterial copy numbers. As an obligate intracellular bacterium it is not possible to directly assess the bacterial copy number using light absorption (OD) measurements as would be done for other model organisms such as E. coli. Plaque-based assays could have been used [20–22] but these are difficult to scale up to test large numbers of conditions and biological repeats. Finally, it would have been possible to quantify bacterial levels using light microscopy techniques such as Giemsa staining or acridine orange as has been done previously [23], but the accuracy and reproducibility of this method is low. We therefore quantified O. tsutsugamushi levels by extracting total DNA and measuring the exact genome copy number using a primer/probe set specific to a short region within the O. tsutsugamushi 47kDa Htra single copy gene. A similar approach has been used to track the growth of spotted fever group Rickettsiae in cultured cells [24].
It would be desirable to measure a complete growth curve over 1–2 weeks for each different condition, as this would give complete information about the degree of uptake and dynamics of intracellular replication of the bacteria. This approach, however, would require a prohibitive number of samples since each condition would need to be measured multiple times to account for biological variation. We therefore examined whether a single time point could give an accurate indication of the number of viable bacteria and the degree of intracellular growth. In order for this to be a robust measurement of uptake and growth the bacterial copy number at a single time point must be linearly proportional to the number of bacteria added at the time of infection.
Fig 1A shows a typical growth curve of O. tsutsugamushi in cultured mouse fibroblast (L929) cells. After 7 days, the bacteria have achieved a substantial level of replication, and this was the time point we selected for all subsequent experiments. In order to validate this choice, we measured the growth after 7 days for a single batch of O. tsutsugamushi inoculated onto host cells at different concentrations. Fig 1B shows that the bacterial copy number after 7 days is roughly proportional to the relative number of bacteria added at the time of infection within a certain range (Here, 1x105–1x107 copy numbers). This is probably because when the bacteria level gets too high they saturate the L929 host cells available for infection and therefore no further bacterial growth is permitted. Fig 1B also demonstrated that bacteria boiled for 30 mins produced no detectable bacteria after 7 days of growth.
In summary, this shows that a single measurement after 7 days of isolate growth can be used to determine the relative number of bacteria available for uptake and growth in L929 host cells and thus to accurately compare conditions for bacterial treatment and growth. Infected cells were always washed 3 hours after infection in order to remove non-infectious extracellular bacteria. In order to account for differences in the exact number of bacteria being used in different conditions within an experiment, a portion of the inoculating bacteria was always retained for each experimental condition. This was quantified by qPCR and used to normalise the 7-day growth between conditions.
Growth assays were performed in 12-well or 24-well cell culture plates. We compared methods for removing and lysing the adherent L929 cells which had been infected with O. tsutsugamushi. Samples grown and infected under identical conditions were treated using RIPA cell lysis buffer, scraping using a sterile inoculating loop or release using trypsin-EDTA. All samples were then processed using the alkaline lysis DNA extraction method and it was found that treatment with RIPA buffer released approximately 3 times more bacteria than the other two methods (Fig 2A). We selected to use this method in all future experiments.
Different DNA extraction methods were compared (Fig 2B). Identical samples were extracted using a Qiagen DNAeasy kit, Alkaline lysis, boiling in water or boiling in cell growth media. Samples were adjusted to the same volume and quantified by qPCR. The alkaline lysis method resulted in the highest level of O. tsutsugamushi DNA, with the other three methods roughly equivalent. The DNAeasy kit, however, resulted in the highest purity DNA as assessed by nanodrop analysis (Fig 2, table). The appropriate method will therefore depend on individual requirements for cost, speed, purity and quantity. The alkaline lysis method is simple and cheap, and yields high levels of detectable bacterial DNA, and we selected to use it in all subsequent experiments described here.
Finally, we determined the effect of boiling time on the DNA yield produced by the alkaline lysis DNA extraction method (Fig 2C). We found that increasing boiling time released more bacterial DNA, but the differences were not very large. We used boiling times ranging between 15 min and 60 min in subsequent experiments, although this was kept constant within a single experiment.
We compared the growth of O. tsutsugamushi after 7 days under a range of different growth conditions. First, we compared the effects of different inoculation conditions (Fig 3). It has previously been reported that O. tsutsugamushi can infect cells both in the adherent and trypsinised state [21]. We repeated this measurement and found a slight decrease in the efficiency of bacterial uptake when infecting newly trypsinised cells compared with adherent cells (Fig 3A). For routine work, therefore, we performed all infections on adherent cultured cells. However, this observation increases the potential for flexibility in future experimental designs as needed.
We measured the growth of bacteria in host cells that had been infected at different levels of confluence (Fig 3B). After 7 days the host cells had all grown to full confluence and any effect is likely to be dominated by a difference in efficiency of bacterial uptake upon infection. We found that host cell confluence had no strong effect on bacterial copy number after 7 days. The bacterial level in cells infected at 50% confluence was slightly lower than the others, but this can be accounted for by the fact that there were fewer host cells available for infection at time 0. From this we concluded that the host cell confluence had no strong effect on bacterial uptake and replication, and that this aspect of future workflow could be considered non-critical.
Next, we compared the growth of O. tsutsugamushi grown in different growth media and with different supplements (Fig 4). We found that addition of fetal bovine serum (FBS) improved the growth of bacteria as shown previously [19]. We found that bacterial growth was increased when FBS was added at 5% and then further when added at 10% or 20%, but that addition of 50% FBS was inhibitory compared with lower concentrations. (Fig 4A). We also quantified the growth of L929 host cells in these experiments using a qPCR assay based on the mouse gene cfd, and found that growth in the presence of 50% FBS was increased around 3-fold compared with any of the lower FBS concentrations (Fig 4B). Therefore, the inhibition of FBS at high concentrations may be due to a reduction in bacterial replication in rapidly dividing host cells. Given the high cost of FBS we used 10% in all standard experiments. In cases where the inoculating level of bacteria is limiting, however, such as in clinical isolates, it may be advantageous to increase the level of FBS supplementation to 20%.
Mouse fibroblast (L929) cells can grow well in both RPMI and DMEM growth media, and we evaluated whether there would be any significant effect on bacterial copy number between the two (Fig 4C). We found no significant difference and therefore concluded that both would be suitable to support growth of O. tsutsugamushi. Daunorubicin hydrochloride is a drug that inhibits DNA and RNA synthesis in mammalian cells. This leads to a growth inhibition, and this has previously been shown to enhance the growth of bacteria in cultured mammalian cells [19,25]. A relationship between O. tsutsugamushi growth and host cell growth rate was also indicated in the FBS experiment, above. The optimum amount of daunorubicin to support bacterial growth varies between strains and host cell types, and we set out to determine the optimum concentration to support growth of O. tsutsugamushi in L929 cells (Fig 4D). Our results show a small increase in bacterial copy number after 7 days growth in the presence of 0.2, 0.4 and 0.8 μg/ml compared with no daunorubicin. This corresponds to a decrease in host cell replication (Fig 4E). At 1 μg/ml and 1.5 μg/ml bacterial growth was greatly reduced, and this is likely explained by excessive damage to host cells (Fig 4E). Whilst the small growth enhancement shown here may not be necessary for routine growth, this could be beneficial in cases where bacterial growth is difficult.
Antibiotics are sometimes added to cultured mammalian cells in order to avoid bacterial contamination. A combination of penicillin and streptomycin is frequently added to O. tsutsugamushi cultures for this purpose [26]. We measured the growth of O. tsutsugamushi in the presence of penicillin-streptomycin and found no effect on growth rate, supporting its use where required (Fig 4E). Chloramphenicol, which is active against O. tsutsugamushi, was used as a positive control and inhibited bacterial growth as expected [25,27].
For certain experiments it is desirable to purify bacteria from the host cells in which they are growing, for example to study growth in a different host cell type or to study the early stages of bacterial infection. This can be performed using percoll density gradient centrifugation [28], or by lysing host cells and separating released bacteria by centrifugation and filtration [29]. The latter approach was employed here. Since O. tsutsugamushi is an obligate intracellular bacterium this process will inevitably cause some damage to some fraction of the bacterial cells purified in this way. Our aim in this set of experiments was to quantify the loss of number and viability of bacteria purified under different conditions, in order to develop improved protocols for extracting bacteria from their host cells when required.
The first step in bacterial purification is host cell lysis. There are a number of ways in which this is achieved, with varying degrees of efficiency and bacterial damage. First, we used fluorescence light microscopy to study the physical effect of a range of different host cell lysis methods (Fig 5A and S1 Fig). We imaged both bacteria and host cells, and found that lysis using the bullet blender with either 0.2 mm glass beads, or no beads at all, led to efficient lysis of host cells and release of bacteria. Surprisingly, addition of 2 mm stainless steel beads appeared to inhibit host cell lysis under these conditions. Where the bullet blender instrument is not available, or where large volumes are required, vortexing for 1 min in the presence of 0.2 mm beads is the best option. Having identified efficient lysis methods, we wanted to measure the effect on bacterial viability of different methods of host cell lysis (Fig 5B). In order to do this we compared the growth of bacteria that had not been lysed, or lysed in a range of different methods. We found that the bacteria from lysed host cells were viable and demonstrated growth to a comparable level as bacteria from host cells that had not been lysed. It is important to note, however, that the uptake efficiency of bacteria from unlysed hosts is likely to be lower than that of purified bacteria, because they need to first exit the original, unlysed host cell before entering the new cell. Therefore we cannot truly compare the uptake and growth between the bacteria from cells that had been lysed, and those that had not. We can, however, conclude that bacterial purification by these methods has not resulted in massive loss of viability.
We compared the viability of bacteria that had been purified at 4°C and 25°C and found no significant difference (Fig 5C). We concluded, therefore, that future bacterial purifications could be conducted at room temperature with no effect on subsequent growth.
We quantified the yield of bacteria at different stages of purification (Fig 5D). After host cell lysis, samples were spun at 50 xg for 3 min to remove host cell debris. This resulted in no detectable loss of bacteria. Samples were then filtered using a 5 μm pore-size filter, and this resulted in a large loss of bacteria (about 90% loss). We concluded that this loss could be tolerated when a high degree of purification was desired (for example when moving between different host cell types) but that in other cases filtration could be excluded where high bacterial levels were a priority.
Finally, we measured the sedimentation of purified O. tsutsugamushi after centrifugation at different speeds (Fig 5E). A sample of purified bacteria was separated into 6 identical Eppendorf tubes, and these were each spun at different speeds. The supernatant was separated from the pellet and the levels of bacteria in each was quantified by qPCR. It was found that at low speeds (50 and 300 xg) there remained high levels of bacteria in the supernatant, but that at speeds of 5,000 xg and above, most of the bacteria were in the cell pellet. There was no difference in bacterial numbers in the pellets of samples spun at 5,000 xg, 10,000 xg and 15,000 xg, but the total amount almost doubled when sample speeds reached the maximum speed of 20,238 xg. From these observations we concluded that the majority of intact bacteria pelleted at 5,000 xg with no significant increase up to 15,000 xg but that at 20,238 xg small fractions of bacteria and free DNA were also pelleted. We then performed growth experiments in order to determine whether high centrifugation speeds had a negative effect on bacterial viability. We pelleted purified bacteria at different speeds and measured the growth after 7 days (Fig 5F). High pelleting speeds had no effect on subsequent growth of bacteria, and therefore we conclude that centrifugation speeds of 5,000–20,238 xg are appropriate for bacterial purification.
In some cases it may be desirable to store isolated bacteria for short periods of time in order to perform certain experiments. It would therefore be advantageous to know exactly how much damage is caused to bacterial cells after storage under different conditions and for different periods of time.
First, we measured the effect of storing purified bacteria in a range of different buffers for 30 mins or 120 mins at room temperature (Fig 6A). Bacteria were isolated from host cells (without filtering), resuspended in DMEM, water, PBS, PBS + 2% BSA, or SPG buffer, and stored at room temperature for 30 min or 120 min before being inoculated onto a lawn of L929s and grown for 7 days. As expected, it was found that water had the greatest negative effect on bacterial growth, probably due to the high osmotic stress to the cells. In support of this, the inhibitory effect of PBS could be somewhat relieved by the addition of BSA, which would increase the osmolarity of this solution. It was found that storage for 120 mins caused a much greater loss of bacteria than storage for 30 mins under all conditions. From this we concluded that O. tsutsugamushi is sensitive to solute osmolarity and that even a short incubation time caused a decrease in bacterial viability. This is not surprising because O. tsutsugamushi is an obligate intracellular bacterium that is unlikely to experience conditions of low osmolarity in its environment and is unlikely to spend any significant portion of its life cycle in a purified, cell-free state.
Second, we aimed to compare the stability of bacteria that had been purified from their host cells with bacteria that were stored within mammalian host cells (Fig 6B). The host cells were scraped from the cell culture flask and all samples were stored in spent culture media in sealed Eppendorf tubes. We compared the viability of purified and non-purified bacteria stored at 37°C or 4°C for 1 day or 7 days. We found that the bacterial viability went down slightly after storage for one day or 7 days at 37°C, and decreased more significantly upon storage at 4°C. In all cases the loss in purified cells was greater than that in intact cells, and in fact when bacteria were stored at 37°C in intact host cells no loss of viability was detected. This is likely to result from the presence of intact, live L929 cells stored at 37°C, even though samples were not grown in ideal culture conditions for the duration of the storage. From these data we concluded that storage in intact hosts was preferable to storage in lysed host cells and that long-term storage up to one week, where required, should be performed at 37°C.
High quality cryopreservation of clinical isolates and laboratory-propagated bacteria is an important aim for all researchers engaged in scrub typhus research [30,31]. We aimed to quantify and optimise the effect of different freezing methodologies on bacterial viability. First, we compared the viability of bacteria cryopreserved in lysed or intact host cells. We observed a clear improvement when bacteria were frozen in intact host cells and recommend this practice where possible (Fig 7A).
We compared the viability of bacteria frozen in intact host cells using two different freezing buffers: FBS/DMSO or sucrose-phosphate-glutamate (SPG) buffer (Fig 7B). We found a clear benefit to using SPG freezing media, as has been recommended for other intracellular bacteria [29,32].
There are certain instances, however, where it is desirable to freeze purified bacteria. One example is when performing cell biology experiments where it would be advantageous to prepare a large batch of purified, validated bacteria to ensure consistency across future experiments. Therefore, we aimed to improve methods for freezing bacteria that had been purified from host cells.
First, we compared the viability of purified O. tsutsugamushi after freezing in a range of different buffers (Fig 7C). We found that SPG-based buffers exhibited a protective effect compared with RPMI alone, but that addition of 15% glycerol or MgCl2 to SPG had no additional benefit for bacterial viability. Freezing in the presence of RPMI + 15% glycerol, in contrast, gave the same degree of cryoprotection as freezing in SPG-based buffers.
Second, we measured the effect of freezing and thawing speed on bacterial viability (Fig 7D). We compared freezing in dry ice (fast) with freezing in isopropanol chambers (slow), and thawing in a 37°C water bath (fast) with thawing on ice (slow). Surprisingly, we found that the speed of freezing and thawing had no observable effect on bacterial viability. Therefore, we concluded that this process was not critical to the outcome of cryopreservation.
The above experiments were performed using short freezing times of 40 mins, as this enabled multiple experiments to be performed in one day and therefore different conditions to be directly compared. In order to determine whether this was sufficient time to measure the effects of freezing, we compared the growth of bacteria that had been stored frozen for 40 mins or 1 week. We found no significant different in recovery rates (Fig 7E). This study did not address the effect of extended freezing periods (months or years), nor the effects of storage in liquid nitrogen compared with a-80°C freezer.
The aim of this study was to develop an evidence-based experimental toolkit to support future research into the important neglected human pathogen O. tsutsugamushi. Here, we have shown that a simplified growth assay can be used to compare the growth of this organism across a range of conditions, and this will be useful in other areas such as comparing the antibiotic sensitivity of a range of isolates.
Using an optimised, simplified growth assay we have shown that growth of O. tsutsugamushi in cultured mouse fibroblast (L929) cells can be improved by the addition of up to 20% FBS and 0.2–0.8 ug/ml daunorubicin, but that uptake and growth is not strongly affected by host cell media (RPMI/DMEM), the confluence of host cells at infection, or whether the host cells are adherent or recently trypsinised at the point of infection. These results suggest that flexibility in these parameters when growing O. tsutsugamushi in cultured cells can be tolerated.
It is frequently desirable to isolate live bacteria from the host cells in which they are growing, in order to perform experiments, for example on the cellular infection cycle. Here, we found that the efficiency of host cell lysis is sensitive to the time and power of lysis, and also to the bead type used. We show that effective host cell lysis has no significant effect on bacterial viability. We attempted to optimise the purification of O. tsutsugamushi from host cells, and found that the temperature of purification (4°C or room temperature) was not important, but that filtering using a 5 μm filter resulted in a loss of bacteria of about 1 order of magnitude. We found that pelleting speeds of 5,000 xg were sufficient to pellet the majority of purified O. tsutsugamushi.
We measured the effect of short-term storage on the viability of O. tsutsugamushi, since certain experiments may require periods of storage while other reagents are being prepared. We found that storage in a range of buffers resulted in a loss of viable bacteria over a short time frame (30–120 min) and that this could be reduced by the use of buffers of higher osmotic strength.
Finally, we compared the number of viable bacteria after cryopreservation across a number of different conditions. We found that preservation in intact host cells led to significantly larger numbers of viable bacteria than preservation of purified bacteria. However, where it is necessary to cryopreserve purified bacteria we found that different buffers could preserve bacteria to varying degrees. Throughout, we found that SPG buffer conferred the greatest protection against cryodamage but that the speeds of freezing and thawing had little effect on preserving bacterial viability.
Taken together, these data can be used to develop experimentally validated protocols for the purification, growth and storage of O. tsutsugamushi, and it is expected that some of these results will be transferable to other obligate intracellular bacteria.
|
10.1371/journal.ppat.1007354 | Phosphatase activity of the control of virulence sensor kinase CovS is critical for the pathogenesis of group A streptococcus | The control of virulence regulator/sensor kinase (CovRS) two-component system is critical to the infectivity of group A streptococcus (GAS), and CovRS inactivating mutations are frequently observed in GAS strains causing severe human infections. CovS modulates the phosphorylation status and with it the regulatory effect of its cognate regulator CovR via its kinase and phosphatase activity. However, the contribution of each aspect of CovS function to GAS pathogenesis is unknown. We created isoallelic GAS strains that differ only by defined mutations which either abrogate CovR phosphorylation, CovS kinase or CovS phosphatase activity in order to test the contribution of CovR phosphorylation levels to GAS virulence, emergence of hypervirulent CovS-inactivated strains during infection, and GAS global gene expression. These sets of strains were created in both serotype M1 and M3 backgrounds, two prevalent GAS disease-causing serotypes, to ascertain whether our observations were serotype-specific. In both serotypes, GAS strains lacking CovS phosphatase activity (CovS-T284A) were profoundly impaired in their ability to cause skin infection or colonize the oropharynx in mice and to survive neutrophil killing in human blood. Further, response to the human cathelicidin LL-37 was abrogated. Hypervirulent GAS isolates harboring inactivating CovRS mutations were not recovered from mice infected with M1 strain M1-CovS-T284A and only sparsely recovered from mice infected with M3 strain M3-CovS-T284A late in the infection course. Consistent with our virulence data, transcriptome analyses revealed increased repression of a broad array of virulence genes in the CovS phosphatase deficient strains, including the genes encoding the key anti-phagocytic M protein and its positive regulator Mga, which are not typically part of the CovRS transcriptome. Taken together, these data establish a key role for CovS phosphatase activity in GAS pathogenesis and suggest that CovS phosphatase activity could be a promising therapeutic target in GAS without promoting emergence of hypervirulent CovS-inactivated strains.
| Group A streptococcus (GAS), also known as Streptococcus pyogenes, causes a broad array of human infections of varying severity. Tight control of production of virulence factors is critical to GAS pathogenesis, and the control of virulence two-component signaling system (CovRS) is central to this process. The activity of the bifunctional histidine kinase CovS determines the phosphorylation status and thereby the activity of its cognate response regulator CovR. Herein, we sought to determine how varying CovR phosphorylation level (CovR~P) impacts GAS pathophysiology. Using three infection models, we discovered that GAS strains lacking CovS phosphatase activity resulting in high CovR~P levels had markedly impaired infectivity. Transcriptome analysis revealed that the hypovirulent phenotype of CovS phosphatase deficient strains is due to down-regulation of numerous genes encoding GAS virulence factors. We identified repression of additional virulence genes that are typically not controlled by CovR, thus expanding the CovR regulon at high CovR~P concentrations. Our data indicate that phosphatase activity of CovS sensor kinase is crucial for spatiotemporal regulation of GAS virulence gene expression. Thus, we propose that targeting the phosphatase activity of CovS sensor kinase could be a promising novel therapeutic approach to combat GAS disease.
| The ability of bacteria to modify gene expression levels in adaptation to external influences is key to many aspects of bacterial pathogenesis [1]. Two-component regulatory systems (TCS) are a major mechanism by which bacteria detect and respond to diverse environmental factors [2, 3]. TCS are absent in humans but abundant in a wide range of bacteria. They usually consist of a membrane-embedded histidine kinase that determines the regulatory activity of its cognate response regulator by altering its phosphorylation status [2, 4].
The control of virulence regulator sensor (CovRS, also called CsrRS for capsule synthesis regulator) system of group A streptococcus (GAS) is one of the best-studied TCS in connection with bacterial pathogenesis [5, 6]. GAS is a strictly human pathogen that causes a variety of diseases from relatively benign to life threatening such as necrotizing fasciitis [7]. GAS strains are classified into >200 serotypes based on variability in the key anti-phagocytic, cell-surface exposed M protein [8, 9]. In tandem with the histidine kinase CovS, CovR is the central regulator of GAS virulence factor production [5, 10]. Similar to other OmpR/PhoB family members, CovR is phosphorylated at a conserved aspartic acid residue (D53) to create CovR~P, which is considered to be the active regulatory form of the protein [11–13].
Several signaling pathways converge to tightly regulate CovR~P levels. CovS primarily serves to increase CovR~P via its kinase activity [14]. As a member of the bifunctional HisKA-family of histidine kinases [15], CovS also possesses phosphatase activity to reduce CovR~P. Extracellular signals (e.g. Mg2+ or LL-37) influence CovS activity to modulate CovR~P levels and CovR-regulated virulence gene expression [16–18]. Deletion of covS reduces but does not completely eliminate CovR~P [14] suggesting a contribution of intracellular metabolites such as acetyl phosphate to CovR phosphorylation. In combination with CovS, the orphan kinase regulator of CovR, RocA, increases CovR~P levels [19]. It has been discovered that conserved mutations in the rocA gene mediate serotype-specific intrinsic CovR~P levels [20, 21]. Finally, serine/threonine kinase (Stk) phosphorylates CovR residue threonine 65 in a fashion that antagonizes D53 phosphorylation [22]. This multi-faceted regulation of CovR~P status underpins its key position in GAS infectivity, which is further highlighted by the observation that the CovRS system is one of the hotspots for mutations in invasive clinical isolates [23, 24]. Given that CovR mainly serves as a transcriptional repressor, emergence of CovRS inactivating mutations during human infection or animal passage result in hypervirulent GAS strains due to relieved repression of virulence gene expression [25–28]. It has been speculated that there is a serotype/strain-specific tendency for the acquisition of covRS mutations that comes along with the underlying invasive potential of different M-types [29]. This tendency is thought to be determined by serotype-specific expression of certain virulence factors, such as the hyaluronic acid capsule, the M protein and the secreted DNase Sda1 [30–33]. Likewise, the intrinsic CovR~P levels of GAS strains might play a role in determining the selective pressure to lower CovR~P via covRS mutations [14].
Although the CovRS system has been extensively studied, several key outstanding issues regarding its impact on GAS pathogenesis remain unanswered. First, it is well established that CovRS-inactivation increases GAS virulence in bacteremia models [5, 21, 22, 25, 34]. However, there are conflicting results regarding how CovR~P levels impact GAS skin/soft tissue infection and nasopharyngeal colonization [20, 25, 26, 35, 36]. Second, the specific influence of CovR kinase and phosphatase activity on GAS virulence is not clear since previous research has focused on the impact of CovR or CovS inactivation. We recently showed that a strain lacking CovS phosphatase activity resulted in increased CovR~P levels [14]. However, how increased CovR~P levels due to abrogating CovS phosphatase activity impact GAS infectivity remains unknown. Finally, virulence studies using strains with varying CovR~P levels have generally been restricted to a single GAS serotype leaving open the question of the applicability of findings in a single strain to the broader GAS population. To address these questions, we employed isoallelic GAS strains of two distinct M serotypes that are common causes of human disease to comprehensively investigate how defined CovR~P levels impact GAS virulence, global gene expression, and emergence of hypervirulent CovRS mutated strains.
We have previously created and characterized isoallelic serotype M3 strains in which single amino acids in CovRS are altered compared to the parental strain MGAS10870 [14, 22]. We recreated the same set of mutations in the parental strain MGAS2221, a contemporary serotype M1 GAS strain (see Table 1) (For clarity of the genetic background of the used strains, we refer to these strains as M1-wild type (WT), M3-wild type (WT) etc. for the remainder of the manuscript). The CovR~P levels (defined as [CovR~P]/[CovR]total) in the respective M1 GAS strains during growth in THY medium were determined by Phostag-Western blot using anti-CovR antibodies (Fig 1). Similar to our M3 studies [14, 22], changing the CovR phosphorylation site aspartate 53 to an alanine resulted in no detectable CovR~P in strain M1-CovR-D53A. Abolishing CovS kinase activity via the E281A mutation decreased CovR~P levels to ~20% in M1-CovS-E281A, resembling a covS deletion strain, whereas the T284A change abrogating the CovS phosphatase activity increased CovR~P levels to ~80% in strain M1-CovS-T284A (Fig 1). Thus, the CovR~P levels observed in the isoallelic mutant strains of M1-WT closely mirrored those previously observed for the serotype M3 derivatives [14]. However, it is important to note that CovR~P levels are higher in the M1-WT strain (~70%) relative to the M3-WT strain (~40%) due to a naturally occurring mutation in the RocA protein in M3 strains [14, 21]. This collection of strains allowed us to study in detail the impact of CovR phosphorylation on GAS virulence and covRS switching rates during infection in two of the most prevalent GAS serotypes causing pharyngitis and severe infections in the US (Center for Disease Control and Prevention [37]).
To investigate the relationship between CovR~P levels and GAS virulence, we first employed a skin/soft tissue mouse model that mimics cellulitis, a common manifestation of GAS infection. For each strain, 20 mice were subcutaneously challenged with 107 colony forming units (CFU) of GAS, and lesion size measurements were performed daily (Fig 2A and 2B). In the serotype M3 background (Fig 2A), all of the strains caused appreciable disease with strain M3-CovR-D53A causing the largest lesions over the course of the entire experiment (P <0.05 in comparison to each of the other three strains). The difference in lesion size generated by strain M3-CovS-E281A compared to the M3-WT strain was not significantly different (P = 0.99). The smallest lesions were produced by strain M3-CovS-T284A (P < 0.05 compared to each of the other three strains). However, it is notable that for some mice inoculated with strain M3-CovS-T284A, lesions started to appear more severe and showed ulceration after day 5 of the experiment, when healing of lesions had already set in for the other strains. We will address this finding later in the manuscript.
Among the M1 strains, the wild type strain was the most virulent with maximal average lesion area of ~150 mm2 on day 4 (P < 0.05 compared to the other three strains), followed by strain M1-CovS-E281A (P < 0.05 compared to strains M1-CovS-D53A and M1-T284A) (Fig 2B). Strain M1-CovS-T284A was the least virulent (Fig 2B) (P < 0.05 compared to the other strains). However, in contrast to its serotype M3 counterpart, only a few mice inoculated with M1-CovS-T284A evidenced any visible infection, and in these cases the lesions were small in size and without any sign of ulceration over the entire length of the experiment. Thus, we did not observe a consistent relationship between CovR~P levels and lesion size. Regardless, our results demonstrate that abrogation of CovS phosphatase activity in both GAS M1 and M3 serotypes attenuated virulence in the skin/soft tissue mouse model of infection.
Next, we investigated the effect of CovR~P levels on oropharyngeal GAS colonization following nasopharyngeal mouse challenge. For serotype M3 strains, 20 mice per strain were inoculated with 108 CFU GAS in a 20 μl volume. The volume was reduced compared to previous studies to avoid aspiration of GAS into the lungs [36]. Unexpectedly, we observed a high death rate in mice inoculated with strains M3-CovS-E281A and M3-CovR-D53A, while no death occurred for mice inoculated with M3-WT or M3-T284A, albeit the same volume and CFU were used. For this reason, we subsequently reduced the inoculum to 107 CFU for all serotype M1 strains. Despite the different inocula used for the M1 and M3 strains, we consistently observed that GAS strains with higher CovR~P colonized at lower rates (Fig 2C and 2D). Specifically, low CovR~P strains CovS-E281A and CovR-D53A colonized at significantly higher rates (~80% and 40–50% for serotype M3 and M1, respectively) compared to the wild type and CovS-T284A strains for both serotypes (P < 0.05 for all specified comparisons). For serotype M3, the wild type strain (medium CovR~P level) had a colonization rate intermediate to the low and high CovR~P strains (~30%, Fig 2C). For serotype M1, the wild type strain and M1-CovS-T284A (both high CovR~P) were rarely recovered shortly after initial inoculation (Fig 2C and 2D). Thus, we observed an inverse relationship between CovR~P levels and the capability of the GAS strains to colonize mouse pharyngeal tissue.
To investigate whether changes in CovR~P levels affect GAS interaction with the human immune system, we performed a Lancefield bactericidal assay using heparinized whole human blood of three non-immune donors. GAS survival in whole human blood was determined by plating serial dilutions on BSA plates after 3 hours exposure of GAS cells to human blood, and multiplication factors compared to the inoculum were calculated for each strain (Fig 2E and 2F). Multiplication factors were generally lower for serotype M1 compared to serotype M3 (Fig 2E and 2F). This is consistent with the previous finding that an intact RocA protein negatively influences GAS survival in blood [39]. The multiplication factors between the respective wild type and CovR-D53A or CovS-E281A mutant strains were not significantly different (P > 0.05 for all comparisons). In stark contrast, CovS-T284A derivatives of both serotype M1 and M3 GAS completely lost the ability to survive and propagate in whole human blood (P < 0.001 for all comparisons) (Fig 2E and 2F). Thus, CovS phosphatase activity is crucial for GAS survival and propagation in whole human blood.
In standard laboratory medium CovR~P levels of strain M1-CovS-T284A are only slightly elevated compared to those in M1-WT, yet there were dramatic differences in virulence between the two strains in the skin/soft tissue infection model and bactericidal assay (Figs 1 and 2). One possible explanation for this discrepancy is that host factors increase CovS phosphatase activity during infection to decrease CovR~P levels and augment virulence factor production. Indeed, CovS senses the human antimicrobial peptide LL-37 and responds by lowering CovR~P levels [14, 18]. Thus, we next addressed the question whether LL-37 signaling is affected by the CovS-T284A mutation. To this end, we measured CovR~P and hasA (hyaluronic acid capsule) transcript levels in GAS strains grown in THY (standard laboratory medium) and THY supplemented with LL-37 (Fig 3A and 3B). Consistent with our previous findings [14], M1-WT CovR~P levels were reduced and consequently hasA transcript levels elevated in the presence of LL-37 compared to unsupplemented THY. In contrast, supplementation with LL-37 did not affect CovR~P or hasA transcript levels in strain M3-WT or strain M1-WT engineered to contain the truncated M3 version of RocA (M1-rocAM3) [39]. In both serotypes, the CovS-T284A strains had higher CovR~P and lower hasA transcript levels compared to the respective wild type strains when grown in unsupplemented THY. However, neither CovR~P nor hasA transcript levels were influenced by LL-37 in the CovS phosphatase deficient strains (Fig 3). Further, CovR~P and hasA transcript levels of CovS phosphatase deficient strains were not affected by an additional mutation in the rocA gene in either medium (see strain M1-CovS-T284A/rocAM3, Fig 3). We conclude that LL-37 increases CovS phosphatase activity in a GAS strain with an intact RocA.
It is well established that strains with covRS mutations can emerge during human infection or animal passage, thereby giving rise to hypervirulent GAS isolates [23, 25–27]. To investigate how CovR~P levels in the context of GAS serotype influence the emergence of CovRS inactivated mutants, we amplified and sequenced the complete covRS operon from at least 25 randomly picked GAS colonies per strain isolated during each infection study (Table 2). In the skin/soft tissue model of infection, we recovered GAS colonies from skin lesions of five mice per strain on day 4. No additional mutations were found in colonies isolated from mice infected with strains M3-WT, M3-CovS-E281A, and M3-CovR-D53A. In contrast, a few of the colonies isolated from mice infected with strain M3-CovS-T284A had mutations in CovR, namely CovR-L155I and CovR-R66H. Since the lesions in some mice infected with M3-CovS-T284A looked more severe after day 5, we sequenced additional colonies isolated from these animals on day 10 (end point of experiment) and were able to detect colonies that had a duplication of covS nucleotides 100 to 131, which is predicted to result in a non-functional CovS due to frameshift (Table 2). Therefore, on rare occasions, it appears that mutations that abrogate CovS activity occurred late in the disease course for strain M3-CovS-T284A which likely accounted for the increase lesion size described above. As observed with the M3 strains, no GAS with additional covRS mutations were recovered from animals infected with strains M1-CovR-D53A and M1-CovS-E281A. Interestingly, however, unlike its serotype M3 counterpart, no additional mutations were found in GAS isolated from mice infected with strain M1-CovS-T284A. In contrast, a high number of colonies recovered from animals inoculated with the M1-WT isolate had mutations in the covRS systems as has been previously reported for this strain [25, 40] (Table 2). Many of the recovered strains had changes that truncated the CovS protein, whereas several colonies had non-synonymous SNPs in covR (A81T in CovR) or covS (R241S or P285S in CovS) (Table 2). No additional mutations were detected in any GAS strain of either serotype during nasopharyngeal mouse challenge or during growth in whole human blood. Thus, elimination of phosphatase activity by the CovS-T284A change abrogated the emergence of covRS mutations in the M1 background during skin/soft tissue infection but increased such emergence in the M3 strain, albeit primarily late in the infection course.
It has previously been shown that truncations in CovS mimic a covS deletion strain (e.g. reduced CovR~P levels), but much less is known about the consequences of non-synonymous single nucleotide polymorphisms (SNPs) in either covR or covS that arise during mouse challenge or human infection [38, 41]. Thus, we next sought to evaluate the effect of some of the previously uncharacterized SNPs isolated during our mouse infection study by generating the isoallelic GAS strains CovR-A81T, CovR-R66H, CovR-L155I, and CovS-P285A in the serotype M3 background (Fig 4). CovR~P levels in strains CovR-A81T and CovS-P285S were strongly reduced compared to the wild type and resembled that of a covS deletion strain. In contrast, strains CovR-R66H and CovR-L155I had CovR~P levels similar to the wild type (Fig 4A). SpeB is an actively secreted broad-spectrum protease whose production is abrogated by CovS inactivation [42, 43]. In accordance with the CovR~P levels, strains CovR-A81T and CovS-P285S had reduced SpeB activity on milk plates, while SpeB activity was not affected in CovR-R66H or CovR-L155I (S1 Fig). Next, we performed TaqMan qRT-PCR of various known CovR-regulated genes that have previously shown to be regulated by CovR via different mechanisms [44] to evaluate the effect of the mutations on CovR-mediated transcription regulation (Fig 4B–4D). Consistent with the CovR~P level analysis, in strains CovR-A81T and CovS-P285S transcript levels of spyM3_0105, prtS, sagB, and cbp (which encode a cell surface protein, an IL-8 degrading protease, a pore-forming toxin, and a pilus protein, respectively) resembled that of a covS deletion strain. Specifically, the transcript levels of spyM3_0105 and prtS were >10 or ~2-fold elevated, respectively, compared to M3-WT (Fig 4B and S2 Fig), while transcript levels of sagB and cbp did not differ significantly compared to M3-WT (Fig 4C and 4D) (P > 0.05). Interestingly, despite CovR~P levels being similar to the wild type, repression of all genes studied was strongly relieved in strains CovR-R66H and CovR-L155I compared to the wild type (Fig 4B–4D) implying that functions aside from CovR phosphorylation (e.g. signal transduction, DNA binding) are affected in these mutants (P < 0.05 for all gene transcript levels).
Next, we determined the transcriptomes of the eight strains used in the animal challenges to obtain mechanistic insights into the observed virulence differences. To this end, four biological replicates per strains were grown to late-logarithmic phase (OD = 0.9), and RNA was extracted and subjected to RNAseq analysis. Transcript levels were considered significantly different if the mean transcript level difference was ≥ 2.0 fold and the final adjusted P value ≤ 0.05 compared to the wild-type strain. By principle component analysis (PCA) biological replicates of each strain clustered together (Fig 5A and 5B). Consistent with the higher CovR~P levels in M1-WT compared to M3-WT, there were more genes differentially regulated compared to the wild type in strains M1-CovR-D53A and M1-CovS-E281A and less genes in M1-CovS-T284A than in the respective M3 serotype strains (Table 3). That is, the number of differentially regulated genes paralleled the differences in CovR~P between the wild type and the distinct isoallelic strains. In accordance with our virulence data, the transcriptomes of M3-CovR-D53A and M3-CovS-E281A were highly similar with 100 and 97 genes being differentially regulated compared to the wild type (Fig 5A, Table 3). Genes up-regulated compared to M3-WT included the known virulence genes prtS and speA (encoding a pyrogenic exotoxin). The transcriptomes of M1-CovR-D53A and M1-CovS-E281A were also similar. However, compared to their M3 counterparts, a larger number of known virulence factor encoding genes were up-regulated in these strains relative to M1-WT including the has operon, nga (NAD glycohydrolase), slo (streptolysin O), and speC (exotoxin). Moreover, transcript levels of several DNA binding proteins and genes involved in amino acid and sugar transport and metabolism were increased compared to the wild type.
In contrast, transcript levels of 41 and 23 genes were decreased in strains M3-CovS-T284A and M1-CovS-T284A compared to the wild type, respectively. Remarkably, genes further repressed in the CovS-T284A strains comprised those encoding nearly the complete repertoire of known GAS virulence factors ranging from secreted toxins to immune-modulating surface proteins (see Fig 5 and S2 Table), which likely explains the striking reduction in virulence of CovS phosphatase deficient GAS strains seen in our infection studies. Many genes with reduced transcript levels in the CovS-T284A strains have not been previously identified as part of the CovRS-regulon including mga, which encodes the multi-gene activator of numerous virulence genes, Mga, and part of its regulon, e.g. emm coding for M protein or grm (gene regulated by Mga) [45]. On the other hand, with the exception of the dpp operon (encoding a dipeptide permease) we did not identify genes involved in metabolism and transport being further repressed in the CovS-T284A strains, suggesting specific modulation of virulence gene expression in these strains. Thus, results of our transcriptome analyses can explain the hypervirulent phenotype of CovS phosphatase deficient strains by identifying (in part novel) CovR-mediated repression of a broad array of virulence factor encoding genes at high CovR~P levels.
We further used our multi-strain, multi-serotype RNAseq data to differentiate three broad classes of CovR-regulated genes depending on their repression (T284A) and de-repression (E281A/D53A) profile (see examples in Fig 5C–5E). The first class (class I) encompasses genes that were de-repressed in the E281A and D53A strains in both serotype M1 and M3. These genes have been identified as part of the CovRS regulon by previous transcriptome studies [10, 17, 22, 25, 28, 38] and include well-known virulence genes like the has operon and prtS. Interestingly, these genes were only further repressed in the T284A strain compared to parent strain M3-WT while they were already fully repressed in M1-WT (Fig 5C). In contrast, we defined class II genes as those, whose transcript levels were affected in the E281A/D53A strains only in the M1-WT background but showed repression in the T284A strains for either serotype. These genes included nga/slo and scpA (C5 peptidase) (Fig 5D). As mentioned earlier, our transcriptome analysis identified novel virulence factor repression in both CovS-T284A strains. These genes, categorized as class III genes, were not increased in the CovS-E281A and CovR-D53A strains and have not previously been identified as part of the CovRS regulon in serotype M1 and serotype M3 GAS strains (Fig 5E). Additionally, transcript levels of several prophage-encoded, serotype-specific virulence factors like speK (exotoxin), sla (extracellular phospholipase) or sdaD2, spd3, and sdn (all secreted DNases) were reduced in the CovS-T284A mutant strains (Fig 5F). Due to their serotype-specific nature these genes could not be unambiguously assigned to one of the described major classes. Extracellular DNases have been shown to contribute to GAS pathogenesis [46]. To gain insight into the physiological consequences of our transcription data, we performed DNase activity tests and found that indeed DNase activity was significantly reduced in supernatants derived from CovS-T284A cells compared to that of the respective parental strains (S3 Fig).
Next, we employed TaqMan qRT-PCR to confirm the transcript level pattern of selected genes exemplifying the different classes of CovR-regulated genes revealed by RNAseq (Fig 6 and S4 Fig). To enable a better comparison of transcript levels vs. CovR~P in both serotypes, we also included strain M1-WT grown in THY supplemented with 100nM LL-37 (CovR~P is ~40% as for M3-WT) in our analyses. The qRT-PCR results were in concert with the RNAseq data. By contrasting the level of CovR~P with the degree of transcriptional regulation we were able to deduce differences in CovR~P dependency for the regulation of distinct promoter classes. The transcript levels of class I genes, exemplified by prtS (Fig 6B) and hasA (S4B Fig), revealed differential regulation over a range of 20–70% CovR~P, above which no further repression was detectable. In contrast, gene regulation of class II genes, such as slo (Fig 6C) and scpA (S4C Fig), was not affected by varying CovR~P below a level of 40%, but was dramatically impacted when CovR~P was increased from 40% to beyond 70%. Mga (Fig 6D) and emm (S4D Fig), as examples for class III genes, showed the highest CovR~P dependency with CovR-mediated repression only observed at CovR~P above 70%.
Mga and its regulated genes have previously not been identified as part of the CovR regulon in serotype M1 or M3 GAS [17, 25]. Our transcriptome data, however, showed that the expression of mga and several Mga-controlled genes such as emm or grm were down-regulated in M3-CovS-T284A and M1-CovS-T284A compared to the respective parental strain (Fig 5E and Fig 6D). Given the presence of potential CovR-binding sites (ATTARA) within the mga promoter, we next performed electrophoretic shift mobility analyses (EMSA) to address the question whether CovR binds to the promoter of mga and whether this binding is dependent on the phosphorylation status of the protein. A PCR fragment of ~500bp encompassing the mga promoter amplified from M1-WT genomic DNA (DNA sequences of mga promoter regions from M1-WT and M3-WT have 94% nucleotide identity) was incubated with increasing concentration of unphosphorylated or in vitro phosphorylated purified CovR protein, and samples were separated on a TBE-PAA gel. Although unphosphorylated CovR was able to bind the mga promoter DNA to create a low molecular weight complex, increasing protein concentrations up to 5 μM did not appreciably change the binding behavior (Fig 7A). By contrast, increasing concentrations of CovR~P progressively resulted in complexes of higher molecular weight as typically seen in CovR/CovR~P promoter binding assays of genes known to be directly regulated by CovR [22, 28] (Fig 7B). In accordance with our transcription data, high concentrations of CovR~P were needed for effective binding of the mga promoter.
Although phosphorylation of response regulator proteins is critical for bacterial pathogenesis, there remains limited understanding of how variation in response regulator phosphorylation impacts bacterial virulence at diverse infection sites. Herein we employed transcriptome and virulence assays of an array of isoallelic serotype M1 and M3 GAS strains to assess the impact of multiple, distinct phosphorylation levels of the key response regulator CovR on GAS pathophysiology.
For our virulence assays, we chose the mouse models of skin/soft tissue infection and nasopharyngeal colonization. Previous studies on the contribution of CovRS to GAS infection using these models have come to varying conclusions [20, 25, 26, 35, 36], which may be explained by serotype-specific differences, the use of strains with diverse inactivated regulators with potentially additional functions (ΔcovS vs. ΔrocA) or possible emergence of hyper-virulent clones during infection. In contrast, all studies to date using the intraperitoneal (i.e. bacteremia) model had consistently found increased GAS virulence with decreasing CovR~P levels [5, 21, 22, 25, 34]. Confirming many previous observations, our assays indicate both serotype- and site-specific effects on the impact of CovRS inactivation on GAS pathogenesis. Similar to results obtained by Dalton et al. [35], we saw an inverse correlation between CovR~P levels and virulence of our M3 strains in the skin/soft tissue model. This result was not mirrored by our M1 strains. However, the hypervirulent phenotype of M1-WT in our mouse skin/soft tissue model likely stems from the emergence of CovS inactivating, SpeB- mutations early in the infection course. It has been shown that a mixture of wild type and covS-inactivated M1 GAS produces larger skin lesions compared to a covS deletion strain [26], consistent with skin-specific increase in virulence of wild-type vs. covS-inactivated M1 GAS described by Sumby et al. [25]. Although inactivating CovS is believed to reduce GAS fitness in the upper respiratory tract [36, 47], we observed a profound increase in nasopharyngeal colonization rates for strains with lower CovR~P levels for both serotypes. Despite the limitations of the nasopharyngeal mouse model, a similar study using an M18 strain found that increasing CovR~P levels by repairing a naturally occurring RocA mutation also decreased colonization rates [20]. While previous studies solely analyzed the effect of lowering CovR~P on GAS infectivity, the most striking conclusion of our virulence data was the consistent hypovirulent phenotype of the CovS-T284A strains. In both M1 and M3 GAS serotype, these CovS phosphatase deficient strains caused the smallest lesions in the skin/soft tissue model, had a strongly reduced capacity to colonize the mouse oropharynx and did not survive neutrophil killing in whole human blood. Thus, CovS phosphatase activity strongly influences GAS overall ability to cause disease.
Another key finding from our animal studies was the effect of CovR~P levels on emergence of GAS strains with CovRS inactivating mutations. Consistent with the concept that the primary selective pressure for CovRS mutations is to decrease CovR~P levels, we only observed mutations in strains whose initial CovR~P levels were ≥ 70%. The absence of additional mutations in the E281A strains suggest that there is no selection pressure to reduce CovR~P beyond ~20% and is in accordance with human data that GAS strains with CovR inactivating mutations are rare compared to CovS. Surprisingly, increasing CovR~P levels via the T284A mutations did not evoke high levels of CovRS mutations. The latter occurred late in the infection course in a very limited number of animals infected with M3-CovS-T284A and not at all in mice infected with M1-CovS-T284A. Given the small lesion sizes, we speculate that the profound hypovirulence induced by the T284A mutation inhibited such emergence.
Our transcriptome analysis offers an explanation for the hypovirulent phenotype by revealing specific down-regulation of nearly the entire repertoire of virulence factor encoding genes in the CovS-T284A strains. Interestingly, plasmid-derived overexpression of RocA in M1-WT produced a similar virulence gene repression profile [39]. Given our finding that RocA impairs CovS phosphatase activity (Fig 3), we speculate that RocA overexpression results in CovR~P levels similar to the CovS-T284A strains. Among the identified novel CovR-repressed virulence genes in the CovS-T284A strains were mga and Mga-regulated genes. Mga is a well-known activator of numerous GAS virulence factors, such as M protein, and is thought to be particularly important in the early stages of infection [45, 48]. Thus it is likely that decreased mga expression played a pivotal role in the observed hypovirulence of our CovS phosphatase deficient strains. In addition, down-regulation of the Mga regulon has been shown to prevent in vivo selection of hypervirulent SpeB negative covRS variants [32], and thus the low mga and emm transcript levels in the CovS-T284A strains may have negatively affected hypervirulent isolate emergence. Although an indirect relationship between CovR and Mga is possible [49], our DNA binding analysis suggests that CovR could directly regulate mga. In accordance with this, we found several potential CovR-binding sites (ATTARA) within the mga promoter region, in particular an ATTARA sequence directly upstream of the P2 and downstream of the P1 promoter [50] as well as one partially overlapping a CodY binding site [51]. Nonetheless, in serotype M1 and M3 GAS, a functional protein-DNA complex seems to be only achieved at high CovR~P levels as seen in the CovS phosphatase deficient strains.
Our multi-strain, multi-serotype approach during our transcription level analysis allowed us to distinguish three distinct classes of CovR-regulated genes on the basis of their CovR~P dependency for gene regulation. In addition, we have previously described a group of CovR-regulated genes (e.g. sagB, cbp, covR), whose transcription regulation is independent of CovS [14, 44]. Hence, for this group, which we designate as class 0 in this context, CovR~P of only 20% is sufficient to repress gene expression. Together our data suggests classes of promoters repressed under distinct CovR~P concentrations increasing from class 0 to class III. This CovR~P dependency is likely determined by a combination of different affinities for CovR binding sites (as suggested by Jain et al. [39]) and diverse DNA-binding mechanisms (cooperativity, CovR oligomerization state, or interaction with other regulators or RNA polymerase) [6, 12, 44, 52–55] and requires further investigations. Regardless, the gradual repression of promoter groups establishes the basis for coordinated expression of GAS virulence factors in response to changing environmental cues.
Adjusting gene expression in adaptation to environmental niches is pivotal for pathogenic bacteria. Recently, this function has been increasingly attributed to the phosphatase activity of bi-functional HisKA-family kinases [56–58]. Thus, besides limiting crosstalk between homologous TCS [59], histidine kinase phosphatase activity evidently fulfills an important role in sensing extracellular signals. The currently established environmental signals that modulate the function of CovRS TCS likewise seem to target CovS phosphatase rather than kinase activity. Previous investigations revealed that inactivation of CovR by CovS is required for survival of GAS under stressful conditions such as the presence of LL-37, iron starvation or acidic stress suggesting that stress signals activate CovS phosphatase activity in M3 or M6 strains [60–62]. Mg2+ reduces CovS phosphatase activity by an unknown mechanism thereby increasing intracellular CovR~P [14]. Herein we show that the presence of accessory protein RocA in MGAS2221 also diminishes CovS phosphatase activity. We speculate that RocA forms a hetero-oligomeric complex with CovS thereby stabilizing a CovS phosphatase incompetent conformation (activated state) [63]. Further, we show that the antimicrobial peptide LL-37 in turn increases CovS phosphatase activity. LL-37 has been demonstrated to bind directly to CovS [64] but its effect on CovR~P and transcription regulation is only observed in GAS strains expressing a functional RocA protein. Thus, we hypothesize that LL-37 increases CovS phosphatase activity indirectly by displacing RocA from the hetero-oligomeric complex to allow formation of a CovS phosphatase competent conformation. The antagonism between Stk mediated phosphorylation of CovR T65 and CovS phosphorylation at D53 adds an additional layer of complexity to regulation of CovR function [22]. The multi-faceted regulation of CovS phosphatase activity highlights its crucial function in adjusting CovR~P status and thus the expression of CovR-controlled virulence genes.
Bacterial TCSs have been proposed as potential therapeutic targets [65–71]. General inhibition of CovS function is unlikely to be desirable in GAS given the hypervirulence of ΔcovS strains. However, the data presented in this study suggests that specifically targeting CovS phosphatase activity might be promising. CovS is present in all GAS serotypes, and abolishing CovS phosphatase activity markedly reduced GAS virulence in all three infection models. Notably, hypovirulence was even detected in serotype M1 GAS, a strain with low intrinsic CovS phosphatase activity and therefore high baseline CovR~P. Further, phosphatase activity of other bifunctional HisKA-family histidine kinases has been shown to play an important role in regulating infectivity of both Gram-positive and Gram-negative pathogens. For example, Liu et al. demonstrated that a T247A mutation (homolog to CovS-T284A) of Salmonella enterica EnvZ, mimicking a conserved pH-controlled mechanism of HK phosphatase activity ablation, increased macrophage infectivity due to accumulation of OmpR~P and downstream activation of ssrA-ssrB genes [58]. Mutations of WalK PAS domain that modulate WalK phosphatase activity also attenuated virulence of S. pneumoniae in a murine infection model [72]. The Spinola group has suggested the use of phosphatase inhibitors towards histidine kinase CpxA to treat certain urinary tract infections caused by uropathogenic E. coli [73, 74]. These examples corroborate the importance of HK phosphatase activity in bacterial virulence and suggest a broader application of this approach with regards to other histidine kinases in pathogen bacteria.
Together our study provides novel insights into mechanisms of GAS virulence factor regulation and establishes an important role of CovS phosphatase activity in controlling GAS pathogenicity.
This study was carried out in strict accordance with the recommendations in the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health. Protocol #00001455-RN00 was approved by The University of Texas MD Anderson Cancer Center Institutional Animal Care and Use Committee. All efforts were made to minimize suffering. Human blood samples were drawn and used under IRB protocol #0110–0015 approved by Houston Methodist Research Institute Review Board. Written informed consent was obtained from all donors.
All GAS strains used in this study are derivatives of either strain MGAS10870 (serotype M3, herein called M3-WT) or strain MGAS2221 (serotype M1, herein called M1-WT), two clinical isolates that are known to have wild type covRS sequence (see Table 1). Primers for strain creation are listed in S1 Table. Single nucleotide exchanges were introduced into chromosomal DNA of strain M1-WT or M3-WT via homologues recombination using the integrative plasmids pBB740 or pJL1055, respectively, as described in [28] to create isoallelic strains that differ only by the presence of a single amino acid replacement in CovR or CovS. GAS strains were statically grown in Todd Hewitt broth supplemented with 0.2% yeast (THY) at 37°C and 5% CO2. When appropriate, chloramphenicol was added to 5μg/ml. Bacteria were plated on tryptic soy agar with 5% sheep blood (BSA) plates.
Recombinant CovR was purified and phosphorylated as described [28] and served as control. GAS lysates were prepared and separated on 12.5% SuperSeq Phostag gels (Wako, USA), and un/phosphorylated CovR species were detected using a polyclonal anti-CovR antibody and the Odyssey imaging system as described previously [14]. Independent Western blots were repeated at least twice.
Strains were grown in 200 ml THY to mid-exponential phase and harvested by centrifugation at 9000 rpm. Cell pellets were washed twice with ice-cold PBS buffer, re-suspended in 4 ml PBS/20% glycerol solution and stored in aliquots at -80° C until use. CFU counts for each strain were determined by plating dilutions of the samples on BSA plates at least three times and confirmed after mice inoculation. All animal experiments were performed in a blinded fashion.
Whole blood was drawn in sodium heparin tubes (Becton Dickinson) from three consented, healthy, non-immune donors under IRB protocol #0110–0015. GAS growth in blood was performed as described in [21]. Indicated strains were grown in THY as described for animal experiments. 20–100 CFU of each GAS strain was inoculated in 300μl human blood containing 10% THY, respectively. Samples were incubated for 3h at 37°C in 5% CO2 with end-to-end rotation. CFU/ml were determined by plating serial dilutions in PBS on BSA plates for enumeration of β-hemolytic colonies. Multiplication factors were determined by dividing CFU/ml after 3h incubation by CFU/ml in the inoculum. The experiment was performed in triple biologic replicates on two separate occasions. Data were analyzed using one-way ANOVA followed by post-hoc analysis using Tukey’s correction for multiple comparisons.
The DNase activity in filtered culture supernatants from M1 and M3 wild type and CovS-T284A strains grown to late-exponential phase (OD = 0.95) was assayed. To this end, 100ng PCR-derived GAS DNA was incubated for 20 min at 37°C in 1x NEB 2 buffer (New England Biolabs) with 0.5μl of the respective supernatant. The remaining DNA was quantified using Quant-IT Pico Green dsDNA reagent (Thermo Fisher Scientific) according to the manufacturer’s instructions. Fluorescence was detected using a fluorescence microplate reader with 480 nm excitation and 520 nm emission wavelength. DNA concentrations were calculated using a standard curve with known DNA concentrations. Three biological replicates were assayed on two separate occasions (n = 6).
The ~500bp encompassing mga promoter region was amplified from M1-WT genomic DNA by PCR using primers listed in S1 Table. The mga promoter region from M1-WT and M3-WT share 94% sequence identity, such that results from our EMSAs using serotype M1 DNA are likely applicable to serotype M3. Purified PCR product was incubated in TBE-buffer with the indicated amount of CovR or CovR phosphorylated with acetyl phosphate at 37°C for 15 min as described [28]. Subsequently, samples were separated on a 5.5% TBE-PAA gel for 2h at 120V and stained with ethidium bromide.
Per strain and experiment, at least 25 colonies were picked, and the complete covRS operon was PCR amplified and sequenced using the primers listed in S1 Table. Sequences were analyzed with Sequencher 5.4.6 using the covRS sequence of M3-WT or M1-WT, respectively, as template.
Strains were grown in quadruplicate to late-exponential phase in THY. RNA was isolated using the RNeasy minikit (Qiagen). RNA sequencing was performed at the MD Anderson Sequencing and Microarray Facility, and data analysis was performed as described using the M3-WT and MGAS5005 (an M1 strain) genome, respectively [22]. A total of 88 out of 1853 (4.7%) genes (M3 serotype) and 73 out of 1849 (4%) genes (M1 serotype) were excluded from the analysis due to low expression levels. Transcript levels were considered significantly different between the isoallelic and wild-type strains if the mean transcript level difference was ≥ 2.0 fold and the final adjusted P value was ≤ 0.05. Transcriptome data have been deposited in the GEO database under accession number GSE121313.
Strains were grown as described for RNA seq. Approximately 300 ng RNA per sample was converted to cDNA using a high-capacity reverse transcription kit (Applied Biosystems). TaqMan quantitative real-time PCR (qRT-PCR) was performed on an Applied Biosystems Step-One Plus system as described [22] using primers and probes listed in S1 Table. Two biological replicates were performed on two separate occasions (n = 4). Transcript levels between wild type and derivative strains were compared using a two-sample t test (unequal variance) with a P value of ≤0.05 following Bonferroni’s adjustment for multiple comparison and a mean transcript level of ≥ 2.0-fold change being considered as statistically significant different.
|
10.1371/journal.pgen.1000208 | Mutability and Importance of a Hypermutable Cell Subpopulation that Produces Stress-Induced Mutants in Escherichia coli | In bacterial, yeast, and human cells, stress-induced mutation mechanisms are induced in growth-limiting environments and produce non-adaptive and adaptive mutations. These mechanisms may accelerate evolution specifically when cells are maladapted to their environments, i.e., when they are are stressed. One mechanism of stress-induced mutagenesis in Escherichia coli occurs by error-prone DNA double-strand break (DSB) repair. This mechanism was linked previously to a differentiated subpopulation of cells with a transiently elevated mutation rate, a hypermutable cell subpopulation (HMS). The HMS could be important, producing essentially all stress-induced mutants. Alternatively, the HMS was proposed to produce only a minority of stress-induced mutants, i.e., it was proposed to be peripheral. We characterize three aspects of the HMS. First, using improved mutation-detection methods, we estimate the number of mutations per genome of HMS-derived cells and find that it is compatible with fitness after the HMS state. This implies that these mutants are not necessarily an evolutionary dead end, and could contribute to adaptive evolution. Second, we show that stress-induced Lac+ mutants, with and without evidence of descent from the HMS, have similar Lac+ mutation sequences. This provides evidence that HMS-descended and most stress-induced mutants form via a common mechanism. Third, mutation-stimulating DSBs introduced via I-SceI endonuclease in vivo do not promote Lac+ mutation independently of the HMS. This and the previous finding support the hypothesis that the HMS underlies most stress-induced mutants, not just a minority of them, i.e., it is important. We consider a model in which HMS differentiation is controlled by stress responses. Differentiation of an HMS potentially limits the risks of mutagenesis in cell clones.
| Mutational processes are being discovered in which bacterial, yeast, and human cells under various stresses activate programs that increase mutagenesis, often under the control of cellular stress responses. These programs may potentially increase genetic variability in populations specifically when they are maladapted to their environments, i.e., when they are stressed. When mutation supply is limiting for evolution (for example, in small populations), these mechanisms might enhance the intrinsic ability of organisms/cells/populations to evolve, specifically during stress. Stress-induced mutagenesis mechanisms recast understanding of, and strategies for combating, problems such as host-pathogen interactions, generation of bacterial antibiotic resistance, cancer progression, and evolution of chemotherapy resistance, all problems of evolution of fitter variant clones fueled by genetic change under stress. A key problem in stress-induced mutagenesis concerns how cells survive the deleterious effects of enhanced mutagenesis. One proposed strategy is the differentiation of a subpopulation of transiently hypermutable cells. This study investigates a previously discovered hypermutable cell subpopulation (HMS) postulated either to underlie most stress-induced mutagenesis in E. coli or only a small fraction of it. First, improved methods allow estimation of mutations per genome accumulated during HMS-generated bursts of mutagenesis and show numbers compatible with fitness after the HMS state. Second, two lines of evidence presented support models in which the HMS is central to this stress-induced mutagenesis pathway. Third, a specific model, with general consequences, for HMS differentiation is discussed.
| Stress-induced mutational processes are responses to growth-limiting environments whereby mutations are produced at an accelerated rate, some of which may confer a growth advantage. The study of stress-induced-mutagenesis mechanisms is expanding our understanding of genome instability and cellular and organismal adaptability to environmental challenges (reviewed [1],2). Whereas classical spontaneous mutations occur in proliferating cells, in a generation-dependent manner, and before cells encounter an environment in which the mutations might prove useful [e.g.],[3],[4,5], stress-induced mutations occur in growth-limiting environments, often under the control of stress responses, via pathways different from those observed in rapidly proliferating cells (reviewed [2]). Stress-induced mutagenesis may potentially accelerate evolution specifically when cells/organisms are maladapted to their environments, i.e., when they are stressed. Stress-induced mutagenesis mechanisms appear to be widespread and important in nature. The vast majority of 787 natural isolates of E. coli show induction of mutagenesis by starvation stress [6]. Stress-induced mutagenesis mechanisms present appealing models for mutagenesis underlying evolution of antibiotic resistance, evasion of the immune response by pathogens, aging, and for genomic instability underlying tumor progression and resistance to chemotherapeutic drugs, all of which are fueled by mutations and occur in stress-provoking environments (reviewed by [2],[7]).
There are multiple molecular mechanisms of stress-induced mutagenesis, observed in different organisms, strains and stresses, but many share important common elements, including control by cellular stress responses (reviewed [2]). In the Escherichia coli Lac assay [8], the mechanism of mutagenesis is a stress-response-controlled switch from high-fidelity to error-prone DNA double-strand-break repair during stress, described below. In the Lac assay, cells carrying a chromosomal deletion of the lac genes and a lac +1 frameshift allele in an F' episome are starved for carbon on solid minimal lactose medium. Over time, Lac+ mutant colonies appear. Many of those visible at day two carry generation-dependent spontaneous mutations that occurred during growth of the culture prior to plating (the Lac+ mutants take two days to form a colony on the selective medium) [8]. Colonies that appear on later days are stress-induced mutants, which form after exposure to starvation-induced stress [8],[9] in a process requiring the RpoS general- or starvation-stress response [10],[11].
At least two independent mechanisms produce the stress-induced Lac+ colonies: Lac+ “point mutagenesis” and stress-induced gene amplification. Point mutagenesis dominates during the first week of incubation and creates compensatory -1 frameshift mutations [12],[13]. Tandem amplification of the lac region to 20–100 copies represents ∼40% of Lac+ colonies at day eight of incubation, and increases thereafter [14],[15]. Amplification allows growth on lactose because multiple copies of the weakly functional mutant lac gene produce enough beta-galactosidase activity to restore growth. Both processes are stress-induced and require the RpoS-controlled general-, stationary-phase- or starvation-stress response [11]. This paper focuses on the mechanism of stress-induced point mutagenesis. Readers are referred to [16]–[18] for recent reviews of the mechanism(s) of stress-induced amplification, and its relevance to genome instability in cancer as well as copy-number variations ubiquitous in human and other genomes.
Work from our lab has provided support for a model in which stress-induced point mutagenesis results from DNA polymerase errors made during acts of DNA double-strand-break repair (DSBR), which is switched to a mutagenic mode, using an error-prone DNA polymerase, specifically during stress [19]:
First, point mutagenesis requires homologous-recombinational (HR)-DSBR proteins RecA (EG10823), RecBC (EG10824 and EG10825), RuvA (EG10923), RuvB (EG10924), and RuvC (EG10925) [20]–[22], and it and stress-induced amplification are greatly stimulated by DSBs made using a regulatable I-SceI (P03882) endonuclease in vivo [19]. Induction of I-SceI cuts next to lac increases mutation rate over 1000-fold; whereas I-SceI-induced DSBs made in another molecule provoke lac reversion only 3-fold. However the DSBs made in a different molecule from lac can again stimulate lac reversion dramatically if one end of the broken DNA molecule contains DNA identical to DNA next to lac, such that homologous repair with the lac region can be initiated [19].
Second, I-SceI-stimulated stress-induced Lac+ point mutagenesis occurs by the same mechanism as “normal” stress-induced mutagenesis in that both require the HR-DSBR proteins [19], the RpoS general-stress-response transcriptional activator [10],[11],[19], induction of the SOS DNA-damage response [19],[23], and functional dinB (EG13141), encoding DNA polymerase (Pol) IV [19],[24] of the Y-superfamily of trans-lesion, error-prone DNA polymerases [25]. These specialized DNA polymerases insert bases opposite otherwise replication-blocking lesions in DNA with reasonably good fidelity, but have low fidelity and are error-prone when synthesizing on undamaged template DNA. Both the SOS response [26],[27] and the RpoS response [10] upregulate dinB, 10- and about 2-fold, respectively. dinB upregulation might account for some or all of the requirement for induction of the SOS and RpoS responses for stress-induced point mutagenesis, though this has not been demonstrated.
The similarity of the proteins required for I-SceI-stimulated and “spontaneous” stress-induced mutagenesis argues that both occur by the same mechanism, as does the finding that I-SceI-induced and “normal” stress-induced Lac+ point mutations are indistinguishable in their Lac+ mutation sequences [19]. All of these data support the idea that stress-induced mutagenesis occurs via error-prone HR-DSBR in which DinB/Pol IV has been licensed to participate in the HR-DSBR reaction [19].
Finally, HR-DSBR is not always mutagenic but rather switches to a mutagenic mode, with DinB/Pol IV participating, under stress. This switch is controlled either by entry of cells into the stationary phase, or, in log-phase cells if the RpoS stationary-phase stress-response transcriptional activator is expressed inappropriately [19]. In both cases, the SOS response should often already be induced by the DSB, given that even well repaired DSBs induce SOS efficiently [28]. (Alternative models for stress-induced Lac point mutagenesis are discussed below.)
Thus, mutagenesis is limited to times of stress via its coupling to two stress responses (SOS and RpoS). Mutagenesis is potentially also restricted in genomic space via being coupled to potentially localized DNA synthesis during DSBR [19]. Both of these restrictions may protect populations from deleterious effects of mutagenesis, and both themes are evident in many different mutagenesis mechanisms in organisms from phage to human, and so appear to be general mutational/evolutionary strategies [reviewed, 2, and Discussion].
In this paper, we investigate a third level of restriction/limitation or regulation of mutagenesis: its limitation to a subpopulation of stressed cells while the main population appears to be unaltered. In the Lac system, there is strong evidence that a subpopulation of cells becomes transiently hypermutable, resulting in mutations in genes throughout the genome. First, E. coli [29]–[31] and Salmonella [32] Lac+ stress-induced point mutants show, respectively, ∼20 and ∼50 times more loss-of-function mutations in chromosomal genes throughout their genomes than are found in Lac− cells that starved for the same length of time: their Lac− neighbors from the same Petri plates. Those Lac− cells represent the main population whereas some or all of the Lac+ mutants arose from a more mutable subpopulation: a hypermutable cell subpopulation (HMS). The evidence that the hypermutability of this HMS is transient is, second, that once the cells have become Lac+, they do not have elevated spontaneous [29]–[31] or stress-induced [33] mutation rates. Moreover, when whole colonies of the initial stress-induced Lac+ mutants were picked and analyzed these colonies were mostly pure, not mosaic, for the unselected mutations that they carried, indicating that they accrued the unselected chromosomal mutations during or before acquiring the Lac+ mutation, not after, further showing that the mutability was transient [29]. The possible evolutionary significance of differentiation of a HMS is that this may protect most members of a clone from the deleterious effects of inducing mutagenesis, an advantage should nutrients suddenly become available, while simultaneously allowing the exploration of evolutionary space when maladapted to an environment.
Although there is consensus in the field regarding the existence of the HMS, both the extent of HMS-cell mutagenicity and the importance of the HMS to most stress-induced mutagenesis are currently unresolved. First, the HMS could either be important or not. On the one hand, the HMS has been hypothesized to give rise to essentially all stress-induced Lac+ point mutants [29], whereas on the other hand, other models suggest that the HMS may contribute to only a small minority, ∼10% or so, of Lac+ point mutants [30],[32], and so be relatively unimportant. Second, it has been argued that too much mutagenesis would occur in the HMS state for it to be adaptive [34]. Here, we first estimate the number of mutations per genome in E. coli cells derived from the HMS and find a level that need not preclude fitness. Second, we provide two lines of experimental support and mathematical modeling that support the idea that the HMS generates most or all, not just a minority of, Lac+ stress-induced point mutants. Finally, we consider a model for a mechanism by which the HMS is differentiated.
To better understand the potential fitness impact of cells' entering into a transient hypermutable state, we wished to estimate the number of mutations expected per genome in cells that have undergone stress-induced mutagenesis. Numbers of unselected secondary mutations among Lac+ mutants are reported in previous studies, but were not used previously to estimate the numbers of mutations per genome. We used the previous data to estimate numbers of mutations per genome (Table 1 and Text S1), and we found that the answer differs between studies that used different organisms and methods for assaying unselected secondary mutations among the Lac+ stress-induced mutants. Whereas the data from three studies in E. coli [29],[30],[35] can be extrapolated to imply that about one unselected mutation cluster (of one or more mutations, discussed below) occurs per genome, in addition to the Lac+ mutation (Table 1/Text S1,), the data from a study using Salmonella enterica and a different mutation-assay method can be extrapolated to indicate about 2.5 unselected mutation clusters, in addition to Lac+, per genome (Table 1/Text S1). In the previous E. coli studies, the secondary mutations were detected by direct transfer of Lac+ colonies (either by replica-plating or patching) directly from the lactose-selection plates to specific indicator media that, for example, showed a different color colony for fermentation-defective mutants. This technique is likely to miss some mutants that are overlapped with wild-type colonies. By contrast, in the previous Salmonella study [32], the authors screened for auxotrophic mutations, using a more sensitive technique. They picked the Lac+ colonies and purified them by streaking, patched them into grids, grew, then replica-plated to media that would indicate auxotrophic mutations by failure of the patch to grow on medium lacking amino acids and bases. This technique is likely to produce fewer false negatives due to overlap of mutant with non-mutant colonies. To understand whether their somewhat different result arose from use of a different organism or the different mutation-detection method, we used their presumably more sensitive method with E. coli to improve estimates of unselected secondary mutations per stress-induced mutant genome.
First, we show that for E. coli, the purify-and-patch method is more sensitive than direct transfer by replica plating for three mutant phenotypes scored (Table 2). Second, using the purify-and-patch method for all of the results presented here, we observed 8/3437 (2.3×10−3) Mal− mutations per Lac+ cell (Table 1). If these occurred in 3178bp (Text S1), then we estimate 3.4 mutations or mutation clusters in addition to Lac+ per 4,639,221bp E. coli genome (Table 1). Third, we found 3/3437 (8.7×10−4) Xyl− mutants per Lac+ point-mutant colony, implying 2.2 mutation clusters in addition to Lac+ per genome (Table 1). Fourth, we assayed for auxotrophic mutations targeting 72 loci providing a mutation target of 28,920bp (Text S1). We observed 28/3437 (8.1×10−3) auxotrophic mutants per Lac+ point mutant (Table 1). This extrapolates to 1.3 mutations or mutation clusters in addition to that conferring Lac+ per E. coli genome (Table 1). These estimates per genome assume that all Lac+ stress-induced point mutants are equally likely to acquire secondary mutations. If only some do then the number of mutation clusters per genome would be higher in those that do (Discussion).
The somewhat higher estimates of secondary mutation clusters per genome in this study compared with those estimated from previous E. coli data (Table 1) is expected to reflect the more sensitive “purify-and-patch” method used here, but alternatively, might reflect the fact that the strains used here differ slightly from that used previously. Unlike the previously used strain, the present strains carry either the chromosomal PBADI-SceI-expression cassette (“Enzyme-only” strain) or the PBAD promoter replacing the phage lambda attachment site (attλ) in the chromosome (“PBAD-only” strain). These strains are negative-control strains for experiments presented below. In Table S1, we show that these slight strain differences are not the relevant variable. We assayed the PBAD-only strain for loss-of-function mutations among Lac+ revertants by direct transfer via replica plating straight from lactose plates onto indicator and selective plates as performed in [29]. We find no significant difference in the proportion of Lac+ mutants with secondary mutations from those previously reported, p = 0.697, (z-test with Yates correction) (Table S1). This rules out the unlikely possibility that the new strains used in this study might have shown enhanced secondary mutation for some reason specific to their genotype, and so confirms that the different mutation-assay method used here is responsible for the somewhat higher frequency of secondary mutations observed relative to previous E. coli studies [29],[30],[35].
Taken together, the data indicate between about one and 3.4 mutation clusters in addition to Lac+ per stress-induced-mutant cell genome.
Mutations in the Lac system appear to be clustered locally in the DNA [35] such that the estimates above are likely to pertain to numbers of mutation clusters per genome. We can make a rough estimate of the number of mutations per mutation cluster from data on the apparent clustering of Lac+ mutants with the linked mutations in the codAB genes (EG11326 and EG11327) 10kb from lac. Previously, loss-of-function mutations in the codAB genes, which confer resistance to the nucleotide analogue 5-fluorocytosine (5-FCR) were shown not to form independently of Lac+ mutations, whereas unlinked chromosomal mutations did, in a study using the direct-transfer-by-replica-plating method [35]. (Note that two E. coli loci confer 5-FCR when mutated, but only codAB mutations confer 5-FCR without also conferring resistance to 5-fluorouracil [5-FU], which is how these mutations were distinguished [35]). Here, we re-quantify coincident mutation of codAB and lac using the purify-and-patch method for detecting 5-FCR mutants (Table 3). We observe that 5-FCR (5-FU-sensitive) mutations in codAB are more frequent among Lac+ mutants than are unlinked mutations (Table 3, first two columns). These and the previous data [35] imply that codAB mutations cluster with lac mutations. In the Text S1, we estimate cluster size, and then estimate mutations per cluster from the data in Table 3, as about 1.67 mutations per cluster.
In models in which the HMS is predicted to produce only a 10% minority of the Lac+ stress-induced mutants, the mutations that occur in the HMS and give rise to Lac+ phenotype are proposed to occur via a different molecular mechanism from that that generates the 90% majority of stress-induced Lac+ mutations [30],[32]. If true, those Lac+ mutations that arise from HMS cells might be predicted to display different reversion-mutation sequences from the majority of stress-induced Lac+ mutations. We examined the Lac+ mutation sequences from stress-induced mutants that demonstrably descended from the HMS, as seen by their carrying an unselected “secondary” chromosomal mutation, and compared these with the published sequences of stress-induced Lac+ mutations [12],[13]. We sequenced a 250bp region spanning the +1 frameshift mutation of the lacI-lacZ (EG10525 and EG10527) fusion gene from 30 independent Lac+ point-mutant isolates carrying secondary mutations. We find that the mutation sequence profile is indistinguishable from those previously reported for stress-induced mutants [12],[13]: dominated by -1 deletions in small mononucleotide repeats with a hotspot at the position of the initial lac frameshift allele (Figure 1). These data support the hypothesis that the mechanism of mutagenesis in the HMS cells is similar to or the same as the stress-induced mutagenesis mechanism that generates all or most Lac+ point mutations. This distinctive mutation spectrum differs from spontaneous generation-dependent reversions of this lac allele, which are more heterogeneous [12],[13]. Summarized in Table S2, these include about half -1 deletions at mononucleotide repeats, and half carrying -1's not at repeats, 2–8 bp insertions, and large insertions and deletions. Instead, the stress-induced Lac+ frameshift-reversion sequences resemble the frameshift component of the error spectrum of DinB/Pol IV [36],[37] which is responsible for ≥85% of Lac+ point mutations in this assay system [24].
The previous demonstration that stress-induced mutations in the Lac system result from error-prone DNA double-strand-break repair (DSBR) and are greatly stimulated by creation of DSBs next to lac in vivo [19], allowed us to make a second test of whether the HMS underlies most stress-induced mutagenesis. In that study [19], DSBs generated near the lac gene by the endonuclease I-SceI were shown to increase Lac+ mutant frequency dramatically: more than 1000-fold above the levels seen in traI (P14565) endonuclease-defective mutants that cannot make nicks in the transfer origin of on the F', and more than 50-fold above levels in TraI+ cells (TraI-generated nicks usually promote mutations in this assay but are more than compensated for by I-SceI-generated DSBs [19]). Most importantly, the I-SceI-induced mutations occurred via the main mechanism of mutagenesis that operates normally (without I-SceI-induced DSBs), not a minority mechanism as shown by the following: the Lac+ sequences were the same; and the mechanism of mutagenesis with I-SceI induction specifically required RecA, RecB and Ruv DSB-repair proteins; DinB error-prone DNA polymerase; the RpoS transcriptional activator of the general stress response; and a functional SOS DNA-damage response, all of which are specifically required for the main mechanism of stress-induced mutagenesis in wild-type cells [19]. Therefore, stimulation of stress-induced mutagenesis by I-SceI cleavage increases the activity of the predominant, normal stress-induced-mutagenesis mechanism. We exploited this fact to examine whether this major increase in Lac+ mutagenesis by I-SceI cleavage of DNA near lac happens independently of the HMS, or inseparably from the HMS, by measuring the frequencies of chromosomal mutations among I-SceI-induced Lac+ mutants.
The idea is as follows: if only 10% of Lac+ mutagenesis were associated with secondary mutagenesis of unselected genes throughout the genome (proposed [30],[32]), and if I-SceI increased the efficiency of most stress-mutagenesis (proposed to form HMS-independently [30],[32]), then I-SceI-induction of stress-induced Lac+ mutagenesis would be expected to increase Lac+ mutagenesis without also increasing secondary mutagenesis of unselected genes throughout the genome (illustrated in Figure 2A, Model 1). I-SceI should “uncouple” Lac+ mutagenesis from secondary mutations such that the frequency of secondary mutations per Lac+ mutant should decrease (Figure 2A). On the other hand, if all stress-induced Lac+ mutagenesis occured in HMS cells [29],[35],[38], then the frequency of secondary mutations per Lac+ mutant cell should be unchanged (Figure 2B, Model 2). I-SceI cleavage might increase the size of the HMS (Discussion), but would not decrease its mutagenicity.
As seen previously [19], we found that a strain carrying both a regulatable chromosomal expression cassette of the I-SceI enzyme and its cutsite on the F' plasmid near lac showed a 70-fold increase in Lac+ mutation rate (Figure 3A,D) above that promoted by TraI-dependent DNA breaks at the transfer origin of the F' in the “wild-type” control cell. As previously, this was not seen in controls with only the enzyme expressed (no cutsite) or only the cutsite present (no enzyme) (Figure 3A,B,D). Lac+ point-mutant colonies from days four and five were assayed for unselected loss-of-function secondary mutations (Materials and Methods, and above).
First, we found that chromosomal loss-of-function mutations conferring inability to ferment maltose (Mal−), or xylose (Xyl−), or a mucoid-colony or auxotrophic phenotypes were not decreased among I-SceI-induced Lac+ point mutants as compared with negative-control strains that did not experience cleavage by I-SceI: the “enzyme-only” or “PBAD-only” controls (Figure 3E and Table 3). Thus, genome-wide mutagenesis was not uncoupled from Lac+ point mutagenesis (Figure 3E and Table 3) even though there was a 70-fold increase in mutagenesis caused by cleavage of DNA near lac by I-SceI (Figure 3A–D). This indicates that the main mechanism of Lac+ point mutagenesis does not occur independently of the HMS. This supports the hypothesis that Lac+ point mutagenesis is inseparable from the HMS (Model 2 of Figure 2B).
Second, there is a small but statistically significant increase in chromosomal secondary mutation frequencies among Lac+ point mutants accompanying I-SceI-mediated DNA breakage. This is discussed below (Discussion).
We assessed the possibility that the induction of I-SceI enzyme might be mutagenic in its own right and therefore might affect the proportion of chromosomal mutations independently of the formation of a DSB. We tested isogenic strains that lack the I-SceI cutsite, and either carry the chromosomal I-SceI-expression cassette Δattλ::PBADI-SceI (“Enzyme only”) or carry the chromosomal regulatable promoter without the I-SceI gene, Δattλ::PBAD (“PBAD only”), for secondary chromosomal mutations. The proportion of Lac+ point mutants carrying a chromosomal secondary mutation was no different for cells expressing I-SceI with no cutsite (enzyme only) compared with the PBAD-only strain, p = 0.697, (z-test with Yates correction) (Table 3). This demonstrates that I-SceI expression does not affect frequencies of chromosomal mutations unless an I-SceI cutsite is also present.
Previous work from our lab showed that cleavage of DNA near lac by I-SceI and repair of the break were not sufficient to increase stress-induced Lac reversion; in addition, the cells had to be either in stationary phase, or expressing the stationary-phase- (general- or starvation-) stress-response transcriptional activator protein RpoS (EG10510) (σS, a sigma factor for RNA polymerase) [19]. Thus, repair of DSBs is not always mutagenic, but becomes so when cells activate their RpoS stress response. As expected from this result, and from the finding that Lac+ and genome-wide secondary mutations are coupled (Table 3, Figure 3E), we found that Lac− unstressed cells do not show dramatically increased secondary mutation frequencies upon I-SceI induction (Table 4). Our results showing no secondary mutations among the 4000 Lac− unstressed cells assayed (Table 4) cannot distinguish whether secondary mutations were increased at all by I-SceI in unstressed cells, but do reveal that secondary mutations are not increased to levels seen among Lac+ mutants. That is, as expected, cleavage near lac with I-SceI is not sufficient to convert unstressed cells into HMS cells.
Second, perhaps surprisingly, we also found that not all Lac− stressed cells are converted into HMS cells upon I-SceI induction. Lac− stressed cells were recovered from the lactose selection plates by sampling agar from between visible Lac+ colonies at day three of incubation, re-suspended and plated on non-selective LBH rifampicin X-gal glucose medium. (Day-three starving cells correspond to day-five Lac+ colonies because colony formation on the lactose medium takes two days after acquisition of the Lac+ mutation [8],[9].) The colonies were then assayed for loss-of-function mutations conferring 5-FCR, Mal−, Xyl−, mucoid and auxotrophic phenotypes. Our results showing no secondary mutations among the 4000 Lac− stressed cells assayed (Table 4) show that secondary mutations are not increased to levels seen among Lac+ mutants. That is, even in starving cells, cleavage near lac with I-SceI apparently does not convert every cell into a HMS cell within the time-frame of an experiment.
Thus the elevated mutability observed among the DSB-induced Lac+ mutants is specific to a subpopulation of cells (i.e., an HMS) and induction of I-SceI-DSBs is not sufficient to render the whole population hypermutable.
The results presented here provide evidence supporting the hypothesis that a previously detected HMS [29]–[32] is important to the genesis of most stress-induced Lac+ revertants, not merely a small fraction as had been suggested [30],[32]. First, the unique sequence spectrum of the majority of stress-induced Lac+ reversion mutations was also observed in those Lac+ mutants demonstrably descended from the HMS: those carrying phenotypically-detectable secondary mutations in their genomes (Figure 1), implying that HMS-descended and most stress-induced Lac+ reversions form via the same mechanism. Second, the main mechanism of stress-induced mutagenesis in the Lac system is an RpoS-controlled switch to error-prone DSBR causing mutations at the sites of repair [19], and requiring HR-DSBR proteins, RpoS, the SOS response, and DinB low-fidelity DNA polymerase ([19] and reviewed [2]). Stimulation of stress-induced HR-DSBR-associated Lac reversion by DSBs delivered next to lac in vivo did not decrease the frequency of secondary mutants among the Lac+ mutants (Table 3, Figure 3) indicating that this main mechanism was inseparable from the HMS (per Figure 2B).
Mathematical modeling of previous data led two groups to favor the hypothesis that the HMS produced only 10% of stress-induced Lac+ revertants in E. coli [30], and in a similar but not identical experimental system in Salmonella enterica [32]. The other 90% of Lac revertants were suggested to arise independently of, and by some other mutagenic mechanism(s) than operates in, the HMS. In a prominent alternative model, the main 90% were proposed to form in cells with no increase in mutation rate relative to that in non-stressed cells, by “standard” generation-dependent mutational processes. The HMS was proposed to generate only few Lac+ mutants via co-amplification of dinB (EG13141), encoding error-prone DNA pol IV, with lac causing a mutator state [32].
The hypothesis that only 10% of Lac+ mutants arose from the HMS (whether via dinB amplification [32] or otherwise [30]) was based on estimation of mutability in Lac+ mutants with no phenotypically detected chromosomal secondary mutations (Lac+ “single” mutants) and finding a lower estimated value than in similar estimates from “double” mutants (Lac+ revertants with one phenotypically detected secondary mutation) [29],[30]. This was interpreted in terms of the HMS generating most double, triple and multiple mutants but few (only 10%) of the Lac+ single mutants [30], an interpretation not supported by the data presented here. We believe that the previous modeling [30] did not allow for cells to exit the HMS immediately upon acquiring an adaptive Lac+ mutation, a point which has been supported experimentally by evidence that Lac+ colonies with secondary mutations are mostly pure, not mixed for those mutations [29],[30], indicating that they generate the secondary mutations before, not after, becoming Lac+. In Text S1, we model a single HMS generating all mutants—single, double, triple, etc.—and ceasing hypermutability upon acquisition of a Lac+ mutation. Our model both predicts the apparent lower mutability of single Lac+ mutants seen previously [29],[30] and is compatible with the data presented here that Lac+ single mutants and multiple mutants arise from a common population by a common mutation mechanism—not two different mutation mechanisms (one involving dinB amplification and one not) as suggested [32].
The existence of a transiently mutable cell subpopulation indicates a differentiated state in a “bi-stable” cell population. We consider a possible model for the origin of the HMS (Figure 4A). We suggest that differentiation into an HMS cell will require three simultaneous events, all known to be required for HR-DSBR-dependent stress-induced mutagenesis in the Lac system: acquisition and repair of a DNA DSB [19]–[22]; induction of the SOS DNA-damage response [23]; and induction of the RpoS-controlled stationary-phase-, starvation- or general-stress response [10],[11]. The two stress responses transcriptionally upregulate DinB error-prone DNA polymerase 10-fold and ∼two-fold respectively [10],[27], which might be why they are required for stress-induced Lac point mutation [11],[23],[24], but this has not been demonstrated. The first two events—double-strand breakage and SOS induction—are probably related; that is, SOS might be induced by the requisite DSB. By contrast, in simple models (Figure 4A), induction of the RpoS response is imagined to occur independently of DSBs and SOS, based on different environmental inputs. That is, cells would have to sense at least two different deleterious conditions: DNA damage and an RpoS-inducing stress—while carrying a DSB—to differentiate into an HMS cell. A recent study from our laboratory showed that the SOS response is induced spontaneously in about 1% of growing cells, about 60% of that due to DSBs or double-strand ends (DSEs, half a DSB) [28]. We suggest that DNA damage provides the first stress-input sensed by the SOS response. We suggest that some of these SOS-induced cells are induced to levels of this graded response [39] appropriate for entry into the HMS at a later time if the RpoS response is induced. RpoS regulates a switch from high-fidelity to error-prone (mutagenic) DSBR mediated by Pol IV [19]. Thus, we propose that the HMS is differentiated by the convergence of these two stress-responses and a DSB/DSE in the observed [28] small subpopulation of cells, as illustrated in Figure 4A [2].
This model predicts that cells will spend differing lengths of time in the HMS. Pennington and Rosenberg [28] found that spontaneously SOS-induced cells, which induced GFP when SOS-induced, spent vastly different lengths of time in that condition. Upon recovery of the SOS-induced cells using fluorescence activated cell sorting (FACS), they found that some apparently repaired or ameliorated whatever DNA damage caused the response, then returned to cell cycling, proliferation, and formed colonies. Others stayed alive for at least eight hours after FACS but were unable to proliferate and form colonies for several days (i.e., did not end their SOS response and resume cell cycling). Friedman et al. also described the basis of the graded SOS response as a temporal gradation in how long individual cells remained induced (transcribing an SOS-GFP reporter gene) [39]. Thus, it seems likely that individual cells might spend varying lengths of time with SOS induced after DNA damage, and would thus, according to our model, spend very different lengths of time in the HMS. Cells would cycle in when they are SOS induced, and concurrently RpoS induced, then cycle out when either stress-response turns off. The SOS response is expected to be turned off when the DNA damage that instigated it is repaired. The RpoS response should turn off if the cells acquire an adaptive (e.g., Lac+) mutation that allows growth, and relief of their nutritional stress.
According to this model, the I-SceI-mediated DSBs given here might be expected to increase the number of cells in the HMS (Figure 4B). In the experiments shown in Figure 3, PBADI-SceI transcription was repressed by glucose in the medium until stationary phase, when glucose would be exhausted and leaky expression from PBAD would ensue, just prior to plating on the selective lactose medium. Leaky expression from PBADI-SceI continues on the lactose selection pates [19]. We do not know what fraction of cells induce I-SceI under these conditions [19], nor how efficiently SOS is induced by I-SceI during stationary phase. However, our results indicate that not all cells become HMS cells as a result of I-SceI-mediated cleavage in these experiments. That is, the Lac− stressed-cell population did not experience the same level of secondary mutagenesis as the I-SceI-induced point mutants (Table 4). This could be either because many cells did not receive an I-SceI-mediated DSB or because many of those that did failed to induce the SOS response. Although SOS-induction by I-SceI-mediated DSBs is efficient in growing cells [28], it is not known whether this is true in starving cells.
I-SceI-generated DSBs caused a small but statistically significant increase in the frequency of secondary mutations among Lac+ point mutants (Figure 3E, Table 3). This suggests a small increase in mutability of cells within the HMS and is not exclusive of the possible proposed increased in HMS population size (above, diagrammed Figure 4B). It is likely that the I-SceI-generated DSBs are repaired using a sister DNA molecule, which would itself carry the I-SceI cutsite. This would cause multiple rounds of I-SceI-mediated DNA cleavage, and, we suggest, prolonged induction of the SOS response, potentially causing cells to stay longer in the HMS condition, accumulating more mutations genome-wide.
Although an HMS can produce adaptive mutations, neutral and deleterious mutations will also be produced. Can an HMS enhance fitness? We suggest here that differentiation of an HMS may enhance fitness of individual cells in it, but also, separately, of the larger population.
Based on findings presented in this study, we estimated that in addition to the selected Lac+ mutation, cells that underwent stress-induced mutagenesis would carry between about one and 3.4 mutation clusters (of one or more mutations) per genome (Results). We also supported previous findings that mutations at lac occur in clusters [35] and estimated the number of mutations per cluster to be about 1.67 (Text S1). If the genome-wide mutations also occur in clusters, as mutations at lac do [35], this would then predict a frequency of between two and six mutations in addition to Lac+ per genome (1 to 3.4 mutations×1.67 mutations per cluster). This is a maximal estimate given that chromosomal mutations might not be clustered similarly, though this hypothesis seems unlikely. Could a developmental program that generates at most 2–6 additional mutations per genome be adaptive for the rare cells that generate an adaptive mutation? This will depend on how many of the additional mutations are not synonymous, and how many of the genes they fall in are relevant to the specific environment the stressed bacterium inhabits. We have no way to assess the latter, but our rough estimate of the former is that about 29.5% of all mutations falling in anywhere in the genome will affect coding (Text S1). Even if every gene mattered for fitness in the bacterium's particular environment—an unlikely prospect—this would mean that on the low end of the estimate for additional mutations (two) the probability of a non-neutral, additional mutation is 1−(1−0.295)2 = 0.503, such that 50% of the Lac+ adaptive mutants would not be harmed by having been through the hypermutable state. On the high end, the probability of a non-neutral, additional mutation 1−(1−0.295)6 = 0.877 (for 6 additional mutations), but this too is probably significantly reduced by the likelihood that many of the genes in the genome are irrelevant to fitness in any given environment (supported by a recent study [40]). The evolutionarily conserved E. coli core genome is only about half of the genes [41], such that it is possible that many of the rest are dispensable in at least some circumstances. At this gross level, it appears plausible that adaptive mutants could be generated without undue burden of coincident maladaptive mutations.
As a nonexclusive alternative, we suggest that HMS cells could produce adaptive and non-adaptive mutations and then sometimes mix their genomes with those of others in the clone, and so enhance populational fitness. A low rate of genetic mixing can allow individual mutations to be selected independently of their genetic background, thus increasing the probability of fixation of adaptive mutations [42] while lowering the probability of fixation of deleterious mutations [43], altogether benefiting the population. The mixing could occur via horizontal transmission for example by conjugation, phage-mediated transduction and natural transformation. Notably, all of these transmission modes are stimulated by stress. Conjugation is promoted by starvation stress (e.g., [44]). Induction of some prophages from the lysogenic state (and so potentially the ability to act as a transductional donor) is activated by the SOS DNA-damage stress response (e.g., [45]). Natural competence is induced by starvation and is controlled in Bacillus subtilis by the same Com gene regulators that also activate a B. subtilis stress-induced mutagenesis program [46]. Perhaps stress provokes both differentiation of an HMS while simultaneously inducing the programs that promote genetic mixing. The HMS cells could act as either donors or recipients. As donors, HMS cells could act as “mutation factories” that export mutations to other cells in the clone. As recipients, HMS cells could potentially lose deleterious mutations by genetic mixing.
Mechanisms of stress-inducible mutagenesis in bacteria, yeast, and human cells appear to limit the dangerous experiment of mutagenizing a genome in at least three important ways, each adding a layer of regulation: in time, specifically to times of stress; in genomic space to localized genome regions; and to a cell subpopulation ([19] reviewed [2]). The first two are now well documented in many different organisms and circumstances (reviewed below) and the third, so far, is demonstrated only in two circumstances of bacterial mutation. All three strategies may enhance inherent “evolvability” of cells and organisms that employ them [2],[19],[47],[48].
First, the coupling of mutagenesis mechanisms/programs to cellular stress responses limits mutagenesis to times of stress, when cells/organisms are maladapted to their environments. The bacterial RpoS-controlled general-, starvation-, or stationary-phase-stress response, positively regulates many mutagenic processes: the fidelity of DSBR, promoting point mutagenesis during stress in E. coli [19]; stress-induced mutagenesis in aging colonies of an E. coli natural isolate [6]; stress-induced few-base deletions in Pseudomonas putida [49]; and genome rearrangements such as stress-induced lac-amplification [11]; phage Mu-transposon mediated deletions in E. coli [50],[51]; and starvation-promoted transpositions in P. putida [52], among others [2]. The diversity of these processes, and the fact that even among point-mutation pathways at least two different DNA polymerases are involved (DinB for Lac [19],[24] and P. putida [53] and Pol II for mutagenesis in aging colonies [6]), suggests that RpoS promotes genome instability by more than one mechanism. The competence (natural-transformation) stress response to starvation in B. subtilis is required for starvation-stress-induced mutagenesis in that organism [46]. Two different human stress responses to hypoxia transcriptionally down-regulate mismatch-repair proteins, causing increased genome instability [54]–[57], and transcriptionally down-regulate BRCA1 and RAD51 homologous-recombinational (HR-) DSB-repair proteins, potentially promoting genome rearrangements in response to hypoxic stress [58]–[60]. The SOS DNA-damage response is the classic stress response that promotes mutagenesis both at sites of DNA damage and elsewhere [reviewed, 61] including in many stress-induced mutagenesis pathways in various bacteria [reviewed, 2]. Similarly a eukaryotic DNA-damage response to shortened telomeres promotes transposition in yeast [62]. All of these stress-response-controlled mutation mechanisms promote genetic change specifically when cells/organisms are maladapted to their environments, i.e., are stressed, potentially accelerating evolution specifically then. They are varied and suggest multiple independent evolutions of this strategy.
Second, in many systems, mutagenesis is limited in genomic space to small genomic regions. This may also be evolutionarily advantageous in potentially limiting accumulation of deleterious mutations in rare adaptive mutants, as well as promoting concerted evolution within linked genes and gene families [2],[19],[47],[48]. Restriction of mutagenesis in genomic space is evident in the coupling of both stress-induced point mutagenesis and gene amplification/genome rearrangement to acts of HR-DSBR in the Lac system [19]; and DSB-repair associated mutations in yeast [63], and is implied in E. coli ciprofloxacin-induced resistance mutations [64], Salmonella bile-induced resistance mutations [65],[66], and yeast stress-induced mutations [67], all of which require DSB-repair proteins and so may occur during localized DSBR. Similarly, the potential genome instability in human cells caused by a switch to non-homologous DSBR is suggested by down-regulation of human homologous-DSBR genes during stress and could potentially also localize mutagenesis [59],[60]. The association of transcription with mutagenesis also implies mutational localization in the genome in stressed in E. coli [68], yeast [69], and this association is also implied in B. subtilis [70] and in more indirect E. coli data [71]. Mutational clustering is observed generally in many organisms [72], including mouse [73], and also in somatic hypermutation of immunoglobulin genes [74]. Thus, many systems display localization of mutagenesis in genomic space, a potentially adaptive strategy [2],[19],[47],[48].
Finally in the E. coli Lac system, we see a third layer of limitation/regulation of mutagenesis: its restriction to a small cell subpopulation [29],[30],[31, and here]. This strategy may further buffer populations against the deleterious effects of mutagenesis by exposing only a minority of the members to these effects. Though dangerous to individual organisms, this differentiation of a bi-stable population can be advantageous to the clone, allowing the population to hedge its bets should stress be relieved suddenly [75]. Moreover differentiation of a HMS could allow some cells to both generate mutations and mix their genomes with others in the clone, as discussed in the previous section, reducing the risk of deleterious-mutation load. Competence for natural transformation in B. subtilis, which promotes genetic diversity by recombination, similarly engages only a subpopulation of stressed bacterial cells, as does sporulation [75]. How general the HMS strategy may be is not known. One other mechanism of mutagenesis in E. coli has shown evidence of engaging a HMS: stress-induced Trp reversions [76], which did not require HR-DSBR proteins, and so occurred by a mechanism different from the HR-DSBR-associated stress-induced point mutagenesis studied here. A report of mutation “showers” in mouse somatic cells [73] suggests bouts of localized transient mutability, which might be limited to a HMS, but this has not been investigated. Given the prevalence of bi-stable (subpopulation) states in bacteria [75] and the ability of all organisms to differentiate, the possible generality of HMS strategies seems likely.
E. coli strains used are shown in Table 5. Stress-induced mutation assays were performed as described [21] with two exceptions. First, the M9 glycerol medium in which cells are grown prior to plating on lactose medium was supplemented with 0.001% glucose to repress PBAD, controlling the I-SceI endonuclease, as were LBH rifampicin plates onto which cfu were spread for daily viable cell measurements. Second, in order to be able to recover Lac+ mutants carrying secondary auxotrophic mutations, the usual minimal lactose medium on which Lac+ mutants are selected was supplemented with the following additions that cannot be used as a carbon source [32],[77] at the following concentrations (mM): histidine, 0.1; isoleucine, 0.3; leucine, 0.3; lysine, 0.3; methionine, 0.3; phenylalanine, 0.3; threonine, 0.3; tryptophan, 0.1; tyrosine, 0.1; valine, 0.3; adenine hydrochloride, 0.5; guanine, 0.3; thymine, 0.32; and uracil, 0.1.
Unselected secondary mutations among Lac+ mutants were assayed by purifying Lac+ point mutants from days 4 and 5 on LBH plates containing 1% glucose (Glu), 100 µg/ml rifampicin (Rif), 40 µg/ml 5-bromo-4-chloro-3-indoyl β-D-galactoside (X-gal), (glucose for repression of PBAD, Rif to exclude FC29 scavenger cells, and X-gal to screen out lac-amplified clones per [14]. These plates were incubated overnight at 37°C. Isolated colonies were patched onto grids on the same medium (LBH Rif X-gal Glu plates) for replica plating and incubated overnight at 37°C. These master plates were replica-plated (printed) via velvets to M9 vitamin B1 minimal glucose (0.1%) plates to screen for auxotrophs; M9 B1 minimal glucose 5-fluorocytosine (5-FC, 50 µg/ml) plates to screen for 5-FC resistance (caused by mutation in the F'-borne codAB genes, per [29], confirmed by sensitivity to 5-fluorouracil) and MacConkey maltose and MacConkey xylose pH indicator plates to screen for defects in maltose and xylose fermentation (per [29]). Mutants were confirmed by purifying from the LBH-Rif-X-gal-Glu master plate and retesting on the appropriate selective or indicator medium. Unselected secondary mutations in Lac− unstressed cells were assayed by plating aliquots on LBH-Rif-X-gal-Glu plates, incubating overnight at 37°C, followed by patching isolated colonies onto the same medium, and treating as above. Similarly, unselected secondary mutations in Lac− starved cells were assayed by taking plugs of agar from between visible colonies at day 3 of incubation (comparable to day-5 Lac+ colonies due to the 2-day colony-formation time) on M9 B1 lactose plates, with supplements as above, and suspending in M9 buffer. Aliquots were plated on LBH-Rif-X-gal-Glu medium, incubated overnight at 37°C, and isolated colonies were patched, grown and replica plated as described above.
To increase Lac+ mutant frequency in the Lac assay, we employed the chromosomal E. coli I-SceI endonuclease system constructed by our lab [78] and used by us and others [e.g.],[19],[79,80]. I-SceI endonuclease makes a specific DSB at an 18bp cutsite, not normally present in the E. coli genome [81]. In this construct [78], the I-SceI-endonuclease open reading frame is cloned in front of the E. coli arabinose-inducible PBAD promoter and the expression cassette is present in the E. coli chromosome, replacing the phage lambda attachment site, attλ. We used strains carrying a chromosomal cassette of the PBAD promoter with or without the I-SceI gene and strains with or without the I-SceI cutsite on the F' episome, 4.5 kb from of the lac allele in the mhpA (EG20273) gene [19] (Table 5).
The lac region of Lac+ mutants containing chromosomal secondary mutations was PCR amplified with primers 5′-ATATCCCGCCGTTAACCACC-3′ and 5′-CGGAGAAGCGATAATGCGGTCGA-3′ and sequenced (Lone Star Labs Inc., Houston, TX) with primer 5′-ATATCCCGCCGTTAACCACC-3′.
|
10.1371/journal.ppat.1007488 | Hepatitis B virus core protein phosphorylation: Identification of the SRPK1 target sites and impact of their occupancy on RNA binding and capsid structure | Hepatitis B virus (HBV) replicates its 3 kb DNA genome through capsid-internal reverse transcription, initiated by assembly of 120 core protein (HBc) dimers around a complex of viral pregenomic (pg) RNA and polymerase. Following synthesis of relaxed circular (RC) DNA capsids can be enveloped and secreted as stable virions. Upon infection of a new cell, however, the capsid disintegrates to release the RC-DNA into the nucleus for conversion into covalently closed circular (ccc) DNA. HBc´s interactions with nucleic acids are mediated by an arginine-rich C terminal domain (CTD) with intrinsically strong non-specific RNA binding activity. Adaptation to the changing demands for nucleic acid binding during the viral life cycle is thought to involve dynamic phosphorylation / dephosphorylation events. However, neither the relevant enzymes nor their target sites in HBc are firmly established. Here we developed a bacterial coexpression system enabling access to definably phosphorylated HBc. Combining Phos-tag gel electrophoresis, mass spectrometry and mutagenesis we identified seven of the eight hydroxy amino acids in the CTD as target sites for serine-arginine rich protein kinase 1 (SRPK1); fewer sites were phosphorylated by PKA and PKC. Phosphorylation of all seven sites reduced nonspecific RNA encapsidation as drastically as deletion of the entire CTD and altered CTD surface accessibility, without major structure changes in the capsid shell. The bulk of capsids from human hepatoma cells was similarly highly, yet non-identically, phosphorylated as by SRPK1. While not proving SRPK1 as the infection-relevant HBc kinase the data suggest a mechanism whereby high-level HBc phosphorylation principally suppresses RNA binding whereas one or few strategic dephosphorylation events enable selective packaging of the pgRNA/polymerase complex. The tools developed in this study should greatly facilitate the further deciphering of the role of HBc phosphorylation in HBV infection and its evaluation as a potential new therapeutic target.
| The liver-pathogenic hepatitis B virus (HBV) is a small enveloped DNA virus that replicates through reverse transcription of a pregenomic (pg)RNA. This requires specific encapsidation of pgRNA and viral polymerase into a shell of 240 core protein (HBc) subunits. Capsid-internal formation of relaxed circular (RC) DNA enables the particle to leave the cell as stable virion; yet, when infecting a new cell it must release the RC-DNA for conversion into another, plasmid-like DNA that templates new viral RNAs. This up and down in nucleic acid interactions is presumably regulated by transient phosphorylation of HBc, mainly in its arginine-rich C terminal domain (CTD) which displays strong non-sequence-specific RNA binding. However, neither the phosphorylation sites nor the relevant enzymes are well defined. We developed a recombinant system to produce kinase-specific phospho-HBc species, and adapted a feasible gel assay for their separation. By mutagenesis and mass spectrometry we identified seven target sites for a major candidate kinase, SRPK1, in the CTD. As full SRPK1 phosphorylation thwarted non-specific RNA binding the comparably high phosphorylation of HBc in human cells suggests how specific pgRNA encapsidation might be achieved. Our new tool set will facilitate disentangling the role of HBc phosphorylation in HBV infection and exploiting it as potential therapeutic target.
| Chronic infection with hepatitis B virus (HBV) puts more than 250 million people at a greatly increased risk to develop terminal liver disease [1]. HBV, the prototypic hepadnavirus, is a small enveloped virus that replicates its 3 kb DNA genome through capsid-internal reverse transcription of a pregenomic (pg) RNA (reviewed in [2, 3]). The virion envelope consists of a lipid bilayer into which the small (S), middle (M; PreS2/S) and large (L; PreS1/PreS2/S) surface proteins are embedded [4–6]. Binding of L to the HBV receptor sodium taurocholate cotransporting polypeptide (NTCP) is essential for infection (reviewed in [7]); in addition, L contributes a "matrix domain" that interacts with the capsid for virion morphogenesis (reviewed in [8]).
The icosahedral HBV capsid (core particle) is composed of 120 dimers (triangulation number T = 4) of a single core protein (HBc) species of 183–185 amino acids (aa) in length; a minor capsid class (T = 3) comprises 90 HBc dimers. The HBc monomer encompasses an N terminal assembly domain [9], linked through residues 141–149 [10] to an arginine-rich C terminal domain (CTD; Fig 1A). The CTD is crucial for specific co-encapsidation of a complex of pgRNA and viral polymerase (P protein) during replication but it can also mediate non-sequence-specific packaging of RNA (Fig 1B), e.g. when HBc is expressed in E. coli [9, 11–13]. Structural analyses of such capsid-like particles (CLPs), mostly from CTD-less variants like HBc149 [9], revealed five α-helices [14–16]. Helices α3 and α4 form an antiparallel hairpin; for dimerization, two such hairpins associate into four-helix bundles which protrude as spikes from the capsid surface (Fig 1B). Helix α5 plus the downstream sequence to position 140 harbor the major inter-dimer contacts.
Little is known about the structure of the CTD; in current CLP structures the visible sequence commonly ends within the linker [10, 14–19]. Regarding CTD disposition most data support a luminal localization [20–23]. However, a permanent internal disposition as well as a static nucleic binding capacity are incompatible with the full set of HBc functions in the viral life-cycle (Fig 1C). Beyond pgRNA/P protein encapsidation these include CTD-mediated reverse transcription [24] into single-stranded (ss) minus-DNA and then partly double-stranded (ds) relaxed circular (RC) DNA [24–26]; capsid envelopment for secretion of virions [8]; and, upon infection of a new cell, transport of the viral genome to the nuclear pore so as to release the RC-DNA into the nucleoplasm [27, 28] for conversion into covalently closed circular (ccc) DNA [29, 30]. Nuclear transport requires binding of CTD-encoded nuclear localization signals (NLSs; [31]) to cytosolic transport receptors [32–35] such that at least one CTD per capsid must become exposed [18, 27, 32, 36, 37].
Hence the capsid must safely stow ss and also ds nucleic acid with twice as many negative charges and then orderly let go of it.
A likely mechanism underlying these CTD dynamics is transient phosphorylation, early-on hinted at by a capsid-associated protein kinase activity [38, 39]. However, neither this kinase nor other potentially HBc-relevant kinases have unambiguously been identified. Proposed candidates include Ca2+ and/or lipid-activated protein kinase C (PKC; [36, 40]); cyclic AMP dependent protein kinase A (PKA; [41, 42]); serine/arginine-rich protein kinase 1 and 2 (SRPK1, SRPK2; [43]); cyclin-dependent kinase 2 (CDK2; [44]); and polo-like kinase 1 (PLK1; [45]). Major analytical challenges are the presence in human cells of >500 protein kinases plus >200 phosphatases [46]; the repetitve nature of the HBc CTD sequence (Fig 1A); the limited specificity of pharmacological kinase inhibitors; and especially the lack of assays that feasibly distinguish non-phosphorylated and differently phosphorylated HBc species.
As a surrogate, various studies assessed the impact of serine and/or threonine (S/T) replacements in HBc by alanine (A) to prevent phosphorylation, and by aspartic acid (D) or glutamic acid (E) to mimic phosphorylation [47–50]. Collectively these data suggest that phosphorylation is necessary early and some dephosphorylation late during replication. However, genetic mimics cannot model the dynamics of phosphorylation. Phospho-HBc-specific antibodies [51, 52] are valuable but they interrogate only few of the many options for phosphorylation site occupancy. Correlation with specific replication states is further convoluted by the variety of capsid forms, including nonenveloped capsids [53] and the recently found enveloped genome-less capsids ("empty virions"; [52]) which seem to by far outnumber infectious virions (reviewed in [54]).
The most direct evidence for the dynamics in core protein phosphorylation comes from duck HBV (DHBV). Intracellular DHBV core protein (DHBc) showed up to four distinct phosphorylation-dependent bands in normal SDS-PAGE [55] whereas virion-derived DHBc displayed a single, non-phosphorylated band, as confirmed by mass spectrometry (MS) [56]. The presence in virions of mostly mature viral ds DNA [57], similarly observed for HBV [58], supported the "maturation signal" hypothesis [57] whereby capsid-internal genome maturation, perhaps sensed by the CTDs and/or their phosphorylation status, exposes interaction sites for L protein on the capsid surface (Fig 1D). The finding that also empty though not ss nucleic acid containing capsids can be enveloped prompted a revised model whereby RNA or ssDNA actively inhibit envelopment [52, 59]. A cryoEM comparison between recombinant RNA containing capsids and serum virus-derived, supposedly dsDNA bearing capsids had indeed revealed subtle differences [17]. However, at the time empty virions in serum [52, 60] were not known; hence potential envelopment-relevant structural differences and their correlation with HBc phosphorylation remain open (Fig 1C, i-iii).
Some natural HBc mutations such as F97L [61–63] promote secretion of immature, ssDNA containing virions. Such capsid proteins might by default adopt an envelopment-proficient conformation (Fig 1D). Yet, while affecting in vitro assembly [64] no differences in stability or morphology between wild-type and F97L HBc particles were detectable [63, 65]. However, a potential role of phosphorylation could not be assessed at the time.
In sum a large body of data supports correlations between HBc phosphorylation, genome maturation and capsid structure, including CTD disposition and envelopment-competence; however, fundamental information for disentangling these interdependencies is lacking.
In order to generate such information we here established the efficient recombinant production of distinct phospho-HBc species and showed the feasibility of Phos-tag SDS-PAGE [66] for their differentiation. Using MS and mutagenesis we identified seven hydroxy amino acid residues in the CTD as target sites for SRPK1 [34, 43] while PKA and PKC modified fewer, yet to be mapped sites. Full occupancy of the SRPK1 sites massively reduced HBc183 CLP RNA content while abrogation of one phospho-site restored substantial RNA binding. Full phosphorylation also caused differences in CTD accessibility to trypsin, regardless of the F97L mutation. Importantly, in human hepatoma cells the bulk of HBc appeared as highly phosphorylated as SRPK1-coexpressed HBc. These data pave the way for a comprehensive characterization of the substrate properties of hepadnaviral core proteins for individual kinases and phosphatases. Virologically they suggest a mechanism whereby high-level HBc phosphorylation principally suppresses RNA binding whereas one or few dephosphorylation events could enable specific packaging of the pgRNA/P protein complex.
Eukaryote-like S/T protein kinases are rare in eubacteria [67] and apparently absent from E. coli laboratory strains. Hence co-expression of HBc with a mammalian kinase should allow the background-free evaluation of its action on HBc. In a first step we created T7 promoter-based vectors carrying an E. coli codon usage adapted HBc gene (HBc183opt) to boost expression of full-length HBc183 (Fig 1A). CLP yields were indeed around three-fold higher (S1 Fig) compared to the older non-optimized HBc gene [68] in the same pET28a2 vector [69, 70], and they increased further using pRSF-Duet derived vectors (see below).
For the coexpression approach the recombinant kinase must be enzymatically active and target sites in HBc must not be sequestered by CLP formation. We thus chose human SRPK1 as the first candidate HBc kinase. Beyond its very high affinity for the HBc CTD [34, 43, 71] we had already shown HBc phosphorylating activity of the full-length enzyme in bacteria [11]; here we used an N terminally His6-tagged variant, NHisSRPK1ΔNS1, in which an internal deletion boosts soluble expression in bacteria [72].
To minimize experimental variations seen when HBc and kinase were expressed from two separate plasmids [11] we devised a single-vector dual expression system; it features a Tet repressor / anhydrotetracycline (AHT) inducible Tet promoter cassette plus an isopropyl β-D-1-thiogalactopyranoside (IPTG) regulatable T7 promoter cassette (S2 Fig). Functionality was shown by inducer-specific expression of eGFP from one and mCherry from the other cassette (S2 Fig). HBc183opt in the T7 cassette was expressed at up to five-fold higher levels than from the pET28a2 vector (S3 Fig). Of note, expression from the Tet cassette was affected by the specific ORF sequence. For instance, the basal level of NHisSRPK1ΔNS1 expression without AHT was still about one third as high as with inducer (S3A Fig); hence in subsequent SRPK1 coexpression experiments AHT was omitted.
Either way, in line with its high affinity for HBc [34, 43] most of the SRPK1 cosedimented with the CLPs, both from wild-type and F97L HBc (S3 Fig); separation was achieved by immobilized metal ion chromatography (IMAC) under semi-denaturing conditions (S4 Fig). Altogether, the system provided easy access to milligram amounts of CLPs per 100 ml of induced bacterial culture.
Low level co-expression in E. coli of full-length SRPK1 with HBc183 had generated a mixture of HBc species containing from one to several phosphoryl groups [11]; this prompted the follow-up use of a triple S>E variant, HBc183-EEE (S155E_S162E_S170E) as a homogeneous genetic mimic of such partially phosphorylated HBc183 [34, 35]. HBc183-EEE CLPs contained ~25% less RNA than wild-type HBc183 CLPs [73]. To evaluate the impact on CLP RNA content of the much higher NHisSRPK1ΔNS1 expression in our new system we separated the CLPs by NAGE, then stained packaged RNA by ethidium bromide (EB) and the protein shell by Coomassie Brillant Blue (CB) [9, 74]; other stains, e.g. Sybr Green 2 (SG2) for RNA and Sypro Ruby (SR) for protein [75], could be used instead (see below). The intensity of RNA vs. protein stain per band is proportional to the CLP´s RNA content.
Fig 2A shows such an analysis for HBc183_F97L expressed in the absence (-SRPK) vs. presence of SRPK1. Despite much weaker EB staining the +SRPK1 samples showed even stronger CB staining; in either case abundant CLPs were visible in negative staining EM; wild-type HBc183 CLPs gave comparable results. To appreciate the extent of reduction in RNA content caused by NHisSRPK1ΔNS1 coexpression we used a reference set of CLPs from C terminally truncated HBc proteins, starting with the CTD-less variant HBc140. SG2 and SR allowed semiquantitative assessment of the relative fluorescence intensities of each band (Fig 2B) using a laser scanner (Typhoon FL7000, GE Healthcare). The ratio of RNA: protein fluorescence for non-phosphorylated HBc183 CLPs was set to 100%, to which the ratios for the other CLPs were normalized. Accordingly, SRPK1-coexpressed HBc183 CLPs contained nominally 27% as much RNA as the HBc183 reference CLPs, very similar to CTD-less HBc140 CLPs. For the other variants, the ratios increased with increasing CTD length, in accord with the electrostatic "charge balance" hypothesis [13, 26, 76].
An estimate of absolute CLP RNA content as molar ratio of nucleotides (N) per core protein (P) subunit (N/P) was obtained from UV-VIS spectra by deconvoluting the superimposed RNA and protein absorbances [73]; spectra were recorded using gradient-enriched CLPs incubated with 1/100 the amount of (w/w) of RNase A and subsequently dialyzed to remove non-packaged RNA. For HBc183 CLPs we obtained (N/P) values ± standard deviation (SD) of 16.4±2.2 (n = 8); for a T = 4 CLP this corresponds to an RNA content equivalent to ~4,000 nt, similar to other reports [12, 73]. Much in contrast, SRPK1-coexpressed HBc183 CLPs gave a ratio of only 2.3±0.5 (n = 8), not significantly different from CTD-less HBc140 (2.9±1.3; n = 4) and HBc149 CLPs (1.8±1.5; n = 4). Intermediate RNA contents were found for HBc161, 163, 166 and 173 CLPs, as shown in Fig 2C in comparison with the NAGE/SG2:SR procedure. While the apparent overestimation of low RNA contents by the latter method could later be reduced by several improvements (Materials and Methods), both data sets corroborated that SRPK1 coexpression reduced RNA content as strongly as deletion of the entire CTD. This implied a phosphorylation extent sufficient to shift assembly to an RNA-independent, protein-protein interaction driven mechanism [37, 77], perhaps by neutralizing most of the positive CTD charges [13].
The Phos-tag chelator coordinated with two Me2+ ions binds to phosphoryl groups, including in proteins [78]. In SDS-PAGE gels containing copolymerized Phos-tag acrylamide and loaded with Mn2+ or Zn2+ phospho-proteins migrate more slowly than their non-phosphorylated forms [66]. Hence this technique appeared attractive to detect HBc phosphorylation. After optimization we routinely used home-made gels containing 75 μM Phos-tag acrylamide loaded with 150 μM Mn2+. As shown in Fig 3, in normal SDS-PAGE both HBc183 and HBc183_F97L migrated to the same ~21 kDa position, regardless of SRPK1 coexpression. In Phos-tag SDS-PAGE, however, exclusively the kinase-coexpressed proteins exerted a drastically reduced mobility.
Band ladders produced in in vitro SRPK1 phosphorylation and lambda phosphatase dephosphorylation reactions with a non-assembling GFP-CTD fusion protein suggested phosphorylation at five or more positions. Analogous coexpression of HBc183 with the catalytic domains of PKA (Genbank NP_032880.1) and PKC theta (Genbank NP_001269573.1), chosen owing to their reportedly functional expression in E. coli [79, 80], caused a less pronounced retardation (see below).
The number of phosphoryl groups introduced into HBc183 by SRPK1 was analyzed by matrix-assisted laser desorption/ionization (MALDI) time-of-flight (TOF) MS. The major MH+ peaks in the relevant 21 kDa region of the spectra had mass/charge (m/z) values of 21,103 for mono-expressed HBc183 and 21,664 for SRPK1-coexpressed HBc183. These values match most closely to the calculated masses of 21,116 Da for completely unmodified HBc183, and to 21,676 Da for a seven-fold phosphorylated HBc183 (S5 Fig; each phosphoryl group contributes ~80 Da). Notably, for SRPK1-coexpressed HBc140 and HBc149 the main peaks still corresponded to the non-phosphorylated proteins; hence the assembly domain and the linker sequence lack SRPK1 target sites.
For more accurate mass determination we devised a method (S1 protocol) to isolate just the CTD peptide from a tobacco etch virus (TEV) protease cleavable NHisGFP-TEV-CTD fusion protein (Fig 4A), expressed with or without kinase. Cleavage reduces the relevant masses to ~5,000 Da, with a synthetic N-acetylated CTD peptide serving as reference (Fig 4B). Methods development included replacement of the C terminal Cys residue by Ala (C183A in full-length HBc numbering). In brief, the NHisGFP-TEV-CTD protein was enriched by Ni2+ IMAC under semi-denaturing conditions, then incubated with His-tagged TEV protease [81] to release the CTD peptide; all His-tagged components (uncleaved GFP fusion protein, the GFP-containing cleavage product and TEV protease) were removed by another round of IMAC. The clipped-off CTD peptide in the flow-through was concentrated and further enriched by ultrafiltration. An SDS-PAGE analysis of the final products obtained upon coexpression with SRPK1ΔNS1 or the PKA catalytic domain (PKAcd) or without kinase, alongside the synthetic CTD peptide, is shown in Fig 4C.
Importantly, MS of the SRPK1 sample (Fig 4D) showed one major peak with excellent agreement to a seven-fold but not six- or eight-fold phosphorylated CTD peptide. Hence there are seven SRPK1 target sites in the HBc sequence 146–183. The m/z values for the synthetic and the non-phosphorylated fusion protein-derived CTD peptide were <1 Da off the calculated masses (S6 Fig). The PKA-coexpressed sample showed two main peaks, matching four and five phosphoryl groups, respectively; for full-length HBc183 as substrate the products from PKA and PKC coexpression carried only three phosphoryl groups (S7 Fig), possibly due to assembly-mediated target site sequestration.
To identify the seven SRPK1 sites in HBc we individually mutated the seven S residues plus one T residue in the CTD (S155, T160, S162, S168, S170, S172, S178 and S181) to A, coexpressed the mutant proteins with SRPK1 and monitored phosphorylation by MS (Fig 5A) and by immunoblotting after Phos-tag SDS-PAGE.
For all mutants but one MS revealed a best match to the six-fold phosphorylated forms, i.e. loss of one phosphorylation site. The exception was mutant S181A which still matched best to a seven-fold phosphorylated species. Hence all hydroxy amino acids in the HBc CTD including T160 yet except S181 are substrates for SRPK1. Double and triple S/T>A mutants confirmed the data as the number of lost phosphorylation sites always corresponded to the number of S/T replacements by A (or by G or R), except when S181 was mutated (S7 Fig).
In line with their at least six-fold phosphorylation all single S/T variants were similarly strongly retarded in Phos-tag SDS-PAGE as SRPK1-coexpressed wild-type HBc183, as shown by immunoblotting (Fig 5B) with the assembly domain-directed anti-HBc mAb 1D8 [82]. Notably, though, the correlation between the number of phosphorylation sites and mobility was not strictly linear. The six-fold phosphorylated variants S155A, S176A, and S178A migrated about as slowly as the seven-fold phosphorylated wild-type protein and variant S181A. In contrast, the likewise six-fold phosphorylated variants T160A, S162A and S170 (marked with asterisks in Fig 5B) had distinctly higher mobilities. Hence beyond the number of phosphoryl groups also their sequence context contributes to the interaction with the Phos-tag. Knowledge of the S/T>A variants´ phosphorylation status also enabled a more detailed characterization of the epitope requirements of mAb T2212 (see below).
We then evaluated the impact of the S/T>A mutations on CLP RNA content by the NAGE based SG2/SR staining assay (Fig 5C, S8 Fig) but using a more elaborate protocol (Material and Methods). Owing to the lower background signals the relative RNA contents of HBc140, HBc149 and SRPK1-coexpressed HBc183 as well as HBc183_S181A CLPs were now estimated to ≤10% of that in unmodified HBc183 CLPs (Fig 5C). The mostly three-fold phosphorylation by PKA and PKC caused only minor reductions (significant only for PKC). Remarkably, preventing phosphorylation at one of the seven SRPK1 sites was sufficient to increase RNA contents to generally ~20–30% of non-phosphorylated HBc183 CLPs; additional phosphorylation site mutations further increased RNA content (S8 Fig). Intriguingly, the single S170A mutation restored RNA content to ~50% of unmodified wild-type HBc183 CLPs, and this difference to the other single variants was significant (p<0.001; n = 6). Hence phosphorylation of S170 may be particularly effective in suppressing RNA binding by the HBc CTD.
HBc183 CLPs are sensitive against SDS [9, 65]; hence titration with SDS might reveal whether high-level phosphorylation and/or the concomitant low RNA content affect CLP stability. To this end we incubated HBc183 CLPs and HBc183_F97L CLPs expressed in the absence vs. presence of SRPK1 with either 1x TAE electrophoresis buffer (Ø), or a native glycerol-based DNA loading buffer (Ø*), or increasing amounts of a 6x DNA loading buffer (NEB Purple) containing 0.48% SDS (0.08% at 1x concentration). RNA and protein were monitored by EB and CB staining; expectedly, SRPK1-coexpressed CLPs showed only faint EB versus CB signals (Fig 6, right panels).
At 0.024% SDS the nonphosphorylated CLPs released some of the RNA (faster migrating EB smear) although the major band remained at the original position. At 0.048% SDS, the fast RNA smear increased, while the original CLP band disappeared in favor of a slower, less intense band from which EB signals emanated upwards to the cathode; still higher SDS concentrations enhanced these effects. CB staining of the latter bands suggests they may represent complexes of HBc183 proteins with exposed CTDs (forced to move towards the cathode) to which some of the initially encapsidated RNA is bound. In accord with previous data [65] no clear difference was detectable for wild-type vs. F97L CLPs. Interestingly the SRPK1-phosphorylated low RNA CLPs also showed a sharp mobility transition between 0.024% and 0.048% SDS, although the new products migrated a bit slower than those from the nonphosphorylated CLPs. In sum, SDS sensitivity of the HBc CLPs was neither affected by seven-fold phosphorylation of the CTD, nor by RNA content or the F97L mutation.
The high arginine content makes the entire HBc CTD an excellent substrate for trypsin; however, protease accessibility is affected by the CLP structure [37, 65, 83]. In our hands, trypsin had converted part of the unmodified wild-type HBc183 into a major product with HBc149-like mobility [83]. To reveal potential alterations of this pattern by phosphorylation, RNA content and/or the F97L mutation, we monitored the kinetics of trypsin action on non-phosphorylated vs. phosphorylated CLPs from wild-type and F97L HBc183 protein.
First we addressed a potential inhibition of trypsin cleavage by CTD phosphorylation by comparing the trypsin sensitivity of the non-assembling NHis-GFP-TEV-CTD fusion protein (see Fig 4A) expressed in the absence vs. presence of SRPK1. Both proteins were completely cleaved with similar kinetics (Fig 7A, left panel). Under the same conditions, HBc183 in non-phosphorylated CLPs was only partially cleaved within 30 min (Fig 7A, right panel), with about half the molecules remaining intact and the other half displaying HBc149-like mobility; this pattern remained stable over 2 h.
Higher temporal resolution (Fig 7B) revealed the transitory formation of intermediate mobility bands which had mostly disappeared at 30 min in favor of the stable HBc149-like product (labeled "4"); thereafter the ratio between presumably intact HBc183 (labeled "1") and product "4" remained constant for at least 4 h. NAGE did not reveal significant alterations in EB vs. CB staining over time (Fig 7B, lower panels). Hence loss of half the CTDs affected neither RNA content nor surface charge, indicating the overall CLP structure remained intact. Comparable results with HBc183_F97L CLPs (Fig 7C) ruled out major effects of the F97L mutation.
In contrast, SRPK1-coexpressed HBc183 CLPs (Fig 7D, top panel) presented two additional discrete, stable products ("2" and "3") of which product 3 eventually reached similar levels as the remaining full-length protein. By comparison with HBc marker proteins (Fig 7D, lower panel) products 3 and 2 resulted from cleavage between positions 157–159 and 166–172, respectively. HBc_F97L CLPs produced a very similar four-band pattern in SDS-PAGE (Fig 7E, top panel); in NAGE CLP mobility and RNA content again remained unaltered (Fig 7E, bottom panels). While quantitation of the band 1 to 4 proportions could be disturbed by the presence of not fully intact CLPs, formation of bands 2 and 3 exclusively from the phospho-CLPs strongly implies distinct steric constraints for the band 2 and 3 processing sites compared to unmodified CLPs. Scenarios compatible with the data (complete CTD extrusion vs. looping out of CTD parts) are shown in S9 Fig.
To directly assess the impact of seven-fold CTD phosphorylation and/or low RNA content on capsid structure we analyzed nonphosphorylated vs. SRPK1-phosphorylated HBc183 CLPs by cryoEM (S10A and S10 Fig) and calculated image reconstructions (Fig 8) at 7.8 Å (B = 707 Å2) and 7.9 Å (B = 979 Å2) resolution (S10C Fig). Most striking was a distinct shell of internal density exclusively in the non-phosphorylated particles (Fig 8B) which must represent CTDs plus packaged RNA. The relatively high B-factors suggested some structural heterogeneity within each CLP class, hence the gross features in the shells of unmodified vs. phosphorylated particles appeared very similar. However, in defined areas significantly different grey value distributions occurred (Fig 8C; see S2 Protocol for significance calculations). For the SRPK1 coexpressed CLPs they could be best described as reduced order in the N terminal regions which flank the spikes (Fig 8C bottom vs. top: less red color around the three-fold axes), plus an apparent stretching of the spike helices and the inter-dimer contact regions (Fig 8C bottom vs. top: more blue color within the dimer contours).
To improve resolution for fitting these differences into atomic models of HBc we vitrified additional aliquots of the same sample preparations which, fortuitously, had been stored for 1.5 years at 4°C. Corroborating the high stability of HBc CLPs, NAGE, SDS-PAGE and Phos-tag SDS-PAGE revealed no signs of degradation or loss of phosphorylation. Surprisingly, however, by 3D-classification the majority of these "aged" CLPs, especially the phosphorylated ones, grouped into a distinct class from the bulk of fresh particles (vitrified within two weeks post preparation) used above. The aged particle reconstructions had higher resolution and lower B-factors (6.3 Å and 6.6 Å, and 555 Å2 and 466 Å2, respectively, for non-phosphorylated vs. SRPK1-phosphorylated CLPs), in line with increased ordering with time. For the non-phosphorylated CLPs the differences were only minor but for the phospho-HBc particles they were pronounced. As in the comparison with fresh unmodified CLPs (Fig 8C) the freshly vitrified phospho-CLPs showed extended spikes and less ordered N termini (S11A Fig, top) whereas after ageing they appeared very similar to the non-phosphorylated ones (S11A Fig, bottom). This suggests a conformational maturation process, although not necessarily in the classic sense of gaining envelopment competence.
Fitting the HBc149-derived crystal structure 1QGT [16] into these reconstructions accounted for most density in the assembly domains but left unaccounted internal density close to the inner CLP surface, likely related to CTDs plus RNA in the non-phosphorylated CLPs yet only CTDs in the phospho-CLPs. The undisturbed view onto the inner shell surface of the latter particles revealed extensive tube-like structures around the 5-fold symmetry axes (Fig 8D, bottom vs. top). Connecting to the last visible residues in the crystal structure they probably represent residues from the linker plus the CTD. The respective density extended towards the feet of the spikes, with apparently direct contacts to R112 in both fresh and matured CLPs (S11B Fig). Their visibility implies rather stable interactions, perhaps including electrostatic binding between R112 and phosphorylated CTD residues in the phospho-CLPs (S9C and S9D Fig). Similar though less prominent extra density was also seen in the non-phosphorylated CLPs, although their high RNA content makes a clear distinction between nucleic acid and protein residues difficult (S11B Fig).
Recent data suggest that most HBc in mammalian cells from replicating as well as non-replicating vectors is phosphorylated [13, 52] yet the number of phospho-sites per HBc monomer is unknown. We therefore analyzed HBc in intracellular capsids from the stable HBV producing TetOFF HepG2.117 line [84] by Phos-tag SDS-PAGE, using our recombinant phospho-HBc183 proteins as markers. To minimize band distortions seen with crude cytoplasmic lysate we first enriched capsids by sedimentation through 10–60% (w/v) Nycodenz gradients (Fig 9A; S12A Fig). The presence of HBV-typical replicative DNA intermediates was corroborated by Southern blotting (Fig 9B); however, less vs. more mature species were barely separated and the presence of empty particles [54] remains to be addressed.
Regardless of this, Phos-tag SDS-PAGE showed a comparably strong retardation of the bulk HBc signals as for SRPK1-phosphorylated recombinant HBc183 (Fig 9C, short exposure). Longer exposure revealed several very weak intermediate mobility bands in the HepG2.117 samples, similar to those seen for CLPs coexpressed with PKA and PKC, and for non-phosphorylated CLPs; however, >95% of the band intensity was concentrated in the slowest migrating band. Though this would be in line with an SRPK1-like phosphorylation further investigation using the phospho-CTD specific mAb T2212 [51] indicated a non-identical phosphorylation pattern. As reported mAb T2212 did not recognize unmodified recombinant HBc183, yet it reacted strongly with the PKA- and some of the PKC-phosphorylated HBc species. Interestingly, fully SRPK1-phosphorylated HBc gave almost no signal whereas single alanine replacements of the phospho acceptor sites S162, S168 and, weakly, S176 and S178 but not S170 restored reactivity (S12 Fig). Hence the T2212 epitope appears to depend on the presence of phosphoryl groups at some positions, including S170, yet their absence from nearby sites.
To assess whether mAb T2212-reactive phospho-HBc species are present in enveloped particles we separated extracellular particles from HepG2.117 cells and from a new HBc183-only expressing Huh7 line by NAGE; recombinant HBc183 CLPs served as control (Fig 9D). Immunoblotting revealed a fast migrating band in the human cell-derived samples which corresponded to naked capsids, as indicated by comigration with bacterial HBc183 CLPs; as before, the latter did not react with mAb T2212 but were detected by the anti-assembly domain mAb 312 (marked by **). Importantly, mAb T2212 revealed a second, slower migrating band exclusively in the HepG2.117 samples which also stained with the anti-HBs mAb 9H9 [85]. Hence capsids containing phosphorylated HBc can be enveloped. More detailed conclusions will require separation from the naked capsids and efficient fractionation of enveloped capsids with differing nucleic acid content.
Dynamic (de)phosphorylation events are long thought of as enabling hepadnaviral core proteins to execute their multiple functions in the viral life-cycle but many details remained obscure. The tools developed in this study, especially in combination, are beginning to shed new light on these unresolved issues.
Our optimized E.coli coexpression system provides robust access to defined HBc species phosphorylated by individual kinases, as shown here for SRPK1, PKA and PKC. Varying the HBc part allows to pinpoint how specific mutations and/or the assembly status affect HBc´s substrate quality for a given kinase; in addition, the recombinant phospho-HBc proteins will make well-defined substrates to study phosphatases. Phos-tag SDS-PAGE was able to separate HBc species containing from no to seven phosphoryl groups which together with MALDI-TOF MS enabled us to map the seven SRPK1 target sites in the CTD. This, in turn, allowed correlating the number of CTD phosphoryl groups with CLP RNA content, stability and CTD exposure. Importantly, the new tools indicated that the bulk of HBc in human hepatoma cells is similarly highly phosphorylated as the recombinant SRPK1 phosphorylated protein, albeit in a non-identical pattern. Here we first discuss technical aspects of our study and then its implications for HBc phosphorylation in the viral life-cycle.
Our MS data confirmed the absence from E. coli BL21 cells of enzymes capable of phosphorylating any residue in HBc183 (Figs 3 and 4; S5 and S6 Figs). Hence all previously reported properties of E. coli derived HBc CLPs relate to the fully unphosphorylated state. Vice versa, co-expressing a eukaryotic kinase then allows to single out how that specific kinase acts on HBc, as shown here for SRPK1 and per proof-of-principle for PKA and PKC.
There are still caveats for interpretation, including the concentration and ratio of kinase to HBc substrate in the bacteria, plus the potential sequestration of CTD-embedded target sites in the capsid interior. ATP is present in E. coli at concentrations of ≥1 mM [86] and thus not limiting. However, a low translation rate of a kinase compared to HBc, or poor solubility and/or low affinity for HBc could all affect phosphorylation efficiency. The higher levels of soluble SRPK1ΔNS1 (S3 and S4 Figs) likely explain the homogeneous seven-fold phosphorylation found here compared to the mixed phospho-HBc species seen upon coexpressing full-length SRPK1 [11]. An impact of target sequestration on phosphorylation efficiency is supported by the MS data for PKA coexpression which indicated predominantly three-fold phosphorylation of self-assembling HBc183 (S7 Fig), yet four- and five-fold phosphorylation of the non-assembling GFP-CTD protein (S6 Fig).
For SRPK1 the seven identified phosphorylation sites likely represent the maximum number; notably they include T160 (Fig 5) although SRPK1 is mainly considered as a serine kinase [72, 87]. Furthermore, the apparent stability of bacterial phospho-HBc in the absence of Phos-Stop inhibitor suggests the absence of phosphatases able to revert the mammalian-type phosphorylations.
A compromised substrate specificity in the E.coli system is unlikely because none of the eleven S and twelve T residues upstream of the CTD showed any indication of phosphorylation when HBc140 or HBc149 were coexpressed with SRPK1 (S7 Fig).
In sum, these data advocate a broad applicability of our recombinant phospho-HBc system to reveal the impact of individual kinases (and with adaptations, of phosphatases) on HBc.
Different from conventional SDS-PAGE Phos-tag SDS-PAGE resolved HBc species containing from none to five or six phosphoryl groups. While higher phosphorylation generally correlated with stronger retardation, discrimination amongst highly phosphorylated species was less clear, also owing to sequence context specific effects. (Fig 5A and 5B). It will be interesting to see whether variants with unusual Phos-tag mobility shifts also display specific biological features. For instance, two of the three single-site mutations causing lower than average retardation, S162A and S170A (Fig 5C), had the strongest negative impact on pgRNA encapsidation [47, 48, 88]. Coexpression with SRPK1 of additional HBc183 S/T>A mutants would provide a unique, comprehensive panel of reference proteins for more systematic investigations.
Uniform seven-fold phosphorylation by SRPK1 enabled to correlate a defined phospho-status with basic properties of HBc. Most striking was the drastic reduction in CLP RNA content (Figs 2, 5, 6 and 7). This has not been seen with partly phosphorylated HBc or with phosphorylation mimicking S/T>D,E variants. For instance, a variant termed 7E with seven S/T>E replacements in the CTD, expressed in HEK293 cells, reportedly contained still half as much RNA as nonphosphorylated bacterial CLPs [89]; similar results were seen for an all-S>D variant (S7D) expressed in E.coli [13]. Likely, the much stronger impairment of RNA binding by seven phosphoryl groups versus seven carboxyl groups is owed to their different chemistries. Seryl- and threonyl-phosphomonoesters are dibasic acids. With pKa values of 1.2 and 6.5 [90] even the second phosphoryl hydroxy group will be mostly deprotonated at physiological pH; hence each phosphoryl group can contribute nearly two negative charges. Seven-fold CTD phosphorylation would thus suffice to largely neutralize the 15 net positive charges (Fig 1), minimize the CTD´s general RNA binding capacity, and explain the low RNA content of the resulting CLPs.This interpretation is supported by inverse experiments in eukaryotic settings where complete inhibition of CTD phosphorylation by S/T>A mutations led to the formation of non-specifically RNA-filled rather than empty HBc CLPs [13, 89]. Practically, low RNA CLPs from SRPK1 coexpression in E. coli of full-length HBc variants bearing heterologous sequences could be of interest for vaccine applications [74].
Seven-fold phosphorylation and low RNA content had little impact on CLP stability against SDS (Fig 6) whereas trypsin treatment revealed specific phosphorylation and/or RNA content-dependent differences; their independence from the F97L mutation suggests that the structural changes causing the premature virion secretion phenotype of the F97L variant are either very subtle or not sufficiently long-lived to be observed [91], as corroborated in a recent high resolution cryoEM study [92].
Non-phosphorylated CLPs yielded one major HBc149-like product ("band 4" in Fig 7B and 7C), whereas the SRPK1-phosphorylated CLPs gave two additional, stable intermediates (Fig 7C and 7D; S9 Fig). These data are reminiscent of but also distinct from artificially RNA-depleted HBc-EEE vs. WT CLPs [37]; there, a band similar in size to our band 3 also accumulated over time, however, in both wild-type and HBc-EEE CLPst. The more pronounced differences in our experiments may relate to the higher number of modifications per CTD plus the chemical differences outlined above. Two models that could explain formation of distinct trypsin processing products (S9C and S9D Fig) would invoke either complete extrusion of some CTDs, or the looping-out of selective trypsin target sites with the very C terminal CTD residues still internally disposed. In the phospho-HBc CLPs this could be mediated by electrostatic interactions between the phosphoryl groups and positively charged side-chains in the CTDs and/or on the inner CLP lining, as proposed for HBc-EEE CLPs [37]. Maintained integrity and RNA content of the trypsin-treated CLPs (Fig 7B, 7C and 7E) appear in better accord with the looping-out model.
The current cryoEM reconstructions did not resolve this issue but they revealed distinct phosphorylation-dependent differences. The lack in SRPK1-coexpressed HBc183 CLPs of internal RNA density (Fig 8B) enabled allocating the remaining extra density to the linker plus CTD residues which formed tube-like structures around the five-folds, in apparent contact with residues lining the capsid lumen, i.e. R112 (S11B Fig). Freshly vitrified phospho-CLPs showed well detectable local differences in their assembly domains compared to unmodified CLPs which largely disappeared upon long-term storage. In particular the very N termini adopted a more ordered conformation, as in both fresh and aged unmodified CLPs. A potential physiological relevance of this slow rearrangement in the phospho-CLPs remains to be determined. Notably, though, replacement of the N-proximal P5 lowered virus secretion [93] and P5 is close in space to residues 95–97 implied in interactions with L protein [94].
The bulk of HBc183 in capsids from human cells was strongly retarded in Phos-tag SDS-PAGE (Fig 9C, S12B Fig), in line with an independent report [13]. Comparison with the recombinant phospho-HBc proteins indicated an SRPK1-like seven-fold phosphorylation status, in line with a physiological role of SRPK1 as a HBc kinase [34, 43, 71]. However, Phos-tag retardation alone would be also compatible with a slightly lower or higher phosphorylation extent (Fig 5). Also, the strong reactivity of mAb T2212 with the human cell-derived but not the recombinant phospho-HBc (S12 Fig) indicates at least one phosphorylation-dependent difference. If SRPK1 was indeed the main HBc kinase, (a) compatible phosphorylation pattern(s) could arise from one or few selective dephosphorylation events. Alternatively, however, the combined action of other kinases (and phosphatases) could lead to a similar pattern.
Notably, mAb T2212 detected phosphorylated HBc in enveloped particles (Fig 9), as proposed for empty virions [52]. Further exploration using our new tools, with more efficient separation of the various HBV particles [54], will help to accurately ascribe specific phosphorylation patterns to individual particle types.
At any rate does the high phosphorylation level of most HBc183 in human cells imply a similar suppression of general RNA binding capacity as in recombinant SRPK1 phosphorylated HBc183. As outlined in Fig 10, in favor of viral replication this would counteract encapsidation of irrelevant RNAs [13, 89] yet also impair pgRNA packaging, giving empty capsids and empty virions. This dilemma could plausibly be resolved by partial dephosphorylation, possibly at just one site (Fig 5C). Most effective would be a phosphatase activity associated with the pgRNA—P protein complex, providing spatio-temporal control of the dephosphorylation event(s); HBc dephosphorylation coincident with pgRNA packaging has indeed very recently been suggested [95]. HBc dimers joining the nascent capsid, but not bulk HBc, could then as well become partly dephosphorylated, enabling stable packaging of the entire 3.5 kb pgRNA. With progressing dsDNA synthesis the co-packaged phosphatase activity could gradually release more CTD phosphoryl groups until this electrostatic buffer is emptied [13, 26, 75]. One might further speculate that timely envelopment blocks the supply of dNTPs into the capsid before the (+)-DNA gap is completely filled, and that this represents the most stable state for a DNA containing nucleocapsid ready to leave the cell as virion. Conversely, after infection of a new cell continued DNA synthesis plus re-phosphorylation might create an excess of negative charge and destabilize the capsid in preparation for uncoating.
More work will be needed to substantiate this model, but the concept as such lets previous statements on the importance of HBc phosphorylation for HBV replication appear oversimplified. Rather than directly promoting pgRNA encapsidation HBc phosphorylation seems to act indirectly by blocking competing interactions with non-specific RNA. In turn, with high-level HBc phosphorylation as default, dephoshorylation becomes as important for viral replication as phosphorylation. Hence production of replication-competent HBV nucleocapsids appears to depend on an intricately balanced level of HBc phosphorylation. The excess of empty over genome-containing HB virions [52, 60] indicates that proper execution of this program is delicate; hence even small perturbations might have severe consequences, inviting therapeutic exploitation, e.g. by kinase [96] or phosphatase inhibitors. However, such strategies will require a deeper understanding of the involved host enzymes and the substrate properties of HBc for which our study provides valuable new tools. On the basic side, these tools can feasibly be adapted to other viruses, including but not limited to relatives of HBV [97]. Preliminary data indicate, for instance, that the core protein of DHBV, a virus capable of replicating in human cells [98], is also highly phosphorylated (at eight yet to be mapped sites) by human SRPK1 in the recombinant coexpression system whereas the core protein of the fish Cichlid nackedna virus (CNDV; [97]) is not. Compatibility with the cellular kinases and phosphatases may thus well be a determinant of viral host and tissue tropism.
HBc expression vector pET28a2-HBc183 [69] carries a T7 promoter controlled synthetic HBc gene [68] for a genotype D (GenBank: V01460.1) core protein. The HBc183opt ORF used here encodes the same aa sequence but was adapted to E. coli codon usage (GeneOptimizer software; ThermoFisher/GeneArt), increasing CLP yields ~3-fold (S1 Fig). The new pRSF expression vectors are based on pRSFDuet-1 (Novagen) which harbors two T7 promoter cassettes and the lacI repressor gene for IPTG inducible coexpression of two ORFs. For HBc mono-expression the HBc183opt ORF and its derivatives were usually inserted after the 5´ T7 promoter, with simultaneous deletion of the 3´ T7 promoter cassette (plasmids pRSF_T7-HBcNNNopt; NNN specifies the last HBc aa present). For coexpression, the upstream T7 promoter was replaced by a Tet promoter, and in addition a gene for the Tet repressor was inserted into the plasmid (S2A Fig) for separate inducibility by anhydrotetracycline (AHT). Kinase ORFs were inserted under Tet promoter control, HBc ORFs under T7 promoter control (plasmids pRSF_Tet-X_T7-HBcNNNopt, where X denotes the ORF under Tet promoter control). Employed kinase ORFs encoded an N-terminally His6-tagged version of SRPK1ΔNS1 [72]; the catalytic subunit alpha isoform 1 (aa 1–351) of PKA (NP_032880.1); and the C-terminal catalytic domain (aa 397–740) of human PKC theta isoform X1 (XP_005252553.1). The latter two ORFs had previously been expressed in bacteria [79, 80] and here were obtained as E. coli expression optimized DNA strings (ThermoFisher/GeneArt). In the non-assembling CTD control constructs the HBc aa 1–145 part was replaced by the ORF for N-terminally His6-tagged eGFP [99] followed by a Gly2 linker and a recognition site for TEV protease (see Fig 4A). Cloning was done by conventional restriction-based methods or using the Q5 mutagenesis kit (NEB). All constructs were verified by Sanger sequencing.
Expression of HBc CLPs followed previously described procedures [99, 100], as detailed in reference [101]. In brief, E. coli BL21*CP served as expression host. T7 promoter and Tet promoter driven expression from the pRSF plasmids were induced by IPTG (1 mM final concentration) and/or 0.2 μg/ml AHT, respectively; cultures (usually 200 ml) were then shaken for 12–16 h at 20–25°C. Cell lysis included treatment with lysozyme, Triton X-100 and benzonase (Merck-Millipore) in the presence of protease-inhibitor cocktail (Roche) with subsequent sonication; in kinase coexpressions, Phos-Stop phosphatase inhibitor (Roche) was included. Cleared cell lysates were subjected to sedimentation (TST41.14 rotor; 41,000 rpm for 2 h at 20°C) through 10%-60% sucrose step gradients in TN150 buffer (25 mM Tris/HCl, 150 mM NaCl, pH 7.4). For longer term storage at -80°C, peak fractions were dialysed into storage buffer (50 mM Tris/HCl pH 7.5, 5 mM EDTA, 5% (w/v) sucrose, 2 mM DTT). Isolated CTD peptides from TEV-cleavable GFP fusions were obtained as detailed in S1 Protocol.
NAGE was performed in 1% agarose gels in 1x TAE buffer (40 mM Tris, 20 mM acetic acid, 1 mM EDTA) as previously described [9, 23, 99]. For routine detection of nucleic acids gels contained 0.5 μg/ml ethidium bromide (EB); subsequent protein staining was done using Coomassie Brilliant Blue R250 (CB) followed by extensive destaining in fixing buffer (50% MeOH, 7% acetic acid, v/v in H2O).
Alternatively, gels run without EB were stained for RNA using 1x Sybr Green 2 (SG2) RNA stain in TAE buffer (from 10,000x in DMSO; FMC Bioproducts), followed by protein staining with ready-to-use Sypro Ruby (SR) Protein Gel Stain (BioRad). Fluorescence signals were recorded using a laser Scanner (Typhoon FLA 7000; GE Healthcare) set at excitation 473 nm / filter Y520 nm (SG2) and 473 nm / filter O580 nm (SR). Signals were quantified using ImageQuant software. Higher quantitation accuracy was achieved by several adapations of the protocol, as described next.
Critical issues for robust signal quantitation included weak SG2 staining (especially for low RNA content CLPs), concentration-dependent variations in SR signal intensities, SG2 to SR signal carry-over during Laser scanning, and high background staining. Implemented countermeasures were to use gels no thicker than 5–6 mm; loading similar protein amounts of different CLPs and possibly two different amounts of the same CLPs; and including on each gel a well-characterized wild-type CLP preparation as standard to account for variations in staining intensity. SG2 staining over background was improved by doubling the dye concentration (1:5,000 dilution), a 1 h staining period, plus ≥3 washes with TAE buffer prior to SG2 laser scanning. For improved SR staining gels were washed twice in fixing buffer, after which the SG2 staining was no longer detectable; fixation also prevented band broadening by diffusion during the subsequent overnight incubation in SR protein gel stain. After two additional washes in distilled water SR fluorescence was recorded as described above. For each gel, the ratio of SG2: SR fluorescence intensity in the wild-type HBc183 standard was set to 100% to which the respective intensity ratios from the test CLPs were normalized. Applying this procedure reduced the calculated relative RNA content of CTD-less and SRPK1-phosphorylated CLPs from >25% of that of unmodified HBc183 CLPs (Fig 2C) to ≤10% (Fig 5C, S8 Fig).
RNA content was calculated from CLP sample absorbances at 260, 280, 340 and 360 nm [73]. For high purity CLP preparations, e.g. HBc183 CLPs from two sequential sucrose gradients [101], ten-fold dilutions in H2O were measured in 1 cm path length cuvettes in an Ultrospec 7000 instrument (GE Healthcare). For single gradient CLP preparations which may contain free RNA [101] aliquots from the respective fractions (routinely containing 0.5–4 mg/ml of HBc protein) were incubated with 1/100 (w/w) of RNase A for 30 min at room temperature, then dialyzed (30 kDa MWCO) against TN 150 buffer. UV/VIS absorbances of the resulting solutions were then monitored without further dilution in a QIAxpert instrument operating with 0.1 cm path length (2 μl volume) microcuvettes. Molar nucleotide per core protein monomer (N/P) ratios were calculated as described [73].
Mn2+-Phos-tag SDS-PAGE was performed by adding Phos-tag acrylamide AAL-107 (Wako Pure Chemical Corporation) plus MnCl2 to the acrylamide solutions for conventional Lämmli SDS-PAGE resolving gels, as recommended by the manufacturer. In our hands, the best separation of differently phosphorylated HBc183 species was obtained using 15% acrylamide gels supplemented with 75 μM Phos-tag acrylamide and 150 μM MnCl2.
Immunoblotting after NAGE or conventional SDS-PAGE was performed as previously described [69, 102]. Low transfer efficiency to polyvinylidene difluoride (PVDF) membrane after Phos-tag SDS-PAGE gels was improved by sequentially soaking the gels for 10 min each in transfer buffer A (39 mM glycine, 48 mM Tris, 3.75% (w/v) SDS, 20% (v/v) MeOH) containing 10 mM EDTA, 1 mM EDTA, and no EDTA, plus the use of a wet blot rather than a semi-dry transfer system. HBc specific antibodies employed were the anti-assembly domain mouse mAbs 312 [103] and 1D8 [82], both recognizing a linear epitope exposed on intact CLPs and SDS-denatured HBc protein; the capsid-specific mAbs 275 [104] and 3120 [105], obtained from Tokyo Future Style, Inc. (catalogue no.: 2AHC22); and the anti-phospho-CTD mAb T2212 [51], also from Tokyo Future Style (catalogue no.: 2AHC23). HBsAg on NAGE blots was detected by human mAb 9H9 [85] which recognizes a conformational epitope in S. Bound mAbs were visualized using horse raddish peroxidase (PO), either as direct mAb conjugate or via appropriate secondary antibody-PO conjugates, and chemiluminescent substrates.
About 5 μg of sucrose gradient enriched HBc CLPs were incubated for 30 min with increasing amounts of a 6x DNA loading buffer containing 0.48% SDS (NEB Purple) to yield final SDS concentrations from 0.024% to 0.24%; samples incubated in a non-denaturing 6x loading buffer served as control. Reaction products were monitored by NAGE in 1% gels; EB fluorescence signals were recorded using a laser scanner (excitation 532 nm / filter O580 nm). Proteins were subsequently stained by CB.
For partial trypsin digests the respective HBc CLPs were incubated with 1/100 (w/w) the amount of sequencing grade trypsin (Promega) in TN50 buffer (25 mM Tris/HCl, 50 mM NaCl, pH 7.5) at 30°C. At various time points aliquots were withdrawn and digestion was stopped by adding 4-(2-aminoethyl)-benzenesulfonyl fluoride hydrochloride (AEBSF; Applichem) to a final concentration of 0.2 mM. Formation of cleavage products and intactness of particles were analyzed by SDS-PAGE and NAGE.
MS analyses were performed at the Institut de Biologie et Chimie des Protéines, UMR5086, CNRS/Université Lyon 1, France on a Sciex Voyager DE-PRO MALDI-TOF instrument, using sucrose gradient enriched CLPs diluted 1:100 (v/v) in sinapic acid as matrix.
Electron microscopy was carried out in the Edinburgh, UK, cryoEM facility. Detailed procedures for sample preparation, micrograph recording, particle selection, image processing, refinement, difference map calculation, and molecular fitting are given in S2 Protocol. In brief, gradient-enriched CLP samples were vitrifed as described [106] using a manual freezing apparatus with an environmental chamber [107] at room temperature. Before use, grids (Quantifoil R1.3/1.2) were glow discharged in air for 1 min with a current of 30 μA using a Quorumtec sputter coater. Micrographs were recorded on a FEI Tecnai F20 microscope operated at 200 kV and a 8192 pixels x 8192 pixels CMOS camera (TVIPS F816). Only particles with round shape and crisp appearance were subsequently used for reconstructions. Resolution and B-factors were calculated with Relion [108]. Fitting was done using Chimera [109], based on pdb 1QGT [16].
The human hepatoma cell line HepG2 and its TetOFF HBV derivative HepG2.117 [84] were cultured as described [84, 110]. The stable, constitutively HBc183 producing cell line H4-15 was derived from the human hepatoma cell line Huh7 by CRISPR/Cas9-mediated homologous recombination of a HBc expression cassette into the AAVS1 locus (Peter Zimmermann, MSc thesis; University of Freiburg, 2017).
Cytoplasmic HBV capsids were obtained from TetOFF HepG2.117 cells cultured for 10 days in the absence of doxycycline [84]. Cytoplasmic lysates were prepared using NP40 lysis buffer (50 mM Tris/HCl, 140 mM NaCl, pH 8, with 0.5% (v/v) NP40 detergent) supplemented with Phos-Stop. To enrich capsids, lysates were sedimented through a 10% - 60% (w/v) Nycodenz step gradient for 2 h at 4°C in a TST 41.14 rotor at 41,000 rpm (302,500 g) and harvested in 14 fractions. Detection of HBV DNA in capsids and by Southern blotting was performed using 32P labeled HBV DNA probes as described [84, 111]. Extracellular particles in the culture supernatants of HepG2.117 and H4-15 cells were enriched by precipitation with PEG8000 [84, 112].
Unless indicated otherwise data are expressed as mean ± standard deviation (SD) from ≥3 experiments. Comparisons between multiple groups were performed using One-way ANOVA and Tukey´s post test (Graphpad Prism 5). Differences between means of two paired groups were evaluated using Student's t-test. P-values of p<0.05 were regarded as statistically significant.
|
10.1371/journal.pgen.1006823 | Allelic variants of OsHKT1;1 underlie the divergence between indica and japonica subspecies of rice (Oryza sativa) for root sodium content | Salinity is a major factor limiting crop productivity. Rice (Oryza sativa), a staple crop for the majority of the world, is highly sensitive to salinity stress. To discover novel sources of genetic variation for salt tolerance-related traits in rice, we screened 390 diverse accessions under 14 days of moderate (9 dS·m-1) salinity. In this study, shoot growth responses to moderate levels of salinity were independent of tissue Na+ content. A significant difference in root Na+ content was observed between the major subpopulations of rice, with indica accessions displaying higher root Na+ and japonica accessions exhibiting lower root Na+ content. The genetic basis of the observed variation in phenotypes was elucidated through genome-wide association (GWA). The strongest associations were identified for root Na+:K+ ratio and root Na+ content in a region spanning ~575 Kb on chromosome 4, named Root Na+ Content 4 (RNC4). Two Na+ transporters, HKT1;1 and HKT1;4 were identified as candidates for RNC4. Reduced expression of both HKT1;1 and HKT1;4 through RNA interference indicated that HKT1;1 regulates shoot and root Na+ content, and is likely the causal gene underlying RNC4. Three non-synonymous mutations within HKT1;1 were present at higher frequency in the indica subpopulation. When expressed in Xenopus oocytes the indica-predominant isoform exhibited higher inward (negative) currents and a less negative voltage threshold of inward rectifying current activation compared to the japonica-predominant isoform. The introduction of a 4.5kb fragment containing the HKT1;1 promoter and CDS from an indica variety into a japonica background, resulted in a phenotype similar to the indica subpopulation, with higher root Na+ and Na+:K+. This study provides evidence that HKT1;1 regulates root Na+ content, and underlies the divergence in root Na+ content between the two major subspecies in rice.
| Despite intensive research, few genes have been identified that underlie natural variation for salinity responses in rice. In this study, we used a rice diversity panel for genome wide association mapping to identify HKT1;1 as a factor regulating Na+ distribution. Within the rice diversity panel we observed higher Na+ levels in root tissue in the indica subpopulation compared to japonica accessions. Three non-synonymous variants were identified within HKT1;1 that were associated with altered Na+ accumulation in root tissue, and displayed contrasting frequencies between indica and japonica subspecies. The introduction of HKT1;1 from an indica accession that contained the three non-synonymous variants into a japonica background resulted in a phenotype similar to that exhibited by the indica subpopulation. This work suggests that these allelic variants are likely responsible for the higher root Na+ observed in indica accessions. This study has identified a genetic resource for modifying Na+ content rice, and provides evidence that HKT1;1 underlies the divergence between indica and japonica subspecies in root Na+ content.
| Salinity is a widespread limitation for agricultural productivity, especially for irrigated agriculture and coastal lowlands prone to seawater ingress [1,2]. By definition, salinity occurs when there is a high concentration of soluble salts in soil [3]. More than 800 million hectares worldwide is affected by salt, which accounts for 6% of the total land area [3]. Besides natural causes such as rising sea levels during the dry and wet cropping seasons, the poor quality of irrigation water and improper drainage, also collectively increases soluble salt concentration in the root zone [2,4].
Rice (Oryza sativa L.) is one of the most important crop species and is a staple food for more than half of the world’s population. Salinity is a major impediment to increasing production in many rice growing regions, including temperate and tropical environments, around the world [5,6]. Rice is the most salt-sensitive species among major cereal crops [3]. The susceptibility of rice to salinity stress varies with growth stages [7,8]. Rice is less sensitive to saline conditions at germination, active tillering and maturity stage [7,9,10]. Vegetative growth during the early seedling stage is highly sensitive to saline conditions, and often translates to reduced stand density in salt-affected fields [11,12]. Some rice varieties are most sensitive to salt stress during early tillering and panicle initiation stages of growth [8]. This developmentally-dependent salt-sensitivity, in context of yield reduction, was associated with a significant decrease in tiller number per plant, spikelet number per panicle, fertility, panicle length and primary branches per panicle [7,8,13,14].
Despite the overall high salt-sensitivity of rice, several studies have demonstrated that considerable natural variation for salinity tolerance exists in rice germplasm [15,16]. Traditional landraces or cultivars such as ‘Pokkali’, ‘Nona Bokra’, ‘Cheriviruppu’ and ‘SR26B’ have originated or have been selected in coastal regions and are more tolerant to saline conditions [5,12,17,18]. Quantitative trait loci (QTL) underlying salinity tolerance have undergone intensive investigations [16,18–24]. Although many QTL have been identified across the rice genome, the most well-characterized QTL is Saltol/SKC1, which harbors HKT1;5, on the short arm of chromosome 1 [18,20–22]. The SKC1 gene (HKT1;5) was subsequently cloned from a salt-tolerant indica landrace, ‘Nona Bokra’, and encodes a Na+ transporter that regulates shoot Na+:K+ homeostasis during salt stress [25].
Salinity tolerance is a complex polygenic trait, and several physiological mechanisms, including tissue tolerance, sodium exclusion, osmotic stress tolerance, and tissue-specific sodium sequestration can be utilized for improving salinity tolerance [3]. While many QTL have been reported for salinity tolerance in rice, few studies have identified the causal genes and confirmed the importance of these resources for improving salinity tolerance. Hence, the genetic resources (QTL and genes) available to rice breeders for improving salt tolerance are limited. Identification of loci that regulate salt accumulation and/or distribution will enable the introgression of favorable genic combinations and greatly accelerate the development of robust salt-tolerant rice varieties.
Genetic variation within the rice germplasm collection can be utilized to identify important loci controlling variation for salinity tolerance through genome-wide association (GWA) analysis, which provides greater mapping resolution and evaluates greater allelic diversity compared to linkage mapping strategies [16,24,26,27]. In this study, we used GWA to investigate the genetic architecture of salinity tolerance using the Rice Diversity Panel 1 (RDP1) [28–30]. RDP1 is comprised of 421 accessions collected from 85 countries and was developed to identify alleles associated with morphological, physiological and agronomic traits [28–30]. RDP1 captures much of the diversity in the rice germplasm collection worldwide [28–30]. We used several quantitative measures to characterize the rice diversity panel for physiological and morphological responses to salinity stress. Here, we show that allelic variants of a sodium transporter (HKT1;1) underlie natural variation for root Na+ content in rice. Using a multifaceted approach, we demonstrate that variants within HKT1;1 alter Na+ transport and can explain the basis of divergence in root Na+ content between the indica and japonica subspecies of cultivated rice.
To assess the degree of natural variation for salinity tolerance associated traits in rice, a 9 dS·m-1 (~90 mM NaCl) salt stress was imposed gradually over a period of four days (in four increments of 20–30 mM) to two-week old rice seedlings. A rice diversity panel, consisting of 390 rice accessions (383 from RDP1 and seven check varieties), was scored for ten phenotypic traits at the end of a two-week 9 dS·m-1 stress period (the plants were 28 days old). The ten traits recorded were root biomass (control and salt conditions), shoot biomass (control and salt conditions), root and shoot Na+ content, root and shoot K+ content, and root and shoot Na+:K+ (S1 File). To control for inherent differences in growth rate between lines, we expressed the saline-induced growth response for each accession as the ratio of biomass in saline conditions over biomass in control conditions. Broad sense heritability for the eight phenotypic traits ranged from 0.32 for root K+ content and 0.83 for shoot biomass in control conditions (S1 Table).
To examine the relationships between each of the eight traits, Pearson correlation analysis was performed across all accessions. No significant relationship was observed between shoot biomass and ion traits (S3 Table). Moreover, root and shoot ion content showed no significant relationship when the analysis was performed with all accessions (S3 Table). Due to the deep population structure in rice, correlation analysis was also performed for each of the five major subpopulations in RDP1 (here, admixed accessions were considered a separate subpopulation; S4–S8 Tables) [29,31]. Root growth response (the ratio of root biomass in salt to control) showed a weak negative correlation with shoot Na+:K+ in admix and tropical japonica (trj) accessions (S4 and S5 Tables, respectively). In trj, aus, and tej subpopulations significant, albeit weak, positive correlations were observed between shoot Na+ and root Na+:K+ (S5–S7 Tables).
Comparisons between each of the subpopulations showed significant differences for shoot and root Na+, K+ and Na+:K+ (Fig 1). Indica accessions exhibited significantly higher root Na+ content and Na+:K+ compared to the other four subpopulations (Fig 1A and 1B). Significantly lower shoot Na+ and Na+:K+ were observed in indica and aus subpopulations compared to temperate japonica (tej), tropical japonica (trj) and admix accessions (Fig 1C and 1D). These results suggest that there are inherent differences in root and shoot ion homeostasis between subpopulations, with indica accessions generally displaying higher root Na+ and Na+:K+, and indica and aus accessions exhibiting lower shoot Na+ content and Na+:K+.
To identify loci associated with salt tolerance-related phenotypes, GWA mapping was conducted using 397,812 SNPs and eight salinity-related phenotypes collected on 365 rice accessions (Fig 2; S1–S4 Figs) [32]. A linear mixed model implemented in EMMA was used for the association analysis [33]. A total of 90 highly significant QTL (245 SNPs; p < 10−5) were identified for salinity-related traits with the strongest associations detected for root Na+ content followed by root Na+:K+ (Fig 2A and 2B, respectively). A region located at ~30.6 Mb on chromosome 4 was found to have the largest effect and explained 15% of the phenotypic variation beyond that explained by population structure for root Na+:K+ (S2 File). An additional 25% of the phenotypic variance for root Na+:K+ was explained by population structure suggesting that this trait may be heavily influenced by the differences between the major subpopulations in rice.
For each trait, the number of significant QTL ranged from 3–24, with the highest number of QTL identified for root biomass ratio (24 QTL). Many of these QTL had small effects, explaining ~4.7–7.5% of phenotypic variation for root growth response. These results indicate a polygenetic architecture for root growth responses to salinity. A large number of QTL with minor effects (explaining < 7% phenotypic variation) were identified for shoot Na+ content and Na+:K+, suggesting a polygenic architecture for these traits in rice. This trend was observed for all traits, with the exception of root Na+ and Na+:K+, suggesting that salinity tolerance in terms of growth and shoot ion homeostasis in rice is regulated by many loci with small effects. Twenty QTL were commonly detected for two or more traits. Shoot Na+ and Na+:K+ showed the largest number of shared QTL (12 QTL), however much of this similarity is likely driven by the strong phenotypic and genetic correlation observed between these traits within tissues (Table 1; S2 Table).
The most significant QTL for root Na+ content and Na+:K+, named Root Na+ Content 4 (RNC4), spans a region of ~575 Kb (30,481,871–31,057,205) on chromosome 4 (Fig 3A). To characterize this region further and identify candidate genes that may be underlying natural variation for this trait, this region was segmented into haplotype blocks and the contributions of each block to root Na+ content and Na+:K+ were determined using ANOVA. A total of 36 blocks were identified in this 575 Kb region (S5–S8 Figs). A single 9.7 Kb block from 30,727,920–30,737,580 bp was found to have the largest contribution to root Na+ and Na+:K+, with approximately 16% of phenotypic variation explained for root Na+ and 17.5% explained for root Na+:K+ (Fig 3B; Table 2). The region spanning from the 5’ boundary of block 2 to the 5’ boundary of block 3 harbored only two genes, both of which were annotated as sodium transporters, HKT1;1 and HKT1;4 (LOC_Os04g51820 and LOC_Os04g51830 respectively; Fig 3C).
To further characterize HKT1;1 and HKT1;4, the expression patterns of both genes were examined in twelve tissues at three developmental time points (early seedling, early tillering and anthesis). The expression of both HKT1;1 and HKT1;4 were higher in leaf tissue compared to root tissue during the seedling stage (Fig 4, S9 Fig). However, the expression of HKT1;1 and HKT1;4 within aerial tissues differed across developmental stages. HKT1;1 was highly expressed in the leaf blade and leaf sheath during the early seedling stage (Fig 4A). HKT1;4, on the other hand displayed the highest expression during reproductive stage, specifically in culm tissue at ~7 days after anthesis (Fig 4B).
To examine whether transcript abundance may be a component of the phenotypic differences observed between allelic groups at RNC4, RNA sequencing was performed on shoot tissue of 32 accessions in control and saline conditions, and the expression of both genes was compared between allelic groups at RNC4 (Fig 4C and 4D; S3 File). For both genes, accessions that showed higher root Na+ content (T allele at SNP-4-30535352), also showed higher expression in both control and saline conditions compared to accessions with low root Na+ content (G allele at SNP-4-30535352). The expression of HKT1;1 was approximately 92% higher in high root Na+ lines in control conditions compared to low root Na+ lines, while a 44% higher expression was observed in saline conditions (Fig 4C). While the overall expression level was much lower for HKT1;4 compared to HKT1;1, a similar trend in gene expression was also observed between the two allelic groups of HKT1;4 (Fig 4D). A 46% and 57% higher expression was observed in lines with high root Na+ content compared to lines with low root Na+ content in control and saline conditions, respectively (Fig 4D). These results suggest that differences in expression of HKT1;1 and/or HKT1;4 may be a component underlying variation in root Na+ content at RNC4.
To determine if these two HKTs within RNC4 regulate Na+ content during salinity stress at the early tillering stage, three independent RNA-interference (RNAi) lines were generated for both genes. Transcript levels in the leaf tissue was reduced by approximately 2.9–6.2 and 2–2.2 fold in HKT1;4RNAi and HKT1;1RNAi lines compared to wild-type (WT) ‘Kitaake’, respectively (S10 Fig). A 9 dS·m-1 (~90 mM NaCl) was gradually imposed at 10 DAT for 14 days to replicate the stress treatment for the large-scale screening. Reduced expression of HKT1;1 had severe phenotypic effects on shoot and root ion homeostasis as well as shoot and root growth under salinity. Shoot Na+ and Na+:K+ were 31–41% and 27–41% higher, respectively, in HKT1;1RNAi lines compared to WT (p < 0.0001, p < 0.05 respectively; Fig 5A–5C). A 21–27% reduction in root Na+ was observed in HKT1;1RNAi and 31–33% lower root Na+:K+ was observed in HKT1;1RNAi compared to WT (p < 0.05 and p < 0.0001, respectively; Fig 5D–5F). In RNAi plants, shoot and root growth was reduced by 44–55% and 78–72% respectively in salt treated plants relative to those in control conditions, while in WT a 26% and 45% reduction in shoot and root growth, respectively was observed in WT plants (S11 Fig). No differences were observed between HKT1;4RNAi and WT plants (Fig 5, S11 Fig). These results suggest that HKT1;1 may influence the shoot and root Na+ content during the early tillering stage, and is likely the causal gene underlying RNC4.
To determine whether there were sequence differences between allelic groups at RNC4, sequencing data was mined for variants in HKT1;1 (S4 File). Nine variants were detected in the coding region of HKT1;1 with four SNPs resulting in non-synonymous amino acid substitutions in HKT1;1 (Fig 6; S12 Fig) [34]. Of the nine variants, only M4 displayed a significant deviation from the expected frequency in the minor allelic group, indicating that it is unlikely to be important for the high root Na+ phenotype exhibited by accessions in the minor allelic group (Pearson’s chi squared test, p < 1.26 x 10−5). The remaining three non-synonymous mutations (M3, M5 and M8) were detected in thirteen accessions all belonging to the minor allelic group, which is characterized by high root Na+ content, at the most significant SNP for root Na+ content (SNP-4-30535352). The higher frequency of these three non-synonymous mutations observed in minor allele accessions (T) suggests that allelic variation in HKT1;1 could be a component in the genetic basis of the observed difference in root Na+ content between major and minor alleles. No sequence differences in HKT1;4 were observed between allelic groups at RNC4.
To characterize the biophysical properties of the two major isoforms identified between allelic groups at RNC4, HKT1;1 was isolated from two representative accessions, ‘Nipponbare’ and ‘Zhenshan 2’, which have the reference and the three non-synonymous mutations at the three locations (M3, M5 and M8), respectively (S12 Fig). At the transporter structure level, two non-synonymous SNPs (M8 and M5) lead to amino acid substitutions in cytosolic regions of HKT1;1: proline to leucine within the N-terminal cytosolic region, phenylalanine to serine in the cytosolic loop between the first and second transmembrane segment-pore region-transmembrane segment (MPM) domains (Fig 6A and 6B; S12 Fig). The third non-synonymous SNP results in an asparagine to serine substitution in the external part of the pore-forming region of the second MPM (Fig 6A and 6B; S12 Fig). Functional analysis was performed by voltage-clamp electrophysiology using Xenopus oocytes for the two variants of HKT1;1 (Fig 6). The amount of expressed transporters targeted to the oocyte membrane was similar for the two variants, as indicated by the mean GFP fluorescence intensity emitted by either of the tagged transporters at the membrane (Fig 6C and 6D). In agreement with previous reports, both isoforms of HKT1;1 displayed low affinity, high Na+ versus K+ selectivity, inward rectifying activity and no time-dependent kinetics (Fig 6E; S13 and S14 Figs) [34]. However, the two allelic variants displayed considerable differences in Na+ transport activity. The variant from the accessions with high root Na+, HKT1;1-Zh, exhibited higher inward (negative) currents compared to that from ‘Nipponbare’, HKT1;1-Ni (Fig 6E and 6F), essentially due to a less negative voltage threshold of inward rectifying current activation by 20–25 mV in all ionic conditions (Fig 6E and 6F, S13 and S14 Figs). This latter feature was especially expected to favor transport activity of HKT1;1-Zh compared to HKT1;1-Ni during salinity stress where the high concentration of Na+ in the apoplast results in a depolarization of the plasma membrane [35,36]. Thus, at a weak negative voltage, the current could be more than six-fold higher in HKT1;1-Zh, compared to HKT1;1-Ni (Fig 6G).
To determine if these differences in transport activity have physiological effects in vivo, native overexpression lines were generated for each variant (HKT1;1Ni, HKT1;1Zh). A ~4.3 kb genomic region was isolated from ‘Nipponbare’ and ‘Zhenshan 2’, which included the entire CDS of HKT1;1 and a 1.9 kb promoter, and was expressed in ‘Kitaake’. The endogenous HKT1;1 in ‘Kitaake’, at the protein level, is identical to HKT1;1-Ni, and thus lacks the three non-synonymous variants. Two independent transformants for each variant (HKT1;1Ni, HKT1;1Zh), each containing only a single copy of the transgene, were evaluated under a 9 dS·m-1 salt stress for a period of two-weeks. The expression of HKT1;1Zh resulted in an increase in root Na+ and Na+:K+ compared to HKT1;1Ni, while no differences were observed between variants for root K+ (Fig 7). A considerable increase in both root Na+ and Na+:K+, as well as a reduction in root K+ was observed in both native overexpression lines (HKT1;1Ni and HKT1;1Zh) compared to ‘Kitaake’, which is opposite to the root phenotype observed in the HKT1;1RNAi lines (Fig 7). However, expression under the native promoter had no effects on shoot Na+ or Na+:K+ (Fig 7). Together, these results provide further evidence that HKT1;1 is responsible for the higher root Na+ phenotype, and that the difference in Na+ content between the allelic groups at RNC4 is likely due to functional differences in Na+ transport by HKT1;1 alleles, with the three non-synonymous SNPs in HKT1;1-Zh resulting in higher Na+ transport activity.
A difference in allele frequencies of the three non-synonymous mutations in HKT1;1 was observed between the major subpopulations in the 32 sequenced accessions of RDP1. However, since it was difficult to examine subpopulation differentiation with this small of a sample size, the differences were explored in more depth using resequencing data from a larger diversity panel of 3,024 accessions [37]. A total of 206 SNPs spanning a ~38 Kb region around HKT1;1 was used for haplotype analysis. In agreement with the allele frequency observed in the 32 accessions of RDP1, a clear differentiation could be observed between indica (ind1A, ind1B, ind2, ind3 and indx) and japonica (temp, trop1, trop2 and japx) subspecies in the larger diversity panel (Fig 8). Haplotypes H1, H5 and H8 harbored the three non-synonymous alleles and were found in nearly 85% of the indica accessions. The sequence similarity between high root Na+ haplotypes was very high, ranging from ~88–94% identity. Haplotypes containing high root Na+ alleles of HKT1;1 were also found in the japonica (temp, trop1, trop2 and japx), aus and aromatic subpopulations, albeit at a much lower frequency (0–3%). In contrast, haplotypes H2, H3, H4, and H7 were found predominately in the japonica accessions and lacked the high root Na+ allelic form of HKT1;1. Within the low root Na+ group, haplotypes exhibited high sequence similarity (~65–94%). Given the clear divergence between indica and japonica for HKT1;1 haplotypes and the effects of HKT1;1 isoforms on root Na+ content, collectively these results strongly suggest that a significant proportion of the difference between rice subpopulations in root Na+ in RDP1 is due to differences in frequency of HKT1;1 variants.
Given the contrasting haplotype frequencies of high and low root Na+ variants of HKT1;1 between subpopulations of cultivated rice, we explored the origins of these haplotypes by examining their frequencies in a collection of 446 Oryza rufipogon accessions collected throughout South and Southeast Asia [38]. These accessions represent three major populations (Or-I, Or-II and Or-III) and provide an adequate representation of the ancestral populations of cultivated rice [38]. Two haplotypes (H1 and H5) were identified that harbored the high root Na+ variants of HKT1;1, and were found in nearly 70% of the O. rufipogon accessions. The H1 haplotype displayed the highest frequency in the Or-II clade and was also found in the majority of indica accessions, suggesting that the indica allele is likely derived from Or-II. In contrast, two haplotypes (H2 and H6) were identified with the low root Na+ variant and were present in only 19% of the O. rufipogon accessions. The H6 haplotype was the most frequent and present in 18% of the O. rufipogon accessions, but absent from the japonica cultivated rice accessions. In contrast, H2 occurred at high frequency (44%) in cultivated japonica, particularly the tropical japonica subpopulation, suggesting that H2 is potentially the ancestral haplotype for the japonica subspecies. Interestingly, the haplotypes found at high frequencies in the japonica subspecies were present at considerably lower frequencies in wild rice accessions (the highest frequency observed was 0.16), indicating that these haplotypes in japonica subspecies may be derived from a relatively small population of wild progenitors.
Salinity tolerance is a complex polygenic trait and is regulated by several physiological mechanisms [3]. Salinity reduces plant growth through osmotic effects, which are experienced shortly after the addition of Na+ to the external media, and ionic effects, which are experienced later in the stress as Na+ accumulates in the leaves to toxic levels. The ability to maintain growth in saline conditions involves a suite of physiological mechanisms including osmotic adjustment, the exclusion of sodium from leaf tissues by sequestration in the root or leaf sheath, the storage of Na+ into vacuoles or partitioning in tissues where the toxic effect of Na+ is reduced [3]. In this study, the complex polygenic nature of salinity tolerance in rice is evidenced by the large number of loci with small effects identified for shoot and root growth responses.
Although shoot Na+ exclusion is often used as a parameter for salt tolerance, the relationship between low shoot Na+ and the ability to maintain growth in saline conditions does not always hold true [39,40]. Here, no significant relationships were observed between ion traits and growth responses across all the subpopulations, suggesting that in the current experimental conditions other tolerance mechanisms besides Na+ exclusion may be important for salt tolerance in rice. In low to moderate salinity, the osmotic effects of high Na+ in the external media are likely to have a much greater impact on plant growth, compared to ionic effects [3]. During the ionic phase of salt stress, Na+ must accumulate to toxic levels to cause cell death and impede growth. Thus, ionic tolerance mechanisms begin to play a role much later in the stress, or when the concentration of Na+ in the external media is high. Other studies that have exposed diverse rice accessions to higher concentrations of NaCl and/or for longer periods have reported weak to moderate relationships between ion traits and growth responses to salinity [24,41,42]. Thus, Na+ exclusion may be important during more severe stress treatments than was used in the current study. The relatively moderate salt stress imposed in the current study may not be enough for Na+ to accumulate to toxic levels to significantly inhibit growth, and may partially explain the lack of correlation, both phenotypic and genetic, between ion traits and growth responses.
RNC4 harbors two Na+ transporter genes, HKT1;1 and HKT1;4. HKTs are well-known components of salinity tolerance in several plant species including rice (HKT1;5 is likely the causal gene in the SalTol QTL), wheat and Arabidopsis [25,43–51]. Although both HKT1;1 and HKT1;4 displayed significant differences in expression between allelic groups at RNC4, several key findings suggest that HKT1;1 is more important for root Na+ content during the early tillering stage and for the salinity level imposed in our experimental set-up. First, the genes are expressed at different developmental stages. HKT1;1 was expressed at the highest levels in blade and leaf sheath tissues of seedlings, while HKT1;4 showed the highest expression in culms of mature plants (Fig 4B). Second, reduced expression of HKT1;1 in transgenic RNAi lines resulted in a greater sensitivity to salinity compared to WT, while HKT1;4RNAi and WT plants displayed similar phenotypes under salinity (Fig 5). In a recent report, Suzuki et al showed that HKT1;4 is primarily expressed in peduncles during flowering (14 week old plants) and, through RNAi, showed that HKT1;4 is primarily involved in Na+ homeostasis only during the reproductive phase [51]. Since the current study was conducted during the early tillering stage (< 1 month old plants), it is unlikely that this gene would have an impact on salinity tolerance in this developmental window. Finally, increased expression of HKT1;1 with the native promoter resulted in higher Na+ in root tissue, which is identical to the phenotype associated with RNC4. Together, these data suggests that HKT1;1 is the causal gene underlying RNC4 and contributes to root Na+ content during the early tillering stage.
The differences in Na+ content observed between allelic groups at RNC4 is likely due to functional differences in Na+ transport by HKT1;1 alleles, with the three non-synonymous SNPs in HKT1;1-Zh resulting in higher Na+ transport activity. Na+ transport occurred at less negative voltages in the isoform found in accessions with high root Na+ compared to that isolated from accessions with low root Na+. During salt stress, the accumulation of Na+ in the apoplastic space increases HKT1;1 Na+ transport activity, the apparent affinity for Na+ of this transporter type is particularly low (Km ~ 80 mM; S12 Fig), but in the meantime, uptake of Na+ from the apoplast results in membrane depolarization, which reduces HKT1;1 conductance due to inward rectification property [34]. In the high root Na+ isoform, a higher (less negative) voltage threshold of current activation was observed, for instance in the presence of 10 mM external Na+ noticeable Na+ transport was observed between -75 and -90 mV, while in the low root Na+ isoforms, activation occurred at more negative voltages (Fig 6C and 6D). Thus, lower Na+ concentrations are required to induce Na+ uptake in the high root Na+ isoform of HKT1;1. In summary, the enhanced ability to transport Na+ in accessions harboring the high root Na+ isoform of HKT1;1 is likely due to the early activation of Na+ transport.
Indica varieties have long been recognized to as a source of salt tolerance, largely due to Na+ exclusion from leaf tissue. The most widely used QTL, SalTol, was identified by Lin et al. using a biparental population derived from the salt tolerant indica landrace ‘Nona Bokra’ and sensitive japonica variety ‘Koshihikari’ [21]. Tolerance mediated by SalTol is associated with the exclusion of Na+ from shoot tissue, through the removal of Na+ from the xylem and sequestration in xylem parenchyma cells in the root tissue [18,25]. While several studies have demonstrated that the indica subspecies harbors many varieties exhibiting high shoot Na+ exclusion ability, tolerant alleles in SalTol have only been utilized from a few indica landraces, and it is likely that other loci are contributing to Na+ exclusion in the indica subspecies [12,52].
In agreement with previous studies, a considerable difference among the five subpopulations was observed in root and shoot Na+ content and Na+:K+, with indica accessions generally displaying higher root Na+ content and Na+:K+, as well as slightly lower shoot Na+ and Na+:K+. The relationship between root and shoot ion traits (specifically Na+ and Na+:K+) differed considerably within each of the subpopulations. For instance, positive correlations were observed between tissues for Na+ and Na+:K+ in the tej, trj and aus subpopulations. However, in the indica and admix subpopulations no relationships were observed between tissues for Na+ and Na+:K+. The moderate positive genetic correlation observed between tissues across all accessions of RDP1 indicates that these traits may be regulated in part by common genes. However, this may be highly dependent on the subpopulation. The high frequency of the Na+ accumulating isoform for of HKT1;1 in the indica and admix subpopulations may “uncouple” the relationship between tissues for Na+ and Na+:K+.
The contrasting root Na+ content observed between indica and japonica accessions of RDP1 is consistent with the differences in transport activity and the frequencies of the high and low root Na+ isoforms of HKT1;1. The haplotypes of HKT1;1 could be clearly separated into two distinct groups, corresponding to the japonica (H2, H3, H4 and H7) and indica predominate forms (H1 and H5). The high root Na+ haplotypes (H1, H5 and H8) were most frequent in Oryza rufipogon, while the low root Na+ haplotypes were identified in only ~31% of the Oryza rufipogon accessions and were nearly fixed in japonica accessions. The two major subspecies of Oryza sativa were domesticated from two geographically isolated populations of Oryza rufipogon [38,53]. The low diversity in japonica germplasm reported by several studies is consistent with a bottleneck during domestication, and suggests that the japonica subspecies may be derived from a relatively small founding population of Oryza rufipogon [38,54–56] (S15 Fig). Although the high root Na+ isoform was found in ~30% of the Or-III subpopulation, the founding subpopulation of Oryza rufipogon, it is plausible that the bottleneck experienced during domestication may have resulted in the loss of the high root Na+ HKT1;1 variant from japonica subspecies.
Like many other HKT members, HKT1;1 is well-expressed in the vascular tissue of the shoot, and to a lesser extent in the root [34,48,50]. In the current study, HKT1;1RNAi lines were more sensitive to salt stress, and exhibited higher shoot Na+ content and lower root Na+ content compared to WT plants. The expression patterns of HKT1;1, as well as the phenotypes exhibited by HKT1;1RNAi lines are in agreement with those reported by Mäser et al for AtHKT1;1 in Arabidopsis, suggesting that the genes may have similar physiological functions [57]. Like HKT1;1RNAi, athkt1;1 knockout mutants are hypersensitive to salt stress and exhibit higher shoot Na+ and lower root Na+ [57,58]. In rice, Wang et al showed that hkt1;1 knockout mutants accumulate Na+ in xylem sap and display a reduction in Na+ in phloem sap compared to WT [50]. These observations together with the observed accumulation of Na+ in shoot tissue prompted Wang et al to suggest that HKT1;1 may regulate sodium exclusion from the shoot of seedlings possibly through xylem-to-phloem or parenchyma-to-xylem transfer of Na+ [50]. Such xylem-to-phloem transfer of Na+ by a HKT member has been debated in Arabidopsis [44,45,58]. In agreement with hkt1;1 mutant phenotype reported by Wang et al, athkt1;1 knockout mutants also exhibit higher xylem Na+ and lower phloem Na+ [44,45,50]. Although AtHKT1;1 was initially proposed to function in the recirculation of Na+ from the root to the shoot (via loading of Na+ into the phloem in the shoots), Sunarpi et al later proposed that AtHKT1;1 functions primarily in the removal of Na+ from the xylem sap and eventually to the phloem through symplastic diffusion [44,45]. However, a later study showed that AtHKT1;1 was primarily involved in the retrieval of Na+ from the xylem in root tissue, and suggested that the function of AtHKT1;1 in shoot tissue may be dependent on the experimental conditions (discussed in [3]) [58]. For the case of HKT1;1 in rice, further studies (outside the scope of this manuscript) are required to provide the exact mechanism for the regulation of root Na+ content and/or shoot Na+ exclusion.
Given the phenotypes exhibited by HKT1;1RNAi lines, as well as the proposed function described by Wang et al., the absence of an association of HKT1;1 with shoot Na+ or Na+:K+ is surprising [50]. If HKT1;1 regulates retrieval of Na+ from the parenchyma or xylem in shoot tissues, one would expect that the high root Na+ allele would also have a large impact on shoot Na+ content. However, the concentration of Na+ in shoot tissue is likely more dependent on the amount of Na+ loaded into the xylem, and thus mechanisms which limit the delivery of Na+ to xylem stream would likely be more effective mechanism for shoot Na+ exclusion [3]. Without an effective mechanism to limit Na+ entry into the xylem stream in the root, very high expression of HKT1;1, or a highly active variant of HKT1;1 would likely be necessary to reduce shoot Na+ content. While the indica (high root Na+ content) variant of HKT1;1 displayed higher transport activity compared to japonica variant (low root Na content), it is likely that these biophysical differences are not sufficient to have an impact on shoot Na+ content.
Other members of the HKT family have been identified that are expressed in the vascular tissue of the root, and primarily function to remove Na+ from the xylem to limit the delivery of Na+ to the shoot. In rice, this function is largely achieved through the action of HKT1;5 [25,59]. In contrast to HKT1;1, HKT1;5 is mostly expressed in the root and therefore is essentially involved in xylem sap desalinization [25]. In the current study, the SalTol QTL that harbors SKC1/HKT1;5 explained only a small portion of phenotypic variation for shoot Na+ and shoot Na+:K+ (~6%; SNP-1.11472400). Several studies have identified alleles within SKC1/HKT1;5 that are associated with Na+ exclusion and salt tolerance, but it is unclear whether the effects of these alleles are as strong as those reported by Gregorio et al. and Bonilla et al. [18,20,21,23,25,60]. Given the small effect of this QTL in the current study, as well as the large number of QTL identified for shoot Na+ and Na+:K+, it is likely that natural variation for shoot Na+ and Na+:K+ involves additional genetic components in addition to SKC1/HKT1;5.
This study included 383 of the 421 original RDP1 accessions, as well as seven check varieties [28–30]. Accessions were obtained from the USDA-ARS Dale Bumpers Rice Research Center and purified through single seed descent before they were phenotyped. Thirty-eight accessions of RDP1 were not included because of lack of seed availability and/or poor seed quality. The set of accessions from RDP1 included 77 indica, 52 aus, 92 temperate japonica, 85 tropical japonica, 12 groupV/aromatic, and 56 highly admixed accessions (nine accessions were unassigned), according to the classification by Famoso et al [29]. A total of 365 accessions from RDP1 were genotyped using 700,000 SNPs [32]. Filtering SNPs based on minor-allele frequency (MAF > 0.05) left ~397,812 high quality SNPs (depending on the trait analyzed) [32]. Previous results indicated LD decays to 0.20 between 0.5–1.0 Mb, indicating the marker density provided by the SNP array has suitable power to detect linked causal variants of moderate to large effect QTL [32].
The experiment was conducted between July to Sep 2013 in a controlled green house at Lincoln, NE. Rice (Oryza sativa) seeds were dehusked manually and germinated in the dark for two days at 28°C in a growth cabinet (Percival Scientific). Twelve hours before transplanting seeds were exposed to light (120 μmol m−2 s−1). The green house conditions were as follows: photoperiod (16:8 day:night), temperature 25–28°C and humidity 50–80%. Seedlings were transplanted into the pots filled with Turface (Profile Products, LLC) and were grown in tap water for four days after transplanting. For the remainder of the experiment the plants were supplemented with half strength Yoshida solution (pH 5.8) [61]. Salt treatment was applied as described previously by Walia et al. with minor modifications [62]. Briefly, NaCl was mixed with CaCl2 in a 6:1 molar ratio and was added after 10 d of seedling growth. The stress treatment was started at 2.5 dS·m-1 which increased gradually up to 9.5 dS·m-1 in 4 steps over a period of four days (~2 dS·m-1 or 20 mM NaCl per day) to avoid any osmotic shock to the plants. The stress treatment was stabilized at 9.5 dS·m-1 for next two weeks. The nutrient solution pH and electrical conductivity (EC) were monitored and maintained twice daily. The pH of the nutrient solution was maintained at 5.8 using H2SO4 and KOH. Root and shoot samples were collected separately and rinsed 3 times in tap water and once in deionized water to remove excess NaCl at the completion of the experiment (14 days of 9.5 dS·m-1; 28 days after transplant). The samples were oven dried at 60°C for one week prior to measuring root and shoot biomass. Shoot and roots from two plants were taken for biomass measurement.
For the large-scale screening of RDP1 dried shoot samples were ground and 200–300 mg of total material was digested with 0.1N Nitric acid (Fisher Scientific) at 70°C for 8 hrs. Root samples were weighed and digested without any grinding. Samples were diluted and cation (Na+ and K+) concentrations in the plant extract were determined with appropriate standard by dual Flame photometry (Cole Parmer, USA).
Data was combined across periods and a linear model was fit to calculate adjusted means for individual accession using the PROC GLM procedure of the Statistical Analysis System (SAS Institute, Inc.). The linear model included period (i.e., June-July or Aug-Sept), replication nested within period, tub nested within replication, accession, and accession-by-period interaction.
For the purpose of estimating variance components, a second similar linear model was fit using PROC MIXED in SAS. This time, all effects were assumed to be random effects. Broad-sense heritability (H2) on an entry-mean basis was calculated as H2=σG2/(σG2+σGP2/2+σe2/6) Where σG2 is the variance among accessions, σGP2 is the accession-by-period interaction variance, and σe2 is the error variance. In this context, the divisor 2 is equal to the number of periods and the divisor 6 is equal the number of replications per period (three) multiplied by the number of periods. Broad-sense heritability provides a sense of how much of the total variation observed is due to genetic variation among accession, and indicates the power of GWAS.
Marker-trait associations were tested using the linear mixed model y = Xβ + Cγ + Zu + e where y is a vector of phenotype; β is a vector of fixed marker effects; γ is a vector of principal component (PC) effects fit in order to account for population structure; u is a vector of polygenic effects caused by relatedness; e is a vector of residuals; X is a marker incidence matrix relating β to y; C is an incidence matrix relating γ to y which consists of the first four principal components (PCs) resulting from a PC analysis; Z is the corresponding design matrix relating y to u. It is assumed u∼MVN(0,Kσu2) and e∼MVN(0,Iσe2) where K is a standardized kinship matrix estimated using an allele-sharing matrix calculated from the SNP data. The above model was implemented using the efficient mixed-model association (EMMA) algorithm of Kang et al [33].
The method published by Li and Ji was used to determine a comparison-wise error rate to control the experiment-wise error rate [63]. Briefly, the correlation matrix and eigenvalue decomposition among 397,812 SNPs were calculated to determine effective number of independent tests (Meff). The test criteria was then adjusted using the Meff with the Sidak correction below
αp=1–(1-αe)1/Meff,
where αp is the comparison-wise error rate and αe is the experiment-wise error rate [64]. An αe = 0.05 was used in this study.
Analysis of variance (ANOVA) was used to estimate proportion of phenotypic variance accounted for by significant SNPs after adjusting for population structure effects. A 200 kb window was used to define groups of significant SNPs tagging the same locus. Only the most significant SNP within a 200 kb window was used to tag that locus. The percent variation explained by each significant SNPs was determined by comparing the linear models, y = Xβ + Cγ + e, and y = Cγ + e, where β is the SNP effect; γ is a vector of PCs effects to account for population structure; X is a vector of SNP genotypes; C is an incidence matrix relating γ to y which consists of the first four principal components (PCs). Therefore, the effect of each SNP is reported after accounting for the effects of population structure.
Genetic correlations between traits were estimated with and without correcting for population structure and family relatedness. The rationale behind correcting genetic correlations for population structure is to measure the correlation independent of long-range LD between loci caused by population structure [65]. To accomplish this, a multivariate mixed model was fit as described by Wisser et al. including all traits as response variables; fixed experimental design effects (replication and tub nested within replication); fixed population structure effects modeled using the four PCs as above; random polygenic effects modeled using the kinship matrix as in the GWAS model described above; and random residuals assumed to independent and identically distributed [65]. Restricted maximum likelihood implemented in ASReml-R v.3.0 was used to estimate genetic and residual variances, and genetic and residual covariances among traits [66]. Estimates of genetic variances and covariances were used to calculate genetic correlations among traits. For estimation of genetic correlations uncorrected for population structure, the same methods were used except population structure and polygenic effects were not included in the mixed linear model.
Haplotype blocks were constructed using the four gamete method (4gamete) implemented in the software Haploview [67]. The method creates block boundaries where there is evidence of recombination between adjacent SNPs based on the presence of all four gametic types. We used a cut-off of 2%, meaning that if addition of a SNP to a block resulted in recombinant alleles at a frequency exceeding 2%, the SNP was not included in the block.
To examine the frequency of high and low root Na+ forms of HKT1;1 in a set of 3,023 cultivated rice and 446 Oryza rufipogon accessions, a set of 206 SNPs was extracted from a ~37 kb (30,700,524–30,737,580) region on chromosome 4. Sequence data for the cultivated rice was obtained from ~9 million genome-wide SNPs generated by the 3000 Rice Genomes Project (3K RGP) [37]. The 206 SNPs for Oryza rufipogon was obtained from riceHap3 (www.ncgr.ac.cn/ricehap3/) [38]. Since SNPs were mapped to different genome builds (IRGSP4.0 to IRGSP1.0 for 3kg and RiceHap3, respectively), the coordinates were converted by aligning a 37 kb region from IRGSP4.0 to IRGSP1.0 using BLAT [68].
Haplotype block analysis was performed using the 4Gamete rule, with a cutoff of 1% in Haploview [67]. The frequency of each haplotype within in each subpopulation was determined in R. A haplotype network for this 37 kb region was built with PopArt [69]. Nucleotide diversity (π) was determined at each position for indica (ind1A, ind1B, ind2, ind3 and indx), japonica (temp, trop 1, trop 2, and japx) and wild rice using the "site-pi" function in VCFtools [70].
For gene expression analysis, plants were grown in a controlled environment growth chamber. Temperatures were maintained at 28°C and 25°C in day and night respectively, relative humidity was maintained at 60% in both day and night. Lighting was maintained at 800 μmoles·m−2·s−1 using high pressure sodium lights (Phillips). Seeds preparation and salt treatment were performed as described above. Eight day (four days after transplant) old rice seedlings were subjected to 6 dS·m-1 for a period of 24h. Salinity stress was increased to 6 dS·m-1 gradually in two 3 dS·m-1 intervals over a period of 24h. After 24h of 6 dS·m-1, aerial parts of the seedlings were excised from the roots and frozen immediately in liquid nitrogen. The samples were ground with Tissuelyser II (Invitrogen) and total RNA was isolated with RNAeasy isolation kit (Qiagen) according to manufacturer’s instructions. On-column DNAse treatment was performed to remove genomic DNA contamination (Qiagen). Sequencing was performed using Illumina HiSeq 2500. Sixteen cDNA libraries were combined in each lane.
After being examined using the package FastQC, short reads, obtained from Illumina 101-bp single-end RNA sequencing, were screened and trimmed using Trimmomatic to ensure each read has average quality score larger than 30 and longer than 15 bp [71,72]. The trimmed short reads were mapped against to the rice genome (Oryza sativa MSU Release 6.0) using TopHat (v.2.0.10), allowing up to two base mismatches per read. Reads mapped to multiple locations were discarded [73]. Numbers of reads in genes were counted by the HTSeq-count tool using gene annotations for the same version of rice genome and the “union” resolution mode was used [74].
For a given genotype, all mapped RNA-seq short reads were sorted and indexed by Samtools (Version: 0.1.18) [75]. Single nucleotide polymorphisms (SNPs) and small insertions/deletions (Indels) were identified based on differences between short reads from the given genotype and the reference genome sequence with mapping quality larger than 25, read depth more than 30, but less than 500. Variations in regions of interest in the rice genome were selected with their coordinates and gene annotations.
First strand cDNA synthesis for real-time quantitative PCR (qRT-PCR) was performed using iScript Reverse Transcription Supermix (Bio-Rad Laboratories, Inc., Hercules, CA, USA) using 2 μg of total RNA. For the qPCR reaction, 3 μL of the diluted cDNA (1:20) was used in the 15 μL reaction mixture. In the qPCR reaction volume, 7.5 μL of LightCycler 480 SYBR Green I Mastermix was used (Roche Diagnostics, Indianapolis, IN, USA). The qRT-PCR was carried out using Roche LightCycle 480 II with the following parameter settings (Roche Diagnostics, Indianapolis, IN, USA): 95 oC pre-incubation for 5 min, amplification was done for 40 cycles at 95 oC for 20 sec and 60 oC for 15 sec and extension at 72 oC for 15 sec; the melting curve was set-up for 95 oC, 65 oC, 97 oC; cooling was set-up at 40 oC for 30 sec. We used two independent tissue samples, with tissue from two to three plants pooled for each sample. LOC_Os04g02820 was used as an internal reference gene, which displayed stable expression in all samples analyzed. Relative expression was determined using the delta-delta Ct method [76]. Primer sequences are provided as S9 Table.
For HKT1;1, a 112 bp region was amplified from genomic DNA of the japonica rice variety ‘Kitaake’, while a 95 bp region was amplified for HKT1;4. The fragment from HKT1;4 was ligated into the pENTR-D-TOPO vector, while for HKT1;1 the fragment was inserted into pDONR221 using the BP reaction following the manufacturer’s instructions (Invitrogen). Finally, each fragment was introduced into the pANDA RNAi expression vector [77,78]. Transformation of ‘Kitaake’ calli was performed according to the methods outlined by Cheng et al. using the EHA-105 strain of Agrobactrium [79]. Calli and plants were selected on ½ strength MS media supplemented with 50 μg/ml hygromycin. The expression of HKT1;1 and HKT1;4 in shoot and flag leaf tissue of T1 plants, respectively, was determined using realtime PCR using the same conditions as described above. Primer sequences are provided in S9 Table.
To generate native overexpression lines for the two isoforms of HKT1;1, a ~4.3 kb fragment was amplified from ‘Nipponbare’ and ‘Zhenshan 2’. The fragments were cloned into pDONR221 vector via the BP reaction, and were subsequently cloned into a pMDC99 backbone with a NOS terminator [80]. Agrobacterium-mediated transformation and selection of transformants was performed as described above. T1 plants with a single insertion were selected on ½ strength MS media supplemented with 50 μg/ml hygromycin and used for phenotyping.
Phenotyping of transgenic plants was performed in a controlled environment growth chamber. Three independent RNAi lines (T2) for HKT1;1 and HKT1;4 was screened for salinity tolerance, while two independent native overexpression lines (T1 generation) were evaluated for each isoform of HKT1;1 (HKT1;1Ni and HKT1;1Zh). Temperatures were maintained at 28°C and 25°C in day and night respectively, relative humidity was maintained at 60% in day and night. Lighting was maintained at 800 μM using high pressure sodium lights (Phillips). Seeds were surface sterilized in a 40% bleach solution for 20 min, rinsed in sterile water and were germinated on ½ MS media supplemented with 50 ug/ml of hygromycin. WT seeds received the same treatment, but were grown on ½ MS. The seeds were germinated for 24h in complete darkness then were transferred to a growth cabinet (Percival Scientific) and grown for four days at 28°C in 16/8h light (120 μmol m−2 s−1). Seedlings were transplanted into the pots filled with Turface (Profile Products, LLC) and were grown in tap water for four days after transplanting. For the remainder of the experiment the plants were supplemented with half strength Yoshida solution (pH 5.8) [61]. Eight days after transplanting a gradual salt stress was applied in three 3 dS·m-1 intervals over a period of 24h. The final 9 dS·m-1 salt level was maintained for two weeks. Sample collection was performed as described above.
To generate constructs for assessing Na+ transport activities in Xenopus laevis oocytes, HKT1;1 was amplified from cDNA from ‘Nipponbare’ and ‘Zhenshan 2’, which are representative accessions for the high and low root Na+ groups at RNC4, respectively, and ligated into the pGEM-Xho vector [81]. The pGEM-Xho contains the T7 promoter and 5′- and 3′-untranslated regions of the Xenopus β-globin gene, which enhances expression in Xenopus oocytes. For N-terminal GFP::HKT1;1 fusion constructs, HKT1;1 was amplified from cDNA from ‘Nipponbare’ and ‘Zhenshan 2’ and cloned into pGWB6 using the Gateway LR reaction. GFP::HKT1;1 was then amplified from each construct using primers with SpeI and SalI restriction sites, and introduced into the pGEM-Xho vector [82].
Capped and polyadenylated RNA were obtained from linearized vector by in vitro transcription, using the mMESSAGE mMACHINE T7 kit (Ambion, USA). Oocytes isolated as previously described were injected with 50 ng of HKT1;1-Ni or HKT1;1-Zh cRNA (equivalent amount of transporter cRNA in GFP-tagged form) in 50 nL of RNase-free water, or with 50 nL of RNase-free water (for control oocytes), and then kept for 24 to 48 h at 19°C in ND96 medium (96 mM NaCl, 2 mM KCl, 1.8 mM CaCl2, 1 mM MgCl2, 2.5 mM sodium pyruvate, and 5 mM HEPES/NaOH, pH 7.4) supplemented with 0.5 mg·L–1 of gentamicin, until experiments [81]. Whole oocyte currents and membrane potential were recorded using the two-electrode voltage-clamp technique with a GeneClamp 500B amplifier (Axon Instruments, USA) 1 to 2 days after cRNA injection. Voltage-pulse protocols, data acquisition and analysis were performed using pClamp9 software (Axon Instruments). Correction was made for voltage drop through the series resistance of the bath and the reference electrode using two external electrodes connected to a bath probe (VG-2A x100 Virtual-ground bath clamp; Axon Instruments). Electrodes were filled with 3 M KCl. The oocytes were continuously perfused during the voltage-clamp experiment with bath solutions containing varying concentrations of monovalent cations (as glutamate or chloride salts) in a background of 6 mM MgCl2, 1.8 mM CaCl2, and 10 mM MES-1,3-bis[tris(hydroxymethyl) methylamino]propane, pH 5.5. The chloride concentration was constant in each set of solutions. D-Mannitol was added when necessary to adjust the osmolarity, which was set to 220–240 mosM in each set of solutions. Voltage-clamp protocol consisted in successive steps of membrane voltage application from -165 to +15 mV in +15 mV increments during 0.5 s, each step beginning with 0.15 s and ending with 0.3 s at the resting potential of the oocyte membrane in the tested bath solution. Mean currents recorded in water-injected control oocytes from the same batch and in the same ionic conditions as HKT-expressing ones were subtracted from those recorded in HKT-expressing oocytes in order to extract HKT-mediated currents from total oocyte currents. HKT1;1-Ni and -Zh current–voltage (I–V) relationships were constructed with transporter extracted currents. The activation potential of HKT currents was estimated as the lowest voltage at which the current in HKT-expressing oocytes reached twice that in control oocytes.
Confocal observations were made on dark poles of oocytes of similar sizes on a Leica SP8 microscope, using a 20x/0.7dry objective. GFP was excited with a 488 nm laser, and spectral acquisitions of emitted fluorescent light were performed between 495 and 645 nm using a bandwidth of 5 nm, to assert GFP specificity. For each oocyte, mean fluorescence intensity at the membrane was determined from at least 2 optical sections, analyzing 3 ROIs per section using ImageJ (https://imagej.nih.gov/ij/) software.
|
10.1371/journal.pbio.1000509 | Genome-Wide Analyses Reveal a Role for Peptide Hormones in Planarian Germline Development | Bioactive peptides (i.e., neuropeptides or peptide hormones) represent the largest class of cell-cell signaling molecules in metazoans and are potent regulators of neural and physiological function. In vertebrates, peptide hormones play an integral role in endocrine signaling between the brain and the gonads that controls reproductive development, yet few of these molecules have been shown to influence reproductive development in invertebrates. Here, we define a role for peptide hormones in controlling reproductive physiology of the model flatworm, the planarian Schmidtea mediterranea. Based on our observation that defective neuropeptide processing results in defects in reproductive system development, we employed peptidomic and functional genomic approaches to characterize the planarian peptide hormone complement, identifying 51 prohormone genes and validating 142 peptides biochemically. Comprehensive in situ hybridization analyses of prohormone gene expression revealed the unanticipated complexity of the flatworm nervous system and identified a prohormone specifically expressed in the nervous system of sexually reproducing planarians. We show that this member of the neuropeptide Y superfamily is required for the maintenance of mature reproductive organs and differentiated germ cells in the testes. Additionally, comparative analyses of our biochemically validated prohormones with the genomes of the parasitic flatworms Schistosoma mansoni and Schistosoma japonicum identified new schistosome prohormones and validated half of all predicted peptide-encoding genes in these parasites. These studies describe the peptide hormone complement of a flatworm on a genome-wide scale and reveal a previously uncharacterized role for peptide hormones in flatworm reproduction. Furthermore, they suggest new opportunities for using planarians as free-living models for understanding the reproductive biology of flatworm parasites.
| Flatworms cause diseases affecting hundreds of millions of people, so understanding what influences their reproductive activity is of fundamental importance. Neurally derived signals have been suggested to coordinate sexual reproduction in free-living flatworms, yet the neuroendocrine signaling repertoire has not been characterized comprehensively for any flatworm. Neuropeptides are a large diverse group of cell-cell signaling molecules and play many roles in vertebrate reproductive development; however, little is known about their function in reproductive development among invertebrates. Here we use biochemical and bioinformatic techniques to identify bioactive peptides in the genome of the planarian flatworm Schmidtea mediterranea and identify 51 genes encoding >200 peptides. Analysis of these genes in both sexual and asexual strains of S. mediterranea identified a neuropeptide Y superfamily member as important for the normal development and maintenance of the planarian reproductive system. We suggest that understanding peptide hormone function in planarian reproduction could have practical implications in the treatment of parasitic flatworms.
| Platyhelminthes (flatworms) inhabit a variety of aquatic and terrestrial environments and members of the phylum are thought to parasitize most vertebrate species [1]. The remarkable ability of flatworms to maintain plasticity in their reproductive cycles is a key to their success. As an example, free-living planarian flatworms are capable of reproducing sexually as cross-fertilizing hermaphrodites or asexually by transverse fission [2]. Some planarian species even maintain the ability to switch between modes of sexual and asexual reproduction, resorbing and regenerating their reproductive organs, depending on the environmental context [3]. This dynamic regulation of reproductive development is not limited to free-living platyhelminths; parasitic flatworms can also undergo dramatic changes in their reproductive development in response to external stimuli. In dioecious parasites of the genus Schistosoma, female reproductive development requires pairing with a male worm [4]–[8]. Thus, female schistosomes derived from single-sex infections have underdeveloped ovaries and accessory reproductive organs when compared to females from mixed sex infections. Interestingly, the reproductive organs of mature females deprived of their male counterpart regress and are capable of regrowing once male-female pairing is reestablished [9]. Because flatworms, including schistosomes, are responsible for causing important neglected tropical diseases, understanding the mechanisms that coordinate the reproduction of both free-living and parasitic members of the phylum is of fundamental importance.
Peptide hormones (i.e. neuropeptides) are among the most structurally and functionally diverse class of metazoan signaling molecules [10]. In vertebrates, a neuroendocrine axis involving peptide hormone signaling between the brain and the gonads controls the maturation and long-term maintenance of reproductive development and function [10]–[13]. A similar role for neuroendocrine signals in controlling flatworm reproduction is suggested by studies exploiting the well-known regeneration abilities of planarians. Head amputation (i.e. removal of the brain/cephalic ganglia) of sexually reproducing planarians results in regression of the testes [14],[15] to their germ cell primordia [16], which re-grow only when cephalic ganglia regeneration is complete. These observations suggest that neural signals control the dynamics of planarian reproduction. Thus, flatworms may employ peptide-based mechanisms, similar to vertebrates, to synchronize their reproductive development.
To date only limited data exist to support a “vertebrate-like” role for peptide hormones in invertebrate reproductive maturation. Insulin-like peptides influence germline stem cell proliferation in Drosophila [17],[18] and C. elegans [19] and promote oocyte maturation in the starfish Asterina pectinifera [20] and the mosquito Aedes aegypti [21]. In locusts, treatment with the peptide hormones ovary maturing parsin (OVP) [22] or short Neuropeptide F (sNPF) [23],[24] can stimulate ovarian development and vitellogenesis. Because of this paucity of data linking neuroendocrine function to invertebrate reproductive development, additional studies are required to determine how invertebrates modulate their reproductive output in response to external and metabolic cues.
Peptide hormones are processed proteolytically from longer secretory prohormone precursors and often require covalent modifications before becoming biologically active [10],[25]. As a result of this extensive processing, and because the biologically relevant signaling units are encoded by short stretches of amino acid sequence (usually 3–40 amino acids), predicting genes encoding these molecules represents a major challenge for bioinformatics-driven genome annotation. The recent application of bioinformatic approaches coupled to mass spectrometry-based peptide characterization techniques (an approach called peptidomics [26]–[28]) has revolutionized discovery efforts, uncovering hundreds of new genes encoding metazoan bioactive peptides. Among invertebrates, however, much of this recent progress has been focused on genome-wide studies of nematodes [29]–[31], arthropods [32]–[36], and mollusks [37],[38]. Thus, little is known of the peptide hormones present in phyla such as Platyhelminthes. Despite recent bioinformatic efforts to characterize flatworm peptide-encoding genes [39],[40], only three distinct peptides have been characterized extensively at the biochemical level in all flatworms [41].
Owing to a wealth of functional genomic tools [42] and a sequenced genome [43], the planarian S. mediterranea represents an ideal model to characterize flatworm neuropeptides. Furthermore, this species exists as two distinct strains: an asexual strain that lacks reproductive organs and propagates exclusively by fission and a sexual strain that reproduces as cross-fertilizing hermaphrodites [44]. This dichotomy presents a unique opportunity to explore the extent to which peptide hormones are associated with distinct reproductive states. To address the possibility that peptide signals influence planarian reproductive development, we began by disrupting a gene encoding a prohormone processing enzyme, Smed-prohormone convertase 2 (Smed-pc2, GB: BK007043), in sexual planarians. Consistent with a role for peptide hormones in controlling planarian reproduction, knockdown of Smed-pc2 led to a depletion of differentiated germ cells in the planarian testes. To identify potential peptide mediators of this effect, we used peptidomic approaches to characterize the peptide hormone complement of S. mediterranea. This analysis identified 51 genes predicted to encode more than 200 peptides, 142 of which we characterized biochemically by mass spectrometry. Global analysis of the expression of these genes by whole mount in situ hybridization revealed a distinct distribution of some peptide prohormones between sexual and asexual strains of S. mediterranea. We find one prohormone gene, npy-8, to be enriched in the nervous system of sexual planarians and show that this gene is required for the proper development and maintenance of reproductive tissues. These results demonstrate the utility of S. mediterranea as a model to characterize metazoan peptides and suggest that flatworm reproductive development is controlled by neuroendocrine signals.
To explore potential roles for peptide signaling in regulating planarian reproductive physiology, we characterized Smed-pc2 (Figure S1), whose orthologues are required in both vertebrate and invertebrate models for the proteolytic processing of prohormones to mature neuropeptides (in the interest of brevity, we will drop the prefix “Smed” from the remainder of the genes described below) [30],[45],[46]. A large-scale RNA interference (RNAi) screen determined that this gene was essential for coordinated movement and normal regeneration in asexual planarians [47]. Whole-mount in situ hybridization in sexual planarians revealed expression of pc2 in the central nervous system [48], the pharynx, sub-muscular cells, the photoreceptors, the copulatory apparatus, and the testes (Figure 1A–C).
To determine if peptide signals are likely to play a functional role in coordinating reproductive development, we monitored the effects of pc2 RNAi on the dynamics of germ cells within the planarian testes. Individual testis lobes consist of an outer spermatogonial layer in which cells divide to form cysts of eight spermatocytes that, after meiosis, give rise to spermatids and, ultimately, sperm [44],[49]. After 17 d of RNAi treatment, pc2(RNAi) animals displayed a decrease in both testis size (Figure 1E) and the number of animals producing mature sperm (28/29 for controls versus 2/36 for pc2 RNAi; p<0.0001, Student's t test). To establish which cell types are affected by pc2 RNAi, we performed fluorescence in situ hybridization (FISH) to detect germinal histone H4 (gH4) (GB: DN306099) and nanos (GB: EF035555) mRNAs, which are expressed in spermatogonia and germline stem cells (GSCs), respectively [16],[50],[51]. In developed testes of control animals, relatively few cells within the outer spermatogonial layer are identifiable as nanos-positive GSCs (Figure 1F). However in pc2(RNAi) animals, regressed testes clusters almost always co-expressed both gH4 and nanos (Figure 1G) (n = 16/17 animals). These results suggest that pc2 is required for proper germ cell differentiation and/or for the maintenance of differentiated germ cells in the testes.
Since our analysis of pc2 implicated peptide signaling in regulating planarian reproductive development, we characterized the peptide hormone complement of S. mediterranea. We employed bioinformatic and mass spectrometry (MS)-based methodologies to identify peptide prohormone genes from the S. mediterranea genome [43] and predict their processing into bioactive peptides (Figure 2A) [52]. With these approaches, we identified 51 prohormone genes in S. mediterranea, with peptides from 40 prohormones detected by MS (Tables S1–S5, gene names and abbreviations are shown in Table 1). In most cases, MS confirmed multiple distinct peptides from a single prohormone, and in five prohormones we detected every predicted peptide by MS (Figure 2B). In total, we characterized 142 peptides biochemically, corresponding to ∼45% of the distinct peptides predicted from our collection of 51 prohormone genes (Table S5). This analysis identified genes encoding relatives of all previously characterized flatworm neuropeptides (YIRFamide [53], spp-11; FRFamide [54], npp-4; and neuropeptide Y-like [55], npy-1 to npy-11) and provided biochemical validation for 10 prohormones previously predicted from the S. mediterranea genome [39].
The neuropeptide Y (NPY)-superfamily represents a large family of neuropeptides that influence diverse processes in both vertebrate and invertebrate taxa [10],[41],[56]. This family is considered to consist of two types of peptides: the NPY-like peptides that possess a C-terminal amidated tyrosine (Y) residue and the NPF peptides that possess a C-terminal amidated phenylalanine (F) residue [55]. Vertebrate genomes typically encode NPY-like peptides [57], whereas invertebrate genomes encode NPF peptides [55],[58],[59]. Our studies found that the planarian genome possesses an expanded family of npy genes predicted to encode both NPY-like and NPF-like peptides (Figure 2C). Prohormones NPY-5, -7, -9, and -10 possess a C-terminal tyrosine residue, similar to vertebrate NPY peptides, and prohormones SMED-NPY-1, -2, -3, -4, -6, and -8 contain a C-terminal phenylalanine residue, similar to invertebrate NPF peptides. Three of these planarian npy genes (npy-1, -4, and -9) have been described previously [39],[60]. Additionally, our studies, and those of others [39],[61], find evidence of conservation in the genomic organization of flatworm NPY genes. NPY genes from vertebrates possess an intron that separates the exon encoding the RXR motif from the penultimate amidated amino acid residue (Figure 2D) [62]. We found an identical architecture for S. mediterranea genes npy-1, -2, -3, -4, -5, -6, -8, -9, -10, and -11, indicating a close evolutionary relationship between chordate and platyhelminth npy genes (Figure 2C,D).
The planarian genome also encodes peptides with sequence similarities to those from other invertebrate taxa, including mollusks (ppp-1, GB:BK007041; ppp-2, GB:BK007018; mpl-1, GB: BK007017; mpl-2, GB: BK007016; and, cpp-1, GB: BK007012) and arthropods (ppl-1, GB: BK007007). Furthermore, our analysis found that previously characterized, novel planarian genes encode peptide prohormones. Homologues of prohormones eye53-1,2 (GB: BK007033 and GB: BK007024, respectively) and 1020-1,2 (GB: GU295180 and GB:BK007025, respectively) from the planarian Dugesia japonica are required for proper visual system function following amputation; knockdown animals show no morphological defects after injury yet are unable to respond properly to light [63]. These previous observations, together with our findings that these genes encode neuropeptides, suggest a role for peptide signaling in the functional recovery of the planarian nervous system following injury.
To examine if pc2 is required for prohormone processing in planarians, we disrupted pc2 expression using RNAi and performed MS to analyze the peptide complement of pc2(RNAi) animals. Consistent with pc2 encoding a genuine prohormone convertase, analysis of peptide profiles in planarian tissue extracts by MALDI-TOF MS analysis demonstrated that pc2 RNAi resulted in a significant decrease in the signal intensity of a specific set of peptides in sexual animals (Figure 2E,F and Table S6). Interestingly, the levels of some peptides were increased following pc2(RNAi); whether this alteration reflects a compensatory mechanism for regulating peptide levels or an altered threshold of detection for certain peptides caused by a global reduction in neuropeptide levels remains to be determined. However, these data parallel studies of pc2 knockout mice, in which the abundance of some peptides was either increased or decreased [45]. Given that the S. mediterranea genome is predicted to encode at least three additional proteins with similarity to prohormone convertases (Figure S2), it is possible that compensatory mechanisms are responsible for the observed elevation in the levels of some peptides. This redundancy among prohormone convertases is also likely to explain why we only observed changes in a subset of peptides following pc2 RNAi. These data suggest that the reproductive defects observed in pc2(RNAi) animals may be due to altered levels of specific peptides.
To determine the extent to which peptides may regulate flatworm reproduction, we took advantage of the fact that S. mediterranea exists as both sexually and asexually reproducing strains. By comparing prohormone gene expression between these strains we sought to uncover expression patterns specific to sexually or asexually reproducing animals. Thus, we began by performing comprehensive whole-mount in situ hybridization analyses of prohormone genes in asexual planarians (Figure 3).
Our studies indicate that in asexual planarians ∼85% (44/51) of prohormone genes are expressed in the central nervous system (CNS) (Table S5), which consists of bi-lobed cephalic ganglia and two ventral nerve cords (VNCs) that run the length of the body [64]. Of the prohormones expressed in the CNS, 20% (10/51) were detected only in the cephalic ganglia. Notably, the expression of individual prohormones was often enriched in specific cell types or regions within the CNS. For example, the expression of some prohormones was enriched in either lateral (e.g. npp-4, GB: BK007037; npp-8, GB: GU295189; spp-4, GB: GU295179; and 1020HH-2), medial (e.g. spp-2, GB: BK007032; and spp-6, GB: GU295177) or posterior (e.g. npy-1, GB: GU295175) regions of the cephalic ganglia (Figure 3). Strikingly, a large fraction of prohormone mRNAs were detected in restricted cell populations within the CNS (e.g. npy-1; npy-2, GB: BK007019; cpp-1; spp-6; spp-9, GB: BK007026; spp-10, GB: BK007028; grh-1, GB: GU295185; and ilp-1, GB: BK007034) (Figure 3).
Consistent with peptide signaling having a role in processes outside the CNS, we also detected prohormone expression in: the pharynx (e.g. npp-1, GB: BK007036; npp-22, GB: BK007038; npy-11, BK007021; and ppp-1); photoreceptors (e.g. eye53-1,-2; npp-12, GB: GU295182; and mpl-2); sub-epidermal marginal adhesive glands (e.g. mpl-2); an anterior domain between the VNCs (e.g. spp-6; spp-7, GB: GU295178; spp-8, GB: GU295181; spp-9; cpp-1; and spp-10, GB: BK007028); cells surrounding the ventral midline (e.g. npp-5, BK007015); the intestine (e.g. npp-8, GB: GU295189; and npy-10, GB: BK007011); and various sub-epidermal cell types (e.g. npp-18, GB: BK007027; spp-4; spp-16, GB: BK007042; and npy-4, BK007039) (Figure 3).
To investigate the extent to which prohormones are expressed in overlapping or distinct cell types in the CNS, we compared the expression of prohormone genes using triple FISH. Prohormone genes spp-1 (GB: GU295176), npp-2 (GB: BK007035), and ppp-1 encode unrelated peptides (Tables S1 and S5) that appear to be expressed ubiquitously in the CNS (Figure 3). Comparison of the expression domains of these prohormone genes revealed that spp-1, npp-2, and ppp-1 are expressed in largely non-overlapping populations of cells of the cephalic ganglia and VNCs (Figure 4A–C). We also analyzed the expression of a family of paralogous prohormone genes (spp-6; spp-7; spp-8; spp-9; and spp-17, GB: GU295183) that encode similar neuropeptides (Figure S3). Because this gene family has been expanded in the S. mediterranea genome, we refer to these prohormones as the Planarins. Examination of Planarins spp-6, -7, and -9 expression by FISH demonstrated that these genes are expressed in a common set of cells distributed between the VNCs and surrounding the pharynx (Figure 4D). Despite being co-expressed in cells outside the CNS, spp-6 and spp-9 transcripts were detected in distinct groups of cells within the cephalic ganglia (Figure 4E,F). These findings, with earlier observations [48],[64], suggest a level of complexity not previously appreciated for the patterning of the flatworm nervous system (see Figure S4).
We also examined four prohormone genes (eye53-1,-2; npp-12, and mpl-2) expressed within the photoreceptors. The planarian photoreceptors are comprised of two distinct cell types: neuronal photoreceptive cells and pigment cells that envelop the rhabdomeric projections of the photoreceptor neurons [2],[65],[66]. Analysis of prohormone gene expression within the photoreceptors revealed that the planarian photoreceptor neurons are patterned along the anterior-posterior axis. Specifically, prohormone genes npp-12 and eye53-1 were expressed exclusively in the anterior photoreceptor neurons, whereas mpl-2 and eye53-2 were expressed exclusively in posterior neurons (Figure 5A). These findings are consistent with dye-tracing studies demonstrating that anterior and posterior photoreceptor neurons project to distinct anatomical regions [67]. In addition, we detected mpl-2 expression in a ventral population of cells that was separate from the expression of eye53-2 (Figure 5B); this result suggests that the photoreceptors are also patterned along the dorsal-ventral axis. Together, these data indicate that at least three chemically and anatomically distinct sets of neurons are present in the planarian photoreceptors.
To determine if peptide expression is correlated with reproductive state, we next analyzed the expression of a subset of prohormones in the sexual strain of S. mediterranea. The reproductive system of this animal is comprised of a pair of ovaries located posterior to the cephalic ganglia, numerous dorsolateral testes lobes, as well as a variety of accessory reproductive organs (i.e. oviducts, sperm ducts, copulatory apparatus, and accessory glands) (Figure 6A). We found several prohormones expressed in sexual reproductive organs, including the oviducts (Figure 6B,C), the copulatory apparatus (Figure 6B,C, and D), gland cells surrounding the copulatory apparatus (Figure 6E,F), and the testes (Figure 6G,H). These expression patterns implicate peptide signaling in reproductive processes such as copulation, fertilization, egg-laying, and gonadal function.
Our expression analyses also found evidence of differential prohormone expression within the nervous system of sexual S. mediterranea. ppl-1 encodes peptides related to the pyrokinin peptides originally isolated from arthropods [68],[69]. In contrast to asexual planarians in which ppl-1 expression was detected almost exclusively in the cephalic ganglia and the distal region of the pharynx (Figure 3), ppl-1 was expressed widely in the VNCs and surrounding the copulatory apparatus of mature sexual animals (Figure 7A). To explore if ppl-1 expression was linked to sexual maturation, we determined the distribution of ppl-1 in immature sexual animals. In sexual animals analyzed within one week of hatching from the egg capsule, ppl-1 was expressed in a pattern similar to that of asexual animals (Figure 7A); thus, ppl-1 expression undergoes a change in spatial distribution during the process of maturation.
The prohormone gene npy-8 (GB: BK007010) is predicted to encode a 29 AA NPF-like peptide (NPY-8A) and a novel C-terminal peptide (NPY-8B) (Figure 8A). By in situ hybridization we failed to detect npy-8 expression in asexual animals (Figures 3 and 7B). In mature sexual animals, however, npy-8 RNA was detected in a variety of cells within the central and peripheral nervous systems including the cephalic ganglia, the VNCs, the sub-muscular plexus, and the pharyngeal nervous system (Figure 7B). Additionally, in a majority of animals (13/18) we detected npy-8 RNA in a dorsal population of cells (Figure 6C). Analysis of this dorsal cell population by FISH localized npy-8 expression to cells often, but not exclusively, found in association with testes lobes (Figure 7D). To determine if npy-8 levels changed with sexual maturation we examined npy-8 expression in sexual hatchlings. In recently hatched animals npy-8 was detected in tissues similar to those of mature sexual animals including the cephalic ganglia, the VNCs, the sub-muscular plexus, and the pharyngeal nervous system (Figure 7B). Furthermore, we observed dorsal cells expressing npy-8 in a majority of animals (8/13) (Figure 7C).
The lack of observable expression of npy-8 in asexual animals by in situ hybridization suggested a relationship between npy-8 expression and the ability to reproduce sexually. Because we initially cloned the npy-8 gene by 3′ RACE with cDNA derived from asexual animals (Table S5), we wished to confirm our in situ hybridization results using an alternative approach. Therefore, we performed northern blot analyses to detect npy-8 transcript in asexual, recently hatched sexual, juvenile sexual, and mature sexual animals (Figure 7E). Consistent with our in situ hybridization results, we detected high levels of npy-8 in sexual animals of all developmental stages but not in asexual animals, suggesting that npy-8 is expressed at negligible levels in asexual planarians.
Because npy-8 was expressed at high levels only in sexually reproducing planarians, we reasoned that peptides encoded from this gene may be important for reproduction. Therefore, we determined the knockdown phenotype of npy-8 using RNAi. For this analysis we employed two distinct RNAi feeding regimens. First, we measured the effect of npy-8 depletion on the maintenance of the reproductive system by feeding mature sexual animals bacterially expressed npy-8 dsRNA and observing the structure of the reproductive system at 4- and 7-wk time points. As a complementary approach, we fed juvenile sexual planarians in vitro synthesized dsRNA and observed the development of the reproductive system after 1 mo of feeding. Mature sexual animals fed npy-8 dsRNA over the course of 4–7 wk displayed a range of phenotypes consistent with loss of sexual maturity (data are summarized in Table 2). Specifically, in comparison to controls, a majority of npy-8(RNAi) animals had regressed testes and failed to produce mature sperm (1/18 for controls versus 14/21 for npy-8 RNAi) (Figure 8B). In addition to testes defects, npy-8(RNAi) treatment resulted in regression of the copulatory organs (0/18 for controls versus 13/20 for npy-8 RNAi) (Figure 8B,C) and a decrease in the size (or complete disappearance) of the gonopore (unpublished data). Similar to mature sexual animals, juvenile planarians fed npy-8 dsRNA for 1 mo displayed stunted testes growth, failed to produce mature sperm (0/8 for controls and 6/8 for npy-8(RNAi)), and had shrunken or absent gonopores (0/20 for controls and 16/20 for npy-8(RNAi), Figure 8D). Importantly, these effects on reproductive maturation were not due to an overall defect in growth since npy-8(RNAi) and control animals grew to similar sizes over this time period (Figure S5A).
Since npy-8 is a member of an expanded family of NPY-like genes in S. mediterranea (Figure 2C), we examined both the effectiveness and the specificity of our npy-8 knockdowns. We fed juvenile planarians dsRNA specific to npy-8 and monitored the transcript levels of npy-8 and its closest relative, npy-1, by quantitative RT-PCR. This analysis found that npy-8 RNAi treatment resulted in a statistically significant decrease in npy-8 transcript levels while having no effect on npy-1 mRNA levels (Figure S5B). To further explore the specificity of the npy-8(RNAi) phenotype, we performed a long-term feeding experiment in which we fed juvenile animals dsRNA specific to npy-8 or either of its two closest relatives, npy-1 or npy-2. In contrast to npy-8 RNAi, neither npy-1 nor npy-2 RNAi treatments produced observable defects in the maturation of the planarian reproductive organs (Figure S5C). Collectively, these studies suggest that the effects of npy-8(RNAi) on reproductive development are due to specific disruption of npy-8 function and suggest that off-target effects are unlikely.
To examine the regressed testes of npy-8(RNAi) animals, we performed FISH to detect nanos and gH4 expression. This analysis uncovered a range of phenotypes associated with npy-8 RNAi (Figure 8E). Some npy-8(RNAi) animals had clusters of gH4-positive cells that were also nanos-positive; these testes clusters are similar to those observed in pc2(RNAi) animals (Figure 1G). In other animals we found gH4-positive clusters in which a subset of cells expressed nanos. We interpret the former to represent a “severe” npy-8 knockdown phenotype, whereas we suggest that the latter represents an “intermediate” phenotype resulting from incomplete npy-8 knockdown and/or perdurance of the peptide.
In the most severe cases, the testes regression phenotypes seen in pc2(RNAi) or npy-8(RNAi) animals were similar. One model to explain this observation is that PC2 is required for proteolytic processing of the NPY-8 prohormone, and loss of a mature peptide (or peptides) encoded by npy-8 results in loss of the ability to achieve or maintain sexual maturity. Since our MS analysis did not identify any peptides encoded by npy-8 in extracts from either asexual or sexual animals (Tables S1–S3), we used FISH to determine if npy-8 and pc2 transcripts are localized to similar cell types in the planarian nervous system. We found that npy-8-expressing cells within the cephalic ganglia, the VNCs, the pharynx, and the sub-muscular plexus also express high levels of pc2 (Figure 8F; and unpublished data). This observation is consistent with PC2 being required for the processing of peptides encoded by the npy-8 gene.
Related flatworms of the genus Schistosoma currently infect over 200 million people worldwide [70]. Because of their complicated life cycles, schistosomes are not readily amenable to the types of large-scale biochemical analyses that we have employed to characterize the planarian peptidome. As an indirect means of biochemically validating peptide sequences from these animals, we compared our MS-validated prohormones with predicted proteins from the genomes of the trematodes Schistosoma mansoni [40] and Schistosoma japonicum [71]. With this approach we validated the sequences of peptides from eight previously characterized schistosome prohormone genes (Tables 3 and S7) [39],[40]. Furthermore, we identified eight additional Schistosoma genes not previously annotated as peptide prohormones (Tables 3 and S7). Among these newly annotated prohormones are schistosome genes that encode the peptide YIRFamide, a well-conserved flatworm peptide that has potent stimulatory effects on schistosome muscle fibers [41] that was not identified in previous bioinformatic efforts [39],[40]. Together, these data provide biochemical validation for roughly half of the predicted prohormones in Schistosoma and demonstrate the utility of using planarians to understand flatworm parasites.
Traditional studies of neuropeptides have relied on the biochemical purification of individual peptides possessing interesting biological activities [72]. However, with the application of genomic and peptidomic technologies, a major bottleneck has been the characterization of this expanded collection of neuropeptide-encoding genes (and their encoded peptides) in vivo. Here we characterized peptide hormones in S. mediterranea using genomic, molecular, and biochemical approaches and determined the tissue-specific expression patterns for the entire collection of prohormone genes. Comparing the distribution of prohormone expression between asexual and sexual planarians, we identified a single prohormone gene, npy-8, as important for the maintenance of reproductive function. While our main focus was to understand the role of peptide hormones in planarian reproductive development, these studies lay the groundwork for using S. mediterranea as an experimental model for studies aimed at understanding the diverse functions of metazoan bioactive peptides.
Although previous studies have characterized the expression of subsets of prohormones or their corresponding peptides [73]–[76], a comprehensive accounting of the expression of these genes at the level of the whole animal has not yet been performed. Here we describe the distribution of all known neuropeptide-encoding genes in the planarian S. mediterranea by whole mount in situ hybridization. One surprising finding from these studies was the complexity of prohormone expression within the planarian CNS, which is considered to be among the most primitive centralized nervous systems in the animal kingdom [77]. We find that prohormone gene expression is localized to distinct regions of the cephalic ganglia and that many individual prohormones are expressed in unique CNS cell types. These results parallel observations in the planarian D. japonica in which small molecule neurotransmitters (e.g. serotonin and dopamine) are found in separate CNS cell populations [64]. The expression of prohormone genes in distinct regions/cell-types in the CNS suggests that processing centers for different neural functions (e.g. sensory, motor, and neuroendocrine) may be localized to chemically and spatially distinct domains of the flatworm CNS. In support of this idea, a “visual center” has been proposed to exist at the medial regions of the cephalic ganglia to which visual axons send their projections [78]. Elucidation of the functions of peptides expressed in these discrete CNS foci may help relate specific anatomical positions to distinct neural functionalities and allow for the dissection of planarian neural circuits.
Our analysis of prohormone expression also revealed that many prohormone genes are expressed in tissues of the reproductive tract. Expression of peptide prohormones has also been observed in the somatic reproductive organs of C. elegans [73]. Interestingly, the expression pattern of some planarian prohormones parallels the immunohistochemical localization of similar gene products in other invertebrates. The NPY family member Smed-npy-9 was expressed in the cement glands (or shell glands) surrounding the copulatory apparatus that are thought to be involved in egg capsule synthesis and deposition [2],[79]. Studies of S. mansoni observed NPY-like immunoreactivity in the region of Mehlis' gland [80], which is morphologically, and likely functionally [79], similar to the glands labeled by npy-9. cpp-1 encodes VPGWamide and TPGWamide, peptides that are related to the APGWamide peptides first described in molluscs [81]. We found cpp-1 to be expressed around the penis papillia and the oviducts of sexual planarians, which mirrors APGWamide localization in the molluscan oviducts and male copulatory organs [82],[83]. While specific functions for any of these peptides in planarian reproductive function remain to be elucidated, these results suggest evolutionarily conserved roles for peptides in several reproductive organs.
Two prohormone genes (ppl-1 and npy-8) were expressed differentially in the nervous systems of mature sexual versus asexual planarians. The expression of ppl-1 was similar in asexual and immature sexual animals but underwent a dramatic change in distribution during sexual maturation. Conversely, npy-8 expression was detected at similar levels and distribution in sexual animals yet was not detected in asexual animals. Interestingly, our biochemical analyses detected a number of peptides uniquely in either mature sexual or asexual planarians (Tables S1–S3). Taken together, these results indicate that sexually mature planarians possess unique signatures in both the composition and spatial distribution of peptide hormones relative to asexual and immature sexual animals.
To address the role of peptide signaling in planarian reproductive physiology we first examined the planarian prohormone convertase 2 orthologue, pc2. This analysis suggested that prohormone processing is required for regulating the dynamics of germ cell differentiation. A similar requirement for prohormone processing in germ cell development has not been described in other animal models. Loss-of-function mutations in the C. elegans pc2 orthologue egl-3 result in a range of neuromuscular defects [84],[85], but mutant animals are capable of germ cell development since they produce viable progeny. The role of the Drosophila pc2 orthologue Amontillado has not been assessed in adult reproductive development due to a requirement for this gene at multiple points during embryonic and larval development [86],[87]. Despite the fact that peptide hormones are known to regulate vertebrate germ cells [11],[12], extensive studies of prohormone convertase knockout mice have also not revealed roles for prohormone processing in germ cell development [46]. Therefore, it is likely that functional redundancies exist among the enzymes responsible for processing hormones involved in vertebrate reproduction. Given this possibility of genetic redundancy in vertebrates, we suggest systematic characterization of prohormone processing in other invertebrate models (e.g. C. elegans and Drosophila) may help address the extent to which peptide signaling regulates reproductive development in other animals.
Our studies suggest that NPY-8 may be among the prohormones processed by PC2 that are required for normal sexual development. At present it is not known which of the two predicted peptides encoded by NPY-8 influence planarian reproductive physiology. Prohormones that encode NPY-like peptides, including NPY-8, often also encode a C-terminal peptide or CPON (C-flanking peptide of NPY) [39],[58],[88],[89]. Because the functions of both vertebrate and invertebrate CPON peptides remain elusive, we speculate that the NPY-related peptide NPY-8A is the functional unit of this prohormone. In vertebrates, NPY signaling is thought to elicit diverse effects on the neuroendocrine axis regulating reproduction. Depending on the hormonal milieu, NPY administration can either promote or inhibit surges of luteinizing hormone [90], a gonadotropin that regulates multiple functions in the male and female reproductive systems [10],[11],[13]. The hypothalamic gonadotropin-releasing hormone, which promotes luteinizing hormone release from the pituitary, can also be influenced by NPY [91],[92]. Additionally, NPY may influence the timing of sexual maturation in mammals since it has been suggested to either induce or inhibit the onset of puberty [93]. Since NPY is a well-known regulator of energy homeostasis, NPY has been suggested to coordinate reproductive function with nutrient status [94]. Studies of Drosophila and Aplysia indicate similar roles for NPY-like peptides in processes related to nutrient homeostasis, such as feeding behavior [56],[95]. However, functional analyses in vertebrate [96] and invertebrate models [56] have not described obvious reproductive deficits in animals deficient for NPY-like peptides. Given the fact that S. mediterranea possesses an expanded collection of NPY-like peptides relative to other animals, additional work will be required to determine whether the function of NPY-8 represents an ancestral or derived function for NPY-like peptides.
Coordinated signaling between the hypothalamus, the pituitary, and the gonads controls vertebrate reproduction. Although our initial observation with pc2 RNAi implicated prohormone processing in planarian germ cell development, the site of action of this effect was difficult to interpret since pc2 expression was detected in both the nervous system and the testes. Our studies of npy-8 have clarified the role of the nervous system in planarian reproduction. npy-8 is expressed in both the central and peripheral nervous systems, and its transcripts are not detected in tissues affected by npy-8 RNAi, such as the testes. Therefore, peptides from NPY-8 are likely to act in a neuroendocrine fashion to influence reproductive development. Since amputation studies suggest that signals from the cephalic ganglia are essential for the maintenance of mature gonads in planarians [14],[15], one possible source of NPY-8 is from the cephalic ganglia.
The function of pc2 within the testes is presently not known, but testes are likely to be a site of prohormone processing since we detect the expression of multiple peptide prohormones (ilp-1 and spp-10) in this organ. Because peptide hormones can act as endocrine and paracrine signaling molecules in the vertebrate testes [12], it is possible that peptides play similar roles in planarians. Therefore, we propose that peptides (e.g. NPY-8 peptides) from the nervous system promote events associated with reproductive maturation (i.e. the production of differentiated germ cells) and peptides produced in the testes may provide feedback to the CNS and other organ systems about the physiological state of the gonads. Additionally, peptides expressed within the testes may serve as paracrine factors that regulate germ cell maturation. This possibility of coordinated signaling between CNS and the gonads may explain why the effects of pc2 RNAi on the reproductive system are more severe than those of npy-8 RNAi. Due to a lack of sufficient markers our studies have not examined the effects of neuropeptide signaling on ovarian development; future efforts will be directed at examining this question.
Although a chromosomal translocation distinguishes sexual and asexual S. mediterranea [42],[97], the strain-specific differences that account for their divergent modes of reproduction remain uncharacterized. With the exception of genes expressed in the reproductive system [98], little is known about the transcriptional differences between these strains. Here we identify npy-8 as enriched in sexual animals and show an important role for this gene in sexual development. Interestingly, the regressed testes of mature sexual animals treated with either pc2 RNAi or npy-8 RNAi resemble the primordial germ cell clusters of asexual planarians that also label exclusively with gH4 and nanos [44]. These observations, together with the loss of somatic reproductive structures in npy-8(RNAi) animals, suggest that lack of NPY-8 expression in asexual planarians may, in part, account for their inability to promote germ cell differentiation and initiate sexual maturation. However, because the phenotypes observed with pc2(RNAi) were more severe than those observed with npy-8(RNAi), we anticipate future studies may uncover additional factors that act in concert with npy-8 to influence planarian reproductive maturation.
According to one estimate, schistosomiasis (infection by Schistosoma) can be directly attributed to as many as 280,000 deaths per year in sub-Saharan Africa alone [99]. Despite the medical and economic impact of schistosomiasis, only a single chemotherapeutic agent (praziquantel) is currently used in treatment of this disease [100]. Therefore, identifying novel anthelmintic agents is an important goal of flatworm research. Schistosome eggs can become lodged in host tissues, such as the liver and bladder, forming granulomas that are the major cause of the pathology associated with schistosomiasis [100]. Thus, targeting reproductive function in adult animals represents a promising means by which to treat and control schistosome infection. The S. mansoni genome is predicted to encode two NPY-like prohormone genes: Sm-npp-20a and Sm-npp-20b [39],[101]. Comparison of the predicted peptides from these prohormones with NPY-like peptides from S. mediterranea found that the NPY-like peptide encoded from Sm-npp-20a shares its closest similarity to NPY-8A (∼48% identity, ClustalW) (Figure 2C). Given this observation, and the similarities in the reproductive anatomy between planarians and trematodes [2], it is possible that these animals employ similar mechanisms to control their reproductive output. Therefore, our results justify efforts aimed at understanding the role of peptide hormones in flatworm reproductive physiology and suggest that neuropeptide signaling may represent a viable target for the treatment and eradication of flatworm parasites.
Sexual and asexual S. mediterranea were maintained at 20°C in 0.75× and 1.0× Montjuïc salts, respectively [102]. To minimize non-specific background from gut contents after feeding, animals were starved at least 1 wk prior to use. For all experiments with sexual S. mediterranea, sexually mature animals (∼1 cm in length, unless otherwise specified) with a well-developed gonopore were used, unless otherwise specified.
All chemicals were obtained from Sigma-Aldrich (St. Louis, MO) unless otherwise stated. The peptide standards for Matrix-assisted laser desorption/ionization time-of-flight mass spectrometry (MALDI-TOF MS) calibration were purchased from Bruker Daltonics (Billerica, MA).
For LC/MS analysis, peptide extracts were prepared from 80–100 sexual or asexual planarians. Whole animals were mechanically homogenized in 8–10 mL of acidified acetone (40∶6∶1 acetone/water/HCl) or acidified methanol (90∶9∶1 methanol/acetic acid/water). After sonication, vortexing, and centrifugation of the homogenate, the supernatant was collected and the organic solvent was removed by evaporation in a SpeedVac concentrator (Thermo Scientific, San Jose, CA). The supernatant was then filtered through a Microcon centrifugal filter with a 10 kDa cutoff (Millipore, Billerica, MA), evaluated for peptide content by MALDI-TOF MS sampling of 0.5 µL and subjected to sequential separations by HPLC prior to tandem MS for peptide identification.
Peptide extracts were fractionated using a microbore HPLC system Magic 2000 (Michrom BioResources, Inc., Aubum, CA) with a C18 reverse phase column (Dionex, 1,000 µm i.d., particle size 3 µm, and pore size 100 Å) at a 20 µL/min flow rate over a 70 min run. A four-step linear solvent gradient was generated by mixing mobile phases A (95% water and 5% acetonitrile (ACN), 0.1% formic acid (FA) and 0.01% trifluoroacetic acid (TFA), and B (95% ACN, 5% water, 0.1% FA, and 0.01% TFA) as follows: 5%–10% B in 20 min, 10%–50% B in next 30 min, 50%–80% B in next 10 min, isocratic 80% B for 5 min, 80%–5% B in 4 min. Fractions were manually collected, evaluated for peptide content by MALDI-TOF MS, and subjected to 2nd stage separation using a Micromass HPLC system (Manchester, U.K.) equipped with a C18 reverse phase column (Dionex, 300 µm i.d., particle size 3 µm, and pore size 100 Å) and coupled to a HCT Ultra ion-trap mass spectrometer via an electrospray ionization source (ESI) (Bruker Daltonics, Bremen, Germany). Second stage separation parameters were optimized individually for each fraction using either the same water/ACN solvent system or water/methanol with 0.1% FA as a counter-ion. Mass spectrometric detection of eluting peptides was controlled by the Esquire software (Bruker Daltonics, Bremen, Germany) in a data-dependent manner. Tandem MS ion precursor selection was limited to 3 ions per min sorted by signal intensity, preferred charge state was set to +2, and the active dynamic exclusion of previously fragmented precursor ions limited to 2 spectra per minute. The scan m/z ranges for MS and MS/MS were 300–1,800 and 50–3,000, respectively.
For peptide identification, tandem mass spectra were converted to the .mgf file format (Mascot generic file) and processed for sequencing automatically using the PEAKS Studio 4.5 software (Bioinformatics Solutions, Inc., Waterloo, CA). PEAKS generated data were manually inspected and verified. Automatic sequencing was performed against an in-house planarian prohormone database using the following search parameters: cleavage sites, variable Post-Translational Modifications (PTMs) (including N-terminal pyro-Glu and pyro-Gln, C-terminal amidation, and disulfide bond; the maximum number of PTMs on a single peptide was set to four), mass tolerance equal 0.3 Da for the precursor ion, and 0.5 Da for fragments.
Criteria for peptide assignments and prohormone confirmation were based on confidence scores generated by PEAKS for each sequenced peptide and detection mass error. A PEAKS confidence score is given as a percentage value from 1% to 99% and represents the statistical likelihood that an amino acid sequence matches a given MS fragmentation spectrum. The PEAKS statistical algorithm considers factors such as signal to noise, total intensity, and spectrum tagging (PEAKS Studio Manual 4.5 http://www.bioinformaticssolutions.com/products/peaks/support/PEAKSStudioManual4.5.pdf). Our results are based on the current database of 51 prohormones. Our criteria for the validation of a prohormone include the identification of at least one peptide from the prohormone with a PEAKS score >80% and a mass accuracy ≤300 ppm, or with a score of >50% and a mass accuracy within 150 ppm. In addition, we manually verified automatic sequencing results, examined prohormone cleavage sites, and evaluated the possible PTMs of the identified peptides. A match of at least three consecutive fragments in an ion series from manual sequencing to an automatically generated peptide sequence was considered sufficient to validate the peptide assignment. As prohormone identification increases with the number of detected encoded peptides, we employed high identification criteria for the first peptide but allowed lower standards for assignment of additional peptides from the same prohormone (PEAKS score >20%, mass accuracy ≤500 ppm) provided the fragmentation spectrum satisfied manual verification.
In cases in which a prohormone had already been confirmed by tandem MS, occasionally we assigned peptides by mass match with MALDI-TOF-MS data. Such assignments were based on a mass-match within 200 ppm to protonated molecular ions of peptides predicted by NeuroPred (http://neuroproteomics.scs.uiuc.edu/neuropred.html) [52]. These assignments are tentative since they are not accompanied by sequencing data.
Two distinct bioinformatic approaches were used to identify prohormone genes in the S. mediterranea genome. First, similarity searches were performed with collections of peptides or prohormones from invertebrate species such as Drosophila melanogaster, Aplysia californica, Apis mellifera [32], Caenorhabditis elegans [73], and various Platyhelminthes [39] with stand-alone BLAST (BLOSSUM62 or PAM30 matrices and Expect values ≥10). Peptides and prohormones were obtained from genome databases (i.e. Wormbase, http://www.wormbase.org), from NCBI, or from an online catalog of bioactive peptides (http://www.peptides.be, [103]). Additionally, sequence tags generated by de novo MS sequencing of unassigned peptides were also used as queries for genomic BLAST searches (BLOSSUM62 or PAM30 matrices and Expect values ≥10). As an alternative to similarity searching we analyzed translated S. mediterranea EST [98],[104] and 454 (Roche, Mannheim, Germany) sequence data (Y. Wang and P.A. Newmark, unpublished) for sequences that possessed characteristics of prohormone genes including multiple dibasic cleavage sites and a signal sequence (www.cbs.dtu.dk/services/SignalP). Translations of nucleotide sequences were performed with longorf.pl, a script that translates the longest open reading frame in a nucleotide sequence (www.bioperl.org/wiki/Bioperl_scripts). Putative prohormone genes identified using these two approaches were used as queries to search the S. mediterranea genome to determine if additional related prohormones existed in the genome. The full-length coding sequences of prohormone genes were predicted using a variety of gene and splice-site prediction tools, including NetGene2 (http://www.cbs.dtu.dk/services/NetGene2), FSPLICE (http://www.softberry.com), GENSCAN (http://genes.mit.edu/GENESCAN.html), and GeneQuest (v8.0.2, DNASTAR, Madison, WI). Where full-length sequences could not be predicted in silico, 5′ and 3′ Rapid Amplification of cDNA Ends (RACE) (FirstChoice RLM-Race Kit, Ambion, Austin, TX) analyses were performed following the manufacturer's protocol. The predictions of all genes reported here were independently verified by cDNA analysis (see below). Once verified, genes were considered to be genuine prohormone genes if they (1) possessed a signal sequence, (2) possessed basic cleavage sites that flanked predicted or MS-confirmed peptides, and (3) were less than 200 amino acids in length. Sequences were excluded if they shared similarity with genes previously annotated to be other than neuropeptide prohormones. All genes were named according to the S. mediterranea genome nomenclature guidelines [105].
Translated nucleotide sequences were downloaded either from the Schistosoma mansoni FTP server (ftp.sanger.ac.ik/pub/pathogens/Schistosoma/mansoni) or from the NCBI taxonomy browser (http://www.ncbi.nlm.nih.gov/Taxonomy/). These sequences were then compared to the sequences of MS-confirmed S. mediterranea prohormones using BLASTP. NPY-family members were not included in this analysis, although three NPY-like proteins have been previously described in Schistosoma [39],[101]. Additionally we analyzed EST sequences in the NCBI database to identify schistosome prohormone genes. Newly annotated schistosome prohormones were analyzed further with SignalP and Neuropred to predict final gene products. These genes were named as described previously [39].
To facilitate efficient analyses of prohormone genes, we constructed a plasmid vector that permits TA-mediated cloning of PCR-amplified cDNAs. To generate a suitable vector backbone, oligonucleotide primers 5′-GATCACGCGTCGATTTCGGCCTATTGGTTA-3′ and 5′-GATCACGCGTGCTTCCTCGCTCACTGACTC-3′ were used to amplify the kanamycin and ampicillin resistance markers and the origin of replication of plasmid pCRII (Invitrogen, Carlsbad, CA); this PCR product was digested with MluI and ligated to generate a circular plasmid. Following circularization, an Eam1105I restriction site was removed from the β-lactamase gene of this plasmid by introduction of a silent mutation using site-directed mutagenesis (Quickchange II, Statagene, La Jolla, CA). For the functional elements of the vector, two mini genes were synthesized (Integrated DNA Technologies, Coralville, IA): T7TermSP6 and T7TermT3. T7TermSP6 included (5′ to 3′) KpnI, MluI, T7-terminator, AscI, T7 Promoter, SP6 promoter, GACCTTAGGCT (an Eam1105I site), and XhoI. T7TermT3 included (5′ to 3′) SacI, MluI, T7 terminator, T7 promoter, T3 promoter, GACCTTAGGCT (an Eam1105I site), and NotI. T7TermSP6 and T7TermT3 were shuttled to pBluescript SK II+ using the KpnI and XhoI sites from T7TermSP6 or the SacI and NotI sites from T7TermT3. These plasmids were digested with MluI and EcoRI and ligated with the MluI site of the vector backbone. A XhoI and NotI-digested PCR fragment including the ccdB and camR genes from plasmid pPR244 [47] were inserted to generate the final plasmid-pJC53.2. Eam1105I (Fermentas, Burlington, Ontario) restriction digest of this plasmid generates 3′ T overhangs that can be ligated to an A-tailed Taq polymerase-amplified PCR product [106]. The ccdB gene prevents any undigested plasmid from giving rise to viable clones [107]. Once cDNAs have been inserted into pJC53.2, riboprobes for in situ hybridization analysis can be generated by in vitro transcription with SP6 or T3 RNA polymerases and dsRNA for RNAi knockdowns can be generated by in vitro transcription with T7 RNA polymerase, or by transformation of E. coli (HT115[DE3]) [108].
To generate riboprobes for in situ hybridization, prohormone genes not represented by EST clones [98] were PCR amplified (Platinum Taq, Invitrogen, Carlsbad, CA) from cDNA generated from total RNA (iScript cDNA Synthesis Kit, Bio-Rad, Hercules, CA) or 3′ RACE cDNA (RLM-RACE Kit, Ambion, Austin, TX) generated from either total or poly-(A)+ RNA (Poly-A Purist, Ambion, Austin, TX). For cDNA preparations, RNA was extracted using Trizol Reagent (Invitrogen, Carlsbad, CA). For cloning, 2–3 µL of PCR product was ligated with 70 ng of Eam1105I-digested pJC53.2 (Rapid DNA Ligation Kit, Roche, Mannheim, Germany) and used to transform DH5α. In vitro transcriptions with the appropriate RNA polymerase were performed using standard approaches with the addition of Digoxigenin-12-UTP (Roche, Mannheim, Germany), Fluorescein-12-UTP (Roche, Mannheim, Germany), or Dinitrophenol-11-UTP (Perkin Elmer, Waltham, MA).
In situ hybridizations were performed using the formaldehyde-based fixation procedure essentially as described previously [109]. However, due to their large size, sexual animals were killed in 10% N-Acetyl Cysteine, fixed for 20–30 min in 4% Formaldehyde in PBSTx (PBS+0.3% Triton X-100), permeabilized with 1% SDS (10 min at RT) prior to reduction (10 min at RT), and treated with 10 µg/mL Proteinase K (10–20 min at RT) after bleaching. Some samples were processed in either a BioLane HTI (Hölle & Hüttner, Tübingen, Germany) [98] or an Insitu Pro (Intavis, Koeln, Germany) hybridization robot [102]. Sexual animals were imaged with either a Microfire digital camera (Optronics, Goleta, CA) mounted on a Leica MZ12.5 stereomicroscope or a Leica DFC420 camera mounted on a Leica M205A stereomicroscope (Leica, Wetzlar, Germany). Both microscopes were equipped with a Leica TL RC base. Asexual animals were imaged over a piece of white filter paper and illuminated from above with an LED light source.
For FISH, following post-hybridization washes and blocking, animals were incubated in α-Digoxigenin-POD (1∶1000, Roche, Mannheim, Germany), α-Fluorescein-POD (1∶1000, Roche, Mannheim, Germany), or α-Dinitrophenol-HRP (1∶100, Perkin Elmer, Waltham, MA) overnight at 4°C, washed in MABT, equilibrated in TNT (100 mM Tris pH 7.5, 150 mM NaCl, and 0.05% Tween-20), and developed in Amplification Diluent containing a fluorescent-tyramide conjugate (Cy3-tyramide, Cy5-tyramide, or Fluorescein-tyramide; TSA-Plus, Perkin Elmer, Waltham, MA). Following development, animals were washed in TNT and HRP activity was quenched by a 1 h incubation in 1.5%–2.0% H2O2 dissolved in TNT. Following HRP inactivation, animals were washed in MABT, incubated in a different α-hapten-HRP antibody, and the process was repeated with a different fluorescent-tyramide conjugate. Samples were mounted in Vectashield (Vector Laboratories, Burlingame, CA) and imaged on a Zeiss LSM 710 confocal microscope (Carl Zeiss, Germany) (Plan-Apochromat 20×/0.8, C-Apochromat 40×/1.2 W korr UV-VIS-IR, or Plan-Apochromat 63×/1.4 Oil DIC objectives). Fluorescein, Cy3, and Cy5 were excited with 488 nm, 561 nm, and 633 nm lasers, respectively. Images were processed using either Zen 2008 (Carl Zeiss, Germany) or ImageJ [110].
Northern blot procedures were performed essentially as previously described [111] and hybridization signals were detected using an anti-digoxigenin alkaline phosphatase-conjugated antibody and chemiluminescence (CDP-STAR, Roche, Mannheim, Germany). Chemiluminescent signals were detected using a FluorChem Q (Alpha Innotech, San Leandro, CA).
Sequences of EST clones corresponding to pc2 [43],[98] were assembled with one another and the S. mediterranea genome (Sequencher 4.7, Gene Codes, Ann Arbor, MI) to determine the full-length sequence and genomic structure of the pc2 gene.
For RNAi analysis of pc2, EST clone PL05006A1C09 [98], which corresponds to pc2, was shuttled to plasmid pPR244 using a Gateway reaction (Invitrogen, Carlsbad, CA) [47]. For npy-8 RNAi, a 3′ RACE product specific to npy-8 was cloned in pJC53.2. RNAi feedings were performed essentially as described previously [112], with some modifications. In pc2 RNAi experiments, ∼6.25 mL of IPTG-induced culture was pelleted, frozen at −80°C, and resupended in 30 µL of a mixture of homogenized beef liver and water. ∼5 mature sexual animals (>1 cm in length) received 1–2 feedings over the course of ∼48 h. npy-8 RNAi experiments were performed similarly to pc2 RNAi except feedings included 50% less bacteria and animals were fed every 5–7 d over the indicated time course; for some feedings, bacteria were omitted. On occasion, because of either refusal to feed or improper nutrition, some animals (both controls and treatment groups) decreased in size over the long time courses of the npy-8 RNAi experiments. Therefore, only animals >1 cm in length at the time of fixation were included in our analyses at time points greater than 4 wk. For all RNAi experiments with bacterially expressed dsRNA, control feedings were performed with bacteria containing empty plasmid pPR242.
For RNAi experiments conducted with juvenile planarians, dsRNA was generated by in vitro transcription [113],[114]. To generate dsRNA, templates cloned in pJC53.2 were amplified with a modified T7 oligonucleotide (GGATCCTAATACGACTCACTATAGGG), cleaned up using the DNA Clean & Concentrator kit (Zymo Research, Orange, CA, D4003), and eluted in 10 µL of water. 4 µL of each PCR product was used as template for in vitro transcription in a reaction containing 5.5 µL DEPC-treated water, 5 µL 100 mM mix of rNTPs (Promega, E6000), 2 µL high-yield transcription buffer (0.4 M Tris pH 8.0, 0.1 M MgCl2, 20 mM spermidine, 0.1 M DTT), 1 µL thermostable inorganic pyrophosphatase (New England Biolabs, Madison, WI, M0296S), 0.5 µL Optizyme recombinant ribonuclease inhibitor (Fisher Scientific, Pittsburg, PA, BP3222-5), and 2 µL HIS-Tagged T7 RNA polymerase [115]. Samples were incubated at 37°C for 4–5 h and then treated with RNase-free DNase (Fisher Scientific, Pittsburg, PA, FP2231). Synthesized RNA was then melted by heating at 75°C, 50°C, and 37°C each for 3 min. 2.5–10 µg of each dsRNA solution was mixed with 45 µL of 3∶1 liver to water mix and used to feed up to 8 worms. For these experiments, animals without visible gonopores (juveniles) were fed every 4–5 d for the indicated time period and starved 1 wk before fixation. Unless otherwise specified, as a negative control, animals were fed dsRNA synthesized from the ccdB and camR-containing insert of pJC53.2.
To analyze the structure of the testes, animals were killed in 2% HCl for 3 min, fixed in either Methacarn (6 MeOH:3 Chloroform: 1 Glacial Acetic Acid) or 4% formaldehyde for 1–2 h, dehydrated in MeOH, bleached in 6% H2O2 in MeOH, and stained with 4′,6-diamidino-2-phenylindole (DAPI) (Sigma-Aldrich, St. Louis, MO). Alternatively, samples were processed for in situ hybridization, as described above. Following staining, animals were mounted in Vectashield, flattened, and imaged on either a Zeiss SteREO Lumar (Carl Zeiss, Germany) or a Zeiss LSM 710 confocal microscope (DAPI was excited with a 405 nm laser).
To examine if npy-8(RNAi) affected overall growth, animals were immobilized on ice and imaged on a Leica M205A stereomicroscope. The area of each animal was determined using ImageJ.
To examine transcript levels in npy-8 knockdowns, juvenile animals were fed either liver homogenate or 45 µL of liver homogenate mixed with 2.5 µg of in vitro synthesized npy-8 dsRNA. 7 d later RNA was extracted from individual planarians using Trizol Reagent (Invitrogen, Carlsbad, CA). Following DNase treatment (DNA-free RNA Kit, Zymo Research, Orange, CA), reverse transcription was performed (iScript cDNA Synthesis Kit, Bio-Rad, Hercules, CA) and quantitative PCR was conducted using Power SYBR Green PCR Master Mix (Applied Biosystems, Warrington, UK) and a 7900HT real-time PCR system (Applied Biosystems). Standard curves were generated from serial dilutions of either plasmid DNA containing the gene of interest (npy-8 and npy-1) or from genomic DNA (β-tubulin GB: DN305397). All samples were measured in triplicate to account for pipetting error. Absolute quantities of each transcript were determined from the standard curves and the levels of npy-8 or npy-1 were normalized to the level of β-tubulin in each sample. The mean value (i.e. npy-8/β-tubulin or npy-1/β-tubulin) for each treatment (i.e. control or npy-8(RNAi)) was then compared using a Student's t test. The primers used for these studies were npy-8 Forward AATCAGAAAAGGCCGATGTTTG, Reverse CAAATAGTTCCGAAAGGCATCAG; npy-1 Forward GTCGACCAAGATTCGGTAAACG, Reverse CATTCTTTTATGAAAATCCCCTGT; β-tubulin F TGGCTGCTTGTGATCCAAGA R AAATTGCCGCAACAGTCAAATA.
To investigate the effect of pc2 RNAi on the proteolytic processing of prohormones, peptide profiles were measured by MALDI-TOF MS and compared by principal component analysis followed by a t test in tissue extracts prepared from 7 individual control and 7 individual RNAi-treated animals. Extracts were prepared by homogenizing each specimen in 100 µL of acidified acetone (see above). Following centrifugation at 14,000× g for 15 min, supernatant was collected, dried in SpeedVac concentrator (Thermo Scientific, San Jose, CA), and reconstituted in 30 µL of 0.01% TFA. For MALDI-TOF MS analysis, 0.7 µL of each extract was spotted on a stainless steel sample holder and co-crystallized with 0.7 µL of freshly prepared concentrated DHB matrix (DHB: 2,5-dihydroxybenzoic acid, 50 mg/mL 50% acetone). Three technical replicates were sampled for each biological sample, 42 spots total. Positive ion mass spectra were acquired manually in 600–6,000 m/z region using a Bruker Ultraflex II mass spectrometer in linear mode with external calibration. For each spot 700 laser shots in 7 acquisitions were accumulated into a sum spectrum representative of a replicate.
For comparison of peptide profiles in control and pc2(RNAi) animals, raw MALDI-TOF MS data were loaded into an evaluation version of ClinProTools software (Bruker Daltonics, Bremen, Germany) using the following processing parameters: convex hull baseline subtraction, baseline flatness 0.2, mass range 1,000–6,000 m/z, Savizky-Golay smoothing over 1 m/z width with 11 cycles, data reduction factor of 10, null spectra exclusion enabled, recalibration with maximum peak shift of 200 ppm. All spectra were normalized to the total ion count (TIC) prior to PCA calculations. Sum spectra from technical replicates were grouped into a representative sample spectrum in ClinProTools, thus representing a biological replicate for statistical calculations. From representative sample spectra a mean spectrum was generated by ClinProTools to reveal general peptide features for control and pc2(RNAi) groups. Standard deviation of signal intensities among biological replicates was derived for each peak in the group profile. Unlimited peak picking on the base of maximal peak intensity and minimal signal-to-noise ratio of 6 was done on the mean spectrum representative of each sample group in order to take advantage of noise reduction effect due to spectra addition. Peptide profiles of mean spectra representative of biological replicates were compared by principal component analysis followed by Anderson-Darling (AD) normality test and paired Student's t test for peaks showing normal distribution. Peaks not showing a normal distribution (pAD≤0.05) were evaluated by the Wilcoxon or Kruskal-Wallis tests, respectively [116]–[118]. To decrease the number of false positives while computing individual peak statistics on highly complex spectra, the Benjamini-Hochberg procedure incorporated into ClinProTool was automatically applied for p value adjustment during analysis [119].
|
10.1371/journal.ppat.1001061 | Variations in TcdB Activity and the Hypervirulence of Emerging Strains of Clostridium difficile | Hypervirulent strains of Clostridium difficile have emerged over the past decade, increasing the morbidity and mortality of patients infected by this opportunistic pathogen. Recent work suggested the major C. difficile virulence factor, TcdB, from hypervirulent strains (TcdBHV) was more cytotoxic in vitro than TcdB from historical strains (TcdBHIST). The current study investigated the in vivo impact of altered TcdB tropism, and the underlying mechanism responsible for the differences in activity between the two forms of this toxin. A combination of protein sequence analyses, in vivo studies using a Danio rerio model system, and cell entry combined with fluorescence assays were used to define the critical differences between TcdBHV and TcdBHIST. Sequence analysis found that TcdB was the most variable protein expressed from the pathogenicity locus of C. difficile. In line with these sequence differences, the in vivo effects of TcdBHV were found to be substantially broader and more pronounced than those caused by TcdBHIST. The increased toxicity of TcdBHV was related to the toxin's ability to enter cells more rapidly and at an earlier stage in endocytosis than TcdBHIST. The underlying biochemical mechanism for more rapid cell entry was identified in experiments demonstrating that TcdBHV undergoes acid-induced conformational changes at a pH much higher than that of TcdBHIST. Such pH-related conformational changes are known to be the inciting step in membrane insertion and translocation for TcdB. These data provide insight into a critical change in TcdB activity that contributes to the emerging hypervirulence of C. difficile.
| Clostridium difficile is a spore-forming bacterium that contaminates hospitals and infects patients undergoing antibiotic therapy. C. difficile is now the leading cause of hospital-acquired diarrhea in developed countries. Most concerning has been the recent increase in mortality of C. difficile patients due to the emergence of a hypervirulent strain of this pathogen. Results from the current study suggest this change in disease severity may be due to new strains producing a variant form of C. difficile's major virulence factor, TcdB. The findings indicate TcdB from hypervirulent strains targets a much broader range of cells in vivo and is able to translocate into target cells more quickly than TcdB from historical strains of C. difficile. The more rapid cell entry by TcdB from hypervirulent C. difficile appears to be due to the toxin's capacity to undergo conformational changes necessary for membrane translocation at a higher pH than TcdB from historical strains. To date, very little has been learned about the underlying reasons for the increased virulence of emerging C. difficile strains. These findings provide insight into this problem and suggest variations in TcdB activity could be an important contributing factor to the hypervirulence of emerging strains of C. difficile.
| Clostridium difficile is a gram-positive, spore-forming anaerobe, first described by Hall and O'Toole over 75 years ago [1]; however, the organism was not associated with human disease until 1978 [2], [3]. Over the past three decades C. difficile has become a major nosocomial pathogen and is the leading cause of diarrhea in hospitalized patients [4]. C. difficile associated disease (CDAD) is routinely treated by supportive therapy and regimens of vancomycin and metronidazole, but treatment of CDAD has become more difficult due to the emergence of hypervirulent (NAP1/BI/027) strains of C. difficile [5], [6], [7]. Elucidating the major differences between historical strains of C. difficile and the NAP1/BI/027-related strains of C. difficile is critical to understanding how this serious human pathogen continues to emerge.
The phenotypes of hypervirulent and historical strains of C. difficile are different [7], [8], [9]. C. difficile NAP1/BI/027 produces more toxin and sporulates with higher efficiency than historical strains [6], [7], [8], [9], [10]. NAP1/BI/027 strains also produce a binary toxin, CDT, which is thought to enhance colonization of C. difficile by triggering the formation of microtubule protrusions on cells of the gastrointestinal epithelium [11], [12], [13]. Finally, C. difficile NAP1/BI/027 strains are resistant to fluoroquinolones due to mutations in DNA gyrase genes [7], [14], [15], [16]. The extent to which one or more of these differences between the two strains contributes to hypervirulence has not been determined.
Recent work from Stabler and colleagues identified several genetic variations between epidemic and historical strains of C. difficile [17]. For example, the historical C. difficile strain, 630, was found to contain 505 unique coding sequences compared to hypervirulent strains. This analysis also identified differences in flagellar genes, metabolic genes, phage islands, and transcriptional regulators. Of interest to our work was the finding that TcdB from C. difficile hypervirulent strains had a greater cytopathic effect on a variety of cell types than TcdB isolated from a C. difficile historical strain. The steps in cellular intoxication that account for these differences in TcdB activity, and whether in vivo tropism varies between the historical and hypervirulent TcdB have not been reported.
TcdB (∼269 kDa) is a 2366 residue single polypeptide toxin encoded on a C. difficile pathogenicity locus (PaLoc) that also includes genes for two regulators (TcdC and TcdR) of toxin expression, a putative holin (TcdE), and TcdA [18], [19]. TcdB has at least four functional domains that contribute to cell entry and glucosylation of small-GTPases within the cytosol of the cell [20]. TcdB's glucosyltransferase domain is included in the first 516 residues of the toxin, which also includes a conserved DXD motif (Asp286/Asp288) and Trp102, which form a complex with Mn2+ and UDP-Glucose [21], [22], [23], [24], [25]. A substrate recognition domain is located between residues 365–516 [26]. The cysteine protease domain at residues 544–955 is necessary for autoproteolytic activity and delivery of the enzymatic domain into the cytosol [27], [28], [29]. A putative membrane-spanning domain resides between residues 956–1128, yet whether this domain is required for intoxication is not known. Finally, the fourth functional domain of TcdB is located within the carboxy-terminal region of the toxin, and is predicted to interact with receptors on target cells [30], [31], [32], [33].
Sequence variations in one or more of the functional domains of TcdB could account for the differences in cytotoxicity between historical and hypervirulent isolates. In the current work we test this hypothesis and demonstrate that TcdB from hypervirulent strains exhibits broader tropism in vivo. We also demonstrate TcdB from hypervirulent C. difficile undergoes hydrophobic conformational changes at a higher pH than toxin from the historical strain, and this correlates with more rapid cell entry. These findings provide insight into a possible mechanism through which hypervirulent C. difficile causes more severe illness than historical strains of this organism.
The carboxy-terminal sequence of TcdB varies between isolates of C. difficile, including hypervirulent and historical strains [17], [34]. Yet, whether sequence variations are more extensive in TcdB compared to other genes in the PaLoc or if the sequences outside of the carboxy-terminal domain of TcdB also varied among different strains of C. difficile has not been reported.
We compared the sequences of proteins encoded within the PaLoc of C. difficile 630 (a non-NAP1/BI/027 strain) and C. difficile R20291 (a 027 strain). The sequence of TcdR, a positive regulator of toxin expression was found to be 100% identical between the two strains of C. difficile. TcdE, the putative holin encoded in the middle of the PaLoc exhibited 99% identity and 100% similarity between the two strains of C. difficile. The enterotoxin, TcdA, exhibited 98% identity and 99% similarity between the two strains. The gene encoding TcdC from the hypervirulent strain encodes a stop codon and contains a deletion, which made it difficult to precisely compare this protein in the two strains. However, at the DNA level the gene was 95% homologous in the intact coding regions of tcdC. In contrast to these almost exact identities of TcdR, TcdE, and TcdA from the two strains, the amino-acid sequence of TcdB from the two strains was found to have the most variation with 92% identity and 96% similarity.
We next compared the functional regions of TcdBHIST and TcdBHV (Fig. 1). The enzymatic region of TcdB (encompassing residues 1–543) was found to be 96% identical and 98% similar between the two strains of C. difficile. Residues critical for catalytic activity, W102 and the DXD motif, did not vary between the two forms of TcdB (Fig. 1A). The substrate specificity domain of TcdB (residues 365 to 516) [26] exhibited 99% identity and 100% similarity (Fig. 1A). The autoproteolytic region (residues 544 to 955) was found to contain 96% identity and 98% similarity. Moreover, the reported catalytic triad (D587, H653, and C698) was conserved between the two forms of TcdB. Interestingly however, the analysis found a rearrangement of a second cysteine residue in this region of TcdB. TcdBHIST contains a cysteine at residue 870, but this residue is a tyrosine in TcdBHV (Fig. 1B). Conversely, TcdBHV has a cysteine residue at 1477, but this was found to be a glycine residue in TcdBHIST. The third putative functional domain of TcdB is between residues 956 and 1644, and encodes a hydrophobic region thought to mediate membrane insertion. Comparison of this region found 91% identity and 96% similarity (Fig. 1C).
In line with earlier reports [17], [34] the carboxy-terminal region, encompassing residues 1645 to 2366, exhibited the highest degree of sequence variation in the toxin. The carboxy-terminal region showed 88% identity and 95% similarity between the two forms of TcdB. The number of CROP regions is identical, with TcdBHIST and TcdBHV containing 24 regions based on the YF consensus motif [30], [32], [35], [36]. However, eight of these regions in TcdBHV were found to exhibit less than 80% sequence identity to TcdBHIST (Fig. 1D).
Fig. 1E shows an SDS-PAGE analysis of TcdBHIST and TcdBHV purified from wild-type strains of C. difficile as described in the materials and methods. Both forms of the toxin were obtained at greater than 95% purity based on minimal detection of contaminating proteins.
We next used a zebrafish model to compare the in vivo effects of the two forms of this toxin. Our group has previously utilized the zebrafish embryo as a model to examine the effects of TcdBHIST in real time, and found that this toxin had potent cardiotoxic effects [37]. The zebrafish provides a distinct advantage for the purpose of examining tissue damage and tropism because it is possible to visualize these events directly with this model.
Zebrafish embryos were arrayed in a 48-well plate in embryo water and TcdBHIST or TcdBHV across a range of concentrations was applied to the individual wells. At 24 h following treatment, a minimum of 20 zebrafish larvae per condition were examined by light microscopy for physiological changes, tissue damage, and viability (Fig. 2). Extensive necrosis was evident in all embryos exposed to TcdBHV, with broad tissue damage caused to the yolk sac, body, and head at concentrations as low as 1 nM (Fig. 2B and 2D). Furthermore, all zebrafish treated with TcdBHV succumbed to the effects of the toxin within 48 h. In contrast, treatment with TcdBHIST resulted in more specific damage at the cardiac region in approximately 75% of embryos, and was not immediately lethal (Fig. 2A). Incubation with higher doses of TcdBHIST or for longer periods of time increased toxicity but did not alter the physiological damage from this toxin. These findings indicate that TcdBHV impacts a broader number of cell types in vivo compared to TcdBHIST. However, corresponding to our previous report TcdBHIST preferentially targets cardiac cells in the zebrafish embryo system.
Recent studies determined the relative cytotoxicity of TcdBHV and TcdBHIST on eight different cell types [17]. Because this analysis did not include cells of cardiac lineage, we compared the two toxins on HL-1 cells, which are derived from mouse cardiac tissue [38]. We also examined the effects of the two toxins on CHO cells for a relative comparison to the cardiomyocytes. As shown in Fig. 3, similar to previous observations, TcdBHV was more cytotoxic to CHO cells (TCD50 2.37×10−13 M) than was TcdBHIST (TCD50 2.53×10−11 M). In contrast, TcdBHV was not more cytotoxic on cardiomyocytes and displayed a very similar activity to TcdBHIST. Upon further investigation of the cardiomyocytes, the cytotoxicity of TcdBHV was found to be slightly lower than TcdBHIST (p<0.05) with a TCD50 approximately 10-fold higher (3.37×10−10 M) than TcdBHIST (TCD50 2.80×10−11 M). These data indicate that while TcdBHV has a broader cell tropism and is most likely more cytotoxic overall, TcdBHIST cardiotropism is more pronounced between the two forms of this toxin.
We next determined if the variation in cytotoxicity was due to differences in the cytosolic activities of the two forms of TcdB. As an approach to this problem we took advantage of a previously described system used for heterologous delivery of proteins and protein fragments into the cytosol of target cells [39], [40]. This system is composed of the cell entry components of anthrax lethal toxin. Briefly, protective antigen (PA) delivers lethal factor (LF) into the cytosol of mammalian cells. The heterologous delivery system is derived from the amino-terminus of LF (LFn), which interacts with PA and can be delivered into cells, but lacks enzymatic activity. In our experiments, the DNA fragment encoding the enzymatic domain of TcdB was genetically fused to lfn, yielding a DNA construct that expresses the cell entry portion of LF with the enzymatic component of TcdB. This heterologous delivery system allowed us to regulate the cell entry of the enzymatic component of TcdBHV and TcdBHIST so that these domains were identical in the way in which they entered the cell. We predicted that if the differences in cytotoxicity were due to factors other than intracellular activity of these forms of TcdB, then the fusions should exhibit identical cytotoxic effects.
The results of the PA, LFn-TcdB fusion experiments are shown in Fig. 4. CHO cells were treated with a fixed amount of PA (500 nM) plus a range of concentrations of LFnTcdBHV(enz) or LFnTcdBHIST(enz) in order to generate a standard killing curve for this assay. As controls, CHO cells were treated with PA, LFnTcdBHV(enz), or LFnTcdBHIST(enz) separately. Following 24 h of treatment the cells were assayed for viability using WST-8 colorimetric assay and the percent survival was plotted versus concentration of the fusion protein. Treatment with each of the components alone had no effect on cell viability in this assay (data not shown). Treatments with PA plus LFnTcdBHV(enz) or PA plus LFnTcdBHIST(enz) resulted in similar (p<0.05) cytotoxicity at each of the concentrations tested (Fig. 4). To confirm that PA was not limiting in these experiments, cytotoxicity of the fusions was tested with 10-fold higher amounts of PA, and this additional amount of PA did not change the level of cytotoxicity for either fusion (data not shown). The results from this experiment suggested that the differences in the cytotoxicity of LFnTcdBHV(enz) and LFnTcdBHIST(enz) were not due to variations in intracellular activities of the enzymatic domains.
The results from the experiment using an identical method of cell entry, suggested the differences in cytotoxicity might be associated with early steps in cell binding and cell entry. To address this hypothesis, we compared the interaction of TcdBHV and TcdBHIST with cultured cells. Cultured cells were incubated with Alexa-647-labeled TcdBHV or Alexa-647-labeled TcdBHIST and the extent of toxin binding was determined by flow cytometry. This analysis was performed on CHO cell and HL-1 cardiomyocytes. As shown in Fig. 5, CHO cells and HL-1 cells exhibited a higher degree of fluorescence when incubated with labeled TcdBHIST than when incubated with labeled TcdBHV. A biphasic profile was detected in CHO cells with a smaller population of cells exhibiting a distinct, reduced, toxin-binding pattern. In contrast, binding to cardiomyocytes was uniform and revealed a profile expected for a single population of cells.
Experiments were next performed to determine the apparent Kd for binding of TcdBHIST and TcdBHV. Interestingly, within the constraints of these experimental conditions we were not able to achieve saturable binding of either form of the toxin to target cells. Fig. 5C shows a nearly linear correlation between the increase in toxin concentration and the mean fluorescence intensity (MFI) of HL-1 cells despite reaching toxin concentrations of over 300 nM. Additionally, Fig. 5C further emphasizes the extremely low level of interaction of TcdBHV with target cells in comparison to the high MFI achieved with TcdBHIST. These data suggest that cell binding involves a higher order and more complex process than expected for a single receptor-ligand interaction.
Experiments were next performed to assess the difference in the rates of cell entry between the two toxins. In previous work on historical TcdB, we found that lysosomotropic inhibitors could completely block cytopathic effects of the toxin for up to 16 h, even if added up to 20 min following exposure of the cells to the toxin [41]. These findings indicate interaction with the cell, uptake, and then translocation into the cytosol requires at least 20 min and acidification of endosomes is necessary. To determine if TcdBHV differed from TcdBHIST in rates of cell entry, cultured CHO cells were treated with the two forms of the toxin and a lysosomotropic agent was added to the cells at time-points ranging from 5 to 60 min following treatment with toxin. The lysosomotropic agent was also added prior to or at the same time cells were treated with the toxins. The effect of the lysosomotropic agent was then assessed by determining the level of cytopathic effects (CPE) either 2 h or 12 h after treatment with the toxin. For this experiment CPE was determined rather than cytotoxicity due to toxicity of ammonium chloride at the later time points necessary for cytotoxicity assays. As shown in Fig. 6, based on the extent of cell rounding, there appeared to be a clear difference in the rates of translocation between TcdBHV and TcdBHIST. Unlike our earlier findings on TcdBHIST, the cytotoxic effects of TcdBHV could not be prevented when the lysosomotropic agent was added as soon as 10 min following treatment with the toxin (Fig. 6A). Furthermore, addition of the lysosomotropic agent within 10 min of treatment of TcdBHV only provided a slight delay in CPE, as all inhibitor treated cells showed complete rounding by 12 h (Fig. 6B). In contrast, the CPE of TcdBHIST could be prevented by adding the inhibitor up to 30 min following treatment with the toxin. These findings indicate TcdBHV translocates to the cytosol more rapidly than TcdBHIST.
Previous studies from our group found that acidic pH triggers hydrophobic transitions in TcdBHIST [41]. Studies by Barth et al. found that this hydrophobic transition in TcdB correlated with membrane insertion by the toxin [42]. These conformational changes corresponded to the decrease in endosome pH that led to translocation of the toxin into the cytosol. Thus, it was reasonable to suspect that TcdBHV translocates more quickly into the cytosol because the hydrophobic transition was induced at a higher pH and thus at an earlier stage of endocytosis. To address this possibility, in the next series of experiments we identified the pH dependent conformational transitions of TcdBHV by observing changes in TNS fluorescence when the toxin was incubated at various pHs. To identify whether TcdBHV exhibits differential transitions compared to TcdBHIST, the proteins were preincubated with 150 µM TNS at pH 4.0, 5.0, 6.0, and 7.0, and then analyzed for changes in TNS fluorescence. As shown in Fig. 7, TcdBHV exhibited a significant increase in hydrophobicity at pH 5.0, while TcdBHIST did not undergo this transition until pH 4.0. Further examination of a narrower pH range revealed that a significant shift occurred between pH 5.4 and 5.6 in TcdBHV (Fig. 7D). In comparison, TNS fluorescence of TcdBHIST at these pHs was just above background levels.
These pH transitions were also studied using the inherent fluorescence of TcdBHIST and TcdBHV from the emission of tryptophan residues. Unfolding of the hydrophobic region should expose portions of the protein to a more aqueous environment, quenching tryptophan fluorescence. Environmental changes surrounding the tryptophan residues over a broad range of pH are shown in Fig. 8A and 8B. A gradual quenching of fluorescence was detected in TcdBHIST from pH 7 to pH 4, while the tryptophan emission spectra of TcdBHV indicated a sudden shift between pH 5 and pH 6. Fig. 8D reveals that this shift took place between pH 5.4 and 5.2, similar to the increase in TNS fluorescence seen at pH 5.4.
In the current study we compared the sequences and activities of TcdB from hypervirulent and historical strains of C. difficile. Because TcdB has been shown to be the major virulence factor of C. difficile [43], we reasoned that changes in the activity of this toxin could have a profound impact on the severity of disease. The findings support this notion, as TcdBHV exhibited a broader tropism and higher potency than TcdBHIST. Among the possible explanations for this increased toxicity are the observations that TcdBHV enters cells more rapidly than TcdBHIST, and TcdBHV undergoes conformational changes at a higher pH than TcdBHIST.
Based on the sequence comparisons and the results of the experiments using the heterologous delivery system (Figs. 1 and 3), it appears that the differences in tropism and cytotoxicity are due to changes in regions outside of the enzymatic domain. Rapid cell entry could lead to more efficient cell killing by providing the toxin an endocytic condition in which the toxin is not subject to possible destruction by lysosomal proteases. The data from the lysosomotropic inhibitor assays (Fig. 6) support the idea that TcdBHV does not reside within the endosome as long as TcdBHIST. Among the possible reasons for more rapid cell entry is a differential sensitivity to levels of IP6 that trigger autoproteolytic processing associated with translocation. We also noted a difference in the sequence of the hydrophobic region of TcdB, and if, as has been proposed [41], [42], this region mediates membrane insertion, such differences could allow TcdBHV to insert into the membrane at an earlier stage of cell entry. We reasoned that if this possibility were true, there should be a difference in the pH-induced transitions of the two forms of TcdB, with the hydrophobic regions of TcdBHV becoming exposed at a pH higher than the pH necessary for triggering this transition in TcdBHIST. The results from the TNS experiments (Fig. 7) indicate that TcdBHV is able to undergo the hydrophobic transition at a higher pH than TcdBHIST, providing further evidence that TcdBHV has higher translocation efficiency than TcdBHIST. Studies looking at the environment surrounding tryptophan residues of TcdBHIST and TcdBHV at lower pH (Fig. 8) support the idea that TcdBHV undergoes a structural change at higher pH than TcdBHIST. Additionally, these experiments revealed that the transition of TcdBHIST occurs gradually, while TcdBHV demonstrates sudden shifts upon lowering the pH. This could be indicative of a more efficient unfolding of TcdBHV, which may contribute to an enhanced ability to traverse the endosomal membrane. Our current working model is that TcdBHV is able to translocate at an earlier point in endocytosis and this contributes, at least in part, to a more efficient intoxication.
We also recognize that the expanded tropism, along with more efficient cell entry could combine to enhance the in vivo toxicity of TcdBHV. The results from the zebrafish experiments (Fig. 2) indicate TcdBHV targets a broader array of cells in vivo than does TcdBHIST. Defining the specific tropism in the murine model or an infection model is more difficult, but it is reasonable to consider the possibility that TcdBHV is more lethal because the toxin targets an extensive variety of cell types systemically. Unfortunately, the TcdB receptor has been difficult to identify. Several attempts by our group to identify the TcdB receptor using standard techniques that have been successful with other toxins have failed. The results from the flow analyses in the current study suggest that the interaction of TcdB with the cell surface does not fit a single ligand-receptor model; this observation may explain why it has been so difficult to identify a receptor for this toxin. We were not able to achieve saturable binding, and interestingly TcdBHV interacted less efficiently than TcdBHIST, despite the fact that TcdBHV is clearly more cytotoxic than TcdBHIST. Undoubtedly, future studies on characterizing this complex interaction with target cells will provide important insight into a novel mechanism of TcdB intoxication.
Previous work by Razaq et al. found that C. difficile BI/NAP1/027 strains were more lethal than historical strains of C. difficile [44]. As mentioned in the introduction of this paper, there are several differences in the phenotypes of the hypervirulent and historical strains of C. difficile. NAP1 strains sporulate at a higher efficiency and are resistant to fluoroquinolones. Both of these characteristics may make the NAP1 strains more difficult to manage in the hospital setting and increase the frequency of disease, but are unlikely to increase virulence. Likewise, the binary toxin has been shown to enhance colonization [13], but clinical data have revealed little correlation between the increase in disease severity and production of this toxin [45], [46]. In addition, previous work found binary toxin to be enterotoxic, but strains producing binary toxin alone did not cause disease in hamsters [47]. Clearly, an increase in toxin production such as that reported for NAP1 strains could enhance virulence, but a recent report suggests that the tcdC mutation in epidemic strains does not always correlate with the overexpression of TcdA and TcdB [48]. Based on the findings from the current study, we suggest that variations in TcdB sequence and activity could be an important determining factor in the hypervirulence of NAP1 strains.
The recent work of Lyras et al. [43] found that TcdB is critical to C. difficile virulence in a hamster model of CDAD. Thus, variations in the antigenic region (e.g. carboxy terminus) of TcdB could allow repeated C. difficile infections of the same host by strains with antigenic variants of this toxin. In a recent publication by He and colleagues it was estimated that C. difficile diverged into a distinct species between 1.1 and 85 million years ago, and has gone through remarkable genetic variation over time [49]. The authors also posited that immune selection could have influenced the genetic variation, and they examined candidate immunogenic proteins that might fit this profile and 12 such proteins were identified. TcdB was not among these candidate proteins. It is unclear whether TcdB fits the criteria established for a positively selected core gene of C. difficile in this study, but it is reasonable to suspect the gene may have varied to avoid immune responses and this hypervariability enriched for a more potent form of the toxin. It is worth noting that while the protein identity was around 92%, the DNA homology was 93%. Nearly all of the residue changes occur as a single nucleotide substitution that result in amino acid substitutions. This further suggests a possible change in the sequence of TcdB that has been selected through an enhancement in virulence and perhaps by immune evasion.
Chinese hamster ovary-K1 (CHO) cells were maintained in F-12K medium (American Tissue and Culture Collection; ATCC) along with 10% fetal bovine serum (ATCC). HL-1 cardiomyocytes were obtained from the Claycomb laboratory [38] and maintained in Claycomb medium (Sigma) supplemented with 10% fetal bovine serum (ATCC), 0.1 mM Norepinephrine (Sigma), and 2 mM L-glutamine (Invitrogen). Cultures were grown at 37°C in the presence of 6% CO2. C. difficile VPI 10463 (produces TcdB with identical sequence to the 630 strain) and C. difficile BI17 6493 (a gift from Dr. Dale Gerding), were used in this study for the purification of TcdBHIST and TcdBHV. The tcdB gene was sequenced from both of these strains and the sequence was confirmed as exact matches to Genbank deposited sequences of strain 630 and R20291 (Genbank numbers AM180355 and FN545816). Cultures were grown as previously described [41], and TcdB was isolated by consecutive steps of anion-exchange (Q-Sepharose) and high-resolution anion-exchange (Mono-Q) chromatography in 20 mM Tris-HCl, 20 mM CaCl2, pH 8.0. Purification steps were followed by protein determination using the Bradford method, visualization of a single band by SDS-PAGE, and LC/MS/MS analysis (University of Oklahoma Health Science Center) to confirm protein identity. Cytotoxicity was determined using a WST-8 [2-(2-methoxy-4-nitrophenyl)-3-(4-nitrophenyl)-5-(2,4-disulfophenyl)-2H-tetrazolium, monosodiumsalt] (Dojindo Laboratories) according to manufacturer's instructions.
Zebrafish maintenance and experiments were performed in accordance with the PHS Principles for The Utilization and Care of Vertebrate Animals Used in Testing, Research, and Training, and followed the recommendations in the Guide for the Care and Use of Laboratory Animals under the approval of The University of Oklahoma Health Sciences Center Campus IACUC (OUHSC #06-126). Zebrafish were obtained from Aquatic Eco-System (Apopka, FL). Zebrafish were maintained at 28.5°C on a 14 h light/10 h dark cycle in 10 gallon tanks equipped with pumps for mechanical and chemical filtration. Matings were performed in false bottom tanks, and embryos were washed briefly with 0.5% bleach after collection. Embryos were incubated in embryo water (60 mM NaCl, 1.2 mM NaHCO3, 0.9 mM CaCl2, 0.7 mM KCl) in petri dishes at 28.5°C, and water was changed daily. For TcdB treatment experiments, embryos were used between 48 and 72 h post fertilization, with chorions removed. Embryos were placed (5 embryos per well) into 48-well plates and treated with TcdBHIST or TcdBHV in embryo water at concentrations ranging from 50 nM to 0.01 nM. The embryos were observed for 72 h after treatment for morphological changes by using a SZX-7 microscope with a DP70 camera (Olympus). All images were captured and processed by using DP controller and DP manager software (Olympus).
The region encoding the enzymatic domain of TcdBHV was amplified from C. difficile NAP1 genomic DNA by PCR using the forward primer 5′-ACGTCCCGGGATGAGTTTAGTTAATA-3′ and the reverse primer 5′-ACTGGATCCTCATTATACTGTATTTTG-3′ to generate the tcdB gene fragment encoding residues 1 to 1668 of tcdB (tcdB1–1668) with a 3′ XmaI/SmaI and a 5′ BamHI site. The restricted gene fragment was fused to lfn by overnight ligation at 16°C with a Xma1/BamHI-restricted pET15b derivative containing lfn. The resulting plasmid was cloned into Escherichia coli XL-1 blue (Novagen) and candidate clones were screened for the correct insert and orientation by restriction analysis and DNA sequencing. LFnTcdBHIST(enz) which had been previously cloned and described [40] and the newly synthesized LFnTcdBHV(enz) were expressed using E. coli BL-21 Star (Invitrogen). Both fusions were purified by Ni2+ affinity chromatography (His-Trap, GE Life Sciences) and the purified protein migrated within the predicted size range of ∼94 kDa on SDS-PAGE. Protective antigen was expressed and purified as previously described [50].
TcdBHIST or TcdBHV were labeled with Alexa Fluor 647 C5 maleimide (Invitrogen) according to manufacturer's instructions. Briefly, a 10 M excess of dye was added to TcdB in 20 mM Tris-HCl, pH 8.0, and incubated overnight at 4°C. Conjugated protein was separated from unincorporated dye using Sephadex G-25, and efficiency of labeling was confirmed to be between 80% and 100%. The activity of labeled TcdB was confirmed by cytotoxicity on CHO and HL-1 cells and was not reduced by >10%. Binding of each toxin to CHO and HL-1 cells was examined as follows. Cells were dissociated from flasks using 1 mM EDTA in PBS, centrifuged at 500× g, and washed once with PBS. One hundred thousand cells were incubated with a range from 10 nM to 320 nM of labeled toxin in 1 mL of PBS on ice for 1 h, washed twice, and the pellets were resuspended in 1 mL of PBS. The samples were analyzed using a FACSCalibur flow cytometer (University of Oklahoma Health Sciences Center) and FLOWJO software (Tree Star, San Carlos, CA). The emission wavelength was set to 665 nm, and the excitation was set at 633 nm with a bandpass of 30 nm.
CHO cells were plated at 5×104 cells/well in a 96-well plate and incubated overnight. The following day, TcdBHIST or TcdBHV was added to the cells at a final concentration of 0.1 µg/mL. At the indicated time points, the cells were washed to remove unbound toxin, and ammonium chloride (Sigma) was added to the cells at final concentration of 100 mM. Each sample was monitored for 24 h, and CPE (cytopathic effect) was determined by visualization. Percent CPE was calculated by counting a minimum of 100 cells in 3 different fields for each sample. Cells scored positive for CPE only when fully rounded, and the percent CPE was calculated as % rounded cellstest - % rounded cellscontrol, where control refers to cells treated with media alone.
2-(p-Toluidinyl) naphthalene-6-sulfonic acid, sodium salt (TNS; Invitrogen) solutions were prepared to a final concentration of 150 µM in pH specific buffers. For pHs ranging from 4.0 to 6.0, 100 mM NaCl-100 mM ammonium acetate-1 mM EDTA was used. For pH 6.0 to 7.0, 100 mM NaCl-100 mM MES-1 mM EDTA was used. For pH 7.0 to 8.0, 100 mM NaCl- 100 mM HEPES-1mM EDTA was used. 40 pmol of TcdBHIST or TcdBHV was added to the buffer/TNS mixture in a final volume of 2.5 mL and allowed to incubate for 20 min and 37°C. Each sample was analyzed on a Fluorolog R928P PMT fluorometer (HORIBA Jobin Yvon) with an excitation of 366 nm and an emission scan of 380 to 500 nm with a slit width of 2.0. Tryptophan fluorescence of TcdBHIST and TcdBHV was also compared in the same manner, using an excitation of 270 nm and an emission scan of 310 nm to 400 nm.
Data are expressed as the means ± S.E.M. Statistical analyses were performed using two-tailed unpaired Student's t-test in GraphPad Prism (La Jolla, CA). Statistical significance is indicated as * p<0.05; ** p<0.01; *** p<0.001.
|
10.1371/journal.pntd.0000446 | Inhibition of Lassa Virus Glycoprotein Cleavage and Multicycle Replication by Site 1 Protease-Adapted α1-Antitrypsin Variants | Proteolytic processing of the Lassa virus envelope glycoprotein precursor GP-C by the host proprotein convertase site 1 protease (S1P) is a prerequisite for the incorporation of the subunits GP-1 and GP-2 into viral particles and, hence, essential for infectivity and virus spread. Therefore, we tested in this study the concept of using S1P as a target to block efficient virus replication.
We demonstrate that stable cell lines inducibly expressing S1P-adapted α1-antitrypsin variants inhibit the proteolytic maturation of GP-C. Introduction of the S1P recognition motifs RRIL and RRLL into the reactive center loop of α1-antitrypsin resulted in abrogation of GP-C processing by endogenous S1P to a similar level observed in S1P-deficient cells. Moreover, S1P-specific α1-antitrypsins significantly inhibited replication and spread of a replication-competent recombinant vesicular stomatitis virus expressing the Lassa virus glycoprotein GP as well as authentic Lassa virus. Inhibition of viral replication correlated with the ability of the different α1-antitrypsin variants to inhibit the processing of the Lassa virus glycoprotein precursor.
Our data suggest that glycoprotein cleavage by S1P is a promising target for the development of novel anti-arenaviral strategies.
| The virus family Arenaviridae includes several hemorrhagic fever causing agents such as Lassa, Guanarito, Junin, Machupo, and Sabia virus that pose a major public health concern to the human population in West African and South American countries. Current treatment options to control fatal outcome of disease are limited to the ribonucleoside analogue ribavirin, although its use has some significant limitations. The lack of effective treatment alternatives emphasizes the need for novel antiviral therapeutics to counteract these life-threatening infections. Maturation cleavage of the viral envelope glycoprotein by the host cell proprotein convertase site 1 protease (S1P) is critical for infectious virion production of several pathogenic arenaviruses. This finding makes this protease an attractive target for the development of novel anti-arenaviral therapeutics. We demonstrate here that highly selective S1P-adapted α1-antitrypsins have the potential to efficiently inhibit glycoprotein processing, which resulted in reduced Lassa virus replication. Our findings suggest that S1P should be considered as an antiviral target and that further optimization of modified α1-antitrypsins could lead to potent and specific S1P inhibitors with the potential for treatment of certain viral hemorrhagic fevers.
| Lassa virus (LASV) belongs to the family Arenaviridae, which are enveloped, single-stranded RNA viruses distributed worldwide. Based on their antigenic relationships and geographic distribution, arenaviruses are divided into two major groups. The Old World group includes the prototype of this family, lymphocytic choriomeningitis virus (LCMV), and LASV, which is endemic in West African countries and causes every year thousands of human infections with hemorrhagic fever as a severe clinical manifestation [1]. The New World group includes among others Machupo, Junin, Guanarito and Sabia viruses which can cause viral hemorrhagic fever (VHF). With the exception of the New World virus Tacaribe, which was isolated from Artibeus bats, arenaviruses are rodent-borne viruses [2].
Over the past few years great efforts have been made to find potential therapeutic and vaccination approaches in the arenavirus field (reviewed in [3],[4],[5]). Until now there is no specific and effective treatment available to combat hemorrhagic fevers caused by arenaviruses. Administration of convalescent plasma has been reported to reduce the mortality rates of patients with Argentine hemorrhagic fever, however, 10% of immune-plasma recipients developed a late neurological syndrome of unknown origin [6]. The only existing drug used to treat Lassa fever and certain South American hemorrhagic fevers is the broad-spectrum antiviral agent ribavirin, a ribonucleoside analogue, which has shown to be partially effective if given early in the course of illness [7],[8],[9],[10]. Even though the drug is relatively inexpensive for patients in high-developed countries, it is still unaffordable for many of those living in West Africa and South America. Moreover, several adverse effects have been associated with ribavirin therapy in patient studies and animal models [11],[12],[13],[14],[15]. The lack of effective disease control measures as well as the discovery of new fatal arenavirus species that pose a risk of epidemic potential [16],[17], emphasize the need for novel therapeutic interventions.
Lassa virions are pleomorphic lipid-enveloped particles that contain two single-stranded RNA segments, designated L (large) and S (small), encoding four viral proteins in a unique ambisense coding strategy. The L segment encodes the viral RNA-dependent RNA polymerase (L) and the small zinc finger matrix protein (Z) [18]; the S segment encodes the virus nucleoprotein (NP) and the virus surface glycoprotein precursor (preGP-C) [19]. preGP-C is cleaved co-translationally into a stable signal peptide and GP-C [20]. Post-translational maturation cleavage of GP-C by the proprotein convertase site 1 protease (S1P, [21]), also known as subtilisin kexin isozyme-1 (SKI-1, [22]), leads then to the generation of the distal receptor-binding subunit GP-1 and the transmembrane-spanning fusion competent subunit GP-2 [23]. Together with the signal peptide these subunits form the tripartite glycoprotein spike complex on the viral surface [24],[25].
The glycoproteins of the Old World arenaviruses LASV and LCMV were the first viral glycoproteins that were shown to be proteolytically processed by S1P [23],[26], which normally plays important physiological regulatory roles in cholesterol metabolism, ER stress response, cartilage development and other cellular processes [21],[27],[28],[29],[30],[31]. Using systematic mutational analysis of the LCMV GP cleavage site, the consensus motif R-(R/K/H)-L-(A/L/S/T/F) was determined, which is conserved in the glycoprotein sequences of the Old World viruses LASV, Mopeia and Mobala, as well as the New World virus Pichinde, suggesting that all arenavirus glycoproteins are cleaved by S1P [26],[32]. Indeed, more recently Rojek et al. reported that glycoproteins from the New World hemorrhagic fever viruses Junin, Machupo and Guanarito are also processed by S1P, although Guanarito possesses a protease recognition motif that differs from known arenavirus GP consensus cleavage sequences, indicating a broader substrate specificity of S1P than previously anticipated [33].
Proteolytic activation of LASV GP-C by S1P is not necessary for transport of GP-C to the cell surface, where budding of arenaviruses occurs, but is essential for incorporation of the cleaved subunits into virions, and thus, for the formation of infectious viral particles. In the absence of GP-C cleavage, enveloped non-infectious LASV-like particles are released containing L, NP, Z protein and viral RNA but are devoid of viral glycoproteins [23]. Similar results were described for LCMV and New World hemorrhagic fever viruses [33],[34].
In addition to its important role in the arenaviral life cycle, S1P is critical for the infectivity of Crimean-Congo hemorrhagic fever virus (CCHFV), a member of the Bunyaviridae family, through processing of the glycoprotein Gn [35],[36]. These findings make the inhibition of S1P particularly interesting for the development of a novel antiviral therapeutic that will target pathogenic viruses known to be processed by S1P.
A successful approach to inhibit proprotein convertases involves genetically engineered antitrypsins, which are derived from α1-antitrypsin (α1-AT). α1-AT is a serine protease inhibitor (serpin) with a characteristic exposed reactive center loop (RCL), which mediates binding to the active site of its target protease. The exploration for the potential use of modified antitrypsins with an altered inhibitory spectrum has been guided by the discovery of a natural variant of α1-AT, known as Pittsburgh (α1-AT-PIT), found in a patient who had a severe bleeding disorder caused by mutation of the P1 reactive center residue of antitrypsin from methionine to arginine [37]. This substitution changed its specificity from elastase to thrombin and other coagulation proteases. Due to the introduction of a second mutation from alanine to arginine at P4 of the RCL, the engineered α1-antitrypsin variant Portland (α1-AT-PDX) showed high affinity for furin [38]. α1-AT-PDX efficiently inhibited the formation of infectious HIV, measles virus, and human cytomegalovirus progeny by blocking furin-dependent processing of glycoproteins gp160, F0 and gB, respectively [38],[39],[40],[41]. Pullikotil and co-workers used this approach for the generation of highly selective α1-antitrypsin variants specific for S1P by introducing various S1P recognition motifs into the RCL of α1-antitrypsin [42]. The adaptation of α1-antitrypsin towards S1P efficiently inhibited the processing of the S1P substrates SREBP-2 (sterol regulatory element binding protein), ATF6 (activating transcription factor 6) as well as CCHFV glycoprotein [42]. However, the effect of these inhibitors on CCHFV infection was not analyzed in that study. To block cleavage of the LASV glycoprotein, we generated here recombinant α1-antitrypsin variants mimicking the S1P recognition motifs RRIL, RRVL and RRYL that exhibited the greatest inhibitory potential based on immunoblot quantification. In addition, we used an α1-AT construct that contains the LASV GP cleavage motif RRLL in its RCL. Using a doxycycline regulated expression system we demonstrate that S1P-adapted α1-antitrypsin variants efficiently block proteolytic maturation of the glycoprotein precursor GP-C, whereas a furin-specific α1-AT had no effect on GP-C processing. Virus replication of both a replication-competent recombinant vesicular stomatitis virus expressing the LASV glycoprotein GP-C (VSVΔG/LASVGP) and authentic LASV was significantly inhibited in the presence of S1P-specific α1-antitrypsins. The degree of inhibition of viral replication correlated with the ability of the different α1-antitrypsin variants to inhibit the processing of LASV GP-C.
Since glycoprotein processing by the endoprotease S1P is not only critical for virus infectivity of LASV [23], and other arenaviruses causing hemorrhagic fever [33], but also for members of the Bunyaviridae family [36], further optimization based on our findings could lead to a potent and specific S1P inhibitor with the potential treatment of certain VHFs.
cDNA of the open reading frame of rat α1-antitrypsin (Gene Bank Accession Number NM_022519) (a kind gift from Dr. G. Thomas, Vollum Institute, Oregon Health & Science University, Portland, USA) was inserted into pSG5 and used as a template to generate S1P-specific α1-antitrypsin variants by recombinant polymerase chain reaction (PCR) using overlapping oligonucleotides [43]. The sequences of the oligonucleotides used are listed in Table S1. The resulting full-length PCR products were digested with BamHI and NheI and cloned into the tetracycline (Tet)-controlled inducible mammalian expression vector pTRE2hyg (Clontech). The accuracy of all constructs was confirmed by DNA sequencing.
To generate stably expressing cell lines, Chinese hamster ovary (CHO)-K1 Tet-On cells (Clontech) were transfected with pTRE2hyg containing the α1-antitrypsin constructs using Lipofectamine 2000 (Invitrogen) according to manufacturer's instructions. Cells were then cultured for 2 weeks under selective pressure in the presence of 500 µg/ml Hygromycin B, the selection agent for the α1-antitrypsin expressing plasmid, and 500 µg/ml G418, the selection agent for the rtTA (reverse Tet-controlled transactivator) cassette. The selective media were replaced every 3 days. Well-separated antibiotic-resistant cell clones were individually isolated with cloning cylinders (Sigma). Therefore, a small volume of Trypsin-EDTA (Sigma) was added and the culture dish was incubated briefly at 37°C until cells detach. Cells were then collected from inside the cylinder and transferred to individual wells of a 24-well plate for further growth in selective medium. When grown to confluence, cells were transferred to larger flasks. Protein expression was induced with 1 µg/ml doxycycline (Clontech) and analyzed by Western Blot and immunofluorescence. Stable cell lines showing similar expression levels of the various α1-antitrypsins were chosen for further experiments.
Vero E6 cells (green monkey kidney) were cultured in Dulbecco's modified Eagle medium (DMEM, Gibco) and CHO-K1 Tet-On cells in DMEM/F12 (Gibco), both media containing penicillin (100 U/ml), streptomycin (100 µg/ml), and L-glutamine (2 mmol/l) (all from Invitrogen) as well as 10% fetal bovine serum (PAN Biotech). S1P-deficient SRD-12B cells (a generous gift from Dr. J. L. Goldstein, Department of Molecular Genetics, University of Texas Southwestern Medical Center, Dallas, USA) were maintained as CHO cells but supplemented with 5 µg/ml of cholesterol (Sigma), 1 mM sodium mevalonate (Sigma), and 20 µM sodium oleate (Sigma) [44].
The vesicular stomatitis virus reverse genetics system (VSV, Indiana serotype) was kindly provided by Dr. J.K. Rose (Department of Pathology, Yale University School of Medicine, New Haven, USA) and was described in detail earlier [45],[46],[47]. Recombinant VSV expressing the glycoprotein GP-C of Lassa virus (LASV, strain Josiah) designated as VSVΔG/LASVGP and wild-type VSV (VSVwt) were propagated in Vero E6 cells as described previously [48]. Influenza virus A/FPV/Rostock/34 (H7N1), designated as fowl plague virus (FPV), was propagated in embryonated hen eggs and stored at −80°C until further use. Virus titration of FPV was described previously [49]. All experiments with infectious FPV were done under biological safety level 3 conditions. VSVΔG/LASVGP titration was performed using a microplate format plaque assay with subsequent immunostaining as described before [50]. In brief, virus dilutions were incubated on Vero E6 cells with an overlay of 3% carboxymethylcellulose (CMC) during plaque formation. Infected cells were visualized after cell fixation with paraformaldehyde (PFA, 4%) and permeabilization with 0.3% Triton-X 100 using a specific LASV GP-C/GP-2 antibody followed by incubation with horseradish peroxidase-labeled secondary anti-rabbit antibody (DAKO). Finally, cells were stained with True Blue Peroxidase substrate (KPL).
For virus spread experiments, CHO cell lines were seeded into 96-well plates in the presence or absence of doxycycline. 24 h after induction, cells were infected with VSVΔG/LASVGP or FPV and were grown without solid overlay. Cells were fixed at different time points post-infection and immunostaining was performed as described above using rabbit sera against VSV (kindly provided by Dr. G. Herrler, Institut für Virologie, Zentrum für Infektionsmedizin, Stiftung Tierärztliche Hochschule Hannover, Germany), for the detection of VSVΔG/LASVGP infected cells, and against FPV, for cells infected with FPV, respectively.
Virus titration of LASV (strain Josiah, Gene Bank Accession Number NC_004297 and NC_004296) was performed by defining the 50% tissue culture infectious dose (TCID50). For this, Vero cells were grown in 96-well plates to 30 to 40% confluence. Cells were inoculated with 10-fold serial dilutions of supernatants from LASV-infected CHO cell lines grown in the presence or absence of doxycycline. The assays were evaluated at 7 to 9 days post-infection. TCID50 values were calculated using the Spearman-Karber method [51]. All experiments involving LASV-infected samples were performed under biological safety level 4 conditions at the Philipps-University Marburg.
At 24 h post-infection, cell culture supernatants from infected cells were cleared from cell debris and pelleted in an SW-60 rotor through a 20% sucrose cushion at 52000 rpm at 4°C for 2 h. The pellet was then resuspended in PBS buffer and mixed with SDS-PAGE sample buffer. To control the intracellular expression level, cell lysates were collected simultaneously. Samples were analyzed by SDS-PAGE and Western blotting using protein-specific antibodies as indicated.
Proteins were separated by SDS-PAGE using 10% polyacrylamide gels. Immunoblotting was performed as described previously [52]. Antiserum against Lassa virus GP-C/GP-2 was also described previously [32]. Polyclonal rabbit anti-ß-tubulin antibody was purchased from Abcam (UK), and monoclonal mouse anti-Flag antibody from Sigma-Aldrich. Secondary antibodies labeled with Alexa680 or IRDye800 were from Molecular Probes Invitrogen and Biomol, respectively, and were used for visualization and quantification of detected proteins using the Odyssey Infrared Imaging System (LI-COR Biosciences).
CHO cell lines were grown on coverslips and 24 h after doxycycline-induction, cells were washed with PBS and fixed with 4% PFA in DMEM for 30 min. The fixative was removed, and free aldehydes were quenched with 100 mM glycine in PBS. Then, samples were washed with PBS and permeabilized for 10 min with PBS containing 0.1% Triton X-100. Cells were incubated in blocking solution (2% bovine serum albumin, 0.2% Tween 20, 5% glycerol, and 0.05% sodium azide in PBS) and subsequently stained with a primary mouse-anti-flag antibody (1∶400) and a secondary anti-mouse antibody coupled to rhodamine (1∶200, Jackson Immunoresearch). Cell nuclei were stained with DAPI (4′,6′-diamidino-2-phenylindole, Sigma). Microscopic analysis was performed with a Zeiss ApoTome/Axiovert 200 M microscope using a magnification of 1∶40.
Replication-competent recombinant vesicular stomatitis virus (rVSV) expressing foreign envelope glycoproteins has been demonstrated to be a suitable model system to study the role of viral glycoproteins in the context of virus replication [47],[53],[54]. In the present study, we took advantage of a rVSV expressing the LASV glycoprotein GP (designated VSVΔG/LASVGP) [48]. In this system biosynthesis and processing of GP was shown to be authentic compared to LASV [48].
In an initial experiment we wanted to determine whether CHO-K1 cells are susceptible to VSVΔG/LASVGP infection. The reason we chose CHO-K1 cells for our studies is the availability of a site 1 protease-deficient CHO cell line (designated SRD-12B cells), in which GP maturation is abolished and only GP-deficient non-infectious LASV particles are released [23]. Thus, this cell clone provides an ideal control for inhibition studies. Vero E6, CHO-K1, and SRD-12B cells were infected with either VSVΔG/LASVGP or wild-type VSV (VSVwt) as a control. Aliquots of cell culture supernatants were collected at different times after infection and were analyzed by plaque assay. Growth kinetics revealed that VSVΔG/LASVGP grows to similar titers in CHO-K1 cells compared to Vero E6 cells which have been used in earlier studies (Fig. 1A) [48]. These data demonstrated that CHO-K1 cells support efficient VSVΔG/LASVGP replication, and thus are useful tools for further investigations. As expected, VSVΔG/LASVGP lacks efficient replication in SRD-12B cells, whereas virus growth of VSVwt remained unaffected in these cells (Fig. 1A). The reason for the low but detectable virus titers in the supernatant of VSVΔG/LASVGP-infected SRD-12B cells is currently not known but has been also observed for LASV ([23] and present study), LCMV [34] and New Word arenaviruses [33]. Glycoprotein activation by a yet unknown protease though with only very low efficiency might explain this phenomenon.
To mimic the conditions of short-term treatment, we decided to use the inducible doxycycline-dependent Tet-On expression system, which allows regulated expression of the protein of interest [55]. To determine whether treatment of cells with doxycycline interferes with viral replication, we cultivated VSVΔG/LASVGP-infected CHO-K1 Tet-On cells in the presence or absence of doxycycline (1 µg/ml) for 24 h and 48 h, respectively. As shown in Fig. 1B, CHO-K1 Tet-On cells treated with doxycycline produced a virus titer comparable to cells that were cultivated in the absence of doxycycline, indicating that these conditions used in our experiments have no influence on efficient virus replication.
Pullikotil and colleagues recently reported that various antitrypsins mimicking S1P recognition motifs are able to block processing of the S1P substrates SREBP and ATF6, although to different degrees [42]. In addition to the α1-AT variants shown to be most effective in that study we have chosen the LASV GP-C cleavage motif RRLL to investigate whether they also inhibit LASV GP-C cleavage. Therefore, we generated various S1P-specific α1-ATs, and as a specificity control, a furin-adapted α1-AT, by recombinant PCR technology using the rat α1-AT-PIT as a template (Fig. 2A). To facilitate their detection, we introduced a flag epitope at the C-termini of the constructs. Stable cell lines were generated and individual clones were isolated and screened for α1-antitrypsin expression after doxycycline induction by immunoblotting and immunofluorescence analysis. Cell lines that showed similar expression levels of α1-antitrypsins were chosen for further experiments (Fig. 2B and 2C).
To test the inhibitory potential of S1P-specific α1-antitrypsins on proteolytic processing of LASV GP, stably transfected CHO-K1 Tet-On cells, and non-transfected wild-type CHO-K1 Tet-On cells as well as SRD-12B cells were infected with VSVΔG/LASVGP at an MOI of 0.2 in the presence or absence of doxycycline. To allow only one replication cycle, cell lysates were analyzed 10 h post-infection for detection of LASV GP cleavage by Western blot analysis using a GP-specific antiserum that recognizes both the precursor GP-C and the cleaved subunit GP-2. In CHO-K1 Tet-On cells LASV GP was efficiently cleaved, regardless of whether doxycycline was present or not. In contrast, virtually no detectable cleavage of GP was observed in SRD-12B cells that are deficient in S1P (Fig. 3A, lanes 1–4). Without expression of the various antitrypsins efficient cleavage was detected in these stably transfected cell lines, similar to the processing of GP in wild-type CHO-K1 Tet-On cells (Fig. 3A, lanes 1, 5, 7, 9, 11, and 13). In contrast, cells expressing the S1P-adapted α1-antitrypsins inhibited proteolytic maturation of LASV GP (Fig. 3A, lanes 6, 8, 10, and 12). Furthermore, our results show that the presence of a furin-specific α1-AT did not influence LASV GP-C processing, demonstrating the specificity of the generated S1P-adapted α1-antitrypsins (Fig. 3A, lanes 13 and 14). Quantification of GP-2 cleavage revealed that the α1-AT variant RRIL exhibited the greatest inhibitory effect on GP processing (>80% inhibition) followed by α1-AT RRLL (>60% inhibition), which possesses the amino acid cleavage motif of the LASV GP-C. Also α1-AT variants RRVL and RRYL were found to be inhibitory, although to a lesser extent (inhibition less than 50%) than the variants RRIL and RRLL (Fig. 3B). Taken together, these data clearly demonstrate that S1P-specific α1-antitrypsins efficiently block the maturation cleavage of LASV GP, however, they differ in regard to their inhibitory potential.
We have shown earlier that S1P-mediated cleavage of GP-C is absolutely required for incorporation of the glycoprotein subunits into the virion envelope and thus for production of infectious LASV [23]. Therefore, we addressed the question of whether a S1P-specific α1-AT has the potential to prevent GP incorporation by blocking glycoprotein processing. To this end, α1-AT RRIL cells were infected in the presence or absence of doxycycline with either VSVΔG/LASVGP or VSVwt as a control. At 24 h post-infection, viral particles released into the cell culture supernatant were purified over a 20% sucrose cushion and analyzed by means of immunoblotting. In viral particles from supernatants of non-induced α1-AT RRIL cells and CHO-K1 Tet-On control cells cleaved GP-2 was readily observed, whereas in the particulate material isolated from the supernatant of α1-AT RRIL expressing cells no glycoprotein was detected (Fig. 4A). However, Western Blot analysis for VSV proteins revealed the release of these viral proteins into the supernatant of α1-AT RRIL expressing cells, which is consistent with our earlier findings that, in the absence of GP-C cleavage, enveloped non-infectious LASV-like particles containing the matrix protein Z and the ribonucleoprotein (RNP) complex, but devoid of viral glycoproteins, are still released [23]. The lower amount of VSV proteins observed in the cell lysate and supernatant of α1-AT RRIL expressing cells reflect lower levels of viral replication, which is due to less efficient virus spread (Fig. 4A). In contrast to its ability to efficiently block incorporation of LASV GP into virions, the presence of α1-AT RRIL had no effect on the release of glycoprotein G containing wild-type VSV particles. The amount of VSV proteins detected in the supernatant from α1-AT RRIL expressing cells was similar to the amount of viral proteins observed in supernatants of non-induced cells and CHO-K1 cells, indicating efficient viral replication and cell-to-cell spread of VSVwt despite the presence of α1-AT RRIL (Fig. 4B). Taken together, these data demonstrate that S1P-specific α1-antitrypsins have the potential to prevent LASV GP incorporation by inhibiting glycoprotein cleavage, which is an essential prerequisite for infectious progeny.
Next, we wanted to know whether the observed inhibition of LASV GP processing correlates with the ability of the different α1-antitrypsin variants to inhibit virus spread. To investigate this, we established a 96-well plate assay in which infected cells are immunostained with True Blue substrate as described in Materials and Methods. Virus spread can be monitored by the appearance of characteristic comet-shaped foci, showing that the virus progeny is carried over the cell monolayer, while prevention of virus spread results in limited radial growth, due to infection of only neighbouring cells. This assay allows rapid screening of potential inhibitors [50]. To this end, doxycycline-induced and non-induced CHO cell lines, as well as CHO-K1 Tet-On cells and SRD-12B cells, were infected with VSVΔG/LASVGP. At 24 h post-infection, cells were fixed and immunostained. Under non-induced conditions efficient virus spread was observed in all CHO-K1 Tet-On α1-AT variant cell lines as well as in CHO-K1 Tet-On wild-type cells (Fig. 5A, upper panel). In contrast, virus spread was significantly diminished in cells expressing α1-AT specific for S1P (Fig. 5A, lower panel). These data indicate that S1P-adapted α1-antitrypsins have the potential to specifically inhibit the processing of LASV GP, which in turn is required for efficient virus spread. It should be noted that the infectious foci observed in α1-AT RRIL expressing cells were larger compared to SRD-12B cells in which virtually no virus spread of VSVΔG/LASVGP was observed, resulting in only single infected cells (Fig. 5A). Although similar inhibition values were observed by means of immunoblot quantification (Fig. 3), a few remaining non-detectable cleavage events may count for this limited cell-to-cell spread in α1-AT RRIL expressing cells. Cells expressing the furin-adapted α1-AT variant RVKR did not prevent virus spread. At first glance, we rather observed an enhancement of infectivity compared to non-induced cells, which might be due to an increase in the LASV cellular receptor α-dystroglycan on the cell surface [56].
To further confirm the specificity of the α1-AT variants, we used fowl plague virus (FPV), which contains a hemagglutinin with a multibasic cleavage motif recognized by furin [57]. Thus, the furin-adapted α1-AT should prevent virus spread of FPV, while virus spread in the presence of S1P-specific α1-antitrypsins should remain unaffected. Fig. 5B clearly demonstrates that the most potent S1P-specific α1-AT variant RRIL had no effect on FPV replication, and that virus spread was found to be similar to that observed in wild type CHO-K1 Tet-On cells. In contrast, in cells expressing the furin-adapted α1-AT variant RVKR virus spread of FPV was drastically reduced, whereas FPV replication occurred efficiently under doxycycline-free conditions in these cells. These results demonstrate that the generated α1-AT variants exhibit high specificity for the corresponding proteases, which are essential for virus spread in cell culture.
To further elucidate the effect of the different α1-AT variants on multicycle replication, viral titers were determined. To this end, cells were infected with VSVΔG/LASVGP at an MOI of 0.02 in the presence or absence of doxycycline. Cell culture supernatants were collected 24 h and 48 h post-infection and virus titers were determined by plaque assay. As shown in Table 1, non-induced S1P-specific α1-AT cell lines permitted unaffected growth of VSVΔG/LASVGP to comparable titers, whereas virus titers were reduced in cells expressing the S1P-specific α1-AT variant. At 24 h post-infection virus production decreased about 100 fold in cells expressing the α1-AT variant RRIL compared to non-induced control cultures. The presence of α1-AT variant RRLL reduced the virus titer in the supernatant about 10 fold, followed by a 6.2 fold reduction of virus production in α1-AT variant RRVL expressing cells. The presence of the α1-AT variant RRYL only exhibited a very moderate inhibitory effect on viral replication (inhibition <2 fold). Again, the presence of the furin-adapted α1-AT variant RVKR did not affect VSVΔG/LASVGP replication compared to non-induced control cells. Our results indicate that the various S1P-adapted α1-antitrypsins exhibit different inhibitory potentials, due to their different recognition motifs. However, the degree of inhibition of virus replication correlated well with the inhibitory potential of the various S1P-adapted α1-antitrypsin variants to block LASV GP processing. Interestingly, following the inhibition of virus progeny over a time period of 48 h only the S1P-adapted α1-AT variants RRIL and RRLL sustained their inhibitory capacity, whereas in cells expressing α1-antitrypsin variants RRVL and RRYL virus production was found to recover although the initial expression levels of α1-antitrypsin variants were similar (Table 1). These data indicate that the inhibitory potential of the α1-AT variants RRVL and RRYL is not sufficient to efficiently suppress the formation of infectious particles by effectively blocking LASV GP-C cleavage, whereas the α1-AT variants RRIL and RRLL seem to be appropriate candidates for efficient inhibition of LASV propagation.
Finally, we wanted to investigate the impact of blocking S1P-mediated GP processing on virus progeny of authentic LASV. Therefore, we assessed the inhibitory potential of the most potent variant, α1-AT RRIL, on the multiplication of LASV (strain Josiah). For this purpose α1-AT RRIL cells and, as controls, CHO-K1-Tet-On and SRD-12B cells were infected with LASV at an MOI of 0.1. To induce α1-AT expression, α1-AT RRIL cells and, as a control for off-target effects, CHO-K1 Tet-On cells were cultivated in the presence of doxycycline. To determine virus titers, infectious virions released into the cell culture supernatant were analyzed by defining the 50% tissue culture infectious dose (TCID50) at various times post-infection, as indicated. In non-induced α1-AT RRIL cells, LASV revealed a growth kinetic similar to that observed in CHO-K1 Tet-On control cultures, while expression of α1-AT RRIL resulted in an average 2 log10 reduction in viral titer (Fig. 6). The difference between infectious LASV titers in the supernatant of α1-AT RRIL expressing cells and SRD-12B cells correlated with the limited virus spread observed in α1-AT RRIL expressing cells compared to single cell infections in S1P null cells (Fig. 5). Taken together, this result highlights the inhibitory activity of modified α1-antitrypsins against LASV and demonstrates that inhibition of endogenous S1P is a potent strategy to reduce the production of infectious LASV progeny.
Current drug treatment of Lassa fever and certain New World hemorrhagic fevers is limited to the guanosine analogue ribavirin [7],[8],[9]. Although ribavirin therapy can reduce the mortality rates of severe clinical cases, its unavailability to most patients in West Africa and South America as well as its association with severe adverse effects including anaemia [11],[13], teratogenicity and embryo lethality [12], argues for the development of new alternative treatment options.
In principle, every step in the viral life cycle is a potential target for antiviral inhibitors. While current antiviral strategies in the arenavirus field mainly target virus entry [58],[59],[60],[61] or replication and assembly [62],[63],[64],[65],[66],[67], inhibition studies of the glycoprotein activating endoprotease and its impact on viral replication are largely unexploited. Due to its central role in the arenavirus life cycle [23],[26],[33],[34], S1P should be considered as a cellular target for antiviral drug development. In the present study we analyzed the inhibitory effect of S1P-adapted α1-antitrypsins on proteolytic processing of LASV GP-C and its consequences for viral replication. To our knowledge, this is the first report that addresses the impact of protein-based S1P inhibition on LASV GP-C cleavage and multicycle replication. Furin-adapted α1-ATs have been shown to efficiently inhibit the formation of infectious progeny of other viruses (e.g. HIV, measles virus and human cytomegalovirus) [38],[39],[40],[41],[68],[69].
Using a replication-competent recombinant VSV pseudotyped with the LASV glycoprotein GP [48], we demonstrate that proteolytic maturation of the precursor GP-C is sensitive to S1P-adapted α1-ATs. Mutagenesis of the reactive centre loop (RCL) into the S1P recognition motif RRIL resulted in an abrogation of GP-C processing similar to that observed in S1P-deficient SRD-12B cells. The inhibitory activity of the α1-AT variant RRIL on LASV GP cleavage described here is in agreement with a previous study showing its inhibitory potential on the processing of the natural S1P substrates SREBP-2 and ATF6 [42]. Also an α1-AT variant that contains the LASV GP-C cleavage motif RRLL exhibited a high S1P inhibitory potential and was found to drastically reduce GP processing. Interestingly, this variant exhibited a 100% inhibition activity on maturation cleavage of an artificial pro-PDGF (precursor of platelet-derived growth factor) mutant that is processed by S1P due to introduction of a RRLL cleavage site, but failed to inhibit cleavage of endogenously expressed SREBP-2 [42]. These data indicate that various substrates differ in their sensitivity towards S1P inhibition.
The outcome of severe illness increased significantly with the level of viremia in Lassa fever patients [70]. Therefore, the extent of multicycle replication of LASV and thus, the load of infectious particles in its host organism have an important impact for the progress of disease. Our studies revealed that α1-AT variants RRIL and RRLL have a potency sufficient to sustain their inhibitory capacity during multicycle replication, which resulted in a significant reduction of virus infectivity. Inhibition of viral replication correlated with the ability of the α1-AT variants RRIL and RRLL to efficiently inhibit the processing of the LASV glycoprotein precursor. Although our data demonstrated that inhibition of glycoprotein cleavage by α1-AT RRIL reduced incorporation of the subunits GP-1 and GP-2 into virions to below detectable levels, the viral titer from α1-AT RRIL expressing cells was found to be greater than that obtained from S1P null cells. Based on this observation, we consider that even the most potent α1-AT variant RRIL failed to entirely inhibit S1P activity. However, given that S1P has important biological functions in the regulation of various cellular processes, a complete inhibition of the catalytic activity of S1P is not desirable. For α1-AT variants RRVL and RRYL, we observed similar inhibition values by immunoblot quantification analysis as described for CCHFV GP cleavage [42]. Though, their inhibitory activity on LASV GP-C cleavage was not sufficient to efficiently reduce virus replication of VSVΔG/LASVGP. These results should be taken into consideration for experimental setups in future studies that address the impact of S1P inhibition in arenavirus replication.
The most potent α1-AT variant RRIL revealed a similar inhibitory potential on virus release of authentic LASV to that observed with the corresponding VSVΔG/LASVGP pseudotype. Therefore, this study also demonstrates that the replication-competent VSV expressing the LASV glycoprotein is an excellent surrogate model for analyzing potential antivirals that target the biological function of GP and its consequence for virus replication. These studies can be performed under biosafety level 2 laboratory conditions that would otherwise require biosafety level 4 laboratory conditions [71]. Taken together, our data indicate that S1P-adapted α1-antitrypsins may represent a promising lead compound for the development of a new class of anti-arenavirus inhibitors.
In recent years improvements were made in the application of bioengineered serpins to combat bacterial and viral infections [39],[72]. For example, the addition of exogenous α1-AT-PDX, a potent and selective furin inhibitor, was found to efficiently block human cytomegalovirus infection [39]. However, in contrast to furin, which is known to recycle between the plasma membrane and the TGN via endosomal compartments, membrane-bound S1P is localized in the secretory pathway and can be sorted to endosomal compartments but not to the cell surface [73],[74],[75]. Follow-up studies with small synthetic peptides, which are derived from S1P-specific α1-antitrypsins described in the present work, are currently in progress and will address cellular delivery and organelle specific targeting, as well as their inhibitory potential on authentic LASV replication. In analogy to inhibition strategies of the eukaryotic subtilase furin, we previously designed and developed a cell-permeable peptidyl chloromethylketone S1P inhibitor, which contained the LASV GP-C cleavage site, designated dec-RRLL-cmk [76],[77],[78]. This irreversible inhibitor efficiently blocked the processing of LASV GP at nanomolar concentrations, however, because of cell type-dependent toxicity observed by us and others, its potential in vitro use requires further investigation [79],[80].
Due to the essential roles of S1P in cholesterol metabolism and fatty acid synthesis, this enzyme has attracted great attention by the pharmaceutical industry. Research efforts are currently directed towards the development of S1P inhibitors that may be used in the treatment of dyslipidemia and a variety of cardiometabolic risk factors associated with diabetes and obesity [81]. Identification of specific S1P inhibitors in this therapeutic area may also be beneficial in treatment of hemorrhagic fevers caused by viruses known to be processed by S1P. Future studies will have to elucidate the anti-viral efficacy of these and other novel S1P inhibitors that have been developed [82],[83].
While most conventional antiviral drugs target proteins that are virus-encoded, cellular proteins essential for viral replication are currently considered to be alternative potential targets for antiviral therapy [84],[85],[86]. With the exception of Ebola virus, whose glycoprotein cleavage by the proprotein convertase furin is not essential for virus replication in cell culture and virulence in nonhuman primates [71],[87],[88],[89], maturation cleavage of surface glycoproteins of several virus species by endoproteases is a key determinant for host cell tropism and pathogenicity [90]. Thus, the emergence of viral escape mutants that confer resistance due to targeted inhibition of an endogenous protease is rather unlikely. In S1P-deficient SRD-12B cells, which were persistently infected with Junin virus vaccine strain Candid 1, no virus escape variants possessing a cleavage motif other than a S1P recognition motif have evolved, indicating a low potential of arenaviruses to develop de novo a different glycoprotein maturation pathway [33]. This observation together with our findings that inhibition of S1P significantly affects LASV GP processing and virus infectivity should encourage the development of S1P inhibitors as a potential drug target to counteract infections caused by pathogenic arenaviruses.
|
10.1371/journal.pbio.1001832 | Endocytic Crosstalk: Cavins, Caveolins, and Caveolae Regulate Clathrin-Independent Endocytosis | Several studies have suggested crosstalk between different clathrin-independent endocytic pathways. However, the molecular mechanisms and functional relevance of these interactions are unclear. Caveolins and cavins are crucial components of caveolae, specialized microdomains that also constitute an endocytic route. Here we show that specific caveolar proteins are independently acting negative regulators of clathrin-independent endocytosis. Cavin-1 and Cavin-3, but not Cavin-2 or Cavin-4, are potent inhibitors of the clathrin-independent carriers/GPI-AP enriched early endosomal compartment (CLIC/GEEC) endocytic pathway, in a process independent of caveola formation. Caveolin-1 (CAV1) and CAV3 also inhibit the CLIC/GEEC pathway upon over-expression. Expression of caveolar protein leads to reduction in formation of early CLIC/GEEC carriers, as detected by quantitative electron microscopy analysis. Furthermore, the CLIC/GEEC pathway is upregulated in cells lacking CAV1/Cavin-1 or with reduced expression of Cavin-1 and Cavin-3. Inhibition by caveolins can be mimicked by the isolated caveolin scaffolding domain and is associated with perturbed diffusion of lipid microdomain components, as revealed by fluorescence recovery after photobleaching (FRAP) studies. In the absence of cavins (and caveolae) CAV1 is itself endocytosed preferentially through the CLIC/GEEC pathway, but the pathway loses polarization and sorting attributes with consequences for membrane dynamics and endocytic polarization in migrating cells and adult muscle tissue. We also found that noncaveolar Cavin-1 can act as a modulator for the activity of the key regulator of the CLIC/GEEC pathway, Cdc42. This work provides new insights into the regulation of noncaveolar clathrin-independent endocytosis by specific caveolar proteins, illustrating multiple levels of crosstalk between these pathways. We show for the first time a role for specific cavins in regulating the CLIC/GEEC pathway, provide a new tool to study this pathway, identify caveola-independent functions of the cavins and propose a novel mechanism for inhibition of the CLIC/GEEC pathway by caveolin.
| Endocytosis is the process that allows cells to take up molecules from the environment. Several endocytic pathways exist in mammalian cells. While the best understood endocytic pathway uses clathrin, recent years have seen a great increase in our understanding of clathrin-independent endocytic pathways. Here we characterize the crosstalk between caveolae, flask-shaped specialized microdomains present at the plasma membrane, and a second clathrin-independent pathway, the CLIC/GEEC Cdc42-regulated endocytic pathway. These pathways are segregated in migrating cells with caveolae at the rear and CLIC/GEEC endocytosis at the leading edge. Here we find that specific caveolar proteins, caveolins and cavins, can also negatively regulate the CLIC/GEEC pathway. With the help of several techniques, including quantitative electron microscopy analysis and real-time live-cell imaging, we demonstrate that expression of caveolar proteins affects early carrier formation, causes cellular lipid changes, and changes the activity of the key regulator of the CLIC/GEEC pathway, Cdc42. The functional consequences of loss of caveolar proteins on the CLIC/GEEC pathway included inhibition of polarized cell migration and increased endocytosis in tissue explants.
| Endocytosis encompasses a number of distinct internalization pathways with clathrin-mediated endocytosis (CME) of receptors and their bound ligands being the best understood [1]–[7]. Caveolae, cup-shaped invaginations of the cell surface, have also received much attention as endocytic vehicles [8]–[10]. However, the contribution to cellular endocytic uptake may vary greatly between cell types and conditions, reflecting striking tissue-specific distribution and, presumably, functions of caveolae [11]–[13]. Caveola biogenesis involves the core structural membrane proteins CAV1 and muscle-specific CAV3, which are essential for caveola formation. A coat protein complex, consisting of cavin family proteins, Cavin-1, Cavin-2, and Cavin-3, and muscle-specific Cavin-4 has been shown to associate with caveolae at the plasma membrane (PM) [14]–[19], and together with the GTPase dynamin, the ATPase EHD2, and pacsin2, regulates caveola formation and dynamics [20],[21]. Studies of caveolae have traditionally followed expressed caveolins as markers of caveolae, but it is now apparent that caveolins depend on cavins and associated proteins for association with, and formation of, caveolae. Excess non caveolar caveolin can be rapidly internalized and degraded, for example when caveolin is over-expressed or upon down-regulation of cavin proteins [18],[22]. In contrast, caveolae generally bud off from the PM and recycle back to the surface, transitioning through classical Rab5-positive early endosomes. While caveolar endocytosis may not be a high capacity route in most cell types, the cycle of endocytosis and recycling is required for maintaining a constant caveolar density at the cell surface [13],[23],[24].
Caveolin-independent, clathrin-independent (CI) endocytic routes have, until recently, escaped extensive characterization because of the lack of specific cargo or regulators limiting the available biochemical or molecular tools to study their unique features. Hence, insights into clathrin-independent endocytosis (CIE) have been derived from combining general markers, such as fluid phase probes or membrane markers [25], with cellular systems that lack or have been manipulated to inhibit the classical clathrin and caveolin routes. More recently, a number of endogenous cargo molecules trafficking by distinct, constitutive CI pathways have been identified, including glycosylphosphatidylinositol-anchored proteins (GPI-APs) [26], the interleukin-2 receptor [27], and the major histocompatibility complex I [28]. These pathways differ based upon their dependence on dynamin function and reliance on small GTPases, namely Cdc42, RhoA, or Arf6, for cargo internalization [1],[4],[5].
Here, we have focused on the Cdc42-regulated, GPI-AP-positive CLIC/GEEC (clathrin-independent carriers/GPI-AP enriched early endosomal compartment) pathway, which constitutes a high capacity route for bulk fluid intake in fibroblasts [29]. Previous work has shown the unique, tubular morphology of the primary carriers (CLICs) in this pathway [23] and defined the early stages in the formation of these carriers: clustering of CLIC cargo, such as GPI-APs, within Cdc42-positive regions of the plasma membrane and requirement of local actin polymerization during formation of the carriers [26],[30],[31]. Using multiple fluid phase markers and pulse-chase experiments, it was shown that after CLICs are formed they bud rapidly, within 15 seconds, from the PM and acquire Rab5 and EEA-1, maturing into the GEEC stage, before fusion with early endosomes and mixing with cargo from the CME pathway, such as transferrin (Tfn) [23],[32]. Other recently revealed key regulators of the CLIC/GEEC pathway include a regulator of secretory traffic, Arf1, and GRAF1 (GTPase regulator associated with focal adhesion kinase-1) [33],[34].
CLIC/GEEC endocytosis and caveolae share several similar properties, including the involvement of actin machinery, an important role for free cholesterol, and the action of specific protein/lipids residing in sphingolipid-rich membrane rafts [25],[30],[35]–[37]. However, the two pathways show striking differences in migrating cells: CLIC-mediated endocytosis occurs at the leading edge of the cell while caveolae are localized to the rear [29],[38]. Interestingly, over-expression of CAV1, in cells either lacking CAV1 or containing endogenous CAV1, has been shown to inhibit dynamin-independent internalization of cholera toxin subunit b (CTxB) and fluid phase markers [23],[39]. Whether this has physiological relevance is not yet known.
In this study we have refined and validated our systems that follow CLIC endocytosis and used them to study potential crosstalk with caveolae. Using internalized antibodies to the hyaluronan receptor, CD44, as a CLIC-specific marker [29], our studies reveal complex crosstalk between the two membrane systems at multiple levels.
Various markers have been used to follow CIE pathways but CD44 has emerged as a highly specific cargo of the CLIC/GEEC pathway [29],[40]. We therefore first optimized and validated the conditions for quantitatively studying CIE using antibodies against CD44 as a marker. An anti-CD44 monoclonal antibody (mAb) was added together with fluorescent transferrin (Tfn-647) to cells for 2 and 10 min at 37°C. Prior to fixation, cells were placed on ice and acid stripped to remove any residual surface label, and internalized CD44 mAb was labeled with fluorescently tagged secondary antibody. For quantitative analysis of internalization, the fluorescence intensity of the internalized markers (over the entire cell) was normalized against the average fluorescence intensity of the internalized markers in control samples. By the following criteria, this procedure allowed us to use CD44 as a specific marker of the CLIC/GEEC pathway: (1) Anti-CD44 mAb internalization was distinct from uptake via clathrin coated pits, labeled with Tfn-647, and did not colocalize with caveolae, indicated by probing for endogenous caveolar proteins, CAV1 and Cavin-1 (Figure 1A). (2) CD44 mAb uptake was dynamin-independent, as determined by the use of the small molecule inhibitor dynasore. Under our experimental conditions, dynasore treatment inhibited Tfn-647 uptake but had no effect on CD44 mAb internalization (Figure 1B) [29]. (3) Uptake of CD44 mAb was completely inhibited by 7-ketocholesterol (7-KC) treatment, shown in previous studies to result in reduced membrane order [41], while Tfn-647 uptake was not significantly affected at this concentration (Figure 1B). These results confirmed that in our system anti-CD44 antibody was internalized via the cholesterol-dependent, dynamin-independent CLIC/GEEC pathway [23],[30] (Figure 1B). Internalized anti-CD44 mAb therefore represents a specific marker of the CLIC/GEEC pathway that does not require a low temperature prebinding step and does not affect the kinetics or the magnitude of the pathway being studied [29]. Specific internalization of anti-CD44 mAb was confirmed by using a control antibody (anti-GFP mAb) at the same concentration in wild type mouse embryonic fibroblast cells (WT MEFs) and with anti-CD44 mAb in COS-7 cells, which lack CD44 receptor on the cell surface [42]. Under both conditions, no internalized antibody could be detected after 2 min at 37°C whereas COS-7 cells expressing CD44-GFP showed specific uptake of the anti-CD44 mAb (Figure S1A,B). Additionally, using live cell imaging, we also observed colocalization between CD44-GFP and the fluid phase marker (Dex-647), suggesting that CD44-GFP is internalized via the CLIC pathway (Figure S2, Movie S1). These internalization conditions were subsequently used to investigate the crosstalk between the caveolar and the CLIC/GEEC endocytic systems.
Transient CAV1 over-expression, which leads to caveola formation in cells that express Cavin-1, has been shown to inhibit CIE [23],[39],[43]–[46]. To investigate whether this reflects a general role of caveolar proteins in the CLIC/GEEC pathway, we used CAV1 null (CAV1−/−) and Cavin-1 null (Cavin-1−/−) MEFs (note these cells also have lower Cavin-1 and CAV1 protein levels, respectively, and so do not allow discrimination of specific effects of loss of either protein (Figure S3A) [18]). We compared the constitutive uptake of CD44 mAb and Tfn-647 in WT, CAV1−/−, and Cavin-1−/− MEFs after 2 min at 37°C; at this early time point of internalization, peripheral labeling for CD44 mAb was observed, consistent with a previous study [29] (Figure 1C). A significant 2-fold increase in anti-CD44 mAb fluorescence, as a measure of CLIC internalization, was observed in CAV1−/− and Cavin-1−/− MEFs in comparison with WT MEFs (Figure 1D). However, there was no significant effect on Tfn-647 uptake, indicating that CME was not affected (Figure 1C,D). Additionally, CD44 mAb and Tfn-647 uptake was also performed in CAV1−/− MEFs expressing Cavin-1-specific siRNA to characterize the functional consequences, if any, of loss of Cavin-1 in CAV1−/− MEFs. No significant difference was observed in either CD44 mAb or Tfn-647 uptake between Cavin-1 siRNA (80% knock down) and control siRNA-transfected CAV1−/− cells (Figure S3B). We also examined whether the increase in anti-CD44 mAb internalization was related to altered CD44 protein expression levels in WT, CAV1−/−, and Cavin-1−/− MEFs. Western blot analysis showed no statistically significant increase in the expression level of CD44 in CAV1−/− MEFs but an increase in Cavin-1−/− MEFs when compared with WT cells (Figure 1E). However, quantitative immunofluorescence (IF) analysis showed no significant difference in the ratio of surface/total levels of anti-CD44 mAb labeling between WT, CAV1−/−, and Cavin-1−/− MEFs (Figure 1F).
Altogether, these results suggest that the CLIC/GEEC pathway may be upregulated in the absence of caveolar proteins. We therefore investigated the effect of loss of caveolar proteins on internalization of the fluid phase marker fluorescent dextran (Dex-488), which is partially internalized through the CLIC/GEEC pathway [23],[26]. Dex-488 was internalized at 37°C for 5 min to provide sufficient signal-to-noise ratio. Similar to CD44 mAb uptake, a significant increase in Dex-488 internalization was observed in CAV1−/− and Cavin-1−/− MEFs in comparison with WT MEFs (Figure 1G). Together, these data suggest that lack of caveolae, through two distinct genetic manipulations, causes upregulation of the CLIC/GEEC pathway. Thus, we next investigated the role of specific caveolar proteins in this regulation.
We restored the levels of CAV1 and Cavin-1 in CAV1−/− MEFs by expressing the fluorescent protein (YFP/GFP) tagged constructs, and then compared uptake of the CD44 mAb (2 min at 37°C) between transfected and untransfected cells (Figure 2A; Figure S4). We observed 30–40% transfection efficiency for almost all transient transfections and only low transgene-expressing cells were analyzed. CD44 mAb endocytosis was drastically reduced in CAV1-expressing cells (Figure 2B). More surprisingly Cavin-1 expressing CAV1−/− MEFs also showed a dramatic decrease in CD44 mAb endocytosis (Figure 2C). As CAV1−/− cells lack morphological caveolae [23] this strongly suggests that Cavin-1 inhibits CLIC endocytosis independent of CAV1 and caveolae. Additionally, exogenous expression of Cavin-1 in Cavin-1−/− MEFs was also observed to cause a significant decrease in CD44 mAb uptake (Figure 2D). Neither CAV1 nor Cavin-1 expression affected Tfn-647 uptake. Compared with untransfected cells, CAV1 over-expression resulted in 95±0.3% inhibition of CD44 mAb endocytosis, while Cavin-1 caused a 70±2.2% inhibition (Figure 2B,C). Since loss of caveolar proteins resulted in an increase in Dex-488 uptake, we also investigated the effects of reconstituted CAV1 and Cavin-1 expression in CAV1−/− MEFs on Dex-488 uptake (5 min at 37°C). Similar to CD44 mAb uptake, expression of either CAV1 or Cavin-1 resulted in a significant decrease in Dex-488 uptake (Figure S5A,B; 40±7% inhibition in CAV1-expressing cells; 44±6% inhibition in Cavin-1 expressing cells) although not as high inhibition as observed with the specific CLIC marker, CD44 mAb. We also characterized the effect of other members of the caveolin family, CAV2 and CAV3, on the CLIC/GEEC pathway. The ectopic expression of CAV3, but not CAV2, inhibited CD44 mAb uptake in CAV1−/− MEFs, while Tfn-647 uptake was not affected by expression of either of the proteins (Figure 2E,F). To test whether inhibition of the CLIC/GEEC pathway activity upon expression of caveolar proteins in our system (CAV1−/− MEFs) is due to inhibition of CLIC/GEEC carrier formation, we used electron microscopy (EM). CAV1−/− MEFs were transiently transfected with CAV1-YFP, Cavin-1-GFP, and with GFP alone respectively and then the fluid phase marker horseradish peroxidase (HRP) was added for 2 min at 37°C to label the putative CLIC/GEEC carriers [23],[29]. In comparison with cells expressing GFP alone, a significant decrease was observed in the number of CLICs in CAV1 and Cavin-1 transfected cells (GFP: 100%; CAV1: 54±4% decrease; Cavin-1: 55±8% decrease; mean ± SEM), indicating that expression of caveolar proteins can limit the formation of carriers of the CLIC/GEEC pathway (Figure 2G).
These results suggested that both caveolins (CAV1 and CAV3) and Cavin-1 are capable of negatively regulating the CLIC/GEEC pathway. This raised the question of whether inhibition is caused exclusively by free caveolar proteins, independent of caveolar formation. We addressed this question by testing whether inhibition of the CLIC/GEEC pathway occurred when levels of caveolin and cavin were balanced, as indicated by their colocalization and by the immobilization of CAV1 within caveolae. Cavin-1−/− MEFs were transfected either with CAV1-YFP or with Cavin-1-mCherry alone or with both constructs simultaneously. Both CAV1 and Cavin-1 proteins were expressed in similar amounts in single and double transfections, as confirmed by quantitative Western blot analysis (Figure 3A). Moreover, confocal images showed negligible levels of cytosolic Cavin-1 present in CAV1 and Cavin-1 co-expressing cells, and quantitative colocalization analysis showed a significantly higher degree of colocalization between CAV1 and Cavin-1 at the PM in comparison with analysis of randomized pixels for the same regions (Pearson coefficient, PM: 0.80±0.02; random: −0.007±0.005; mean ± SEM, n = 40 cells, p<0.0001) (Figure 3B). In the same set of cells, anti-CD44 mAb and Tfn-647 uptake studies were performed. A significant decrease in CD44 mAb endocytosis was observed in cells co-transfected with CAV1 and Cavin-1, whereas Tfn-647 uptake was not affected (Figure 3C). To gain a quantitative estimation of free caveolar proteins present in our system upon co-expression of CAV1 and Cavin-1, we made use of fluorescence recovery after photobleaching (FRAP) to analyze the mobility and diffusion properties of CAV1. Cavin-1−/− MEFs were transiently transfected with CAV1-YFP or co-transfected with CAV1-YFP and Cavin-1-mCherry respectively. After photobleaching a defined region of interest (ROI) at the PM, we compared the rate of fluorescence recovery of CAV1 between the single and co-transfected cells. A significant decrease in the mobile fraction (CAV1: 0.82±0.03; CAV1+Cavin-1: 0.58±0.05, p<0.05) and lateral diffusion (CAV1: 0.13±0.08; CAV1+Cavin-1: 0.039±0.002; mean ± SD, n = 15 cells, p<0.001, see Text S1) of CAV1 was observed when co-expressed with Cavin-1, in comparison with expression alone (Figure 3D), suggesting that most of the CAV1 in the presence of Cavin-1 had been immobilized through the formation of caveolae. Taken together, these results suggest that the caveolar proteins, independent of caveolae, can inhibit CLIC endocytosis, but in addition, when incorporated into caveolae, there is also potent inhibition of CLIC endocytosis.
To further investigate the unexpected regulatory role of cavins in CLIC/GEEC endocytosis, we analyzed the effects of ectopic expression of each of the four members of the cavin family (Figure S6A) on CD44 mAb internalization. In CAV1−/− MEFs, moderate over-expression of Cavin-1 and Cavin-3 caused a significant decrease in CD44 mAb endocytosis, whereas in Cavin-2 and Cavin-4-expressing cells no effect on CD44 mAb endocytosis was observed (Figure 4A). Expression of both Cavin-1 and Cavin-3 resulted in significant inhibition (Cavin-1: 77.9±3.3% decrease; Cavin-3: 76.9±3.9% decrease; mean ± SEM) of CD44 mAb endocytosis when compared with control cells (100±13%; Figure 4B). These results reaffirmed the noncaveolar role of Cavin-1 as a negative regulator of CLIC/GEEC endocytosis. Furthermore, Cavin-3 was also observed to have a CAV1- and caveola-independent regulatory role in CLIC/GEEC endocytosis.
We next assessed the effect of reducing cavin protein levels on the CLIC/GEEC pathway using a siRNA directed approach in 3T3-L1 fibroblasts, as we obtained more efficient knock down in this cell type compared with MEFs. Reduction of Cavin-1 levels (by approximately 80%) and Cavin-3 levels (by approximately 50%) resulted in an increase in CD44 mAb uptake in comparison with control siRNA-treated cells (Figure 4C,D; Figure S6B). Reduction of Cavin-3 levels also led to an increase in Tfn-647 uptake while reduction in Cavin-1 did not, suggesting Cavin-3 might influence CME.
To characterize the regulatory mechanism by which CAV1 inhibits the CLIC/GEEC pathway, we investigated the role of caveolin scaffolding domain (CSD) of CAV1, as this domain has been shown to mediate several regulatory roles of CAV1 [47],[48]. We tested CAV1 and CAV3 CSD point substitution mutants (CAV1G83S; CAV3G55S) and the deletion mutant (CAV1Δ80–100) as used in previous studies [43]. Expression of the CSD mutants, unlike the WT proteins, did not inhibit CD44 mAb internalization in CAV1−/− MEFs (Figure 5A–C), suggesting that the CSD is required for inhibition of the CLIC/GEEC pathway. We next investigated whether the CSD is sufficient for inhibition by expressing the minimal CAV1 scaffolding domain as a fusion protein with GFP (CAV1-SD, amino acid 82–101) in CAV1−/− MEFs. CD44 mAb uptake was significantly decreased (58±3% decrease; mean ± SEM) in cells expressing CAV1-SD while Tfn-647 uptake was unaffected (Figure 5D). Taken together, these results suggest that inhibition of the CLIC/GEEC pathway by caveolin proteins requires an intact scaffolding domain and that this domain alone has significant inhibitory activity on CLIC endocytosis. The fact that the scaffolding domain has inhibitory activity when expressed in cells lacking endogenous CAV1 shows that the inhibitory activity of the mutant is not mediated through inhibition of interactions of proteins with endogenous CAV1 and caveolae but is an inherent property of this polypeptide. This suggests that CAV1 may affect fundamental membrane properties, as investigated in the following section.
CAV1 has been shown to interfere with the mobility of membrane microdomain-associated proteins, which further blocks the integrin-mediated internalization of bacteria [49]. We investigated whether similar effects could underlie the influence of expressed CAV1 and cavins on the CLIC/GEEC pathway.
We first analyzed, using FRAP, the degree of mobility of a lipid raft-associated CLIC/GEEC cargo protein, GPI-YFP, in the presence and absence of CAV1. WT and CAV1−/− MEFs were transiently transfected with GPI-YFP or co-transfected with GPI-YFP and CAV1-mCherry. To analyze the mobility of GPI-YFP, a small ROI was bleached at the PM, half-life times of fluorescence recovery were recorded and the resulting diffusion coefficients were calculated. When compared with WT MEFs or CAV1−/− MEFs expressing CAV1, a significant increase was observed in the mobility of GPI-YFP in CAV1−/− MEFs (diffusion coefficient for WT: 0.48±0.03; CAV1−/−+CAV1: 0.50±0.02; CAV1−/− 0.67±0.03; mean ± SEM, n>20 cells, p<0.001) (Figure 6A). We also analyzed the mobility of a model GPI-anchored protein in CAV1−/− MEFs co-expressing GPI-YFP and the CSD mutant CAV1G83S. As expected, the mutant failed to decrease the mobility of GPI-YFP (diffusion coefficient 0.78±0.06) (Figure 6A), suggesting that localization of CAV1 at the PM with an intact scaffolding domain is required to restrict the mobility of the tested microdomain-associated proteins. The very short half-life times (around 8.5 seconds) observed in the above analysis make it unlikely that the changes are due to a defect in trafficking of GPI-anchored proteins, as exocytosis and endocytosis occur at longer time scales.
We next tested the effect of CAV1 on the mobility of the CLIC-specific cargo protein CD44. To characterize CD44 membrane diffusion ability, we made use of an expression construct generated by fusing a photo-activatable variant of green fluorescent protein (mRFP-PAGFP) to CD44 (mRFP-PAGFP-CD44, termed PA-CD44). First, we tested the internalization route of the PA-CD44 construct. PA-CD44 was photo-activated at a specific region of the PM of PA-CD44-expressing COS-7 cells. Internalized PA-CD44 showed significant colocalization with Dex-647-labeled endocytic vesicles in the photo-activated region consistent with endocytosis via the CLIC pathway (Figure S7, Movie S2). For diffusion studies, PA-CD44 was photo-activated in a small ROI at the PM and decay of fluorescence over time was measured in CAV1−/− MEFs and CAV1−/− MEFs expressing CAV1, respectively. In the presence of CAV1, the CD44 diffusion rate was significantly decreased in comparison with CAV1−/− MEFs and was on a similar time scale to that observed in the FRAP experiments (Figure 6B).
As the effects of CAV1 on CD44 and GPI-YFP may be mediated through its impact on physical properties of the PM, we then investigated whether stimulation of membrane fluidity by chemical or physical means would have any effect on internalization of cargo proteins of the CLIC/GEEC pathway using two independent treatments [49],[50]. As a chemical stimulus, we treated CAV1−/− MEFs with 0.05% Tween20 for 15 min at 37°C. After treatment, cells were incubated with anti-CD44 mAb and Tfn-647 for 2 min at 37°C. Secondly, we used high temperature, which can also alter the membrane mobility of proteins. CAV1−/− MEFs were incubated with anti-CD44 mAb and Tfn-647 for 2 min at 41°C. Interestingly, both physical and chemical stimuli caused a significant increase in CD44 mAb uptake while Tfn-647 uptake was unaffected, compared with untreated cells (Figure 6C). We further analyzed whether alterations in membrane fluidity or mobility of microdomain-associated proteins could overcome the inhibitory effect of CAV1 and Cavin-1 on their internalization by the CLIC/GEEC pathway. CAV1−/− MEFs were transiently transfected with CAV1-YFP and Cavin-1-GFP respectively. Cells were either treated with Tween20 prior to uptake or incubated with endocytic markers at 41°C, as described above. Both treatments significantly increased CD44 mAb uptake, overcoming the inhibitory effect of CAV1 (Figure 6D). In contrast the treatments did not affect endocytosis in Cavin-1 expressing cells (Figure 6E). Tfn-647 internalization was also unaffected in both treated and untreated cells (Figure 6D,E). We also tested the effect of an increase in membrane fluidity on fluid phase endocytosis. CAV1−/− MEFs were transiently transfected with CAV1-mCherry, and Dex-488 uptake (5 min) was performed at 41°C. At 41°C, Dex-488 uptake was unaffected by CAV1 expression, suggesting that higher temperature rescues the inhibition of both a specific CLIC marker and a fluid phase marker (Figure 6F). Taken together, these results suggest that CAV1 expression can modulate the mobility of membrane microdomain-associated proteins. Treatments that increase the mobility of these proteins can at least partially restore endocytosis through the CLIC pathway.
The physical and chemical stimuli described above alter mobility or fluidity of microdomain-associated proteins, an effect that might indicate lipid changes in the PM. As cholesterol is crucial for the function of the CLIC pathway [30], we investigated the effect of these treatments on cellular cholesterol distribution using the polyene antibiotic, filipin. CAV1−/− MEFs were transiently transfected with CAV1-YFP, treated with Tween20 or incubated with endocytic markers at 41°C, and labeled with filipin. In CAV1-expressing and treated cells, higher filipin labeling was observed (135.7±4.8% for Tween20 and 132.7±4.1% for 41°C; mean ± SEM) in comparison with treated, untransfected CAV1−/− MEFs (100±1.6%; mean ± SEM) (Figure 6G) or untreated CAV1-expressing cells. This suggests that CAV1 expression affects membrane cholesterol distribution/availability, which inhibits internalization via the CLIC/GEEC pathway. The physical and chemical stimuli restore cholesterol distribution allowing up-regulation of the CLIC/GEEC pathway. This effect is more complex than simply the availability of free cholesterol in the bulk membrane because cholesterol addition in the form of a cholesterol–cyclodextrin complex did not rescue the CAV1-mediated inhibition (unpublished data).
Unlike CAV1, physical and chemical stimuli could not rescue the internalization of CD44 mAb in Cavin-1-expressing cells (Figure 6E), suggesting Cavin-1 might not act by regulating the mobility of microdomain-associated surface proteins in these cells. Cavins have also been reported to respond to cholesterol levels and alter cholesterol cellular distribution [51],[52]. Thus, we next tested whether expression of Cavin-1 in cells lacking CAV1 could affect cellular cholesterol distribution by labeling Cavin-1-GFP-expressing CAV1−/− MEFs with filipin to analyze the distribution of free cholesterol. In Cavin-1-expressing cells we observed a significant increase (138±8.4%; mean ± SEM) in filipin labeling in comparison with untransfected cells (100±6.5%; mean ± SEM) (Figure 7A). These results suggest that Cavin-1 expression can alter the distribution of filipin staining, indicating a change in distribution of cholesterol in cells.
To examine the possible implications of such alteration in lipid distribution, we focused on Cdc42, whose localization and activation has been shown to depend on specific lipid organization at the PM [53]. CAV1−/− MEFs were transiently transfected with Cavin-1-GFP and endogenous Cdc42 protein levels were analyzed by Western blotting. No effect was observed on total Cdc42 protein levels in Cavin-1-expressing cells (Figure 7B, Figure S8). We next characterized the correlation between Cavin-1 and Cdc42 at the PM. CAV1−/− MEFs expressing Cavin-1-GFP were labeled for endogenous Cdc42 and subjected to Pearson coefficient analysis to calculate the correlation between the two proteins. This revealed a significant spatial correlation, compared with randomized values, between Cdc42 and Cavin-1 at the cell surface, although less than that between CAV1 and Cavin-1 (Cavin-1–Cdc42: 0.21±0.02, randomized: −0.01±0.01, CAV1–Cavin-1 PM: 0.78±0.02; mean ± SEM, n = 40, p<0.0001) (Figure 7C). This suggested that Cdc42 and Cavin-1 might colocalize within regions of the cell surface. Indeed, live-cell imaging of CAV1−/− MEFs co-transfected with Cdc42-GFP and Cavin-1-mCherry revealed that these two proteins co-accumulated in PM ruffles (Figure S9A). Further live-cell imaging demonstrated that GPI-YFP also co-accumulated with Cavin-1-mCherry in PM ruffles, suggesting that these are sites for the possible association of Cavin-1 with CLIC components (Figure S9B).
We next tested whether Cavin-1 also affected Cdc42 activity, using a fluorescently tagged CRIB domain (Cdc42/Rac-interacting binding domain from N-WASP). This construct (CRIB-YPet) binds to GTP-loaded Cdc42 and thus can act as a location biosensor for active Cdc42. CAV1−/− MEFs were either transfected with CRIB-YPet alone or co-transfected with both CRIB-Ypet and Cavin-1-mCherry. Quantitative line scan analysis of fluorescence intensity showed a high degree of colocalization between CRIB and Cavin-1 at the PM ruffles in comparison with cytosol (Pearson coefficient, PM: 0.82±0.04; cytosol: 0.35±0.01; mean ± SEM, p<0.001) (Figure 7D). This supported the notion that Cavin-1 colocalized with active Cdc42 in ruffles.
We then tested whether Cdc42 could be activated by Cavin-1 expression. We used the FRET pair of Cdc42-CyPet and CRIB-YPet, whose association on GTP-loading of Cdc42 leads to a decrease in the lifetime of CyPet that can be measured by FLIM-FRET. CAV1−/− MEFs were co-transfected with Cdc42-CyPet and CRIB-YPet either in the presence or absence of Cavin-1. The lifetime of CyPet was significantly reduced, consistent with energy transfer in Cavin-1-expressing cells (1.5±0.02 nanoseconds), compared with cells lacking Cavin-1 (1.64±0.01 nanoseconds; mean ± SEM) (Figure 7E). Therefore Cavin-1 expression promotes activation of Cdc42 in CAV1-deficient cells. All together these results suggest that Cdc42 is selectively activated in PM ruffles when non-caveolar Cavin-1 is expressed in cells.
We also noted additional cellular effects upon the loss of Cavin-1. In WT MEFs, neither CAV1 nor Cavin-1 colocalized significantly with either anti-CD44 mAb or Tfn-647 after 2 or 10 min of uptake (Figure 1A). However, in Cavin-1−/− MEFs colocalization was observed between CAV1 and the internalized CD44 mAb but not with Tfn-647 after 2 min of uptake (Figure 7F). We further confirmed this association of CAV1 with CLICs by assessing its localization following inhibition of either CLICs or CME. Upon inhibition of CME by the small molecule dynamin inhibitor dyngo4a, CAV1 still colocalized with the CD44 mAb, whereas this colocalization was lost when the CLIC/GEEC pathway was inhibited by 7-KC (Figure S10). This suggests that noncaveolar caveolin is preferentially recruited into the CLIC/GEEC pathway in cells lacking Cavin-1.
Loss of Cavin-1 also affects trafficking through the CLIC/GEEC pathway. After 10 min of uptake in the Cavin-1−/− MEFs, internalized CD44 mAb and Tfn-647 were seen in the same endocytic vesicles (Figure 7F); this was not observed in WT or in CAV1−/− MEFs. These results provide evidence for a novel role for caveolar proteins in regulating not only the magnitude but also specific features of the endosomal system. In this respect we noted a significant increase in the expression levels of Tfn receptor in Cavin-1−/− MEFs (Figure S11).
The inhibitory roles of caveolae, caveolins, and cavins suggest an important role for caveolae in regulating the CLIC/GEEC pathway. We therefore examined the physiological consequences of this inhibition. We first examined whether caveolae could play a role in spatial organization of the CLIC/GEEC pathway in migrating cells. A scratch-wound assay was applied to a confluent monolayer of CAV1−/− MEFs co-expressing CAV1 and Cavin-1. Fluorescently tagged CAV1 and Cavin-1 showed complete colocalization at the rear of the migrating cells (Figure 8A), consistent with previous studies showing caveolae enriched in this domain [38],[54]. CLIC endocytosis as detected by CD44 mAb uptake was dramatically reduced by the co-expression of the two caveolar proteins but the few carriers observed were invariably in areas lacking caveolae (Figure 8A). Concomitantly, in confluent monolayers of CAV1-expressing CAV1−/− MEFs, we also checked the cellular distribution of Cdc42, a pivotal regulator of polarity and the CLIC/GEEC pathway. In CAV1-YFP-expressing cells Cdc42 expression, detected by a Cdc42-specific antibody, was observed to be excluded from CAV1-positive areas. Quantification of fluorescence intensity at the cell surface, by line scan analysis, showed a significant decrease in Cdc42 expression in regions of PM expressing high levels of CAV1 compared with low- or nonCAV1-expressing regions (CAV1 high expression region: 14±0.32; low/no expression regions: 26±0.60; mean ± SEM, p<0.0001), within the same cell (Figure 8B). However, overall Cdc42 protein levels were unaltered in CAV1-YFP-expressing cells, as observed by Western blotting (Figure 7B, Figure S8).
To further characterize the correlation between Cdc42 and CAV1 at the cell surface, we performed the Pearson correlation coefficient analysis on the same set of images that were used for line scan analysis. This revealed significantly less correlation between CAV1 and Cdc42 at the PM in comparison with CAV1 and Cavin-1 at the PM (Pearson coefficient, CAV1–Cdc42: 0.11±0.01, randomized: −0.002±0.001, CAV1–Cavin-1 PM: 0.78±0.02; mean ± SEM, n = 40, p<0.0001) (Figure 8C). As expected from the above results, the activity of Cdc42, monitored by recruitment of the CRIB domain, was also negatively affected by CAV1 expression. Quantitative line scan analysis showed significantly fewer CRIB protein in regions of the PM expressing high levels of CAV1 in comparison with lower-expressing regions (PM region enriched in CAV1: 45±1.9; PM regions with low/no CAV1 expression: 57±2.4; mean ± SEM, p<0.001) within the same cell (Figure 8D). This suggests that CAV1 expression can result in differential distribution of CLIC components; in addition, caveolae can locally inhibit CLIC endocytosis and help polarize the pathway to the leading edge.
To test this hypothesis, we next examined the effect of loss of caveolar proteins on polarization of CLICs to the leading edge of 2D migrating fibroblasts, a key feature of the CLIC/GEEC pathway [29]. First we analyzed whether lack of caveolar proteins can alter cell migration by using a scratch-wound assay applied to confluent monolayers of WT, CAV1−/−, and Cavin-1−/− MEFs. As shown in a previous study [55], we observed a significant decrease in ability of CAV1−/− MEFs to close the wound compared with WT MEFs (after wound closure at 12 h, WT: 84±2.1%; CAV1−/−: 54±4.1%; mean ± SEM, p<0.0001) (Figure 8E). Similarly, Cavin-1−/− MEFs also showed less efficient wound closure compared with WT MEFs and showed defects in cell migration (after wound closure at 12 h, Cavin-1−/−: 60±2.5%; mean ± SEM, p<0.001) (Figure 8E). To characterize the polarization of CLICs in migrating cells, confluent monolayers of WT, CAV1−/−, and Cavin-1−/− MEFs were scratch-wounded and uptake of CD44 mAb and Tfn-647 was compared with the localization of CAV1 or Cavin-1 as cells migrated into wound. In both CAV1−/− and Cavin-1−/− MEFs, CD44 mAb internalization occurred at both the leading and trailing edge, indicating the CLIC/GEEC endocytosis was no longer polarized in these cells (Figure 8F).
Loss of caveolins and cavins has profound effects on specific tissues in vivo, including skeletal and striated muscle (reviewed by [56]). To examine whether the loss of caveolar components also affected endocytic activity in a differentiated tissue relevant to disease, we isolated mature adult muscle fibers from the flexor digitorum brevis muscle of WT and Cavin-1−/− mice. We assessed endocytic activity by adding either anti-CD44 mAb to the isolated fibers and visualizing uptake by IF, or by using HRP as a fluid phase marker for EM analysis. No significant difference was seen in CD44 surface labeling of the fibers, suggesting endogenous levels and surface accessibility of CD44 are not altered (not shown). CD44 mAb internalization was dramatically increased in fibers isolated from Cavin-1−/− mice (23.4±2.9, compared with WT: 2.2±0.3; mean ± SEM) (Figure 9A). Similarly, isolated fibers incubated with HRP as a fluid phase marker showed a highly significant 4.5-fold increase in the volume of HRP-labeled structures, as compared with WT fibers (Figure 9B–D), indicative of greatly increased fluid phase endocytosis. Taken together, these data show that the CLIC/GEEC pathway is fundamentally altered by the loss of Cavin-1 demonstrating a dramatic in vivo consequence of loss of caveolar components.
In this study we have identified extensive crosstalk, at several levels, between the caveolae and the CLIC/GEEC endocytic pathway in mammalian cells. Previous studies have shown that caveolin over-expression inhibits CIE [23],[39]. We now show that both expression comparable with physiological levels and down-regulation of caveolin can regulate endocytosis, demonstrating that CAV1 is an important cellular endocytic regulator. Previous studies suggested a role for Tyr14 phosphorylation of caveolin in inhibition of plasma membrane Cdc42 required for fluid phase endocytosis [39] but this cannot explain the findings presented here. Firstly, CAV3, which lacks a tyrosine residue equivalent to Tyr14, is an equally potent inhibitor of CIE as is CAV1. Secondly, we observed no significant CAV1 colocalization with Cdc42. Finally, expression of the isolated caveolin scaffolding domain as a fusion protein with GFP also significantly inhibited the CLIC/GEEC endocytosis. This inhibition was observed even in cells lacking endogenous caveolins, arguing against an effect on caveolin-signaling protein interactions. These results hinted at a more general mechanism of inhibition, and in view of previous studies [49],[57] prompted us to test whether specific properties of the plasma membrane were affected by CAV1 expression. Expression of CAV1 significantly decreased the membrane mobility of CD44 and GPI-YFP membrane microdomain-associated CLIC cargo proteins. Interestingly, effects of CAV1 on both diffusion properties and internalization of CLIC cargo proteins required an intact scaffolding domain of CAV1, as inhibitory effects were lost with single-point mutations in this region. A role for CAV1, dependent on the scaffolding domain, has been reported in regulation of dynamin-dependent, raft-mediated internalization of the CTxB as well as in endocytosis of bacterial engaged integrins [48],[49]. It is therefore tempting to speculate that perturbation of membrane properties by CAV1 expression underlies the effect on the CLIC/GEEC endocytosis. We could significantly rescue the inhibition of CLIC/GEEC endocytosis exerted by CAV1 on CD44 mAb and fluid phase marker uptake by using previously characterized chemical and physical stimuli, which enhanced the fluidity of the PM [49]. Under these conditions, we also observed increased cellular staining for free cholesterol, but only in CAV1-expressing cells. We conclude that complex changes in cholesterol trafficking/accessibility may accompany CAV1 expression, specifically perturb CLIC/GEEC endocytosis through effects on membrane lipid composition, and be rescued by experimental manipulation of membrane fluidity. Independent studies from our laboratory have shown that loss of CAV1 also has striking effects on nanoscale organization of the plasma membrane, including increased clustering of phosphatidylserine and farnesylated K-ras but decreased nanoclustering of dually palmitoylated H-ras [58]. Additionally, quantitative EM analysis pointed at a direct inhibitory effect of CAV1 on early carrier formation in the CLIC/GEEC pathway. However, we cannot rule out additional CAV1 effects on the mobility of microdomain-associated cargo proteins affecting access into the CLIC pathway.
The effect of the 20-amino acid scaffolding domain of CAV1 on CIE indicates that this domain of caveolin can have potent biological activity even in cells lacking endogenous caveolin. The activities of corresponding peptides added to cells in the form of a cell-penetrating fusion protein are well described as potent activators of endothelial nitric-oxide synthase (eNOS) through inhibition of caveolin–eNOS interactions and implicated in regulation of Rac1 signaling, a putative CLIC-interacting protein, by CAV1 [29],[59],[60]. However, these effects are observed only in cells with endogenous caveolin [47], unlike the effects described here. Nevertheless, the potent effects of the isolated scaffolding domain on endocytic trafficking described here should be taken into account when evaluating the effects of these peptides in cells and tissues. In addition, the role of CAV1 in many signaling pathways could potentially be explained by the inhibition of this major endocytic pathway in view of the importance of endocytic trafficking in many signaling events [61]. Again, consistent with this, both inhibition of signaling [60],[62],[63] and CLIC endocytosis (this study; [39]) show a similar dependence on the caveolin scaffolding domain.
To our surprise, our studies also revealed a novel noncaveolar regulatory role for cavin family proteins, specifically Cavin-1 and Cavin-3. While expression of non-caveolar Cavin-1 inhibited the formation of CLIC/GEEC carriers we could not detect any changes in the physical properties of the membrane upon Cavin-1 expression, although more subtle changes below our detection limits certainly cannot be excluded. Instead, single cell-based analysis of Cavin-1 expressing CAV1−/− MEFs consistently demonstrated an increased intracellular filipin labeling. Our recent studies showed that Cavin-1 expression in prostate cancer cells that express endogenous CAV1-modulated secretion pathways and cholesterol distribution, with decreased levels of cholesterol and impaired recruitment of actin to the detergent resistant membranes [51]. In view of the dependence of the CLIC/GEEC pathway on actin and cholesterol [30], including perturbation by 7-KC, as shown in the present study, it is likely that Cavin-1 regulates the CLIC/GEEC pathway through effects on cholesterol trafficking and/or distribution. However, it is intriguing that this inhibition can occur in the absence of CAV1 and caveolae. In this context it is noteworthy that detailed real-time observations of Cavin-1 localization in CAV1−/− cells revealed colocalization with active Cdc42 and GPI-AP in membrane ruffles. Availability of phosphatidylserine (PS) at the PM has been shown to be critical for membrane localization and activation of Cdc42 [53], and previous work from our laboratory has shown that Cavin-1 can bind to the PS at PM [18]. Hence it seems plausible that by interacting with CLIC components/regulators at the PM Cavin-1 regulates the activity of the pathway. In support of this hypothesis we observed that Cavin-1 could modulate Cdc42 activity and so perturb the Cdc42 activation–deactivation cycle. This would be limiting for the activity of the CLIC/GEEC pathway, which is dependent on a functional Cdc42 cycle with inhibition by an activated form of Cdc42 [30],[33].
Cavin-1 appears to be a highly specific regulator of the CLIC/GEEC pathway in comparison with CME and may be a useful tool to further characterize the CLIC/GEEC endocytic route. In wild-type cells, CD44 is rapidly sorted into a distinct, transferrin-negative recycling route [29] but this sorting ability of CLICs was reduced in Cavin-1−/− cells, suggesting a novel role for Cavin-1 in endosomal sorting possibly through indirect effects on cholesterol. Similar observations were not made in CAV1−/− cells even though these cells have reduced Cavin-1 levels. An intriguing possibility is that the low levels of noncaveolar CAV1 in CAV1−/−, which we show is internalized predominantly through the CLIC/GEEC pathway, has a modulatory effect on sorting in the endocytic pathway.
In our system Cavin-3 also showed a potent and specific inhibition of the CLIC/GEEC endocytosis when expressed alone, while knockdown led to a significant increase in both CLIC/GEEC and clathrin-dependent endocytosis. A systemic study of endocytosis, exploiting multi-parametric image analysis and multi-dimensional gene profiling, has also implicated Cavin-3 in endocytic trafficking of transferrin [64]. Together, these studies suggest a more general role for Cavin-3 in regulation of endocytic trafficking.
Co-expression of caveolar proteins in cells generated caveolae and negatively regulated the CLIC/GEEC pathway. Our results suggest that association of caveolins and cavins with caveolae does not perturb their inhibitory activity and hence identifies cavins, caveolins and caveolae as regulators of CI endocytosis. What might be the function of this complex regulatory network? Caveolae are polarized in migrating cells and highly concentrated at the rear of the cell [18],[29],[38],[54]. In contrast, CLIC/GEEC endocytosis predominantly occurs at the leading edge of migrating cells [29]. In line with the above studies, we observed that expression and activation of the CLIC regulator, Cdc42, was differentially distributed in areas of the PM significantly devoid of CAV1 expression. Upon expression of CAV1 and Cavin-1, CLIC/GEEC endocytosis is reduced but residual endocytosis occurs exclusively in regions devoid of caveolae, suggesting potent local inhibition of endocytosis. Conversely, loss of CAV1 and/or Cavin-1 causes a reduction in polarized migration and polarization of CLIC/GEEC endocytosis leading to increased endocytosis at the rear of the cell. This suggests that polarization of the CLIC/GEEC pathway is dependent directly or indirectly on the caveolar membrane system and that caveolae can “dampen” endocytic activity at the rear of the cell. This model is further strengthened by examination of explant tissues from mice lacking caveolae due to genetic deletion of Cavin-1. WT skeletal muscle fibers show a remarkable density of caveolae and relatively low endocytic activity. Loss of caveolae in the skeletal muscle of Cavin-1−/− mice causes a dramatic increase in endocytic activity as monitored by CD44 mAb uptake and by fluid phase uptake.
It is now clear that the consequences of expression of caveolins or cavins are highly complex and modulation of membrane traffic and lipids must be taken into account in understanding their actions. The evolution of two sets of protein components associated with caveolae as independent inhibitory agents, when free or incorporated into caveolae, and the profound cellular consequences of their loss emphasizes the importance of the caveolar system as a key regulator of endocytosis in mammalian cells.
All the animal experiments were conducted in accordance with the guidelines of the ethics committee at The University of Queensland.
Mouse anti-CD44 (clone 5035-41.1D, Novus Biologicals), rabbit anti-CAV1 (BD Biosciences), rabbit anti-Caveolin2 (Sigma Aldrich), rabbit anti-Cavin-1 and Cavin-4 antibody were raised as described previously [16], rabbit anti-Cavin-3 (ProteinTech group), mouse anti-Cavin-2 (Sigma Aldrich), rabbit anti-HA (Sigma Aldrich), mouse anti-Cdc42 (Becton Dickinson), mouse anti-GFP (Roche), mouse anti-transferrin receptor antibody (Zymed), Alexa Fluor conjugated dextran (Life Technologies), anti-rabbit and anti-mouse Alexa Fluor antibodies (Invitrogen), Alexa Fluor conjugated transferrin (Invitrogen), Alexa Fluor conjugated phalloidin (Invitrogen), Filipin III (Sigma Aldrich), Dynasore (Sigma Aldrich), 7-Ketocholesterol (Sigma Aldrich), Protease Inhibitor Cocktail Set III (Merck Millipore), PhosSTOP Phosphatase Inhibitor Cocktail (Roche), Dyngo-4a (Sigma). Stealth RNAi siRNA duplex oligonucleotides targeted against mouse Cavin-1 (5′CCGCUGUCUACAAGGUGCCGCCUUU3′;5′AAAGGCGGCACCUUGUAGACAGCGG3′) and Cavin-3 (5′CCGGAGCUCUGAAGGCCCAUCAGAA3′; 5′UUCUGAUGGGCCUUCAGAGCUCCGG3′) (Invitrogen).
WT, immortalized CAV1-null (CAV1−/−), Cavin-1 null (Cavin-1−/−) mouse embryonic fibroblasts (MEFs), 3T3-L1 and COS-7 cells were maintained in Dubelcco's modified Eagle's medium (DMEM; Gibco) supplemented with 10% fetal bovine serum (FBS; Cambrix) and 2 mM L-Glutamine at 37°C and 5% CO2. For transfection, cells were either electroporated (BIORAD) or Lipofectamine 2000 (Invitrogen) was used, as per the manufacturer's instructions, and by this method 30–40% transfection efficiency was achieved. For siRNA transfection, Cavin-1 and Cavin-3-specific oligonucleotides were transfected at a final concentration of 600 pM using Lipofectamine 2000.
Cells grown on 12 mm coverslips were placed on 50 µl pre-warmed drops containing 10 µg/ml anti-CD44 antibody/3 mg/ml Dex-488/10 µg/ml anti-GFP antibody+10 µg/ml Tfn-647 diluted in culture media and incubated at 37°C in a 5% CO2 incubator for time indicated. To remove any surface bound markers (CD44 mAb, GFP mAb, or Tfn-647), 2×30-second acid stripping was performed with 0.5 M glycine (pH 2.2), before fixation with 4% paraformaldehyde. The cell migration (scratch-wound) assay and immunocytochemistry were performed as previously described [29]. For cholesterol staining, cells were labeled with filipin, 50 µg/ml, for 2 h in the dark. For Cdc42 staining, cells were fixed with 10% TCA and quenched using 3×5-min washes of 30 mM glycine on ice. Permeabilized cells were blocked with 3% bovine serum albumin (BSA) and incubated in primary antibody overnight at 4°C followed by Alexa Fluor conjugated secondary antibody incubation for 40 min at room temperature.
In transfected CAV1−/− MEFs, HRP uptake (10 µg/ml) was performed at 37°C for 2 min, and after brief washing with DMEM containing 1% BSA, diaminobenzidine (DAB) (10 mg/ml) reaction was performed on the live cell as previously described in [29]. Fixation, embedding, and sectioning were performed as follows. MEFs were fixed with 2.5% glutaraldehyde in 0.1 M cacodylate buffer (pH 7.4). Cells were post-fixed with 1% osmium tetroxide for 1 h at room temperature and serially dehydrated with ethanol. Cells were embedded in increasing ratios of LX-112 resin∶ethanol to 100% resin, and polymerized overnight at 60°C. Ultrathin (60 nm) sections were cut on a Leica UC6 microtome and imaged on a JEOL1011 electron microscope at 80 kV. Quantifications were performed as follows: the perimeters of approximately 16 cells (per experimental condition) were imaged and the number of CLICs/GEEC carriers from each cell was quantified and averaged across all 16 cells. The average number of CLICs/GEEC carriers per cell was generated from two separate repeats of the same experimental conditions.
Cells were grown either on 12 mm coverslips or on 35 mm glass bottom dishes (Mat-Tek Corporation) and fluorescence micrographs were captured for random fields containing transfected cells on a confocal laser-scanning microscope (Zeiss 710 META; 510 META; Carl Zeiss Inc.). Images were captured with a 63× plan Apochromat 1.4 NA Oil objective (Zeiss, Jena, Germany), using a 488 nm laser line for excitation and a 505–530 nm band pass emission filter to capture GFP and Alexa Fluor 488 fluorescence, a 561 nm laser line for excitation, and a 580–620 nm band pass emission filter to capture RFP/mCherry/Alexa Fluor 555 fluorescence, and a 633 nm laser line for excitation and a 650 long pass emission filter for emission to capture Alexa Fluor 647 fluorescence. For live cell co-imaging of Cherry and YFP, YFP fluorescence was captured using the 488 nm laser line for excitation and a 505–530 nm band pass for emission. For live cell imaging a region of interest was chosen and 2× digital zoom was applied. Optical path and emission filters were applied as appropriate for each fluorophore. Images were processed with Adobe Photoshop CS3 (Adobe, San Jose, CA) and fluorescent intensity of fluorophores on images was measured using ImageJ (National Institutes of Health, Bethesda, MD). Where indicated, correlation between fluorescence intensities between two fluorophores in different subcellular location was determined using Pearson coefficient analysis. For this, binary masks (values 0 and 1) were created to isolate pixels belonging to different regions of interest (such as the plasma membrane and/or cytosol) by multiplying (pixel by pixel) the mask image by the raw images. This method was chosen in preference to the Pearson analysis over the entire cell as it provides a better estimation of whether or not a high correlation exists between different fluorophores at a specific subcellular location, for example the plasma membrane. Then, each individual image (two channels) was used to calculate the Pearson correlation coefficient between the two fluorophores in each specific subcellular location, and the data represent an average Pearson correlation coefficient determined for at least 40 cells from at least three independent experiments.
To assess GPI-YFP dynamics, cells were seeded on 35 mm glass bottom dishes and transfected with GPI-YFP construct [65]. Images were captured on an inverted confocal microscope using a 63× plan Apo 1.4 NA, Oil objective with 4× digital zoom, with a resolution of 0.15 µm/pixel. A circular region of interest (ROI, 4.7 µm radius) was bleached to ∼70% using the 488 nm argon laser line and 405 nm laser line at 100% transmission. Time-lapse images of the same region and a reference region of identical size were acquired before (20 frames, 5 seconds) and after (300 frames, 90 seconds) photobleaching with an interval of ∼250 milliseconds per frame using the 488 nm laser line at 2% transmission and emission detected between 500 and 530 nm. Half-life time and diffusion rate were calculated as described previously [49],[66]; also see Text S1 for details.
CAV1 dynamics were determined in Cavin-1−/− cells expressing CAV1-YFP alone or co-transfected with CAV1-YFP and Cavin-1-mCherry. Images were acquired as described above and half-times were calculated as described previously for CAV1-GFP in HeLa cells [67],[68]. The lateral diffusion nature of CAV1-YFP recovery curves at the plasma membrane was determined by performing FRAP experiments at variable ROI areas (9.6, 14.4, and 16.6 µm2) to analyze the dependency of half times with area size. Linear regression was performed and the diffusion coefficient for simple lateral diffusion was extracted from the slope as described in Text S1.
To achieve spatially defined photoactivation of PAGFP, a Ti:sapphire two-photon laser (1600–1800 mW Chameleon Ultra, Coherent Scientific) tuned at 775 nm was used [69]. For lateral diffusion studies of CD44, the plasma membrane region of cells was identified with the mRFP signal of CD44-mRFP-PA-GFP. PAGFP fluorescence was activated using a constant circular region of interest (ROI, 1.98 µm radius) by a single scan with infrared laser irradiation (30% transmission), and time-lapse images of photo-activated GFP at the same region were captured every ∼250 millisecond using a 60× plan Apo 1.4 NA, Oil objective at 4× digital zoom and appropriate filter sets to capture GFP fluorescence. For each time point, the average fluorescence intensity was calculated and normalized to the value at the first frame after photo-activation. Decay curves were fitted to a double exponential decay curve and a global half-time of fluorescence decay was obtained numerically as described for FRAP experiments.
CAV1−/− MEFs expressing FRET pair (Cdc42-CyPet and CRIB-YPet) either in presence or absence of FLAG-tagged Cavin-1 were subjected to FLIM microscopy as described previously [18].
Cells were serum starved for 3 h in serum-free DMEM before incubation with 60 µM dynasore/Dyngo-4a. Internalization assay was performed in the presence of 40 µM dynasore for the desired amount of time. Cells were treated with 30 µM 7-ketocholesterol (7-KC) for 30 min at 37°C, followed by the internalization assay.
Whole cell lysates were subjected to SDS-PAGE and further to Western blotting. Membranes were probed with primary antibody at the desired concentration for 1 h at room temperature or overnight at 4°C, followed by incubation with either secondary HRP-conjugated antibodies or infrared dye-labeled Odyssey secondary antibodies for 1 h at room temperature. For detection, either the Licor Odyssey infrared imaging system was used, as per the manufacturer's instructions (Licor Biotechnology), or the SuperSignal West Pico chemiluminescent substrate (Pierce) was captured on film (Kodak). Densitometric analysis of protein bands was performed either by ImageJ or by using Licor Odyssey analysis software. Also see Text S1 for details.
Muscle fibers were isolated from WT and Cavin-1−/− adult mice using a method described previously [70], with modifications. For EM analysis of HRP uptake, isolated fibers cultured overnight on matrigel-coated plastic dishes were incubated in HRP (10 mg/ml) at 37°C for 5 min, washed briefly, then fixed in glutaraldehyde before DAB visualization of the HRP reaction product. Quantitation of the volume of HRP-labeled elements relative to the sampled cytoplasmic volume (volume density) was determined by point counting of peripheral areas of WT and Cavin-1−/− muscle fibers, as shown in previous studies [29]. For procedure details see Text S1.
Statistical analyses were conducted using Microsoft Excel and Prism (GraphPad). Error bars represent either standard error of the mean (SEM) or standard deviation (SD) for at least three independent experiments, as indicated in figure legends. Statistical significance was determined either by two-tailed Student's t-test or by one-way ANOVA, as indicated in the figure legends.
|
10.1371/journal.ppat.0030074 | Centrosomal Latency of Incoming Foamy Viruses in Resting Cells | Completion of early stages of retrovirus infection depends on the cell cycle. While gammaretroviruses require mitosis for proviral integration, lentiviruses are able to replicate in post-mitotic non-dividing cells. Resting cells such as naive resting T lymphocytes from peripheral blood cannot be productively infected by retroviruses, including lentiviruses, but the molecular basis of this restriction remains poorly understood. We demonstrate that in G0 resting cells (primary fibroblasts or peripheral T cells), incoming foamy retroviruses accumulate in close proximity to the centrosome, where they lie as structured and assembled capsids for several weeks. Under these settings, virus uncoating is impaired, but upon cell stimulation, Gag proteolysis and capsid disassembly occur, which allows viral infection to proceed. The data imply that foamy virus uncoating is the rate-limiting step for productive infection of primary G0 cells. Incoming foamy retroviruses can stably persist at the centrosome, awaiting cell stimulation to initiate capsid cleavage, nuclear import, and viral gene expression.
| Naive quiescent CD4-positive T cells or monocytes that are in the G0 stage of the cell cycle cannot be productively infected by retroviruses in vitro, but the molecular basis of this restriction remains poorly understood. In this report, we demonstrate that incoming foamy retroviruses remain around the centrosome as structured and assembled capsids for weeks in resting cultures. Under these conditions, virus uncoating is impaired, but upon cell activation, viral capsids undergo proteolysis and disassembly, allowing infection to proceed. Maintenance of incoming viral capsids at the centrosome in resting cells could be a strategy that viruses have evolved to rapidly respond to stimuli received by the cell. The cellular signal triggering the uncoating process upon cell stimulation remains unclear, but is likely linked to the centrosome cycle.
| After entry into their host cells, retroviruses undergo uncoating and reverse transcription, leading to the formation of pre-integration complexes, which must then gain access to the nuclear compartment and integrate the provirus into host chromosomes [1]. The cell cycle status is a key determinant for completion of these early stages of infection [2], and retroviruses have been classified based on their ability to productively infect non-cycling cells. Gammaretroviruses, like murine leukemia virus, require mitosis for proviral integration [3], while lentiviruses such as human immunodeficiency virus type 1 (HIV-1) show almost no difference between dividing and non-dividing cells [4]. Indeed, HIV-1 and other lentiviruses can replicate in terminally differentiated and post-mitotic cells such as neurons or macrophages [5,6]. However, like murine leukemia virus, HIV-1 cannot infect naive quiescent CD4-positive T cells or monocytes isolated from peripheral blood that are in the G0 stage of the cell cycle [2,7]. Since reverse transcription does occur in these conditions [8], it is conceivable that other host cell proteins or processes are necessary for the completion of the viral cycle [9].
Foamy viruses (FVs) are complex retroviruses isolated from different animal species, mainly in non-human primates. FVs share with all retroviruses the same organization of the genome, which encodes Gag, Pol, and Env proteins. In addition, the FV genome encodes at least two other proteins, Tas and Bet, which are not incorporated into the viral particle. Remarkably, FVs exhibit some features related to the hepatitis B virus [10]. In particular, they reverse transcribe their RNA genome during the late stages of infection, leading to the presence of infectious viral DNA in extracellular virions [11]. Moreover, the structural FV Gag presents specific characteristics that set it clearly apart from other retroviral Gag proteins. In particular, FV Gag maturation by the viral protease does not lead to the formation of the canonical matrix, capsid, and nucleocapsid products. Rather, the Gag precursor is partially cleaved by the viral protease near its C terminus into a mature product, before or during budding. This results in the presence of two Gag proteins of 71 kDa and 68 kDa in extracellular virions [12]. We have previously reported that upon entry into target cells and prior to nuclear translocation, incoming FV capsids traffic along the microtubule network to reach the microtubule-organizing centre (MTOC) [13,14], which includes the centrosome in animal cells [15]. A similar route has been described for HIV-1 [16]. Drugs disrupting the microtubule network, such as nocodazole or colchicine, largely prevent early intracellular FV trafficking [13], as well as that of HIV-1 [16]. Gag also targets the pericentrosomal region for capsid assembly during the late stages of infection [17]. Similar to murine leukemia virus, productive FV infection requires passage through mitosis [18,19]. Like all other animal retroviruses, FVs do not productively infect cells arrested in the G0 stage of the cell cycle, such as peripheral T lymphocytes or growth-arrested fibroblasts in vitro [19]. To gain insight into this restriction, we have focused our attention on early stages of FV replication in G0 cells.
Here, we demonstrate that incoming FVs stably localize at the vicinity of the MTOC as structured and assembled capsids for several weeks in resting cultures in vitro. Upon cell stimulation, Gag proteolysis and capsid disassembly take place, allowing infection to proceed. Altogether, these data demonstrate that the centrosome represents a cellular site around which incoming viruses persist as a stable pre-integration intermediate in resting cells, and that virus uncoating is the rate-limiting step for FV infection in growth-arrested cells.
Cycling and resting human primary MRC5 fibroblastic cells were infected with the prototypic primate foamy virus (PFV) at a multiplicity of infection (m.o.i.) of 1, and cell-free supernatant was collected 48 h later for virus titration. Consistent with previous reports [18,19], we found that PFV does not productively infect G0 resting cells (unpublished data). To investigate the molecular mechanisms involved in the restriction of FV replication cycle in G0 resting cells, we first analyzed the distribution of incoming viral components in resting MRC5 cells. As early as 4 h post-infection (p.i.)., we observed that incoming Gag proteins localized at the centrosome, which was revealed with anti-γ-tubulin antibodies (abs). These findings are consistent with previous observations showing the pericentrosomal concentration of incoming FVs after trafficking along the microtubule network in cycling cells [13]. Remarkably, while in cycling cells, incoming Gag antigens are no longer observed near this organelle 10 h p.i. ([20]; unpublished data); in resting cells, centrosomal localization of viral antigens was consistently observed up to 30 d p.i. (Figure 1A). We then investigated the localization of the viral genome in infected resting cells. By fluorescent in situ hybridization, using the entire PFV DNA genome as a probe, we found that the incoming viral DNA genome also localized at the centrosome in resting MRC5 cells 15 d p.i. (Figure 1B). Therefore, both incoming FV antigens and the viral DNA genome reside at the MTOC of resting cells for several weeks after infection.
To visualize the status (assembled capsids or not) of incoming viruses in resting cells, MRC5 cells were analyzed by electron microscopy (EM) at different time points after infection. In cycling cells (Figure 2A), incoming FVs were observed at the centrosome 4 h p.i., mainly as structured and assembled capsids, confirming that uncoating is not complete at this time point [20]. At later time points, these viral capsids completely disassembled in cycling cells as already reported [20], and viral capsids were never detected in uninfected cells (Figure 2A). Importantly, assembled and structured viral capsids were observed around the centrosome in resting cells 5 or 15 d p.i. (Figure 2A), strongly suggesting that virus uncoating is impaired under these settings.
FV uncoating requires the enzymatic activity of both viral and cellular proteases, which cleave the major structural components of viral capsids (the 71–68 kDa Gag doublet), into shorter fragments [20]. Among these cleavage products, a 38-kDa Gag-derived product specifically results from the action of the viral protease [20]. To confirm at the biochemical level that FV uncoating is inhibited in resting cells, the status of the Gag polyprotein was determined. For that purpose, resting and cycling MRC5 cells were infected with PFV at an m.o.i. of 1, and intracellular viral proteins were analyzed by Western blot using mouse anti-Gag abs. Figure 2B shows that Gag cleavage products, in particular the 38-kDa fragment, were easily detected in cycling cells. On the contrary, FV Gag was not cleaved following infection of resting MRC5 cells and remained as an intact doublet of 71 and 68 kDa (Figure 2B). These results demonstrate the absence of Gag cleavage in resting cells, reflecting the absence of viral uncoating observed by EM.
To assess whether incoming FV capsids at the centrosome of resting cells constitute a stable pre-integration intermediate, which can later be reactivated for productive infection, resting PFV-infected MRC5 cell cultures were stimulated to divide by splitting and serum addition. At different time points following this activation, Gag expression and distribution were analyzed by indirect immunofluorescence using mouse anti-Gag abs. Twenty-four hours after cell activation, we observed that, although Gag still associated to the centrosome, it could be detected in both the cytoplasm and the nucleus of 20% of the cells (Figure 3A). Gag was detected in both compartments in the entire culture 48 h after cell activation, and the formation of numerous syncytia was detected 96 h post-activation (Figure 3A). These results demonstrate that viral replication, which was inhibited in resting cells, resumes upon cell activation. To confirm that cell activation actually triggered virus uncoating, the status of the Gag polyprotein was studied by Western blot. To exclusively analyze incoming viral antigens, avoiding contamination from Gag synthesis and degradation, which might occur following reactivation, these experiments were performed under cycloheximide (CHX) treatment, a translation inhibitor. At 24 h post-reactivation, several Gag cleavage products, notably the 38-kDa fragment [20], were clearly detected in CHX-treated cells (Figure 3B). Moreover, under these settings, accumulation of Gag was observed only in untreated reactivated cells. Altogether, these observations demonstrated that entry into the cell cycle triggered virus uncoating, as assessed by Gag cleavage, and productive infection.
Several reports have demonstrated that FVs infect lymphocytes in vivo and that infectious particles can be recovered from peripheral T cells of infected animals [21−24]. Therefore, the intracellular distribution of incoming FVs was assessed in one of its natural targets. To this aim, primary resting human CD4-positive T cells were infected with PFV at an m.o.i of 1 and were maintained in a resting state in culture for 5 d without addition of exogenous lymphokines. PFV infection of resting CD4-positive T cells was non-productive (unpublished data) and did not trigger cell activation (Figure 4A). In these cells, localization of incoming capsids was analyzed by confocal microscopy. Consistent with our previous observation of infected resting MRC5 cells, incoming Gag strictly localized at the centrosome from day 2 to day 5 p.i. (Figure 4B). On the contrary, Gag was diffusely distributed in the cytoplasm and the nucleus of activated CD4-positive T cells, indicating that productive replication was taking place in these cells (Figure 4B) [25]. When resting CD4-positive T cells were stimulated to divide 5 d p.i., Gag was no longer observed uniquely at the centrosome but localized diffusely in the cytoplasm and the nucleus at 24 h post-activation (Figure 4B). Taken together, these results confirm that a stable PFV intermediate can persist in the vicinity of the MTOC in primary resting CD4-positive T cells and that T cell activation allows completion of the viral replication cycle.
For all retroviruses, completion of the early steps of the replication cycle depends on the cell cycle status (reviewed in [2,9]). We show that in vitro FV infection is restricted in G0 resting cells, either from fibroblastic (MRC5) or lymphocytic (CD4-positive T cells) origin. We further demonstrate that incoming viral capsids persist in infected G0 resting cells as a stable pre-integration intermediate over a period of at least 30 d in MRC5 cells. Under these conditions, Gag proteolysis, and consequently virus uncoating, is blocked, and incoming viruses are maintained as assembled and structured capsids around the centrosome. Upon cell activation, Gag is cleaved, viral capsids disassemble, and infection proceeds.
Post-mitotic cells such as neurons or macrophages are productively infected by lentiviruses. In contrast, resting G0 cultures in vitro, such as naive T lymphocytes isolated from peripheral blood, cannot be productively infected by any classes of retroviruses, including HIV-1 [8,26−30]. Since reverse transcription is completed in these cells [8,28], additional blocks seem to occur during the early stages of the virus life cycle. Several hypotheses have been raised to elucidate the molecular basis of this restriction. It has been suggested that APOBEC3G, a cellular antiretroviral protein that is associated with the hypermutation of viral DNA through cytidine deamination [31], could inhibit HIV replication as part of a low molecular mass ribonucleoprotein complex in resting T cells. This seems to impair the formation of HIV-1 late reverse transcription products [29]. A recent alternative hypothesis suggests that virus uncoating represents the main rate limiting stage in resting T CD4+ cells, since cellular extracts from activated, but not resting cells, support uncoating of HIV cores in vitro [30,32]. Clearly, our observations suggest this second scenario in the case of FVs. Interestingly, our observations might explain the efficient in vivo transduction of haematopoietic stem cells by FV-derived vectors in mice [33,34]. Indeed, despite the fact that FVs cannot productively infect resting stem cells in vitro, FV vectors can repopulate bone marrow [33−36]. In fact, in vivo implantation of transduced resting stem cells likely triggers their activation, allowing the infection to proceed.
The remarkable stability of the FV intermediate in resting cells that we have evidenced here could be related to the particular mode of replication of these viruses. First, virus uncoating is a relatively late event during FV replication, allowing incoming viral antigens to accumulate in close vicinity to the nuclear compartment as assembled and structured capsids [13,20]. On the contrary, for HIV-1, capsid disassembly in activated and cycling cells seems to start as soon as the viral particles enter into the cytoplasm [16,37]. Second, in contrast to other retroviruses, the presence of an infectious viral DNA genome in incoming capsids could make FV less dependent on the metabolism of the target cell [11]. Our data also demonstrate that the centrosome is a cellular site around which incoming FVs can stably persist, awaiting further cell stimulation for completion of the viral cycle. Recent studies have shown that the centrosome is not a mere spectator of the cell cycle but can exert significant control over it [38]. By providing a scaffold for many cell cycle regulators and their activities [38−40], the centrosome influences cell cycle progression, especially during the transition from G1 to S phase [41,42]. Therefore, this organelle receives and integrates signals from outside the cell and facilitates conversion of these signals into cellular functions. Maintenance of viral capsids at the vicinity of the centrosome in resting cells could be a strategy that some viruses have evolved to rapidly respond to growth stimuli received by the cell. The cellular signal(s) triggering the uncoating process upon cell stimulation remains unclear, but is likely linked to the centrosome cycle.
Human MRC5 fibroblasts were cultured as described [14]. Quiescent MRC5 fibroblasts were generated following confluence, serum starvation, and addition of 10−6 M dexamethasone. The cells were cultured in these conditions for 10 d before infection. The cell cycle status was analyzed by pyronine staining as described by [43]. In resting MRC5 cells, less than 3% of cells remain activated, as already reported ([44]; unpublished data). For cell activation, cells were stimulated to divide by subculture and serum addition at different times after infection. CHX treatment was performed at 150 μg/ml immediately after cell activation and maintained during the entire experiment.
Peripheral blood mononuclear cells obtained from healthy donors after informed consent were separated on lymphocytes separation medium (Eurobio, http://www.eurobio.fr). CD4-positive T cells were negatively selected using the CD4-positive T Cell Isolation Kit II (Miltenyi Biotec, http://www.miltenyibiotec.com). Briefly, cells were labeled using a cocktail of biotin-conjugated abs against CD8, CD16, CD19, CD36, CD56, CD123, TCRγδ, and glycophorin A, and anti-biotin magnetic beads. After washing, negative cells were selected by magnetic separation with the autoMACS Separator (Miltenyi Biotec). The cells were then labeled with FITC-anti-CD8 (clone SK1; BD Biosciences, http://www.bdbiosciences.com), anti-CD14 (Clone TUK4, Miltenyi Biotec), Phycoerythrin-conjugated anti-CD25 (clone 4E3, Miltenyi Biotec), and anti HLA-DR (L243, BD Biosciences) ab, washed, and sorted on a FACSVantage. The purity was examined by flow cytometry with the same ab and an allophycocyanin-conjugated anti-CD4 (clone RPA-T4, BD Biosciences). Typically, cells were more than 99% negative for the activation markers (CD25 and HLA-DR) and more than 98% positive for the CD4. Activated CD4-positive T cells were prepared by incubating negatively selected CD4 T cells in RPMI 10% normal human serum and 1 μg/ml PHA-L (Sigma, http://www.sigmaaldrich.com). At day 3, 150 UI/ml of interleukin 2 (IL-2; Promocell, http://www.promocell.com) was added to the cells. Activation status of the control and infected CD4 T cells were checked by flow cytometry with the ab anti-CD4 FITC and an anti-CD25 or HLA DR PE ab.
MRC5 cultures or CD4-positive T cells were infected by spinoculation at an m.o.i. of 1 for 1 h 30 min at 30 °C. Titres were determined by infection using FAG cells (BHK cells stably harbouring the GFP gene under the control of the PFVU3 promoter) [45].
MRC5 cells grown on glass coverslips were infected with PFV at an m.o.i. of 1 by spinoculation for 1 h 30 min at 30 °C. After different times p.i., cells were rinsed with PBS, fixed for 10 min at 4 °C with 4% PFA, and permeabilized for 5 min at −20 °C with ice-cold methanol. Cells were incubated successively with mouse polyclonal anti-Gag serum overnight at 4 °C (1/250) and with rabbit polyclonal γ-tubulin antiserum (1/2000; Abcam, http://www.abcam.com) for 1 h at 37 °C. Cells were washed and incubated for 1 h with a 1/500 dilution of appropriate fluorescent-labeled secondary abs. Nuclei were stained with 4'-6-diamidino-2-phenylindole (DAPI) and the coverslips were mounted in Moviol. Confocal microscopy observations were performed with a laser-scanning confocal microscope (LSM510 Meta; Carl Zeiss, http://www.zeiss.com) equipped with an Axiovert 200 M inverted microscope using a Plan Apo 63_/1.4-N oil immersion objective.
Fifteen days following PFV infection (m.o.i. of 1), arrested MRC5 cells were fixed with 4% PFA, permeabilized with 0.2% Triton X-100, and incubated with γ-tubulin antiserum (Abcam). Then, cells were fixed a second time in 4% PFA for 20 min at room temperature to cross-link bound abs. Incubation with the secondary ab was performed during the fluorescent in situ hybridization detection step, performed as described [20]. Briefly, cells were treated with RNase at 100 μg/ml in PBS for 30 min at 37 °C and incubated with probe (plasmid p13 containing the entire PFV genome) overnight at 37 °C. Probes were labeled with FITC-avidin DN (1/200; Vector Laboratories, http://www.vectorlabs.com) and signals were amplified with biotinylated anti-avidin D (1/500, Vector Laboratories), followed by another round of FITC-avidin staining. Finally, cells were stained for DNA with DAPI and mounted in Vectashield. Confocal observations were performed as previously described.
Cell pellets were lysed in Triton buffer (10 mM Tris [pH 7.4]; 50 mM NaCl; 3 mM MgCl2; 1 mM CaCl2; orthovanadate, benzamidine, and protease inhibitor cocktail [Roche, http://www.roche.com] at 1 mM each; 10 mM NaF; and 0.5% Triton X-100) for 30 min at 4 °C and centrifuged for 15 min at 20,000g. Resulting pellets were treated with radioimmunoprecipitation buffer (10 mM Tris [pH 7.4]; 150 mM NaCl; orthovanadate, benzamidine, and protease inhibitor cocktail at 1 mM each; 10 mM NaF; 1% deoxycholate; 1% Triton X-100; and 0.1% sodium dodecyl sulfate [SDS]) during an additional 30 min at 4 °C, centrifuged for 15 min at 20,000g, collected, and diluted in Laemmli buffer. Samples were migrated on an SDS–10% polyacrylamide gel, and proteins were transferred onto cellulose nitrate membrane (Optitran BA-S83; Schleicher-Schuell, http://www.schleicher-schuell.com), incubated with appropriated abs, and detected by enhanced chemiluminescence (Amersham ECL Advance Western Blotting Detection Kit, http://www.gelifesciences.com).
For EM studies, MRC5 cells were infected at an m.o.i. of 5 as described above and fixed in situ by incubation for 48 h in 4% PFA and 1% glutaraldehyde in 0.1 M phosphate buffer (pH 7.2); they were then post-fixed by incubation for 1 h with 2% osmium tetroxide (Electron Microscopy Science, http://www.emsdiasum.com). Next, MRC5 cells were dehydrated in a graded ethanol series, cleared in propylene oxyde, and then embedded in Epon resin (Sigma), which was allowed to polymerize for 48 h at 60 °C. Ultrathin sections were cut, stained with 5% uranyl acetate/5% lead citrate, and then placed on EM grids coated with collodion membrane. They were then observed with a Jeol 1010 transmission electron microscope (Jeol, http://www.jeol.com). |
10.1371/journal.pcbi.1002226 | Phosphorylation of the Arp2 Subunit Relieves Auto-inhibitory Interactions for Arp2/3 Complex Activation | Actin filament assembly by the actin-related protein (Arp) 2/3 complex is necessary to build many cellular structures, including lamellipodia at the leading edge of motile cells and phagocytic cups, and to move endosomes and intracellular pathogens. The crucial role of the Arp2/3 complex in cellular processes requires precise spatiotemporal regulation of its activity. While binding of nucleation-promoting factors (NPFs) has long been considered essential to Arp2/3 complex activity, we recently showed that phosphorylation of the Arp2 subunit is also necessary for Arp2/3 complex activation. Using molecular dynamics simulations and biochemical assays with recombinant Arp2/3 complex, we now show how phosphorylation of Arp2 induces conformational changes permitting activation. The simulations suggest that phosphorylation causes reorientation of Arp2 relative to Arp3 by destabilizing a network of salt-bridge interactions at the interface of the Arp2, Arp3, and ARPC4 subunits. Simulations also suggest a gain-of-function ARPC4 mutant that we show experimentally to have substantial activity in the absence of NPFs. We propose a model in which a network of auto-inhibitory salt-bridge interactions holds the Arp2 subunit in an inactive orientation. These auto-inhibitory interactions are destabilized upon phosphorylation of Arp2, allowing Arp2 to reorient to an activation-competent state.
| The Arp2/3 complex consists of seven associated protein subunits including Arp2 and Arp3 that play a central role in the formation of actin filaments. Filament formation by the Arp2/3 complex drives important cell processes such as cell movement and endocytosis. The function of the Arp2/3 complex is highly regulated, and improper regulation of its activity has been linked to cancer metastasis. One level of regulation is post-translational phosphorylation, in which a −2 charged phosphate group is added to the uncharged amino acids threonine 237 and 238 of Arp2. We use molecular dynamics simulations and biochemical studies to show that Arp2 phosphorylation results in large structural changes of the Arp2/3 complex consistent with low-resolution structural studies. The simulations suggest phosphorylation allows the complex to reorient to an activation competent state by destabilizing interactions that hold Arp2 in an inactive position. Further simulations suggested that mutation of the Arp2/3 complex could allow complex activation, and we verified this gain-of-function mutation biochemically. We propose a model for Arp2/3 complex activation in which phosphorylation destabilizes the inactive state of the complex, allowing structural changes that are permissive for activation by nucleation-promoting factors and binding to the mother filament.
| Spatial and temporal control of the assembly and disassembly of actin filaments is crucial for a number of distinct cell processes, including endocytosis and cell migration [1]. The spontaneous assembly of actin filaments from a pool of actin monomers requires the formation of an unstable actin trimer nucleus, from which further polymerization is thermodynamically favorable [2]. Fast filament assembly is achieved by several classes of proteins that act as actin nucleators, constituting one level of regulation. Formins and the spire proteins nucleate unbranched filaments [3], [4] and the Arp2/3 complex facilitates assembly of branched filaments by nucleating a new “daughter” filament from the side of an existing “mother” filament [5], [6]. Branched filament networks generated by the Arp2/3 complex are required to build many cellular structures, and Arp2/3 complex is the primary nucleator of new actin filaments in most crawling cells (reviewed in [1], [7], [8]). Aberrant Arp2/3 complex function has been implicated in a number of disease conditions, most notably cancer metastasis [7], [9].
Direct regulation of the nucleating activity of the Arp2/3 complex leads to a second level of control over actin filament assembly. The Arp2/3 complex is composed of seven subunits: the actin-related proteins Arp2 and Arp3, and ARPC1–5. While binding of the mother filament contributes significantly to activation [10], full activity of the Arp2/3 complex also requires ATP binding to Arp2 and Arp3 [11], [12] and binding of a nucleation promoting factor (NPF), such as WASP [13], [14], N-WASP [15], SCAR/WAVE [16], [17], and the pathogenic proteins ActA from Listeria monocytogenes [18] and RickA from Rickettsia [19], [20]. NPF binding to actin monomers facilitates the nucleation reaction, and NPFs couple Arp2/3 complex activity to that of Rho-family GTPases [21], [22].
The structures of the apo and nucleotide-bound states of Arp2/3 were revealed by X-ray crystallography [23], [24], [25], and these lead to the hypothesis that activation required large structural changes [25]. The structure of the active Arp2/3 complex at the junction of the mother filament and the newly nucleated daughter filament (the branch junction) was recently revealed in reconstructions from electron micrographs of negatively-stained specimens [26], [27]. In support of the hypothesized structural changes, docking of the inactive Arp2/3 complex crystal structure into the branch junction density revealed substantial rearrangements of subunits, particularly of the Arp2 and ARPC3 subunits [27]. For the Arp2/3 complex to incorporate into the daughter filament and increase filament assembly, Arp3 and Arp2 appear to undergo a large change in their relative orientation from their arrangement in the inactive crystal structure to their conformation in the branch junction density, in which they appear to mimic the short-pitch of an actin dimer [25], [27].
An additional, more recently identified requirement for activating nucleation by the Arp2/3 complex is threonine or tyrosine phosphorylation of the Arp2 subunit [28]. Mass spectrometry of purified Arp2/3 complex revealed phosphorylation of Arp2 T237 and T238. Although no phosphorylated tyrosine was identified by mass spectrometry, mutagenesis studies suggested Arp2 Y202 as the likely phosphorylation site. Mapping of the Arp2 phosphorylation sites onto the crystal structure reveals that these residues are near the interface of Arp2, Arp3, and ARPC4, and we predicted that phosphorylation of these residues could play a role in the large conformational changes predicted upon activation [28]. Consistent with this prediction, our biochemical assays suggested that Arp2 phosphorylation primes the complex for activation to allow conformational changes predicted to be necessary for activation [28]. However, the mechanism by which phosphorylation permits activation of the Arp2/3 complex remains poorly understood.
Computational studies have the potential to elucidate aspects of Arp2/3 complex function and regulation. While the impact of computation in this regard has been limited thus far due to the large system size, molecular dynamics simulations have been used to examine dynamics of the ATP binding cleft in Arp2 and Arp3 [29], [30]. In addition, homology modeling of the structures of the Arp2/3 complex from different species has generated hypotheses about functionally important surfaces [31]. Recently, steered molecular dynamics simulations were used to investigate potential pathways of Arp2/3 complex in the absence of phosphorylation [32], and molecular dynamics and protein-protein docking was used to generate a model of mother filament bound to the Arp2/3 complex, which was then validated experimentally [33]. Computational methods, molecular dynamics methods in particular, have also been used previously to study conformational changes of other proteins upon phosphorylation (reviewed in ref. [34]). Examples include the study of structural changes caused by phosphorylation in the activation and glycine-rich loops of protein kinases [35], [36], [37], [38], changes in peptide conformations [39], [40], and in membrane proteins such as phospholamban [41].
Here, we use unbiased molecular dynamics simulations to determine how phosphorylation at Arp2 T237 and T238 may change the structure of the Arp2/3 complex and permit activation by NPFs. We find large conformational changes in the Arp2/3 complex upon phosphorylation, including the reorientation of Arp2 relative to Arp3, toward the short-pitch dimer orientation. Our simulations suggest a mechanism by which a complex network of positively and negatively charged amino acids at the Arp2/Arp3/ARPC4 interface holds the complex in an inactive configuration, and phosphorylation disrupts these auto-inhibitory interactions. To test this prediction, we designed, based on further computational simulations, mutations of the Arp2/3 complex that we predicted would disrupt the auto-inhibitory interactions. Biochemical assays reveal that this mutant, R105/106A ARPC4, does in fact show nucleation activity even in the absence of NPFs.
We previously reported that phosphorylation of Arp2 T237/238 or Y202 is necessary for activation of the Arp2/3 complex in the presence of NPF [28]. Phosphorylation of T237/238 in endogenous Arp2 was confirmed by mass spectrometry, and heterologous expression of Arp2 with alanine substitutions in T237/238 and Y202 inhibits membrane protrusion. We tested nucleation activity of a mutant Arp2/3 with these residues mutated to alanine using recombinant Arp2/3 complex generated in a baculovirus expression system and purified as previously described [42]. We mutated all three residues because we previously showed that phosphorylation of these sites acts as a logical ‘or gate’ with either being necessary for activation [28]. Subunits of the wild type (WT) and mutant Ala-substituted T237/238-Y202 Arp2 (T237/238A-Y202A) Arp2/3 complex were expressed independently in Spodoptera frugiperdas (Sf21) insect cells, and were confirmed to assemble the seven-subunit complex with equimolar stoichiometry (Fig. 1a). We also confirmed that binding to NPF (N-WASP-VCA) (Kd = 0.5 µM) and actin filaments (Kd = 1.2–1.3 µM) is similar for WT and T237/238A-Y202A Arp2 rArp2/3 complex (Fig. S1a,b).
The rate of assembly of pyrene-labeled actin into filaments, an index of Arp2/3 complex nucleation activity, was similar with WT rArp2/3 (0.193 nM filament ends) and with Arp2/3 complex purified from bovine thymus (0.209 nM filament ends at concentrations of 5 nM Arp2/3 complex with 4 µM actin) (Fig. 1b and Fig. S2a). In the presence of NPF (C-terminal VCA domain of N-WASP), assembly rates increased 19-fold and were 3.30 nM and 2.51 nM filament ends for recombinant and native Arp2/3 complex respectively (Fig. S2a). We previously showed that native Arp2/3 complex from bovine thymus and WT rArp2/3 purified from insect cells are phosphorylated, and that native Arp2/3 complex pretreated with the dual specificity alkaline phosphatase Antarctic phosphatase (AP) is not activated by NPFs [28]. We confirmed that nucleation by native Arp2/3 and WT rArp2/3 complex in the presence of NPFs was reduced after treatment with AP to levels similar to untreated rArp2/3 complex in the absence of NPF (Fig. 1b and Fig. S2a). These data indicate that the activity of WT rArp2/3 in the absence and presence of NPF is similar to that of native Arp2/3 complex and that dephosphorylation inhibits NPF-induced activity.
In the absence of NPF, the rate of actin filament assembly and the concentration of filament ends with mutant rArp2/3 complex containing T237/238A-Y202A Arp2 (t1/2 = 500 s, 0.260 nM) were similar to native and WT complexes (Fig. 1c and Fig. S2b). In the presence of NPF, although the rate of filament assembly with the mutant decreased to t1/2 = 348 s, it was 4-fold slower than the rate of t1/2 = 84 s with WT rArp2/3 and similar to that of NPF-stimulated WT rArp2/3 complex pretreated with AP. The concentration of filament ends for AP-treated Arp2/3 complex and the mutant was similar in samples with or without NPF (Fig. S2a and S2b). Pretreating the mutant with AP completely blocked the increased rate of filament assembly in the presence of NPF, and rates were similar to WT and mutant in the absence of NPF. These findings indicate that phosphorylation of T237/238 or Y202 of the Arp2 subunit is necessary for maximal nucleation activity of the Arp2/3 complex. However, mutant rArp2/3 complex containing T237/238A-Y202A Arp2 retains some residual activity that is abolished by AP.
To nucleate a new actin filament, the Arp2/3 complex binds to filament pointed ends, which reflects its capping activity. We used actin seeds capped at the barbed end with gelsolin to measure pointed end capping by WT and mutant Arp2 rArp2/3 complex. Gelsolin-capped actin filaments elongated from their pointed ends in the presence of 4 µm actin (Fig. 1d). Addition of WT rArp2/3 complex slowed filament assembly from the pointed end. The rArp2/3 complex containing mutant Arp2 also slowed filament assembly, although markedly less than with WT. The measured affinity of the mutant for the pointed end decreased approximately 8-fold (Table S1). These data suggest that phosphorylation of T237/238 or Y202 in Arp2 is necessary for rArp2/3 complex to efficiently bind the pointed end of actin filaments.
We previously hypothesized that phosphorylation may induce a conformational change that allows activation by mother filament and NPF [28]. To test this hypothesis, we performed molecular dynamics simulations on phosphorylated and unphosphorylated wild-type Arp2/3 complex, using the inactive conformation observed in the crystal structure [25] as a starting point (Fig. 2a). We reasoned that, if our hypothesis were correct, the unphosphorylated wild-type structure should remain relatively unperturbed during the molecular dynamics simulation, while phosphorylated Arp2/3 should show conformational changes caused by the strong electrostatic perturbation associated with introducing the phosphate groups.
Due to the large size of the Arp2/3 complex, the simulations required substantial computational resources on a supercomputer. We performed simulations only with Arp2 T237 and T238 individually phosphorylated, as well as the unphosphorylated ‘control’ simulation. For each of these systems, we ran duplicate 30 ns molecular dynamics simulations to control for simulation dependence on the initial conditions and stochastic fluctuations. Observing large conformational changes using molecular dynamics simulations is difficult because of the gap between experimental and computationally feasible timescales, and we do not expect to see the full range of structural change in these simulations. Nonetheless, large conformational changes were observed for Arp2/3 when phosphorylated on either T237 or T238 of Arp2. Backbone root mean square deviations (RMSDs) following global alignment of simulation snapshots to the starting model revealed modest conformational changes of 3–4 Å RMSD for the unphosphorylated simulations compared with larger conformational changes of 4–8 Å for the phosphorylated simulations (Fig. 2b). In general, the directionality of the conformational changes relative to the unphosphorylated simulations with pT237 (phosphorylated T237) or pT238 were qualitatively similar, as were the results in the two duplicate simulations for each system (Fig. S3b). The conclusions we draw are supported by all of the simulations, although the precise details of the dynamical behaviors differed. The unphosphorylated simulations show convergence over the last 20 ns of simulation time. Therefore, the last 20 ns were used in the analyses below. This convergence with regard to simulation time does not indicate equilibrium convergence; it is possible and even likely that the full range of structural changes in the phosphorylated simulation in particular have not been realized (see Discussion). It should also be noted that these structural changes were not due to artifactual steric effects of adding phosphate groups to construct the phosphothreonine side-chains. The phosphate groups were added without causing any steric clashes with surrounding residues (data not shown). The starting structure of Arp2 subdomains 1 and 2, which are disordered in all but one crystal structure, which was stabilized with glutaraldehyde [24], were homology modeled based on the actin monomer structure (PDB 1ATN [43]).
The conformational changes induced by phosphorylation were dominated by changes in the orientation of Arp2, ARPC1, and ARPC3 relative to other subunits (Fig. S3). In particular, phosphorylation induced motion of the Arp2 subunit relative to the Arp3 subunit toward its active position as a mimic of an actin short-pitch dimer [27], a conformation required for polymerization of actin (Fig. 3A and Fig. S4). To quantify this motion, we used the model of active Arp2/3 obtained by orienting Arp2 and Arp3 as in an actin short-pitch dimer, with no changes in the structure or orientation of other subunits [B. Nolen, personal communication]. Specifically, we computed the Cα root mean square deviation (RMSD) of the Arp2 subunit between individual snapshots over the last 20 ns and Arp2 in the starting (Fig. 3b) or active orientation (Fig. 3c) with respect to Arp3 following alignment of experimentally resolved Cα atoms of Arp3 subdomains 1 and 2. Phosphorylation of either T237 or T238 caused conformational changes away from the starting inactive state relative to that observed in the simulation of unphosphorylated Arp2/3. Phosphorylation also appears to cause conformational changes that lower the RMSD of the Arp2 subunit to the active orientation, although further large conformational changes are required for full activation (Fig. 3d). This may be in part to the relatively short timescale of the simulations. However, phosphorylation is not expected to induce a conformational change to the fully active state because our biochemical determinations (Fig. 1b and [28]) indicate phosphorylation is necessary but not sufficient for full activation, which also requires binding of the mother filament and NPFs.
Analyzing the Cα RMSD of each residue of Arp2 after alignment of Arp3 subdomains 1 and 2 reveals that the largest changes in Cα position of Arp2 occur in the C-terminal tail as well as in subdomains 1 and 2, which are disordered in most unphosphorylated Arp2/3 complex crystal structures (Fig. S5a) [23], [24], [25]. Larger structural changes are induced in the phosphorylated simulations than in the unphosphorylated simulations across the entirety of Arp2. Substantial increases in the Arp2 per-residue RMSD after alignment of Arp3 subdomains 1 and 2 compared with those after alignment of the Arp2 Cα atoms suggest that the motion of Arp2 largely consists of a rigid-body movement (Fig. S5b). Phosphorylation induces the loss of contacts between Arp2 and Arp3 subunits, potentially allowing the reorientation of these subunits (Fig. 2e).
The results of the molecular dynamics simulations are consistent with the hypothesis that phosphorylation induces conformational changes that contribute to adopting a nucleation-competent form [28]. We examined the vicinity of the phosphorylation sites in detail to identify the interactions mediating these structural changes. In the unphosphorylated state, T237 and T238 are located near the interface between Arp2 and ARPC4, and are also close to the interfaces with Arp3 and ARPC2. The interactions between Arp2 and ARPC4 in the vicinity of T237/238 are dominated by salt bridges, several of which are highlighted in Fig. 4a and Fig. S6a. In particular, the complex electrostatic network involves E39, R71, E99, R105, R106, and K107 on ARPC4; E236 and K232 on Arp2; and R409 and E121 on Arp3. Unphosphorylated T237 and T238 do not participate in the side-chain hydrogen-bonding network.
By contrast, these interactions were dramatically rearranged with phosphorylation of either T237 or T238 (Fig. 4b and Fig. S6b). LeClaire, et al. hypothesized that R105 and R106 of ARPC4 mediate the effects of phosphorylation at T237 and T238 of Arp2, respectively [28]. The molecular dynamics simulations support this hypothesis but also suggest a much more complex rearrangement of the electrostatic network driven by introduction of the phosphate charge. The strengths of salt bridge interactions between phosphorylated amino acids and lysine/arginine side chains are stronger than those between aspartate/glutamate and lysine/arginine, with phosphate-arginine interactions being particularly stable [44]. Thus, unsurprisingly, the phosphorylated amino acids form both transient and stable interactions with arginine residues in the simulations, such as between pT237 of Arp2 and R105 of ARPC4 (Fig. 4b) and between pT238 of Arp2 and R106 of ARPC4 (Fig. S6b). The incorporation of pT237 and pT238 into the electrostatic network necessitates that other salt bridging interactions are disrupted and a new set of interactions are formed, either as a direct consequence or as an indirect result of the induced conformational changes (Fig. S7).
Because T237 and T238 are near the interfaces with several other subunits, perturbations to the electrostatic network induced by phosphorylation can cause large conformational changes. We hypothesized that mutating key residues that interact with pT237 and pT238 would abolish the ability of phosphorylation to induce these conformational changes, and hence activation of the nucleation activity. This hypothesis is based on studies of phosphorylation-mediated activation in systems such as protein kinases, in which attractive interactions with arginine residues that interact with phosphorylated residues drive conformational changes key to phospho-activation [45]. In particular, the simulations predicted that R105 of ARPC4 would form a specific and stable ion pair with the phosphate on T237 of Arp2. To test this hypothesis, we constructed R105A ARPC4 mutants with unphosphorylated and phosphorylated T237 in silico, and generated two independent 30 ns molecular dynamics simulations for each.
Contrary to our expectations, simulations of the unphosphorylated Arp2/3 complex with the R105A ARPC4 mutation produced a similar, but somewhat smaller, structural change to that induced by phosphorylation at T237 or T238 Arp2 (Fig. 5a and 6). This result suggested that the R105A ARPC4 mutation could allow partial activation of Arp2/3 even in the absence of phosphorylation. Phosphorylation of T237 Arp2 in the context of the R105A ARPC4 mutant produced even larger structural changes than phosphorylation of T237 or T238 Arp2 alone (Fig. 5b and 6). It should again be noted that the structures at the end of these simulations likely do not represent the full range of conformational change of these complexes due to the short timescales available to MD simulation. However, the large conformational changes away from the inactive, initial state observed at these short timescales are similar to the conformational changes caused by phosphorylation, suggesting increased activity of these mutants.
To test predictions from our molecular dynamics simulations on the role of arginines in ARPC4, we generated rArp2/3 complex with R105 and R106 of ARPC4 mutated to alanine (R105/106A ARPC4). While we did not simulate structural changes associated with the R106A ARPC4 mutation, the simulations indicated that R106 ARPC4 upon T238 phosphorylation played a role analogous to that of R105 ARPC4, forming a stable interaction with the phosphate group (Fig. S6b). The mutant R105/106A ARPC4 rArp2/3 complex purified from Sf21 cells showed subunit stoichiometry (Fig. 7a), binding to NPF (Kd = 0.06 µM), and binding to actin filaments (Kd = 1.5 µM) (Fig. S1) similar to WT. In the absence of NPF, the rate of actin filament assembly was markedly faster with rArp2/3 complex containing R105/106A ARPC4 (t1/2 = 173 s) than with WT (t1/2 = 539 s) (Fig. 7b) and there was 2.5-fold more filament ends with 4 µM actin and 5 nM Arp2/3 complex (Fig. S8), indicating that the mutant is constitutively more active. In the presence of NPF, the rate of assembly for the mutant (t1/2 = 71 s) and WT (t1/2 = 84 s) was similar, as was the number of filament ends (Fig. S8), revealing that the mutation has no effect on maximal NPF-induced Arp2/3 complex activity. Preincubating the mutant with AP decreased actin nucleation rates but only slightly (Fig. 7b), and much less than with AP treatment of WT complex. These data and our molecular dynamics simulations suggest that mutating R105/106 disrupts auto-inhibitory interactions and releases the Arp2/3 complex structure to a conformation that is permissive for nucleation activity.
Pointed end capping by the R105/106A ARPC4 rArp2/3 complex appeared similar to WT rArp2/3. Using gelsolin-capped actin seeds, the rate of pointed end elongation was not significantly different between mutant and wild-type Arp2/3 complex, and treating with AP only slightly attenuated capping activity of the mutant (Fig. 7c and Table S1). Hence, the R105/106A ARPC4 mutation likely causes structural differences similar to those seen in activated WT rArp2/3 complex, and phosphorylation of the R105/106A mutant enhances this effect, consistent with the simulations of the R105A ARPC4 mutant.
Our simulations revealed large structural differences between the phosphorylated or mutant complexes and the unphosphorylated wild-type complex. These structural changes included the movement of Arp2 toward the active short-pitch dimer orientation relative to Arp3. However, the conformational changes observed in these short molecular dynamics simulations are much smaller than those assumed to occur upon full activation, and even with longer simulations we would not expect the phosphorylated complex to adopt the putative active conformation, since phosphorylation is necessary but not sufficient for activation. Rather, we propose a model in which phosphorylation destabilizes the inactive state, leading to conformational changes that relieve the auto-inhibition and thus are permissive for full activation by NPF binding.
It is impossible to say whether the actual structural differences in response to phosphorylation or mutation are realized at the end of these simulations, but this is very unlikely. Each of the duplicate simulations of the same state of the complex show differences, even the dual simulations of the unphosphorylated state (see, for example, the two peaks in the Arp2 RMSD to the active orientation (Fig. 3d)), indicating that equilibrium convergence has not been achieved even for the unphosphorylated, wild-type complex. Consequently, we believe the structures at the end of these simulations simply suggest that either phosphorylation or mutation induces large conformational changes that shift Arp2 towards the active short-pitch dimer orientation – they are not a prediction of the structure of the Arp2/3 complex upon phosphorylation or mutation.
Besides the changes in the Arp2-Arp3 orientation, large changes in the orientation of ARPC1 and ARPC3 relative to Arp3 were observed upon phosphorylation (Fig. S3). We can only speculate about the mechanism by which phosphorylation effects these changes due to the long distance of these subunits from the phosphorylation site. Similar to results from recent steered molecular dynamics simulations [32], the movement of ARPC3 appears to be linked to changes in the bilobal structure of Arp3, and the movement of ARPC1 appears to be linked to the change in orientation of Arp2. Unlike their simulations however, we do not observe large changes between ARPC2 and ARPC4 – these changes may appear at larger changes in Arp2 orientation than are observed in our simulations or on a longer timescale. The lack of large changes in ARPC2 and APRC4 are consistent however with the fact that mother filament binding is not increased relative to the unphosphorylated state as ARPC2 and ARPC4 appear to be the main sites of mother filament binding [33]. Additionally, while some contacts between Arp2 and Arp3 were lost, Arp2 was not observed to fully dissociate from Arp3 or ARPC4 (data not shown).
The findings here and previously [28] indicate that phosphorylation is required for activation of the Arp2/3 complex. However, crystal structures of the Arp2/3 complex from preparations found to have nucleating activity do not show phosphorylation of Arp2. We confirmed that phosphatase treatment rendered similar preparations to those used in crystallographic studies inactive (Fig. 1 and [28]). We confirmed that loss of activity was due to dephosphorylation and not phosphatase binding to the complex or ATP dephosphorylation of the subunits, and our findings suggest that several populations of Arp2/3 complex exist in our preparations (data not shown). It is possible that phosphorylated Arp2/3 complex may not form crystals due to differences in conformation.
In all cases in which phosphorylation is required for functional activation, the unphosphorylated state can be considered auto-inhibited. A common mechanism for converting from the auto-inhibited to the activated state is one in which phosphorylation induces attractive interactions between the phosphorylated residue and other residues that are required for activating structural changes. This is seen in Ser/Thr protein kinases, and also in more traditionally auto-inhibited systems such as the phosphorylation of the tail of Tyr protein kinases (reviewed in [45]).
In contrast, mutation of arginine residues in the Arp2/3 complex enhances activation by phosphorylation (Fig. 5 and 6). This suggests a distinct mechanism to that discussed above. Structurally, these results suggested that the effects of phosphorylation are better understood as relieving an auto-inhibitory interaction through repulsive forces rather than driving conversion towards the active state via the formation of attractive interactions. The introduction of phosphate groups with a −2 charge disrupts the complex electrostatic network at the inter-subunit interfaces near the threonine phosphorylation sites that hold the complex in an inactive state. The destabilization of the interaction network driven by the need to accommodate the electrostatic perturbation leads to conformational changes that are permissive for full activation by NPF binding. Mutation of ARPC4 R105 and R106 constitutes an electrostatic perturbation that destabilizes the inter-subunit interfaces akin to phosphorylation. Combining the electrostatic perturbations of T237 or T238 phosphorylation and mutation of ARPC4 R105/R106 appears to lead to larger conformational changes (Fig. 5 and 6) and higher activity in the absence of NPF (Fig. 7). In addition, several other salt-bridge interactions are broken upon phosphorylation (Fig. S7). Other mutations, particularly of positively charged amino acids such as K232 Arp2, R409 Arp3, and R200 Arp2 that show strong interactions in the unphosphorylated states that are then broken upon phosphorylation, could constitute electrostatic perturbations that would also magnify the destabilizing effects of phosphorylation. Further studies will have to determine the extent to which the breakage of these other interactions can contribute to activation.
Based on the activation model laid forth by Dalhaimer and Pollard [32], the phosphorylation-induced relief-of-autoinhibition may provide a reduced energy barrier for conversion to the active state upon mother filament and NPF binding, although there may be thermodynamic effects such as stabilization of the active complex that our current data do not reveal. In this model, mutation of R105/R106 ARPC4 in the context of phosphorylation may further reduce the barrier to forming the fully active complex such that binding to mother filament even in the absence of NPF can still result in a substantial increase in the activation kinetics, though our data indicate that NPF binding further accelerates this process.
Due to the large computational expense of the simulations (∼500,000 cpu-hours during the course of this study), we were unable to perform all of the potentially informative simulations. For example, we have not investigated phosphorylation of Y202 on Arp2, which has been proposed as the site of tyrosine phosphorylation and is located close to T237/238 Arp2. The proximity of Y202 to the salt-bridge network around T238 leads us to speculate that its phosphorylation would exert its effects by a similar mechanism, but this hypothesis remains to be examined. We have also not investigated dual phosphorylation of both T237 and T238 Arp2. Nonetheless, despite these limitations and the short timescale probed by molecular dynamics, the simulations suggested a structural mechanism for phosphoregulation of Arp2/3 and predicted a gain-of-function mutation, which was confirmed experimentally. As such, this study provides an example of how computational simulations can be used to create testable models of regulatory phosphorylation, which is valuable when it is difficult to obtain direct, atomic-resolution structural information, as is often the case. Here, we have provided a new model for Arp2/3 regulation in which a network of electrostatic interactions helps to hold the complex in the inactive state, and this auto-inhibition must be relieved by phosphorylation to permit activation.
Plasmids encoding Arp2/3 complex subunits were obtained from M. Welch (UC Berkeley) and were generated as described [42]. Site directed mutagenesis was performed using a QuikChange Mutagenesis kit (Agilent Technologies) using the appropriate template. Primers used for T237/238A Arp2 mutation: 5′ primer (GAGCAGAAACTGGCCTTAGAAGCCGCAGTATTAGTTGAATCTTATACACTCCC) 3′ primer (GGGAGTGTATAAGATTCAACTAATACTGCGGCTTCTAAGGCCAGTTTCTGCTC), primers for the Y202A Arp2 mutation 5′ primer (CAAGCTACTTCTGTTGCGAGGAGCCGCCTTCAACCACTCTGCTGATTTTGAAAC), 3′ primer (GTTTCAAAATCAGCAGAGTGGTTGAAGGCGGCTCCTCGCAACAGAAGTAGCTTG). Primers used for the R105/106A ARPC4 mutations: 5′ primer (GAGAACTTCTTTATCCTTGCAGCGAAGCCTGTGGAGGGG), 3′ primer (CTCTTGAAGAAATAGGAAACGTCGCTTCGGACACCTCCCC).
Recombinant Arp2/3 complex was expressed and purified as described [42]. Briefly, Sf21 cells at a density 1.0×106 cells/ml were infected with baculoviruses containing cDNA encoding subunits of the Arp2/3 complex at equal infection units. Cells were grown in sf900 media in suspension for 48 hours and then harvested by a 10 min 1000×g centrifugation. Recombinant Arp2/3 complex was affinity purified on Talon resin (Clonetech), and fully assembled complex was collected after passage over a Superdex 200 FPLC gel filtration column.
Pyrene actin polymerization assays were performed with 4 µM monomeric actin containing 5% pyrene-labeled actin in KMEI (50 mM KCl, 1 mM MgCl2, 1 mM EGTA, and 10 mM imidazole, pH 7), 2.5 to 50 nM Arp2/3 complex, and 500 nM N-WASP VCA domain. Measurements were made with an RF-5301PC spectrophotometer (Shimadzu) at 1 s intervals. Growing filament ends were calculated by determining the rate of actin assembly at 80% of polymerization and using the relationship R = k+[A][E] where R is the rate of actin assembly, k+ is the association rate constant (10 µM−1⋅s−1), [A] is the concentration of monomeric actin and [E] is the concentration of growing filament ends as described previously [17]. The concentration of Arp2/3 complex was varied from 0 to 50 nM and the number of growing filaments calculated for each condition. Pointed elongation from gelsolin-capped actin filaments was measured as described previously [6]. Gelsolin-capped actin filaments (100 nM) were used for pointed end binding assays. F-actin binding assays and in vitro dephosphorylation of the Arp2/3 complex were performed as described previously [28].
Binding constants of Arp2/3 complex for NPFs were determined by using GST-NWASP VCA covalently coupled to Activated CH-sepharose 4B (GE Healthcare, Piscataway, NJ). GST-NPF-coupled beads were added to mock-treated or Antarctic phosphatase-treated Arp2/3 complex and incubated at room temperature for 30 min. NPF-coupled beads were spun at 700× g for 5 min, the supernatant removed and beads resuspended in SDS-PAGE sample buffer. Coomasie-stained gels were scanned and quantified using a LabWorks imaging system and LabWorks Software (UVI, CA). The data were plotted and fitted using GraphPad Prism software (GraphPad Software, Inc., San Diego, CA). Binding constants for Arp2/3 complex for actin filaments were determined by actin co-sedimentation as described [46].
Systems were prepared for molecular dynamics simulations starting from the crystal structure of the apo bovine Arp2/3 complex (PDB 1K8K [25]). A complete model of the unphosphorylated, wild-type bovine Arp2/3 complex was generated using the Protein Local Optimization Program (PLOP) [47], [48], [49] by building in all atoms missing in the electron density (except Arp2 subdomains 1 and 2). Subdomains 1 and 2 of Arp2 were modeled in based on homology to the actin monomer structure (PDB 1ATN [43]). The 15-residue unstructured extension at the end of ARPC2 was energy minimized, as were residues 39–51 of Arp3, residues 288–297 and 309–319 of ARPC1, and 41–43 and 65–67 of the Arp2 model. All phosphorylated and mutant models were generated from the unphosphorylated model by removing all side chain atoms from the unmodified residue and optimizing the positions of the side-chain atoms of the modified residue. These models were then solvated in TIP3P water [50] and monovalent counterions were added to neutralize the system using Maestro (Schrodinger LLC).
The full system was then energy minimized using DESMOND [51] (D.E. Shaw Research) in five stages with the following atoms restrained to their positions in the starting model: 1) all heavy atoms; 2) all backbone (N-C〈-C-O) heavy atoms and experimentally determined side-chain heavy atoms; 3) all experimentally determined heavy atoms; 4) all experimentally determined backbone atoms; 5) no restraints. Minimizations were performed with at least 100 steps of Steepest Descent minimization followed by L-BFGS optimization after reaching a gradient of 10.0 kcal·mol−1·Å−1 up to a total of 10,000 steps or a gradient of 0.1 kcal·mol−1·Å−1. After full energy minimization of the system, an equilibration was performed. First, the systems were annealed to a temperature of 300 K using Langevin dynamics at constant temperature and volume over 50 ps with all heavy atoms restrained. Subsequently, Langevin dynamics at constant temperature and pressure with a target temperature and pressure of 300 K and 1 atm were performed in stages: 1) 50 ps with all heavy atoms restrained with 50 kcal·mol−1·Å−1 force constants; 2) 50 ps with all backbone heavy atoms and experimentally determined side-chain atoms restrained with 50 kcal·mol−1·Å−1 force constants; 3) 150 ps with all experimentally-determined heavy atoms restrained with force constants reduced over the course of the simulation from 25 to 5 kcal·mol−1·Å−1; 4) 100 ps of simulation restraining only the experimentally determined backbone heavy atoms, over which the force constants of the restraints were brought to 0 from 5.0 kcal·mol−1·Å−1; 5) 100 ps of the unrestrained system. All Langevin dynamics simulations were performed with a 100 ps−1 damping constant.
Each system was then simulated for 30 ns using the Martyna-Tobias-Klein integrator [52] with a reference temperature of 300 K and a reference pressure of 1 atm. The barostat mass was set with a time constant of 2 ps and an equilibrium temperature of 300 K. The masses of all chain variables were set using a time constant of 1.0 ps. Both the Langevin dynamics and standard molecular dynamics simulations were performed with all bonds involving hydrogens constrained, a 2 fs time step for the bonded and short-range nonbonded interactions and updating of long-range nonbonded interactions every 4 fs using the RESPA multiple time step approach. Non-bonded interactions were tapered using force-switching starting at a distance of 9.0 Å to an interaction cutoff of 9.5 Å. Pairlists were constructed using a distance of 10.5 Å and a migration interval of 12 ps. These parameters were tested in short simulations in the NVE ensemble to ensure good energy conservation. Coordinates of the full system were added to the output trajectory every 10 ps.
Coordinates of the Cα atoms from the last 20 ns of each unphosphorylated, Arp2 pThr237, and Arp2 pThr238 simulation were collected into a single trajectory on which Principal Component Analysis [53], [54] was performed using the Bio3D package for the R statistical software package [55]. All Cα coordinates were used after superimposing the Cα atoms of Arp3 subdomains 1 and 2 resolved in the starting crystal structure (residues 3–39, 51–151, 376–410) of each frame in each trajectory. The first and second principal components (PCs) account for 62.4% of the variation in atomic coordinates, and the first 4 principal components account for 84.0% (Fig. S3a). The major differences between the unphosphorylated and phosphorylated simulations are largely localized to the first PC, with the second PC capturing variation between the duplicate simulations of the complex in the same phosphorylation state.
A contact between the Arp2 and Arp3 subunits was defined as the number of heavy atoms in Arp3 that were within 3.5 Å of any heavy atom in Arp2. The number of contacts between the Arp2 and Arp3 subunits was calculated for every 10th frame (100 ps) of each simulation. The results for the duplicate simulations of each wild-type or mutant complex were then pooled and compared.
|
10.1371/journal.pgen.1004116 | E3 Ubiquitin Ligase CHIP and NBR1-Mediated Selective Autophagy Protect Additively against Proteotoxicity in Plant Stress Responses | Plant stress responses require both protective measures that reduce or restore stress-inflicted damage to cellular structures and mechanisms that efficiently remove damaged and toxic macromolecules, such as misfolded and damaged proteins. We have recently reported that NBR1, the first identified plant autophagy adaptor with a ubiquitin-association domain, plays a critical role in plant stress tolerance by targeting stress-induced, ubiquitinated protein aggregates for degradation by autophagy. Here we report a comprehensive genetic analysis of CHIP, a chaperone-associated E3 ubiquitin ligase from Arabidopsis thaliana implicated in mediating degradation of nonnative proteins by 26S proteasomes. We isolated two chip knockout mutants and discovered that they had the same phenotypes as the nbr1 mutants with compromised tolerance to heat, oxidative and salt stresses and increased accumulation of insoluble proteins under heat stress. To determine their functional interactions, we generated chip nbr1 double mutants and found them to be further compromised in stress tolerance and in clearance of stress-induced protein aggregates, indicating additive roles of CHIP and NBR1. Furthermore, stress-induced protein aggregates were still ubiquitinated in the chip mutants. Through proteomic profiling, we systemically identified heat-induced protein aggregates in the chip and nbr1 single and double mutants. These experiments revealed that highly aggregate-prone proteins such as Rubisco activase and catalases preferentially accumulated in the nbr1 mutant while a number of light-harvesting complex proteins accumulated at high levels in the chip mutant after a relatively short period of heat stress. With extended heat stress, aggregates for a large number of intracellular proteins accumulated in both chip and nbr1 mutants and, to a greater extent, in the chip nbr1 double mutant. Based on these results, we propose that CHIP and NBR1 mediate two distinct but complementary anti-proteotoxic pathways and protein's propensity to aggregate under stress conditions is one of the critical factors for pathway selection of protein degradation.
| Environmental stresses such as heat cause generation of misfolded and damaged proteins, which are highly toxic and must be efficiently removed. In plants, NBR1, the first isolated autophagy receptor with an ubiquitin-association domain, plays a critical role in plant stress tolerance by targeting ubiquitinated protein aggregates under stress conditions for degradation by autophagy. To study how stress-induced misfolded and damaged proteins are detected and ubiquitinated in plant cells, we analyzed the chaperone-associated E3 ubiquitin ligase CHIP from Arabidopsis thaliana for its role in protection against proteotoxicity in plant stress responses. Disruption of Arabidopsis CHIP caused increased sensitivity to a spectrum of abiotic stresses as found in the Arabidopsis nbr1 mutants. Disruption of both Arabidopsis CHIP and NBR1 further compromised plant stress tolerance, indicating that their roles are additive. Furthermore, in the chip nbr1 double mutant, compromised heat tolerance was associated with increased accumulation of insoluble proteins derived mostly from heat-sensitive but biologically important proteins such as Rubisco activase, catalases and proteins required for protein synthesis and folding. Importantly, stress-induced protein aggregates were still highly ubiquitinated in the chip mutants. These results strongly suggest that CHIP and NBR1 function in two distinct but complementary anti-proteotoxic pathways in plant stress responses.
| A newly synthesized polypeptide must fold and at times refold into a proper conformation in order to function properly in the cell. However, folding into a proper conformation is a complex process and misfolding is inevitable. In addition, cellular proteins can be damaged by stress conditions such as high temperature, high salt concentrations and reactive molecules. Misfolded and damaged proteins are highly toxic to the cell as they can engage in inappropriate interactions with important cellular components. To monitor, repair or degrade misfolded and damaged proteins, the cell relies on an elaborately regulated protein quality control system, which consists of molecular chaperones such as Hsp70 that promote folding and refolding of nonnative proteins and the protein degradation systems that remove misfolded and damaged proteins [1]. Most soluble misfolded proteins are degraded through the ubiquitin proteasome system (UPS) in which a misfolded protein is ubiquitinated by an enzymatic E1/E2/E3 ubiquitination cascade for targeted destruction by the 26S proteasome. Misfolded and damaged proteins can also be degraded by selective autophagy, which often also relies on ubiquitination for cargo recognition and delivery through such autophagy receptors as P62 and NBR1, which recognize ubiquitinated misfolded and damaged proteins through their ubiquitin association domain [1].
As sessile organisms, plants are constantly exposed to a variety of stress conditions that cause damage to cellular molecules and structures. To survive, plants have evolved complex mechanisms for sensing, responding and adapting to the ever-changing and often stressed environmental conditions. A large number of studies have shown that protein ubiquitination plays a critical role in plant stress responses [2]. Expression of multiple polyubiquitin genes is induced by stress conditions such as high temperatures in plants [3], [4], [5] and overexpression of a single mono-ubiquitin gene increased tolerance of transgenic plants to multiple stresses [6]. Furthermore, mutations of the genes for the 19S regulatory particle subunits of the 26S proteasome reduces plant tolerance to salt, UV radiation and heat shock, indicating that UPS plays a critical role in general plant stress responses [7], [8], [9], [10], [11]. Interestingly, almost all analyzed ubiquitin E3 ligases that regulate plant abiotic stress responses function through modulating the levels of regulatory proteins, mostly transcription factors [2]. By contrast, very little is known about the roles of plant UPS in removing misfolded and damaged proteins that accumulate under stress conditions. This is contrary to the intensive interest and extensive knowledge about cellular pathways for clearance of damaged and toxic macromolecules in mammalian organisms, which are linked with a large number of human diseases including prominent neurodegenerative disorders such as Alzheimer, Parkinson and Huntington's diseases [12], [13].
Autophagy is another evolutionarily conserved major route of protein degradation [1], [14]. Autophagy plays a critical role in plant nutrient recycling and utilization and responses to both biotic and abiotic stresses [15], [16], [17]. Under nutrient starvation, autophagy provides an internal source of nutrients under starvation through nonselective, bulk degradation of cytoplasmic constituents including proteins and organelles. However, autophagy also functions as a quality control mechanism that selectively targets damaged organelles and toxic macromolecules [17], [18]. Selective autophagy is mediated by autophagy receptors/adaptors that recognize specific autophagy substrates [17], [18]. To understand plant selective autophagy, we have recently analyzed Arabidopsis NBR1, the first isolated plant autophagy receptor [19]. We have discovered that NBR1 has a selective role in plant tolerance to heat, oxidative, drought and salt stresses but not in age- and darkness-induced senescence and in resistance to necrotrophic pathogens, which also involve autophagy [20]. The compromised heat tolerance of atg5, atg7, and nbr1 mutants was associated with increased accumulation of insoluble, detergent-resistant proteins that were highly ubiquitinated under heat stress [20]. NBR1, which contains an ubiquitin-association domain, also accumulated to high levels with an increasing enrichment in the insoluble protein fraction in the autophagy-deficient mutants under heat stress [20]. These results suggest that NBR1-mediated autophagy targets ubiquitinated protein aggregates most likely derived from denatured and damaged proteins generated under stress conditions.
In animal cells, the carboxyl terminus of the Hsc70-interacting protein (CHIP) plays a critical role in protein quality control by ubiquitinating Hsp70-bound misfolded proteins [21], [22]. CHIP acts as both an Hsp70 co-chaperone through its N-terminal tetratricopeptide repeat (TPR) domain and an E3 ubiquitin ligase through the C-terminal U-box domain [23]. If refolding of denatured or damaged proteins assisted by molecular chaperones such as Hsp70 is not successful, CHIP E3 ubiquitin ligase can introduce ubiquitination and thereby target denatured and damaged proteins for degradation by both UPS and autophagy [1]. For example, α-synuclein, a major component of the protein aggregates associated with Parkinson's disease, is degraded by both UPS and autophagy after ubiquitination by CHIP [24], [25], [26]. Overexpression of CHIP decreases aggregation of proteins and cell death associated with chronic neurodegenerative diseases including Parkinson's and Alzheimer [27], [28], [29]. By contrast, deficiency of CHIP in knockout mice decreases longevity associated with accelerated aging phenotypes accompanied by altered protein quality control [30], [31].
Arabidopsis CHIP protein is structurally highly similar to animal CHIPs with three tetratricopeptide repeats at the N-terminal side and a U-box domain at the C-terminal side [32]. In addition, Arabidopsis CHIP interacts with molecular chaperones such as Hsc70 and has E3 ubiquitin ligase activity in vitro [32]. A number of studies have previously reported functional analysis of Arabidopsis CHIP in protein turnover and stress responses. CHIP is induced by high cytosolic levels of chloroplast-destined precursor proteins and promotes their degradation by UPS when co-overexpressed with Hsc70-4 in protoplasts [33]. CHIP also interacts with and ubiquitinates chloroplast FtsH and Clp4 proteases and protein phosphatase 2A (PP2A) and the levels of FtsH and Clp4 proteins were reduced but those of PP2A were increased in transgenic plants overexpressing CHIP [34], [35], [36]. CHIP is induced by stress conditions including cold temperature, heat and salt [32]. However, overexpression of CHIP in transgenic Arabidopsis plants increases sensitivity to both low and high temperatures [32]. The phenotypes of the CHIP-overexpressing transgenic plants are unexpected given the positive roles of the E3 ligase in protein quality control. However, even in human cells, when chronically overexpressed, CHIP can affect essential signaling pathways and causes deleterious effects on cellular health [37]. In the present study, we isolated two independent T-DNA knockout mutants for the CHIP gene from Arabidopsis. The chip mutants are normal under normal growth conditions but are hypersensitive to heat, salt and oxidative stresses. To study its functional interaction with NBR1, we generated chip nbr1 double mutants and provided genetic evidence that CHIP functions additively with NBR1 in plant stress responses. Further analysis including proteomic profiling of stress-induced protein aggregates strongly suggest that CHIP and NBR1 mediate two distinct but complementary anti-proteotoxic pathways in plant stress responses.
To genetically analyze the biological functions of Arabidopsis CHIP, we screened T-DNA insertion stocks and identified two independent T-DNA insertion mutants for Arabidopsis CHIP. The chip-1 mutant (Salk_048371) contains a T-DNA insertion in the seventh exon and the chip-2 mutant (Salk_059253) contains a T-DNA insertion in the sixth exon of the CHIP gene (Figure 1A). Quantitative RT-PCR showed that the two mutants had less than 1% of the wild-type level of CHIP transcript (Figure 1B), indicating that they are likely null mutants. Both chip-1 and chip-2 mutants were normal in growth and development and displayed no detectable morphological phenotypes throughout the entire life cycle when grown under normal growth conditions.
As a chaperone-associated E3 ubiquitin ligase, animal CHIP proteins ubiquitinate misfolded and damaged proteins and target their degradation by both 26S proteasomes and P62/NBR1-mediated selective autophagy. To examine whether Arabidopsis CHIP plays a similar role in plant stress tolerance through ubiquitination of stress-induced protein aggregates in NBR1-mediated selective autophagy, we first analyzed the chip mutants for responses to a spectrum of abiotic stresses. For testing heat tolerance, three-weeks-old seedlings of wild type and chip mutants were placed in a 45°C growth chamber for 9 hours and scored for survival rates after recovery for 5 days at the room temperature. As shown in Figure 1C & 1D, more than 80% of wild-type seedlings but only about 30% of chip mutant seedlings survived after the heat stress. When 5-weeks old mature plants were heat-stressed for 9 hours at 45°C, 70–80% of wild-type leaves but less than 20% of chip mutant leaves remained green after 5-day recovery at the room temperature (Figure 1E). These results indicated that the heat tolerance of the chip mutants was substantially compromised. To test tolerance to oxidative stress, we sprayed 5-weeks old wild type, chip-1, chip-2 mutants with 20 µM methyl viologen (MV), a reactive oxygen species (ROS)-generating herbicide, and kept the plants under light for two days. For wild-type plants, only old leaves were significantly bleached but more than 90% of leaf areas remained green (Figure 1F). By contrast, more than 50% of leaf tissues of chip-1 and chip-2 mutant plants were bleached after MV treatment (Figure 1F). Thus, the chip mutant plants were also compromised in tolerance to oxidative stress.
We also compared the wild type and chip mutant plants for responses to abscisic acid (ABA) and salt stress. Wild-type and chip mutant seeds were sown on 0.5× Murashige and Skoog (MS) agar medium with or without ABA (0.5 µM) or NaCl (150 mM). Germination rates were determined daily for the following 8 days through scoring of green cotyledons. On the MS medium with no added ABA or NaCl, there was no difference in germination rates between wild type and chip mutants, with close to 100% germination after 4 days on the medium (Figure 2A). On the MS medium containing 0.5 µM ABA, however, germination of the chip mutants were substantially delayed when compare to that of wild-type (Figure 2B). For example, more than 30% of wild-type seeds but almost no chip mutant seeds had green cotyledons after 4 days (Figure 2B). Likewise, more than 60% of wild type but only 20% of chip mutants had green cotyledons after 5 days on the ABA-containing MS medium (Figure 2B). After 8 days, however, almost 100% of wild type and chip mutant seeds had green cotyledons (Figure 2B). On the MS medium containing 150 mM NaCl, the percentages of wild-type seeds with green cotyledons increased steadily to about 80% after 8 days (Figure 2C). However, only about 20–30% of chip mutants had green cotyledons after 8 days on the NaCl-containing medium (Figure 2C). Thus, the chip mutants were highly compromised in salt stress.
We also analyzed the responses of the chip mutants to the hemibiotrophic bacterial pathogen Pseudomonas syringae and the necrotrophic fungal pathogen Botrytis cinerea and found the mutants to be normal in resistance to both pathogens based on both symptom development and pathogen growth. Furthermore, we found that the chip mutants were normal in age- and dark-induced senescence. Thus, the chip mutants appeared to be compromised specifically in tolerance to abiotic stresses and ABA.
The compromised phenotypes in tolerance to a spectrum of abiotic stresses but normal phenotypes in disease resistance and senescence of the chip mutants were strikingly similar to those shown by the nbr1 mutants [20], supporting that the chaperone-associated E3 ubiquitin ligase might function in the same pathway as NBR1 in plant stress responses. In the nbr1 mutants, compromised heat tolerance was associated with accumulation of insoluble, detergent-resistant protein aggregates that are most likely derived from heat-denatured or damaged proteins [20]. To determine whether the chip mutants also shared this phenotype with the nbr1 mutants, we compared wild type with the chip mutants for the levels of insoluble, detergent-resistant protein aggregates during the 9-hour heat treatment. The plants were placed in a 45°C growth chamber and leaves were collected at various time points for isolation of both total and insoluble proteins. As shown in Figure 3A, insoluble proteins as percentages of total proteins in wild-type plants increased only slightly, from 1.6% to about 3%, after 9-hour heat stress (Figure 3A). By contrast, in both chip-1 and chip-2 mutants, insoluble proteins increased as percentages of total proteins from 2% to more than 8% after 9 hours at 45°C (Figure 3A). As a result, the levels of insoluble protein aggregates in the chip-1 and chip-2 mutants were about three times of those in wild type after 9-hour heat stress (Figure 3A). Thus, as in nbr1 mutants, compromised heat tolerance in the chip mutants was associated with increased accumulation of heat-induced insoluble protein aggregates.
The strikingly similar phenotypes of the chip and nbr1 mutants in compromised stress tolerance and in accumulation of stress-induced protein aggregates strongly suggested that Arabidopsis CHIP and NBR1 have similar roles in plant stress responses. To analyze genetically the functional relationship between CHIP and NBR1, we generated the chip-1 nbr1-1 double mutant and compared it with its parental chip and nbr1 single mutants for heat tolerance. When 5-weeks old mature plants were heat-stressed for 9 hours at 45°C, more than 80% of wild-type leaves and 20–30% of chip and nbr1 single mutant leaves remained green after 5-day recovery at room temperature (Figure 4A). By contrast, no green leaves from the mature chip nbr1 double mutants were detected after 5-day recovery at room temperature following 9-hour heat stress (Figure 4A). We also placed two-weeks-old seedlings of wild type and mutants in a 45°C growth chamber for 9 hours and scored for survival rates after recovery for 5 days at room temperature. More than 80% of wild-type seedlings and about 30% of chip and nbr1 single mutant seedlings survived after the heat stress (Figure 4B & 4C). By contrast, only about 5% of the chip nbr1 double mutant seedlings survived after the heat stress (Figure 4B & 4C). Thus, assays of both mature plants and seedlings indicated that the chip nbr1 double mutants were more compromised in heat tolerance than the parental chip and nbr1 single mutants. In addition, we compared the chip nbr1 double mutant with the chip and nbr1 single mutants for the levels of insoluble, detergent-resistant protein aggregates during the 9-hour heat treatment. At the three time points assayed (3, 6 and 9 hours), the levels of insoluble protein aggregates in the chip nbr1 double mutant were about 30–40% higher than those in the chip and nbr1 single mutants (Figure 5A). This result supported that the roles of CHIP and NBR1 in plant heat tolerance were additive.
Previously it has been shown that CHIP E3 ubiquitin ligase and its interacting Hsc70 mediate degradation of plastid-destined precursor proteins by the 26S proteasomes to prevent cytosolic precursor accumulation [33]. With the important role of CHIP in plant tolerance to abiotic stresses, we examined whether CHIP is involved in ubiquitination of heat-induced protein aggregates, which accumulated in the chip mutants under heat stress. We isolated both soluble and insoluble proteins from the wild type and chip mutants collected before and after 9-hour heat stress. The proteins were fractionated by SDS-PAGE and analyzed for ubiquitinated proteins using an anti-ubiquitin monoclonal antibody. As shown in Figure 3B, we observed similar levels of ubiquitinated proteins in the soluble fractions in these plants with or without heat stress. In the insoluble fractions, we observed similar levels of ubiquitinated proteins in the chip-1 and chip-2 mutants and in wild type before heat treatment (Figure 3B). However, after 9-hour heat stress, we observed a drastic increase in the levels of ubiquitinated proteins in the chip-1 and chip-2 mutants but not in the wild-type plants (Figure 3B). Thus, stress-induced insoluble proteins from heat-stressed chip mutants are still highly ubiquitinated.
To examine combined effects of CHIP and NBR1, we also compared the chip nbr1 double mutant with its parental chip and nbr1 single mutants for the changes in the levels of insoluble ubiquitinated proteins after 3-, 6-, and 9-hour heat treatment. As shown in Figure 5B, there were little or slight increases in the levels of ubiquitinated insoluble proteins in the wild-type plants during the 9-hour heat stress. On the other hand, ubiquitinated insoluble proteins increased steadily and to similar levels in the chip and nbr1 single mutants with time of heat stress (Figure 5B). In the chip nbr1 double mutant plants, ubiquitinated insoluble proteins increased to even higher levels than those in their parental chip and nbr1 single mutants during the 9-hour heat treatment (Figure 5B). These results demonstrated the additive roles of CHIP and NBR1 in protection against proteotoxicity in plant stress responses and provided further evidence that CHIP is dispensable for ubiquitination of heat-induced protein aggregates targeted by NBR1-mediated selective autophagy.
Compromised heat tolerance of chip and nbr1 single and double mutants was associated with increased accumulation of insoluble ubiquitinated protein aggregates under heat stress (Figures 5). To identify proteins targeted by CHIP- and NBR1-mediated degradation pathways under heat stress, we isolated insoluble protein aggregates from wild type, chip and nbr1 single and double mutant plants after 6 and 9-hour heat stress and subjected them to shotgun LC-MS/MS profiling. From the proteomic profiling of the four genotypes with two time points of heat stress, more than 16,000 tryptic peptide sequences were obtained, which belong to 440 non-redundant candidate proteins (Tables S1 & S2). Because a substantial percentage of the identified peptides matched several members of the same protein family, more than 554 proteins were identified (Tables S1 & S2). Functional categorization of the identified proteins showed that almost 20% of them are components of chloroplasts/plastids but less than 3% are associated with each of the other major organelles including nucleus, mitochondria, Golgi apparatus and endoplasmic reticulum (Figure 6A). Intriguingly, almost 20% of identified proteins are involved in responses to biotic and abiotic stimuli or other stress signals (Figure 6B), making it one of the largest categories for important biological processes in plants.
To investigate differential accumulation of aggregated proteins in the chip and nbr1 mutants, we estimated the relative abundance of aggregated proteins by normalizing the numbers of detected peptides to the amount of total proteins from which aggregated proteins were isolated (Tables S1 & S2). Survey of the most abundant protein aggregates based on the normalized peptide numbers revealed both expected and surprising results. As expected, a large percentage of heat-induced protein aggregates are chloroplast/plastid proteins but their relative abundances in the aggregates were not totally correlated with their relative levels in leaf tissues. For example, Rubisco proteins are the most abundant proteins in leaf tissues but were not the most abundant aggregated proteins after 6-hour heat stress (Table S1). Rubisco activase was one of the most abundant aggregated proteins accumulated in heat-stressed chip nbr1 mutant plants (Tables S1 & S2). CAT3 and CAT2, two major leaf catalase isoforms in Arabidopsis [38], were also present abundantly as aggregated proteins (Tables S1 & S2). Other abundant protein aggregates accumulated in heat-stressed plants include a number of chloroplast-localized light-harvesting complex proteins (Tables S1 & S2). Interestingly, a substantial number of abundant aggregated proteins are involved in protein synthesis, folding/refolding and maturation. These proteins included translation initiation and elongation factors, cyclophilin-type peptidyl-protyl cis-trans isomerases and heat shock proteins (Tables S1 & S2).
We also estimated heat-induced enrichment of aggregated proteins in the chip and nbr1 mutants by calculating the ratio of their normalized peptide numbers to those in the wild type. As shown in Figure 7A, among the 10 most abundant aggregated proteins a number of them differentially accumulated in the chip or nbr1 mutants after 6-hour heat stress. For example, the abundances of protein aggregates with peptide sequences matching several highly similar plastid light-harvesting complex (LHC) proteins including those of At1g29910 and At2g34420 were 2 to 3 times higher in the chip mutant than in the nbr1 mutant (Figure 7A). On the other hand, the levels of protein aggregates for Rubisco activase (RCA, At2g39730), CAT3 (At1g20620) and CAT2 (At4g35090) in the nbr1 mutant were 3 to 5 times higher than those in the chip mutant (Figure 7A). Survey of all detected aggregated proteins also showed a substantial percentage of them accumulating differentially in the two mutants after 6-hour heat stress (Tables S1 & S2). Intriguingly, after 9-hour heat stress, relative abundances of Rubisco activase and CAT proteins became similar in the two mutants (Figure 7B). Consistent with the heat sensitive phenotypes, the abundances of protein aggregates in the chip nbr1 double mutant were generally higher than those in the chip and nbr1 single mutants after both 6- and 9-hour heat stress (Figure 7; Tables S1 & S2).
To confirm the differential accumulation of protein aggregates from the proteomic profiling, we analyzed the accumulation of RCR and CAT protein aggregates in the chip and nbr1 mutants after 0- and 6-hour heat stress using western blotting. In the western blotting, we included two T-DNA insertion mutants (rpn1a-4 and rpn1a-5) for the Arabidopsis 26S proteasome subunit RPN1a, which are also sensitive to heat stress [39]. When both soluble and insoluble proteins were probed with a previously generated RCA antibody [40], two RCA isoforms of 43 and 47 kD arising from mRNA alternative splicing were detected (Figure 8A). Without heat stress, high levels of RCA proteins were detected in the soluble fraction but little RCA proteins were present in the insoluble fractions (Figure 8A). After 6-hour heat stress, high levels of RCA proteins were still present in the soluble fraction of wild type, chip and rpn1a mutants, although a significant level of the proteins was also detected in the insoluble fraction (Figure 8A). In the nbr1 mutant, on the other hand, the soluble RCA was reduced while insoluble RCA, particularly the 43-kD isoform, was substantially increased (Figure 8A). Thus, insoluble RCA accumulated to a higher level in the nbr1 mutant than in the chip and rpn1a mutants. Likewise, using a catalase monoclonal antibody that recognize both Arabidopsis CAT2 and CAT3 (Figure S1), we observed that after 6-hour heat stress, insoluble catalase proteins accumulated at high levels in the nbr1 mutant but not in wild type, chip or rpn1a mutants (Figure 8B).
Catalases are antioxidant enzymes with a crucial role in cellular responses to oxidative stress, which are linked with a wide variety of biotic and abiotic stresses including heat stress [41]. Arabidopsis CAT3 and CAT2 are two major catalase isoforms in Arabidopsis leaves [38] and their abundant presence as aggregated forms in heat-stressed chip nbr1 mutant plants suggested a possible mechanistic link between stress-induced oxidative stresses and protein quality control. To examine the connection between heat-induced proteotoxic and oxidative stresses, we compared wild type and the chip nbr1 double mutant for the changes of both the catalase proteins and activity. Before heat stress, wild type and chip nbr1 mutant plants had similar levels of catalase proteins, which were mostly soluble (Figure 9A). After 6- and 9-hour heat treatment, a majority of catalases remained soluble and no major increases in catalase proteins were observed in wild-type plants (Figure 9A). On the other hand, heat stress increased total catalase protein levels in the chip nbr1 mutant but a majority of these catalase proteins existed as insoluble proteins (Figure 9A). Importantly, despite higher levels of total catalase proteins, heat-stressed chip nbr1 mutant plants had lower levels of soluble catalase proteins than heat-stressed wild-type plants (Figure 9A). Furthermore, a majority of aggregated catalases accumulated in heat-stressed chip nbr1 mutant plants migrated normally on SDS-PAGES when compared to soluble catalase proteins (Figure 9A), indicating that they were not ubiquitinated. A significant fraction of the insoluble catalase aggregates in the chip nbr1 mutant plants were present as high molecular weight proteins, most likely due to polyubiquitination (Figure 9A).
To determine whether aggregated catalase proteins were catalytically active, we first attempted to measure directly the catalase activity of the insoluble protein fractions but found the assays to be difficult once the protein aggregates were pelleted after centrifugation. Therefore, we used an indirect approach by comparing the catalase activities of the total protein fractions with those of the soluble protein fractions. As shown in Figure 9B, the total and soluble catalase activities in wild-type plants were very similar before or after heat treatment, which were expected given that catalase proteins in wild type remained mostly soluble after heat treatment (Figure 9A). In heat-stressed chip nbr1 mutant plants, the catalase activities from the total and soluble protein fractions were also very similar (Figure 9B) even though the levels of catalase proteins in the total protein factions were much higher than those in the soluble protein factions (Figure 9A). This result indicated that only the soluble catalases were catalytically active. Consistent with this interpretation, the total catalase activities increased slightly in heat-treated wild-type plants but reduced substantially in heat-treated chip nbr1 mutant plants. After 9-hour heat treatment, the catalase activities in the chip nbr1 mutant plants were only about 35% of those in the wild-type plants. In both wild type and chip nbr1 mutant, the levels of catalase activities were closely correlated with the levels of soluble catalase proteins (Figure 9). Thus, reduced degradation of denatured catalase aggregates in the chip nbr1 mutant plants not only led to increased accumulation of catalase protein aggregates but also caused reduction of soluble, active catalase proteins.
The additive phenotypes and the differential accumulation of aggregated proteins in the chip and nbr1 mutants suggest that CHIP and NBR1 participate in two distinct pathways in degrading misfolded and damaged proteins under stress conditions. This interpretation is consistent with the normal ubiquitination of NBR1-targeted protein aggregates in the chip mutants (Figures 3B & 5B). However, since CHIP- and NBR1-mediated anti-proteotoxic pathways contribute to the same process of removing misfolded and damaged proteins under stress conditions, they may be coordinated not only in functions but also in regulation. To test this, we examined the effect of CHIP deficiency on heat-induced autophagy by comparing wild type and chip mutants for heat-induced autophagosome formation and autophagy gene expression. We examined the effect of heat stress on induction of autophagosome accumulation using green fluorescent protein (GFP)-tagged ATG8a, which is associated with autophagosomes and has been used as a marker of autophagosomes in Arabidopsis [42], [43], [44]. Transgenic wild-type and chip plants expressing GFP-ATG8a were exposed to 45°C for 0, 1.5 and 3 hours, recovered for 0.5 hour at room temperature and then observed by confocal fluorescence microscopy for autophagosomes. In both the wild-type and chip mutant plants, the numbers of punctate GFP signals were low before heat stress (Figure 10). The punctate fluorescent structures in wild-type plants did not increase significantly during the first 1.5 hours of heat stress but elevated by 6-fold after 3-hour heat stress (Figure 10). In the chip mutant plants, the punctate fluorescent structures increased by almost 4-fold during the first 1.5 hours of heat stress (Figure 10). By the third hour of heat stress, however, there were similar numbers of punctate fluorescent structures in the wild-type and chip mutant plants (Figure 10). Thus, deficiency of CHIP caused an earlier induction of autophagosome accumulation by heat stress.
We also compared wild type and chip mutant plants for expression patterns of seven Arabidopsis autophagy genes (ATG5, ATG6, ATG7, ATG8a, ATG9, ATG10, ATG18a) and NBR1 in response to high temperature. Arabidopsis wild-type and chip mutant plants were placed in a 45°C chamber and total RNA was isolated from rosette leaves for detection of ATG and NBR1 gene transcripts using qRT-PCR. At 45°C, the transcript levels of the ATG genes were elevated with varying kinetics (Figure S2). For most of the ATG genes, the increased levels of transcripts were detected as early as 2 hours after initiation of the heat stress (Figure S2). ATG8a exhibited increased transcript levels after 2-hour exposure to the high temperature (Figure 10). In the chip mutant, we observed increased levels of transcripts for ATG6, ATG7, ATG9 similar to those in the wild-type plants (Figure S2). For the other five genes tested, however, changes in their transcripts in the chip mutant responded more sensitively to heat stress than those in the wild-type plants (Figure S2). For example, transcripts for ATG5, ATG8a, ATG10 and NBR1 in the chip mutant increased more rapidly and to higher levels than those in the wild-type plants during the first 4–6 hours of heat stress (Figure S2). During the later stage of heat stress, decline in the transcript levels for some of the ATG and NBR1 genes after initial increase in the chip mutant also occurred earlier and to greater extents than in the wild-type plants (Figure S2). ATG18a was an exception because its transcripts declined earlier in the chip mutant than in the wild-type plants after an initial induction by heat stress (Figure S2).
We also compared wild type and chip mutant plants for changes in the transcript levels for CAT2 and CAT3 under heat stress. For the wild-type plants, the transcript levels for CAT2 were little changed during the first 6 hour heat stress but declined for the remaining four hours to about 50% of its control levels (Figure S3). The transcript levels for CAT2 in the chip mutant remained unchanged during the first 2 hours but declined rapidly during the remaining 8-hour heat stress to about 10% of its control levels. For CAT3, heat stress increased transcript levels in both wild-type and chip mutant plants but this increase was more pronounced in the wild-type plants than in the chip mutant (Figure S3).
The animal CHIP ubiquitin E3 ligase has been extensively analyzed for its critical role in protein quality control. CHIP deficiency in knockout mice reduces median survival from 25 months to 10 months, representing a 60% decrease in longevity [30], [31]. Decreased longevity in the knockout mice is associated with accelerated aging-related pathophysiological phenotypes including reduced body weight, increased and accelerated skeletal muscle atrophy, decreased body fat stores, increased signs of cardiac hypertrophy, osteoporosis and kyphosis [30], [31]. The Arabidopsis CHIP has also been characterized for its expression, E3 ligase activity and interaction with the Hspc70 chaperones and potential ubiquitination substrates [32], . However, functional analysis of Arabidopsis CHIP has been exclusively through overexpression, which, surprisingly, caused cold- and heat-hypersensitivity in transgenic plants [32], [33]. In the present study, we have isolated two independent T-DNA knockout mutants for Arabidopsis CHIP and conducted a comprehensive genetic analysis of their phenotypes under both normal and stress conditions. Contrary to the strong phenotypes of knockout mice, Arabidopsis chip and chip nbr1 mutants display no detectable alternation in growth or development throughout the life cycle. A large number of studies have shown that misfolded or damaged proteins are prevalent even under normal growth conditions with roughly 30% of all newly synthesized proteins degraded by the UPS and related pathways [45], [46]. In Arabidopsis, mutant deficient for autophagy or other ubiquitin E3 ligases display premature senescence under normal growth conditions [47], [48], [49]. Therefore, there are likely other ubiquitin E3 ligases and autophagy adaptors/receptors that act independently or redundantly with CHIP and NBR1 for basal protein quality control in plants.
Arabidopsis chip mutants are also normal to both the hemibiotrophic bacterial pathogen P. syringae and necrotrophic pathogen B. cinerea, suggesting that it is also dispensable in plant immunity. On the other hand, the chip mutants are hypersensitive to heat, oxidative and salt stresses (Figures 1 & 2). Increased heat sensitivity of the chip mutant plants was associated with increased accumulation of insoluble protein aggregates (Figure 3). Therefore, CHIP is important for removal of stress-damaged proteins during plant responses to abiotic stresses. CHIP functions as both a co-chaperone and an E3 ubiquitin ligase, thereby linking cellular protein folding with protein degradation. Through physical interactions with molecular chaperones Hsp70/Hsc70 proteins, CHIP can ubiquitinate those chaperone-bound nonnative proteins and target them for degradation by UPS or selective autophagy. Deficiency of CHIP in the chip mutants would lead to accumulation of denatured or damaged proteins and cytotoxicity under stress conditions. Intriguingly, transgenic CHIP-overexpressing plants are hypersensitive to both cold and high temperatures [32], most likely because of the deleterious effect to cellular heath due to a number of possible mechanisms. First, molecular chaperones associated with CHIP is also involved in folding and refolding of nonnative proteins that may form under stress conditions. Excess levels of CHIP could interfere with molecular chaperones for their protein folding or refolding activities or prematurely ubiquitinate those chaperone-bound but refoldable proteins and targets them for unnecessary degradation. Second, CHIP may interact with other proteins and regulate their degradation independent of chaperones. In Arabidopsis, a number of proteins including subunits of protein phosphatase 2A, chloroplast FtsH and ClpP4 proteases interact directly with and act as substrates of the CHIP ubiquitin E3 ligase [34], [35], [36]. Chronic overexpression of CHIP could affect important signaling pathways involving PP2A and alter protein degradation and other important functions of chloroplasts, leading to deleterious effects on plant health [34], [35], [36]. These results indicate that a delicate balance in cellular CHIP levels is important for protein quality control and for plant stress tolerance.
Despite the highly similar phenotypes of the chip and nbr1 mutants, additional analysis strongly suggests that CHIP and NBR1 mediate two distinct but complementary anti-proteotoxic pathways. First, compromised tolerance of the chip nbr1 double mutant to heat stresses was consistently more severe than those of the chip and nbr1 single mutants (Figure 4). The additive nature of the chip nbr1 double mutant phenotypes indicated that the roles of CHIP and NBR1 in plant stress tolerance do not completely overlap. Second, insoluble protein aggregates accumulated in the chip single and chip nbr1 double mutant plants were still highly ubiquitinated under heat stress (Figures 3 & 5), thereby implicating a CHIP-independent mechanism for ubiquitination of stress-induced protein aggregates in NBR1-mediated selective autophagy. Third, proteomic profiling revealed that after 6-hour heat stress, aggregates for a substantial number of proteins differentially accumulated in the nbr1 and chip mutants (Figure 7A; Tables S1 & S2). The levels of protein aggregates in the nbr1 mutant for a substantial number of proteins including Rubisco actvase and catalases were 3–5 times higher than in the chip mutant after 6-hr heat stress (Figure 7A). Other proteins including a group of light-harvesting complex subunits, on the other hand, preferentially accumulates as aggregates in the chip mutant after the same period of heat stress (Figure 7A). A previously reported study has shown that CHIP and HSC70-4 specifically target degradation of the plastid-destined light-harvesting protein subunit proteins in a plastid import mutant [50].
Rubisco activase and catalases are known to be highly heat sensitive and prone to form aggregates [51], [52], [53]. After 6-hour heat stress, the two proteins preferentially accumulated as protein aggregates in the nbr1 mutants but not in the chip or rpn1a proteasome mutant (Figure 7A; Figure 8). This finding suggests a critical factor for selection of protein substrates by the two pathways: those highly aggregate-prone proteins such as Rubisco activase and catalases are efficiently cleared by selective autophagy only most likely because these protein aggregates/oligomers are difficult to be dissociated and unfolded to pass through the small 13 Å wide central cavity of the barrel-shaped 20S proteolytic core [54] (Figure 11). Soluble misfolded proteins such as the cytosolic precursors of LHC proteins can apparently be unfolded and, therefore, can be efficiently degraded by CHIP-mediated UPS [50] (Figure 11). After extended heat stress, differential accumulation of protein aggregates was reduced even though the total levels of protein aggregates increased (Figure 5A). This result is consistent with a complementary nature of the two pathways in clearing misfolded proteins (Figure 11). When there is an extended heat stress and, consequently, increased levels of misfolded proteins, soluble misfolded proteins normally targeted by CHIP-mediated UPS can increase and form aggregate due to limited capacity of UPS and accumulated if selective autophagy is blocked as in the nbr1 mutant (Figure 11). Likewise, aggregate-prone proteins such as Rubisco activase and catalases normally targeted by NBR-mediated selective autophagy can also increase in the chip mutant after extended heat stress because CHIP deficiency leads to accumulation and aggregation of soluble misfolded proteins at increased levels, which could overwhelm the capacity of NBR1-mediated selective autophagy. Consistent with the complementary nature of the two pathways, the levels of aggregates for a majority of detected proteins were higher in the chip nbr1 double mutant than in the chip and nbr1 single mutants (Figures S1 & S2).
Using the well-established GFP-ATG8 system, we showed that chip mutant plants became more sensitive to heat stress for induced autophagosome formation (Figure 10). Furthermore, changes in the transcript levels for some of the autophagy genes in the chip mutant plants occurred earlier than in wild-type plants in response to heat stress (Figure S2). Thus, CHIP deficiency led to quicker induction of autophagy in stressed plants for degradation of accumulated toxic protein species. These results indicated that CHIP- and NBR1-mediated pathways are not only complementary in function but also coordinated in regulation. Similarly, impairment of the UPS has been found to induce autophagy in vitro in non-plant organisms. When cultured human cells are challenged with excess misfolded proteins that overwhelms the UPS or treated with proteasome inhibitors, induction of autophagy is observed as evidenced by redistribution of ATG8/LC3 into punctate structures and accumulation of autophagosomes [55], [56]. Similar induction of autophagy is also observed in response to genetic impairment of the 26S proteasome in Drosophila [57]. How the UPS and autophagy are coordinated in both function and regulation is little understood. Some studies have suggested that misfolded proteins, if accumulated, can be actively transformed into a cytoplasmic, juxtanuclear structure called aggresomes [58]. Several proteins including histone-deacetylase 6 (HDAC6), P62, and Alfy (autophagy-linked FYVE) implicated in aggresome formation and clearance have also emerged as potential players in mediating the crosstalk between the UPS and autophagy [58]. In Arabidopsis, compromised heat tolerance due to CHIP or NBR1 deficiency is associated with increased accumulation of insoluble ubiquitinated protein aggregates (Figures 3 & 5). Identifying components important for the formation, detection and ubiquitination of stress-induced protein aggregates will provide valuable insights into the functional and regulatory coordination between the UPS and selective autophagy in plant stress responses.
Protein aggregates have been extensively studied in animal systems because of their roles in a wide variety of diseases called amyloidosis including Alzheimer's, Parkinsons's and prion disease [58]. By contrast, there has been no reported effort to identify systematically stress-denatured or damaged proteins and protein aggregates in plants. From proteomic profiling, we have identified a number of proteins that were highly accumulated as insoluble protein aggregates in heat-stressed chip and nbr1 mutants (Tables S1 & S2). One of these abundant protein aggregates is Rubisco activase, which activates Rubisco by facilitating the ATP-dependent removal of sugar phosphates from Rubisco active sites. Rubisco activase is known to be highly heat-sensitive and aggregate-prone with a temperature optimum for ATP hydrolysis of 44°C compared to >60°C for carboxylation by Rubisco [53]. The finding that Rubisco activase is among the most abundant aggregated proteins in heat-stressed plants supported its highly aggregate-prone nature and its critical role for the sensitivity of photosynthesis to inhibition by heat [53]. Rubsco activase may have additional biological functions including protecting the thylakoid associated protein synthesis machinery against heat inactivation and repressing leaf senescence [59]. Arabidopsis knockout mutant seedlings for Rubisco activase were yellow, severely stunted and unable to set seeds, supporting its crucial role in plant growth [59]. Proteomic profiling also revealed that CAT3 and CAT2, the major catalase isoforms in photosynthetic Arabidopsis tissues [38], accumulated abundantly as insoluble protein aggregates in heat-stressed chip nbr1 mutants (Tables S1 & S2). Western blotting confirmed that catalases were preferentially accumulated as insoluble, inactive proteins in the chip nbr1 mutants after 6- and 9-hour heat stress (Figure 9). Reduction in soluble, active forms of Rubisco activase and catalases would lead to inhibited photosynthesis and increased oxidative stress, which are known to occur in heat-stressed plants [60], [61].
Proteotoxicity has been generally attributed to non-specific interactions of nonnative proteins with functional proteins and cellular structures such as membranes [1]. Monitoring the changes of the transcripts, proteins and activity of Arabidopsis CAT2 and CAT3 revealed a possible regulatory circuit at multiple levels that ultimately leads to large reduction of cellular catalase activity under heat stress. Western blotting revealed that CHIP and NBR1 deficiency in the chip nbr1 mutant increased the levels of total and insoluble catalase proteins but reduced the levels of soluble catalase proteins after heat stress (Figure 9). Reduced levels of soluble, active catalases in heat-stressed chip nbr1 mutant plants indicated that CHIP- and NBR-mediated pathways impact not only the levels of misfolded and aggregated protein targets but also the levels of their native counterparts. This effect on soluble native catalases could be due to direct physical interactions with accumulated misfolded proteins, thereby promoting catalase protein aggregation and inactivation. In addition, we observed that the levels of CAT2 and CAT3 transcripts in the chip mutant plants were substantially lower than those in the wild-type plants even during the relatively early stages of heat stress (Figure S3). Therefore, reduced levels of soluble, active catalases could also be due to reduced expression of the CAT genes caused by compromised protein degradation in heat-stressed chip mutant plants. In light of these findings, proteotoxicity due to compromised protein quality control can be mediated not only by direct physical interactions with cellular molecules but also by indirect effects on important cellular processes such as gene expression.
A substantial number of proteins identified from aggregated proteins are involved in protein synthesis, folding and maturation (Tables S1 & S2). The highly heat-sensitive and aggregation-prone nature of some of these proteins such as translation factors could underlie the detrimental effects of heat stress on plant growth and development. Aggregation of some of these proteins may also act as adaptive mechanisms for reprograming transcription and protein synthesis in response to abiotic stresses. Other proteins such as heat shock proteins function as molecular chaperones that may be involved in folding or refolding of soluble misfolded proteins and became associated with aggregated proteins when refolding is unsuccessful. Other proteins with activities in protein folding and maturation may be associated with insoluble proteins for facilitation of protein aggregation. During severe and prolonged stress conditions, a large number of denatured and damaged proteins are expected to be generated and can overwhelm UPS due to limited capacity of ubiquitination and degradation, leading to a highly proteotoxic environment in stressed cells. Active promotion of aggregation of denatured and damaged proteins may reduce proteotoxicity, particularly if the protein aggregates are sequestered subcellularly. From western blotting of catalases, only a small fraction of the high levels of aggregated catalases accumulated in heat-stressed chip nbr1 mutant was ubiquitinated (Figure 9A). It appears that unlike in UPS, ubiquitination of only a fraction of proteins in a protein aggregate is sufficient for targeted degradation by NBR1-mediated selective autophagy. As ubiquitination is ATP-dependent, targeted degradation of aggregated proteins by selective autophagy is, therefore, more energetically efficient than UPS and this advantage could be potentially important for plants under stress conditions, which often inhibit photosynthesis and promote respiration.
The Arabidopsis mutants and wild-type plants used in the study are all in the Col-0 background. The nbr1 and rpn1a mutants have been previously described [39], [62]. Homozygous chip-1 (Salk_048371), chip-2 (Salk_059253), cat2-1 (Salk_076998) and cat3-1 (GABI_110C11) mutants were identified by PCR using primers flanking the T-DNA insertions listed in Table S3. The chip nbr1 double mutant was generated through a genetic cross between the chip-1 and nbr1-1 single mutants. Arabidopsis plants were grown in growth chambers at 22°C, 120 µE m−2 light on a photoperiod of 12-hour light and 12 h dark.
Total plant RNA isolation and reverse transcription were performed as previously described [20]. qRT-PCR was performed using the iCycler iQTM real-time PCR detection system (Bio-Rad, Hercules, CA, USA) and the relative gene expression was calculated as previously described [55]. The Arabidopsis ACTIN2 gene was used as internal control. Gene-specific primers for qRT-PCR are listed in Table S4.
For testing heat tolerance, Arabidopsis Col-0 wild type and mutant plants were placed in 22°C and 45°C growth for 9 hours and then moved to room temperature for 3–5 day recovery for observation of heat stress symptoms or survival rates. For testing tolerance to oxidative stress, six weeks-old Arabidopsis plants were sprayed with 20 µM methyl viologen (MV) and kept under light for two days before the picture of representative plants was taken. For testing ABA sensitivity or salt tolerance, sterilized Arabidopsis seeds were sown on solid ½× MS medium or on the ½× MS medium containing 0.5 µM ABA or 150 mM NaCl. Germination rates were determined by scoring green cotyledons for the following 8 days.
Arabidopsis leaves were collected before and after heat treatment, ground in liquid nitrogen and homogenized in a detergent containing extraction buffer (100 mMTris/HCl, pH 8.0, 10 mM NaCl, 1 mM EDTA, 1% Triton X-100, 0.2% ß-mercaptoethanol). Soluble and detergent-resistant insoluble proteins were separated through low-speed centrifugation as previously described [20]. Protein fractionation by SDS-PAGE and western blotting for detection of ubiquitinated proteins and catalases were performed as previously described [20]. Ubiquitinated proteins were detected by protein blotting using an anti-ubiquitin monoclonal antibody (Sigma, USA). RCA was detected with a previously generated monoclonal antibody [40]. Catalases were detected using an anti-catalase monoclonal antibody (3B6) that was raised against tobacco catalases [51], [63]. The antigen-antibody complexes were detected by enhanced chemiluminescence using luminal as substrate as previously described [20].
Transgenic Col-0 wild-type and atg7 plants expressing a GFP–ATG8a fusion construct were previously described [20]. To generate transgenic chip mutant plants expressing GFP–ATG8a, the fusion construct was transformed into the chip-1 mutant using the floral-dip method and transgenic plants were identified on the basis of kanamycin resistance and confirmed by RNA blotting using the GFP DNA fragment as a probe. For visualization of induction of autophagy, 5-weeks old transgenic plants expressing the GFP-ATG8a fusion gene were treated with or without heat shock for various amounts of time and recovered for 0.5 hour. The leaves of transgenic plants were observed using LSM710 confocal microscope with excitation at 488 nm, and images were superimposed using ZEISS LSM710 software.
Insoluble proteins were isolated and separated by SDS-PAGE. Destained gels were dried in a vacuum centrifuge and the in-gel proteins were reduced with 10 mM dithiothreitol (DTT) in 100 mM NH4HCO3 for 30 min at 56°C and then alkylated with 200 mM iodoacetamide in 100 mM NH4HCO3 in the dark at room temperature for 30 minutes. In-gel proteins were digested overnight in 12.5 ng/µl trypsin in 25 mM NH4HCO3. The peptides were extracted three times with 60% acetonitrile (ACN)/0.1% trifluoroacetic acid (TFA), pooled and dried by a vacuum centrifuge. LTQ Velos (Thermo Scientific) equipped with a micro-spray interface was connected for eluted peptides detection. Data-dependent MS/MS spectra were obtained simultaneously. Each scan cycle consisted of one full scan mass spectrum (m/z 300–1800) followed by 20 MS/MS events of the most intense ions with the following dynamic exclusion settings: repeat count 2, repeat duration 30 seconds, exclusion duration 90 seconds. MS/MS spectra were automatically searched against the Unprot ARATH database (53847 sequences, March 8th, 2013) using the BioworksBrowser rev. 3.1(Thermo Electron, San Jose, CA.). Protein identification results were extracted from SEQUEST outfiles with BuildSummary. The peptides were constrained to be tryptic and up to two missed cleavages were allowed. Carbamidomethylation of cysteines were treated as a fixed modification, whereas oxidation of methionine residues was considered as variable modifications. The mass tolerance allowed for the precursor and fragment ions were 2.0 and 0.8 Da, respectively. The protein identification criteria were based on Delta CN (≥0.1) and cross-correlation scores (Xcorr, one charge≥1.9, two charges ≥2.2, three charges ≥3.75).
Total and soluble proteins from Arabidopsis leaves were prepared as previously described [20] and used for the determination of the catalase activity. All of the steps were performed at 4°C. An aliquot of the extract was used to determine the protein content, following the method as previously described [64]. Catalase activity was measured as a decline in A240 using the method as previously described [65]. The spectrophotometric assays were conducted using a SHIMADZU UV-2410PC spectrophotometer (Shimadzu Co., Kyoto, Japan).
Sequence data for the genes described in this study can be found in the GenBank/EMBL data libraries under the accession numbers shown in parentheses: CHIP (At3g07370), ACTIN2 (AT3G18780), ATG5 (At5g17290), ATG6 (At3g61710), ATG7 (At5g45900), ATG8a (AT4G21980), ATG9 (At2g31260), ATG10 (At3g07525), ATG18a (At3g62770), NBR1 (AT4G24690), CAT2 (At4g35090), CAT3(At1g20620), RPN1a (At2g20580).
|
10.1371/journal.pntd.0004998 | Analgesic Effect of Photobiomodulation on Bothrops Moojeni Venom-Induced Hyperalgesia: A Mechanism Dependent on Neuronal Inhibition, Cytokines and Kinin Receptors Modulation | Envenoming induced by Bothrops snakebites is characterized by drastic local tissue damage that involves an intense inflammatory reaction and local hyperalgesia which are not neutralized by conventional antivenom treatment. Herein, the effectiveness of photobiomodulation to reduce inflammatory hyperalgesia induced by Bothrops moojeni venom (Bmv), as well as the mechanisms involved was investigated.
Bmv (1 μg) was injected through the intraplantar route in the right hind paw of mice. Mechanical hyperalgesia and allodynia were evaluated by von Frey filaments at different time points after venom injection. Low level laser therapy (LLLT) was applied at the site of Bmv injection at wavelength of red 685 nm with energy density of 2.2 J/cm2 at 30 min and 3 h after venom inoculation. Neuronal activation in the dorsal horn spinal cord was determined by immunohistochemistry of Fos protein and the mRNA expression of IL-6, TNF-α, IL-10, B1 and B2 kinin receptors were evaluated by Real time-PCR 6 h after venom injection. Photobiomodulation reversed Bmv-induced mechanical hyperalgesia and allodynia and decreased Fos expression, induced by Bmv as well as the mRNA levels of IL-6, TNF-α and B1 and B2 kinin receptors. Finally, an increase on IL-10, was observed following LLLT.
These data demonstrate that LLLT interferes with mechanisms involved in nociception and hyperalgesia and modulates Bmv-induced nociceptive signal. The use of photobiomodulation in reducing local pain induced by Bothropic venoms should be considered as a novel therapeutic tool for the treatment of local symptoms induced after bothropic snakebites.
| Envenoming caused by Bothrops snakes is characterized by drastic local tissue damage involving hemorrhage, blistering, myonecrosis, prominent inflammatory response and intense pain. The most effective treatment for Bothrops snakebites is antivenom therapy, which is very efficient in reversing systemic effects of envenomation but not the severe local effects. Thus, there exists a need to find novel complementary therapies that may further assist in the prevention or even counteract the severe local effects of bothrops snakebite. Several studies have shown the effectiveness of photobiomodulation in reducing local effects induced by Bothropic venoms, however its mechanisms still remain unknown. In this study, we analyzed the effectiveness of photobiomodulation in reducing BmV-induced mechanical allodynia and hyperalgesia as well as part of the mechanisms involved in such effect. Results demonstrate that photobiomodulation reduces venom-induced mechanical allodynia and hyperalgesia and this effect depends on a decrease of nociceptor activation at the spinal cord level and by a modulation of pro- and anti- inflammatory cytokines as well as kinin receptors at mRNA transcriptional levels. These findings make phtobiomodulation a promising candidate to be associated to antivenom therapy for the treatment of the local response induced by Bothrops venoms.
| Bothropic envenomation is characterized by severe local manifestation associated with oedema, myonecrosis, hemorrhage and intense pain [1–4] caused by the toxic action of venom components and aggravated by induced-inflammation. The local effects induced by bothropic venoms are the result of multifactorial and synergistic actions of toxins, which are still poorly understood. Bothrops moojeni is a venomous snake responsible for most of the snakebites in the Central region of Brazil [5]. Despite the medical importance, there are only a few studies related to the local inflammatory reaction caused by Bothrops moojeni venom (Bmv). In this sense, the literature shows that in the accidents caused by these snakes serious local complications occur, including a prominent edema formation, intense pain, swelling and pallor, which may develop into more severe outcomes such as muscle mass loss, neuropathy, and amputation [6, 7].
Currently, the most effective treatment for Bothrops snakebites accidents is the antivenom therapy (AV). However, although AV has proven to be effective in reversal the systemic response, its administration does not prevent local effects and resultant disabilities [3]. Consequently, there is a need to find therapeutic approaches associated with AV treatment that can be effective in reducing the local effects caused by Bothrops snakes envenoming in order to minimize or prevent the progression to a severe clinical status observed after Bothrops snakebites [8, 9].
Photobiomodulation is a form of light that triggers biochemical changes within cells, where the photons are absorbed by cellular photoreceptors and triggers chemical alterations [10]. The mechanisms of photobiomodulation essentially rely on particular visible red and infrared light waves in photoreceptors within sub-cellular components, particularly the respiratory chain within mitochondrial membranes due to the activation of various transcription factors by the immediate chemical signaling molecules produced from mitochondrial stimulation [11]. The most important of these signaling molecules are thought to be Adenosine Triphosphate (ATP), cyclic-AMP, nitric oxide (NO) and Reactive Oxygen Species (ROS) [12].
Many studies have demonstrated analgesic and anti-inflammatory effects provided by photobiomodulation in both experimental [13, 14] and clinical trials [15, 16]. Photobiomodulation has also proven to be an interesting and efficient complementary alternative for the treatment of local effects caused by bothropic venom through the ability of decreasing the observed local effects, such as myonecrosis [17, 18]; inflammation [19–22] hemorrhage [21] and pain [20, 23]. In this context, we have recently demonstrated that photobiostimulation with LLLT and light emitting diode (LED) reverse edema formation, local hemorrhage and inflammatory hyperalgesia induced by Bohtrops moojeni venom (BmV) in mice [18, 24].
Although some studies have demonstrated the effectiveness of photobiomodulation in reducing hyperalgesia and allodynia induced by bothropic venom, the mechanism involved in this effect still remains unknown. In this context, the present experiments were designed to investigate the antinociceptive effect of photobiomodulation on BmV-induced allodynia and hyperalgesia and to explore possible underlying mechanisms.
Male Swiss mice weighing 20–25 g, age-matched, were used throughout this study. Animals were maintained under controlled light cycle (12/12 h) and temperature (21 ± 2°C) with free access to food and water.
All animal experimentation protocols received the approval by the Ethics Committee on the Use of Animals at of Hospital Sírio-Libanês (Protocol no. (CEUA 2010/01), in agreement with Brazilian federal law (11.794/2008, Decreto n° 6.899/2009). We followed institutional guidelines on animal manipulation, adhering to the “Principles of Laboratory Animal Care” (National Society for Medical Research, USA) and the “Guide for the Care and Use of Laboratory Animals” (National Academy of Sciences, USA).
Bothrops moojeni venom (Bmv) was supplied by the Serpentarium of the Center of Studies of Nature at UNIVAP. Bmv was lyophilized, kept refrigerated at 4°C and diluted in sterile saline solution (0.9%) immediately before use. Bmv was injected into the subplantar surface of the right hind paw at the concentration of 1.0 μg/50 μL. Equine antivenom (AV) used in the experiments was a polyvalent Bothrops AV (lot# 990504–18) raised against a pool of venom from B. alternatus, B. jararaca, B. jararacussu, B. cotiara, B. moojeni and B. neuwiedi obtained from the Butantan Institute (São Paulo, SP, Brazil). AV was injected through the intravenous route (0.2 μL of AV diluted in saline; final volume of 50 μL, considering that 1 mL of AV neutralizes 5 mg of Bothropic venom [25] 30 min after BmV injection.
Hyperalgesia and allodynia of the hind paw were assessed as described by Takasaki et al. [17]. Mice were placed individually in plastic cages with a wire bottom, which allowed access to their paws. To reduce stress, mice were habituated to the experimental environment one day before the first measurement. At the day of the test, the animals were placed in the cages 30 min before the beginning of each measurement and received an injection of 1.0 μg of crude Bmv diluted in 50 μL of sterile saline into the subplantar surface of the right hind paw. Control group animals received the same volume of sterile saline. Von Frey filaments with bending forces of 0.407 g (3.61 filament—allodynia stimulus), 0.692 g and 1.202 g (3.84 and 4.08 filaments—hyperalgesia stimulus) were pressed perpendicularly against the plantar skin and held for 5 s, at 1, 3, 6 and 24 h after venom injection. A stimulation of the same intensity was applied three times to each hind paw at intervals of 5 s. The responses to these stimuli were ranked as follows: 0, no response; 1, move away from von Frey filament and 2, immediate flinching or licking of the hind foot. The nociceptive score was calculated as follows:
Nociceptive score (%) = Σ(average score of each animal) x 1002 x number tested animals
Animals were returned to their home cages with free access to food and water between the 1 and 3 h, 3 and 6 h and 6 and 24 h measurements.
A low-level semiconductor Ga-As laser, Theralaser D.M.C. (São Carlos, SP, Brazil), operating with a wavelength of red 685 nm, was used through the experiments with a beam spot of 0,2 cm2 and an output power of 30 mW, energy density of 2.2 J/cm2 and exposure time of 15 s. Laser doses, low enough to avoid any thermal effect, were chosen on the basis of previous study from our laboratory [18]. Animals were gently manually restrained and the LLLT was applied to the same area as the injection of Bmv or saline solution. A control group was treated using the same experimental procedure but with the laser turned off. Animals were irradiated 30 min and 3 h after subplantar injection of either Bmv or saline and were immediately returned to their home cages with free access to food and water after each application.
Experiments were conducted in an environment with partial obscurity to not suffer interference from external light. The output power of the laser equipment was measured using the Laser Check1power meter (MM Optics, São Carlos, Brazil).
Six hours after the intraplantar (i.pl.) injection of Bmv or saline, mice were deeply anesthetized with ketamine hydrochloride (100 mg/kg) and xylazine (10 mg/kg) and transcardially perfused with phosphate-buffered saline and 4% paraformaldehyde in 0.1 M phosphate buffer (PB; pH 7.4). The spinal cord (L4 and L5) was removed, left in the same fixative for 5–8 h and then cryoprotected overnight in 30% sucrose. Thirty μm frozen sections were immunostained for Fos expression. The spinal cord sections were incubated free floating with a rabbit polyclonal antibody against the nuclear protein which is the product of the early response gene c-fos (Ab-5; Calbiochem, CA/USA), and diluted 1:1000 in PB containing 0.3% Triton X-100 plus 5% of normal goat serum. Incubation with the primary antibody was conducted overnight at 24°C. After three washes (10 min each) in PB, the sections were incubated with biotinylated goat anti-rabbit sera (Vector Labs, Burlingame, CA) diluted 1:200 in PB for 2 h at 24°C. The sections were washed again in PB and incubated with the avidin-biotin-peroxidase complex (ABC Elite; Vector Labs). After the reaction with 0.05% 3–3’ diaminobenzidine and a 0.01% solution of hydrogen peroxide in PB and intensification with 0.05% osmium tetroxide in water, the sections were mounted on gelatin- and chromoalumen-coated slides, dehydrated, cleared, and coversliped. The material was then analyzed on a light microscope, and digital images were collected. A quantitative analysis of the immunolabeled material was analyzed using a light microscope and the NIS Elements F3.0 Image analysis system (Nikon Instruments Inc., USA). A quantitative analysis was performed on the density of nuclei representative of thle immunoreactivity for Fos (Fos-IR) in: a) the dorsal horn of the spinal cord (DHSC; superficial laminae-I to IV according to the classification of Rexed. Measurements were taken from 10 different sections for each animal analyzed, including areas that were defined for each structure by using a 20 x objective for the DHSC. Measurements were performed with the program Image J and the operator was blinded to the animal treatment group.
Total RNA was isolated from subplantar muscles and spinal cord by TRIzol reagent (Gibco BRL, Gaithersburg, MD), according to the manufacturer's protocol. RNA was subjected to DNase I digestion, followed by reverse transcription to cDNA, as previously described [26]. PCR was performed in a 7000 Sequence Detection System (ABI Prism, Applied Biosystems, Foster City, CA) using the SYBRGreen core reaction kit (Applied Biosystems). Primers used are described in Table 1.
Quantitative values for IL-6, IL-10, TNF-α, kinin B1 and B2 receptors, CAPDH and mRNA transcription were obtained from the threshold cycle number, where the increase in the signal associated with an exponential growth of PCR products begins to be detected. Melting curves were generated at the end of every run to ensure product uniformity. The relative target gene expression level was normalized on the basis of GADPH expression as endogenous RNA control [27]. Results are expressed as a ratio relative to the sum of GAPDH transcript levels as internal control.
Results were expressed as the mean±SEM. Statistical analyses of data were generated by using GraphPad Prism, version 4.02 (GraphPad). A value of p<0.05 indicated a significant difference. Statistical comparison of more than two groups was performed using analysis of variance (ANOVA), followed by Bonferroni’s test. Statistical comparison for treatment over time was performed using two way ANOVA followed by Bonferroni’s test.
We initially investigated the effects of photobiomodulation on the allodynia and hyperalgesia induced by Bmv. We found that animals injected with Bmv showed significant mechanical allodynia and hyperalgesia when compared with baseline measurement taken before the test, as indicated by basal threshold in response to stimulation by von Frey filaments observed from 1st h after Bmv injection up to 24 h (Fig 1). Photobiomodulation treatment applied 30 min and 3 h after Bmv injection reversed mechanical allodynia of mice in all evaluated times (Fig 1A). Regarding hyperalgesia, LLL was able to interfere with mechanical sensitivity evaluated by 3.84 filament in all evaluated times (Fig 1B) however, for the 4.08 filament the reversion of hyperalgesia was observed only at the 3rd h of evaluation (Fig 1C). AV treatment itself did not interfere with mechanical sensitivity of mice (Fig 1).
As demonstrated in Fig 2, intraplantar administration of Bmv induced a significant increase of Fos immunoreactivity observed in the dorsal horn of the spinal cord of animals injected with Bmv (42.75 ± 3.26) when compared to the saline group (10.65 ± 1.61). Photobiomodulation treatment significantly decreased Fos expression (26.58 ± 3.58; Fig 2).
Cytokine production was evaluated on samples obtained from either spinal cord or footpad of animals previously evaluated at the nociceptive tests. As shown in Fig 3, the mRNA concentrations of IL-6 and TNF-α increased significantly at 6 h after Bmv injection in the footpad of mice when compared with control group (Fig 3A and 3B). After laser treatment, a significant reduction of both IL-6 and TNF-α mRNA levels was found. Moreover, treatment with AV did not significantly interfere with either IL-6 or TNF-α mRNA levels. However, concomitant treatment of mice with AV and phtobiomodulation decreased both IL-6 and TNF-α mRNA levels (Fig 3A and 3B). Furthermore, no changes on IL-6 and TNF-α were observed in samples from spinal cord of mice (Fig 3D and 3E). IL-10 mRNA levels were decreased after Bmv injection on both footpad and spinal cord of mice. Photobiomodulation treatment increased IL-10 levels in both footpad and spinal cord samples (Fig 3C and 3F). AV treatment did not interfere with IL-10 levels, however it prevented the decrease of this cytokine on samples from spinal cord (Fig 3F).
A significant increase on mRNA expression of kinin B1 receptors was observed on Bmv-treated mice when compared to the control group (Fig 4A). LLLT, AV and the association of LLLT and AV induced a significant decrease of mRNA levels of kinin B1 receptors when compared with Bmv-treated animals (Fig 4A). Kinin B2 receptors mRNA expression was also significantly increased in envenomed mice paw when compared to control group (Fig 4B). Once again, LLLT or AV treatment decreased mRNA levels of B2 kinin receptors. More interestingly, the combination of LLLT and AV was more effective in decreasing B2 levels when compared with AV itself (Fig 4B).
The most effective treatment for venomous snakebites accidents is antivenom therapy. However, it is well known that such therapy is effective in neutralizing only the systemic effects of envenomation, without interfering with the severe local effects induced by these venoms [3]. Thus, given the importance framework triggered by local envenoming caused by bothropic venom and the incapacity of the antivenon to neutralize them, it is essential to investigate alternative therapies, with the greatest effectiveness in delaying progression and decreasing local symptoms of envenomed victims.
Among clinical symptoms induced by bothrops snakebites, local pain is a common and clinically relevant manifestation to the patient [28, 29]. Therefore, in this study, we investigated the capacity of photobiomodulation in reducing the nociceptive response caused by Bmv in mice footpad as well as the mechanisms involved. Herein, the intraplantar injection of Bmv induced mechanical allodynia and hyperalgesia. These results are in accordance with previous data demonstrating that Bmv induces potent mechanical allodynia and hyperalgesia in mice [24, 30]. Photobiomodulation applied 30 min and 3 h after Bmv reversed both mechanical allodynia and hyperalgesia. From these data, we confirmed that photobiomodulation, in fact, is effective in reducing Bmv-induced local pain. In our study, as in previous studies [24, 30], we observed that antinociception was not related to AV treatment, since it was not able to interfere with mechanical sensitivity of mice. Also, the association of LLLT and AV did not modify the effect of LLLT alone, reinforcing the therapeutic potential of LLL in treating local effects induced by bothrops venoms.
To better understand the capacity of photobiomodulation to decrease nociception, we evaluated the expression of Fos protein in the dorsal horn of the spinal cord of mice. The expression of proto-oncogenes from the c-fos, c-jun, and erg-1 family are extensively used as tools for the expression of enhanced activity of nociceptive neurons [20, 21]. Our results demonstrate that the intraplantar administration of Bmv induced a significant increase of Fos expression, observed in the dorsal horn of the spinal cord, which is characteristic of nociceptor activation. According to the results of this study, photobiomodulation not only significantly inhibited Bmv-induced mechanical allodynia and hyperalgesia, but also decreased nociceptor activation at the spinal level. More interestingly, we showed here that photobiomodulation is able to interfere with the transmission of Bmv-induced pain message to the central nervous system, reducing nociceptor activation at the central level. This result reveals sensory neurons as an important cellular target for photobiomodulation in the context of pain. In addition to nociceptor-mediated effects, other mechanism(s) may also take part in the antinociception observed in our experimental model. We hypothesized that photobiomodulation may reduce the inflammatory cytokines in the paw and spinal cord. Therefore, the next experiment was designed to further validate the proposed hypothesis.
It is commonly believed that proinflammatory cytokines such as TNF-α and IL-6 are involved in the pain process and that their peripheral and central levels are up-regulated in many pain models [31, 32]. In addition, as described in previous studies, Bothropic venom induces the accumulation of pro-inflammatory IL-6 and TNF-α cytokines in the local of venom injection, which contributes to the enhancement of local tissue damage [1, 33, 34]. Moreover, some studies suggest that the analgesic effect of LLLT may be due to the anti-inflammatory activity by the inhibition of inflammatory mediators [13, 35, 36]. Hence, to further analyze the mechanism by which photobiomodulation reduces nociception of mice induced by Bmv, the expression of pro-inflammatory IL-6 and TNF-α cytokines was evaluated on samples obtained from either footpad or spinal cord of animals. Our results showed that photobiomodulation was able to reduce IL-6 and TNF-α gene expression in the footpad of animals. Also, we showed that associated treatment of AV and LLLT induced the same decrease on IL-6 and TNF-α mRNA levels as the observed with LLLT alone. Moreover, no changes on IL-6 and TNF-α mRNA levels were observed in samples from spinal cord of mice, thus suggesting that inhibition of hyperalgesia depends on a peripheral inhibition of inflammatory cytokines. This result corroborates the study of Ferreira et al. (2005) [13] that proposed that the analgesic effect of LLLT involves the inhibition of hyperalgesic mediators.
Regarding IL-10, we observed that Bmv injection decreased IL-10 mRNA levels on both footpad and spinal cord samples. Also, LLLT increased IL-10 mRNA levels in both footpad and spinal cord. AV treatment did not interfere with IL-10 levels on samples from footpad of mice. However it prevented the decrease of this cytokine on samples from spinal cord. From these data, we confirmed that AV prevents systemic effects induced by Bmv however it did not protect against local hyperalgesia. IL-10 is considered a regulatory cytokine, related to the control of the inflammatory process due to its capacity of inhibiting the proinflammatory cytokine secretion [37]. Results presented herein suggest that laser irradiation was able to modulate the expression of this regulatory cytokine, both in the local of venom injection and in the spinal cord, and it appears likely that this modulation plays a role in the anti-nociception observed after bothropic venom in response to photobiomodulation.
To further analyze the mechanism by which photobiomodulation reduced Bmv-induced nociception, we evaluated the kinin receptors levels in the footpad of mice. Both kinin B1 and B2 receptors, evaluated here, play a central role in the pathophysiology of inflammation [38]. Kinin B2 receptors are broadly and constitutively expressed in most tissues, whereas B1 receptor is weakly expressed in most tissues under basal conditions but strongly upregulated following inflammation [39]. The involvement of bradykinin on Bmv-induced hyperalgesia and edema has been demonstrated [7, 40]. In addition, it was already demonstrated that the kinin B2 receptors are involved in hyperalgesic response induce by B. jararaca and B. asper venoms [22, 41]. Our results demonstrate that both B1 and B2 kinin receptors are increased in the footpad of animals injected with Bmv. Among the treatments, we found that both LLLT and AV were able to reduce the expression of B1 and B2 kinin mRNA levels. However, the association of LLLT and AV showed greater effectiveness in reducing B2 kinin receptors. Considering that kinin receptors are important mediators on Bothrops-induced hyperalgesia [22, 23] it is feasible to suggest that photobiostimulation reverses Bmv-induced hyperagesia, at least in part, by modulating bradikinin receptors involved in the process.
We conclude that photobiomodulation with low level laser is effective in decreasing nociceptor activation at the spinal level. Moreover LLL is effective in modulating pro- and anti-inflammatory cytokines as well as kinin receptors at mRNA transcriptional level. These effects, at least in part, contribute to the decrease of hyperalgesia observed after Bmv. Photobioestimulation with the parameters used herein should be considered as a potential therapeutic approach for the treatment of local effects of Bothrops snakebite.
|
10.1371/journal.pgen.1007401 | CRP-cAMP mediates silencing of Salmonella virulence at the post-transcriptional level | Invasion of epithelial cells by Salmonella enterica requires expression of genes located in the pathogenicity island I (SPI-1). The expression of SPI-1 genes is very tightly regulated and activated only under specific conditions. Most studies have focused on the regulatory pathways that induce SPI-1 expression. Here, we describe a new regulatory circuit involving CRP-cAMP, a widely established metabolic regulator, in silencing of SPI-1 genes under non-permissive conditions. In CRP-cAMP-deficient strains we detected a strong upregulation of SPI-1 genes in the mid-logarithmic growth phase. Genetic analyses revealed that CRP-cAMP modulates the level of HilD, the master regulator of Salmonella invasion. This regulation occurs at the post-transcriptional level and requires the presence of a newly identified regulatory motif within the hilD 3’UTR. We further demonstrate that in Salmonella the Hfq-dependent sRNA Spot 42 is under the transcriptional repression of CRP-cAMP and, when this transcriptional repression is relieved, Spot 42 exerts a positive effect on hilD expression. In vivo and in vitro assays indicate that Spot 42 targets, through its unstructured region III, the 3’UTR of the hilD transcript. Together, our results highlight the biological relevance of the hilD 3’UTR as a hub for post-transcriptional control of Salmonella invasion gene expression.
| Salmonella infection is one of the major causes of foodborne illness worldwide. During infection, Salmonella expresses a set of virulence genes encoded in discrete regions of the genome. The expression of these genes is tightly regulated, being specific for different stages of the Salmonella infection process. While many regulatory mechanisms that lead to the activation of infection-related gene expression have been described, little is known about silencing mechanisms under conditions when expression is not needed and may rather represent a burden than a benefit for the bacterial fitness. Here, we report a condition-specific silencing mechanism of bacterial virulence. That is, the global transcriptional regulator CRP-cAMP represses, indirectly through a post-transcriptional mechanism, the expression of the major Salmonella virulence regulator HilD. In bacteria, post-transcriptional regulation has so far been mainly focused on 5’ untranslated regions (5’UTR). Remarkably, here we describe a molecular mechanism targeting the 3’untranslated region (3’UTR) of the mRNA of the major regulator of Salmonella virulence by a small non-coding RNA under the transcriptional control of the global regulator CRP-cAMP. Our data highlight the importance of 3’UTR in the regulation of gene expression in bacteria.
| Salmonella enterica serovar Typhimurium is a prevalent gastrointestinal pathogen. Upon arrival in the intestinal lumen, Salmonella is able to both invade epithelial cells and survive within phagocytic cells. Genomic studies revealed the presence of several pathogenicity islands in the Salmonella chromosome (SPIs). Among them, SPI-1 and SPI-2 are the best studied and known to encode factors required for invasion of non-phagocytic cells and survival within macrophages, respectively [1,2]. Induction of virulence programs is generally associated with significant energetic costs for the bacterial cell. For example, induction of SPI-1 under non-infectious conditions in vitro has a negative impact on cell physiology, resulting in a deleterious effect on Salmonella’s growth [3]. Consequently, the expression of virulence programs is generally tightly regulated and induction occurs only upon sensing of a variety of defined environmental and physiological signals.
The complex regulatory circuit that controls SPI-1 expression has attracted much attention [2,4] and become a model to understand how the activities of multiple molecular factors converge to achieve a precise timing of virulence gene activation. The majority of the multiple signal transduction systems that modulate SPI-1 regulation converge at the level of HilA expression, a SPI-1 encoded transcriptional regulator required for the expression of most SPI-1 genes [5]. Salmonella does not express HilA when it is growing exponentially in LB cultures, a condition stated as non-permissive in the present study. However, HilA expression is induced at early stationary phase, when growth conditions become nutrient-limiting [6], a condition here referred to as SPI-1-permissive. Transcription of hilA is controlled by three AraC-like transcriptional activators: HilD, HilC and RtsA. The first two are encoded within SPI-1 itself, while RtsA is encoded outside this locus [7]. HilD, HilC and RtsA form a feed-forward regulatory loop, whereby each activator induces the two other genes, but also auto-regulates its own expression [8]. This regulatory triad responds to a wide range of physiological and environmental stimuli that are sensed by a variety of cellular factors, including both global and specific regulators (as reviewed by Fabrega and Vila [2]). Within this triad, a prominent role has been attributed to HilD, the main target for signaling pathways controlling SPI-1 expression [8,9]. Regulatory mechanisms have been described, acting at all levels of hilD gene expression—transcriptional, post-transcriptional, translational and post-translational [10–13]. Most studies focused on the mechanisms required for full induction of SPI-1 genes, whereas very little is known on the regulatory pathways involved in the shutdown of the SPI-1 expression under non-permissive conditions. Here, we report that general transcription factor CRP is required to silence SPI-1 genes in exponential growing cells and propose a new regulatory axis formed by CRP and the broadly conserved small RNA (sRNA) Spot 42 that contributes to growth phase-specific activation of SPI-1 genes.
CRP is a global transcriptional regulator that acts as a metabolic sensor upon binding of intracellular cAMP (cyclic adenosine monophosphate), which is synthesized by the adenylate cyclase CyaA [14]. CRP-cAMP-deficient Salmonella strains are unable to secrete SPI-1 T3SS effector proteins and are avirulent in a mouse model, suggesting a role for CRP-cAMP in the regulation of Salmonella virulence [15,16]. Indeed, CRP-cAMP indirectly regulates virulence by affecting the post-transcriptional regulation of hilD. The sRNAs CsrB and CsrC are under the transcriptional control of Bar/SirA and are upregulated in a crp knockout mutant. CsrB and CsrC are antagonists of CsrA, a post-transcriptional repressor of hilD mRNA [10,17,18]. Therefore, in early stationary phase (permissive conditions for SPI-1 expression), CRP-cAMP generally promotes SPI-1 expression by indirectly repressing the activity of CsrA [17–19]. Here, we report that in mid-logarithmic growth phase (non-permissive conditions), CRP-cAMP represses hilD expression by a mechanism requiring Hfq and the 3’UTR of hilD mRNA. Given the established primary role of Hfq in mediating the base pairing interactions of sRNAs [20,21], it is tempting to speculate that hilD may be post-transcriptionally regulated by a CRP-cAMP controlled sRNA; this control, however, would be unusual in light of the fact that almost all Hfq-associated sRNAs characterized to date recognize mRNAs in the 5’ region. Of several candidates for CRP-cAMP-dependent sRNAs known in enteric bacteria [22], Spot 42 has been best characterized in Escherichia coli, where together with CRP-cAMP, it forms a multi-output feedforward loop to enact catabolite repression [23–27]. In Salmonella, Spot 42 has been known as one of the most abundant Hfq-associated sRNAs in fast-growing cells [28], but except for a repression of the sugar-related mglB mRNA [29] its activity has not been characterized.
By dissecting the molecular mechanism of CRP-cAMP-mediated SPI-1 repression in exponentially growing Salmonella, we here reveal novel mechanisms by which sRNAs target mRNAs. Our data point towards an unusual post-transcriptional stimulation of the hilD mRNA by Spot 42. Different from other trans-acting sRNA characterized, Spot 42-mediated activation occurs in the 3’ UTR of the hilD mRNA, adding to a growing appreciation of mRNA 3’ ends as sites for post-transcriptional control in bacteria.
To characterize the role of the metabolic sensor CRP-cAMP in SPI-1 expression, we monitored transcription of the main regulator HilA in wild-type and Δcrp derivative strains grown in LB at 37°C. The expression pattern was determined in mid-logarithmic cultures (OD600nm 0.4, non-permissive conditions for SPI-1 expression) and at early stationary phase (OD600nm 2.0, permissive conditions for SPI-1 expression) (Fig 1A). Consistent with previous reports, a growth phase dependent profile in SPI-1 expression was observed [6]. In the wild-type strain, hilA expression levels were 8-fold higher at early stationary phase when compared to mid-logarithmic cultures. Remarkably, we also observed a growth-phase dependent effect of the Δcrp mutation. In agreement with previous work [18], the Δcrp mutation reduces hilA transcription in early stationary phase. In mid-logarithmic cultures, however, the Δcrp mutation caused an upregulation of hilA expression (4-fold, as compared to the wild-type). Using a chromosomally encoded FLAG-tagged HilA variant, these transcriptional profiles were corroborated on the protein level. More HilA protein accumulated in the Δcrp mutant in mid-logarithmic cultures and less in early stationary cultures, relative to wild-type levels (Fig 1B).
CRP is active upon binding of the cofactor cAMP [14]. Therefore, lack of either CRP or cAMP should have a similar effect on SPI-1 expression. HilA levels were monitored in a Δcya mutant strain, which is deficient in the synthesis of cAMP (Fig 1C). As expected, Δcya mutation caused an almost 4-fold increase in HilA levels in mid-logarithmic phase cells. Chemical complementation was performed by monitoring HilA abundance after addition of cAMP (Fig 1C). An 8-fold decrease in HilA levels was observed when cAMP was added to cultures of the Δcya strain. Interestingly, when cAMP was added to a culture of a cya+ (i.e. wild-type) strain, a 2-fold drop in HilA levels was observed. These results may indicate that the intracellular cAMP levels were not saturating all CRP molecules. Consequently, external addition of the cofactor to wild-type cultures would lead to an increase in the number of CRP-cAMP complexes, causing further repression of HilA expression.
To further corroborate the involvement of cAMP in the control of HilA expression, the intracellular levels of cAMP were lowered by ectopically over-expressing CpdA in Salmonella, a putative cAMP phosphodiesterase [30]. Over-expression of CpdA, confirmed by immunodetection (S1 Fig), caused a 4-fold increase in HilA expression in the wild-type background. This clearly depended on cAMP turnover, since CpdA over-expression had no effect in a Δcya strain (Fig 1D).
HilA regulates the transcriptional expression of most SPI-1 genes, including those required for the synthesis of a type III secretion system (T3SS) and several effector proteins that are translocated to the host cell during Salmonella infection [2]. A ΔhilA mutation impairs secretion of SPI-1 effector proteins [31]. Comparative studies of the secreted protein profile between wild-type and ΔhilA strains were performed to identify protein bands corresponding to SPI-1 effectors (S2 Fig). Major protein bands exclusively detected in extracts of the wild-type strain were identified by LC-MS/MS as the SPI-1-encoded proteins SipA and SipC. The secretome of wild-type, Δcrp and Δcya derivative strains was characterized in LB cultures grown to mid-logarithmic and early stationary phase (Fig 2A). Consistent with previous reports [16], CRP-cAMP-deficient cells in early stationary phase showed a lower amount of those secreted proteins. Yet, CRP-cAMP-deficient Salmonella hyper-secreted SPI-1 effector proteins in mid-logarithmic phase cultures. The Δcrp-dependent overproduction in mid-logarithmic phase of the larger protein, the effector protein SipA, was confirmed by using a SipA-3xFLAG variant (Fig 2B).
HilA has also been reported to regulate the expression of SopE, an effector protein that is encoded outside the SPI-1 locus but it is secreted by the SPI-1 encoded T3SS [32]. SopE levels were monitored in secreted protein extracts of wild-type and Δcrp mutant strains (Fig 2A). Again, the Δcrp strain secreted more SopE protein in the mid-logarithmic phase and less in early stationary phase, as compared to wild-type. The fact that CRP-cAMP-mediated repression of hilA expression has a concomitant effect on the expression and secretion of SPI-1 effector proteins highlights the biological relevance of CRP-cAMP in the control of Salmonella virulence under non-permissive conditions. In support of this notion, a Δcrp mutant grown to mid-logarithmic phase prior to infection, invaded HeLa cells more efficiently (>10-fold) than the wild-type (Fig 2C). In contrast, the wild-type strain showed a higher rate (>4.5-fold) than the Δcrp derivative when cultures were grown to early stationary phase prior to infection.
Three AraC-like transcriptional activators, HilD, HilC and RtsA, are directly involved in hilA activation [8]. To determine at which level CRP-cAMP controls SPI-1 through HilA, the expression of hilD, hilC and rtsA mRNA was monitored. RNA was extracted from mid-logarithmic cultures (OD600nm 0.4) of both wild-type and Δcrp derivative strains and the relative amounts of mRNA of all three AraC-like regulators were determined by qRT-PCR. As shown in Fig 3A, in the Δcrp mutant higher transcripts levels of hilD, hilC and rtsA were detected than in the wild-type, indicating that the effect of CRP-cAMP on SPI-1 expression occurs upstream of HilA. The hilA transcript was also monitored by qRT-PCR as a control; as expected, it too over-accumulated in the Δcrp strain (S3 Fig).
HilD, HilC and RtsA form a feed-forward regulatory loop to stimulate hilA expression. In order to elucidate the direct target of CRP-mediated regulation of SPI-1, mutants of each of the three regulators in a crp proficient and deficient background were generated and hilA expression monitored in mid-logarithmic cultures (Fig 3B). In the crp+ strains, hilA expression was very low in all genetic backgrounds, further validating the silenced hilA expression during logarithmic growth. The hilA derepression in the absence of CRP was altered in the different mutants. Remarkably, in the absence of HilC and HilD the derepression of hilA transcription was greatly reduced. The expression of hilC and hilD was further studied using transcriptional lacZ fusions. As expected, both hilC and hilD were deregulated in a Δcrp mutant background (Fig 3C). Particularly, HilD seems to be required for the observed upregulation of hilC in the Δcrp background, whereas hilD induction does not require HilC. Taken together, these results suggest that HilD is the direct target of CRP-cAMP-mediated regulation of SPI-1 expression. Northern blot detection further corroborates an increase in the levels of hilD mRNA in the Δcrp strain as compared to wild-type in mid-logarithmic phase (Fig 3D).
The hilD mRNA possess an unusually long (310 nt) 3’UTR that has an overall negative effect on hilD expression [11]. If the 3’UTR is deleted, the hilD mRNA accumulates and the SPI-1 genes are induced concomitantly [11]. Of note, the above-described effect of CRP-cAMP on hilD transcription (Fig 3C) was elucidated using a hilD-lacZ fusion at position +1,235 (relative to the transcription start site), containing the hilD coding sequence and the full-length 3’UTR. To determine whether the hilD 3’UTR is important in CRP-mediated regulation, a proximal fusion at position +76 was constructed. Remarkably, the Δcrp mutation had no effect on this proximal fusion (Fig 4A). This indicates that either CRP-cAMP does not regulate hilD expression at the level of transcription initiation or that the HilD protein is required for the induction of transcription initiation in a crp-deficient strain. To discriminate between these two possibilities, a hilD-lacZ fusion at position +965 was constructed, carrying the whole hilD coding sequence but lacking the hilD 3’UTR. As shown in Fig 4A, the Δcrp mutation did not lead to a significant induction even when the full coding sequence was included (hilD965-lacZ), as compared to a 6-fold induction detected using the hilD1235-lacZ fusion which includes both the hilD coding sequence and its 3’UTR. Although we cannot fully rule out a potential effect of CRP on hilD transcription, the different behavior of the hilD965-lacZ and hilD1235-lacZ reporters clearly points towards the 3’UTR being crucial for CRP-mediated post-transcriptional regulation of hilD.
The relevance of the 3’UTR in CRP-mediated regulation of HilD expression was supported by i) a Δcrp-dependent increase in the levels of HilD-3xFLAG protein was only detected when the hilD-3xFLAG mRNA contained the 3’UTR (Fig 4B) and ii) similarly, the Δcrp-dependent increase in SipA levels was only detected in strains carrying a hilD allele with the 3’UTR (Fig 4C). Additionally, the transcriptional expression of the SPI-1 gene sipC can be monitored as a proxy for HilD activity in the cell, since sipC upregulation in a crp mutant strain requires the presence of HilD (S4 Fig). Consistently, sipC-lacZ was upregulated in a Δcrp mutant background only when the hilD allele carried its native 3’UTR (Fig 4D). Our data also demonstrate that, according to the role attributed to the 3’UTR in the expression of hilD mRNA [11], there was an increase in the levels of HilD-3xFLAG, SipA-3xFLAG and sipC-lacZ when the 3’UTR was removed as compared to the parental strains carrying the native hilD mRNA containing the 3’UTR.
Based on the fact that CRP-cAMP is a transcriptional factor, it is surprising that the CRP-mediated regulation of HilD occurs at the post-transcriptional (and not the transcriptional) level, requiring the hilD 3’UTR. In other words, the data shown suggest that CRP-cAMP modulates hilD expression by an indirect mechanism. In line with previous reports [11,33], we found that the Δcrp-dependent activation of hilD expression, as judged by the hilD1235-lacZ fusion (containing the 3’UTR), was impaired in the absence of the sRNA chaperone Hfq (Fig 5A). Similarly, the drastic increase (16-fold) in sipC expression caused by the deletion of crp was abolished in the absence of Hfq (Fig 5B). We thus hypothesized that Hfq-dependent sRNA may be involved in the CRP-mediated regulation of hilD.
In search for candidate sRNAs in Salmonella, we focused on Spot 42 (encoded by the spf gene) which is transcriptionally controlled by CRP-cAMP in the closely related species, E. coli [34]. Work by the Storz and Valentin-Hansen laboratories had established this sRNA to be a general repressor of sugar-related mRNAs during CRP-mediated catabolite repression [23]. In Salmonella, Spot 42 is highly abundant, with maximal expression in mid-logarithmic phase and reduced upon entry into stationary phase [35]. We tested by Northern blot whether Spot 42 is under CRP-cAMP control also in Salmonella (Fig 5C). In the mid-logarithmic growth phase a 12-fold upregulation of Spot 42 sRNA was detected in the CRP-deficient compared to the wild-type strain. This pattern was further validated using a chromosomal spf-lacZ fusion (S5 Fig). Additionally, spf expression was assessed at early stationary phase. Interestingly, the spf induction detected in the Δcrp derivative strain in mid-logarithmic phase was no longer observed in early stationary phase (S5 Fig), reflecting the divergent effects observed for CRP on SPI-1 expression in exponential versus early stationary phase.
To establish whether Spot 42 is involved in the CRP-cAMP-mediated regulation of hilD, expression studies in strains either deficient in Spot 42 or over-expressing the sRNA were performed. In the absence of CRP, a partial but significant drop in the upregulation of hilD in the Spot 42-deficient background (Δspf) was detected (Fig 5D). In contrast, ectopically expressing Spot 42 stimulated hilD expression. This suggests that Spot 42 is indeed involved in CRP-mediated repression of hilD. Importantly, over-expression of Spot 42 caused a 3-fold increase in hilD expression only when the 3’UTR was present, which implicates the hilD 3’UTR as a previously unknown target of this sRNA (Fig 5E). Consistently, a 2.5-fold increase in HilD-3xFLAG levels was detected upon the over-expression of Spot 42 (S6 Fig).
In agreement with previous data (Fig 5A, [29,35,36]), the positive effect of Spot 42 on hilD requires the chaperone Hfq (Fig 5E). As it has been shown before [29], Hfq binds to both Spot 42 and the hilD 3’UTR (S7 Fig). In addition, the major endoribonuclease RNase E has been suggested to play a role in 3’UTR mediated silencing of hilD expression [11]. Accordingly, hilD induction upon over-expression of Spot 42 was partially lost in the rne537 background encoding an RNase E with a truncated C-terminal domain that is defective in degradosome assembly [37] (Fig 5E). These results suggest that both Hfq and RNase E are involved in the Spot 42-mediated effect on hilD expression.
The involvement of Spot 42 in the control of SPI-1 gene expression was further assessed by examining the transcriptional activity of a sipC-lacZ reporter. Transcription was monitored in either strains carrying the native hilD (+UTR) or strains from which the hilD 3’UTR had been removed (-UTR). As shown in Fig 5F, there was a 5-fold induction of sipC-lacZ upon over-expression of Spot 42 in the +UTR background, whereas sipC transcription was unaffected when Spot 42 was over-expressed in a background lacking the hilD 3’UTR (-UTR).
To confirm that the hilD 3’UTR is targeted by Spot 42, the hilD 3’UTR was cloned downstream of the gfp coding sequence expressed from a constitutive promoter. Expression of this genetic reporter was monitored in either the presence or absence of Spot 42. Co-expression of the sRNA led to a nearly two-fold increase in fluorescence, suggesting that Spot 42 targets the hilD 3’UTR regardless of the genomic location of the latter (S8 Fig). Overall, these results led us to conclude that Spot 42 sRNA activates hilD expression (either directly or indirectly) in a manner that requires the presence of the hilD 3’UTR.
Spot 42 from E. coli and Salmonella share 98% sequence identity. Three unstructured regions (denoted I, II and III, Fig 6A) of Spot 42 have been identified in E. coli to participate in gene regulation through base-pairing interactions [24]. To dissect the mechanism of action of Spot 42 on hilD expression, we determined if specific regions within the sRNA were essential for regulation. The software IntaRNA [38], developed to search for putative interaction sites between two given RNA molecules, predicted an interaction between unstructured region III of Spot 42 and positions 1,129–1,138 of the hilD mRNA (i.e. a region within the 3’UTR). To test whether this unstructured region III of Spot 42 is required for the regulation of SPI-1 genes, two Spot 42 mutant variants were generated, spf-mut1 and spf-mut2 (Fig 6B). Over-expression of these Spot 42 derivatives was performed in strains carrying a deletion of the endogenous spf gene, and their effect on hilD expression was monitored by determination of i) hilD1235-lacZ expression, ii) hilD mRNA levels by qRT-PCR, and iii) sipC-lacZ expression as a readout for HilD activity.
In accordance with our previous results, over-expression of wild-type Spot 42 (Spot 42WT) upregulated hilD1235-lacZ and sipC-lacZ expression. Likewise, relative hilD mRNA levels were elevated upon Spot 42WT over-expression (Fig 6C). Conversely, over-expression of neither Spot 42mut1 (spf-mut1) nor Spot 42mut2 (spf-mut2) induced hilD1235-lacZ expression, hilD mRNA levels or sipC expression, demonstrating that mutations in region III disrupt the stimulatory effect of Spot 42 on hilD expression. (Fig 6C). The substitutions introduced to generate spf-mut1 (GUA-CAU) and spf-mut2 (GGA-CAC) have previously been described in E. coli, where those substitutions were shown to retain Spot 42 steady-state levels [37,38]. Similarly, Northern blots showed that these mutations did not dramatically affect Spot 42 stability in Salmonella either (Fig 6D), arguing that the reduced capability of the Spot 42 mutant variants to induce SPI-1 was not due to lowered sRNA levels. Taken together, the results indicate that the unstructured region III of Spot 42 is required for the regulation of hilD expression.
The in silico prediction suggests that region III of Spot 42 interacts within the hilD 3’UTR, between positions 1,129 and 1,138 of the hilD mRNA. Accordingly, we generated two chromosomal compensatory mutations in the hilD 3’UTR that restore the base pairing of Spot 42mut1 or Spot 42mut2 with the putative target sequence within hilD. The mutant alleles were designated hilD 3’UTRmut1 and hilD 3’UTRmut2, respectively (Fig 6B). sipC expression was used as a readout for HilD activity. Over-expression of Spot 42WT in both hilD 3’UTRmut1 and hilD 3’UTRmut2 genetic backgrounds induced expression of sipC-lacZ, indicating that substitution of those residues within the hilD 3’UTR did not impair the positive effect of Spot 42 on SPI-1 expression. Additionally, over-expression of either Spot 42mut1 or Spot 42mut2 in both hilD 3’UTRmut1 and hilD 3’UTRmut2 backgrounds did not reestablish the ability to induce sipC expression (S9 Fig). Despite unstructured region III of Spot 42 being responsible for hilD activation, these results suggest that the in silico predicted interaction site—positions 1,129–1,138 within hilD mRNA—is not the actual target site or, at least, not the unique interaction site with Spot 42. More complex interaction mechanisms cannot be ruled out such as multiple interactions sites of Spot 42 within the hilD 3’UTR.
Biochemical approaches were used to confirm the physical interaction between Spot 42 and the hilD 3’UTR. The ability of Spot 42 to bind to the hilD 3’UTR-derived fragments was assessed by electrophoretic mobility shift assays (EMSAs). EMSAs of radiolabeled full-length hilD 3’UTR incubated with increasing concentrations of Spot 42 confirmed a direct interaction between the two RNA species (Fig 7A). Consistent with our in vivo data, the unstructured region III of Spot 42 is required for the interaction with the hilD 3’UTR, since the binding affinity of the mutant version of Spot 42 (spf-mut2) was markedly reduced. Next, the hilD 3’UTR was divided into two halves, UTRL and UTRR (Fig 7B). The UTRL fragment spans positions +927 to +1114 of the hilD mRNA (roughly the first half of the hilD 3’UTR), while UTRR covers the second half of it (position +1090 to +1275) and includes the putative interaction site with the unstructured region III of Spot 42 (Fig 6B) as well as two Hfq binding sites as inferred from CLIP-seq [29]. EMSAs with radiolabelled Spot 42 and increasing concentrations of either one of the two UTR fragments were conducted. A concentration dependent upshift of Spot 42 was only observed upon addition of the UTRR fragment with an apparent Kd of 80 nM but not upon addition of the UTRL fragment (Fig 7C), indicating that Spot 42 interacts with the second half of hilD 3’UTR. Again, Spot 42-UTRR interaction was specific as the affinity of the UTRR fragment to the mutant version of Spot 42 (spf-mut2) was markedly reduced (Fig 7D). Further supporting this notion, in the reverse experiment increasing concentrations of Spot 42 did not lead to a band shift of the UTRL but only of the UTRR fragment (S10 Fig). Our results indicate that the loss of hilD activation by Spot 42mut2 in vivo is due to the inability of this mutant sRNA version to directly interact with the hilD mRNA, specifically with the second half of its 3’UTR. Overall, this makes Spot 42 the first Hfq-associated sRNA that potentially activates a trans-encoded target gene via its 3’UTR.
During the infection process, Salmonella relies on the expression of genes encoded on SPI-1 for epithelial cell invasion. Although SPI-1 is therefore crucial for Salmonella infection, it has a retarding effect on the growth rate, presumably as a consequence of the energetically high costs to produce the SPI-1 T3SS [3]. Accordingly, the expression of SPI-1 genes is tightly regulated [1]. Most relevant studies have focused on the regulatory pathways dedicated to induce SPI-1 under permissive conditions. However, as SPI-1 expression affects cell fitness, SPI-1 silencing mechanisms under non-permissive conditions, for instance in fast growing cells in the mid-logarithmic phase, are equally important. In this study, we identified CRP-cAMP, a metabolic sensor and global transcription factor [14,39], as a key player in the repression of SPI-1. CRP-cAMP is involved in a post-transcriptional regulatory circuit, controlling the expression of hilD by a mechanism dependent on its 3’UTR.
Coordination of metabolism and stress-related functions is crucial for the evolutionary success of bacterial populations. In pathogenic bacteria, the cross regulation between virulence factors, which can be considered within-host stress-related factors, and physiology is crucial for efficient colonization. A sudden shift between the expression of genes involved in active growth and genes involved in adaptation to stress might be required for rapid adaptation to changing conditions during the infection process. Secondary messengers such as cAMP, the intracellular levels of which can be altered by the action of both synthetases (adenylate cyclases) and degrading enzymes (phosphodiesterases), provide a rapid response system that can promote rapid changes in the expression profile. Although cAMP has traditionally been described as a regulator of metabolism, its role in the modulation of virulence-related functions has been extensively studied in several pathogens [40]. In E. coli, CRP-cAMP has been described to repress type 1 fimbriae expression during logarithmic growth [41]. Other secondary messengers, such as ppGpp, have also been reported to participate in the interplay between cell metabolism and virulence control [42]. Post-transcriptional regulation by small non-coding RNA confers to the cell another level for a rapid response to environmental conditions, in fact, a number of sRNAs have been found to play a relevant role in the metabolism-virulence crosstalk [43,44].
In this study we found that CRP-cAMP represses SPI-1 expression by modulating the expression of the regulator HilD (Figs 1–3). The role of HilD is not restricted to SPI-1; rather there is a complex cross-talk between HilD and master regulators of other virulence associated pathways. For example, it has been shown that HilD activates, under certain conditions, SPI-2 expression that is required for survival within macrophages [6,45]. Within macrophage-like cells, SPI-1 genes are downregulated and SPI-2 genes are induced [46]. CRP-cAMP is a regulator tailored to mediate rapid responses to environmental changes and may therefore be relevant for HilD-mediated regulation of virulence in response to the environmental conditions that Salmonella encounters through the infection process.
The CRP-mediated regulation of hilD does not occur at the transcriptional initiation level (Fig 4). Rather CRP-cAMP modulates hilD expression at the post-transcriptional level through the long 3’UTR (310 nt) of hilD. Post-transcriptional regulation is an extensively used mechanism to finely regulate virulence in bacterial pathogens [47,48]. The role of 5’UTRs in post-transcriptional gene expression control has been established and it is noteworthy that mRNAs of important SPI-1 regulators, such as invF, hilA and hilE, all carry long 5’UTRs [43,49–51]. In contrast, the involvement of 3’UTRs in post-transcriptional regulation is still poorly understood. Recently, it has been reported in Staphylococcus aureus that one-third of the cellular transcripts carry 3’UTRs longer than 100 nt [52]. In addition to be targets of regulation, 3’UTRs may provide regulators themselves, namely 3’UTR-derived sRNAs [36,53].
The 3’UTR of hilD constitutes a silencing module, since its deletion causes significant hilD upregulation. Although we are yet to elucidate the full molecular mechanism, the observed Hfq dependency suggested that one or several sRNAs are targeting the hilD 3’UTR [11]. Likewise, the CRP-mediated repression of hilD requires both the presence of the hilD 3’UTR and Hfq, indicating that CRP regulates hilD expression in an sRNA-mediated manner (Fig 4 and Fig 5).
Spot 42 is an integral member of the CRP-mediated gene expression network in E. coli [23–27] and its expression is repressed by CRP-cAMP in both E. coli and Salmonella ([47], Fig 5). Here we found that Spot 42 is involved in the CRP-mediated regulation of hilD expression, since the derepression of hilD in a Δcrp strain was diminished in the absence of Spot 42 and over-expressing Spot 42 caused a concomitant upregulation of hilD expression. Remarkably, Spot 42-mediated regulation targets the long hilD 3’UTR. The fact that the absence of Spot 42 did not completely abolish the hilD deregulation caused by the Δcrp mutation points at additional factors that could be involved in the described regulation. Although the nature of these factors remains fully elusive, it should be noted that these putative factors seem to also act through the hilD 3’UTR. Further studies will be required to determine whether other sRNAs or proteins plays a role in the CRP-mediated repression of hilD expression.
Both genetic and biochemical approaches point towards a direct interaction between Spot 42 and the hilD 3’UTR, involving the unstructured region III of Spot 42 (Figs 6 and 7). Although the exact target sequence within the hilD 3’UTR have not been identified, EMSA experiments indicate that the interaction occurs between Spot 42 and the downstream half of the 3’UTR (last 185 nt). This interaction is strongly diminished when the unstructured region III of Spot 42 is altered by base substitution in three positions previously described to be involved in base-pairing [23]. The recent finding that the transcription elongation factors GreA and GreB target the hilD 3’UTR to regulate hilD at permissive conditions [54] led us to speculate that transcriptional pausing might trigger a specific folding of the hilD 3’UTR important for post-transcriptional regulation. Overall, the regulation through the hilD 3’UTR seem to be complex and presumably several factors target the hilD 3’UTR. Although it has been proposed that intrinsic motifs in the long 3’UTR of hilD might confer susceptibility to degradation in a polynucleotide phosphorylase (PNP) and RNase E dependent manner, no effect in the stability of the hilD mRNA was detected [11]. Consistent with these data, we found no difference in the stability of the hilD mRNA between the wild-type and Δcrp strain in mid-logarithmic phase (S11 Fig). The exact mechanism by which these factors converge to regulate hilD expression should be the focus of future studies. Our results highlight the 3’UTR of hilD as a central hub in SPI-1 regulation and indicate that the whole hilD 3’UTR is required for the post-transcriptional regulation of hilD.
To our knowledge, there are no other examples of trans-encoded sRNAs targeting 3’UTRs. Of note, the cis-encoded sRNA GadY, which is encoded on the opposite strand of gadX 3’UTR, seems to positively regulate gadX through interaction with the gadX-gadW intergenic region [55]. Unlike Spot 42 and hilD which are expressed from regions in the chromosome over 1 Mb apart, GadY and gadX physically overlap. Global screens for Hfq-mediated sRNA-mRNA interactions [56,57] suggest, however, that 3’UTR targeting may be more common than currently appreciated.
In conclusion, our findings imply a novel mechanism in the complex regulatory network of SPI-1 expression. Under non-permissive conditions, very low transcriptional expression from the hilD gene occurs. Additionally, CRP-cAMP represses the transcription of the sRNA Spot 42, thereby maintaining basal levels of hilD expression. Consequently, despite hilD transcription occurs, only low levels of HilD protein arise (Fig 8 panel I). In contrast, environmental and/or physiological signals may relieve CRP-dependent Spot 42 repression. Upon binding to its 3’UTR in an Hfq-dependent manner, Spot 42 may exerts a positive effect on hilD mRNA, thereby activating HilD protein expression. As HilD auto-activates itself by promoting its own transcription, expression of some copies of HilD protein would likely be sufficient to amplify the final output (Fig 8 panel II). Thus, CRP-cAMP seems to play a relevant role by coordinating post-transcriptional virulence control in Salmonella. Somewhat similar to the described Spot 42-mediated regulation of SPI-1, the sRNA PinT acts as a timer of virulence gene expression in Salmonella, regulating SPI-2 genes through the modulation of CRP-cAMP [58]. Altogether, this highlights the emerging importance of collaborative activities of general transcription factors and sRNAs to precisely adjust the costly expression of major virulence factors to internal and external metabolic cues [44].
The bacterial strains and plasmids used in this study are listed in S1 Table.
Salmonella enterica serovar Typhimurium SL1344 and derivative strains were cultivated either in Luria Bertani broth (tryptone 10 g/l, yeast extract 5 g/l and sodium chloride 10 g/l). When required, the media was supplemented with ampicillin (Amp) 100 μg/ml, kanamycin (Km) 50 μg/ml, chloramphenicol (Cm) 15 μg/ml, or tetracycline (Tc) 15 μg/ml. Induction of genes cloned into pTRc99a was achieved by adding 0.1 mM IPTG.
Liquid cultures (20 ml of LB in 100 ml culture flasks) were inoculated to an OD600nm of 0.001 and incubated at 37°C with vigorous shaking (200 rpm). An OD600nm of 0.3–0.4 was considered mid-logarithmic phase of growth, while an OD600nm of 2.0 was considered early stationary phase of growth.
The cpdA (SL1344_3157) gene was cloned into the IPTG inducible vector pTRc99a [59]. The cpdA coding sequence was PCR amplified by Phusion polymerase (Invitrogen) with the primers cpdA_XbaI_Fw and cpdA_SalI6xHis_rev (S2 Table), subsequently digested with XbaI and SalI and ligated into XbaI/SalI digested pTRc99a.
The spf gene encoding Spot 42 sRNA was cloned into pBRplac vector [60]. spf was PCR amplified with the primers spf_AatII_Fw and spf_EcoRI_rev (S2 Table), subsequently digested with AatII and EcoRI and ligated into AatII/EcoRI digested pBRplac. Mutations in Spot 42 (spf-mut1 and spf-mut2) were generated by assembly PCR and subsequent cloning in the pBRplac vector.
A gfp-hilD 3’ UTR construct was cloned in the backbone of plasmid pXG1 [61]. The hilD 3’UTR region was fused to gfp by overlapping PCR using chromosomal SV5015 and plasmid pXG1 as templates and the primers gfp_NheI_Fw, gfp_hilD_rev, hilD_gfp_Fw and hilD_XbaI_rev (S2 Table). The purified PCR fragment was subsequently digested with XbaI/NheI and ligated into an XbaI/NheI digested pXG1 vector, resulting in the plasmid pXG1 gfp-3’UTR.
Deletion mutants were generated by gene replacement as described by Datsenko and Wanner [62]. Briefly, antibiotic resistance cassettes carrying either KmR or CmR resistance genes were amplified from pKD4 and pKD3, respectively. Primers used include a 40 bp sequence complementary to the region where the insertion was desired. Purified fragments were electroporated into strains carrying pKD46. Positive clones were selected in presence of the required antibiotic. When desired, the antibiotic resistance cassette was removed by expression of the Flp recombinase from the pCP20 plasmid, as described [63].
Deletion mutants, where the antibiotic cassette was removed, were further used for the generation of reporter gene fusions. Transcriptional lacZ fusions were generated as described [64], the remaining FRT-site was used to integrate plasmid pKG136.
Epitope tagged proteins were constructed as follows: HilA, HilD and SipA 3xFLAG tagged proteins were generated by a λRed recombinase system as described [65]. When desired, the KmR cassette downstream of the 3xFLAG epitope was removed using the Flp recombinase (see above). In the HilD 3xFLAG construct retaining the KmR cassette, the hilD coding sequence and 3’UTR are split and not co-transcribed. Thus, the KmR cassette was removed, with the hilD 3’UTR now located right downstream of the 3xFLAG epitope. Oligonucleotides used to generate the constructs are listed in S2 Table. All strains were PCR confirmed and integrity of the sequence was checked by DNA sequencing.
Chromosomal modifications in the hilD 3’UTR region were generated by scarless mutation [66]. Briefly, a 1 kb fragment containing the hilD 3’UTR sequence was cloned into pGEM vector, desired point mutations were generated via the Quick change method using the hilD-UTR oligonucleotides listed in S2 Table. Subsequently, the vector was digested with SacI/XbaI and ligated into the suicide plasmid pDMS197 [67]. The derivative pDMS197 was propagated in S17-1 lambda pir and used as donor in matings with SV5015 Δspf sipC-lacZ. Trans-conjugants were selected for tetracycline resistance. Selected clones were grown in salt-free nutrient broth supplemented with 5% sucrose. Individual tetracycline-sensitive clones were checked by PCR and subsequent DNA sequencing to select the clones carrying the desired chromosomal mutation.
To obtain whole cell and secreted protein extracts, LB cultures were grown at 37°C and processed as previously described [54]. Samples in Laemmli sample buffer were subjected to SDS-PAGE separation. Normalization of the loading samples was performed based on the culture biomass (OD600nm). Coomassie stain was used to visualize protein bands.
Protein extracts were subjected to SDS-PAGE separation, transfer to PVDF filter and subsequent immunodetection using monoclonal Anti-FLAG (Sigma), Anti-His (Sigma) or polyclonal Anti-SopE [68] as primary antibodies. Commercial polyclonal anti-mouse (Promega) and anti-rabbit (GE Healthcare) secondary antibodies conjugated to horseradish peroxidase were used. For detection, ECL Prime Western Blotting Detection Reagent (GE Healthcare) served as a substrate. Chemi-luminescence was detected using Chemidoc equipment (Bio-Rad). As a control prior to the immunodetection, all whole cell extract samples were analyzed by SDS-PAGE and Coomassie staining to ensure proper normalization of the loaded amounts.
For protein identification, the protein bands from Coomassie stained SDS-PAGE gels were trypsin digested and analyzed by LC-MS/MS by the Proteomic facility from the Scientific Park of Barcelona (PCB).
β-galactosidase activity was measured as described previously [69]. Activity determination was performed in technical duplicates for each of three biological replicates.
For each strain, samples from three independent LB cultures grown at 37°C to mid-logarithmic phase (OD600nm 0.4) were processed. RNA was extracted by classical hot phenol method. RNA quality and concentration was assessed by an Agilent Technologies Bioanalyzer 2100.
Quantitative reverse transcription-PCR (qRT-PCR) was performed as previously described [54]. The relative amount of target cDNA was normalized using the gapA (GAPDH) gene as an internal control. Oligonucleotides used for qRT-PCR are listed in S2 Table.
Electrophoretic separation of total RNA samples was carried out in Tris-Borate-EDTA (TBE) 8% acrylamide gels containing 8.3 M urea. Samples were prepared by mixing 10 μl of RNA samples with 10 μl of urea dye (2x) loading buffer and incubated for 10 minutes at 65°C, immediately chilled on ice and loaded for electrophoretic separation at 30 mA for 2 hours.
RNAs were transferred to Hybond N+ (GE Healthcare) filters by semi-dry TBE based transfer for 2 hours at 400 mA. RNAs were subsequently fixed to the filter by UV crosslinking. Filters were then hybridized with radiolabeled oligos, sequences are given in S2 Table. Images of radioactive filters were obtained with the FLA-5100 imaging system (Fujifilm) and quantification was performed using Image J software.
RNA-RNA interactions were detected by Electrophoretic Mobility Shift Assay (EMSA) as described in [70]. First, DNA templates for in vitro T7 RNA transcription were generated by PCR, primers used are listed in S2 Table. RNA was produced in vitro by following the Megascript transcription procedure from Ambion. Then, either the sRNA (Spot 42) or the target RNAs (hilD 3’UTR, UTRR or UTRL) was dephosphorylated and 5’ labeled with [(α-32P) ATP]. The putatively interacting RNAs were next incubated in structure buffer (Ambion): In a 10 μl final volume, 4 nM of the radiolabeled RNA was incubated with increasing concentrations of the unlabeled RNA (0, 56, 280, 560, 1700nM). Samples were incubated at 37°C for 1 hour and subjected to electrophoresis in a native 6% acrylamide gel. For specific RNA detection, acrylamide gels were dried and exposed. Images were obtained as for Northern blots.
For single-cell analysis, cell cultures were grown to the desired conditions, pelleted and resuspended in PBS. The bacterial suspensions were then fixed in 4% formaldehyde. The fluorescence of 20,000 bacterial cells was measured by flow cytometry using preset parameters for GFP (excitation wavelength of 484 nm and emission wavelength of 512 nm). Measurements were performed in technical duplicates for each three biological replicates; average was used to compare GFP expression.
HeLa human epithelial cells (ATCC CCL2) were cultured in tissue culture medium (Dulbecco’s Modified Essential Medium (DMEM) supplemented with 10% fetal calf serum and 1mM glutamine). HeLa cells were seeded the day before the infection in 24-well plates containing 0.5 ml of DMEM per well and grown at 370°C, 5% CO2. Bacterial cells grown at 37°C to different phases of growth were prepared in DMEM. The bacterial mixture was added to HeLa cells to reach a multiplicity of infection (MOI) of 75 bacteria per eukaryotic cell. 30 minutes post-infection HeLa cells were washed twice with phosphate buffered saline (PBS) and incubated in fresh DMEM medium containing 100 μg/ml gentamicin for 90 minutes. Numbers of viable intracellular bacteria were obtained after lysis of infected cells with 1% Triton X-100, and subsequent plating. Infections were carried out in triplicate. Invasion rate is defined as the intracellular bacteria recovered versus viable bacteria used to infect the HeLa cells (initial inoculum). Invasion rates were normalized to bacterial culture of a wild type strain.
GraphPad Prism 5.0 software was used for data analysis. Two-tailed Student’s t-test were carried out and p-values < 0.05 were considered significant.
|
10.1371/journal.pntd.0007181 | Genetic evidence of Coxiella burnetii infection in acute febrile illnesses in Iran | Mounting evidence suggests that Q-fever is more prevalent in Iran than originally believed. However, in most parts of the country, clinicians do not pay enough attention to Q fever in their differential diagnosis. The aim of this study was to investigate the prevalence of Coxiella burnetii in suspected cases of acute Q fever in north-western Iran using molecular techniques. Febrile patients were enrolled in the study and investigated for C. burnetii infection. Sera samples were tested using real-time PCR for detection of IS1111 gene, and positive samples were confirmed with nested PCR. Nine patients (4.2%) out of 216 suspected cases were positive for C. burnetii. Weakness and fatigue, headache, and lethargy were the most prevalent clinical symptoms in acute Q fever patients. According to the results of this study and other reports of human cases in Iran, the diagnosis system of Q fever in Iran should be urgently revamped.
| Q fever is a zoonotic infectious disease caused by Coxiella burnetii. Domestic ruminants are the most common source of human infection. Main route of transmission to humans is inhalation of infected aerosols and dust with C. burnetii. Acute Q fever is usually presented as a non-specific febrile and self-limiting influenza-like illness, but in severe acute cases, may manifest as atypical pneumonia or hepatitis. In Iran, Q fever is an endemic disease with high seroprevalence among humans and domestic animals. However, human Q fever cases remain undiagnosed in most regions of Iran, especially because most clinicians fail to spot this disease in their differential diagnosis. The aim of this study was to investigate the prevalence of acute Q fever in suspected cases (216 suspected cases) using molecular techniques. Nine acute Q fever patients were diagnosed by Real-time PCR and Nested PCR. Weakness and fatigue, headache, and lethargy were the most prevalent clinical symptoms in positive cases. Human Q fever cases described in this, and previous studies, indicate the need to implement diagnostic techniques for this disease across the country.
| Q fever is a zoonotic infectious disease caused by Coxiella burnetii, an obligate intracellular bacterium. Outbreaks have been reported in both developing and developed countries. Domestic ruminants (cattle, sheep, and goats) are the most common source of human infection [1]. In animals, C. burnetii infection is mostly asymptomatic, but can lead to abortion, stillbirth, infertility, endometritis and metritis. Infected animals shed C. burnetii in urine, feces, milk, and especially in birth or abortion products. Main route of transmission to humans is inhalation of infected aerosols and dust with C. burnetii. Ingestion of contaminated raw milk and dairy products, skin or mucosal contact, tick bites, blood transfusion, sexual transmission, and embryo transfer are less common routes of infection transmission to humans [1, 2].
In humans, the incubation period for primary infection has been estimated to be between 7 and 32 days after exposure [3]. Acute Q fever is the primary form of infection by C. burnetii, and more than half of the patients are asymptomatic. Acute Q fever usually presents itself as a non-specific febrile self-limiting influenza-like illness, but may also manifest as atypical pneumonia or hepatitis [4]. Cardiac involvement, neurological signs, acute lymphadenitis, cholecystitis, autoimmunity, bone marrow involvement, and dermatological signs have also been reported in some acute cases of Q fever [1]. The main manifestation of chronic Q fever is life-threatening endocarditis and vascular infection. Other less common forms of chronic Q fever includes abortion, lymphadenitis, osteomyelitis, prosthetic joint arthritis and osteoarticular infection [1, 5].
Due to the wide range of clinical symptoms of Q fever in humans, clinical diagnosis of the disease can be challenging based on symptoms alone. Therefore, laboratory confirmation is a major and crucial part in the diagnosis of clinical cases [4]. Laboratory diagnosis of Q fever in humans is mainly based on serological tests, ELISA and IFA as the gold standard test. C. burnetii isolation from clinical samples is not performed in most diagnostic laboratories because it requires eukaryotic cell cultures and access to BSL3 facilities. In recent years, PCR-based molecular assays were developed to detect C. burnetii in clinical specimens. PCR-based techniques are more adapted than serology for early diagnosis of acute Q fever because of delay in the antibody response, which is detectable only after 2–3 weeks following infection [1, 6].
In Iran, Q fever is an endemic disease with high seroprevalence among humans and domestic animals [7]. In recent years, many acute and chronic Q fever cases have been reported in Iran [8–11]. Furthermore, several investigations have been published on the prevalence of Q fever among domestic livestock in Iran [7]. However, human cases of Q fever remain undiagnosed in most regions of Iran, especially because most clinicians do not consider this disease in their differential diagnosis.
The incidence of acute Q fever is underestimated in most parts of the world. The clinical presentations in acute Q fever patients is very pleomorphic, nonspecific and confusing. Less than 4% of patients with acute fever require hospitalization [1, 12]. This disease is often disregarded by physicians and healthcare system and diagnosis relies upon the physicians’ awareness of the clinical symptoms of acute Q fever and access to reliable diagnostic laboratory facilities including serology and PCR [4]. Diagnosed acute cases with C. burnetii must be treated promptly to avoid to chronic Q fever [13]. The rapid and timely diagnosis of acute fever can help cure patients and avoid the spreading of the disease. Conducting molecular studies, such as the current study, can help to rapidly diagnosis of patients with acute febrile illness, as well as can raise awareness and sensitivity of clinicians and the health care system about Q fever in Iran. The aim of this study was to investigate the prevalence of C. burnetii in suspected cases of acute Q fever by molecular methods.
The samples of this study were collected from two surveys carried out in Tabriz County in the East Azerbaijan Province (North West of Iran) in 2013 and Ghaemshahr County in the Mazandaran province (Northern Iran) in 2015–2016.
Patients which met the following criteria, were enrolled to the study as suspected acute Q fever cases:
Suspected patients were examined by clinical practitioners, and all the symptoms were diagnosed by them. The clinical symptoms and epidemiological evidences were recorded by practitioners in the questionnaires. Eligible individuals were selected by the practitioners and enrolled to the study based on the inclusion criteria. Demographic characteristics, clinical signs and risk factors were recorded for each participant by a standardized questionnaire developed for this study (S1 Questionnaire). Blood sample was taken from each patient. Sera were subsequently extracted and used for molecular investigation.
This study was approved by the Ethics Committee for Biomedical Research of Tarbiat Modarres University (Ethic Code: IR.TMU.REC.1395.510). The Ethics Committee for Biomedical Research of Tarbiat Modarres University approved the consent procedure, the proposal and protocol of this study, covering all the samples taken (blood), questionnaire and verbal or written informed consent. All participants signed an informed consent: Written informed consent was obtained from adult’s patients and parents of patients below the age of 18. Also, for participants who were illiterate, the consent form was read aloud to them and the interviewer signed the consent form with the permission of these individuals on their behalf.
A 200 μL aliquot of each serum was used for DNA extraction. Genomic DNA was isolated using the Roche High Pure PCR Template Preparation Kit (Roche, Germany), according to the manufacturer's instruction.
All samples were tested by real-time PCR for detection of IS1111 gene of C. burnetii, and positive samples were confirmed with nested PCR (Table 1). Real-time PCR was performed using specific primers and probe sequences targeting IS1111 gene (Table 1). Real-time PCR reactions were performed using the following reaction mixture: 10 μL of 2x RealQ Plus Master Mix for Probe (Ampliqon, Denmark), 900 nM forward primer, 900 nM reverse primer, 200 nM probe and 4 μL of DNA template. Real-time PCR was performed on the Corbett 6000 Rotor-Gene system (Corbett, Victoria, Australia), with a final volume of 20 μL for each reaction. The PCR amplification program were 10 mins at 95°C, followed by 45 cycles of 15 s at 94°C and 60 s at 60°C [14].
Nested PCR method was performed via two runs of PCR using two sets of primers including Trans1 and Trans2 for first amplification followed by 261F and 463R for second amplification reaction. The products of first PCR were separately used as DNA template in a second round of PCR. Each PCR reaction contained 5μL of DNA, 12.5μL Taq DNA Polymerase Master Mix RED (Ampliqon, Denmark), and 10 pmol/μL from each primer in a final volume of 25μL. PCR was performed in a thermal cycler (Bioneer, South Korea). The first amplification of PCR was done at 95°C for 2 min, followed by five cycles at 94°C for 30 s, 66 to 61°C (touchdown assay) for 1 min and 72°C for 1 min. These cycles were followed by 35 cycles consisting of 94°C for 30 s, 61°C for 30 s, and 72°C for 1 min, then a final extension step of 10 min at 72°C. In the second amplification, the cycling conditions included an initial denaturation of DNA at 94°C for 3 min, followed by 35 cycles at 94°C for 30 s, 50°C for 45 s, 72°C for 1 min, then a final extension step of 10 min at 72°C. The amplicons were electrophoresed on 1.5% agarose gel and visualized under UV light [10].
In total, 8500 patients were invited to participate in the study and were screened by the clinical practitioners; among them 235 patients had the clinical and epidemiological sings to be suspected for Acute Q fever as the discussed inclusion criteria. Participants who matched the inclusion criteria were selected randomly. Finally, 216 out of 235 suspected febrile patients were enrolled (138 patients from Tabriz County in East Azerbaijan Province and 78 patients from Ghaemshahr County in Mazandaran Province) (S1 Fig). The age of participants ranged between 2–82 years with a mean age of 41.5. In total, 61.1% of individuals were male and 38.9% were female. Residency in rural and urban regions among participants was 60.1% and 39.9%, respectively. Also, 61.6% of participants had a history of keeping domestic animals (Table 2).
Nine patients’ samples (4.2%) were positive for C. burnetii. All nine sera samples were positive by nested PCR and real-time PCR (Table 3). Prevalence of acute Q fever in Tabriz County and Ghaemshahr County were 3.6% and 5.1%, respectively. Weakness and fatigue (100%), headache (88. 9%), and lethargy (66.7%) were the most prevalent clinical symptoms in positive cases (Table 4). Seven (77. 8%) of nine identified patients had a history of keeping livestock. Also, Seven (77. 78%) of the nine detected acute Q fever cases were female and five (55.5%) were residents in rural areas. The demographic and epidemiological findings and were not statistically significant risk factors for Q fever infection.
This study is the first molecular investigation of human Q fever cases in the north and north-west of Iran. Among 216 investigated febrile patients in this study, 4.2% were confirmed to be infected with C. burnetii. Based on recent evidence, Q fever shows high prevalence in livestock and milk and also a high seroprevalence in many different human populations in Iran. The seroprevalence of IgG phase I and II antibodies of Q fever in human has been reported to be 19.80% and 32.86%, respectively. Also, the prevalence of C. burnetii antibodies in goat, sheep and cattle were reported to be 31.97%, 24.66% and 13.30%, respectively [7]. Despite all evidence, the disease is underestimated by clinicians and the health system in Iran. In fact, most of the clinically diagnosed cases of Q fever have been the outcome of research projects. The results of this study and other reports about human cases in Iran, suggest that the physicians and health care system should pay more attention to diagnosis of Q fever cases in Iran. Special training should be provided for diagnosis of Q fever for clinicians and infectious diseases specialists. In addition to these measures, laboratory diagnostic facilities for the diagnosis of C. burnetii infection should be expanded throughout the country. It is essential that the healthcare system provides the necessary training for people to understand the disease and to prevent it. Patients with suspected clinical symptoms of acute Q fever must be advised to follow up on specific tests as well as on the completion of appropriate treatment. This way, a higher number of suspected Q fever patients will be diagnosed and treated and thus prevent possible progression of the disease toward chronic Q fever. It is noteworthy that all patients in our study diagnosed with acute Q fever were treated with appropriate antibiotics (Doxycycline and Hydroxychloroquine) and all effectively recovered.
In this study, 4.2% of the 216 suspected febrile patients were positive for IS1111 gene of C. burnetii as confirmed by nested PCR and real-time PCR. Prevalence of acute Q fever in Tabriz county (East Azarbaijan province) and Ghaemshahr county (Mazandaran province) were 3.6% and 5.3%, respectively. In a similar study that was conducted in northeastern Iran, 7.4% of 92 patients were positive for C. burnetii, as confirmed by nested PCR [10]. In similar studies conducted in other countries; molecular prevalence of C. burnetii in acute febrile patients were 0.4% in Senegal [15], 4.5% in India [16] and 14.1% in Poland [17]. The low molecular prevalence of acute Q fever in febrile cases in our study compared to other studies may be due to a number of factors, such as differences in geographical location and climate. More comprehensive studies in this region and other regions of Iran can be helpful for accurate estimation of the C. burnetii infection in acute illness.
Acute Q fever generally presents as a flu-like illness with wide range of nonspecific clinical manifestations [4]. Patients with cute Q fever may develop respiratory illness or hepatitis. Pneumonia is an important clinical manifestation of acute Q fever, and C. burnetii might be an underrecognized cause of community-acquired pneumonia [13, 15]. Based on available information and review of the literature, most clinical data of acute Q fever were obtained from patients with Q fever pneumonia [1, 4, 14, 18–20]. Due to the above reasons, we enrolled acute febrile patients with pneumonia (acute lower respiratory tract infections). For future studies, it is recommended that a wider range of clinical symptoms along with pneumonia and undifferentiated fever be considered in order to cast a wider net for the diagnosis of the clinical cases of acute Q fever.
Serologic tests are known as reference methods for diagnosis of clinical cases of Q fever. The reason for the use of serology as detection method is partially the limitation of culture methods in isolation of C. burnetii and also the strong immune response to the infection (the antibody produced against the bacterium) in the human body, which is easily detectable by serological tests [13]. Unfortunately, serology has limitations in diagnosis of acute febrile illnesses, because it requires two serum specimens (from the acute phase and the convalescent period) and looks for a fourfold increase in antibody content in paired serum samples. Access to the second serum sample takes time (approximately 4 weeks) [4]. Molecular tests are an attractive alternative; they allow for rapid, one-step, diagnosis of patients with acute Q fever and can be performed at an early stage of the C. burnetii infection [1, 14]. In our study, we developed a diagnostics assay based on real-time PCR for diagnosis of suspected patients and we used nested PCR for confirmation of positive results by Real time-PCR. All nine positive cases were confirmed with nested PCR. Employing this laboratory diagnostic protocol (real-time PCR) can improve and accelerate primary molecular detection, after which the initial positive results can be confirmed by the nested PCR. It is worth noting that the initial and confirmation tests identify and amplify different regions of the IS1111 gene of C. burnetii, increasing the fidelity of the detection technique. Based on our results, we recommend that molecular tests be combined with the accepted serological tests to diagnose patients with suspected Q fever in shorter time and at earlier stages of the disease.
One of the limitations of our study was the small number of positive cases, which made us unable to do a proper statistical analysis of risk factors and epidemiologic factors. In addition, more precision in the entry of eligible individuals and those who were more closely related to the criteria for diagnosis of acute Q fever, could provide a more precise prevalence of acute Q fever. The combination of molecular tests with serologic tests (as the gold standard diagnostics method) allows for proper identification of all suspected patients. Another limitation of our study was lack of attention to whether antibiotics against C. burnetii were administered during the sampling time. Therefore, it is suggested that the mentioned limitations should be considered in subsequent studies.
|
10.1371/journal.pcbi.1006646 | Epithelial stratification shapes infection dynamics | Infections of stratified epithelia contribute to a large group of common diseases, such as dermatological conditions and sexually transmitted diseases. To investigate how epithelial structure affects infection dynamics, we develop a general ecology-inspired model for stratified epithelia. Our model allows us to simulate infections, explore new hypotheses and estimate parameters that are difficult to measure with tissue cell cultures. We focus on two contrasting pathogens: Chlamydia trachomatis and Human papillomaviruses (HPV). Using cervicovaginal parameter estimates, we find that key infection symptoms can be explained by differential interactions with the layers, while clearance and pathogen burden appear to be bottom-up processes. Cell protective responses to infections (e.g. mucus trapping) generally lowered pathogen load but there were specific effects based on infection strategies. Our modeling approach opens new perspectives for 3D tissue culture experimental systems of infections and, more generally, for developing and testing hypotheses related to infections of stratified epithelia.
| Many epithelia are stratified in layers of cells and their infection can result in many pathologies, from rashes to cancer. It is important to understand to what extent the epithelial structure determines infection dynamics and outcomes. To aid experimental and clinical studies, we develop a mathematical model that recreates epithelial and infection dynamics. By applying it to a virus, human papillomavirus (HPV), and a bacteria, chlamydia, we show that considering stratification improves our general understanding of disease patterns. For instance, the duration of infection can be driven by the rate at which the stem cells of the epithelium divide. Having a general model also allows us to investigate and compare hypotheses. This ecological framework can be modified to study specific pathogens or to estimate parameters from data generated in 3D skin cell culture experiments.
| Stratified epithelia cover most of the human body’s exterior and line the inner cavities, such as the mouth and vagina. Localized (non-systemic) infections of these epithelia can cause a wide range of conditions that collectively represent a major burden on global public health systems. For instance, skin conditions are ranked 4th in global years lost due to disability (YLDs) and are in the top 10 most prevalent diseases globally [1]. Infections (viral, fungal, bacterial, etc.) are either the etiological agents or are secondary opportunistic infections (e.g. scabies, eczema) of many skin conditions and thus play a major role in their burden and outcomes. While stratified epithelia are often the first line of defense against infections [2], their cells are the primary target for many viruses or bacteria. This is why understanding epithelial life-cycles, signaling, and dynamics is an active line of research [3].
Epithelial infections are very heterogeneous in their outcomes, ranging from short sub-clinical acute infections to chronic pathologies [1]. Our hypothesis is that the stratified structure is one of the keys to understanding these patterns. Though experimental and clinical methods used for studying these infections are increasingly quantitative (e.g. flow cytometry or -omics technologies), theoretical frameworks for understanding infection properties and dynamics in stratified epithelia are lacking since most models consider infections of monolayers or blood. Here, we build on the analogy between a host and an ecological system [4, 5] to investigate how the stratification of the epithelium drives infection dynamics. We focus on keratinocyte epithelia as an example as it is a well-studied stratified system with important public health implications.
Localized infections of stratified epithelia such as the cervicovaginal mucosa are involved in a range of health concerns, such as decreasing fertility [6–9] or carcinogenesis [10]. Studying the cervical epithelium has greatly helped improve women’s health [11] and histological studies of cervical infections have characterized both healthy and diseased cells. The ectocervix is a non-keratinized stratified epithelium that acts as an important barrier to prevent infections from entering the upper part of the female genital tract and affecting fertility. The tight packing of the epithelial cells and their migration to the surface are believed to prevent bacteria and viruses from reaching the dermis [12]. Furthermore, the continual production of surface mucus is thought to aid in trapping and removing invaders [13]. Studying these processes using tractable experimental systems has been a challenge given the complexity of recreating stratified epithelia with realistic features, but this is changing rapidly [14]. Mathematical modeling can aid this experimental work by helping to estimate parameters such as changes in cell migration or mucus production rate.
The vast majority of mathematical models of within-host dynamics focus on virulent viruses causing systemic infections, such as HIV (for a review, see [15]), but some investigate pathogens that only (or mainly) target epithelia such as Chlamydia [16–20], HPV [21–24], Epstein-Barr Virus (EBV) [25, 26] or HSV [27]. A common feature of these models is that they focus on the pathogen and the associated immune response, while largely overlooking the epithelium itself. As a consequence, with few exceptions (e.g. [23]), they assume that the population of cells infected by the pathogen is homogeneous and not structured. We take an ecological approach to model the stratified epithelium to investigate the effect of the structure of the life cycle of the host cells on infection dynamics. The analogy between ecological systems and within-host interactions is not new (e.g. [4]), but it is becoming increasingly common and has underlaid successful quantitative tools for understanding viral kinetics [15, 28] and drug resistance [29].
From an ecological perspective, the stratified epithelial structure can be viewed as having stages or age structure (herein called ‘stage-structure’), meaning the full life-cycle of an individual cell is divided up into stages (or ages). Therefore, populations of one stage give rise to another in a successive fashion. Ecological populations with stage-structure have been shown to have rich dynamics [30]. If resource populations (species low in a food chain) are stage-structured, then the resulting dynamics can impact the entire ecological system [30–32]. Generally, oversimplifying (e.g. not considering the stage-structure of the resource) or not considering the resources is known to potentially lead to incorrect predictions about the behavior of the system [33]. Similar importance has been shown in host-pathogen systems. For instance, by combining mathematical models with experimental data Mideo et al. showed that differences between Plasmodium chabaudi strains could be most parsimoniously explained by their different affinity for erythrocytes of different ages, as well as differences in erythropoiesis, i.e. in how red blood cells are made [34]. Target cell heterogeneity has also been put forward to explain the HIV co-receptor switch [35]. While we pursue this analogy, we insist that stratified epithelia exhibit features that differ from traditional populations. For instance, differentiated keratinocytes (or ‘adults’) do not reproduce to make stem cells (or ‘juveniles’) the way free-living species do. Additionally, the epithelium self-regulates its dynamics as a means to maintain homeostasis, which involves the maintenance of constant numbers of cells by physiological processes, such as states of dormancy, proliferation and signaling [3]. Together, this calls for a system-specific approach.
Having a framework for epithelial dynamics allows us to simulate infections. For this, we chose two prevalent stratified epithelium infections with very different biological features: Human papillomaviruses (HPVs) and Chlamydia trachomatis bacteria. In the United States alone, more than 1.5 million cases of C. trachomatis were reported to the Center for Disease Control (CDC) in 2017 and HPVs are the most common sexually transmitted infection in the country [36]. Most HPVs are, in fact, not sexually transmitted and are part of a large family of viruses that infect stratified epithelia throughout the body (of the mucosal and or cutaneous tissues) and are considered part of our virome [37, 38]. While both Chlamydia and HPVs replicate intra-cellularly, these two infections exhibit contrasting strategies for infecting the squamous epithelium: HPVs cause non-lytic basal-up infections, whereas chlamydia infections are from the surface-down and are lytic. As mentioned, there are some previous mathematical models of both HPV and chlamydia [16–24] and, importantly, the biology of these two pathogens have been considerably studied, with well characterized life-histories (HPVs [37] and Chlamydia [39]). Consequently, this provided us with peer-reviewed parameter estimates, biologically grounded assumptions and previous results from mathematical models without epithelium stage-structure with which to compare our results. Finally, to maintain focus on the epithelium, we used a simple model for the immune response, as in earlier studies (e.g. on HSV, [27]).
We address to what extent epithelium dynamics affect infection dynamics and as a result determine infection outcomes. First, we introduce a general epithelium model, which we calibrate using existing data, as well as original cell culture data from a spontaneously immortalized human cell line (NIKS) [40]. With this data we infer parameters that are difficult to measure, such as the fraction of symmetric cell divisions. We then ‘infect’ this epithelial model with chlamydia, wart-associated HPVs and oncogenic (high-risk, HR) HPVs to investigate how protective measures by the epithelium affect infection load and duration, while identifying the parameters that control key infection traits. We find that epithelium stratification plays a key role in the dynamics and outcomes of these infections.
Our model abstracts the stratified epithelium into four phenotypically distinct populations, relevant to clinical and experimental models of the epithelium: stem-like cells in the basal and parabasal layers, and differentiated cells in the mid and surface layers (Fig 1 and equation system 1). These phenotypes can be identified experimentally using immunofluorescence techniques that target genes or proteins expressed differentially as the cells mature and move up the epithelial column. The model is sufficiently generic that it can represent any stratified squamous epithelium, keratinized or not. We considered the cervicovaginal mucosa as an example to parametrize and infect. The model includes 7 parameters, of which 4 are inferred from cervical autoradiographic experiments done in 1970 [41] and one, Nb, is a scaling parameter describing the surface of the basal monolayer considered. The two remaining capture the difference in symmetric divisions probabilities by the basal and parabasal cells (Δp is fixed at zero and Δq is free and calibrated). All the parameters are listed in Table 1.
The parameters for which we have less information are related to the fraction of cells dividing symmetrically (e.g. a parabasal cell produces two daughter parabasal cells or two differentiated cells). Existing data suggests symmetric divisions are expected to be low [42, 43]. This is further reinforced by our estimate of epithelium thickness. Histological studies calculate 26 to 28 cell layers in the vaginal epithelium depending on the stage of the menstrual cycle [44] and in vivo studies of the cervical epithelium count 16 to 17 layers [45]. To achieve comparable values, and assuming that the ranges of the other parameters are biologically plausible, we find that symmetric divisions must be rare. Calibrating Δq ≈ −0.012 gives an epithelium ‘thickness’ of 17 layers, i.e. 17Nb. Analytical results shown in S1 Text revealed the need for some degree of symmetric division biased towards producing differentiated cells (Δq < 0). Furthermore, if we assume each layer of parabasal cells has the same number of basal cells and that the differentiated cells are half the number of cells per layer (because they are twice the size [46]), then 17Nb corresponds to 26 layers. Finally, we found that the mid layers, that is the differentiated, Ud, and parabasal layers, Up, are larger than the basal and superficial, Us, layers. To obtain experimentally relevant parameter estimates, we used our model and the known parameters as priors to estimate values using original data from raft cultures of NIKS (Normal Immortal Keratinocytes) cells. The NIKS cell-line grows into a 3D epithelium structure and is commonly used as a model of cervicovaginal tissue and HPV infections, though they are known to differ from in vivo tissue [40]. Fig 2A and 2B show an example of NIKS cell growth into stratified form. Fig 2C shows the dynamics of the number of basal and suprabasal (non-keratinized and keratinized) cells, along with the inferred dynamics from the model. From this data (of growth from single layer to stratified) the symmetric divisions were inferred to be negligible in the basal layer but important in the parabasal layers (Table 1). This implies then that the constant basal layer assumption, and thus model 1, is appropriate for fitting organotypic culture datasets. The data constrained the replication rate of the parabasal cells, ρp, to be low and the Δq was estimated to be close to −1, suggesting that while the replication rate is low, nearly all parabasal divisions produce two differentiated cells which move up the column (Table 1). This, along with the higher than in vivo estimates for the basal replication rate, ρb, is consistent with a growth phase of an epithelium growing up to homeostasis.
We performed a sensitivity analysis to explore the general behavior of the model and identify the parameters that have the largest effect on homeostasis, i.e. at a dynamic equilibrium without infection (Table 2). This showed that the total number of cells in the layers above the basal layer is mostly governed by the basal cell proliferation rate, ρb. Additionally, the time for the system to reach homeostasis (which is important for repairing damaged tissues) depends on the proliferation rate of the parabasal cells (ρp; S1 Text). Indeed, homeostasis is reached faster when the replication rate, ρp, or the symmetric divisions of the parabasal cells, Δq, are significantly higher, as found from fitting the data and the model simulations (Table 1 and not shown).
Having generated and calibrated a model for epithelial dynamics, we could then simulate infections to investigate how stratification affects important properties of the infection.
Epithelial infections by both chlamydia and HPVs are heterogeneous in their clinical manifestations. Chlamydia infections can be asymptomatic or with clinical manifestations such as cervicitis [47]. The lytic nature of chlamydia infections reduces the epithelium to lower cell numbers than homeostasis, therefore affecting the integrity of the layers (Fig 3). This is consistent with the cervical erosion observed in chlamydia-driven cervicitis or in infections by other lytic pathogens such as HSV [48].
Several HPVs have been found to be associated in wart-like lesions, which are substantive cell overgrowth above homeostasis levels. Among the mucosal α-genus HPVs, HPV6 and HPV11 often (though not always) generate papillary lesions or warts. In cutaneous stratified squamous epithelia several HPVs are associated to warts (e.g. species 2 and 4 of the α-genus and the types of the μ- and ν-papillomaviruses) in various locations, such as the feet and hands [49]. Conversely, HR-HPV types cause flat lesions (yet with a thickening of the epithelium) [37]. How differences between HPV types translate into this observed diversity of clinical manifestations in the epithelium is not always clear. What is clear is that HR-HPVs have stronger cell transforming properties than low-risk (LR) and wart-associated HPVs [37]. The epithelium model allowed us to identify conditions that lead to wart-like manifestations. When assuming that there can be rare events of new virions entering the basal layer (e.g. due to immunosuppression and cytokines loosening epithelial junctions) and that wart-associated types do not drive cell proliferation in lower layers [37], we find that they must either have higher burst sizes (produce more virions per cell) than HR-HPV types or be better at driving differentiated cells back into S-phase in the upper layers (ρa and θ control the peak of infected cells in Table 2). Burst size, θ, also controls how quickly the number of infected cells increases, as does the infection rate, β. This explains why simulations of wart-associated HPVs with higher burst sizes are more effective at reaching basal cells, as illustrated by the differences in shading of basal layers between Fig 4A and 4B. Epidemiological studies that directly compare viral loads of LR vs. HR genital HPVs are needed, however, wart-associated HPVs (mucosal or cutaneous) have higher viral loads in warts than other HPVs [49].
HR-HPVs have enhanced E6 and E7 oncoprotein effects in the lower layers [37]. In spite of this increase in epithelial cell division rate their infections are flat, slow growing, and are often clinically indistinguishable from a normal epithelium for many months. For this to occur, we find that the extra proliferation in the basal, ab, and upper layers, ρa, and the type’s burst size, θ, must be kept low (Table 2 and Fig 4B). This implies HR-HPVs would be less ‘productive’ (shed less virions) than wart-associated HPV during an infection of the same duration (Fig 4C). If HR-HPVs were to have low burst sizes but high oncoprotein-driven proliferation in the lower layers, then their infections would be wart-like (S1 Text, S2 Fig). Thus, to maintain flat lesions the strong HR oncogenes need to be down-regulated. A simulated representation of silent, productive HR-HPV infection is shown in Fig 4B.
For an infection to be sufficiently disruptive to generate a visible manifestation, both the size of inoculum (number of cells infected initially) and how quickly the microabrasion closes from repair appear to matter. For instance, the wart-like overgrowth of cells in Fig 4A can be created either by a small inoculum with slow repair or by a large inoculum and fast repair. When microabrasions close quickly (within a few days) and only a small number of cells are infected initially, both HR and wart-associated types do not cause any visible disruption to homeostasis Fig 4D. Clinically, these infections would be asymptomatic with normal cytology and would likely only be detectable using PCR methods.
Finally, we compared our HPV results with a non-stratified model of HPV infection (see S1 Text) and find that it is unable to reproduce the features associated with HR and non-HR HPV infections if using the same biologically constrained parameter ranges (see S4 and S5 Figs).
For some parameter combinations the kinetics of chlamydia infections had an acute phase only (Fig 3A), as have been observed in guinea pigs and other animal models [18]. We obtain this qualitative pattern most readily when the infection rates are the same for all layers or when the lower layers are difficult to infect (for instance due to the reduced permeability down the epithelium column [12]) and the population of immune effectors grows rapidly. From the sensitivity analysis, duration is longer when the EBs can infect the lower layers, ηu, and shortened when the cell recovery rate is high (Table 2).
We also found an acute phase can be followed by a chronic phase, where a pathogen load stabilizes to a set point value (Fig 3B). How quickly a chronic phase is reached depended on chlamydia’s infection rates of the various layers. Generally, infection rates had to be low to achieve the chronic phase (because if too high then the bacteria burn through the epithelium and its population crashes). Additionally, if the layers are differentially infected by chlamydia (i.e. βb < βp < βu), then the chronic phase is reached earlier (see S3 Fig).
In contrast, long-lasting wart-associated and HR HPV infections did not exhibit an acute and a chronic phase in our model. Instead, they persisted by monotonically reaching an equilibrium (e.g. Fig 4Di and 4Dii). Also, for both HPVs, the immunity killing rate, κ, was the most important factor in determining infection duration (Table 2). With more antigen in the lower layers to detect, the efficiency of immune killing (κ) becomes important for determining duration of infection, speed of growth and size of infected cell accumulation (Table 2).
Finally, while Chlamydia and HPVs can cause either acute or chronic infections [51], our model showed that a clinically detected chronic state is achieved through different underlying dynamic patterns for each pathogen.
Upon infection, epithelia exhibit defense mechanisms such as increasing mucus flow, tightening the packing of cells, migration to the surface [52] and increasing proliferation (promoted by Interleukin-22 cytokines [53–55]). We varied epithelial parameters from their homeostasis value to investigate in detail the effect of such mechanisms on various measures of infection using our infection models for HPV and chlamydia (models 3, 4 and sensitivity analyses in Table 2).
We found some mechanisms had similar effect on both HPVs and chlamydia. First, increasing upward migration of epithelial cells, ν, reduced the maximum pathogen load reached during the infection (Table 2). Second, mucus trapping, ζ, delayed the peak and the duration (although it played a bigger role in decreasing the peak of infection for chlamydia than for HPV). And finally, for all infections, increasing basal or parabasal cell proliferation, ρb and ρp, scored high in affecting all the infection measures, e.g. size of peak or duration (‘effects of epithelial parameters’ in Table 2). However, a pathogen-specific effect was that increasing basal proliferation, ρb, of uninfected cells decreases the time to clear HPVs but not chlamydia. Together, this suggests that epithelial cell features themselves play an important role in infection dynamics and outcomes.
Epithelial infections are a major public health burden, and, in particular, STIs are on the rise causing a worldwide concern [1, 9, 56]. Quantitative models, both experimental and mathematical, are essential in developing our understanding of these infections. As for systemic (and virulent) infections such as HIV and HCV, mathematical models have been developed to predict and analyze the kinetics of epithelial infections. Here, we show that to understand the kinetics of epithelial infections, it is essential to account for the stratified structure of the epithelium, a property that is absent from most models. We illustrated how such a general framework can be combined with 3D cell culture data to estimate key parameters and how it can generate relevant insights regarding the course of epithelial infections.
The rate of basal cell proliferation had a strong effect on the homeostasis of both uninfected and infected epithelia, which suggests an ecological ‘bottom-up controlled’ system [57, 58], analogous to those found in free-living food webs. These bottom-up effects are more apparent if we consider that basal cell replication is strongly determined by the resources that are available in the basal lamina, such as growth factor. While we did not explicitly model the resources of the basal layer (it is implicit in the basal proliferation rates), the growth of the cells in the experimental set-up does depend on concentration and temporo-spatial distribution of growth factors, impacting epithelial thickness and proliferation rates. Therefore, this ecological insight of bottom-up driven systems, could be tested more formally in experimental systems by monitoring resource concentrations.
The key role of bottom-up control is further supported by our finding that accelerating basal cell proliferation, as a response to infection [53, 54], affected all infection measures (e.g. time of peak, total load, duration). This infection response, then, can have a strong effect on the severity and duration of infections. However, using the same response mechanism might be differentially effective depending on the infection strategy of the pathogen. For instance, we found that increasing cell proliferation did not shorten the infection of chlamydia. This is probably because proliferation increases the number of uninfected epithelial cells in the upper layers which, for chlamydia, means more ‘resources’.
Pathogens can have different tropisms for the various cell phenotypes of the stratified epithelium. For instance, EBV more readily infects and replicates in differentiated cells of the upper/mid layers [59], whereas HPV infects the basal layer to establish an infection [37]. We hypothesized that this should impact how effective protective processes (e.g. increased mucus production) of the epithelium are against them. In chlamydia, where the pathogen infects all cell types equally well, we found that tight packing (i.e. epithelial permeability) mattered to the pathology at the site. The speed at which the epithelium shrank and the stability of the infection system (how quickly it can reach chronic phase) depended on how well the bacteria could access cells down the column. If the bacteria were able to infect the bottom of the column quickly, that led to a population crash due to the lack of resources. On the contrary, and somehow unexpectedly, less epithelial permeability stabilized the infection that then lasted much longer and exhibited a clear chronic phase. This stabilizing effect is also observed in ecological systems when one stage is invulnerable to attack, i.e. a stage refugia [32, 60]. For instance, a parasitic wasp was introduced as a biological control of red scales (a common plant pest). It successfully controlled the red scales because one of the mature stage of the red scales was not vulnerable to attack [32]. Such effects from decreasing permeability (protecting the basal replicative stage) would have implications in the context of treatments that bolster cell adhesion and require testing experimentally.
Considering pathogens with contrasted life-histories allowed us to identify how similar infection outcomes arise. In the case of chlamydia, the interaction between free-form chlamydia and its infection rates of the various stages drove the chronic phase, but although the activation of the immune response through the same feedback ultimately led to clearance, this feedback affected several infection features. In contrast, HR-HPV persistence was achieved via a slow growth strategy that delays clearance by decreasing the negative dynamical feedback involving the immune system (i.e. faster growth implies faster immune detection and clearance). Indeed, HPV types appear to evade, or counteract, these immune responses differently. In particular, viral protein E6 of various HPV types differ in their many cellular binding partners resulting in a variety of effects on host processes [61]. We found that the difference between HPV-induced genital warts and epithelial lesions depended most on the number of virions an infected cell releases upon death (or ‘burst size’) and the initial size of inoculum; suggesting that more productive viruses are better colonizers. A ‘colonization’ strategy (in ecology ‘r strategy’) appeared to have a cost for the virus because infecting the basal layer of the epithelium triggers the immune response. While more sites are colonized, each site is exploited less optimally. Another feature that was mediated through the immune response feedback was that mucus trapping delayed the peak of the infection (i.e. the decreased progeny of bacteria and viruses meant less antigen and thus slower immunity detection).
To compare our results to HPV epidemiological studies of acute HPV infections, we see that the model creates underlying patterns (e.g. viral load Fig 4C) that could be looked for using prospective studies of HPV infections with normal cytology. Study designs with dense sampling (with visits every 3 or less months) are best for capturing the dynamics of these infections, particularly for the exponential increase and decay of viral loads. The majority of HPV prospective studies are of persistent infections and with advancing cytological abnormality but there are exceptions. For instance, Marks et al. sampled young women with HPV16 infections every 3 months and found that a greater than 2 log decrease in viral load was associated to clearance and a single viral load measure could not predict clearance [62]. The HR-HPV viral load dynamics from our model (Fig 4C) can provide possible underlying explanations and our exponential decrease is consistent with the decrease found by this epidemiological study. Though, sampling once would not give enough information as to whether the infection is increasing or decreasing at a given point. Consecutive viral load measures, then, are more appropriate to estimate clearance or persistence [63].
The effect of stage-structure on infection dynamics can be interpreted in light of earlier results from ecology or epidemiology. For instance, in epidemiology, it is known that the more a general population of infected host is subdivided into classes, the more rapid the growth rate of the epidemic is and the shorter it lasts [64]. Our model bears even more parallels with age structured models in epidemiology where the age groups of the host population are explicitly considered. In many of these models, children tend to be key to the spread of epidemics [64], a result that echoes the bottom-up effects we identify. However, the driving forces in the two models are different: in our model it is due to the fact that basal cells are the ones replicating, whereas in epidemiology it is usually driven by longer lasting acquired immunity at higher ages.
Spatial structure is a natural extension of our model that could be investigated further. Here, the different cell populations partly capture the vertical structure. A specific consequence of not including space is that the immune system effects are more homogeneous than in reality, where more resident immune cells are present in the lower layers. The assumption of well-mixed populations holds best when the model represents a portion of the squamous epithelium (rather than, for instance, the whole cervix). In the case of patchy infections like HPV, a metapopulation modeling approach may be more appropriate (e.g. [22, 38]) or a full spatial model [21]. We chose not to include space since much of the experimental data available on these systems is not spatial. Instead most are cell population counts from immunofluorescence or flow cytometry techniques. Several mathematical modeling methods, such as agent-based models, are available to study spatial aspects of infections, particularly cell-to-cell spread [65]. These should be of interest to those studying chlaymdia infections. Even though HPVs have not been found to spread cell-to-cell like other viruses [66], studying the spatial aspects of their infections should most certainly still be considered in future studies.
Introducing stochastic aspects in epithelial dynamics have recently refueled the discussion on the determinants of HPV clearance [23]. In general, considering stochastic dynamics could matter most when pathogen populations approach low-levels (i.e. very few infected cells or small loads). For instance, our finding that mucus trapping can delay the peak and the duration of infections could interact with stochasticity. This is similarly true for infections started with a small inoculum, very rapid abrasion closure, and rapid repair with small inoculum. These processes keep the pathogen populations sizes down and thus, as seen in ecological systems, stochasticity should play a larger role in extinction. As for the spatial structure, it is important to stress that there often is little data on the initial stage of the infections, when the pathogen is rare.
Many previous works have used simplified descriptions of the immune response in a similar fashion as we have chosen to model here [15, 27]. Models with simplified immunity usually ask conceptual questions or are used to infer parameter values from data with few measured cell types (e.g. only counting CD8+ and CD4+ T cells). Future work interested in specific questions that are immune related, for instance comparing the respective roles of innate and adaptive immunity in clearance, could benefit from more detailed descriptions of immune effectors. In particular, expression of cytokines are interesting as they are important in the epithelium’s part in innate immunity [52].
Our model does not attempt to capture the progression stages that HPVs can cause in persisting infections. To appropriately model these changes would require several adjustments, including that cell proliferation of infected cells and probabilities of symmetric divisions become time variant. Indeed, our model can be adapted to study other oncoviruses that infect the epithelium, where future studies can consider the contexts of immune evasion and cellular transformation driven by oncogenes [37]. In addition, there is increased interest in how epithelial cell dynamics (e.g. cell competition, mechanisms to maintain homeostasis and repair) interact with our knowledge of how tumor viruses alter cellular programing, in particular changing balanced cell fate ratios, skewing squamous differentiation toward a proliferative phenotype [67]. New modeling methods will require possible evolutionary approaches of cell phenotypes emerging over time.
In many ways, the simultaneous infection of a host by different pathogen strains or even species is the rule rather than the exception [68]. Of particular interest is how different pathogens or strains interact inside a host and how this affects the course of the infection. For instance, HPV infections are often of multiple HPV types and as lesions progress to cancer there is clonal-selection, usually leading to a single type as the main driver of the tumor [69]. One straightforward extension of this model would be to investigate coinfections between pathogens with similar cell tropisms (e.g. chlamydia and EBV) or pathogens that differ in their life-cycles. Our model could consider both infections at once or be adapted to study organotypic models that include multiple pathogen infections (e.g. HSV, EBV and HPV coinfecting the same tissues and cells [70]) or the effects of the resident microbiota.
Finally, opening a dialogue between mathematical modeling and experimental data generates new hypotheses to test. One of the clearest illustrations of this is our result that burst size differences appear as the most parsimonious explanation to explain symptom differences between wart-causing and lesion-causing HPV infections. Technological improvements in clinical and experimental techniques also allow us to test more subtle predictions. Testing hypotheses generated by the model will allow us to move forward by validating the model assumptions that are consistent with the data and rejecting the others. This will allow us to increase the model complexity and test more elaborate predictions. We hope to inspire experimental studies on infections of stratified epithelia to focus more on dynamics and time series approaches (including mathematics) to better understand these varied and broadly impacting pathogens.
The Thunder Bay Regional Health Research Institute’s Biosafety Committee approved all research involving NIKS cell line cultures. The NIKS cell line [40] was obtained from Dr. Paul Lambert, McArdle Laboratory for Cancer Research, University of Wisconsin.
Organotypic culture growing techniques used here have already been described in detail elsewhere [71, 72]. Original experiments were performed to obtain time series data with sufficient replicates for model fitting. Three independent experiments were performed, with rafts harvested at one-week intervals (0, 1, 2, and 3 weeks) starting the day after lifting them to an air-liquid interface. From a total of 12 formalin-fixed, paraffin-embedded (FFPE) rafts, 48 tissue slices were imaged using fluorescence microscopy (DAPI staining for cell nuclei) and resulted in 3 Fields of View (FOV) per slice (n = 144). Counts in each FOV were done semi-automatized using ImageJ cell counting software.
The uninfected epithelial model consists of 4 cell populations of distinct phenotypes to capture epithelial structure (Fig 1): basal cells (assumed to have a constant population size, Ub = Nb, as cells that move up are replaced immediately), parabasal cells (with population size Up), differentiated cells of the mid and upper layers (with population size Ud) and of the surface layer (with population size Us). Since we are interested in cervicovaginal infections of non-keratinized squamous epithelia, we assume the top layer of keratinocytes are close to death and are shed from the surface as they die. The cell population dynamics are captured by three ordinary differential equations (ODE):
U s . = ν U d - μ U s U d . = ρ p ( 1 - Δ q ) U p - ν U d U p . = ρ b ( 1 - Δ p ) U b + ρ p Δ q U p (1)
Dots above the variables indicate time derivatives. Basal cells proliferate at a rate ρb, giving rise to parabasal cells which in turn proliferate at a rate ρp, while entering the mid and upper layers of the squamous column (Eq 1). These cells are differentiated and migrate up to the surface layer at a rate ν. Mature keratinocytes die at the surface of the epithelium at a rate μ. There is an overlap between cell phenotype and spatial structure since an epithelial cell moves up the stages as it ages (Fig 1).
When modeling stem cell divisions, we follow earlier studies [23, 73] and introduce the fraction of basal cell divisions that are symmetric and give rise to two basal cells, p1, and the fraction that creates two parabasal daughter cells is p2. Note that q1 and q2 are the parabasal equivalent terms (see Fig 1). The generation of parabasal cells from basal cells is found by 2 p2 + (1 − p1 − p2) which we simplify to 1 − Δp by assuming Δp = p1 − p2 and the equivalent of this for the generation of differentiated cells is Δq = q1 − q2 in equation system 1. We considered distinct probabilities of divisions for the two layers (ps and qs), even though both the basal and parabasal layers are mostly made up of the same transit-amplifying cells, because the basal layer also contains stem cells which can divide in an unlimited fashion [74]. Thus, the two layers should have distinct properties. Finally, the assumption that the basal layer is constant implies that we must assume Δp = 0 in order for the basal layer to neither grow nor shrink. However, we maintain this structure of the model because Δp would be needed if one were to either relax the assumption of a constant basal layers (e.g. when studying a growing epithelium, as in organotypic cultures) or when it is infected (e.g. HPV infections might alter this parameter and make p1 divisions more frequent [67]; though we do not address this feature of HPV infection directly).
We chose to not include the stochastic nature of these cell divisions, as it has been considered previously [23, 73], and we were interested in understanding deterministic behaviors of the system, such as active repair or active changes to cell ratios. All the variables and parameters used are summarized in Fig 1 and Table 1. Finally, the model is sufficiently general that it can represent different kinds of stratified epithelia, including keratinized and non-keratinized squamous epithelia.
To calibrate parameters (Table 1), we initially relied on a study from 1970 that used in vivo autoradiography techniques to calculate the mean cell cycle time for epithelial cells in cervical and vaginal tissues [41]. They found that basal cells have a relatively slow cycle of approximately 33 days and that 1.14% of these cells are synthesizing DNA at a given time point. Parabasal cells have a much shorter cell cycle (2.6 days) and 14.25% of these cells are synthesizing DNA. Differentiated cells do not divide and have a life expectancy of 4 days (Table 1). A detailed analytical analysis of this uninfected model can be found in the S1 Text.
For fitting raft cell culture data, we did not want to assume a priori that the basal layer starts off as a constant, especially since in the experiments the tissue is grown-up from a single layer cells. So we used a variation of our model by assuming the basal layer was not constant but rather followed this equation:
U b . = ρ b U b Δ p ( 1 - U b N b ) . (2)
Here we assume the basal layer (cells that are touching the basal lamina) are growing until they reach a maximum capacity, Nb, and Δp is not assumed to be zero. There are other changes from the previous model: Us now represents the surface cells that are keratinized, and since the Up and Ud cells cannot be distinguished experimentally we summed these two variables for fitting the ‘suprabasal’ cells counted in the experiment.
Modeling infections of the stratified epithelium requires adding populations of free-forms of the pathogens, infected cells and immune cells. See Fig 5 for the schematics of the models and Table 3 for the parameter estimates.
Nearly all the parameter values could be set using data from the literature, which mostly lay in narrow ranges (Tables 1 and 3). Parameters for which we had little information were either kept free or calibrated. For instance we used Δq to scale all equilibrium population sizes (see Results).
To test the robustness of our results, we performed uncertainty and sensitivity analyses using Latin Hypercube Sampling and Partial Rank Correlation Coefficients (PRCC) via the pse package in R [77], which is popular for disease models [78], and numerical integration was done using deSolve package. We generated 1,000 parameter sets by Latin Hypercube sampling from uniformly distributed parameter values within a specified biologically realistic range. PRCCs were calculated between the rank-transformed samples and the resulting output matrix of the response variables (e.g. duration of infection, maximum pathogen load). 100 bootstraps were performed to generate 95% confidence intervals. The magnitude of the PRCCs determines the effect strength of a given parameter on a specific response variable (0 for no effect and 1 for very strong) and the sign indicates whether the response grows or shrinks with increasing the parameter value.
Monotonicity for each parameter was checked for each response variable, and the parameter ranges were shortened when monotonicity was not obeyed. This was not common and was usually for values very close to zero.
We inferred parameter values from the data we collected over 3 weeks from a 3D raft culture of NIKS cells. Note that cells attached to the basal membrane were considered basal and those above them were counted as suprabasal cells. This was done (rather than use differentiation markers) in order to differentiate between true basal cells and parabasals and to estimate a carrying capacity, Nb. Model parameters were inferred using maximum likelihood estimation and trajectory matching, assuming measurement error follows a Poisson distribution. Fitting and model predictions were performed in R software [79], using packages bbmle [80], deSolve [81], and pomp [82]. Note that the parameter values estimated experimentally were not used for the infection models since the experiments had the tissues growing up into full stratified form while infections usually start with the epithelium already at homeostasis, thus the epithelium parameters from the literature were more appropriate.
|
10.1371/journal.pntd.0004414 | Current Perspectives on Plague Vector Control in Madagascar: Susceptibility Status of Xenopsylla cheopis to 12 Insecticides | Plague is a rodent disease transmissible to humans by infected flea bites, and Madagascar is one of the countries with the highest plague incidence in the world. This study reports the susceptibility of the main plague vector Xenopsylla cheopis to 12 different insecticides belonging to 4 insecticide families (carbamates, organophosphates, pyrethroids and organochlorines). Eight populations from different geographical regions of Madagascar previously resistant to deltamethrin were tested with a World Health Organization standard bioassay. Insecticide susceptibility varied amongst populations, but all of them were resistant to six insecticides belonging to pyrethroid and carbamate insecticides (alphacypermethrin, lambdacyhalothrin, etofenprox, deltamethrin, bendiocarb and propoxur). Only one insecticide (dieldrin) was an efficient pulicide for all flea populations. Cross resistances were suspected. This study proposes at least three alternative insecticides (malathion, fenitrothion and cyfluthrin) to replace deltamethrin during plague epidemic responses, but the most efficient insecticide may be different for each population studied. We highlight the importance of continuous insecticide susceptibility surveillance in the areas of high plague risk in Madagascar.
| In spite of more than 50 years of efforts to control plague, Madagascar is one of the countries with the highest plague incidence. Bubonic plague, the most encountered form, is transmitted by flea bites. The plague control and prevention policy is based on flea control with chemical insecticides. Hence the occurrence of flea resistance is of major concern in the public health context. Our research team conducted laboratory work to assess the resistance level of Xenopsylla cheopis, the main plague vector, to 12 insecticides. The results of this study will contribute to more focused flea population control, and therefore more efficient control of plague outbreaks.
| Arthropod-borne diseases are a major concern worldwide. Every year more than 1 billion cases and over 1 million deaths from vector-borne diseases are estimated [1]. Most of these vectors are bloodsucking arthropods (e.g., mosquitoes, flies, fleas, ticks, lice) living in direct contact with humans or harbored by livestock or commensal animals [2–5]Arthropod-borne diseases such as malaria, chikungunya, dengue fever, Lyme disease, West Nile fever, Rift Valley fever, and plague erupt and cause substantial mortality in humans and livestock [1,5–8]. The risk posed by these diseases can be significantly reduced by the use of insecticides during a public health emergency; insecticide-based intervention can prevent an outbreak or can limit the expansion of the disease [3,5].
In the 1940s, the use of synthetic insecticides led to great improvement in the battle against disease vectors [3,9]. Consequently, the intensive use of insecticides caused selection pressure on insect populations, which developed mechanisms to survive insecticide treatments. All classes of insecticides are currently involved, and the list of pests associated with agriculture and health has been continually increasing [10]. The lack of new insecticidal compounds and the misuse or overuse of insecticides have been identified as reasons of the emergence of insecticide resistance in pests [11]. Hence, an efficient vector control policy must take in account the possibility of insecticide resistance, which can lead to a failure of the control strategy [12]. The early detection and monitoring of insecticide resistance in a vector population may positively impact intervention strategies. Until now, the main defense against resistance is close surveillance of the susceptibility of vector populations [4].
Plague, a rodent disease transmitted to human by infected flea bites, remains an important health problem in Madagascar [13]. The flea, Xenopsylla cheopis is the main plague vector, parasitizing black rats Rattus rattus that live in urban and rural housing [14]. According to the World Health Organization (WHO), the most rapid and effective method for controlling fleas is to apply an appropriate insecticide formulated as a dust or low-volume spray [15,16]. Insecticide dusting in households is the strategy adopted by the National Plague Control Program in Madagascar to control vectors and to limit the expansion of plague epidemics [17,18]. In 1947 the use of DDT (dichlorodiphenyltrichloroethane) insecticide to control fleas in Madagascar gave new hope for combatting plague. Since 1965, resistance of X. cheopis to DDT has been developing in Madagascar [19]. Later, X. cheopis populations were reported to be resistant to different families of insecticide from the early 1980s to 2000 [20–26]. More recently, amongst 32 populations of X. cheopis, only two populations were susceptible to deltamethrin, which is currently the preferred insecticide in Madagascar for flea control [27]. Hence, it is crucial to find insecticide alternatives to deltamethrin. Here we report the results of 12 different insecticide bioassays performed on 8 populations of X. cheopis previously found to be resistant to deltamethrin.
A previous study reported the susceptibility of 32 populations of fleas to deltamethrin [27]. We chose to study eight populations from different geographical regions of Madagascar (Fig 1). Chosen populations were resistant fleas with mortality rates when exposed to deltamethrin of 2.5% to 65% [27]Flea populations (X. cheopis) were collected from the field and reared in an insectarium [27]. Briefly, rodents were trapped alive, fleas were combed into a large container, and fleas were reared in insectarium at 22–27°C and 75–80% relative humidity until having the sufficient number to perform bioassays. [27]. Fleas used in bioassays were subsequent generation of those collected in field.
Bioassays were conducted on fleas populations according to the WHO protocol [28]. Ten adult fleas per tube were exposed to insecticide-impregnated paper (1.5 x 6 cm; Vector Control Research Unit, Penang, Malaysia) for specified times and at predetermined insecticide concentrations (Table 1). Each test was replicated at least four times for a total of 40 fleas per insecticide and per population. Negative controls were performed with paper only impregnated with the carrier of each insecticide family. LT50 (the time by which 50% of fleas were knocked down) were estimated for each insecticide during the diagnostic time. At the end of the exposure time and for all bioassays, the impregnated papers were removed and replaced by non-impregnated papers. Final mortality was recorded 24 hours after the beginning of exposure time. Susceptibility status was established according to the WHO guidelines for insecticide susceptibility test. Mortality rates of 98 to 100% indicated susceptibility, 80 to 98% tolerance or suspected resistance, and less than 80% resistance [29]. The test was not validated, and the data not included, if the negative control mortality rate was over 20%. The mortality rate was corrected with the Abbott formula [30] when control values were between 5% and 20%.
Analysis of Variance (ANOVA) and Tukey’s b test were used to compare mortality rates. Mean LT50 and the standard errors for each flea population and for each insecticide were estimated with a binomial generalized linear model (glm) analysis. This glm including a probit function is a fitted model giving a prediction and a standard error at each response probability (p.model function with the package MASS). High mortality may not occur with some insecticides for some populations and therefore the LT50 would not be estimated (NE) Correlations between the mortality rates were calculated with Pearson tests (packages: corplot, Hmisc and ggplot2 to generate figures). Statistical analyses were done with R software (RStudio) [31].
The mortality rate was different among insecticides and populations: mean mortality was significantly different according to populations (F value = 195.34, p < 0.0001) and insecticide (F value = 36.22, p < 0.0001). A strong correlation between insecticides and populations (F value = 9.10, p < 0.0001) was observed. Nonetheless, all populations were at least somewhat resistant to the six insecticides alphacypermethrin, lambdacyhalothrin, etofenprox, deltamethrin, bendiocarb and propoxur, with mortality rate ranging from 0 to 79% (Fig 2). Dieldrin was the only insecticide with 100% mortality rate for all flea populations (Fig 2). The resistance to DDT was substantial for most populations, with mortality rates varying between 5 and 26.4%, with the exception of one tolerant population, Andranomanalina, which had 90% mortality. Apart from the dieldrin, the highest mortality rates were observed for malathion, fenitrothion, cyfluthrin and permethrin (Fig 2). For these four insecticides, the susceptibility profiles were very different for each population (Fig 3). Almost the same resistance profile was observed for the organophoshates: two populations (Ambohimiandra and Ambohipananina) were susceptible with 100% mortality rate and fleas from Tsararano Ambony and Amparaky were both resistant to malathion and fenitrothion. Mortality induced by cyfluthrin ranged from 67.5 to 100% with one susceptible population (Ambohipananina) and four populations were tolerant. The mortality rate of the eight populations with permethrin varied between 50 and 95%, with two resistant populations.
The curve profile, obtained during exposure time for each insecticide and for each station (Fig 4), and values of LT50 (S1 Table) were in concordance with the results obtained with the average mortality observed after 24 hours. Highly resistant population to insecticide had LT50 values longer than durations of exposure time. For Etofenprox, six tested populations had estimated LT50 > 500 minutes whereas exposure time was 480 minutes. These six populations had mortality below 30% after 24 hours. For DDT, no tested population reached LT50 until the exposure time (LT50 > 360 minutes), except for the tolerant population of Andranomanalina with a LT50 = 142 ± 8.80 minutes (Fig 4). Flea populations susceptible to cyfluthrin (Ambohipananina) had a LT50 equal to 21 ± 1.88 minutes and the tolerant ones had LT50 between 36 ± 4.58 and 128 ± 10.26 minutes. Even though the LT50 value of the population most resistant to cyfluthrin (Tsararano Ambony) was seven times higher than the value for the most susceptible population, the possible emergence of resistant individuals in tolerant populations could be suspected.
Similarly for malathion, susceptible population had an LT50 < 104 ± 5.14 minutes. LT50 was reached after 152 ± 6.81 to 285 ± 6.97 minutes for tolerant populations. Resistant populations’ LT50 values (Amparaky and Tsararano Ambony) exceeded the exposure time (300 minutes): LT50 were 370 ± 31.03 minutes and 464 ± 128.71 minutes, corresponding to 57.7% and 10% mean mortality after 24 hours, respectively.
Fig 4 illustrates the heterogeneity of response to insecticides amongst different flea populations. For example, in dieldrin trials, although 100% mortality rate was observed after 24 hours for every population, one population (Ambohimiandra) did not reach its LT50 value until the end of the exposure time (360 minutes). The LT50 value (425 ± 38.51 minutes), was 4 times higher than the minimal value obtained for dieldrin (99±6.84 minutes).
With propoxur, 100% of exposed fleas were knocked down before the exposure time was elapsed in two populations (Ambohimiandra and Ambohipnanina). These two populations had the shortest values of LT50 and lowest mortality rate values to propoxur after 24 hours (highly resistant populations).
Positive correlations were observed between deltamethrin, etofenprox, cyfluthrin, malathion, fenitrothion and propoxur (Fig 5), suggesting possible insecticide cross-resistance mechanisms in fleas. A strong negative correlation was observed between permethrin and etofenprox (r = -0.74, p<0.05). Significant correlations (p<0.05) were observed between fenitrothion and propoxur (r = 0.76, p = 0.03), propoxur and cyfluthrin (r = 0.77, p = 0.02), malathion and propoxur (r = 0.82, p = 0.01), and fenitrothion and cyfluthrin (r = 0.82, p = 0.01).
X. cheopis populations tested in this study were found highly resistant to DDT. Seven of eight populations showed final mortality rate less than 30% to DDT. These results reflected past observations of DDT resistance amongst X. cheopis populations from Madagascar, even though this product has not been used for many decades. The main argument raised to explain X.cheopis resistance to DDT worldwide was the extensive use of this insecticide in plague and malaria vector control [32–37]. In Madagascar, DDT was widely used against rat fleas since the 1940s [19]. DDT and pyrethroids were used in Indoor Residual Spraying (IRS) and long lasting insecticide impregnated nets against malaria vectors [38]. Furthermore, malaria vector treatment could have effect on flea vector resistance; in fact it was demonstrated elsewhere that insecticides used in IRS programs reduced flea loads on indoor rodents. [39]. However, in areas where malaria and plague are endemic, IRS treatment could have the potential to put selective pressure on fleas to develop resistance [40].
Dieldrin, an organochlorine insecticide, also saw widespread use in countries where plague occurred. Dieldrin was used in Madagascar during the period of DDT use, and X. cheopis was already described as resistant to dieldrin [21,26]. Insecticide susceptibility tests done in India showed that fleas resistant to DDT often were resistant to Dieldrin and other cyclodien insecticides [25, 29]. Yet, in our study, X. cheopis populations were all susceptible to this compound.
Pyrethrum was shown to have lethal effects on rat fleas before synthetic pyrethroids were used [41]. X. cheopis resistance to pyrethroid compounds (deltamethrin 0.025% and cyfluthrine 0.15%) was previously described in Madagascar [23,24]. Fleas from the Central Highlands of Madagascar were resistant to low concentrations of deltamenthrin (0.025%). Recently, 94% of studied populations were not sensitive to higher concentrations of deltamethrin (0.05%) [23,24,27]. The use of deltamethrin in plague control since the 1990s likely led to the development of resistant flea populations.
We present the first data illustrating resistance of X. cheopis populations to alphacypermethrin, lambdacyhalothrin and etofenprox, which were never used in mass vector control. This may suggest the involvement of cross-resistance mechanisms between these insecticides and those that were extensively used.
Organophoshates were also described as inducing resistance in rat flea populations. In India, X. cheopis was indicated as resistant to malathion and fenitrothion, although these compounds were never used in the study areas [36]. It was suggested that flea resistance to these compounds was associated with resistance to DDT [42]. Even if resistance to organophosphates was already described in some areas of Madagascar, the majority of populations studied presently showed less resistance to these compounds [21,43]. On the other hand, X. cheopis populations were previously described as susceptible to carbamate insecticides [24]; however, our study demonstrated a high resistance to propoxur and bendiocarb.
Our results suggest resistance to all insecticides except dieldrin, which produced 100% mortality for all population. However, the LT50 values observed in one population (Ambohimiandra) suggest a progressive development of resistance to this compound. But dieldrin was banned in most of country worldwide because of its high toxicity in mammals and its bioaccumulation in the environment [44,45]. The use of dieldrin was suspended in Madagascar since 1993 [46]. However, other insecticide families having the same mode of action as dieldrin (antagonist of GABBA receptor) such as fiproles could be promising [10,47].
Six insecticides (alphacypermethrin, lambdacyhalothrin, etofenprox, deltamethrin, bendiocarb and propoxur) were relatively ineffective for flea control in all populations. Nonetheless, resistance level to the insecticides (permethrin, cyfluthrin malathion and fenitrothion) was very different among populations, suggesting different selection pressures. Hence, in this study, according to WHO thresholds, some insecticides were still efficient in some localities; thus, insecticides that induce resistance according WHO thresholds still may exhibit high performance in the field [12].
The different responses of populations to each insecticide reflect also the mode of action of insecticide molecules and the mechanism developed by insects to overcome toxic effects. Pyrethroids and DDT belong to a group of neurotoxic chemicals and share a similar mode of action distinct from other classes of insecticide. The studies on kdr mutation demonstrated the same mode of action of pyrethroids and DDT, which is the reduced target-site sensitivity of sodium channel [48]. Thus, the mechanism of resistance may not be specific to a particular insecticide family or group but the molecule structure of each insecticide can play important role.
For instance, the negative correlation between permethrin and etofenprox may involve the different effect induced by a Type I pyrethroid (permethrin) and a pseudo-pyrethroid (nonester pyrethroid) [49]. In addition, different levels of pyrethroid resistances were observed amongst populations. All studied populations were resistant to etofenprox alphacypermethrin, lambdacyhalothrin and deltamethrin; yet cyfluthrin and permethrin were effective in some localities. In a study of cross resistance amongst pyrethroids, cross resistance between 19 pyrethroid insecticides was assessed in bollworm moth, Helicoverpa armigera [50]. Cross resistance between pyrethroids seemed due to enhanced oxidative metabolism induced by pyrethroid with the same structure. The modification or replacement of any compound (aromatic compound) in the molecule structure could modify the susceptibility of the population [50].
Moreover, DDT and dieldrin belong to the oragnochlorine family, but their structures are very different, conferring different mode of action. The first attempt to elucidate cross resistance between chlorinated insecticides in X. cheopis was performed in 1974 [27]; a DDT-selected population was found to be resistant also to insecticides structurally related to DDT, and exhibited variable resistance to cyclodiene insecticides (such as dieldrin, endrin). But biochemical assays did not show significant difference between susceptible and resistant population. [32]. Other mechanisms like Rdl mutation can confer resistance to cyclodiene like dieldrin [47].
Furthermore, the correlations between deltamethrin, etofenprox, propoxur, and between cyfluthrin, malathion and fenitrothion, may be explained by the same mechanism of resistance developed by X. cheopis for these insecticides. The absence of references on this topic in Siphonaptera illustrates a need for further research on insecticide resistance mechanisms in fleas.
Efforts can be undertaken before each epidemic season in order to control the proliferation of vectors and host, such as live rat mass trapping, promotion of rat proofing in houses and environment sanitation [15]. One available method could be the use of insecticide bait box, combining insecticide and delayed toxicity rodenticide [51–55]. The objective is to kill fleas on rodents before the action of the rodenticide. Instead of insecticide dusting in household, the quantity of insecticide is reduced because the insecticide bait box is more focused on fleas harbored by rodent with the host acting as a vehicle for the insecticide, carrying it to its nest. In the same line of thought, the feasibility of “insecticide delivery tubes” in reducing flea loads was studied on commensal rodents, capitalizing on the knowledge of their behavior [56]. Similarly, using rodent bait containing systemic insecticide could be a new avenue for combating or at least, reducing fleas load on rodents in plague endemic area during inter-epidemic season [57]. Besides, novel approaches to fighting vector limiting the use of chemical insecticides should be explored in order to avoid insecticide resistance [58,59]. For instance, research must be undertaken in the way to better understand the interaction between the vector, the pathogen and the insect microbiome. The strategy is based on introduction of microorganism which may affect the insect lifespan or the ability to transmit the pathogen [59–61]. Furthermore, research on bio pesticide is already ongoing with the use of entomopathogen fungi to reduce the survival of flea larvae [62].
The main finding in this study is that X. cheopis populations developed resistance mechanisms to the insecticide families most used in vector control. The description of phenotypic resistance to insecticides is important to help practitioners choose the most efficient strategy in pest management. Hence, in a public health context, insecticide susceptibility status of fleas in each plague risk area may be monitored periodically to conduct more focused and adapted flea control. However information available on the mechanism of resistance and cross-resistance about X. cheopis is scarce or nonexistent. Research must be done to understand the mechanisms conferring resistance to insecticides in plague vectors such X. cheopis.
|
10.1371/journal.pcbi.1004148 | Preferred Supramolecular Organization and Dimer Interfaces of Opioid Receptors from Simulated Self-Association | Substantial evidence in support of the formation of opioid receptor (OR) di-/oligomers suggests previously unknown mechanisms used by these proteins to exert their biological functions. In an attempt to guide experimental assessment of the identity of the minimal signaling unit for ORs, we conducted extensive coarse-grained (CG) molecular dynamics (MD) simulations of different combinations of the three major OR subtypes, i.e., μ-OR, δ-OR, and κ-OR, in an explicit lipid bilayer. Specifically, we ran multiple, independent MD simulations of each homomeric μ-OR/μ-OR, δ-OR/δ-OR, and κ-OR/κ-OR complex, as well as two of the most studied heteromeric complexes, i.e., δ-OR/μ-OR and δ-OR/κ-OR, to derive the preferred supramolecular organization and dimer interfaces of ORs in a cell membrane model. These simulations yielded over 250 microseconds of accumulated data, which correspond to approximately 1 millisecond of effective simulated dynamics according to established scaling factors of the CG model we employed. Analysis of these data indicates similar preferred supramolecular organization and dimer interfaces of ORs across the different receptor subtypes, but also important differences in the kinetics of receptor association at specific dimer interfaces. We also investigated the kinetic properties of interfacial lipids, and explored their possible role in modulating the rate of receptor association and in promoting the formation of filiform aggregates, thus supporting a distinctive role of the membrane in OR oligomerization and, possibly, signaling.
| Opioid receptors associate with each other in the plasma membrane, following a mechanism that has been implicated in either beneficial or adverse effects, depending on the environment and the interacting partners. Whether or not the different opioid receptor subtypes share a similar propensity to form di-/oligomers and specific receptor-receptor interactions is still unclear on the basis of the published data. This information, however, is necessary to predict stabilizing or de-stabilizing mutations and to design experiments to clarify the role of oligomerization in opioid receptor function. The inferences provided by the extensive molecular dynamics simulations reported herein constitute a first step in this direction.
| Experimental evidence accumulated over almost two decades supports the ability of all three major opioid receptor (OR) subtypes (μ-OR, δ-OR, and κ-OR) to form homomeric and heteromeric complexes, a feature that is common to several other G Protein-Coupled Receptors (GPCRs) (see [1] for a recent review). In particular, heteromerization between μ-OR and δ-OR and between δ-OR and κ-OR has been suggested to expand the repertoire of OR signaling by modulating ligand binding, receptor signaling, and/or trafficking properties [2]. This modulation has a direct translational relevance in view of the putative role that OR heteromers play in opioid-induced adverse effects (e.g., the development of analgesic tolerance [3,4]), and in the observed potentiation of morphine-induced analgesia by δ-OR-selective antagonists [5]. Thus, targeting OR oligomers directly may lead to novel drugs with potentially greater selectivity and reduced side effects compared to molecules targeting individual receptors.
While a few ligands have already been proposed to target OR heteromers (e.g., see [6–11]), it remains unclear how they bind and activate these receptor complexes. To begin addressing these questions and eventually use these and other molecules as tools to elucidate the physiological relevance of OR oligomers, ligand-bound models of these receptor complexes are highly desirable. However, building such models is not feasible in the absence of reliable information about their interface of dimerization/oligomerization in a physiologically relevant environment. The recent X-ray crystal structures of the μ-OR [12] and κ-OR [13] have suggested specific receptor-receptor interactions involving transmembrane (TM) helices TM5 and TM6 or TM1, TM2, and helix 8 (H8). Although these interfaces are thermodynamically stable according to our recent free-energy calculations of μ-OR and κ-OR homo-dimers in an explicit lipid-water environment [14], the possibility cannot be ruled out that other receptor-receptor interactions are also feasible in the cell membrane, and perhaps more kinetically favorable than those inferred from crystallography.
Here, we carried out extensive, unbiased coarse-grained (CG) molecular dynamics (MD) simulations of freely diffusing ORs in an explicit lipid-water environment to evaluate differences and similarities in the supramolecular organization and preferred dimeric interfaces of all three major receptor subtypes.
A summary of all the MD simulations we carried out on CG molecular models based on the crystallographic structures of inactive δ-OR [15], μ-OR [12], and κ-OR [13] is provided in S1 Table. Specifically, to enhance the statistical significance of our description of the dimerization process, we ran five independent simulations for each of the homomeric (i.e., μ-OR/μ-OR, δ-OR/ δ-OR, and κ-OR/κ-OR) and the most studied heteromeric (i.e., δ-OR/μ-OR and δ-OR/κ-OR) complexes of ORs, starting from randomized initial orientations of CG models of sixteen individual receptors (8+8 in the case of heteromers) in an explicit CG palmitoyl-oleoyl-phosphatidyl-choline (POPC)/10% cholesterol bilayer. Together, these simulations yielded over 250 μs of accumulated data, which correspond to approximately 1 ms of effective simulated dynamics, according to the established (×4) scaling factor of the CG MARTINI model [16–18] we employed.
All the twenty-five different simulations we carried out showed multiple association events between ORs initially located far apart and moving freely in the lipid bilayer. As an example, Fig. 1. shows the location of each of the 16 receptor molecules in one of the five simulations executed on the κ-OR system (run #1) at different simulation times (specifically 0, 2, 6, and 10 μs).
While the formation of several receptor-receptor complexes could be observed during simulation of the μ-OR/μ-OR, δ-OR/ δ-OR, κ-OR/κ-OR, δ-OR/μ-OR, and δ-OR/κ-OR systems, each individual protomer in the complex seldom shared more than two interfaces, thus favoring filiform rather than branched or compact high-order arrangements (see simulation snapshots at 6 and 10 μs in Fig. 1, as well as S1–S5 Movies). Notably, linear arrays of receptors were observed in early atomic force microscopy images of the prototypic GPCR rhodopsin in native membranes [19], and were further supported by early CG simulations [20]. Below, we speculate that the formation of chains of receptors may be influenced by lipid dynamics (see the Dynamic Behavior of Lipid Molecules section for a rationale).
Importantly, once formed, no receptor complex dissociated over the maximal simulated time of 10 μs, but rather only minor interface rearrangements were observed. This observation is in line with experimental estimates of dimer lifetime (a few seconds) obtained from recent single-molecule imaging studies of different, individually labeled, GPCRs in living cells [21]. Since dissociation is not observed in the simulations reported here, a direct calculation of free-energies is precluded. However, analysis of the executed twenty-five simulations instills confidence in that they might have captured the majority of fastest-forming dimer interfaces of ORs.
The five different simulation trajectories obtained for each OR system were pooled together to derive statistically meaningful information about the various dimer interfaces that formed during simulation. Since each protomer structure was kept fairly rigid by elastic network forces during the simulations (see Materials and Methods for details), no allosteric communication between protomers or inter-dependence between dimer interfaces were expected. Interfaces were defined based on the minimal number of residues on each receptor TM helix within a certain distance cutoff from each other, and were clustered based on the similarity between inter-protomer contact maps (see details of the analysis under Materials and Methods).
Tables 1 and 2 report the preferred OR homo-dimer and hetero-dimer interfaces, as derived from Bayesian analysis of the pooled trajectories. The following interesting observations can be made on the basis of this analysis. First of all, not all possible combinations of TMs were found to be involved in dimer interfaces during the simulated timescale of 10 μs. The only interfaces that formed in all studied homo- and heteromeric systems are: TM1,2,H8/TM1,2,H8, TM1,2/TM4,5 (also TM4,5/TM1,2 for hetero-dimers), and TM1,2/TM5,6 (also TM5,6/TM1,2 for hetero-dimers). Other interfaces, such as TM4,5/TM4,5, TM4,5/TM5,6, TM5/TM5, and TM5/TM1,2 did not form in at least one of the five studied μ-OR/μ-OR, δ-OR/ δ-OR, κ-OR/κ-OR, δ-OR/μ-OR, and δ-OR/κ-OR systems. In particular, the TM4,5/TM4,5 interface, which has been suggested to constitute a possible GPCR dimer interface by various experimental assays (reviewed in [22]), only formed during simulation of the δ-OR/δ-OR system, and with a lower frequency with respect to other interfaces formed by δ-OR homomers. Notably, neither TM3 nor TM7 were ever found to be involved in a dimer interface, whether formed by the same or different receptor subtypes.
Another important observation from our study is that TM helices appear with different frequencies at a dimer interface depending on the OR system. Specifically, the TM1,2 helices appear most frequently at the observed dimer interfaces of δ-OR and κ-OR, followed by TM4,5 and TM5,6, with the latter helix pair being more involved in a dimer interface in μ-OR compared to δ-OR and κ-OR. However, it is noted that the calculated confidence intervals for frequencies of the specific interfaces are quite broad and overlapping, and therefore the estimated differences between the three OR subtypes may not be as relevant as they appear to be.
Several interfaces observed in the simulations reported here are structurally similar to some of the putative dimer interfaces inferred from recent GPCR crystal structures (see S2 Table for a list of currently available GPCR crystal structures showing parallel receptor arrangements). To allow a quantitative comparison, we calculated the minimum Cα root mean square deviation (RMSD) distance between members of the cluster of dimeric complexes that formed during the simulations and each crystal structure listed in S2 Table. The cases where this distance resulted to be less than 10 Å are reported in S3 Table for homo-dimers and in S4 Table for hetero-dimers.
The calculated RMSD values of S3 and S4 Tables suggest that the dimer interface from simulations that is closest to one inferred from crystal structures is the TM1,2,H8/TM1,2,H8 interface. In particular, δ-OR and κ-OR form TM1,2,H8/TM1,2,H8 interfaces in both homo- and hetero-dimers that are very close (RMSDs below 4.3 Å) to that seen in the crystal structure of κ-OR (4DJH [13]). The closest crystal structure to the TM1,2,H8/TM1,2,H8 interface that forms during μ-OR simulations is not the one inferred by the μ-OR crystal structure (4DKL [12]), but rather the one suggested by a β1-adrenergic receptor (B1AR) crystal structure (4GPO [23]). Figs. 2A, 3A, and 4A show, as an example, the TM1,2,H8/TM1,2,H8 homo-dimer interfaces formed during simulation of δ-OR, κ-OR, and μ-OR, respectively, overlapped onto the closest crystal structures, i.e., 4DJH or 4GPO.
The relatively small RMSD values listed in S3 Table, indicate that the simulations of the δ-OR system also reproduced both symmetric and asymmetric dimer interfaces inferred from CXCR4 crystal structures [24] (see S2 Table for details) with reasonable accuracy. Specifically, the interface herein termed TM1,2/TM5,6 deviated only 6.48 Å from the asymmetric interface revealed by 3OE8 (after overlapping the dimer from simulation with chains A and B of 3OE8) whereas the interface herein called TM4,5/TM5,6 deviated only 6.66 Å from an interface inferred from 3ODU (after overlapping the dimer from simulation with chains A and B of 3ODU) and 6.62 Å from an interface inferred from 3OE8 (after overlapping the dimer from simulation with chains B and C of 3OE8). Fig. 2B and 2C show structural overlaps of the TM1,2/TM5,6 and TM4,5/TM5,6 interfaces of δ-OR dimers with 3OE8.
Larger RMSD values were obtained for the identified TM5/TM5 interface in both κ-OR and μ-OR simulations (see S3 and S4 Tables) after comparison with the available GPCR crystal structures of interacting parallel receptors that are listed in S2 Table. Specifically, in κ-OR, this interface is 8.56 Å apart from the putative TM5,6-TM5,6 dimer interface inferred from the μ-OR crystal structure (after overlapping the dimer from simulation with chain A of 4DKL and its periodic image). Notably, μ-OR simulations did not produce dimeric arrangements that were close enough to the crystallographic TM5,6-TM5,6 interface of μ-OR, in spite of it being thermodynamically stable as we recently demonstrated through free-energy calculations [14]. The closest structure to the identified TM5/TM5 in μ-OR simulations was the interface termed TM4,5/TM4,5 in the B1AR crystal structure corresponding to PDB code 4GPO (RMSD of 8.82 Å, after overlapping the formed dimer with chain A of 4GPO and its periodic image). While the calculated RMSD of the identified TM5/TM5 interfaces of κ-OR and μ-OR homo- and hetero-dimers with respect to available crystal structures appear to be quite large, visual inspection of these overlaps (Figs. 3B and 4B for κ-OR and μ-OR, respectively) shows that the slight rotation of one protomer needed to match the two configurations would not dramatically change the nature of those interfaces. Thus, we speculate that the reason why the simulations reported herein are unable to reproduce the crystallographic TM5,6-TM5,6 interface of μ-OR is that this interface may need longer times to form than its slight modifications (more on this below).
To understand whether there are interfaces that are kinetically favored over others, i.e are fast forming in an explicit membrane environment, we calculated interface-specific dimerization rates (kon) for all simulated OR systems by fitting a Poisson model to the association instances observed during simulation. The results are reported in Tables 3 and 4 for the simulated homo- and hetero-dimers of ORs, respectively. The fastest forming homo-dimer interfaces are: TM1,2/TM4,5 and TM1,2/TM5,6 for δ-OR/δ-OR, TM1,2,H8/TM1,2,H8 and TM1,2/TM4,5 for κ-OR/κ-OR, and TM5/TM5 for μ-OR/μ-OR. Notably, TM1,2,H8/TM1,2,H8, TM4,5/TM1,2, and TM4,5/TM5,6 are the fastest interfaces for the δ-OR/κ-OR hetero-dimer, whereas the TM5/TM5 interface is the fastest forming for the δ-OR/μ-OR hetero-dimer. This observation further stresses the higher propensity for κ-OR to have the TM1,2 helices involved in a fast-forming dimer interface. In contrast, TM5 is more likely to be involved in a fast-forming dimer interface when one of the receptor partners is μ-OR.
Analysis of the dynamic behavior of the POPC (herein referred to as ‘lipid’) and cholesterol molecules during the simulations reported here reveals considerable dynamical heterogeneity, in that regions of high mobility appear to be surrounded by slower molecules. Such a behavior, which is typical of glass-forming fluids and supercooled liquids [25], and was also suggested for membrane proteins (e.g., see [26–28]), has also been recently reported for water at the interface of globular proteins [29]. We speculate that this dynamical heterogeneity also controls membrane diffusion and viscosity near OR dimer interfaces, and plays an important role in modulating the rate of receptor association and the structure of the complex.
Visualization of the twenty-five simulation trajectories reported herein showed a clear variation in the dynamic behavior of lipid molecules during receptor dimerization. As seen in S1–S5 Movies, which are provided as representative simulation trajectories of the δ-OR/δ-OR, κ-OR/κ-OR, μ-OR/μ-OR, δ-OR/κ-OR, and δ-OR/μ-OR systems, respectively, the density of slow lipids becomes higher (dark blue in S1–S5 Movies) as protomers approach each other, suggesting that slow lipids may interfere with dimer formation at specific interfaces. To quantitatively investigate the role of lipid molecules in regulating OR dimerization, we calculated exchange and persistence time (tX and tP, respectively) distributions of the lipids at different positions relative to isolated receptors. These two quantities characterize the diffusion properties (D) of the lipids and the effective viscosity (η) of the membrane around the protein, respectively (see details in the Methods section).
Typical observed average exchange times 〈tX〉 of lipids in the bulk membrane, i.e. away from the receptors, are ~10 ns (see S1 Fig.), corresponding to a lipid diffusion coefficient D≈d2/〈tX〉 = 10−6 cm2/s, (or D≈2.5×10−7 cm2/s, when accounting for the effective time scaling for the CG force field we used). In the simulations reported here, the average exchange time of lipids increases up to 〈tX〉≈40–50 ns at specific regions near the OR surface, giving D≈2.5×10−7 cm2/s (or, effectively, D≈6.2×10−8 cm2/s). Notably, similar values of lipid diffusion constants have recently been reported in the literature [26,27] for comparable CG force fields, and a similar behavior was implied.
In the bulk membrane, the equivalence between the exchange and persistence times implies an inverse relationship between viscosity and diffusion coefficient in homogeneous systems, known as the Stokes-Einstein relation. The presence of dynamical heterogeneity, with slow lipids close to the protein surface, corresponds to a breakdown of the Stokes-Einstein relation for lipid dynamics in this region. In other words, correlated lipid motion leads to an increased membrane effective viscosity, decoupled from the diffusional motion of the lipids, so that the inverse relationship between the diffusion coefficient and viscosity (ηD ∝ constant), is no longer homogeneously valid. Analysis of the lipid dynamics around isolated ORs shows longer average persistence times 〈tP〉 (see Fig. 5, panels A, B, and C for δ-OR, κ-OR, and μ-OR, respectively) at specific locations of the protein surface, up to 100 ns. In general, persistence times increase more than exchange times 〈tX〉, so that the ratio 〈tP〉/〈tX〉 is usually larger than 1 (see Fig. 5, panels D, E, and F for δ-OR, κ-OR, and μ-OR, respectively). Usually, lipid molecules in the region of helices TM1,2 display the shortest persistence times during simulation compared to the region of helices TM4,5 and TM5,6. This is interesting in view of the observed prevalence of interfaces that involve helices TM1,2 and the corresponding fastest on-rates, compared to the generally less frequent participation of TM4,5 and TM5,6 in dimeric interfaces of ORs. Based on this observation, it is tempting to speculate that this local effect on the membrane viscosity (i.e., 〈tP〉/〈tX〉>1) and the presence of long-lived lipid microstates with long persistence times near the surface may delay the formation of specific interfaces, making them kinetically disfavored with respect to others.
The viscosity of the environment has a well-established, direct effect on the kinetics of biological processes as indicated by the expression of the rate constant in the widely applied Kramers’ framework [30]:
k≃mω†ω2πηexp(−G†kBT)
where m is the effective mass associated with the order parameter used to describe the biological process, G† is the free-energy of the transition state, ω and ω† are the curvatures of the free-energy profile at the bottom and top of the barrier, and η is the viscosity. According to this expression, the higher viscosity observed in regions of the environment with long persistence times results in slower rates, and therefore slower kinetics of the process.
To quantitatively assess the relationship between regions of slow lipid dynamics around OR protomers and reduced kinetic rates, we calculated the position-dependent translational diffusion coefficients (DP) of μ-OR protomers at the identified homo-dimeric interfaces using a Bayesian inference approach described in the literature [31]. According to these calculations, the diffusion coefficients of μ-OR at homo-dimeric interfaces displaying slower on-rates (i.e., TM1,2,H8/TM1,2,H8, TM1,2/TM4,5, and especially TM4,5/TM5,6) are significantly reduced when the two protomers are within distances of a few nanometers. As reported in S5 Table, average diffusion rates of ~5×10−7 cm2/s estimated when the protomers are far apart (d>50 Å) are reduced to ~1–2×10−7 cm2/s as the inter-protomer distance reaches values between 40 and 50 Å. At shorter distances (d<40 Å), the diffusion rates decreased to less than 10−7 cm2/s for the slow-forming interfaces, but did not change much for the fast-forming interfaces (i.e, TM1,2/TM5,6 and TM5/TM5). This observation provides a possible explanation why the TM5,6/TM5,6 interface seen in the μ-OR crystal structure did not form during the 10 μs simulations, notwithstanding its thermodynamical stability [14]. As mentioned before, the two protomers of the TM5/TM5 dimer identified by simulations need to rotate to obtain the crystallographic TM5,6/TM5,6 configuration. We estimated the free-energy associated with such a rotation using the results of steered MD simulations, and assuming no conformational changes occurring within the individual protomers based on the CG model we employed. These results, which are reported in S4 Fig., are consistent with a high free-energy barrier (~10 kcal/mol) between the identified TM5/TM5 and the crystallographic TM5,6/TM5,6 dimers, further supporting the hypothesis that longer time scales are needed for the latter to form.
As also evident when viewing the S1–S5 Movies, lipid regions of length scales of a few nanometers between approaching receptors prior to dimer formation can also become locally trapped/restricted in motion (so-called “jammed” regions). While a complete kinetic analysis of the dimerization process that includes lipid dynamics cannot be achieved using the data reported herein, the simulations suggest a distinctive role for long-lived lipid microstructures as they appear to decrease the dimerization on-rate at specific interfaces, and kinetically select specific dimeric arrangements among all different possibilities.
We also investigated the dynamics of cholesterol molecules by calculating their persistence and exchange times around isolated δ-OR, κ-OR, and μ-OR, using the same strategy employed for the study of the lipid molecules. Average values of cholesterol exchange times are reported in S2 Fig., panels A-C, for δ-OR, κ-OR, and μ-OR, respectively, whereas average persistence times and persistence-to-exchange ratios of cholesterol molecules surrounding isolated δ-OR, κ-OR, and μ-OR are reported in Fig. 6A-C and 6D-F, respectively. Notably, regions with long cholesterol persistence times appear to be generally co-localized with regions with strong lipid molecule persistence, suggesting that both cholesterol-protein interactions and cholesterol-lipid interactions contribute to the kinetic selection of specific dimer interfaces.
Preferred cholesterol interacting sites at the surface of GPCR molecules have been reported in some of the published crystal structures. For instance, a cholesterol binding pocket was identified in a groove characterized by highly conserved residues (so-called “consensus-motif” residues) between the intracellular ends of helices TM2 and TM4 in two B2AR crystal structures, i.e., the carazolol-bound 2RH1 [32] and the timolol-bound 3D4S [33]. Cholesterol molecules were also observed in the ultra-high resolution crystal structure of the A2A adenosine receptor corresponding to PDB code 4EIY [34]. While the “consensus motif” residues identified in B2AR are conserved in A2A, no cholesterol was observed at the intracellular end of helices TM2 and TM4. In contrast, three cholesterol molecules were found at the extracellular sides of TM2,3, as well as TM5,6 and TM6,7. While no cholesterol molecules were resolved in the κ-OR or δ-OR crystal structures, electron density was attributed to a cholesterol molecule in the μ-OR crystal structure (4DKL), at the same location between TM6 and TM7 as seen in the A2A crystal structure 4EIY. Notably, the aforementioned “consensus motif” residues on TM2 (Y2.41, S2.45) are not conserved in members of the opioid receptor family. Moreover, the calculated high persistence time of cholesterol close to the extracellular ends of helices TM6 and TM7 from simulations of all three OR subtypes (see Fig. 6.) is consistent with the location of cholesterol molecules found in the ultra-high-resolution adenosine A2A crystal structure 4EIY and in the μ-OR crystal structure 4DKL. Although the palmitoylation site C3.55 at the intracellular end of TM3 was proposed to constitute part of a cholesterol preferred binding site in the groove lining the intracellular region between TM4 and TM5 [35], this region does not show increased persistence times of cholesterol molecules in the simulations reported here.
It must be noted that in the simulations reported here, the calculation of persistence and exchange times of both POPC and cholesterol near dimer interfaces is limited by the relative motion of one protomer with respect to one another. Thus, for these specific calculations, we used the results of previous simulations [14] of μ-OR crystallographic dimers, i.e., TM1,2,H8/TM1,2,H8 and TM5,6/TM5,6, where the relative orientation and distance of the protomers in the dimer were maintained constant. As illustrated in Fig. 7, the protein surface adjacent to the TM1,2,H8/TM1,2,H8 and TM5,6/TM5,6 dimer interfaces is in contact with regions of the membrane with slower dynamics. These regions of long persistence lipids right at the dimer interface may provide a mechanistic explanation for the preferential arrangement of receptors into extended linear arrays rather than compact or branched aggregates.
In summary, the simulations reported here suggest that both the formation of specific dimer interfaces and the overall topology of oligomeric aggregates depend on the kinetic kon rates. Through calculation of both persistence and exchange times of lipid and cholesterol molecules during simulation, we show the presence of ‘jammed’ lipid regions that exhibit long persistence times and non-Poissonian dynamics, and we speculate that these regions play an important role in modulating the kinetics of GPCR di-/oligomerization.
Slower diffusion at the interface between integral membrane proteins and the membrane has been extensively investigated, leading to the distinction between annular and “non-annular” lipid behavior (e.g, see [26,27]). Averaged measures of lipid diffusion have been reported (e.g., see [27]) as a function of the radial distance from the receptor, and have showed a generally slower lipid motion at small radial distances from the protein surface. Although informative, these measures fail to discriminate between the different behavior of lipids adjacent to different helices of the receptor, and are therefore of limited use. While reported measurements of the extent of local time-averaged lipid displacement (proportional to the local average velocity of lipid molecules) [26] have allowed to identify regions of the protein surface likely to be in contact with slower lipids, the average velocity is not a direct measure of the membrane viscosity, thus complicating the interpretation of the results. The present analysis provides rigorous information about the local behavior of the viscosity, thus allowing to analyze its relation to OR dimer formation at specific interfaces.
The results reported here complement those of previous studies, and suggest the use of persistence and exchange times to assess the important role of metastable lipid structures in GPCR dimer formation. Our assessment of the rheological properties of the lipid bilayer also complements the analysis of membrane elasticity and mechanical properties via macroscopic empirical models [36–38]. In particular, the results reported here provide direct evidence for the effect of the microscopic dynamics of lipids and cholesterol at dimeric interfaces on the mesoscopic length- and time-scales of receptor interactions, further supporting an essential role of the lipid membrane in determining the identity of homomeric or heteromeric complexes of GPCRs.
The atomic coordinates of non-protein molecules were removed from the PDB files of the crystallographic structures of the mouse δ-OR (PDB ID: 4EJ4 [15]), mouse μ-OR (PDB ID: 4DKL [12]), and chain A of the human κ-OR (PDB ID: 4DJH [13]) receptors. Missing or unresolved residues (specifically, residues 263–270 in μ-OR and residues 262 and 301–307 for κ-OR) were built using ROSETTA [39]. The conformation for the intracellular loop 3 (IL3) was selected as the minimum energy structure reported by ROSETTA. For μ-OR, the lowest energy conformation that also did not interfere (inter-lock) with the IL3 of the adjacent receptor at the TM5/TM6 interface was selected. For δ-OR, the crystallographic IL3 was rebuilt using ROSETTA between residues 241 and 258. Notably, the root mean square deviation (RMSD) of this loop conformation from the resolved loop of the newest high-resolution crystal structure of δ-OR [40] is 0.46 Å overall. Throughout this article we use the Ballesteros-Weinstein numbering scheme to facilitate comparison between the receptor subtype TM regions [41]. Accordingly, the first number of this scheme indicates the TM helix in which the residue in question resides, and the second number indicates the position of the residue relative to the most conserved residue in the helix, which is always numbered 50. The receptors were converted to a CG representation under the MARTINI force field (version 2.1) [16–18] and a modified elastic network was applied, as reported previously in the literature [42]. According to recent experimental findings [35], the receptors were also palmitoylated at the C3.55 position, with a palmitoylate chain consisting of 4 C1 beads, with a bond length of 0.47 nm, a force constant of 1250 kJ mol-1 nm-2 and angles of 180°, with a force constant 25 kJ mol−1.
Eighteen orientations (each rotated 20 degrees with respect to each other, thus covering 360 degrees around the z-axis of the receptor) of the CG receptors were each embedded in explicit CG POPC/10% cholesterol membranes (with a protein/lipid ratio of approximately 1:100) solvated with MARTINI CG water and neutralizing counterions. Temperature coupling at 300 K was achieved with the V-rescale algorithm, and pressure coupling at 1 bar was achieved with the Berendsen algorithm. Simulation were performed with GROMACS version 4.5.3 [43]. The systems were minimized and equilibrated with harmonic restraints of decreasing strength over 10 ns, on protein backbone beads. Water and counter ions were removed and sixteen of these membrane/protomer systems were randomly selected (repeat selections were permitted) and combined to create five 16-receptor setups for each of the studied homomeric (μ-OR/μ-OR, δ-OR/δ-OR, and κ-OR/κ-OR) and heteromeric (δ-OR/κ-OR and μ-OR/δ-OR) systems, which were re-solvated, neutralized, and subsequently minimized. The heteromeric systems contained eight receptors of each subtype. Thus, twenty-five independently constructed 16-receptor systems, each with approximately 57,000 beads, were generated. Each of these systems was simulated for 10 μs with a timestep of 20 fs, to give 50 μs of pooled simulation time for each homo- and heteromeric receptor system, and 250 microseconds in total, across all systems. Periodic boundary conditions were employed, and neighbor lists were updated every 10 steps. The Shift algorithm was used for electrostatic interactions. A single cut-off of 1.2 nm was used for Van der Waals interactions.
For each receptor, we calculated the Cα contact map δ with all possible dimerization partners, and defined as “dimeric” the pairs for which at least 10 residues on each receptor were at a distance below an assigned cutoff (8 Å). Dimeric interfaces were clustered by a k-means algorithm using the distance induced by the Frobenius norm on the contact map matrices, and the clusters automatically labeled according to the TM regions involved in receptor-receptor interactions. Specifically, the dissimilarity of two interfaces k1 and k2 was defined as
dk1k2=‖δijk1−δijk2‖
for hetero-dimers, while for homo-dimers the symmetrized form
dk1k2=min(‖δijk1−δijk2‖,‖δijk1−δjik2‖)
was used to account for the equivalence of the dimer with swapped protomers. The marginal contact map averages over the NC interfaces k in cluster C were calculated for one receptor as
δj(C)=1NC∑k∈C∑iδij(k)
while for the interacting one the analogous expression with i and j swapped inside the sum was used. Helices for which at least 3 residues were involved in the interface, were included in the label of the cluster.
We used a Bayesian inference framework to pool the information from the different trajectories and calculate estimates of the interface prevalence for each dimer and on-rates. The number of dimerization instances in a given trajectory i that yielded an interface belonging to cluster C can be described by a stochastic variable Ni,C with a multinomial distribution
p(NiC)=∑CNiC∏CNiC!∏CpiCNiC
where the probability of each cluster is pi,C = wC/ΣCwC, wC being the overall weight of the cluster. We defined wC = exp(aC), used non-informative normal priors (with zero mean and large standard deviation) for the parameters aC, and employed Gibbs sampling to obtain the posterior distributions of pC, for which we report the average as well as (2.5%,97.5%) confidence intervals. While best estimates of the probabilities pC are well approximated by the pooled fractions, the larger confidence intervals obtained by Bayesian inference reflect the variations in the different trajectories.
Estimates of the on-rate for dimerization kon(C) at each interface C were obtained assuming that associations are independent of each other, and that the number of interface-specific dimerization events n follows a Poisson distribution:
p(n)=(t kon(C) ci)nn!e−t kon(C) ci
where t is the time and ci the concentration of monomers in trajectory i. Again, we defined the parameters bCi so that the Poisson intensity is (t kon ci) = exp(bCi), where bCi have non-informative normal priors, and sampled the posterior distribution of kon using Gibbs sampling. For all estimates, 104 samples from the posterior distributions were obtained after a 5×103 burn-in phase using Markov-chain Monte Carlo techniques [44].
Comparisons with available crystal structures of parallel interacting GPCRs (see S2 Table for a current list) were evaluated by calculating the overall Cα RMSD. In order to ignore the structural differences in the monomeric structures, and capture only the degree of similarity of the OR dimer interfaces from simulation with those inferred by crystal structures, we aligned the individual CG ORs to the receptors in each crystal dimer. For the evaluation of the identified hetero-dimeric interfaces, both pair combinations (e.g., μ-OR/δ-OR and δ-OR/μ-OR) were considered for the superposition onto crystal structures, and the RMSD was defined as the minimum between the two individual RMSD values.
We characterized the dynamical properties of the lipid (POPC) and cholesterol molecules by calculating their exchange and persistence time distributions (tX and tP, respectively) at different positions relative to the protein. Specifically, we applied the equations reported in [25], and calculated local averages of lipid exchange and persistence time distributions, 〈tX〉, and 〈tP〉, respectively.
The persistence time is defined as the time it takes for a lipid molecule to move beyond a given cutoff distance d from its position at time 0, i.e. the minimum time for which ||x(tP)-x(0)||>d, (see S1, S2 Figs, and Figs. 5–6). Formally, this approach consists of viewing the lipid dynamics as a continuous time random walk (CTRW) model [45], defined by a coarse-graining length d, i.e. as a sequence of displacements of length d occurring at different times. The reported results for d = 10 Å do not change significantly for other choices of d. Increased persistence times tP reflect long-lived microscopic arrangements of lipid molecules, thus representing a measure of structural correlation of the lipid motion or the membrane local viscosity ηα〈tP〉. The exchange time is the time elapsing between subsequent displacements of length d. Thus, the first exchange time tx,1 is the minimum time for which ||x(tx,1)-x(tP)||>d, where x is the position of the lipid, and the following values of the exchange time ti are defined similarly by the condition ||x(tx,i)-x(tx,(i-1))||>d.
In short, the persistence time tP is the time the first step of the random walk occurs independent of the choice of the reference t = 0 time, whereas the exchange time tX is the waiting time between two consecutive jump events. The latter characterizes the dynamics of the lipid molecules by a diffusion constant D∝1/〈tX〉. The distributions of persistence and exchange times, pP and pX, respectively, are related by:
pP(t)=1〈tX〉∫t∞ds pX(s)
so that in normal bulk conditions, when pX is Poissonian, so is pP, and 〈tX〉~〈tP〉.
In regions of locally restricted motion (“jammed” regions), such as crevices at the protein surface and regions between two protomers, jump events are clustered in time, typical persistence times can become larger than the exchange times, and 〈tP〉 is increased with respect to 〈tX〉, so that ηD ∝ 〈tP〉/〈tX〉 is no longer constant. The effective viscosity η increases and is decoupled from the diffusional motion of the lipids. We calculated the distributions pP and pX in bins of 1Å×1Å using the dynamics of the phosphate group in the first third of each trajectory, and averaging over upper and lower leaflet lipids, over all homomeric trajectories for a given receptor subtype, and over all protomers in the simulation, after alignment to a common reference protein molecule. Although the heteromeric simulations yielded comparable results, these are not reported here because of their reduced statistical significance.
Position-dependent diffusion coefficients for all μ-OR protomers were calculated from the five 10 μs-long unbiased simulations carried out to simulate receptor homo-dimerization, using the Bayesian inference approach described previously in [31]. Continuous trajectories of pairs of interacting proteins were extracted from the original trajectories containing 16 copies of the receptor. To avoid problems with the interpretation of the results, frames in which a given interacting pair of protomers was in contact with a third protomer were discarded. This was achieved by checking that the distance between protomers in a given pair was smaller than distances from any other nearby protomer (see S3 Fig.). The time dependence of the two angles defining the relative orientation of the protomers (ϕ1, ϕ2) and their distance (d) was calculated for all the frames of the resulting trajectories of protomer pairs. Trajectories relative to the formation of specific interfaces were selected by restricting the analysis to regions I = {α1≤ϕ1≤β1 α2≤ϕ2≤β2}, and the distances {3 nm ≤ d ≤ 9 nm} binned to obtain a set of unbiased binned trajectories X = {dj(τ)}. The diffusion coefficient D was calculated by sampling the posterior distribution of the rate matrix from the posterior distribution p(K|X) = p(X|K) p(K), assuming a uniform prior p(K), and the likelihood
lnp(X|K)=∑jj′Njj′ln(eτK)jj′
where Njj’ is the number of observed transitions between distance bin j and bin j’. From the distribution of the rate matrix K, the distribution of the diffusion coefficients was obtained from
Average values of these quantities and their (25%, 95%) credible intervals are reported in S5 Table.
The free-energy barrier separating the μ-OR TM5/TM5 interface from simulations and the crystallographic TM5,6/TM5,6 interface was estimated using the Jarzynski equality [46] and multiple runs in which the angles ϕ1 = ϕ2 were steered from 1.2 degrees (corresponding to the TM5/TM5 interface) to 2.1 degrees (corresponding to the TM5,6/TM5,6 interface), with a velocity of 0.05 rad/ns and an elastic force of k = 800 (kcal mol−1)/rad2. The Jarzynski expression was applied separately to 5 sets of 5 independent runs to obtain the potential of mean force (PMF) and error estimates.
|
10.1371/journal.pcbi.0030122 | Flexible and Accurate Detection of Genomic Copy-Number Changes from aCGH | Genomic DNA copy-number alterations (CNAs) are associated with complex diseases, including cancer: CNAs are indeed related to tumoral grade, metastasis, and patient survival. CNAs discovered from array-based comparative genomic hybridization (aCGH) data have been instrumental in identifying disease-related genes and potential therapeutic targets. To be immediately useful in both clinical and basic research scenarios, aCGH data analysis requires accurate methods that do not impose unrealistic biological assumptions and that provide direct answers to the key question, “What is the probability that this gene/region has CNAs?” Current approaches fail, however, to meet these requirements. Here, we introduce reversible jump aCGH (RJaCGH), a new method for identifying CNAs from aCGH; we use a nonhomogeneous hidden Markov model fitted via reversible jump Markov chain Monte Carlo; and we incorporate model uncertainty through Bayesian model averaging. RJaCGH provides an estimate of the probability that a gene/region has CNAs while incorporating interprobe distance and the capability to analyze data on a chromosome or genome-wide basis. RJaCGH outperforms alternative methods, and the performance difference is even larger with noisy data and highly variable interprobe distance, both commonly found features in aCGH data. Furthermore, our probabilistic method allows us to identify minimal common regions of CNAs among samples and can be extended to incorporate expression data. In summary, we provide a rigorous statistical framework for locating genes and chromosomal regions with CNAs with potential applications to cancer and other complex human diseases.
| As a consequence of problems during cell division, the number of copies of a gene in a chromosome can either increase or decrease. These copy-number alterations (CNAs) can play a crucial role in the emergence of complex multigenic diseases. For example, in cancer, amplification of oncogenes can drive tumor activation, and CNAs are associated with metastasis development and patient survival. Studies on the relationship between CNAs and disease have been recently fueled by the widespread use of array-based comparative genomic hybridization (aCGH), a technique with much finer resolution than previous experimental approaches. Detection of CNAs from these data depends on methods of analysis that do not impose biologically unrealistic assumptions and that provide direct answers to fundamental research questions. We have developed a statistical method, using a Bayesian approach, that returns estimates of the probabilities of CNAs from aCGH data, the most direct and valuable answer to the key biological question: “What is the probability that this gene/region has an altered copy number?” The output of the method can therefore be immediately used in different settings from clinical to basic research scenarios, and is applicable over a wide variety of aCGH technologies.
| Alterations in the number of copies (gains, losses) of genomic DNA have been associated with several hereditary anomalies and are involved in human cancers [1–7]. For example, amplification of some genes, especially oncogenes, is one well-known mechanism for tumor activation [8,9], and it is involved in the deregulation of cellular control [10,11]. Copy-number alterations (CNAs) have been associated with tumoral grade, metastasis development, and patient survival [1–7], and studies about copy-number changes have been instrumental for identifying relevant genes for cancer development and patient classification [1,2,12].
A widely used technique to identify copy-number changes in genomic DNA is array-based comparative genomic hybridization (aCGH). Two DNA samples (e.g., problem and control) are differentially labeled (often with fluorescent dyes) and competitively hybridized to chromosomal DNA targets. After hybridization, emission from each of the two fluorescent dyes is measured, and the signal intensity ratios are indicative of the relative copy number of the two samples [1,2,13]. Therefore, a key step in any study of the relationship between altered copy numbers and disease is using the fluorescence ratio data to identify genes and contiguous chromosomal regions with altered copy numbers.
The main biomedical problem, both for the study of the CNAs per se and for downstream analysis (e.g., relationship with gene expression changes or patient classification), is the accurate identification of the genes/chromosomal regions that have an altered copy number. Satisfactorily dealing with this problem requires a method that (1) provides direct answers that can be used in different settings (e.g., clinical versus basic research), (2) reflects the underlying biology and accounts for key features of the technological platform, and (3) can accommodate the different levels of analysis (types of questions) addressed with these data.
First, estimates of the probabilities of alteration (instead of p-values or smoothed means) are the most direct and usable answer to this problem [14,15]. Probabilities can be used in contexts that cover basic research to clinical applications [1,2] so that, for instance, a clinician might require high certainty of alteration of a specific gene before more invasive procedures, whereas a basic researcher can consider for further study genes that show only a moderate probability of alteration (e.g., probability >0.5). Finally, appropriately used, probabilities of alteration can account for uncertainty in model building [16,17].
Second, the analysis should incorporate distance between probes [2,15,18–21]: widely used aCGH platforms such as those based on cDNA microarrays, oligonucleotide arrays, and representational oligeonucleotide microarray analysis (ROMA) lead to variable coverage across chromosomes, with unequal distances between probes (i.e., some regions have probes that are very close to each other, whereas in other regions probes are very far apart). As copy-number changes involve chromosome segments, contiguous loci will have the same copy number, unless there is an abrupt change to another copy number [1,22]: the farther apart two loci are, the more likely it is that a copy-number event will have taken place in between them. Thus, in densely covered regions, the copy number of a probe is a good predictor of the copy number of the neighboring probes. In contrast, in poorly covered regions, contiguous probes or loci might be many thousands of kilobases apart, making it more likely that at least one copy-number change has taken place, and consequently, a probe provides less information about the likely state of its neighboring probes. Therefore, unless we use a platform where all probes are equally spaced, we need to use the distance between probes (and not just the order) so that the information that consecutive probes provide is adequately accounted for.
Third, depending on the focus of the study, the analysis should be conducted either chromosome by chromosome, or genome-wide [14–16]. Analyses at the chromosome level are appropriate to detect alterations in copy numbers of loci relative to the rest of the loci in that same chromosome, regardless of that chromosome's ploidy (a trivial example would be detection of copy-number changes in loci of the human Y chromosome in an otherwise diploid genome). On the other hand, detection of copy-number changes that affect most of a chromosome often require genome-wide analysis (in chromosome-wide analysis, as the mean or median chromosome level is used as the reference, detection of such changes is virtually impossible). Moreover, the use of genome-wide analysis can offer statistical advantages (e.g., reduced variance of estimation). As both types of analysis offer complementary information because they focus on different biological phenomena (chromosomal gains/losses versus gains of loci within chromosomes), a suitable method should allow these two approaches.
Available methods for the analysis of aCGH fail some or most of these requirements. Smoothing techniques [21,23–28] do not use interprobe distance, nor do they provide posterior estimates of the likely state of each probe/clone, and data from each chromosome are analyzed independently of each other. Hidden Markov models (HMMs) and related techniques offer a flexible modeling framework, and can provide probabilities of alteration [14–16]. Some HMM-based methods [16,19], however, do not incorporate the distance between probes, assuming instead that interprobe distance is constant. In addition, most of them do not deal satisfactorily with the unknown number of hidden states (the true number of states of copy number). Some methods fix in advance the number of hidden states to three [14,15] or four [16]: prespecification of the number of states has the consequence of jumbling all changes involving multiple gains into a single state with a common mean, which is biologically questionable [22], especially as the resolution of the technology improves. Moreover, the identification of important genes for disease sometimes requires examining the amplitude of CNAs and not just their presence and location [1]; collapsing states into three or four, however, precludes examining in fine enough detail the amplitude of CNAs. A better approach would provide posterior probabilities of the number of states; using such a procedure over many different experiments will tell us whether three- or four-state models are a reasonable simplification. Of those methods that do not assume a fixed number of hidden states [18,19,22], one of them [22] cannot be used for questions about the number of hidden states, or for breaking the data into more categories than gained/lost/no change, which are increasingly important questions with higher-resolution techniques and are needed for distinguishing regions of moderate copy gains from regions of large copy gains; see also above for relationship between amplitude of CNAs and presence of disease genes. The remaining two [18,19] fit HMMs for a range of number of states and then use Akaike information criterion (AIC)–based model selection, but AIC-based selection with HMMs has not been theoretically justified [29] and does not provide a probability of the likely number of states; moreover, selecting a single model leads to underestimation of the true variability in the data. These two methods, in addition, use a final clustering step of hidden states that introduces several ad hoc decisions.
We have developed a method, reversible jump aCGH (RJaCGH), that fulfills the three requirements above, and does not suffer from the limitations discussed for other methods. Our method is applicable to aCGH from platforms including ROMA, oligonucleotide aCGH (oaCGH; including Agilent, NimbleGen, and many noncommercial, in-house oligonucleotide arrays), bacterial artificial chromosome (BAC), and cDNA arrays [1,13]. We start our modeling by noting that, for a given chromosome or genome, the copy numbers of genomic DNA (e.g., 0, 1, 2 copies, . . . ) of different probes or segments are an unknown finite number. Thus, probes or segments could be classified into several groups with respect to their (unknown) copy number. In addition, as mentioned above, we expect that the copy number of a probe will be similar to the copy number of its closest neighbors, with that expected similarity decreasing when probes are farther apart. Finally, for a given copy number, the aCGH fluorescence ratios should be centered around a log2 value, with some random noise. We want to use the observed log-ratios to identify regions with altered copy number.
The biological features of this model (a finite number of unknown or hidden states that are indirectly measured, with states of close elements likely to be similar, and variable distances between probes) can be modeled with a nonhomogeneous HMM [29]. To provide a direct estimate of the probability that a given probe or region has an altered copy number, we use a Bayesian model computed via Markov chain Monte Carlo (MCMC). Since we do not know the true number of hidden states, we fit models with varying numbers of hidden states and, to allow for transdimensional moves between models with different numbers of states, we used reversible jump [30]. After running a large number of MCMC iterations, we can summarize the posterior probabilities. First, we obtain posterior probabilities for the number of states. Conditional on a given number of states, each model provides posterior distributions of the parameters of interest (e.g., means, variances, transition matrices). From the latter, we can obtain posterior probabilities that a probe is gained or lost. To obtain our final estimates, we incorporate the uncertainty in model selection by using Bayesian model averaging [17], with estimates weighted by the posterior probability of each number of states, for the probabilities of probes being gained or lost. We call the complete statistical method RJaCGH (from reversible jump–based analysis of aCGH data).
We applied RJaCGH and the best performing alternative methods (based on two recent reviews [20,31]) to the 500 simulated datasets of [31] (see also Protocol S1). These are data “...simulated to emulate the complexity of real tumor profiles” and designed to become “...a standard for systematic comparisons of computational segmentation approaches,” [31] and are not data simulated under our own model. To assess the effect of variable interprobe distance, we randomly deleted data points (see details in Protocol S1) so that each original simulated dataset gives rise to another four datasets with (an average of) 10%, 25%, 50%, and 65% of observations missing. The length of these gaps is modeled by a Poisson distribution, so larger percentages of missing data correspond to larger variability in interprobe distances.
Results in Figure 1 (see also Figure 1 in Protocol S1) show the excellent performance of RJaCGH, and how it outperforms alternative methods. Moreover, Figure 2 (see also Figures 2 and 3 in Protocol S1) shows that the difference between RJaCGH and alternative approaches is accentuated when we consider jointly the effects of noise and variability in interprobe distance. Analysis using three other performance statistics (false discovery rate, sensitivity, and specificity) show the same overall patterns (see Protocol S1, Figures 2 and 3): for some specific statistics, RJaCGH can be second (but very close) to another approach; this other approach, however, performs poorly with respect to the remaining statistics.
This paper focuses on the statistical performance of the methods compared. In terms of speed, nevertheless, our approach is clearly the slowest one. We are currently working on improving the speed of the execution both by using more efficient algorithms and by using parallel computing.
Similar results are obtained when applying these methods to a real dataset of nine cell lines [32], and when comparing the predicted ploidy with the known ploidy (see Protocol S1, Figure 4). Overall, therefore, there is strong evidence that RJaCGH is the best performing of the existing methods.
The excellent performance of RJaCGH is a result of the statistical method used, which is essentially a careful and rigorous development from first principles. We set out to obtain a method that allows us to seamlessly incorporate interprobe distances (to allow usage over varied technological platforms), that makes no untenable assumptions about the true number of copy levels (since this is likely to vary between datasets), that permits analysis at the chromosome and the genome level, and, finally, that returns posterior probabilities of alteration, because these posterior probabilities constitute the direct answer to the basic biomedical question (“Is this gene likely to have an altered copy number?”).
Based simply on our usage of interprobe distance, we should expect RJaCGH to perform better than all alternative approaches, with the possible exception of BioHMM [18], as interprobe distance variability increases. Moreover, RJaCGH adapts to variable noise in the data, without the need for fine-tuning of parameters (all results reported are obtained from the default settings of RJaCGH). As noted above, the relative advantage of RJaCGH increases as the interprobe variability increases and the noise in the data increases, which shows that our theoretical developments have practical consequences and emphasizes the importance of both accounting for interprobe distance and appropriately modeling variance in the data.
In addition, we use Bayesian model averaging, which has been repeatedly shown [33] not only to account for uncertainty in model selection but also to lead to point estimators and predictions that minimize mean square error. On its own, our usage of Bayesian model averaging could be largely responsible for the better performance of RJaCGH over all other methods, even in the absence of interprobe distance variability and when there is low noise in the data (left of Figure 1, and left of bottom-row panels in Figure 2). In addition, reversible jump allows us to consider a variety of models (regarding number of states), and its birth and split moves are also beneficial for a more thorough exploration of the posterior probability (within a model with a given number of states) when the density is multimodal. Finally, our method, in contrast to other approaches (e.g., DNAcopy), can identify single-clone aberrations, which might be key for large-scale genomic deregulation if the single-clone aberrations affect certain specific genes or promoters; for example, the inability to detect single-gene alterations is shown to have an effect in a study of pancreatic adenocarcinoma [5], where the loss of the SMAD4 tumor suppressor is undetected.
In addition to features that can be compared with other methods, RJaCGH has two unique features that set it apart from most alternative approaches. First, the user can analyze data at either the genome or the chromosome level, thus addressing different types of questions. Some approaches (e.g., BioHMM, HMM, GLAD, DNAcopy) allow us to perform genome-wide inferences, but they use essentially an ad hoc postprocessing of results of analysis that is conducted at the chromosome level. Finally, one of the main features of RJaCGH, its returning of posterior probabilities of CNAs, simply cannot be compared with most alternative methods as they do not provide this type of output. What most alternative approaches return are smoothed means, p-values, or a classification into states without any assessment of the uncertainty of this assignment to states. But a probability of alteration (which RJaCGH returns) is much easier to interpret and to use (with possibly different thresholds depending on the type of research question), and is often the direct answer to the basic biomedical question. The few alternative approaches that return probabilities of alteration [14–16] all make the untenable assumption that the true number of biological states of alteration is three [14,15] or four [16].
Directly returning probabilities of alterations has profound consequences, both for current practices and for future developments. As argued above, these probabilities are the direct answer to the question “Does this gene have an altered copy number?”; p-values or smoothed means are not a direct (and often not even an indirect) answer to that question. In addition, the improvement in the resolution of aCGH techniques [2,13] is increasingly allowing for multiple assayed spots per gene. Probabilities of alteration for each spot can be combined to find the gene-level probability of alteration, a distinct advantage over smoothed means or p-values.
For currently active research areas, the availability of rigorously obtained probabilities of alterations has far-reaching consequences, both in terms of the biological phenomena that can be exposed and as an avenue of further research. First, the availability of probabilities of alteration should improve the identification of regions with consistent alterations across samples [34,35] in a statistically rigorous way (including, if needed, control of false discovery rate), and the detection of subgroups of samples according to recurrence patterns [4,35,36]. Critical disease genes are often located in CNAs that are recurrent across individuals and that have at least some high-amplitude changes [1,35,37], and analysis of aCGH data has allowed us to identify subgroups, within established diseases, that could have therapeutic relevance (e.g., in glioblastoma [4]). Available methods use the assignment of each gene to one state (equivalent to assuming that there is complete certainty in this assignment); however, we would not want to give the same weight, when looking for minimal common regions, to a gene with a probability of being gained of 51% and to a gene with a probability of being gained of 90%, since this practice will lead to a coarser definition of boundaries and can even preclude the detection of some minimal regions altogether. The inherent limitations of methods that use a simple categorization into gain/loss/no change with an assumed 100% certainty have already been recognized by some of the developers of such methods [35]. Moreover, incorporation of amplitude of change, which might be a crucial feature of minimal common regions that harbor critical disease genes [1,5,35,37], is not feasible with some methods [34], but should be straightforward by combining posterior probabilities and posterior means of each state, as returned from RJaCGH.
Second, posterior probabilities of being in a specific state, together with the estimated posterior mean of each state, can be used as the basis for a statistically rigorous and biologically sound approach for identifying breakpoints. At present, the identification of breakpoints depends completely on the resolution of the method, and does not allow us to combine the probability of membership in different states with the biological relevance of an estimated mean difference; however, the precise definition of boundaries and amplification maxima are important not only for the study of genomic copy numbers, but also for understanding the relationship between aCGH and expression data [38].
Third, the model of RJaCGH can be extended to provide rigorous downstream analysis of aCGH, including patient classification [1,31] and the integration of gene expression and proteomic data [12,31]. CNA data analysis, compared with mRNA expression data, can be performed on formalin-fixed paraffin-embedded material, and CNAs define key events that drive tumorigenesis, and thus are probably more valuable as prognostic markers and as predictors of treatment response [39,40]. Improved resolution of CNA data analysis, however, can be crucial in obtaining very valuable classifiers, as evidenced from the “almost success” of some studies attempting to differentiate BRCA2 from BRCA1, BRCAX, and sporadic cases in breast tumors (see discussion in [40]); the finer resolution provided by probabilities and posterior mean estimates might be pivotal here. Incorporating expression and proteomic data, on the other hand, is the basis for the identification of copy-number changes that are significant in the development of disease [1,41,42]. Since changes in copy number are not always reflected in changes in expression [1,5], analytical methods that provide finer resolution are crucial. Moreover, within a probabilistic framework it is possible to systematically and rigorously address questions of how CNAs in a given chromosome affect expression changes in genes located in other chromosomes, an increasingly important research question [43]. Finally, the posterior probabilities and means returned from aCGH can be considered as denoised [44,45] signals from the log2 aCGH ratios that reflect underlying copy number variation; as such, these are highly relevant to the recent studies on the relationship between copy-number variation and complex phenotypes [46,47], which emphasize the importance of copy-number variation in genetic diversity and disease in humans.
We use a nonhomogeneous HMM with Gaussian emissions. We can either fit one model to all the chromosomes of an array, or we can fit a different model for each chromosome of an array. Let n be the number of probes, and k the number of different copy numbers in the collection of probes. Let Si be the true state (copy number) of the probe i: Si = {1, …, k}i =1,...,n. Let Yi be the relative copy number of the probe i, that is the log ratio of fluorescence intensities between tumor and control samples. Let di be the distance in bases between probe i and probe i + 1. How distance is measured depends on the platform: distance can be the distance from the end of the spot to the start of the next, if the length of the spots is proportional to the length of the probe (so we have the same information for every probe), or the distance between the midpoint of the spots, if the length of the spots is not proportional to the length of the probe. We normalize these distances between 0 and 1 to increase numerical stability (with probes in adjacent bases with a scaled distance of 0).
We assume that {Si} follows a nonhomogeneous first-order Markov process, as: P(Si = si | Si−1 = si−1, Xi−1 = xi−1) =
Biologically, we expect that
,the probability of staying in the same hidden state, is a decreasing function of Xi−1, so the dependence of the state of a probe onto the next one is lower the farther the probes are. We also expect that when the distance between two probes is maximal, the state of a probe should be independent from the state of its predecessor. Thus, we model the transition probabilities as:
where β has the form:
with all βij ≥ 0 ∀ i, j. Finally, conditioned on {Si}, {Yi} follows a Gaussian process:
Similar approaches have been used before with nonhomogeneous HMM [48,49]. In our case, the transition matrix should fulfill the following biologically based properties: (1) the probability of remaining in the same hidden state should be a decreasing function of the distance between a probe and the previous probe; and (2) when the distance between two probes is maximal, the state of a probe should not be affected by the state of the previous probe. With the above parameterization, and since the diagonal of β is zero (which is also needed for the parameters to be uniquely defined), the probability of remaining in the same state i is
, a decreasing function of distance (x). Moreover, as distances are scaled between 0 and 1, when the distance between two probes is 1, the probability of staying in the same state is 1 / k, where k is the number of states; therefore, when the distance is maximal, the state of a probe does not depend on the state of the previous probe. (The value of this “maximal distance” beyond which two probes are considered independent is a parameter to the model, and can be adjusted taking into account the specific array characteristics).
For computational reasons and modeling flexibility, we opted for Bayesian methods using MCMC. To fit models with varying number of hidden states, we used reversible jump. Suppose that we have a collection of K HMM models, and each of them has a number of k hidden states, from k = {1, . . ., K}. Let θ(k) be the HMM associated to k, that is, θ(k) = {μ(k), σ2(k), β(k)}. The prior distributions for the model are the usual ones in mixture problems [50]: p(k) is the prior for the number of hidden states with p(k) ~ U(1,k), p(θ(k)| k) is the prior of the HMM conditioned to k, the number of hidden states with u(k) ~ N(α,ϱ2), where α and ϱ are the median and range of Yi; σ2(k) ~ IG(ka, g), where ka is 2 and g is ϱ2(Yi) / 50; βk) ~ Γ(1, 1). The likelihood of the model, L(y; k, θ(k)) can be computed by forward filtering [29], so the joint distribution is p(k)p(θ(k)|k) L(y; k, θ(k)).
We can draw samples from the posterior distribution through a reversible jump MCMC (RJMCMC) algorithm [30]. In RJMCMC, we explore the posterior distribution of possible models, jumping not only within a model but also between models with a different number of parameters. To match the difference between degrees of freedom, some random numbers u with density P(u) are generated, so if we are in state x, the new one is proposed in a deterministic way x′(x,u). The reverse move is the inverse of that function: x(x′,u′). This way, the usual Metropolis-Hastings acceptance probability can be computed [50]:
where L(y | x) is the likelihood, p(x) are the priors, p(u | x) are the densities of the candidates, and
the determinant of the Jacobian of the change of variable. We combine several Metropolis steps in a sweep [29,51].
(1) Update HMM of a model using a series of Metropolis-Hastings moves. (We do not use Gibbs Sampler to avoid the hidden state sequence from becoming part of the state space of the sampler, so dimensionality is reduced and reaching convergence is easier).
(2) Update model (birth/death). When we have r states, a birth/death move is chosen with probabilities pbirth(r) and pdeath(r) (these are 1/2 except in the cases when no movement of that type can be made, [e.g., a death move when there is only one state]). If a birth move is selected, a new state is created from the prior distributions and accepted with probability
If a death move is chosen, a random state is deleted with a probability inverse to Equation 4.
(3) Update model (split/combine). A split/combine move is attempted with probabilities psplit(r) and pcombine(r) (again, 1/2 except when a move cannot be made). If a split move is selected, an existing state i0 is split into two, i1, i2:
Split column
Split row
This move is accepted with probability
The split move must follow the adjacency condition [50]: the resulting states must be closer between them than to any other existing ones. If a combine step is selected, the symmetric move is performed, and the inverse probability of acceptance is computed.
The combination of birth and split moves makes it possible not only to visit models with a different number of parameters, but also to explore more thoroughly the posterior probability in the case of a parameter with a multimodal density.
These moves are common ones [29,51], but we have changed several aspects of their design to improve the probability of acceptance, which is the most difficult step in reversible jump [29,30,51]. We constrain the variance of every state so that it cannot be greater than the variance of the entire data. Also, we have added the adjacency condition mentioned before, and used centering proposals [52]. To prevent label-switching of states, we have ordered the states according to means after every iteration of the sweep [50].
We run the former algorithm a large number of times (e.g., 50,000) and, after discarding the first iterations as burn-in, we keep the last (e.g., 10,000) samples as observations from the joint distribution so that we can make inferences from it. For every model that has been visited, we obtain the posterior probabilities of the mean copy number of every state, the variance of the copy number of every state, and the function of transitions between hidden states. By counting the number of times that each model has been visited, we obtain an estimate of the posterior probability of each model (i.e., we avoid using Bayesian information criterion [BIC] or AIC). Then, applying the Viterbi algorithm [29] to every sample obtained from the MCMC, and, as this sample is a function of the HMM, we can obtain its posterior probability, something that usual Viterbi cannot. From the Viterbi paths for all the samples, we can then compute the posterior probability that a probe belongs to every state or the probability that a sequence of probes is in a given state.
When obtaining posterior probabilities of copy-number change, we use Bayesian model averaging [17] over all models visited. Let Si be the lost, gained, no-change status of probe i, K the set of the models considered (in our case, that would be HMMs with 1, . . ., K number of states), Mk the model with k number of states, and Si | Mk the state of probe i according to model k. We compute the unconditional (with respect to model selection) probability for the probe i as:
As in any MCMC approach, it is crucial to assess convergence of the sampler. We follow common practice [53] of running several chains in parallel. The convergence of the sampler depends strongly on the distribution of the candidates in Metropolis-Hastings. That is, for every iteration, a new value for the parameters is proposed from a distribution centered in their current values. The standard deviation of that distribution must be chosen in a way that samples explore all the parameter space. These standard deviations are not parameters of the model in the sense that different values give different fits, but values that can speed up convergence of the algorithm. The convergence of the posterior probability of the number of hidden states is reached when a large enough number of transdimensional moves is made. This number need not to be large if the likelihood is substantially higher in a particular model and data size is big enough. The birth and death moves only depend on the priors, but the split and combine moves depend also on their own design and the values of τμ and τβ (see Equation 5 and Equation 7). The priors chosen have been extensively tested in mixture models [50]. In addition, the priors and rest of the parameters have very little effect: even small CGH arrays contain thousands of points, so that the likelihood from the data dominates any prior. With the 2,500 simulated datasets analyzed, we have only needed to specify the number of burn-in—50,000—and to-keep samples—10,000, and the number of chains—four, and in only nine cases was there evidence of nonconvergence, which was solved by rerunning the samplers.
We have implemented RJaCGH using C (for the sweep algorithm) and R [54], and all analysis and comparisons have been done in R. The code that implements RJaCGH is freely available from the usual Comprehensive R Archive Network (CRAN) repositories as package RJaCGH (http://cran.r-project.org/src/contrib/Descriptions/RJaCGH.html) or from the repository at Launchpad (https://launchpad.net/rjacgh). All data and code used for this paper are also publicly and freely available (see details in Protocol S1).
|
10.1371/journal.pgen.1008369 | Dosage regulation, and variation in gene expression and copy number of human Y chromosome ampliconic genes | The Y chromosome harbors nine multi-copy ampliconic gene families expressed exclusively in testis. The gene copies within each family are >99% identical to each other, which poses a major challenge in evaluating their copy number. Recent studies demonstrated high variation in Y ampliconic gene copy number among humans. However, how this variation affects expression levels in human testis remains understudied. Here we developed a novel computational tool Ampliconic Copy Number Estimator (AmpliCoNE) that utilizes read sequencing depth information to estimate Y ampliconic gene copy number per family. We applied this tool to whole-genome sequencing data of 149 men with matched testis expression data whose samples are part of the Genotype-Tissue Expression (GTEx) project. We found that the Y ampliconic gene families with low copy number in humans were deleted or pseudogenized in non-human great apes, suggesting relaxation of functional constraints. Among the Y ampliconic gene families, higher copy number leads to higher expression. Within the Y ampliconic gene families, copy number does not influence gene expression, rather a high tolerance for variation in gene expression was observed in testis of presumably healthy men. No differences in gene expression levels were found among major Y haplogroups. Age positively correlated with expression levels of the HSFY and PRY gene families in the African subhaplogroup E1b, but not in the European subhaplogroups R1b and I1. We also found that expression of five out of six Y ampliconic gene families is coordinated with that of their non-Y (i.e. X or autosomal) homologs. Indeed, five ampliconic gene families had consistently lower expression levels when compared to their non-Y homologs suggesting dosage regulation, while the HSFY family had higher expression levels than its X homolog and thus lacked dosage regulation.
| The human genome harbors two sex chromosomes—X and Y. Among them, the Y chromosome is present only in males. Deletions of portions of this chromosome have been linked to male infertility, however exactly why the loss of these genes leads to this condition is not well understood. Here we study a group of Y chromosome genes called ampliconic genes, which are expressed in testis and are frequently deleted in males with infertility. These genes are organized in nine gene families, each of which harbors multiple copies of genes highly similar in sequence. In this study, we aimed to establish a baseline of their variation in copy number and in gene expression—one measure of genes’ functional output—by studying 149 healthy men. We found that testis tolerates a wide range of copy number and expression variation of Y ampliconic genes. Additionally, we demonstrated that gene expression within most Y ampliconic gene families depends on the expression levels of gene family members located outside of the Y chromosome, i.e. they undergo dosage regulation.
| The human Y chromosome harbors 10.2 million bases (Mb) of ampliconic regions containing nine protein-coding multi-copy gene families [1]. These genes are important not only because of their association with male infertility [1,2] but also because they might hold the key to understanding the evolutionary forces that have shaped the Y chromosome. Ampliconic gene families show a high level of copy number variability [3–5] and, possibly, a similar variability in gene expression levels. Understanding the relationship between these two variabilities is an important step in the study of these genes. Yet, there has been no comprehensive investigation to-date that explores expression of these gene families and its connection to copy number at a large, population-level scale.
Studying ampliconic gene families has been a considerable challenge because they exhibit a much higher intra-familial sequence similarity than other gene families. The majority (eight out of nine) of Y ampliconic gene families are located in palindromes—structures composed of highly similar inverted repeats (arms) around a relatively short unique sequence (spacer). The arms within a palindrome are 99.9% identical to each other, which results in a high sequence identity among paralogous genes located on the arms [1]. The ninth family, TSPY, is present as an array of tandem repeats outside of palindromes [1], however its genes still share sequence identity of >99%. It has been hypothesized that the Y chromosome has acquired its ampliconic structure as a way of facilitating gene conversion [6], which can overcome the decay due to a lack of inter-chromosomal recombination [7,8].
Why these ampliconic gene families are preserved on the Y chromosome remains an open question. It has been suggested that this is due to sexual antagonism eventually leading to increased male reproductive fitness [6,7,9]. Sexual antagonism is expected to lead to the accumulation of genes and mutations benefiting males on the Y chromosome [10]. Consistent with the sexual antagonism hypothesis, all ampliconic genes on the Y are expressed exclusively or predominantly in testis. However, it is also possible that these genes have recently evolved under relaxed function constraints. The ability to analyze the expression levels of Y ampliconic genes at a large scale can help in exploring their potential functional constraints via comparing their testis expression level to that of their non-Y homologs (when available). For instance, if a Y ampliconic gene family undergoes neo-functionalization, then its resulting expression level is expected to be independent of and potentially higher than that for its non-Y homologs (which we assume retained the ancestral function).
In support of some functional constraints is the observation that the loss or partial deletion of Y ampliconic gene copies is linked to infertility in humans. For example, TSPY copy number was linked to both infertility [11] and sperm count [11–13]. The long arm of the human Y chromosome includes three azoospermia factor regions (AZFa, AZFb, and AZFc), which cover most of the ampliconic genes families and are active during different phases of spermatogenesis [14]. Complete or partial deletion of these regions is linked to azoospermia and arrest of spermatogenesis [2,12,14–16]. Presumably, copy number decrease linked with infertility is accompanied by a reduction in gene expression of the affected Y ampliconic gene families, however this is yet to be demonstrated.
Recent studies indicated high variation in Y ampliconic gene copy number in healthy men [3–5]. Skov and colleagues [4] studied Y ampliconic gene copy number variation in 62 men of Danish descent and identified multiple copy number changes across all nine gene families among unrelated individuals, as well as copy number differences for the TSPY and VCY gene families between a father and a son. Ye and colleagues [3] assessed Y ampliconic gene copy number variation in 100 individuals from around the world. They observed that the size of gene family is correlated with its variation in copy number: larger families, such as TSPY and RBMY, have higher levels of variation, however the variation appears to be independent of the Y haplogroup. Two men rarely had the same Y ampliconic gene copy number profile and, when they did, this was likely a result of homoplasy. Lucotte and colleagues [5] used the data from the Simons Genome Diversity Project [17] and observed substantial variation in copy number in six out of nine human Y ampliconic gene families [5]. Teitz and colleagues [18] assessed copy number of full-length Y chromosome amplicons located in the AZFc region in men sequenced by the 1000 Genomes Project [18]. Their results suggest that selection has preserved the ancestral ampliconic gene copy number on the Y chromosome in diverse human lineages [18].
These multiple studies of copy number notwithstanding, there has been little investigation of gene expression of Y ampliconic genes. A recent study investigating the expression of Y ampliconic genes during male meiosis found that gene families with high variation in copy number also have high expression levels at different stages of sperm development [5]. Other than the results of this single study, there is a big gap in our understanding of variation in expression of Y ampliconic genes among humans, even though gene expression could be a better predictor of genes’ functions than copy number. Additionally, previous studies have reported that aging affects gene expression [19,20].
Even less is known about how variation in copy number of Y ampliconic genes affects their gene expression. Most parsimoniously, a gain of a complete gene copy should lead to an increase in gene expression levels, unless the extra copy obtains a new function through neo-functionalization, has decreased functional demands due to sub-functionalization or is lost due to pseudogenization. Indeed, this parsimonious hypothesis was supported by the data from the 1000 Genomes Project, where most genes overlapping multiallelic copy number variations (CNVs) display a positive correlation between copy number and gene expression [21]. However, studies across different model organisms have reported that differences in copy number result in increased, decreased or unchanged expression levels among individuals in a population [22]. This more complex relationship can be caused by several scenarios during duplication. For instance, a tandem duplication event may not include regulatory elements, may physically disrupt topologically associated domains (TADs), which prevents the interaction of the gene with its enhancer in 3D space [23,24], or may result in a new copy acting as a negative feedback loop to reduce transcription [22]. Moreover, a non-tandem duplication may occur to a site that is not transcriptionally active [22]. Which of these parsimonious or more complex scenarios occurs on the human Y chromosome ampliconic genes has not been explored.
In this study, we explored the above questions by analyzing the largest data set available to-date consisting of expression data from testis, along with matched whole-genome sequencing data, from 170 men, as generated by the Genotype Tissue Expression (GTEx) consortium [25]. Simultaneously, we developed a novel computational tool AmpliCoNE to estimate the copy number of an ampliconic gene family from sequencing data. Such estimation is complicated by the presence of multiple highly-similar gene copies in the reference, which makes conventional tools inapplicable [26]. Custom strategies have been developed and shown to be effective [4,5,21,27–29], but we did not identify any existing software that could be run directly on Y chromosome ampliconic gene families.
Using AmpliCoNE, we explored whether variation in Y ampliconic gene expression levels could be explained by variation in gene copy number, Y haplogroup, and individual’s age. We correlated the estimated with AmpliCoNE copy numbers of Y ampliconic gene families to their expression levels in testis, and studied how this correlation is affected by Y haplogroups. Additionally, we investigated how testis-specific expression of Y ampliconic genes diverged from their non-Y homologs during evolution.
AmpliCoNE is composed of two programs. The first (AmpliCoNE-build) is executed only once to process the reference genome. It takes the location of all the gene copies in the reference genome, grouped by family, determines which positions in the genes are informative (i.e. where read depth is an effective predictor of copy number) and which positions in the reference can be used as a control (where copy number variation is infrequent and the read depth has limited noise). The second step (AmpliCoNE-count) is then executed separately for every sample. It parses read alignments and measures the GC-corrected read depth at the informative positions. It then accumulates this information at a family-level and reports the copy number for each gene family, using the read depth at control positions as a baseline. We provide further details in the Methods.
To evaluate AmpliCoNE’s accuracy, we ran it on simulated data and whole-genome short-read data from the Genome in a Bottle (GIAB) consortium [30]. Using the hg38 human genome reference, we simulated three datasets with varying copy numbers of RBMY, TSPY, and VCY gene families and kept the copy numbers for the remaining six gene families constant (i.e. with the copy number found in the reference). AmpliCoNE estimated ampliconic copy numbers correctly 100% of the time in the simulated datasets (S1 Table). We then compared gene family copy numbers between different GIAB experimental runs (technical replicates) for the same human sample (S2 Table), as well as between a father and a son (which can be treated as biological replicates because copy number differences between generations are expected to be rare [4]). AmpliCoNE consistently predicted copy numbers with a difference of less than 0.5 copies per family. We tested AmpliCoNE at different depths of coverage and showed that it can predict similar copy numbers (estimates with difference of less than 0.5) even for datasets with the Y chromosome sequencing depth as low as 6x (S3 Table). AmpliCoNE’s runtime is dependent on the number of reads it needs to process. For instance, it took AmpliCoNE 11 minutes to process the GTEx Y-chromosome-specific BAM file (~500 MB in size).
To measure the concordance between AmpliCoNE’s copy number estimates and complementary non-sequencing assays, we used droplet digital PCR (ddPCR). Both AmpliCoNE and ddPCR were applied to estimate Y ampliconic gene copy numbers for four males sequenced by the GIAB consortium (Tables 1 and S4) [30]. The ddPCR estimates were identical to AmpliCoNE estimates for five out of nine gene families (BPY2, DAZ, HSFY, PRY, and XKRY) in all four samples. The CDY and RBMY family copy numbers differed between the two methods in only one and two individuals, respectively. The VCY and TSPY family copy number estimates differed in three and four individuals, respectively. Compared with ddPCR, AmpliCoNE consistently underestimated the copy number for the VCY gene family. Previous studies have indicated presence of X-to-Y gene conversion between VCX and VCY [31,32]. We investigated this case in more detail and discovered that genes from the VCY family harbor only a very short (220-bp) sequence distinguishing them from their VCX paralogs. This sequence has a low sequencing depth even after GC correction, which results in the underestimation of the VCY copy number by AmpliCoNE. In the case of TSPY, it is known to have many highly-similar pseudogene copies which may themselves vary in copy number, which can potentially confound both AmpliCoNE and ddPCR estimates. These caveats notwithstanding, AmpliCoNE’s biases in estimating copy numbers for TSPY and VCY are consistent across samples and thus should not affect our results in a systematic way.
Using AmpliCoNE, we estimated copy numbers of Y chromosome ampliconic genes in 170 presumably healthy men whose genomes were sequenced in their entirety as part of the GTEx project [25]. These individuals (S5 Table) were selected because they had matched testis expression data. The individuals belonged to ten major haplogroups: B, E, G, I, J, L, O, Q, R, and T (Table 2). The majority of the samples in the dataset had European or African Y haplogroups, with a few Asian haplogroups present. We also used AmpliCoNE to estimate the copy numbers of X-degenerate genes, which are expected to be 1 in healthy samples. Three samples had copy number estimates close to zero for two or more ampliconic gene families, or had less than one copy for several X-degenerate genes, which could suggest an individual with a disease or could result from a technical artifact, and thus were removed from the downstream analysis. As a result, we retained 167 samples.
Gene families with higher median copy number had higher variation when compared to gene families with lower median copy number (R2 = 0.91; S1 Fig). RBMY and TSPY were the largest gene families and displayed the highest variation in copy number (5–14 and 20–64 copies for RBMY and TSPY, respectively). HSFY, PRY, VCY, and XKRY were the smallest gene families, which on average had two copies per individual, and displayed low variation in copy number. We observed a positive correlation in copy number among BPY2, CDY, and DAZ gene families, which could be explained by their co-localization on palindrome P1; duplication or deletion involving P1 can affect the copy numbers of all three gene families (Fig 1A).
We expected to observe a higher probability for gene families with lower median copy number to be completely deleted due to random rearrangements. Therefore, we aimed to test whether the gene families with lower copy number in human had a higher chance of being deleted in non-human great ape species. It is known from previous studies that the VCY gene family is missing in bonobo, gorilla, and orangutan, whereas the HSFY, PRY, and XKRY families are missing in bonobo and chimpanzee [33]. Consistent with our hypothesis, the HSFY, PRY, VCY, and XKRY gene families had low copy numbers in humans (S1 Fig; S6 Table).
To explore the relationship between ampliconic gene copy number and their expression levels, we analyzed testis expression data from the same 167 humans whose Y ampliconic gene copy number was estimated with AmpliCoNE. After removing outliers (see Materials and Methods), we retained 149 samples and obtained expression levels for each gene family—the sum of expression of all the gene copies within each family—in each of them. We found that, similar to our observation for copy numbers (S1 Fig), families with higher gene expression levels had higher variation in gene expression (R2 = 0.99; S2 Fig). The TSPY family had the highest gene expression level and the highest variation in expression across individuals, and XKRY—the lowest (S6 Table; S2 Fig). The XKRY gene family could be considered to be not expressed (as its expression levels are zero) in 58 individuals or expressed at very low levels (with DESeq2 normalized read count < 10) in the remaining 91 individuals. DAZ, HSFY, and RBMY gene families had similar median expression levels and variance among themselves (S6 Table; S2 Fig). Within our dataset, we found two sets of ampliconic gene families whose expression levels were positively correlated with each other (Fig 1B). The first set included BPY2, CDY, HSFY, and PRY, and the second set—DAZ, TSPY, RBMY, and VCY (Fig 1B). The expression of these sets of gene families could be co-regulated or might have cell-type specificity.
When we investigated the relationship between expression levels and copy number among all 149 individuals across nine ampliconic gene families, we found that more copious gene families tended to have higher expression levels in comparison to the less copious gene families (Fig 2). Indeed, the expression levels were positively correlated with estimates of copy numbers (Spearman's rank correlation rho = 0.43; P-value < 2.2x10-16). The DAZ, HSFY, and VCY gene families appeared to be outliers in this analysis, as they had gene expression levels similar to the RBMY gene family even though their median copy number estimates were approximately half or less than half of RBMY gene family. The DAZ gene family had similar gene copy number yet higher expression levels when compared to the CDY gene family. The XKRY family consistently had very low expression levels, even though its median copy number per individual was two.
Next, we tested whether copy number, as measured for each individual, is positively correlated with gene expression levels, again measured for each individual, within the same gene family. There was no significant correlation in any of the nine families studied (all P-values were above the Bonferroni-corrected P-value cutoff of 0.05/9 = 0.006; S3 Fig; S7 Table). To control for genetic variation on the Y, we next compared copy number estimates to gene expression levels for individuals with the same Y subhaplogroup. We focused on the European R1b and I1a, and the African E1b subhaplogroups because they had more than 10 individuals in our dataset (77, 15 and 22, respectively; Table 2). We still found no significant correlations between copy number and expression levels in any of the nine gene families for individuals from either of these three subhaplogroups (all P-values were above the Bonferroni-corrected P-value cutoff of 0.05/9 = 0.006; S4–S6 Figs; S7 Table).
We further asked whether the major Y haplogroup could at least in part explain the variation we observed in copy number and in gene expression levels of Y chromosome ampliconic genes. We focused our analysis on major haplogroups R (European), I (European), E (African), and J (Western Asian) because they were represented by at least 10 samples in our dataset (Table 2). Using one-way ANOVA, we found that the copy numbers of BPY2 (P = 2.34x10-3), RBMY (P = 2.97x10-8), and TSPY (P = 1.07x10-22) gene families had significant differences among the four major Y haplogroups analyzed (Bonferroni-corrected P-value cutoff of 0.05/9 = 0.006; Table 3). The remaining six gene families did not display significant differences among Y haplogroups (Table 3). When we compared the mean copy number differences between haplogroups in a pairwise fashion using a permutation test (1 million permutations; 9 gene families are tested for 6 cases—R vs E; R vs I; R vs J; I vs E, I vs J, E vs J—thus we performed 9 x 6 = 54 tests; Bonferroni-corrected P-value cutoff of 0.05/54 = 0.00093), TSPY differed significantly in copy numbers (Fig 3) between major European (R and I) vs. African (E) or vs. Western Asian (J) haplogroups (P = 0 for R vs. E; P = 0 for I vs. E; P = 0 for R vs. J; P = 0.3x10-5 for I vs. J; S8 Table). RBMY copy numbers differed significantly between European (R) vs. African (E) or Western Asian (J) haplogroups (P = 6.94x10-4 for R vs. E; P = 0 for R vs J; S8 Table). No significant differences between the two major European haplogroups (R and I) were observed (S8 Table).
In contrast, we found that gene expression levels of all nine Y ampliconic gene families were not significantly different among major Y haplogroups (all P-values were above the P-value cutoff of 0.05/9≈0.006; one-way ANOVA; Table 3). We observed a trend suggesting differences in expression values among haplogroups for the BPY2 and DAZ gene families, but these differences were small in scale. Nevertheless, out of the nine gene families, BPY2 (P = 0.056) and DAZ (P = 0.01) had low P-values for the ANOVA analysis (Table 3, Fig 3) and for the permutation test comparing mean expression levels between haplogroups (P = 7.09x10-3 for E vs. R for BPY2; P = 1.36x10-2 for E vs. R for DAZ; P cutoff of 0.05/54 = 0.00093; S9 Table). When we compared the trend in copy number and gene expression differentiation among haplogroups, we observed that in the TSPY gene family both copy number and gene expression levels were lower for the European haplogroups (I, R) than for the African (E) or Western Asian (J) haplogroups (Fig 3). This trend was statistically significant for copy number, but not significant for gene expression. Analyzing a larger sample size might lead to finding this trend to be significant also for gene expression.
To examine the potential role of aging in determining Y ampliconic gene expression, we compared the ages of individuals at the time of sample collection to the ampliconic gene expression levels and found no statistically significant relationship (nine gene families were tested for correlation which results in Bonferroni correction P-value cutoff of 0.05/9 = 0.006; S7 Fig; S10 Table). Next, to perform a similar analysis for individuals with the same subhaplogroup, we limited our analysis to individuals with the European R1b and I1a, and African E1b subhaplogroups (77, 15, and 22 individuals, respectively). For the R1b and I1a subhaplogroups we found no significant relationship between age and expression levels for any of the nine Y ampliconic gene families studied (S8 and S9 Figs; S10 Table). However, for the African E1b subhaplogroup, HSFY (Spearman correlation = 0.57; P = 0.0061) and PRY (Spearman correlation = 0.61; P = 0.0028) gene families had a positive correlation between expression levels and age, which was significant after Bonferroni correction (S10 Fig; S10 Table). A larger dataset of African samples should be studied to validate this relationship.
The presence of homologs outside of the Y for two groups of Y ampliconic gene families allows us to study evolution of their gene expression levels [34]. In particular, the CDY and DAZ genes were copied to the Y chromosome from autosomes [34]; the HSFY, RBMY, TSPY, VCY, and XKRY gene families have homologs on the X and were likely present on the ancestral autosomes giving rise to the two sex chromosomes [34]. In the analyses below, we assume that the testis-specific expression of Y ampliconic genes was acquired prior to their amplification on the Y [9] and that their autosomal or X-chromosomal homologs have maintained ancestral expression levels, i.e. they possess expression levels of Y ampliconic genes prior to their Y linkage [35]. The latter assumption is based on the overall slower rates of evolution of X-chromosomal and autosomal genes as compared to their Y-chromosomal homologs.
We envision three possible scenarios for gene expression evolution of Y ampliconic gene families that have non-Y homologs (Fig 4). First, because of sexual antagonism, a gene on the Y could obtain beneficial mutations and diverge in function from its non-Y homolog to acquire new functions in testis (i.e. neo-functionalization). The expression of such a gene family would be independent from, and potentially higher than that for, its non-Y homologs (scenario A). Second, a gene family on the Y could retain function of the non-Y homolog, but acquire testis-specific expression (i.e. sub-functionalization). In this case, either the non-Y copy represents the ancestral expression levels and the Y copies are expected to maintain low expression levels, or the sum of expression from the Y and non-Y copies is regulated to be at levels similar to those of the non-Y copy in the ancestor (scenario B). In this case, the expression of both Y and non-Y homologs might be down-regulated. Third, genes on the Y might be under relaxed selective constraints and thus have low expression levels (scenario C) [36]. Below we test these three scenarios by comparing expression levels of both Y and non-Y ampliconic gene homologs in testis tissue.
In addition to the analysis of such overall differences in the expression level (Fig 4), we can also examine the relationship between the Y ampliconic genes’ and their non-Y homologs’ gene expression across individuals, which should further assist in determining a particular evolutionary scenario (S11 Fig). If the expression levels of Y ampliconic genes are higher than those of their non-Y homologs, and across individuals the expression levels of these two groups of genes are positively correlated, then this pattern is consistent with neo-functionalization of the Y ampliconic genes. This is because higher expression levels of ampliconic genes than those at the ancestral state suggest independent expression of Y ampliconic genes from their non-Y homologs, and a positive correlation between Y ampliconic genes and their non-Y homologs suggests co-regulation, e.g. they might share similar transcription factors [37]. A combination of these two patterns suggests an acquisition of a new function (scenario A) (S11A Fig). If the expression levels of Y ampliconic genes are higher than those of their non-Y homologs, and across individuals the expression levels of these two groups of genes are negatively correlated, then the data are compatible with neo- or sub-functionalization (scenario A or B). Indeed, the observed negative correlation could be explained by neo-functionalization, where ampliconic genes acquired a new function and inhibit the expression of the non-Y homologs. Alternatively, the negative correlation could be explained by sub-functionalization, where ampliconic genes acquired new transcription factors which limit their expression to a few cell types, and the negative correlation is due to the differences in the abundance of cell types in which ampliconic genes are expressed (S11B Fig). If the expression levels of Y ampliconic genes are lower than those of their non-Y homologs, and across individuals the expression levels of these two groups of genes are positively correlated, then this pattern is consistent with any of the three scenarios A-C. This is because the lower expression levels of Y ampliconic genes could be due to down-regulation of gene expression by the Y chromosome to accommodate the multi-copy state of ampliconic genes [38], evolution of which could still be compatible with any of the three scenarios A-C (S11C Fig). If the expression levels of the Y ampliconic genes are lower than those of their non-Y homologs, and across individuals the expression levels of these two groups of genes are negatively correlated, then the data are compatible with scenario A or B. This is because negative correlation eliminates the scenario of relaxed selection, i.e. scenario C (S11D Fig). Finally, if we observe no correlation in expression levels between Y ampliconic genes and their non-Y homologs, then we can conclude that their expression is independent from each other, which could be a result of neo-functionalization, sub-functionalization or random drift in expression levels under relaxed selection.
To test these scenarios, we first compared testis expression levels between Y ampliconic gene families CDY and DAZ, which were copied to the Y from autosomes, and their autosomal homologs (Fig 5). The CDY autosomal homologs CDYL and CDYL2 are ubiquitously expressed; and the DAZ autosomal homolog DAZL has testis-specific expression [34,39–41]. The expression levels of CDY (the sum of expression levels for the whole gene family) were 89% lower than those for their autosomal homologs (the sum of expression of CDYL and CDYL2), and for DAZ they were 63% lower than those for their autosomal homolog DAZL (Fig 5). Next, we tested whether the expression levels for Y ampliconic genes and their autosomal homologs are regulated at the level of each individual. For each gene family, we examined a potential correlation in gene expression levels between the Y ampliconic genes and their non-Y homologs. We observed a significant negative correlation between CDY and CDYL+CDYL2 expression levels (Spearman correlation = -0.31; P = 2x10-4), which indicates that, across individuals, whenever the CDY expression levels increase, the CDYL+CDYL2 expression levels decrease (Fig 6). In case of DAZ, a positive correlation in expression levels (Spearman correlation = 0.57; P = 0) was observed between DAZ and its autosomal homolog DAZL (Fig 6). Lower expression of CDY and DAZ than their non-Y homologs could be a result of down-regulation of gene expression by Y chromosome to maintain the multi-copy state, however the negative correlation in CDY vs. CDYL+CDYL2 expression levels indicates the presence of either neo- or sub-functionalization. DAZ could have undergone any of the three scenarios, which are difficult to differentiate based on the available data.
We next examined how testis-specific gene expression of the HSFY, RBMY, TSPY, VCY, and XKRY gene families diverged from that of their X homologs. Most of the X homologs of ampliconic genes (except for VCY and XKRY) are expressed in multiple tissues along with testis. The XKRX gene, an X homolog of the XKRY gene family, is not expressed in testis and we omitted this gene family from our analysis (S11 Table). Three Y gene families studied (RBMY, TSPY, and VCY) on average had lower expression levels in comparison to their X homologs (66%, 75%, and 71% lower, respectively; Fig 5). HSFY was the only gene family that on average had higher expression in comparison to their homologs on the X (35% higher than X-homologs). This could imply that HSFY might have acquired a new function, which is selected for in testis (scenario A). At the level of the studied individuals, all four studied gene families exhibited positive correlation in gene expression levels between their Y ampliconic and X homolog genes, suggesting a potential co-regulation (Fig 6). This correlation was particularly strong for the HSFY and VCY gene families (Spearman correlation of 0.69 and 0.84, respectively). The observed higher expression of HSFY than of its X homologs, as well as positive correlation in gene expression levels between these two groups of genes, is a strong indicator of neo-functionalization. In the case of RBMY, TSPY, and VCY, it is challenging to differentiate among the three scenarios we propose based on the available data.
Ampliconic genes constitute the majority (80%) of protein-coding genes present on the human Y chromosome and play an important role in spermatogenesis [1]. Yet, very little is known about the significance of Y ampliconic gene copy number variation in determining their expression levels in humans. Here we analyzed both copy number and testis-specific expression of ampliconic gene families in 149 presumably healthy men. Our goal was to understand the relationship between copy number variation and expression levels while accounting for Y chromosome haplogroups.
Our results indicate that smaller Y ampliconic gene families maintain lower variation in copy number and, as the size of gene families increases, variation in copy number also increases, in agreement with previous studies [3–5]. The parsimonious explanation for this observation is that a greater number of gene copies leads to loss or gain of gene copies because of a higher probability of rearrangements via replication slippage and/or non-allelic homologous recombination (NAHR) [42–44]. On the human Y, the larger gene families are either spread across multiple palindromes (e.g., RBMY) or are arranged as a tandem array (TSPY), and such arrangements can result in multiple scenarios of NAHR, which will lead to gain or loss of gene copies. BPY2 has two functional copies on palindrome P1 and one copy outside of palindromes, and such an arrangement can also result in NAHR.
We found that the large TSPY and RBMY gene families have not only a high level of variation in copy number, but also a significantly different number of gene copies among the major Y haplogroups analyzed. An earlier study also found significant differences in copy number for these two gene families among human Y haplogroups across the world and suggested that this observation cannot be explained by selection [3]. However, selection explanation might warrant a further investigation. Indeed, a recent molecular analysis of infertile men indicated a positive correlation between the number of RBMY copies and sperm count and motility [45]. Moreover, RBMY is a male-specific oncogene [46]. Therefore, it will be of interest to investigate whether variation in RBMY copy number across Y haplogroups influences these two disease-related phenotypes and might be subject to natural selection. Similarly, TSPY is a candidate proto-oncogene which can regulate its own expression via a positive feedback loop in gonadoblastoma and a variety of somatic cancers [47]. Thus, additional studies should be performed to test whether variation in TSPY copy number across haplogroups is associated with differential predisposition to gonadoblastoma.
The smaller Y ampliconic gene families (HSFY, PRY, VCY, and XKRY) have lower variation in copy number compared with larger families. These gene families, for which the average family size is only two copies, are each present on an individual palindrome (the two copies are present as inverted repeats on opposite palindrome arms). Recombination between inverted repeats is expected to result in an inversion keeping copy number constant [48]. In addition, the presence of only two copies increases the chances of a complete gene family elimination due to Muller’s ratchet or of rearrangements which involve the whole palindrome. Consistent with this prediction, we find these gene families to be deleted or pseudogenized in several great ape species [33].
Thus, the copy number of ampliconic genes is an important factor in determining the survival of a gene family on the Y chromosome. Too few copies can lead to a complete loss of a gene family (see the preceding paragraph), whereas too many copies can lead to frequent NAHR which can rapidly increase or decrease copy number [49]. Consistent with this expectation, it was suggested that the human Y chromosome evolves under selection to maintain an optimal copy number for its amplicons in diverse human lineages [18]. Most likely both random genetic drift and natural selection contribute to determining the Y chromosome ampliconic gene copy number. Drift leads to smaller-scale changes in copy numbers, whereas selection might act at removing extreme copy numbers because too few copies might lead to infertility and too many copies might lead to genetic instability and thus both are selected against. Variation in Y ampliconic gene copy number in subfertile and infertile males should be investigated in future studies and should shed additional light on the balance between these two evolutionary processes.
Note that in the present study we only examined complete gene copy gains or losses, but insertions and deletions inside a gene can also affect gene expression and functionality, and might be linked to infertility [50]. The effects of such smaller CNVs are more robustly evaluated from long-read data and we leave this exploration to future work.
Here we studied the expression levels of the Y ampliconic gene families in testis tissue of presumably healthy individuals. The vast majority of cells in testis are germline cells in the seminiferous ducts, where spermatogenesis takes place. We primarily captured Y chromosome gene expression in spermatogonia prior to meiosis and throughout different spermatogenesis stages after meiosis [51,52]; this is because Y transcription is silenced at other stages of spermatogenesis due to meiotic sex chromosome inactivation [52,53] and postmeiotic sex chromosome repression [51,52]. As a tissue, testis is a mixture of germline cells at different stages of development, Sertoli cells, myoid peritubular cells, and interstitial Leydig cells. Thus, the expression values generated using testis tissue as a source represent cumulative gene expression of germline cells at different stages of spermatogenesis with a mixture of somatic cells. This potential limitation notwithstanding, our results indicate substantial variation in expression levels for Y ampliconic genes in testis among men and suggest that different levels of Y ampliconic genes’ expression are tolerated by presumably healthy individuals.
When we compared copy number of ampliconic genes to their gene expression values, we found that across gene families the gene families with higher median copy number had higher expression levels. This is consistent with an observation made by Lucotte and colleagues [5] who reported on the expression of Y ampliconic genes at different stages of spermatogenesis with respect to variation in their copy number. Overall, the Y chromosome has higher copy number of genes for those gene families whose median expression levels are higher in testis, however it is important to note that this relationship might be different at individual cell types in testis and should be studied further.
When we examined the relationship between copy number and expression within a gene family, our analysis revealed that expression of Y ampliconic gene families is independent of their copy number. Moreover, no significant differences in Y ampliconic gene expression levels were observed among Y haplogroups, even though we found significant differences among Y haplogroups in copy number for some gene families (BPY2, TSPY, and RBMY). This suggests that testis tissue might have evolved the ability to tolerate different Y ampliconic gene copy numbers, and also variable Y ampliconic gene expression levels.
Approximately 77% of all protein-coding genes in the human genome are expressed in testis [54], and some of these genes could regulate expression of the Y ampliconic genes. Understanding the 3D organization and chromatin structure on the Y is expected to aid in identifying the genomic regions and genes that ampliconic genes interact with and are regulated by in the genome. Future studies analyzing expression data at different stages of spermatogenesis in individuals with different Y ampliconic gene copy numbers will assist in deciphering the role of copy number variation in determining gene expression in more detail. Additionally, our findings should be confirmed by studies of gene expression at the protein level.
A man’s advanced age has significant negative impact on reproduction [55]. Semen parameters such as daily sperm production, total sperm count, and sperm viability are negatively correlated with age [56]. However, within our dataset, we observed mixed results regarding age effects on Y ampliconic gene expression: age did not influence variation in gene expression of these genes in individuals with European Y subhaplogroups I1a and R1b, however HSFY and PRY expression had a positive correlation with age in individuals with an African subhaplogroup E1b. These findings should be validated with a larger data set to examine the role of Y ampliconic genes in changes in spermatogenesis with age.
The Y chromosome degradation, which is common across eutherian mammals, has resulted in the loss of the majority of genes originally present on the proto-sex autosomal pair [57]. To balance the loss of genes on the Y in males, the mammalian X chromosome adapted its expression levels by inactivating one of its copies and increasing the expression of the other copy in females [57–59]. We wondered whether a similar process evolved at Y ampliconic genes that have non-Y homologs, namely whether the expression of Y ampliconic genes and their non-Y homologs has been co-regulated. Alternatively, Y ampliconic genes might have evolved new functions, and thus potentially high expression levels, independent of their non-Y homologs. Yet another alternative would be the overall low expression levels because of the relaxation of functional constraints on the Y ampliconic genes. The precise functions of Y ampliconic genes have been under-characterized (S12 Table) due to the repeated nature of the Y chromosome and scarcity of testable orthologs in model organisms. While Y ampliconic genes have testis-specific expression likely as a result of sexual antagonism, the majority of non-Y homologs of Y ampliconic genes have ubiquitous expression.
Recently, a multi-step model for preservation of tandem duplicate genes was presented. According to this model, the expression of gene duplicates is down-regulated immediately after the duplication event, followed by dosage sharing which could lead to functional adaptations such as sub- or neo-functionalization [38]. Knowing that non-Y homologs of Y ampliconic genes are expressed in testis (except for XKRX), we compared the expression levels of closely related homologs of ampliconic genes on both autosomes and X chromosome to the sum of expression levels for all the copies of a Y gene family. We demonstrated that, with the exception of the HSFY family, Y ampliconic gene families have consistently lower expression levels when compared to their non-Y homologs, thus not elevating the overall expression level of the family. We term this phenomenon dosage regulation of Y ampliconic genes. Lower expression of Y ampliconic gene families could be an adaptation of the Y to maintain the multi-copy state of ampliconic gene families. By lowering the expression of the whole gene family, the Y can buffer sudden loss or gain of gene copies. In addition to dosage regulation, the gene family should be expressed at optimal levels to maintain their functionality during spermatogenesis. Lower optimal expression of Y ampliconic gene families compared to their non-Y homologs could be a result of sub-functionalization (e.g., testis specificity in expression), which benefits germline cell development. Alternatively, such low expression could be a result of relaxed selection, and, in agreement with this possibility, Y ampliconic genes show a higher rate of nonsynonymous to synonymous substitution rates compared to single-copy X degenerate genes on the Y [7]. Alternatively, a gene family could be under positive selection or undergoing neo-functionalization even in their low-level expression state. The expression of ampliconic gene families is important for spermatogenesis because of an association between gene deletions and infertility, but relaxed selection can facilitate rapid differentiation of ampliconic gene function.
We found that expression levels of the CDY ampliconic genes and those of their autosomal homologs are negatively correlated among individual men. This suggests that the CDY gene family might not be expressed at the same time during spermatogenesis as its autosomal homologs or that there is a coordinated down-regulation of CDY expression with a rise in CDYL and CDYL2 expression (and vice versa). In humans, the CDYL and CDYL2 autosomal genes produce the ubiquitously expressed long transcripts, but lost the testis-specific short transcript which is now produced by CDY [40]. The combined tissue expression patterns of CDY, CDYL, and CDYL2 in human recapitulate the expression patterns of CDYL and CDYL2 in mouse or rabbit, which do not have CDY on their Y chromosome [40].
In contrast with CDY, we found that expression levels of DAZ, HSFY, and VCY gene families are strongly positively correlated with their non-Y homolog expression across individuals, which suggests a co-regulation in gene expression levels of these ampliconic gene families and their homologs (the RBMY and TSPY families also show positive correlation, however it is not strong). When we examine the linear relationship between ampliconic gene families and their homologs among individual men, the Y ampliconic gene expression increases at a slower pace when compared to the expression of their non-Y homologs, except for HSFY where the expression increases at a similar rate for both Y and non-Y homologs (Fig 6).
The VCY gene family is the most commonly lost gene family among great apes, however in our dataset the expression of this gene family is higher than for most other gene families on the Y and is higher than is predicted from its copy number (Fig 2). The homologs of VCY on the X chromosome (VCX, VCX2, VCX3A, and VCX3B) are expressed in testis [60,61]—and we show that at higher levels than the VCY family itself. In addition, there is high sequence identity (>95%) between the VCX and VCY gene families, which could imply that both VCX and VCY could have been under selection to maintain function of the gene family, however, to balance the expression of the multi-copy VCX family, VCY might have lowered its expression. The role of both VCX and VCY in ribosome assembly in spermatogenesis has been suggested [62]. The loss of VCY in great ape species might have been compensated by functionally similar VCX family expression in testis. The expression levels of the VCX family across great apes must be studied to understand its role in the loss of VCY.
A recent study found multiple distinct clusters of full-length Y ampliconic gene transcripts, likely originating from different copies of the same family [63]. Therefore, the presence of multiple full-length transcripts [63] and low expression levels for Y ampliconic gene families (the present study) suggest that individual gene copies within a family are down-regulated to accommodate the expression of the whole gene family on the Y chromosome and outside of it (on autosomes and on the X). This hypothesis needs to be examined in future studies in which expression levels of individual gene copies will be evaluated with long-read sequencing technology. It will also be important to decipher the isoforms and their expression levels for Y ampliconic genes and their non-Y homologs to understand whether Y ampliconic genes and their homologs express the same isoforms, or whether Y ampliconic genes express their own, unique, testis-specific isoforms.
It is essential to note that, in addition to evolution of expression levels of the whole gene family including its non-Y homologs, the Y ampliconic genes can diverge to acquire additional male-specific functionality because they are present on the Y, which is susceptible to accumulating genetic differences dictated by sexual antagonism. In other words, Y ampliconic genes could have gained secondary functions independent of their functions on the proto-sex chromosomes. This scenario might be exemplified by the case of the HSFY family, whose expression levels have increased in comparison to its X-chromosome homologs. This pattern suggests that this gene family underwent neo-functionalization. The exact function of HSFY is unknown, but its role in transcription regulation has been suggested because it harbors a DNA-binding domain [64]. In fact, it was shown that HSFY and HSFX share only this DNA-binding domain but not the rest of their sequences and thus indeed might have diverged in their functions [64]. Moreover, HSFY has stage-specific expression during spermatogenesis, suggesting that it acquired a function different than that of heat shock proteins it is homologous to [64]. The loss of HSFY was linked to infertility [64–66]. In another study, under-expression of HSFY was linked to arrest of maturation of nascent germ cells to motile sperm [67]. According to our study, the expression of HSFY gene family was positively correlated with age in the African E1b Y haplogroup, however such a relationship was not found in the R1b haplogroup. Further studies addressing transcription regulation by the HSFY family in individuals of varying age across different Y haplogroups are required to understand the HSFY functionality in more detail.
We assume that non-Y homologs have retained the ancestral expression state because of the overall fast evolutionary rate on the Y chromosome [68]. However, the X chromosome and autosomes have also been evolving, albeit slower than the Y. Evolutionary changes acquired by non-Y homologs since they diverged from the Y homologs have not been addressed in this study due to the lack of ancestral expression data. To address this, future studies should identify species which have orthologs of human ampliconic genes in a single-copy state on their Y chromosome. In the case of CDY and DAZ gene families, future studies should identify species in which these genes’ orthologs are present in a single-copy state on their autosomes and absent from the sex chromosomes. Once such species are identified, their testis-specific expression data for these genes could be used as the ancestral expression state.
To estimate copy number in highly-similar multi-copy gene families, several strategies have been proposed. One can align each read to all possible locations in the reference genome [69], identify sites in the reference genome that uniquely distinguish and tag paralogs of interest [18,21,27,29], use simulated reads for mock genomes with human gene cDNAs at different gene copy counts to obtain a theoretical function of the coverage distribution with respect to copy number [28], or customize the reference to keep a single copy of each gene family [4,5]. While these strategies were effective in their respective papers, we could not find software that could work on human Y ampliconic genes. We therefore combine the ideas from these strategies into AmpliCoNE, a tool for estimating copy number in highly-similar multi-copy gene families. The Results section contains an overview of AmpliCoNE, but we provide more details here.
To evaluate the accuracy of AmpliCoNE, we ran simulations. There are nine TSPY genes (six functional + three pseudogenes), six RBMY genes and two VCY genes in the hg38 reference. We added different copies of these three ampliconic gene families to the Y chromosome (S1 Table) to simulate reads. The total number of gene copies in the three custom references used to generate the simulated reads were 22/7/4 copies (for TSPY/RBMY/VCY, respectively) in set 1; 29/12/2 copies in set 2; 23/9/3 copies in set 3. Using wgsim [0.3.2] [76] we simulated 666 million paired-end reads of length 101 bp and insert size of 260 bp (the exact parameters were "-d 260 -N 666873346–1 101–2 101 -S 9 -e 0 -r 0 -R 0"). The reads from the three simulated datasets were aligned to the hg38 reference genome using BWA MEM[v0.7.15] [77]. The SAM files were sorted and PCR duplicates were removed using the PICARD toolkit [v1.128] [78]. Finally, samtools [v1.3.1] [79] were used to index the alignments. The sorted indexed BAM files were presented as input to AmpliCoNE-count.
We used mRNA sequencing data for 170 testis samples with matched whole-genome sequencing (WGS) data from the GTEx project [25]. The GTEx RNA-seq libraries were generated with the Illumina TruSeq protocol and whole-genome sequencing was performed with paired-end reads ranging from 100 bp to 150 bp in length with target insert size of 350–370 bp [25]. As the validation dataset for AmpliCoNE, we used WGS data from four males (depth of coverage ranging from ~45-50x in HG002 and HG003, ~300x in HG005 and ~100x in HG006) sequenced by the GIAB consortium [30].
The Y-chromosome-specific alignments of the GTEx dataset were extracted from dbGAP using the SRA toolkit [80]. From the alignments, we extracted the reads and aligned them to the hg38 reference genome using bwa-mem [v0.7.15] [77]. The SAM files were sorted and PCR duplicates were removed using PICARD toolkit [v1.128] [78]. Finally, samtools [v1.3.1] [79] were used to index the read alignment files. The generated BAM files were presented as input to AmpliCoNE-count to estimate ampliconic copy number.
AMpliCoNE-build requires the locations of all the gene copies, in the reference genome, for each ampliconic gene family. While the locations of functional copies are already annotated in hg38, these do not include highly similar pseudogenized copies. These are necessary to include since they will affect the read mappings. For each family, we therefore took an arbitrary annotated copy of a gene, and used BLAT [81] to find all sites aligning with >99% identity (S13 Table). These locations were given as input to AmpliCoNE-build.
In order to validate the in silico ampliconic gene copy number count in four individuals sequenced by the GIAB consortium [30], we acquired their DNA (NA24385, NA24149, NA24631, and NA24694) from Coriell and performed ddPCR for all nine Y ampliconic gene families. In order to infer the copy number of these gene families we used SRY, a single-copy gene on the Y chromosome, and RPP30, a two-copy autosomal gene, as references. We ran ddPCR for each sample in triplicates using EvaGreen dsDNA dye (Bio-Rad) on the Biorad QX200 digital droplet platform with the protocol and primers from our previous publication [82]. The results were analyzed using QuantaSoft software. Subsequently, after removal of outliers, the concentration (copies/uL) of each ampliconic gene family of interest was divided by the concentration of the references, SRY and RPP30 (S4 Table).
Gene expression estimates were obtained using the kallisto-DESeq2 pipeline described below. The standard human (hg38) RefSeq transcripts obtained from the UCSC Genome Browser [83] were used as reference. We generated an index for the reference using the kallisto [v0.43.0] index function with default parameters [84]. For each sample we obtained read counts per transcript using the kallisto quant (—bias,—seed = 9,—bootstrap-samples = 100) function. The hg38 refFlat file containing the transcript-to-gene mapping information was obtained from the UCSC Genome Browser [83] annotation database, which was used to convert the transcript-level read counts to gene-level expression levels using tximport package [v1.2.0] [85]. Since there were no replicates for the samples, we set the 170 sample ids as different conditions in the design, and the gene-level read counts for 170 RNA-seq samples were normalized using DESeq2 [v1.14.1] [86]. Additionally, read counts based on the vst (Variance Stabilizing Transformation) function in DESeq2 were used to check for outliers. To identify outliers in the dataset we performed Principal Component Analysis using the prcomp() function on the vst-based normalized read counts. When we plotted the first and second principal components, we found 21 samples outside the main cluster of the remaining 149 samples (S12 Fig). We followed steps described in DEseq2 vignettes and plotted the heatmap of sample-to-sample distance for the top 1,000 highly expressed genes to identify outliers visually and we found the same 21 samples as outliers. Thus, we filtered out these 21 samples and utilized the expression values for the nine ampliconic families in the remaining 149 samples in the downstream analysis. We summed the expression values for all the gene copies within a gene family to obtain family-level expression values.
Yhaplo [v1.0.11] [87] was used to predict Y haplogroup of the samples. The version of Yhaplo[1.0.11] we used expects the SNP coordinates consistent with the hg19 [88] version of the human reference. The Y-chromosome-specific BAM files downloaded from dbGAP were aligned to the hg19 version of the human reference using BWA MEM. We directly converted the downloaded BAM files into pileup format using samtools mpileup function. A custom script was used to convert the pileup file into Yhaplo-compatible input format. We annotated the Y haplotype for all the samples in the dataset using Yhaplo default parameters.
Code used in the manuscript is available at github link: https://github.com/makovalab-psu/GTEx_Testis_Analysis. Steps to install and use AmpliCoNE are available at github: https://github.com/makovalab-psu/AmpliCoNE-tool
|
10.1371/journal.pgen.1005545 | The Type VI Secretion TssEFGK-VgrG Phage-Like Baseplate Is Recruited to the TssJLM Membrane Complex via Multiple Contacts and Serves As Assembly Platform for Tail Tube/Sheath Polymerization | The Type VI secretion system (T6SS) is a widespread weapon dedicated to the delivery of toxin proteins into eukaryotic and prokaryotic cells. The 13 T6SS subunits assemble a cytoplasmic contractile structure anchored to the cell envelope by a membrane-spanning complex. This structure is evolutionarily, structurally and functionally related to the tail of contractile bacteriophages. In bacteriophages, the tail assembles onto a protein complex, referred to as the baseplate, that not only serves as a platform during assembly of the tube and sheath, but also triggers the contraction of the sheath. Although progress has been made in understanding T6SS assembly and function, the composition of the T6SS baseplate remains mostly unknown. Here, we report that six T6SS proteins–TssA, TssE, TssF, TssG, TssK and VgrG–are required for proper assembly of the T6SS tail tube, and a complex between VgrG, TssE,-F and-G could be isolated. In addition, we demonstrate that TssF and TssG share limited sequence homologies with known phage components, and we report the interaction network between these subunits and other baseplate and tail components. In agreement with the baseplate being the assembly platform for the tail, fluorescence microscopy analyses of functional GFP-TssF and TssK-GFP fusion proteins show that these proteins assemble stable and static clusters on which the sheath polymerizes. Finally, we show that recruitment of the baseplate to the apparatus requires initial positioning of the membrane complex and contacts between TssG and the inner membrane TssM protein.
| In the environment, bacteria compete for privileged access to nutrients or to a particular niche. Bacteria have therefore evolved mechanisms to eliminate competitors. Among them, the Type VI secretion system (T6SS) is a contractile machine functionally comparable to a crossbow: an inner tube is wrapped by a contractile structure. Upon contraction of this outer sheath, the inner tube is propelled towards the target cell and delivers anti-bacterial effectors. The tubular structure assembles on a protein complex called the baseplate. Here we define the composition of the baseplate, demonstrating that it is composed of five subunits: TssE, TssF, TssG, TssK and VgrG. We further detail the role of the TssF and TssG proteins by defining their localizations and identifying their partners. We show that, in addition to TssE and VgrG that have been shown to share homologies with the bacteriophage gp25 and gp27-gp5 proteins, the TssF and TssG proteins also have homologies with bacteriophage components. Finally, we show that this baseplate is recruited to the TssJLM membrane complex prior to the assembly of the contractile tail structure. This study allows a better understanding of the early events of the assembly pathway of this molecular weapon.
| In the environment, bacteria endure an intense warfare. Bacteria collaborate or compete to acquire nutrients or to efficiently colonize a niche. The outcome of inter-bacteria interactions depends on several mechanisms including cooperative behaviors or antagonistic activities [1]. The newly identified Type VI secretion system (T6SS) is widely distributed among proteobacteria and has been reported to be a key player in antagonism among bacterial communities [2–4]. Although several T6SSs have been shown to be required for full virulence towards different eukaryotic cells, most T6SSs shape bacterial communities through inter-bacteria interactions [1]. In both cases, T6SSs inject toxic effectors into target/recipient cells. A number of anti-bacterial toxins have been recently identified and carry a versatile repertoire of cytotoxic activities such as peptidoglycan hydrolases, phospholipases or DNases [1,5,6]. Delivery of these toxins into the periplasm or cytoplasm of the target cell leads to a rapid lysis that usually occurs within minutes [7–9].
At a molecular level, the T6SS core apparatus is composed of 13 conserved subunits that assemble a long cytoplasmic tubular structure tethered to the cell envelope by a trans-envelope complex [3,10–12]. The composition, structure and biogenesis of the membrane-associated complex has been well characterized over the last years. It is composed of three proteins: TssL, TssM and TssJ. The TssL and TssM proteins interact in the inner membrane whereas the periplasmic domain of TssM contacts the TssJ outer membrane lipoprotein [13–15]. The current model considers the cytosolic complex of the T6SS to be similar to tails of contractile bacteriophages. These two related structures feature a cell-puncturing syringe and a contractile sheath wrapping an inner tube. The T6SS inner tube is composed of Hcp hexamers stacked on each other [16–18]. The cell-puncturing syringe assembles from a trimer of the VgrG protein tipped by the PAAR protein and is thought to cap the Hcp tube [16,19]. This structure is structurally comparable to the tail tube composed of polymerized gp19 proteins capped by the gp27-gp5 complex–or hub–in the bacteriophage T4 [20]. The TssB and TssC proteins share structural and functional similarities with the bacteriophage T4 gp18 sheath [8,21–25]. Indeed, time-lapse fluorescence microscopy experiments using a TssB-GFP fusion revealed that these structures are highly dynamic: they assemble micrometer-long tubes that sequentially extend in tens of seconds and contract in a few milliseconds [8,9,26,27]. The mechanism of contraction is thought to be similar to that of contractile bacteriophages [22–25]. Recent fluorescence microscopy assays in mixed culture evidenced that contraction of this sheath-like structure correlates with prey killing [7–9]. Based on these data and on the mechanism of bacteriophage infection, the current model proposes that the contraction of sheath-like structure propels the Hcp inner tube towards the target cell, resulting in the cell envelope puncturing and delivery of anti-bacterial toxins [3,10,28]. In tailed phages, tube and sheath polymerize on a structure called the baseplate. The bacteriophage T4 baseplate is composed of 140 polypeptide chains of at least 16 different proteins. This highly complex structure assembles from 6 wedges surrounding the central hub. Seven proteins form the baseplate wedges (gp11, gp10, gp7, gp8, gp6, gp53 and gp25) [22,29–31]. However, in other tailed bacteriophages such as P2, the baseplate is significantly less complex as it is only composed of four different subunits: gpV (the homologue of the hub) and the wedge components W (the homologue of gp25), gpJ (gp6-like) and gpI [32,33]. Based on this observation, Leiman & Shneider formulated the concept of a minimal contractile tail-like structure [22]. In the minimal contractile tail, the baseplate could be significantly “simplified” as long as it performs its main functions: controlling tube assembly, initiating sheath polymerization and triggering sheath contraction. The minimal baseplate should then conserve the central hub and three other wedge proteins: gp6, gp25 and gp53 [22]. The central hub bears the spike and acts as a threefold to sixfold adaptor connecting the tail tube. Gp25 initiates the polymerization of the sheath. Gp6 connects the wedges together maintaining the baseplate integrity during the infection process. The role of gp53 remains unclear. However, gp53 is required for gp25 to assemble onto the gp6 ring [22]. In the T6SS, the assembly of the tail-like tube and sheath must require components that will perform similar functions. With the exceptions of TssE and VgrG, which feature striking homologies to the gp25 protein and the gp27-gp5 hub complex respectively, the components that assemble the T6SS baseplate are unknown. By analogy with the morphogenesis pathway of contractile bacteriophages, we hypothesized that the assembly of the tail tube should be impaired in absence of a functional baseplate. We therefore recently developed a biochemical approach based on inter-molecular disulfide bonds to probe the assembly of the Hcp tube in vivo, in the cytoplasm of enteroaggregative Escherichia coli (EAEC) [18]. We demonstrated that Hcp hexamers stacked in a head-to-tail manner to form bona fide tubular structures in vivo [18]. More importantly, the precise stacking organization of Hcp hexamers became uncontrolled in absence of the other T6SS components. Here, using the collection of nonpolar T6SS gene deletions we provide evidence that six T6SS proteins are required for proper assembly of Hcp tubes: TssA, TssE, TssF, TssG, TssK and VgrG. The identification of TssE and VgrG, two known homologues of bacteriophage baseplate components, validates the experimental approach. We further characterize the TssF and TssG proteins. We report that these two proteins interact and stabilize each other, and make contacts with TssE, TssK and VgrG as well as with tube and sheath components. A bioinformatic analysis suggests that TssF and TssG share similarities with the J and I proteins of the bacteriophage P2 baseplate respectively. Fluorescence microscopy experiments further show that functional GFP-TssF (sfGFPTssF) and TssK-GFP (TssKsfGFP) proteins assemble into static foci near the cell envelope. The integrity of the sfGFPTssF foci is dependent on TssK and its proper localization requires interactions between TssF, TssG and the cytoplasmic loop of TssM. Futhermore, co-localization experiments with mCherry-labeled TssB demonstrate that sfGFPTssF clusters are positioned prior to sheath subunits recruitment and remain at the base of the sheath during elongation and contraction. Taken together the biochemical and cytological approaches presented in this study provide support to the role of TssE, TssF, TssG, TssK and VgrG as T6SS baseplate components and to a sequential recruitment hierarchy (membrane complex, baseplate, tail tube/sheath) during T6SS biogenesis.
Using an in vivo inter-molecular cross-linking approach, based on disulfide bond formation between adjacent cysteine residues, we recently reported that the EAEC Hcp hexamers organize head-to-tail to form tubular structures in the cytoplasm of EAEC. Importantly, we also demonstrated that these hexamers stack randomly in a strain lacking the Sci-1 T6SS subunits, resulting in head-to-tail, tail-to-tail and head-to-head configurations ([18], Fig 1A). During the morphogenesis of tailed bacteriophages, the tail-tube and-sheath structures polymerize on an assembly platform referred to as baseplate [22]. Additionally, the gp19 tail tube of bacteriophage T4 does not polymerize in absence of a fully functional baseplate [34]. By analogy, we reasoned that the aberrant assembly of Hcp hexamers in vivo could report a defective baseplate-like structure in the T6SS. We therefore probed the assembly of Hcp in each nonpolar Δtss strain (each lacking an essential Tss subunit) using the disulfide bond assay. As proof of concept, we previously showed that the Sci-1 T6SS-associated spike protein, VgrG, is required for proper assembly of Hcp tubes [18]. Aside the cysteine-less Hcp C38S protein, three combinations were tested to probe head-to-tail (G96C-S158C), tail-to-tail (Q24C-A95C) or head-to-head (G48C) stacking. As previously reported, SDS-PAGE analyses of cytoplasmic extracts from Δhcp oxidized cells producing these variants showed that the head-to-tail G96C-S158C combination leads to formation of dimers and higher molecular weight complexes while the tail-to-tail Q24C-A95C and head-to-head G48C combinations remain strictly monomeric (Fig 1A). Similar results were obtained for the tssB, tssC and clpV backgrounds, suggesting that tail sheath components do not regulate tail tube assembly (Fig 1B). This is in agreement with the morphogenesis pathway of contractile bacteriophages in which tail tube polymerization immediately precedes that of the sheath. Importantly, we also observed that Hcp hexamers properly assemble in the cytoplasm of tssJ, tssL and tssM mutants (Fig 1B). TssJ, TssL and TssM interact to form the trans-envelope complex that anchors the T6SS tail-like structure to the membranes [12–15]. Proper assembly of the tail tube structure is therefore independent of the membrane complex, a result in agreement with the different evolutionarily history of the T6SS membrane and phage complexes [2,35–37]. However, the controlled assembly of Hcp tubes was impaired in the vgrG, tssA, tssE, tssF, tssG and tssK backgrounds: Hcp hexamers interact in head-to-tail, tail-to-tail or head-to-head packing (Fig 1B). Proper Hcp assembly was restored in these different mutant strains when a wild-type allele of the missing gene was expressed from complementation vectors (Fig 1C). From these results, we concluded that six T6SS components, i.e. VgrG, TssA, TssE, TssF, TssG and TssK, increase the efficiency of tube formation in vivo and therefore control the assembly of Hcp tubes. TssE and VgrG share structural homologies with the gp25 protein and the gp27-gp5 complex (hub) respectively [16,19,35,38]. During the morphogenesis of the bacteriophage T4, the tail tube assembly initiates onto the baseplate only after the (gp27-gp5) hub complex has been recruited, while the gp25 subunit is required for functional baseplate assembly [39]. The observation that Hcp assembly was impaired in vgrG and tssE cells is therefore in agreement with the bacteriophage assembly pathway and further validates the initial hypothesis that Hcp tube proper polymerization depends on a baseplate-like structure. Based on these observations, we hypothesized that TssA, TssF, TssG and TssK may form, along with TssE and VgrG, a platform similar to the bacteriophage baseplate. However, while the TssE, TssK and VgrG subunits have been previously characterized [16,38,40], little information on TssA, TssF and TssG is available. The bioinformatics study published by Boyer et al. demonstrated a high level of co-occurrence between the tssE (COG3518), tssF (COG3519) and tssG (COG3520) genes. tssE and tssF are genetically linked in 87% of the T6SS gene clusters whereas the co-organization of tssF and tssG occurs in 97% of these clusters ([2], Fig 2A). As noted by Boyer and collaborators [2], co-occurrence usually reflects protein-protein interactions. Indeed, we show below that TssF and TssG are two components of the T6SS baseplate. By contrast, no co-occurrence of the tssA gene (COG3515) with tssE, vgrG or hcp was noticed in this study. Although TssA is required for Hcp tube formation, we will report elsewhere that it is not a component of the T6SS per se (Zoued, Durand et al., in preparation). We therefore focused our further work on the two uncharacterized tssF and tssG genes.
Bioinformatic analyses: TssF and TssG are homologues to phage tail proteins. To gain further insights onto TssF and TssG we performed a bioinformatic analysis using HHPred (homology detection and structure prediction, [41]). The Sci-1 EAEC TssF (accession number: EC042_4542; gene ID: 387609963) and TssG (accession number: EC042_4543; gene ID: 387609964) protein sequences were used as baits to identify homologues in bacteriophages. HHpred analyses with TssF reported that the fragment comprising residues 7–144 (over 587) resembles region 9–132 of the phage tail-like protein TIGR02243 (PFAM04865) that has for prototype the protein J of phage P2. The segment encompassing residues 86–179 shares also secondary structures with residues 89–178 of gp6, a baseplate wedge component of bacteriophage T4 (Fig 2B and 2C). Similarly, the TssG fragment comprising residues 37–152 (over 303) resembles region 4–110 of the phage tail-like protein TIGR02242 (PFAM09684) that has for prototype the protein I of phage P2 (Fig 2D). The phage P2 baseplate is composed of four subunits: V, W, I and J [32]. Protein V is an homologue of the bacteriophage T4 gp27-gp5 complex (VgrG) [42] whereas protein W is the homologue of gp25 (TssE). Leiman & Shneider recently hypothesized that the baseplate of a minimal contractile structure assembles from a central hub (gp27-gp5 and the protein V in the bacteriophages T4 and P2 respectively) and three key wedge proteins (gp25, gp6 and gp53) in the bacteriophage T4; W, I and J proteins in the bacteriophage P2) [22]. The predicted structural homologies suggest that the N-terminal region of TssF corresponds to the N-terminal region of gp6 whereas TssG (and phage P2 protein I) corresponds to gp53. We propose therefore that the baseplate of the T6SS is composed of at least four subunits (VgrG (gp27-gp5; V), TssE (gp25; W), TssF (gp6; J) and TssG (gp53; I).
T6SS function: TssF and TssG are required for sheath polymerization and Hcp release. A set of elegant studies coupling genetic and biochemical approaches to electron microscopy imaging demonstrated that during the morphogenesis of bacteriophage particles the absence of baseplate components prevents polymerization of the inner tube and of the outer sheath [31,39,43–45]. Here, we followed the dynamic of a chromosomally-encoded TssB-mCherry fusion protein (TssBmCh) using time-lapse fluorescence microscopy. We observed that sheath assembly is abolished in tssF and tssG cells (S1A Fig). In agreement with the absence of sheath polymerization and contraction, western blot analyses of culture supernatants showed that tssF and tssG cells do not release Hcp in the medium (S1B Fig).
Interaction network: TssF and TssG interact with TssE, VgrG, TssK, Hcp and TssC. To gain further information on TssF and TssG partners, we used a bacterial two-hybrid (BTH)-based systematic approach. T18-TssF/G and TssF/G-T18 translational fusions were tested against the phage-related T6SS core-components (TssB, TssC, Hcp, TssE, VgrG and TssA) fused to the T25 domain. As shown in Fig 3A, TssF interacts with Hcp, while TssG interacts with TssC, Hcp and TssE. These pair-wise interactions were then tested by co-immunoprecipitation in the heterologous host E. coli K-12. Fig 3B shows that HA-tagged Hcp was co-immunoprecipitated with FLAG-tagged TssF or TssG. Fig 3C shows that HA-tagged TssG was co-immunoprecipitated with FLAG-tagged TssE. Interestingly, the HA-tagged TssF was also specifically co-immunoprecipitated with TssE (Fig 3D) even though the BTH assay failed to detect this interaction. Taken together, the BTH and co-immunoprecipitation assays provide evidence that TssF and TssG interact with T6SS tail components, including the TssE baseplate protein, as well as with the tube and sheath proteins Hcp and TssC. These data provide further evidence to support the Hcp assembly assay and the bioinformatics analyses to propose that TssF and TssG form with TssE and VgrG the T6SS baseplate onto which the tube (Hcp) and sheath (TssBC) polymerize.
In bacteriophages such as T4, six wedges formed by the gp6-gp25-gp53 complex assemble around the central hub. We therefore tested whether the TssE-F-G complex interacts with VgrG using a reconstitution approach. Lysates of cells producing VSV-G epitope-tagged VgrG and FLAG-tagged TssE,-F and-G were mixed prior to immunoprecipitation on anti-VSV-G resin. Fig 3E shows that VgrG efficiently precipitates TssE,-F and-G. Based on this result and on the bacterial two-hybrid approach (Fig 3A), we hypothesized that TssE links VgrG and TssF/TssG. We therefore repeated the immunoprecipitation experiments in absence of TssE. However, in these conditions we also observed that both TssF and TssG co-immunoprecipitate with VgrG (Fig 3E). Because we did not detect direct VgrG-TssF and VgrG-TssG interactions in BTH and co-immunoprecipitation assays, these results suggested that formation of a TssF-TssG complex is pre-required to interact with VgrG. Indeed, BTH experiments showed that VgrG-TssF and VgrG-TssG interactions are detected when the third partner, TssG and TssF respectively, is present (Fig 3F). A stable TssKFG complex was recently reported [46]. However, we did not observe interactions between TssK and TssF or TssG neither in this study, nor in the systematic interaction study of TssK [40]. Interestingly, further BTH experiments showed that both TssF and TssG are required to stably interact with TssK (Fig 3F), similarly to what we observed for VgrG. Therefore, we conclude that formation of the TssFG sub-complex is a pre-requisite for further interactions with VgrG and TssK.
The experiments described above and the genetic linkage of the tssF and tssG genes suggest that TssF and TssG interact. To test this hypothesis, we first performed steady-state stability experiments in E. coli K-12 cells producing TssF, TssG or both TssF and TssG. Cells were harvested at different time points after inhibition of protein synthesis, and the stabilities of TssF and TssG were estimated by Western blot. Fig 4A shows that TssF and TssG, when produced alone, are relatively unstable proteins as TssF and TssG were undetectable 120 and 60 minutes after protein synthesis inhibition respectively. However, both proteins were stabilized when co-produced and remained detectable up to 8 hours after protein synthesis arrest. This result shows that TssF and TssG stabilize each other. The co-stabilization of these two proteins is in agreement with a recent work showing that the Serratia TssF and TssG proteins are unstable in absence of the other [46]. To test for direct interaction, we performed BTH and co-immunoprecipitation experiments. First, the BTH assay revealed that (i) both TssF and TssG are involved in homotypic interactions suggesting that these proteins dimerize or multimerize, and (ii) that these two proteins interact (Fig 4B). However, the location of the fusion at the C-terminus of TssF or at the N-terminus of TssG causes a steric hindrance that prevents TssF-TssG complex formation. In addition, the HA epitope-tagged TssF protein was specifically co-immunoprecipitated with FLAG-tagged TssG in the heterologous T6SS- host E. coli K-12 (Fig 4C) demonstrating that this interaction is not mediated by an another T6SS components. Taken together, these experiments demonstrate that TssF and TssG form a complex that stabilizes both subunits. Interestingly, within the uropathogenic E. coli (UPEC) CFT073 T6SS gene cluster, the tssF and tssG genes are fused, leading to a tssF'-'tssG chimeric gene (see Fig 4D). A Clustal W protein sequence alignment of the EAEC TssF and TssG proteins with the UPEC TssF-G fusion showed that the C-terminal PG residues of TssF are fused to the N-terminal MGFP residues of TssG to yield a PGMGFP motif in the fusion protein (see S2 Fig). To test whether a fusion protein might be functional in EAEC, we constructed a ΔtssFG mutant strain and fused the tssF and tssG genes in frame either in the native (F-G fusion) or in the opposite orientation (G-F fusion). Both fusion proteins accumulated at comparable levels. The tssF and tssG genes were also cloned contiguously, mimicking their natural genetic organization (F+G) in EAEC. As expected, the T6SS was nonfunctional in ΔtssFG cells as Hcp was not released in the culture supernatant, nor in the supernatant of ΔtssFG cells producing TssF or TssG alone. The WT phenotype was restored upon production of both TssF and TssG, or of the TssF-TssG fusion protein; however, production the TssG-TssF “inverted” fusion protein failed to complement the ΔtssFG mutation (Fig 4E). Taken together, these data provide evidence that TssF and TssG interact. The observation that the TssF-TssG fusion protein is functional suggests that TssF and TssG form a sub-complex with a 1:1 stoichiometry. However, by using quantitative gel staining, English et al. recently reported a 2:1 molar ratio for the Serratia TssF:G complex [46]. Although our data suggest a 1:1 stoichiometry, we cannot rule out that the TssF-G fusion protein we engineered is subjected to partial degradation. In bacteriophage T4, a 2:1 ratio has been noted for the gp6:gp53 complex [31], in agreement with the Serratia data [46].
To gain insight onto TssF and TssG, we further tested their sub-cellular localizations using cell fractionation experiments. In WT cells, both TssF and TssG mainly co-fractionated with EFTu, a cytoplasmic elongation factor (Fig 5A). Surprisingly, small but reproducible amounts of TssF and TssG were found associated with the membrane fraction (Fig 5A). TssF and TssG co-fractionate with the IM protein TolA, and the NADH oxidase activity in sedimentation density gradient experiments indicating that both proteins associate with the inner membrane (IM) (Fig 5B). These results suggest that TssF and TssG are peripherally associated with the IM, probably through protein-protein contacts. Interestingly, both proteins exclusively localized in the cytoplasmic fractions in E. coli K-12 (i.e., devoid of T6SS genes) (Fig 5C), further supporting the notion that TssF and TssG are tethered to the inner membrane by a T6SS component.
Three proteins of T6SS interact to form a membrane-associated complex that spans the cell envelope: the inner membrane TssL and TssM proteins and the outer membrane TssJ lipoprotein [12–15]. To identify TssF and TssG partners, we tested their interactions with the soluble domains of TssL, TssM and TssJ by BTH. As shown in Fig 5D, TssG interacts with the cytoplasmic loop of the inner membrane protein TssM (TssMc). This result was validated by co-immunoprecipitation: the HA-tagged TssG was co-immunoprecipitated with the FLAG-tagged TssMc domain (Fig 5E). The hypothesis that TssF and TssG were recruited to the membrane through interactions with TssM was tested by cell fractionation in ΔtssM cells. Fig 5F shows that in absence of TssM, TssF and TssG co-fractionate exclusively with the cytoplasmic marker EFTu and do not associate with the membrane fraction anymore. Taken together, these data show that TssF and TssG are recruited at the cytoplasmic face of the inner membrane via interactions with the cytoplasmic loop of TssM. However, association to the membrane complex might be stabilized by additional contacts involving TssFG-TssK and TssK-TssL or TssK-TssM interactions [40].
The previous data suggest that TssE,-F,-G,-K, VgrG assemble a baseplate structure anchored to the T6SS membrane complex. Based on the knowledge on bacteriophages, we hypothesized that this baseplate serves as assembly platform for tail tube/sheath extension. To further gain information on baseplate components cellular locations, recruitment and dynamic behavior, we engineered strains producing super-folder GFP (sfGFP) fused to the N- or C-terminus of TssE, TssF, TssG and TssK. All these chromosomal constructs were introduced at the native, original loci. Only the GFP-TssF (sfGFPTssF) and TssK-GFP (TssKsfGFP) fusions were functional. In WT cells, sfGFPTssF forms 1–3 foci per cell, located close to the cytoplasmic side of the envelope and with a limited dynamic (Figs 6A [upper panel] and S3A and S3B). The number of sfGFPTssF foci and their dynamic remained unchanged in tssBC cells suggesting that assembly of the T6SS sheath do not impact formation of these structures (Figs 6A and S3A and S3B). By contrast, the sfGFPTssF fluorescence was diffuse in tssK cells, demonstrating that TssK is required for proper assembly of the TssF-containing baseplates (Fig 6A). The sfGFPTssF behavior was different in tssM cells. Although ~ 95% of the cells present diffuse fluorescence, sfGFPTssF clusters are present in ~ 5% of the cells. These clusters do not remain tightly associated to the membrane but rather display random dynamics (Fig 6A), suggesting that the membrane complexes stabilize and anchor baseplate complexes to the IM.
Fluorescence microscopy recordings of cells producing both sfGFPTssF and mCherry-labeled TssB (TssBmCh) from their native chromosomal loci further informs the assembly mechanism of the T6SS: i) sfGFPTssF clusters are positioned first (Figs 6B and S3C), and ii) The TssBmCh sheath extends from the preassembled sfGFPTssF cluster (Figs 6B and S3D and S3E). In addition, the recordings show that sfGFPTssF foci remain associated with the sheath during all the cycle (elongation, contraction and disassembly) (S3C and S3D Fig). These observations strongly support a model in which TssF-containing complexes serve as platforms for the polymerization of the sheaths. The TssKsfGFP fusion present similar behavior to sfGFPTssF: it assembles 1–3 stable and static foci per cell and shows diffuse localization in absence of TssM (Figs 6C and S3B and S3F). Conversely, sfGFPTssM forms discrete foci in WT cells [15] that assemble independently of TssK (Figs 6D and S3G). Based on these results, and in agreement with the hypothesis raised by English and co-authors [46], we suggest the TssL-M-J membrane complex serves as the docking area for the T6SS tail-like structure, the assembly of which starts with the subsequent recruitments of TssK, TssFG, Hcp and TssBC (Fig 6E).
Recent studies have evidenced that the T6SS assembles a cytosolic structure similar to the tail tube and sheath of contractile bacteriophages [16,23–26]. During bacteriophage morphogenesis, the tube and sheath polymerize onto the baseplate that initiates and guides the assembly process [29–31]. In addition, the baseplate undergoes large conformational rearrangements upon landing on a target cell that ultimately trigger tail contraction [30]. In the bacteriophage T4, the baseplate is a complex structure; however, simplest baseplates exist in other contractile bacteriophages. Based on baseplates comparison, Leiman & Shneider proposed that only four components will be required to have a functional baseplate: the gp27-gp5 hub complex (or spike) and the gp25, gp6 and gp53 wedge subunits [22]. However, in the T6SS, the identity and composition of the baseplate that controls the assembly of the tube/sheath structure are not known [10–12,37]. To identify these components, we developed an assay to probe formation of Hcp tubes in vivo [18]. Employing systematically this assay in all T6SS gene deletion mutant strains, we identified six T6SS core components required for the controlled polymerization of the Hcp tube. This experimental approach was validated by the fact we identified VgrG and TssE, the T6SS homologues of the bacteriophage spike/hub complex and gp25 subunits respectively. Among the four remaining subunits, TssA, TssK, TssF and TssG, we showed that the two laters share limited but significant homologies with protein J and I, two baseplate components of phage P2, respectively [32,33]. In addition to bacteriophages and T6SS, homologues of P2 baseplate I and J proteins, as well as homologues of the spike/hub and gp25, are found in anti-feeding prophages, Photorhabdus Virulence Cassettes and R-pyocins [47] suggesting that they likely represent the core of phage-like protein translocation machineries. A schematic comparison of the homology and contacts between bacteriophage T4 and T6SS components is shown in S4 Fig. In this study, we focused our work on the TssF and TssG proteins. We demonstrated that TssF and TssG interact with each other and with TssE. Although we have not addressed the biological relevance of these interactions, these results are in agreement with the co-occurrence of these three genes in T6SS gene clusters [2], as well as with the observation that the bacteriophage T4 homologues of TssE and TssF–gp25 and gp6 –interact [48]. Interestingly, Aksyuk and coauthors showed that the gp6-gp25 interaction involves the N-terminal fragment of gp6, the fragment that shares homology with TssF [48]. In addition, contacts were detected with TssK, VgrG and components of the tube (Hcp) and of the sheath (TssC). Interestingly, contacts with TssK and VgrG require the pre-formation of the TssF-G complex. Taken together these data support the idea that the T6SS baseplate is composed of the VgrG, TssE, TssF, TssG and TssK subunits, on which the Hcp tube and the TssB-TssC sheath will sequentially polymerize (Fig 7). Indeed, co-localization studies showed that TssF is recruited to the apparatus prior to sheath extension. Based on the observations that this baseplate is docked to the membrane complex and remains at the base of the extended tail, it likely corresponds to the structure observed at the same location on T6SS electron cryo-tomographs [26].
The phage wedge and baseplate assembly pathways are regulated by the stepwise addition of the different subunits in a strict order [39,44]. Regarding T6SS biogenesis, the results presented here as well as three recent studies [15,40,46] support the proposal that the TssLMJ membrane complex is first assembled and that the T6SS is built by the hierarchical addition of TssK and TssFG. Alternatively, complete baseplates may assemble prior to docking to the membrane complex (Fig 7). Indeed, the baseplate-like structure is anchored to the inner membrane by a network of interactions including contacts between TssK and TssL, TssK and TssM, and TssG and TssM ([40] and this study). One may hypothesize that TssK is recruited to the TssJLM complex, and the interaction between the baseplate-like structure and the membrane complex is then stabilized by additional contacts between TssM and the TssFG complex. This assembly pathway is supported by fluorescence microscopy recordings demonstrating that TssF is not recruited to the apparatus in absence of TssK or TssM, and that TssK is not recruited to the apparatus in absence of TssM (Fig 6). Interestingly, in absence of TssM, TssF-containing complexes (baseplates?) are assembled–albeit at significant lower levels, suggesting that these complexes are stabilized by the membrane complex. The TssM-independent assembly of these TssF-containing complexes is in agreement with the observations that (i) T6SS membrane and baseplate/tail complexes have distinct evolutionarily histories [2] and (ii) that the membrane complex is not required for proper assembly of Hcp tubes (Fig 1C). Strikingly, phage baseplates display 6-fold symmetry [22,29] and one could assume that an identical symmetry will apply for T6SS baseplates. However, the T6SS membrane complex has been recently show to have 5-fold symmetry [15]. Understanding how a 6-fold symmetry structure successfully and functionally associates to a 5-fold symmetry docking station remains to be elucidated. Once the baseplate is docked to the TssJLM complex, the Hcp inner tube/TssBC outer sheath polymerization can proceed (Fig 7). Indeed, interaction studies showed that TssF and TssG are connected to Hcp and TssC, and might initiate tail extension. Data reporting the specific role of TssA during T6SS biogenesis will be reported elsewhere (Zoued, Durand et al., in preparation). The spatio-temporal recruitment of TssE and VgrG during the T6SS assembly pathway and their contributions to the structure of the T6SS baseplate are not yet known and will require further investigations. In contractile tailed bacteriophages, the baseplate serves as an assembly platform for the tube/sheath polymerization and triggers contraction of the sheath by transducing conformational changes from the fibers upon landing on host cells. It will be important to define the contribution of the T6SS baseplate to the control of T6SS sheath dynamics upon contact with prey cells.
Escherichia coli K-12 DH5α was used for cloning procedures, W3110 for co-immunoprecipitation, and BTH101 for the bacterial two-hybrid assay. The enteroaggregative E. coli strain 17–2 was used for this study. Strains were routinely grown in LB broth at 37°C, with aeration. For induction of the sci-1 T6SS gene cluster, cells were grown in Sci-1-inducing medium (SIM: M9 minimal medium supplemented with glycerol (0.2%), vitamin B1 (1 μg/mL), casaminoacids (40 μg/mL), LB (10% v/v)) [49]. Plasmids and mutations were maintained by the addition of ampicillin (100 μg/ml for K-12, 200 μg/ml for EAEC), kanamycin (50 μg/ml for K-12, 50 μg/ml for chromosomal insertion on EAEC, 100 μg/ml for plasmid-bearing EAEC) or chloramphenicol (40 μg/ml). L-arabinose was purchased from Sigma-Aldrich, anhydrotetracyclin (AHT–used at 0.2 μg/ml throughout the study) from IBA. The strains, plasmids and oligonucleotides used in this study are listed in S1 Table.
Δtss deletion mutant strains were constructed using the modified one-step inactivation procedure [50] using red recombinase expressed from pKOBEG [51] as previously described [52]. The kanamycine cassette from pKD4 [50] was amplified with oligonucleotides carrying 50-nucleotide extensions homologous to regions adjacent to the target gene. The Polymerase Chain Reaction (PCR) product was column purified (Promega PCR and Gel Clean up) and electroporated. Kanamycin resistant clones were recovered and the insertion of the kanamycin cassette at the targeted site was verified by PCR. The kanamycin cassette was then excised using plasmid pCP20 [50], and the final strain was verified by PCR. The Δsci-1 deletion strain, which comprises a deletion of the tssB-tssE fragment (i.e., all the T6SS genes), was constructed similarly. Chromosomal fluorescent reporter insertions were obtained by the same procedure using pKD4-gfp, pgfp-KD4 and pmCh-pKD4 as templates for PCR amplification.
PCR were performed with a Biometra thermocycler, using the Pfu Turbo DNA polymerase (Stratagene; La Jolla, CA). Custom oligonucleotides were synthesized by Eurogentec. Constructions of pOK-HcpHA and pUC-HcpFLAG and its derivatives have been previously described [18,52]. pTssF-HA has been constructed by insertion of EcoRI-XhoI PCR fragment into pMS600 digested by the same enzymes. All other plasmids have been constructed by restriction-free cloning [53]: the gene of interest was amplified with oligonucleotides carrying 5’ extensions annealing to the target vector. The product of the first PCR has then been used as oligonucleotides for a second PCR using the target vector as template. All constructs have been verified by DNA sequencing (MWG).
In vivo disulfide cross-linking assay was performed as previously described [18].
We used the adenylate cyclase-based two-hybrid technique using previously published protocols [40,54,55]. Briefly, pairs of proteins to be tested were fused to the two catalytic domains T18 and T25 of the Bordetella adenylate cyclase. After co-transformation of the BTH101 strain with the two plasmids producing the fusion proteins, plates were incubated at 30°C for 2 days. 600 μl of LB medium supplemented with ampicillin, kanamycin and 0.5 mM isopropyl—thio-galactoside (IPTG) were inoculated with independent colonies. Cells were grown at 30°C overnight and spotted on LB agar plates supplemented with ampicillin, kanamycin, IPTG (0.2 mM) and 5-bromo-4-chloro-3-indolyl—D-galactopyrannoside (X-Gal), the chromogenic substrate of the -galactosidase.
1011 exponentially growing cells producing the proteins of interest were harvested, and resuspended in Tris-HCl 20 mM (pH8.0), NaCl 100 mM supplemented with protease inhibitors (Complete, Roche) and broken by three passages at the French press (1000 psi). The total cell extract was ultracentrifuged for 45 min at 20,000 × g to discard unsolubilized material. Supernatants were then incubated overnight at 4°C with anti-FLAG M2 affinity beads (Sigma Aldrich) or with Protein G-Agarose beads (Roche) coupled to the anti-VSVG antibody. Beads were then washed three times with Tris-HCl 20 mM (pH8.0), NaCl 100 mM. The total extract and immunoprecipitated material were resuspended in Laemmli loading buffer prior to analyses by SDS-PAGE and immunoblotting. For reconstitution experiments, cell lysates were mixed for 30 min. at 28°C prior to immune precipitation.
Hcp release assay, Fractionation, SLS differential solubilisation, discontinuous sedimentation sucrose gradients and NADH oxidase activity measurements were performed as previously described [13].
Overnight cultures of entero-aggregative E. coli 17–2 derivative strains were diluted 1:100 in SIM medium and grown for 6 hours to an OD600nm ~ 1.0 to maximize expression of the sci-1 T6SS gene cluster that is up-regulated in iron-depleted conditions [49]. Cells were washed in phosphate buffered saline (PBS), resuspended in PBS to an OD600nm ~ 50 and spotted on a thin pad of 1.5% agarose in PBS and covered with a cover slip. Microscopy recordings and digital image processing have been performed as previously described [9,15,18,40]. For statistical analyses, fluorescent foci were automatically detected. First, noise and background were reduced using the ‘Subtract Background’ (20 pixels Rolling Ball) plugin from Fiji [56]. The sfGFP foci were automatically detected by a simple image processing: (1) create a mask of cell surface and dilate (2) detect the individual cells using the “Analyse particle” plugin of Fiji (3) sfGFP foci were identified by the “Find Maxima” process in Fiji [56]. To avoid false positive, each event was manually controlled in the original raw data. Box-and-whisker representations of the number of foci per cell were made with R software. For sub-pixel resolution tracking, fluorescent foci were detected using a local and sub-pixel resolution maxima detection algorithm and tracked over time with a specifically-developed plug-in for ImageJ [56]. The x and y coordinates were obtained for each fluorescent focus on each frame. The mean square displacement was calculated as the distance of the foci from its location at t0 at each time using R software and plotted over time. For each strain tested, the mean square displacement of at least ten individual focus trajectories was calculated. Kymographs were obtained after background fluorescence substraction and sectioning using the Kymoreslicewide plug-in under Fiji [56].
The protein stability was assessed as previously described [57]. Exponential growing cells producing TssF, TssG or both TssF and TssG were treated with chloramphenicol (40 μg/ml) and spectinomycin (200 μg/ml). Equivalent OD samples were harvested at 0, 5, 15, 30, 60, 120, 240, 480 and 960 minutes after protein synthesis arrest, resuspended in loading buffer prior to analyzes by SDS-PAGE and Western blot immunodetection. Bacterial density (OD600nm) was measured throughout the experiment to verify that no growth occurred.
Proteins suspended in loading buffer were subjected to SDS-PAGE. For detection by immunostaining, proteins were transferred onto nitrocellulose membranes, and immunoblots were probed with primary antibodies, and goat secondary antibodies coupled to alkaline phosphatase, and developed in alkaline buffer in presence of 5-bromo-4-chloro-3-indolylphosphate and nitroblue tetrazolium. Anti-TolA,-TolB,-Pal and-OmpA polyclonal antibodies are from our laboratory collection. Anti-FLAG (Sigma-Aldrich), anti-VSV-G (Sigma-Aldrich), anti-HA (Roche), anti-EFTu (Hycult Biotech) and anti-rabbit,-mouse or-rat alkaline phosphatase-conjugated goat secondary antibodies (Beckman Coulter) have been purchased as indicated and used as recommended by the manufacturer.
|
10.1371/journal.pntd.0002668 | Potential Wildlife Sentinels for Monitoring the Endemic Spread of Human Buruli Ulcer in South-East Australia | The last 20 years has seen a significant series of outbreaks of Buruli/Bairnsdale Ulcer (BU), caused by Mycobacterium ulcerans, in temperate south-eastern Australia (state of Victoria). Here, the prevailing view of M. ulcerans as an aquatic pathogen has been questioned by recent research identifying native wildlife as potential terrestrial reservoirs of infection; specifically, tree-dwelling common ringtail and brushtail possums. In that previous work, sampling of environmental possum faeces detected a high prevalence of M. ulcerans DNA in established endemic areas for human BU on the Bellarine Peninsula, compared with non-endemic areas. Here, we report research from an emergent BU focus recently identified on the Mornington Peninsula, confirming associations between human BU and the presence of the aetiological agent in possum faeces, detected by real-time PCR targeting M. ulcerans IS2404, IS2606 and KR. Mycobacterium ulcerans DNA was detected in 20/216 (9.3%) ground collected ringtail possum faecal samples and 4/6 (66.6%) brushtail possum faecal samples. The distribution of the PCR positive possum faecal samples and human BU cases was highly focal: there was a significant non-random cluster of 16 M. ulcerans positive possum faecal sample points detected by spatial scan statistics (P<0.0001) within a circle of radius 0.42 km, within which were located the addresses of 6/12 human cases reported from the area to date; moreover, the highest sample PCR signal strength (equivalent to ≥106 organisms per gram of faeces) was found in a sample point located within this cluster radius. Corresponding faecal samples collected from closely adjacent BU-free areas were predominantly negative. Possums may be useful sentinels to predict endemic spread of human BU in Victoria, for public health planning. Further research is needed to establish whether spatial associations represent evidence of direct or indirect transmission between possums and humans, and the mechanism by which this may occur.
| Mycobacterium ulcerans causes the disfiguring human skin disease Buruli ulcer (BU). The mechanism of transmission and reservoir for human infection remain unknown. In previous research, we reported the detection of M. ulcerans DNA in the faeces of possums (small tree-dwelling marsupials) in an area of South-East Australia (the Bellarine peninsula) where the largest recorded outbreak of human BU has been in progress for the last decade. The current study was carried out in a new outbreak area (the Mornington peninsula), and describes the detection of M. ulcerans DNA in possum faeces collected from the ground, in locations which correspond closely with the addresses of human BU cases. The association of new human BU cases with areas where M. ulcerans positive possum faeces are found contributes further evidence to the possible role of possums as an environmental reservoir of infection. Possums may be useful sentinel animals to monitor the spread of BU in Australia.
| Mycobacterium ulcerans is an environmental, potentially zoonotic bacterial pathogen, which in humans causes the progressive ulcerative skin condition Buruli Ulcer (BU), a neglected tropical disease which is endemic in at least 30 countries worldwide [1]. The majority of the disease burden is in West and sub Saharan Africa, however there is a significant and ongoing outbreak in temperate south-eastern Australia in the state of Victoria (where the disease is also referred to as Bairnsdale ulcer) [2], [3]. In all settings, the geographic distribution of human BU case clusters is highly focal. The exact method of disease transmission is unknown: BU foci in Africa have been associated with natural bodies of fresh water such as rivers and lakes, prompting the hypothesis that human infection is acquired through skin abrasions by physical contact with contaminated water [4], [5] or from the bites of infected aquatic insects such as water bugs (Naucoridae) [6]. It has also been observed that new endemic areas emerge in areas adjacent to recent soil disturbance and flooding [7]. In south-east Australia, infection is consistently associated with coastal areas [8], [9] and the mechanism of transmission remains elusive, although several studies have indicated that mosquitoes may have a role [2], [10]. The bacterium infects a wide range of terrestrial mammals in Australia, including both domestic animals [11]–[13] and native wildlife [14], [15]. More recently, an extensive survey conducted in an area endemic for human BU on the Bellarine Peninsula (Point Lonsdale; see map, Figure 1) revealed that a large proportion of faecal samples from common ringtail (Pseudocheirus peregrinus) and common brushtail (Trichosurus vulpecula) possums contained high concentrations of M. ulcerans DNA [16]. That study showed a strong association between BU endemicity of an area and the proportion and DNA concentration of M. ulcerans positive possum faecal samples. More recently, BU has emerged in a previously non-endemic area of the Mornington Peninsula, in the towns of Sorrento and Blairgowrie (Figure 1) distant from previous historical foci further to the east (Phillip Island, and the Frankston-Langwarrin area of outer Melbourne). From 2006 to the present, 12 new cases of human BU have been confirmed in patients who were either residents (n = 6) or visitors (n = 6) to this region, with no known contact with any of the established endemic areas for BU such as the Bellarine Peninsula, and no recent history of travel to endemic areas either interstate or overseas. In light of previous research showing the potential for possums in BU endemic areas to excrete M. ulcerans DNA, the aim of this study was to determine whether the presence and/or relative abundance of M. ulcerans DNA in possum faecal samples could be associated with a newly established focus of human BU cases in a previously non-endemic area. We undertook a systematic survey of ground collected possum faeces in the area of the emergent Mornington Peninsula BU focus with the objective to analyze the distribution of M. ulcerans positive possum faecal samples and to look for spatial associations with human BU case addresses.
The towns of Sorrento (population 1448) and Blairgowrie (population 2161), are located near the western tip of the 750 km2 Mornington Peninsula, approximately 90 km south of Melbourne (Figure 1). The terrain is predominantly low-lying coastal scrubland (<50 m above sea level), with an average annual rainfall of approximately 730 mm. There are no substantial water courses or large bodies of fresh water in either of the two towns; open drainage ditches with accumulations of standing water are uncommon. The sample sites in each town consist of similar networks of single-track asphalt or gravel roads with grass verges, connecting rows of large dwellings set in spacious fenced grounds with fairly abundant scrub and tree cover including coastal tea trees (Leptospermum laevigatum) as well as numerous introduced cultivar species in gardens. A significant proportion of properties in this region are not occupied by permanent residents, but used as holiday homes or temporary tourist accommodation.
A case of Buruli Ulcer (BU) was defined as a human patient with at least one suggestive clinical lesion from which M. ulcerans DNA was detected by real-time IS2404 PCR. The identity of the M. ulcerans strain was confirmed via Variable Number of Tandem Repeat (VNTR) typing [17] of the cultured isolate or using DNA extracted from the clinical specimen at the Victorian Infectious Diseases Reference Laboratory (VIDRL). A patient was suspected of having acquired BU from the Sorrento/Blairgowrie area if he/she was a resident of (n = 6), or a visitor to (n = 6), that area and had not reported recent contact (<12 months) with any other BU endemic area. Addresses were available for all 6 residents, whereas the addresses of visitors' holiday homes were available in 3 cases. The remaining 3 non-resident patients visited holiday homes of unknown location within the Sorrento/Blairgowrie area.
Samples of possum faeces were collected from ground level (roadside verges) at points arranged in a predetermined grid pattern 200 m apart. Where there was no tree cover at the indicated grid point (and hence, no possum faeces), samples were collected from the nearest available location where faecal pellets could be found. Faeces originating from common ringtail and common brushtail possums (hereafter referred to as ringtail and brushtail possums) were collected and distinguished based on their characteristic size and shape by experienced field workers and with the aid of a track and scat manual [18]. Where possible intact scats, which had not started to break down due to weather and invertebrates, and estimated to be less than a week old, were selected. The sample spacing interval was chosen in an attempt to minimize resampling of faeces from the same animals between adjacent points, since possums are highly territorial and radio-tracking data show that they generally restrict their movements within a radius of approximately 100 m or less (A. Legione, unpublished data). Faecal samples from each sampling location were stored separately in sterile ziplock plastic bags and transported cool to the laboratory for storage at +4°C prior to DNA extraction, typically within a week of collection.
DNA was extracted from possum faecal material using the FastDNA SPIN Kit for Soil (MP Biomedicals, Solon, OH). Faeces (approximately 100 mg) was added to the kit-supplied Sodium Phosphate and MT Buffer in Lysing Matrix E tubes, and was homogenized for 40 s at setting 6 on a FastPrep Instrument (MP Biomedicals, Solon, OH). Tubes were then centrifuged at maximum speed in a bench microfuge for 10 minutes to pellet debris, before the supernatant was removed and mixed with the supplied protein precipitation solution. After centrifugation at maximum speed for 5 min, 200 µl supernatant was transferred for extraction using an automated robotic system (Corbett X-tractor gene, Qiagen), following the manufacturer's recommendations. Extracted DNA (100 µl) was stored at −20°C. Two microliters of DNA template were used in subsequent real time PCR reactions targeting three independent regions in the M. ulcerans genome (IS2404, IS2606 and KR), as described previously [19]. Based on the difference in cycle threshold (Ct) values between IS2606 and IS2404 (ΔCt [IS2606-IS2404]) these assays are able to distinguish between M. ulcerans strains, which typically cause disease in mammals, and other members of the M. ulcerans/M. marinum complex (with fewer copies of IS2606) which may be present in the environment, but are not associated with the human outbreak. An estimate of M. ulcerans bacterial load per gram of possum faeces was obtained based on the previously established correlations between IS2404 PCR Ct values and bacterial loads in spiked possum faeces [16]. These calculations enabled comparison of the relative amounts of M. ulcerans DNA between samples in the present survey, expressed in 10-fold orders of magnitude up to ≥106 organisms per gram of faeces, and should be considered semi-quantitative rather than absolute. Culture of M. ulcerans was not attempted since our previous research has shown this to be an insensitive diagnostic method when applied to possum faeces, due to overgrowth of contaminants [16].
VNTR typing was performed using 1 µl DNA template in 25 µl reaction volume, using conditions described previously [17]. PCR products were visualized on 2% agarose gel with ethidium bromide staining, and product size was estimated with reference to a 100 bp DNA ladder (Promega, Wisconsin, USA). Products of the expected size were purified using the High Pure PCR Purification Kit (Roche Diagnostics, Australia) and sequenced using the BigDye Terminator v3.1 Cycle Sequencing Kit (Applied Biosystems, Foster City, CA) according to the manufacturer's instructions. VNTR sequences were compared with those from a well characterized Victorian M. ulcerans isolate (Strain Ref: MU_JKD8049), which was obtained from a BU patient linked to Point Lonsdale in 2004 [20].
Scan statistics were used to detect and evaluate clusters of positive possum faecal samples in a purely spatial setting, using a Bernoulli model (binary outcome) [21]. This analysis was carried out in SaTScan V8.0 (http://www.satscan.org/).
Mycobacterium ulcerans DNA was detected by real-time PCR in 20/216 (9.3%) ground collected ringtail possum faecal samples and 4/6 (66.6%) brushtail possum faecal samples. There was a significant non-random clustering of positive possum faecal samples identified by spatial scan statistics (P<0.0001; 16/30 samples were positive within a circle of radius 0.42 km; see Figure 2). There was a visually apparent spatial correlation between the occurrence of positive possum faeces and the addresses of 6 human BU cases. Four patients who resided locally and two patients who had holiday homes in the area were located within the cluster radius described above: due to requirements of patient confidentiality we are unable to show the specific locations of patients' addresses in Figure 2. Due to the small number of human cases to date, it is not yet possible to confirm statistically significant clustering of human BU cases in this area. Additionally, one human case (resident) was located adjacent to an outlier positive faecal sampling point (again we are unable to depict this case location due to confidentiality requirements) in an area of predominantly negative possum faecal samples. Finally, two human BU case addresses were located in areas where possum faeces was not sampled (one holiday house address, and one resident). The calculated values for ΔCt (IS2606-IS2404) from M. ulcerans PCR positive faecal samples were ≤3.32 (95% CI = 1.56–2.61), confirming that all the sequences detected were attributable to M. ulcerans and not another member of the M. ulcerans/M. marinum complex which typically give higher ΔCt values (95% CI = 6.94–8.07) [19]. IS2404 real-time PCR Ct values ranged from 24–39, corresponding with M. ulcerans burdens in faeces (estimated as described previously) ranging from ≥106 to 100 organisms per gram of possum faeces. The median estimated bacterial load in M. ulcerans positive ringtail faeces was 103–104 organisms per gram, which was similar to the median load in positive brushtail faeces. The two faecal samples with the highest M. ulcerans DNA concentrations (≥106 organisms per gram) were from ringtail possums sampled within the cluster of positive possum faecal samples as described above. The DNA concentration in these two samples was sufficiently high to allow sequencing of VNTR locus 14 (both samples) and 9 (1 sample only). The nucleotide sequences obtained were identical to those from the strain of M. ulcerans causing human BU disease in Victoria (MU_JKD8049).
Human BU incidence in south-eastern Australia is on the increase, particularly in the last two decades [2], [8]. The progressive extension of the westernmost extremity of the endemic area from the original Bairnsdale region more than 260 km to the east, and the frequent emergence of new endemic foci are current public health concerns. In the most recently studied focus on the Bellarine Peninsula, BU was first reported in 1998, and is now endemic in three small towns near the Eastern tip of the peninsula: in Point Lonsdale (see Figure 1 for location map), the infection rate calculated in 2011 (26 cases) was equivalent to 770/100,000 population (C. Lavender, unpublished data). Tracking of the geographic shift of endemic areas (such as the recent emergence of BU on the western extremity of the Mornington Peninsula) using traditional epidemiological survey methods is complicated by the long incubation period of the disease in humans (median 4.5 months, IQR = 109–160 days) [22], requiring time-consuming analysis of patients' historical movement patterns over an extended period to identify their exposure location. This is particularly challenging in patients who acquire infection from very brief visits to an endemic area as in one documented case, during a stay lasting only a few hours [3]. Further complications in tracking the location of patients' exposure arise because many BU-affected areas are popular holiday resorts that experience high numbers of visiting non-residents particularly in summer. Conversely, survey sampling of roadside-collected possum faeces is straightforward, detection of M. ulcerans DNA by real-time PCR can be done within hours using an automated robotic platform, and such a sampling procedure does not raise issues of informed consent, since it does not require examination or interview of human patients. There is increasing interest in the use of wildlife sentinels to monitor the emergence and spread of a number of zoonotic diseases such as West Nile disease, rabies, and anaplasmosis [23], [24]. As a first step towards validating possum faecal surveys as a public health tool to monitor BU emergence, we show here that detection of M. ulcerans DNA in possum faeces was associated with a recent outbreak of BU in a previously non-endemic area of the Mornington Peninsula. A significant non-random cluster of Mycobacterium ulcerans PCR positive possum faeces was closely adjacent to the addresses of 6 of the total 9 Sorrento/Blairgowrie human BU patients for whom we have obtained residential and holiday home addresses, and the highest M. ulcerans bacterial loads in possum faeces coincided with this cluster. It should be noted that in the present study, sampling was carried out on roadside verges underneath overhanging branches of trees growing along the fence line of residences: we cannot rule out the possibility that conditions within fences and boundaries differ from those outside, however this seems unlikely since possum movement is not restricted by such artificial barriers at ground level as being arboreal, they are highly adapted for climbing. Although it was not possible to accurately determine the age of the individual faecal pellets, all were collected from areas exposed to rain and invertebrates which increase the rate of degradation of such samples [25] and on this basis were estimated to be up to a week old. Since pre-outbreak sampling (before 2006) was not done, it is not yet possible to confirm the temporal relationship between possum and human infections. It is noteworthy that we have identified a small number of positive possum scat in a survey of nearby area of the Mornington peninsula (approximately 2 km distant from the present study site) which as yet has no human BU cases (data not shown) – any developments will be reported in future research. Interestingly, the geographic location of the current outbreak area of Sorrento is adjacent to a new housing development, built on the site of a golf course in the mid-1990s. It is highly likely that significant soil disturbance would have taken place at that time.
VNTR typing showed that the M. ulcerans in possum faeces on the Mornington Peninsula was indistinguishable from the strain causing human disease in south-east Australia, as was previously demonstrated in possum faeces collected on the Bellarine Peninsula [16]. It is not yet known whether this finding reflects transmission of M. ulcerans between possums and humans, or simply a common environmental source of infection. Consistent with previous research [19], we found that sequencing of VNTR loci could be achieved only from IS2404 positive faecal samples with high M. ulcerans bacterial loads (≥106 organisms per gram in the present study). Also in agreement with previous work sampling possum faeces in Victoria [26], no other members of the M. ulcerans/M. marinum complex were detected in faecal samples collected in Sorrento and Blairgowrie.
Relative to the ubiquitous nature of ringtail possum faeces in the environment, brushtail possum faecal specimens were found rarely (in 6 sampling grid locations only). However the proportion of M. ulcerans IS2404 positive brushtail faecal specimens was higher than that of ringtail samples (66.6%; 4/6 vs 9.3%; 20/216), and it is interesting that positive samples from both species of possum coincided spatially in Sorrento, adjacent to a focus of human BU. Population survey work would be required to determine if ringtail and brushtail possums do indeed coexist in the outbreak area and not elsewhere, which may support the hypothesis of M. ulcerans as a cyclozoonosis.
Conversely, in previous work in Point Lonsdale on the Bellarine Peninsula (for map, see Figure 1), the proportion of positive brushtail samples was lower than that of ringtail possum faecal samples (29%; 8/28 vs 43%; 70/164) [16]. The overall number of sample points with brushtail faecal specimens in Point Lonsdale (n = 28) was greater than in the current study in Sorrento/Blairgowrie (n = 6), despite similar sizes of the two sampling areas (approximately 5 km2). It is not known if this finding reflects a higher population density of brushtail possums in a well-established BU endemic area, than in the location of a more recent outbreak: confirmation of this would require a survey of the live possum population to estimate overall numbers using established techniques such as spotlighting or trapping [27].
The limited distribution (i.e. highly focal nature) of both human cases and positive possum samples at Sorrento/Blairgowrie contrast the distribution pattern at Point Lonsdale, where human cases and infected possums were more widespread across the whole township [16]. This distribution pattern probably reflects the recent nature of the outbreak in the former, and the considerably longer term presence of the disease agent in populations of both possums and humans in the latter. It will be insightful to reassess the Sorrento/Blairgowrie site again in several years.
As a future research priority, we need information on the degree to which relative population density of ringtail and brushtail possums influences endemicity and/or emergence of BU in humans. Specifically, longitudinal follow-up is needed of human BU disease incidence, possum population dynamics and the prevalence of possum faecal M. ulcerans DNA, in the above described emergent endemic area on the Mornington Peninsula. Live trapping of possums has not yet been done in this area to confirm the presence of possums with M. ulcerans positive skin lesions, as described in the previous work on the Bellarine Peninsula [16], as distinct from those showing only positive faecal samples. This distinction is important since faecal shedding could occur by simple ingestion of the pathogen e.g. on vegetation, and subsequent excretion, whereas the development of active clinical disease in possums shows the potential for establishment of a long lasting infectious reservoir host. Mycobacterium ulcerans in superficial skin lesions would also be available for uptake by biting insects which could potentially act as vectors of BU, as discussed in previous research [2], [3], Overall, a better understanding of spatial and temporal associations between human and possum M. ulcerans infection is likely to be the key to elucidation of the transmission mechanism of BU in south-east Australia.
|
10.1371/journal.ppat.1005963 | Tissue Degeneration following Loss of Schistosoma mansoni cbp1 Is Associated with Increased Stem Cell Proliferation and Parasite Death In Vivo | Schistosomiasis is second only to malaria in terms of the global impact among diseases caused by parasites. A striking feature of schistosomes are their ability to thrive in their hosts for decades. We have previously demonstrated that stem cells, called neoblasts, promote homeostatic tissue maintenance in adult schistosomes and suggested these cells likely contribute to parasite longevity. Whether these schistosome neoblasts have functions independent of homeostatic tissue maintenance, for example in processes such as tissue regeneration following injury, remains unexplored. Here we characterize the schistosome CBP/p300 homolog, Sm-cbp1. We found that depleting cbp1 transcript levels with RNA interference (RNAi) resulted in increased neoblast proliferation and cell death, eventually leading to organ degeneration. Based on these observations we speculated this increased rate of neoblast proliferation may be a response to mitigate tissue damage due to increased cell death. Therefore, we tested if mechanical injury was sufficient to stimulate neoblast proliferation. We found that mechanical injury induced both cell death and neoblast proliferation at wound sites, suggesting that schistosome neoblasts are capable of mounting proliferative responses to injury. Furthermore, we observed that the health of cbp1(RNAi) parasites progressively declined during the course of our in vitro experiments. To determine the fate of cbp1(RNAi) parasites in the context of a mammalian host, we coupled RNAi with an established technique to transplant schistosomes into the mesenteric veins of uninfected mice. We found transplanted cbp1(RNAi) parasites were cleared from vasculature of recipient mice and were incapable of inducing measurable pathology in their recipient hosts. Together our data suggest that injury is sufficient to induce neoblast proliferation and that cbp1 is essential for parasite survival in vivo. These studies present a new methodology to study schistosome gene function in vivo and highlight a potential role for schistosome neoblasts in promoting tissue repair following injury.
| Schistosomes are parasitic flatworms that infect more than 200 million people in the developing world. Once these parasites infect a human, they are capable of living in the bloodstream for decades. Previously, our group has shown that these parasites have stem cells that are capable of renewing worn out cells in these parasites. Although we know these stem cells can continuously restore aging tissues, whether these stem cells can respond to injuries in the parasite is not clear. Here, we show that reducing levels of a gene called cbp1 results in cell death and degeneration of certain schistosome tissues and causes dramatic increases in the level of stem cell proliferation. This result suggested that the parasites might perceive elevations in cell death as a form of injury, which then triggers stem cell proliferation to enhance the rate of tissue repair. To explore this idea further, we physically injured parasites and observed that physical injury was indeed capable of inducing stem cell proliferation. We also found that loss of cbp1 resulted in the progressive decline in the health of parasites cultured in the laboratory, likely as a result of increased cell death and tissue degeneration. To determine if cbp1 was also important for parasites living inside a mammalian host, we coupled our gene disruption approaches with a classic technique to transplant schistosomes into the veins of uninfected mice. Using this novel methodology, we found that reducing cbp1 levels resulted in parasite death inside mice. Taken together, these observations demonstrate that cbp1 is important for parasite survival inside a mammalian host and show that schistosome stem cells are capable of responding to injury. These data have important implications for the development of new therapies and for understanding the roles stem cells play in the biology of schistosome infection.
| Schistosomes infect over 200 million people and are a major cause of morbidity in the developing world. The primary driver of this morbidity is the prodigious egg production of these parasites, which can lay several hundred eggs every day while living in the vasculature of their hosts [1]. A large fraction of these eggs are swept into the circulation and become lodged in host organs (such as the liver and bladder), leading to inflammatory responses that can compromise organ function [2]. The pathological consequences of schistosome egg production are compounded by the fact that schistosomes can survive and produce eggs for decades inside their human hosts [1, 3]. Understanding the developmental forces that promote parasite longevity is essential for understanding the chronic nature of this disease.
Schistosomes possess a population of somatic stem cells similar to the neoblasts found in free-living flatworms (e.g., freshwater planarians) [3, 4]. In schistosomes, these neoblast-like cells appear to represent the only proliferative somatic cell type [4] and support the homeostatic renewal of tissues such as the intestine [4] and tegument [5]. Together, these data suggest that schistosome neoblasts are likely critical for long-term parasite survival in their hosts. What is not clear is whether neoblasts serve other important functions in these parasites. In free-living planarians, neoblasts are essential for both homeostatic tissue maintenance and for tissue regeneration [6, 7]. Following amputation, there is a burst in planarian neoblast proliferation, which fuels the regeneration of damaged and missing tissues [8, 9]. Unlike planarians, schistosomes live exclusively in the vasculature of mammalian hosts and are unlikely to face the same types of mechanical insults (e.g., amputation) that planarians do [3]. Therefore, whether schistosome neoblasts are capable of interpreting injury signals and modulating their behavior to repair damage is not clear. However, since schistosomes are likely subjected to a myriad of immunological and chemical insults inside their mammalian host, it is possible that neoblasts could possess the capacity to respond to various types of injury. Thus, understanding how parasites respond to injury, and the role of neoblasts in tissue repair, would provide important new insights into the mechanisms that support parasite longevity in vivo.
During the course of a systematic effort to identify factors with the potential to regulate schistosome neoblast function we characterized Sm-cbp1 (for brevity, we will refer to this gene as cbp1), a gene that encodes a homolog of the mammalian CBP/p300 family of proteins [10]. In mammals, these proteins serve as transcriptional co-activators that possess histone acetyltransferase (HAT) activity [11]. In schistosomes, cbp1 was previously demonstrated to act as a transcriptional co-activator in vitro [10] and suggested to regulate genes important for schistosome egg production via its HAT activity [12]. Here we show that abrogation of cbp1 function leads to simultaneous increases in cell death and neoblast proliferation. Based on our observation that physical injury similarly induces parasite cell death and neoblast proliferation, we suggest that increases in neoblast proliferation following cbp1(RNAi) is a strategy by the parasite to cope with cell death-mediated tissue damage. In addition, we report a novel application of existing techniques to examine adult schistosome gene function in vivo and show that cbp1 is essential for parasite survival in mice. These data suggest an important function for cbp1 in parasite survival and highlight a potential role for neoblasts in regenerative processes in schistosomes.
Using whole-mount in situ hybridization (WISH) we found that cbp1 was expressed in adult parasites in a variety of cells throughout the worm’s parenchyma and in cells within the male testes and female ovaries (Fig 1A and 1B). To characterize how broadly cbp1 was expressed in the parenchyma, we performed fluorescence in situ hybridization (FISH) with two markers of somatic cells residing in the parenchyma: Histone H2B to mark neoblasts [4] and tsp-2 to label tegument-associated cells [5]. In addition to being expressed in both Histone H2B+ and tsp-2+ cells, we weakly detected cbp1 transcripts in most cells within the schistosome parenchyma (Fig 1C and 1D). While we cannot conclude that cbp1 is expressed in every cell in the worm, our data suggest this gene is expressed in a large number of schistosome cell types.
To explore a role for cbp1 in regulating schistosome stem cells, we performed RNAi experiments. In comparison to controls, depletion of cbp1 mRNA levels (S1A and S1B Fig) led to a dramatic (Fig 1E and 1F) and statistically significant (Fig 1G and 1H) increase in the number of neoblasts that incorporated the thymidine analog EdU. Similar increases in cell proliferation were observed with dsRNAs targeting two distinct regions of the cbp1 gene, indicating these effects are specific to the reduction of cbp1 levels (S1A and S1C Fig) and not due to off-target effects. To explore this observation further, we also performed WISH with the neoblast markers Histone H2B and fgfrA [4] (Fig 1I and 1J) and FISH (Fig 1K) with Histone H2B. Similar to our observations with EdU incorporation, we noted an increase in the number of cells expressing neoblast markers (Fig 1I–1K). Together, these data suggest that loss of cbp1 increases the number of proliferative neoblasts.
Two simple stem cell behaviors can explain our observations following cbp1(RNAi). First, loss of cbp1 could block the ability of neoblasts to differentiate, effectively locking the cells in a proliferative state. This type of behavior is observed following perturbations that block planarian neoblast differentiation [13]. Alternatively, the cells could maintain the capacity for differentiation but the size of the stem cell pool is expanded via an increased rate of cell proliferation. To distinguish between these possibilities, we performed WISH for the neoblast differentiation progeny marker tsp-2. Previously we demonstrated that tsp-2 is expressed in a tegument-associated cell population that is the primary differentiation progeny of schistosome neoblasts [5]. Since tsp-2+ cells are short lived and rapidly renewed by neoblasts [5], they are a sensitive measure of the capacity for neoblasts to differentiate. Consistent with neoblasts in cbp1(RNAi) parasites maintaining the ability to differentiate, we observed substantial increases in the number of tsp-2+ cells in cbp1(RNAi) parasites (Fig 1L). Together, these data suggest that loss of cbp1 expands the size of the neoblast pool and this results in an increased rate of production of at least one differentiated cell type.
The schistosome esophageal glands are located anterior to the intestine (Fig 2A) and are thought to secrete factors that aid in the digestion of blood cells [14, 15]. By both EdU labeling (S1C Fig) and FISH for Histone H2B (Fig 2B) we noted a focus of proliferative neoblasts in the vicinity of the esophageal glands in cbp1(RNAi) animals. We explored this observation more closely by double FISH for Histone H2B and the esophageal gland marker meg-4 [16, 17]. Consistent with our prediction, at D11 of RNAi-treatment, masses of neoblasts are observed surrounding the esophageal gland of cbp1(RNAi) parasites (Fig 2C). In some cases we observed “holes” in the esophageal gland that were occupied by Histone H2B+ neoblasts (Fig 2C, top cbp1(RNAi) panels). In the most severe cases, the esophageal glands were degenerated and only small numbers of meg-4+ cells remained (Fig 2C, bottom cbp1(RNAi) panels). To explore the degeneration of the esophageal glands in more detail, we performed time course analyses examining the expression of meg-4 by WISH. We observed a progressive degeneration of the esophageal gland in cbp1(RNAi) parasites and by D18 cbp1(RNAi) parasites possessed few traces of meg-4+ gland cells (Fig 2D and 2E).
We next explored the relationship between neoblast proliferation and the degeneration of the esophageal glands in cbp1(RNAi) parasites. In principle, the observed masses of neoblasts (Fig 2B and 2C) could either be a cause of esophageal gland degeneration, an effect of this degeneration, or unrelated to the disappearance of the gland. Given how prominent the masses of proliferative neoblasts are surrounding the gland (Fig 2B and 2C), we believe the latter of these possibilities is unlikely. Therefore, to determine if neoblast proliferation is a cause or an effect of gland degeneration, we treated parasites with γ-irradiation, which rapidly depletes neoblasts [4], and examined meg-4 expression by FISH. In control(RNAi) parasites, neoblast depletion had no observable effect on the morphology of the esophageal gland at D11 (Fig 2F). In contrast to control(RNAi) parasites, irradiated and unirradiated cbp1(RNAi) parasites displayed extensive degeneration of the esophageal glands (Fig 2F), suggesting that neoblast over proliferation is not likely a direct cause of gland loss. Although we observed substantial gland degeneration in both irradiated and unirradiated cbp(RNAi) parasites, we noted more scattered meg-4+ cells in unirradiated cbp1(RNAi) where neoblasts were present (arrowheads, Fig 2F). Based on this observation we speculate that many of these remaining meg-4+ cells in unirradiated cbp1(RNAi) parasites represent newly born differentiation progeny of the neoblasts.
Our data indicated that between D8 and D14 a large fraction of parasites had esophageal glands that were in intermediate stages of degeneration (Fig 2D and 2E). To determine if programmed cell death was playing a role in this degeneration, we developed a whole-mount assay to examine Terminal deoxynucleotidyl transferase dUTP Nick-End Labeling (TUNEL). TUNEL is a methodology to detect double stranded breaks in the DNA of cells undergoing the process of programmed cell death [18], and has been successfully used to detect apoptosis in both free-living flatworms [19] and in sectioned adult female schistosomes [20]. Using this assay we determined that at D10 28% of cbp1(RNAi) parasites had large clusters of TUNEL+ cells within their esophageal glands (Fig 3A–3C). Visualizing glands with the lectin PNA [21], large pockets of TUNEL+ cells were not observed in cbp1(RNAi) parasites with largely intact glands nor in parasites with severely degenerated glands. Rather the presence of large numbers of TUNEL+ cells was restricted to glands that appeared to be in the early to intermediate stages of degeneration. These data suggest that programmed cell death is a likely driver of esophageal gland cell loss.
In the esophageal glands of cbp1(RNAi) parasites we noted elevations in cell proliferation and cell death at roughly similar time points after beginning dsRNA treatment (Figs 2C and 3C). Since we also noted increases in cell proliferation throughout the bodies of cbp1(RNAi) parasites (Fig 1F–1H) we explored if cell death was similarly elevated in the trunks and tails of cbp1(RNAi) parasites (Fig 3A). Although we did not note measurable changes by D4 of RNAi, at both D10 and D14 we observed statistically significant increases in TUNEL+ cells in cbp1(RNAi) parasites (Fig 3D–3F); by D14 cbp1 RNAi treatment on average resulted in 4.6 and 4.8-fold elevations in TUNEL+ nuclei in trunks and tails of male parasites, respectively (Fig 3D and 3E). Interestingly, at both D10 and D14 the levels of cell death varied considerably among individual cbp1(RNAi) parasites: some cbp1(RNAi) parasites possessed levels of TUNEL+ nuclei comparable to controls, whereas in other parasites the number of TUNEL+ nuclei was dramatically elevated (Fig 3D and 3E). This observation mirrors what we observed in the esophageal glands where large numbers of dying cells were only present in a subset of parasites in which the glands were in the process of degenerating (Fig 3C). Therefore, these elevations in cell death observed in the trunks and tails may similarly reflect the sudden degeneration of one or more tissue types in cbp1(RNAi) worms. Unfortunately, given the paucity of cell type-specific markers that are compatible with TUNEL staining, it is presently not possible to determine if this elevated rate of cell death was restricted to a specific cell/tissue type or whether all tissues were undergoing similar levels of cell death. Nevertheless, our data suggest that in addition to being required for preventing cell death, and degeneration of the esophageal gland cells, cbp1 is important for maintaining normal levels of cell death in other tissues within the parasite.
In diverse organisms (e.g., Hydra [22], Drosophila [23], planarians [19]) tissue injury induces apoptosis and precedes increases in stem cell proliferation [24]. Therefore, one attractive model to explain the simultaneous elevations of both neoblast proliferation and cell death observed in cbp1(RNAi) parasites could be that cbp1 is required for the survival of various cell types in the worm (e.g., esophageal gland cells) and that death of these cells induces neoblast proliferation. Alternatively, cbp1 could be acting in some cells (e.g., esophageal gland cells) to promote cell survival and acting independently in neoblasts to repress proliferation. To indirectly distinguish between these possibilities, and examine if tissue injury can induce both cell death and neoblast proliferation in schistosomes, we physically injured male parasites. For these experiments, parasites were immobilized on an agarose pad and poked with a sharpened tungsten needle (Fig 4A). Consistent with this injury regime inducing tissue damage and subsequent cell death in the worm, we noted substantial numbers of TUNEL+ nuclei at wound sites 4 hours post-injury (Fig 4B). We next examined injured parasites with neoblast and cell proliferation markers 48–72 hours following injury. Consistent with injury inducing neoblast proliferation, we noted accumulations of EdU-incorporating cells (Fig 4C) and Histone H2B+ neoblasts (Fig 4D) surrounding wound sites at both 48 and 72 hours post-injury. Similarly, by immunofluorescence we noted increases in cells positive for M-phase specific marker Phospho-Histone H3 at sites adjacent to wounds (Fig 4E). Interestingly, examination of parasites at 48-hours post-injury by TUNEL staining found that although increases in cell death could often still be detected at the wound site, rates of cell death were depressed in the tissues immediately adjacent to wounds relative to the rest of the parasite (Fig 4E). This suggests that injury may repress physiological rates of cell death in tissues near wound sites. This repression of cell death may serve as a mechanism to preserve the function of tissues undergoing repair. Taken together, our data suggest that injury, and perhaps cell death, is capable of stimulating neoblast proliferation. Furthermore, these data suggest that schistosomes may be capable of utilizing neoblast-mediated tissue renewal to fuel tissue repair following injury.
Presumably due to elevations in cell death and declining tissue function, we observed that cbp1(RNAi) parasites became progressively sicker during in vitro culture (Fig 5A, S1 Movie). By D8, male and female cbp1(RNAi) parasites became unpaired and lost the ability to attach to the surface of the tissue culture dish (Fig 5A). By D15, parasite movement became uncoordinated and often times the heads of male worms curled ventrally (Fig 5A, S1 Movie). At D19, movement in cbp1(RNAi) parasites was limited to irregular and jerky motions (S1 Movie). The progressive decline in the vitality of parasites was not likely due to elevations in cell proliferation since irradiated cbp1(RNAi) parasites were indistinguishable from unirradiated cbp1(RNAi) parasites with regards to male-female pairing and attachment to the substrate (S1 Movie).
Given the complexity of the schistosome lifecycle and the lack of robust transgenic tools, few studies to date have examined adult schistosome gene function in the context of a mammalian host [25]. To explore if cbp1 is essential for parasite survival in vivo, we coupled in vitro RNAi treatment with a procedure pioneered by Cioli in the 1970’s for the surgical transplantation of schistosomes into the mesenteric veins of rodent hosts [26]. For these experiments, 4 to 5 week old parasites were recovered from mice, treated for 4 days with control or cbp1 dsRNA in vitro, and then surgically transplanted into the mesenteric veins of recipient mice (Fig 5B). At D26 post-transplantation, we euthanized the mice, performed hepatic portal vein perfusion, and measured both the percent recovery of transplanted parasites and extent of schistosome induced host pathology. In mice that received control(RNAi) worms, we noted hepatosplenomegaly consistent with the transplanted parasites establishing a productive infection (Fig 5C). Following hepatic portal vein perfusion, we recovered about 70% of the male control(RNAi) parasites originally transplanted (Fig 5D). In contrast to controls, mice receiving cbp1(RNAi) parasites did not display hepatosplenomegaly (Fig 5C) and we failed to recover any male parasites following hepatic portal vein perfusion (Fig 5D). We also noted obvious signs of egg-induced liver pathology in control(RNAi) recipient mice (Fig 5E); no evidence of egg-induced granuloma formation was observed in cbp1(RNAi) recipient mice (Fig 5E). Examination of histological sections from the livers of control and cbp1(RNAi) recipient mice confirmed that control parasites were capable of generating egg-induced pathology whereas no egg-induced inflammation was observed in cbp1(RNAi) recipient mice (Fig 5F and 5G).
Although we detected no signs of egg-induced inflammation, we did note large masses located at the periphery of the livers of cbp1(RNAi) recipient mice (Fig 5E, arrowhead). Examination of these livers in histological sections found these masses to be cbp1(RNAi) parasites trapped in the liver of these mice (Fig 5H). Observing these sections in more detail, we identified worms at several stages of deterioration: some parasites were relatively intact with an uninterrupted tegument (Fig 5H, left panel) whereas others were severely degenerated with virtually no organized schistosome tissues (Fig 5H, right panel). The composition of host cells surrounding the parasite, and the apparent maturity of the immunological response to the worms, correlated with the structural integrity of the worms. More intact worms were surrounded by large numbers of neutrophils and lymphocytes (indicative of early host response) whereas more degenerated worms were found in lesions encased in fibroblasts (indicative of a mature host response to the parasites). These data suggest cbp1(RNAi) parasites are incapable of establishing an infection. Based on what we observe in vitro (Fig 5A and S1 Movie), we hypothesize that within 4–5 days following transplantation these parasites lose the ability to attach to the host endothelium and are washed into the liver. In the liver, the health of the parasites continues to decline and they are susceptible to being killed by the host immune system, perhaps in a similar fashion as schistosomes treated with praziquantel in vivo [27]. Based on these data, we suggest that cbp1 is essential for schistosome survival in vivo.
Aside from supporting new cell birth during the physiological turnover of tissues (e.g., the tegument [5]), we know relatively little about the roles that neoblasts play in the biology of adult schistosomes. Here, we report that reductions in cbp1 levels result in simultaneous elevations of both cell proliferation and cell death. The esophageal glands were emblematic of this: apoptosis driven cell death was accompanied by massive accumulations of proliferative neoblasts. These observations suggested that neoblasts might be equipped to respond to lost or damaged tissues, an observation we confirmed by demonstrating that physical wounding induced proliferative neoblasts to accumulate around wound sites. Based on these data we suggest a model in which reduction of cbp1 levels leads to cell death and tissue loss throughout the parasite (Fig 6). This cell loss is (directly or indirectly) sensed by neoblasts resulting in an increased rate of neoblast proliferation. Since we observe large increases in the number of cells expressing the neoblast progeny marker tsp-2, it is likely the neoblasts then differentiate to restore lost cells. Because cbp1 levels remain depressed due to the effects of RNAi these newly differentiated cells die, inducing more neoblast proliferation. Tissue degeneration and the inability of neoblasts to restore tissue function eventually results in parasite death.
While physical injury induces schistosome neoblast proliferation, the precise role apoptosis and other types of cell death (e.g., necrosis) play in this process are not known. In the cnidarian Hydra, programmed cell death releases Wnt molecules that are required to induce stem cell proliferation and regeneration following amputation [22]. In Drosophila, genetic induction of apoptosis stimulates proliferation of intestinal stem cells [23]. In planarians, injury induces apoptosis although the requirement for cell death in fueling regeneration is not clear [19]. Therefore, dying cells in schistosomes may directly signal to induce neoblast proliferation. Alternatively, a myriad of other factors (e.g., loss of tissue integrity and/or loss of cell-cell contacts) may stimulate neoblast proliferation. As tools to study schistosome cell death mature, it should be possible in the future to determine precisely how apoptosis influences neoblast behavior.
Mammalian cbp1 homologs serve as transcriptional co-activators linking transcription factors to the core transcriptional machinery [11]. These mammalian cbp1 relatives also possess acetyltransferase activity and can acetylate a variety of substrates including histones and non-histone proteins [11]. Whether these activities of cbp1 are important for maintaining schistosome cellular viability is not presently clear. However, previous studies have shown that pharmacological inhibition of histone deacetylase activity induces apoptosis in larval schistosomes [28, 29]. Thus, not unexpectedly, maintaining normal chromatin structure is likely important for schistosome cellular survival. Since cbp1 possesses histone acetyltransferase activity in vitro [10], the cell death induced by cbp1 may be due to alterations in chromatin landscape in certain cell types. Further exploration of chromatin-modifying enzymes may represent fertile ground for the development of novel therapeutics.
Here, we combine a previously described method for the surgical transplantation of schistosomes and RNA interference to demonstrate that cbp1 is required for parasite survival in vivo. Not only do these studies validate the potential to target cpb1 therapeutically, they provide a novel methodology to explore the functions of schistosome genes in vivo. A potentially useful application of this approach is for studies of schistosome reproduction. Since schistosome reproduction ceases within one week of in vitro culture [20], this approach could help identify genes required for the development and maintenance of the schistosome reproductive system. One potential limitation of this approach is the persistence of the effects of RNAi. Although the effects of RNAi have been reported to last for several weeks in larval schistosomes in vitro [30], how this translates to older parasites in vivo is not known. However, as tools to manipulate the schistosome genome (i.e., transgenic expression and genome editing) continue to mature, we suggest surgical transplantation could become an invaluable tool to explore gene function in vivo.
Our observation that injury is met with increases in neoblast proliferation indicates schistosomes may possess the capacity to regenerate following certain types of injury in vivo. The regenerative potential of schistosomes has not been extensively characterized and conflicting reports exist. Senft and Weller reported that schistosomes amputated during recovery from mice were capable of regenerating new tails in vitro [31]. However, this conflicts with another account where in vitro cultured worms were capable of rapidly healing wounds, but incapable of regeneration [32]. Thus, the ability of schistosomes to perform whole-body regeneration (i.e., regenerating new heads and/or tails) is unresolved and may be a function of culture conditions and the nature of the injury. What is less controversial is the ability of schistosomes to repair tissues following in vivo exposure to sublethal doses of the anthelminthic drug praziquantel [33]. Thus, future studies exploring roles for neoblasts in tissue repair, following a variety of injuries (e.g., amputation and drug treatment) and in a variety of culture conditions, are necessary and could have important implications for understanding the longevity and resilience of these parasites in vivo.
In adherence to the Animal Welfare Act and the Public Health Service Policy on Humane Care and Use of Laboratory Animals, all experiments with and care of vertebrate animals were performed in accordance with protocols approved by the Institutional Animal Care and Use Committee (IACUC) of the UT Southwestern Medical Center (protocol approval number APN 2014–0072).
Adult S. mansoni (6–8 weeks post-infection) were obtained from infected female mice by hepatic portal vein perfusion with 37°C DMEM (Sigma-Aldrich, St. Louis, MO) plus 10% Serum (either Fetal Calf Serum or Horse Serum) and heparin. Parasites were cultured as previously described [5]. Unless otherwise noted, all experiments were performed with male parasites.
cDNAs used for in situ hybridization and RNA interference were cloned as previously described [34]. Quantitative PCR analyses were performed as previous described [5]. Oligonucleotide sequences are listed in S1 Table.
EdU labeling, whole-mount in situ hybridization and fluorescence in situ hybridization analyses were performed as previously described [4, 5]. For RNAi experiments, 5–10 freshly perfused male parasites (either as single worms or paired with females) were treated with 30 μg/ml dsRNA for 4 days in Basch Media 169 [35]. dsRNA was generated by in vitro transcription [4] and was replaced every day. As a negative control for RNAi, we used a non-specific dsRNA containing two bacterial genes [4]. Sequences used for dsRNA synthesis are listed in S2 Fig. For irradiation of RNAi-treated parasites, worms were exposed to 100 Gy of Gamma Irradiation using a J.L. Shepard Mark I-30 Cs137 source. Lectin labeling was performed as previously described [21]. For TUNEL labeling, parasites were fixed for 4 hours in 4% Formaldehyde in PBS + 0.3% Triton X100 (PBSTx), dehydrated in methanol, and stored at -20°C. Parasites were subsequently rehydrated with PBSTx, permeabilized with 20ug/ml Proteinase K (Invitrogen, Carlsbad, CA) in PBSTx for 45 min, and post-fixed with 4% Formaldehyde in PBSTx. Following fixation parasites were processed for TUNEL labeling using the In situ BrdU-Red DNA Fragmentation (TUNEL) Assay Kit (Abcam). For this procedure, post-fixed worms were briefly incubated in the kit provided “wash” buffer, incubated in “DNA labeling solution” (2 to 3 male worms per 50 ul) for 4 hours at 37°C, rinsed twice in PBSTx, blocked with “FISH Block” (0.1 M Tris pH 7.5, 0.15 M NaCl and 0.1% Tween-20 with 5% Horse Serum and 0.5% Roche Western Blocking Reagent [36]), and incubated overnight in Anti-BrdU-Red Antibody (1:20) in “rinse buffer”. After several PBSTx washes, worms were either mounted on slides in Vectashield (Vector Labs, Burlingame, Ca) or further processed for immunofluorescence or lectin labeling. For immunofluorescence, permeabilized worms were blocked in FISH Block and incubated overnight at 4°C in Anti-Phospho-Histone H3 (Ser10) (Rabbit mAB, D2C8, Cell Signaling) diluted 1:1000 in FISH block. Following 6 x 1 hour washes in PBSTx worms were incubated overnight at 4°C in Goat anti-Mouse IgG secondary antibody conjugated to AlexaFluor 488 diluted in FISH block (Thermo Fisher). Following several washes in PBSTx, parasites were mounted on slides in Vectashield.
Confocal imaging of fluorescently labeled samples and brightfield imaging (i.e, whole-mount in situ hybridizations and histological sections) were performed using a Zeiss LSM700 Laser Scanning Confocal Microscope or a Zeiss AxioZoom V16 equipped with a transmitted light base and a Zeiss AxioCam 105 Color camera, respectively. All images of fluorescently-labeled samples represent maximum intensity projections. To perform counts of EdU+ and TUNEL+ cells, cells were manually counted in maximum intensity projections derived from confocal stacks; to normalize between samples cell counts were divided by the total volume of the stack in μm3. All plots and statistical analyses were performed using GraphPad Prism.
For injury, worms were gently pipetted onto the surface of a 35 mm Petri dish filled with solidified 4% agarose diluted in H2O. After removal of excess liquid, worms were perforated with a sharpened tungsten needle. The impaled parasites were then carefully removed from the needle into fresh media using a pipette tip. As a control, “mock” injured parasites were similarly transferred to Petri dishes but were not injured; we observed no changes in cell death or cell proliferation in these parasites.
Methods for surgical transplantation of schistosomes are based on a procedure originally developed for hamsters [26]. 4 to 5 days prior to surgery parasites 4–5 weeks post-infection were recovered from mice and treated with 30 μg/ml dsRNA for 4 days in Basch Media 169 [35] as previously described [4]. Media and dsRNA were changed daily. Before mice were anesthetized, 8 male parasites (either paired or unpaired with female, see below) were sucked into a 1ml syringe, the syringe was fitted with a custom 25G extra thin wall hypodermic needle (Cadence, Cranston, RI), the air and all but ~300 μL of media were purged from the needle, and the syringe was placed needle down in a test tube to settle the parasites to the bottom of the syringe. We attempted to inject male/female worm pairs, but it was not always clear if females were present in the gynecophoral canal. Therefore, each injection also included a few unpaired female parasites to ensure maximal potential for mating. Once the syringe was loaded with parasites, young male Swiss Webster mice (~25–30G) were anesthetized with Isoflurane using a vaporizer system equipped with both an induction chamber and nose cone. Abdomens of anesthetized mice were shaved and the area was sterilized with three alternating scrubs with betadine and ethanol. A single longitudinal incision (~1.5 cm) centered on the navel was made to expose the intestines. A sterile piece of gauze with a 2 cm slit in the center was dampened with sterile saline and placed over the incision. The intestines were gently fed through the gauze to expose the large vein running along the cecum. The intestines were kept damp throughout the entire procedure with sterile saline. Making sure the bevel of the needle remained facing down, the worms were injected into the cecal vein. To avoid hemorrhage, prior to removing the needle a small piece of hemostatic gauze (Blood Stop) was placed over the injection site. As the needle was removed, gentle pressure was applied to the injection site. Once bleeding stopped (~1–2 minutes) the hemostatic gauze was removed and the intestines returned into the abdominal cavity. The cavity was filled with sterile saline and abdominal muscles and skin were sutured (Maxon, Absorbable Sutures, Taper Point, Size 4–0, Needle V-20, ½ Circle). Following wound closure, mice received a single subcutaneous dose of buprenorphine for pain (30 μl of 1 mg/ml) and were allowed to recover on a warm heating pad. After transplant, needles were flushed with media to determine how many parasites had been injected into each mouse. Mice were group housed and individual mice were tracked by marking their tails with a permanent marker. On day 26 post-transplantation mice were sacrificed and perfused to recover parasites. Male and female parasites were counted and livers were removed and fixed for 30–40 hours in 4% formaldehyde in PBS. The percentage parasite recovery was determined by dividing the number of male worms transplanted by the number of male parasites recovered following perfusion. Counting male parasites was the most informative since the initial number of female parasites was not accurately quantified (see above). Livers from individual mice were sectioned and processed for Haematoxylin and Eosin staining by the UT Southwestern Molecular Pathology Core.
|
10.1371/journal.pcbi.1001003 | Structure Learning in Human Sequential Decision-Making | Studies of sequential decision-making in humans frequently find suboptimal performance relative to an ideal actor that has perfect knowledge of the model of how rewards and events are generated in the environment. Rather than being suboptimal, we argue that the learning problem humans face is more complex, in that it also involves learning the structure of reward generation in the environment. We formulate the problem of structure learning in sequential decision tasks using Bayesian reinforcement learning, and show that learning the generative model for rewards qualitatively changes the behavior of an optimal learning agent. To test whether people exhibit structure learning, we performed experiments involving a mixture of one-armed and two-armed bandit reward models, where structure learning produces many of the qualitative behaviors deemed suboptimal in previous studies. Our results demonstrate humans can perform structure learning in a near-optimal manner.
| Every decision-making experiment has a structure that specifies how rewards are obtained, which is usually explained to the subject at the beginning of the experiment. Participants frequently fail to act as if they understand the experimental structure, even in tasks as simple as determining which of two biased coins they should choose to maximize the number of trials that produce “heads”. We hypothesize that participants' behavior is not driven by top-down instructions—rather, participants must learn through experience how the rewards are generated. We formalize this hypothesis using a fully rational optimal Bayesian reinforcement learning approach that models optimal structure learning in sequential decision making. In an experimental test of structure learning in humans, we show that humans learn reward structure from experience in a near optimal manner. Our results demonstrate that behavior purported to show that humans are error-prone and suboptimal decision makers can result from an optimal learning approach. Our findings provide a compelling new family of rational hypotheses for behavior previously deemed irrational, including under- and over-exploration.
| From a squirrel deciding where to bury its nuts to a scientist selecting the next experiment, all decision-making organisms must balance exploration of alternatives against exploitation of known options in developing action plans. Finding a balance is equivalent to knowing when you can profit from learning about new options and knowing when you know enough. However, determining when exploration is profitable is itself a decision problem that requires understanding or learning about the statistical structure of the environment. Theoretical work on optimal exploration [1], [2] shows that assessing the long-term value of exploration involves integrating the predicted informational value of exploration with primary reward. Predicting the value of future information requires having a model of the reward generation process for the domain.
The structure learning problem may be present in tasks with as few as two options. Suppose, for example, that you interact with the environment by choosing one of the two options at discrete choice points and that the option chosen generates a stochastic binary reward. As a rational agent, your aim is to maximize the total reward from the environment, but the difficulty is that the rate of reward for each option is unknown and must be learned. In this simple setting, there may be several hypothesis about how the reward generation process works—how actions, observations and unknowns are structurally “connected.” We propose three kinds of structures that capture several versions of sequential decision-making tasks available in the literature. The first structure has temporal dependency between the present probability of reward and the past probability of reward, investigated in the context of Multi-Armed Bandit problems [3]–[5]. When this dependency involves a random walk, the environment becomes non-stationary and a rational agent will discount both past reward observations [6] and potential future reward (equivalent to discounting) and it will exhibit a higher learning rate in the sense of a greater dependence on recent reward information. In the second structure, reward probabilities can be affected by actions. For example, choosing an option may temporarily decrease the reward probability. Different kinds of action-reward probability contingencies can produce a range of different rational responses, from probability matching (foraging) to maximization. The third structure is reward coupling and is the primary focus of this paper.
To illustrate what structure learning entails, Fig. 1A shows a probabilistic graphical model representing the possible relationships between variables for a typical sequential decision task with two outcomes. In the graph, nodes represent unknown or observable quantities and links represent statistical contingencies between them. The unknown probabilities of reward at a given time for both option 1 and 2 are represented by and , respectively. Taking action at time produces a reward that can be either 0 (failure) or 1 (success). Learning the success probabilities must be balanced with the desire to maximize expected future reward. Different assumptions about the connectivity (structure) between variables produce a surprising range of rational responses. One of those structures is temporal dependency (see Fig. 1B) between success probabilities. In this case, rather than being fixed, the success probabilities and depend on past values and [3], [4]. The second structure includes an effect of actions on reward probabilities (see Fig. 1C). Different kinds of action-reward probability contingencies can produce a range of different rational responses, from matching to maximization [7], [8]. Fig. 1D illustrates Reward coupling which determines whether the reward probabilities are related to each other. For example, options may be probabilistically coupled so that if one option is “good” the other must be “bad”. This type of structure has profound consequences on exploratory and exploitative behavior.
To illustrate reward coupling, imagine you are serving a ball in tennis against an opponent who almost always adopts the same position near the center of the court. How do you choose whether you serve left or right? Assume the defender must anticipate and make its choice to defend left or right before it sees your serve. Clearly you should take advantage of the previous history of successful and unsuccessful serves against this opponent to try to exploit any weakness, but how you should make use of this history depends on what you can learn from your choices. For example, if you last served left and failed, can you infer it would have been better to serve right? The answer depends critically on the way options are probabilistically related. The outcomes of an anticipatory defender are probabilistically coupled - its probability of selecting left is one minus its probability of selecting right (similar to a coin flip). For coupled outcomes, what can be learned on each trial is independent of your actions and no active exploration is needed.
Imagine instead you throw a ball at one of two targets: left or right—with the goal of determining which target is easier to hit. In this case, you can infer little from a failure on the left target about your success on the right. The options are independent, which means that observing one option tells you little or nothing about the other. Exploration is then necessary for learning, and your choices impact what can be learned. Thus, the kind of probabilistic dependence between options determines whether passive (action independent) or active learning strategies are needed.
An organism with initial ignorance about the environment will not have a model of the probabilistic coupling, and thus will not know the value of exploration. But how can it know what kind of probabilistic dependence is present?
In this work, we investigate the possibility that people learn models of reward generation using rational analysis. From a rational perspective, actions should be selected both to increase reward and to provide information about the reward generation process. Probabilistic methods for learning dependencies between variables are termed structure learning or causal learning, and has been an active topic within the machine learning community. We argue that structure learning plays a major role in human sequential decision making. Because structure denotes the statistical relationships between entities and events, it forms the basis for generating future predictions, and it enables model-based approaches to reinforcement learning.
Using model-based (Bayesian) reinforcement learning [9]–[12] optimal exploration can be extended to handle uncertainty across a set of plausible reward generation models. In one formulation we follow here, latent parameters on model structure are treated as a hidden state, such that the algorithm tries to find values of the hidden state that maximize expected discounted reward. In essence, at the beginning of a set of tasks, we assume there is initial uncertainty over a parametric family of structures—causal models of reward generation. The learning of this causal structure is then incorporated into acting. This is a natural extension of causal induction (predictive of behavior in simpler tasks [13]) to sequential experimentation.
To maximize the differences that uncertainty about the causal relationships between options would produce, we exposed subjects to one of two possible models that represent two extremes in the exploration– exploitation trade-off in a slot-machine gambling environment, where the probabilistic coupling between the payoffs between machines must be learned. Using Bayesian RL to generate an optimal exploratory agent for this environment, we show that optimal actions with reward model uncertainty include exploratory actions that are specific to model learning, and exhibit patterns that would be considered over- and under- exploration for an agent without reward model uncertainty. We demonstrate that humans are able to learn the probabilistic coupling structure for this environment, and that they exhibit exploratory choice behavior predicted by reward model learning.
Participants made decisions in a set of 32 two-option tasks, each terminating stochastically, with an average of 48 trials. For each task, an option produced an stochastic binary reward with a fixed probability that had to be estimated by the participant. Participants were asked to maximized their reward gathered for the whole experiment and were compensated in proportion to the total reward.
Formally, the choice of option 1 or 2 transitions the agent into that state, and generates an observable binary reward and , respectively. The reward distributions are initially unknown but remain constant within a task, which ends stochastically with a probability . At the end of each task the reward distributions are reset. The tasks are analogous to playing slot machines in a casino. There are two slot machines. The state of the environment represents which of the slot machines is active. Actions involve selecting which of the machines to activate (pull the slot machine lever), and active machines generate binary rewards probabilistically.
To experimentally test how well humans can learn the probabilistic coupling structure of an environment, we used two environments with different reward structure designed to generate clear differences in decisions and exploratory behavior. In the first environment, which we term independent, the reward distributions for each machine are independent. In the second environment, called coupled, the two reward distributions are coupled by sharing a common cause: when one option gives reward, the other will not. The optimal policies for these environments generate exploratory behavior that span the range of possibilities, from independent where exploration is necessary to coupled, where exploration is superfluous. An agent with uncertainty about whether the environment is coupled or independent will need to learn both the coupling structure and the reward values of the options.
The environments were presented as two distinctive “blocks” of tasks. Each block was presented as a “game room” and machines in that game room had a unique color (blue in one room and yellow in the other). Unknown to the subjects, however, the first block of 16 tasks corresponded to one reward structure and the second block of 16 tasks corresponded to other reward structure.
We argue that it would be unreasonable for participants to assume a reward structure beforehand. They, instead, have to perform an estimation of this structure through a block of tasks while jointly learning the reward rates within the task. To predict human decisions in the task, we develop a normative model that makes decisions while actively gathering evidence about both task structure and the rewards available at each option and compare its performance both to other normative models that assume a fixed task structure and to model-free RL based on Q-learning with soft-max action selection.
In general, structure learning involves estimating the underlying dependency structure between variables. Such learning has been formulated as a probabilistic inference problem, where inference is performed over a family of hypothesized dependencies. Within machine learning, it is common to represent these dependencies using graphical models, in which nodes are variables and conditional dependencies between variables can be represented as edges.
More specifically, a graphical model conveys knowledge on how a joint probability distribution can be factored into multiple known conditional probabilities. For example, in Fig. 2A, and ignoring all the plates, the edge from node to node would indicate that the joint probability distribution can be equivalently written as the product of two known distributions . Additionally, a plate is a shorthand notation for replicating variables inside it while sharing conditional relationships and distribution functions. For example, the node inside the plate with means that there are variables () that have the same known distribution function. The node is inside a plate with and inside the plate, which indicates—quite compactly—that the total set of nodes is for each . Finally, the conditional probabilities , for any and , correspond to the same distribution function.
A variety of Machine Learning methods have been developed to perform structure learning in graphical models (e.g., [14], [15]), and these have provided a compelling account of human causal inference and learning in cognitive tasks [13], [16]. Below we show human structure learning in a sequential decision-making task. However, formulating the structure learning problem within sequential decision making is significantly more difficult, requiring a combination of probabilistic inference with reinforcement learning commonly called Bayesian reinforcement learning.
Bayesian reinforcement learning (BRL) can be used to describe an agent that learns the structure of rewards in the environment while performing optimal action selection that balances exploration and exploitation. Agents interact with a stochastic environment by performing an action that affect the state of the environment by transitioning to a new state with probability . Rewards are received with a probability that depends on the action and the outcome of the action. For the agents we are interested in describing, the goal is to maximize the reward accumulated across participation in a set of tasks which end stochastically with a probability . The optimal BRL agent schedules actions that maximize the expected reward received during the task: , where is the reward to be received immediately, the reward received next, the reward received two steps into the future, and so on, and is current model of the environment. In standard model-based reinforcement learning, the agent uses a probabilistic model of reward sources and environment to compute this expectation. In BRL, the agent does not know either the reward sources and environment precisely, but rather generates beliefs over a family of possible models.
After each observation, the belief distribution is updated using Bayesian inference. By considering the set of possible future observations, this belief updating can be used to “look ahead” to predict future rewards that can be achieved from different plans of action. The value of a belief can be found using the Bellman equation [17](1)where represents the belief “update” by Bayes' rule(2)
In the context of reinforcement learning, a policy is a prescription of what action should be taken at a particular state. One of the key ideas in BRL is that the optimal policy can be described as a mapping from belief states to actions. In particular, an optimal policy can be recovered by(3)
We specialized this framework to model structure learning in sequential decision experiments (see Materials and Methods for more details). For the BRL agent with structure learning, uncertainty about reward dynamics and contingencies can be modeled by including within the belief state not only reward probabilities, but also the possibility of independent or coupled structure. Maximizing the expected reward over this belief state yields the optimal balance of exploration and exploitation, resulting in action selection that seeks to simultaneously maximize (1) immediate expected rewards, (2) information about reward dynamics and (3) information about task structure.
Fig. 2A represents a graphical model for the generation of rewards in an independent environment. Rewards are samples from Bernoulli distributions with separate Beta prior distributed reward probabilities for each option. The belief state about is summarized by the counts of the number of successes and failures for each option. Fig. 2B shows a graphical model for a coupled environment. Coupling is represented as a “shared” probability of reward from which the rewards and are sampled. However, the probability of reward follows a Bernoulli distribution with parameter whereas follows a Bernoulli distribution with parameter .
To model learning coupling structure, we introduce a hidden binary state , representing whether the options are independent or coupled in the environment. Uncertainty about the coupling structure generates a mixture between the independent and coupled environment models. Fig. 2C shows the full graphical model that incorporates uncertainty about the environment structure. It is a mixture model of the independent and coupled environments (Fig. 2A and B.) The parameter switches between a coupled environment for and an independent environment for (see Materials and Methods for details.). Structure uncertainty is captured by a Bernoulli distribution on with parameter , which will change solely based on the rewards observed.
Without uncertainty, the optimal decision-making strategies for both the independent and coupled environments are well-known and relatively simple. The optimal policy for a coupled environment is purely exploitative—it simply chooses the option with the greater number of successes (including failures of the other option as successes) because the reward observed in one option tells us everything the reward that would have been received in the other option. Optimal action selection for an independent environment, however, involves balancing the exploration–exploitation trade-off. Exploration is required because choosing one option does not provide information about the other. The optimal policy for an independent environment involves computing a quality index for each option, called the Gittins index [18], and selecting the highest quality option. The Gittins index computes the maximum expected reward per unit discounted time for each option, and is the result of the following optimization problem:With uncertainty, optimal action selection depends on the belief that the environment is coupled, as captured by the parameter . In the methods section, we show that the optimal policy for structure learning can be expressed as a mixture of the optimal policies for the independent and coupled environments. For all the models, the optimal policy is a function of the observed counts of successes, , and failures, , for each option, and priors .
To illustrate the behavior of the structure learning model, we expose the model to a sequence of tasks. The model is placed in either a coupled or independent environment (Fig. 3A & B). Every 50 trials the reward probabilities on the options are randomly reset, but the type of environment stays fixed. For both environments, the structure learning model learns the environment type, as expressed by the convergence of the posterior distribution on the parameter to its true value. For the parameters and , the marginal probability is indicated by the color, with brighter indicating higher relative probability mass. The structure learning model quickly learns in both environments, although it is frequently easier to detect an independent environment—whenever both options are significantly above or below chance, the coupled structure can be quickly ruled out. Once there is high certainty on the structure ( or , where is the data), beliefs are concentrated on the parameters that matter for that structure— and becomes concentrated on the reward probabilities of each option in the independent environment, and becomes uniform in the coupled environment.
The effect of structure uncertainty on the behavior of the structure learning model is evident by looking at the expected reward. For action , this expected reward iswhere is the posterior probability on the structure given the data represented by the counts , , and . If the probability that the structure is coupled is high (), then the expected reward accrues regardless of which action is chosen. If the probability that the structure is independent () is high, then the expected reward depends only on the option chosen. Thus the belief about coupling gates the need for exploration. In an independent model, there is a value attached to choosing the option with less evidence even if the current evidence suggests it has a lower probability of success. The expected reward for action is similarlyIn Fig. 4, we perform a simulation that shows how the structure learning model described can behave as a independent or coupled model depending on the uncertainty about coupling belief. We purposely chose evidence values for which the independent model would pick one option while the coupled model would pick the other. When a curve dips below 0, it means that the learning model would choose option 1, and when it does above 0, it would pick option 2. Note that the structure learning model can sometimes behave as a coupled or independent model depending on the uncertainty about the structure. This difference between the structure learning model vs. fixed models will play an important role later when we show that people change their policy in accord with structure learning.
To quantify structure learning in participant's decisions, we compared the predictions of the structure learning model with models that capture the decisions expected from knowledge of structure in the absence of learning (fixed independent and coupled structure). Additionally, we used Q-learning algorithm [19] with a soft-max action selection [20] as a base model. Q-learning is a model-free RL method that does not model the reward probabilities or structure, rather it estimates the value of an action by compiling over experienced outcomes (called Temporal Difference learning). However, Q-learning does not balance exploration and exploitation in a principled way, but rather performs heuristic explorations based on random actions. It is proven to estimate the optimal value of an action after infinitely many observations for every action and state [19]. The temporal difference aspect of Q-learning as well as the exploratory interpretation of the soft-max rule have been shown to correlate with brain activity [4], [21], [22].
Fitting the models to all the response data, we find that the structure learning model prediction rate () is better than the coupled model prediction rate (), exact binomial test , better than the fixed independent model prediction rate (), , and better than Q-learning model (), ). Note that the Bayesian models have no free parameters, with the exception of the initial value of the prior belief about coupling structure for the structure learning model, which is quickly swamped by the evidence. However, we allowed for individual differences in all five parameters of the Q-learning model. For all models, we assumed uniform priors on probabilities of reward (, , at the beginning of tasks).
The remainder of the results are organized as follows. Because essentially all models predict well a large number of trials that occur later in blocks (where evidence is high and the better option is easy to identify), we focus on the set of trials for which there is at least one disagreement between the models so that we can better tell them apart. We call this set of trials diagnostic. We show the structure learning model can better account for several aspects of decision-making on diagnostic trials. In particular, we show how uncertainty in task structure tracks qualitative and quantitative changes in choice behavior. Then we show that the structure learning model gives a principled explanation for strategies that appear suboptimal. Finally, we analyze decisions that are specifically diagnostic for the structure learning model (structure learning predicts differently than fixed models) and show that the structure learning model predicts human choice behavior better than models with fixed structure.
We have provided evidence that structure learning may be an important missing piece in evaluating human sequential decision making. The idea of modeling sequential decision making under uncertainty as a structure learning problem is a natural extension of previous work on structure learning in models of cognition [13], [16] (also see [32]), animal learning [33] and motor control (e.g., see [34]). It also extends previous work on Bayesian approaches to modeling sequential decision making in the Multi-armed bandits [35] by adding structure learning. It is important to note that we have intentionally focused on reward structure, ignoring issues involving dependencies across trials. Clearly reward structure learning must be integrated with learning about temporal dependencies [36] (e.g. assumptions of a non-stationary environment [5], [37], [38]).
Interestingly, there were a set of participants' decisions that none of the models were able to capture and that constitute 9.4% of the data. These trials are predominately localized on positive evidence (Eq. 4), but negative confidence (Eq. 5) levels (see Fig. 5A and B, left panel, people column.). These choices corresponded to persisting in choosing the worst option despite statistical evidence supporting the better option. None of the models considered would choose the worse option under these conditions. Participants may have limited memory or may be considering a larger space of possible models; for example nonconstant reward rates (allowing for nonstationary reward probabilities).
Although we focused on learning coupling between options, there are other kinds of reward structure learning that may account for a broad variety of human decision making performance. In particular, allowing dependence between the probability of reward at a site and previous actions can produce large changes in decision making behavior. For example, in a “foraging” model where reward is collected from a site and probabilistically replenished, optimal strategies will produce choice sequences that alternate between reward sites [39]. Thus uncertainty about the independence of reward on previous actions can produce a continuum of behavior, from maximization to probability matching. Note that structure learning explanations for probability matching are significantly different than explanations based on reinforcing previously successful actions (the “law of effect”) [40]. Instead of explaining behavior in terms of the idiosyncrasies of a learning rule, structure learning constitutes a fully rational response to uncertainty about the causal structure of rewards in the environment. By expanding the range of normative hypotheses for human decision-making, we believe we can begin to develop more principled accounts of human sequential decision-making.
The general alternative to the rational approach is to assume that choice behavior reflects some fundamental limitations in sensing, neural computation or storage. It is possible that the decisions we could not predict in any dependent environment result from human processing limitations. For example, one of the key decision patterns that does not fit in the normative approach is choice stickiness, a persistence in choosing the same option despite evidence suggesting it would be better to switch. This could reflect a transition to model-free learning in the independent environment. Participants may have learned a policy for choosing that option based on early reward evidence. However, we find no evidence for this possibility in our data. Another possibility is that participants have memory limitations that prevent them from compiling all of the evidence [35]—the observed persistence may be sensitivity to local reward. While limitations to human decision-making surely exist, and people are bounded rational, our results provide evidence that decisions are also driven by sophisticated structure learning. We believe that many aspects of human decision-making that appears mysterious may be the result of the brain's attempts to acquire compact and useful representations of the structure of its environment.
We foresee an adoption of more sophisticated models of sequential decision-making to account for the compact representation that humans might be using to act in diverse reward structures. While we believe that the theory to analyze these representations is available, it has only been cautiously adopted in Psychology and Neuroscience [35], [41]–[43]. We have already seen this pattern of adoption occur in Artificial Intelligence where the development of efficient computational methods to solve Bellman's equation (i.e. model-free RL methods like Q-learning) led to the rapid development and application of RL methods starting in the 1980s, despite the fact that the theoretical foundations had been laid by control theorist more than two decades prior [1], [44], [45]. While Robotics, for example, today hardly uses model-free reinforcement learning to think about tasks of any level of complexity, much work remains for model-based reinforcement learning to make its way into mainstream human and animal sequential decision-making analysis.
Informed consent was obtained and all investigations were conducted according to the principles expressed in the Declaration of Helsinki, under the Assurance of Compliance number FWA00000312.
Sixteen volunteers solve 32 bandit tasks, 16 for each environment. The probabilities of rewards were randomly sampled from a uniform distribution, and the stopping times for each bandit task were sampled from a Geometric distribution . The average stopping time was 48. The order of the tasks within an environment was randomized, and the order of presentation of the environments was randomized as well. All subjects were exposed to the same probabilities of rewards and stopping times.
Each option is shown in the screen as a slot machine. Subjects pull a machine by pressing a key in the keyboard. When pulled, an animation of the lever is shown, 200 msec later the reward appears in the machine's screen, and a sound mimicking dropping coins lasts proportionally to the amount gathered. We provide several cues, some redundant, to help subjects keep track of previous rewards. We display the number of pulls, total reward, and the current average reward per pull. Reward magnitudes were 0 or 100 points. The machine's screen changes the color according to the average reward, from red (zero points), through yellow (fifty points), and green (one hundred points). The machine's total reward is shown as a pile of coins underneath it. The total score, total pulls, and rankings within a game were presented.
All participants finished all tasks. Each participant performed 1194 trials on independent environment and 925 on the coupled environment, for a total of 33904 trials. In general, participants understood the task well. No apparent outliers were found nor missed trials.
The language of graphical models provides a useful framework for describing the possible structure of rewards in the environment. Consider an environment with several distinct reward sites that can be sampled, but the way models generate these rewards is unknown. In particular, rewards at each site may be independent, or there may be a latent cause which accounts for the presence of rewards at both sites. Uncertainty about which reward model is correct naturally produces a mixture as the appropriate learning model. This structure learning model is a special case of Bayesian Reinforcement Learning (BRL), where the states of the environment are the reward sites and the transitions between states are determined by the action of sampling a reward site. Uncertainty about reward dynamics and contingencies can be modeled by including within the belief state not only reward probabilities, but also the possibility of independent or coupled rewards. Then, the optimal balance of exploration and exploitation in BRL results in action selection that seeks to maximize (1) expected rewards (2) information about rewards dynamics, and (3) information about task structure.
The belief over dynamics is effectively a probability distribution over possible Markov Decision Processes that would explain observables. As such, the optimal policy can be described as a mapping from belief states to actions. In principle, the optimal solution can be found by solving Bellman optimality equations but generally there are countably or uncountably infinitely many states and solutions need approximations. If we were certain which of the two models were right, the action selection problem has known solution for both cases, presented below.
|
10.1371/journal.ppat.1007413 | Hypermutation-induced in vivo oxidative stress resistance enhances Vibrio cholerae host adaptation | Bacterial pathogens are highly adaptable organisms, a quality that enables them to overcome changing hostile environments. For example, Vibrio cholerae, the causative agent of cholera, is able to colonize host small intestines and combat host-produced reactive oxygen species (ROS) during infection. To dissect the molecular mechanisms utilized by V. cholerae to overcome ROS in vivo, we performed a whole-genome transposon sequencing analysis (Tn-seq) by comparing gene requirements for colonization using adult mice with and without the treatment of the antioxidant, N-acetyl cysteine. We found that mutants of the methyl-directed mismatch repair (MMR) system, such as MutS, displayed significant colonization advantages in untreated, ROS-rich mice, but not in NAC-treated mice. Further analyses suggest that the accumulation of both catalase-overproducing mutants and rugose colony variants in NAC- mice was the leading cause of mutS mutant enrichment caused by oxidative stress during infection. We also found that rugose variants could revert back to smooth colonies upon aerobic, in vitro culture. Additionally, the mutation rate of wildtype colonized in NAC- mice was significantly higher than that in NAC+ mice. Taken together, these findings support a paradigm in which V. cholerae employs a temporal adaptive strategy to battle ROS during infection, resulting in enriched phenotypes. Moreover, ΔmutS passage and complementation can be used to model hypermuation in diverse pathogens to identify novel stress resistance mechanisms.
| Cholera is a devastating diarrheal disease that is still endemic to many developing nations, with the worst outbreak in history having occurred recently in Yemen. Vibrio cholerae, the causative agent of cholera, transitions from aquatic reservoirs to the human gastrointestinal tract, where it expresses virulence factors to facilitate colonization of the small intestines and to combat host innate immune effectors, such as reactive oxygen species (ROS). We applied a genome-wide transposon screen (Tn-seq) and identified that deletion of mutS, which is part of DNA mismatch repair system, drastically increased colonization in ROS-rich mice. The deletion of mutS led to the accumulation of catalase-overproducing mutants and a high frequency rugose phenotype when exposed to ROS selective pressures in vivo. Additionally, ROS elevated mutation frequency in wildtype, both in vitro and in vivo. Our data imply that V. cholerae may modulate mutation frequency as a temporal adaptive strategy to overcome oxidative stress and to enhance infectivity.
| Vibrio cholerae, the etiological agent of the pandemic disease cholera, resides in aquatic environments and can also colonize human intestines following ingestion of contaminated food and water. In order to survive in both aquatic and host environments, V. cholerae has the ability to cope with harsh conditions during the transition to the host gut and during subsequent growth [1]. For example, upon infection, V. cholerae senses host signals and is able to coordinate both virulence gene activation and repression to evade host defenses and successfully colonize intestines [2–5]. Late in the infection, V. cholerae also optimally modulates its genetic programs for the forthcoming dissemination into the aquatic environment [6, 7] where it is often associated with abiotic or biotic surfaces such as phytoplankton and zooplankton. These associations enable the formation of biofilms, which provide protection from a number of environmental stresses; including nutrient limitation, protozoa predation, and bacteriophage infection [8]. Additionally, biofilms may enhance infectivity due to their acid-resistant properties and higher growth rate during infection [9, 10].
One of the major stresses V. cholerae must overcome is exposure to reactive radical species. Reactive compounds, including reactive oxygen species (ROS), are abundant in marine systems [11]. V. cholerae also encounters oxidative stress during the later stages of infection, as demonstrated by an increase in ROS levels and a decrease in the levels of host antioxidant enzymes during V. cholerae-induced diarrhea [12, 13]. It has been previously demonstrated that catalases (KatG and KatB), peroxiredoxin (PrxA), organic hydroperoxide resistance protein (OhrA), a redox-regulated chaperone (Hsp33), and a DNA-binding protein from starved cells (DPS) are important for V. cholerae ROS resistance [14–17]. ROS resistance in V. cholerae is known to be tightly regulated through a variety of mechanisms. OxyR is required to activate catalase genes and dps, and is modulated by another OxyR homolog, OxyR2 [14, 16, 18]. Quorum sensing systems [19], PhoB/PhoR two-component systems [20], and the virulence regulator, AphB, also play important roles in oxidative stress response [21]. Further identifying bacterial stress responses to host-derived ROS is important for understanding V. cholerae pathogenesis.
In this study, we used Tn-seq to screen for V. cholerae genes that are involved in ROS resistance during infection. By comparing colonization in control mice to mice treated with antioxidant N-acetyl cysteine (NAC) that reduces the production of ROS in murine intestines [15], we found that deletion of mutS, encoding a key component in the DNA methyl-directed mismatch repair (MMR) system, results in a significant colonization advantage compared to wildtype in ROS-rich mice. The MMR system is highly conserved from bacteria to humans and is critical for maintaining the overall stability of the genetic material [22]. Mutations in this pathway lead to hypermutation rates across the genome. It has been shown that inactivation of the MMR system of various bacterial pathogens, such as Escherichia coli, Salmonella enterica serovar Typhimurium, and Pseudomonas aeruginosa leads to better adaptation and persistence of these pathogens in murine models [23–26]. It has been proposed that under certain stressful conditions, hypermutators are selected in the total population by hitchhiking with the adaptive mutations that they produce. However, the mechanism(s) by which hypermutators become better persistors is less clear. In this work, we developed a strategy to study bacterial temporal hypermutation in vivo and found that mutations resulting in increased catalase production and increased biofilm formation, demonstrated by rugose colony phenotypes, may lead V. cholerae hypermutators to display colonization advantages in ROS-rich mouse intestines.
To investigate V. cholerae genes involved in ROS resistance during colonization, we performed a Tn-seq screen in a streptomycin-treated adult mouse model, in which bacteria experience host-generated oxidative and nitrosative stress [15, 27, 28]. As a comparison, we also treated a set of mice with N-acetyl cysteine (NAC), an antioxidant widely used in human and animal studies to artificially reduce ROS levels [29, 30]. Previously we have shown that NAC significantly reduces the production of ROS related biomarkers in mice [15]. We mutagenized V. cholerae with a Tn5 transposon and inoculated the Tn5 library into adult mice with NAC treatment as a variable. At the 3-day post-infection (PI) time point, passaged mutants were recovered from fecal pellets. We then extracted bacterial DNA and used Illumina sequencing [6] to determine the number of transposon insertions in the input and output mutant libraries. We compared the output/input ratios of mutants colonized in NAC-treated mice (NAC+ mice) to mice without NAC treatment (NAC- mice) (Fig 1A). Several mutations that have Tn insertions in previously-known genes required for ROS resistance were found colonizing poorly in NAC- mice but not in NAC+ mice (S1 Data), validating the NAC treatment and suggesting that these genes are important for overcoming ROS in vivo. These genes include prxA (VC2637)[14], ohrA (VCA1006)[15], dps (VC0139)[16], and rpoS (VC0534)[21]. In addition, we identified iron transport systems (VC0776-VC0780, VC1264), efflux pumps (VC0629, VC1410, VC1675, VC2761, VCA0183, VCA0267), and a number of transcriptional regulators (such as VC0068, VC2301, VCA0182) that are important for colonizing in NAC- mice (S1 Data). These genes are subject for independent confirmation and further investigation.
Interestingly, the Tn-seq screen revealed that a number of mutations are highly enriched in NAC-mice but not in NAC+ mice (Fig 1A), suggesting that mutants containing disruptions in these genes have colonization advantages in ROS-rich intestines. Among them, several mutations in DNA methyl-directed mismatch repair (MMR) pathways displayed significantly higher number of reads in the pools isolated from NAC- mice than those of NAC+ mice (Fig 1A). MMR is highly conserved in all organisms and repairs mispaired bases in DNA generated by replication errors [22]. In E. coli, MutS recognizes mispairs and coordinates with MutL and MutH to direct excision of the newly synthesized DNA strand [31](Fig 1B). We found that the reads of insertions in mutS, mutL, and mutH from NAC- mice were all higher when compared to NAC+ mice, whereas reads of insertions in the downstream MMR pathway (uvrD, recJ and dinB) were similar between these two conditions (Fig 1C). It has been reported that UvrD, RecJ, and DinB play less critical roles in bacterial DNA repair than MutSLH [22, 32]. We confirmed that in V. cholerae, deletion of dinB did not affect colonization, nor spontaneous mutation frequency (Fig A in S1 Text). Therefore, in this study, we selected MutS for further investigation to decipher the possible role of hypermutation on ROS resistance. Of note, the Tn-seq screen also revealed that other mutations are significantly enriched in NAC- but not in NAC+ mice. These mutations included genes in the flagellar biosynthesis pathway (VC2120-VC2134) and the MSHA pilin biogenesis pathway (VC0398-VC0411)(S1 Data). The mechanisms are subjected to another study, but we speculate that since both flagella and MSHA pilins activate host innate immunity [2, 33, 34], which is activated by reactive oxygen species synergistically [35, 36], deletion in flagellar synthesis or MSHA synthesis may therefore have localized colonization advantages. Removing ROS in the gut abolishes the advantage of these mutants.
To confirm the Tn-seq results, we constructed an in-frame deletion of mutS. We first compared spontaneous rifampicin resistance by colony enumeration of the ΔmutS mutant with that of wildtype as a proxy for mutation frequency. As predicted, the mutation frequency in ΔmutS mutants was approximately 100-fold higher than that in wildtype (Fig 2A). Complementation of mutS on a plasmid restored the mutS mutation frequency to wildtype levels (Fig 2A, grey bar). We then performed a competition experiment by mixing differentially-labeled wildtype and ΔmutS mutants in a 1:1 ratio and inoculated them into streptomycin-treated mice with NAC treatment as a variable. Fecal pellets were collected daily and colony forming units (CFU) of wildtype and ΔmutS mutants were determined by serial dilution and colony enumeration on selective LB agar plates. Fig 2B shows that in the NAC- mice, ΔmutS mutants colonized similarly to wildtype initially, but outcompeted wildtype later in the infection. At day 6, the competitive index (ΔmutS/WT) exceeded a 1,000-fold advantage. On the other hand, in the NAC+ mice, ΔmutS mutants did not display a colonization advantage over wildtype throughout the course of infection (Fig 2C). Note that the total number of colonized bacteria was similar between different infection time, mice, and conditions. These data confirm the Tn-seq study suggesting that ΔmutS mutants are advantageous over wildtype in NAC- mice, which is predicted to have relatively higher levels of ROS compared to NAC+ treated mice.
To investigate the possible mechanisms that enable a ΔmutS colonization advantage in NAC- mice, we further examined these isolates in vitro and in vivo. To avoid additional accumulation of mutations after in vivo passage, we introduced a copy of mutS into the lacZ locus of mutS mutants immediately after being isolated from mice. Introducing the chromosomal copy of mutS into mutS mutants restored the mutation frequency to wildtype levels (Fig B in S1 Text). We then tested 24 mutS (lacZ::mutS) isolates (annotated as ΔmutS*) from NAC- mice. We first performed competition colonization experiments to examine whether these individual isolates maintain colonization advantages over wildtype. We found that all 24 ΔmutS* tested colonized NAC- mice better than wildtype and the competitive indexes ranged from ~10–1000 (Fig 3A, light green squares). In NAC+ mice, these isolates gained little, if any, competitive advantage. As a control, we also tested 5 wildtype isolates (WT*) that were passaged through NAC- mice. These isolates colonized at a comparable level to the wildtype parental strain in both types of mice (Fig 3A, orange triangles). These data suggest that the ΔmutS competitive advantage in ROS-rich mice is heritable.
We then measured ROS resistance of these ΔmutS* isolates in vitro. Parental ΔmutS mutants had a similar in vitro growth rate as wildtype in LB medium and AKI virulence-inducing medium [37](Fig C in S1 Text). WT* and ΔmutS* also grew similarly under these conditions (Fig C (C) in S1 Text). When cultured in LB until mid-log and then treated with H2O2, we found that ΔmutS had similar ROS resistance as that of wildtype (Fig 3B). However, approximately half of ΔmutS* isolates tested displayed significantly more resilience to H2O2 exposure than that of parental ΔmutS (Fig 3B circles, one-way ANOVA P value = 0.0005), whereas WT* were similar to the parental wildtype strain (Fig 3B triangles, P value > 0.99). Of note, most of those ΔmutS* isolates that did not produce more catalase displayed different colony morphology (Fig 3B and 3C, squares) (see next section). Correspondingly, about half of ΔmutS* were detected to have more catalase activity (Fig 3C, circles, one-way ANOVA P value = 0.0074). The mutations that led to overproduction of catalase in these ΔmutS mutants were not determined. We selected five such high-catalase-producing ΔmutS* isolates and examined transcription of catalase genes (katG and katB)[14] induced by H2O2 using qPCR and found transcription of both catalase genes was elevated in three of these mutants (Fig D in S1 Text). For the other two ΔmutS* isolates that did not displayed increasing catalase gene expression, it is possible that mutations involved in post-transcriptional regulation of KatGB activity are accumulated in these isolates. Taken together, these data suggest that mutations leading to increased catalase production are a contributing factor to the observed colonization advantage gained by ΔmutS during colonization in NAC- mice. To test this hypothesis, we deleted two catalase genes katG and katB [14] in ΔmutS and the resulting strain was competed with wildtype in NAC- mice. We found that deletion of katG and katB in ΔmutS mutants reduced colonization advantage of ΔmutS mutants significantly (Fig E (A) in S1 Text). To further confirm the importance of ROS resistance for V. cholerae in vivo, we examined the colonization of ΔoxyR mutants in NAC- mice. OxyR activates a number of ROS resistance genes in V. cholerae [14, 16]. Fig E (B) in S1 Text shows that ΔoxyR mutants colonized poorly in this mouse model. These results suggest that ROS is important for V. cholerae colonization of NAC- mice.
Upon enumeration of bacteria from fecal mouse pellets, an unusually high number of rugose (wrinkled) colonies, originating from smooth ΔmutS mutants, were observed on LB plates (Fig 4A). It has been reported that V. cholerae can switch its colony morphology from smooth to rugose phenotypes due to the overproduction of exopolysaccharide. This phenotypic switch is reversible and confers greater resistance to environmental stresses compared to strains that undergo this transition at low frequency [38–40]. We thus determined the frequency of rugose colony formation in wildtype and ΔmutS isolates from NAC- and NAC+ mice. Fig 4B shows that from NAC- mice, a significant number of output ΔmutS colonies displayed the rugose phenotype, ranging from ~5% to ~30% of total colonies isolated form each mouse. In NAC+ mice, however, the percentage of rugose colonies recovered from ΔmutS mutants was much lower (Fig 4B, blue circles). As for wildtype that were isolated from either NAC- or NAC+ mice, a relatively low number of colonies displayed the rugose phenotype (Fig 4B, squares). These data suggest that the lack of a functional DNA repair system may increase the frequency of rugose colony formation, which may lead to enhanced survival in ROS-rich, in vivo environments. Interestingly, when the rugose variants were cultured in liquid LB with aeration, a majority of them reversed to smooth colonies with high reversion rates (Fig 4C, left panel). However, if incubated anaerobically, which mimics the in vivo growth condition, the reversion rates were less prominent as compared to aerobic incubation (Fig 4C, right panel), implying that anaerobiosis may be one of the in vivo selective pressures that promote rugose colony formation. These data suggest the involvement of temporal phenotypic switches during V. cholerae infection possibly mediated or enhanced by genetic adaptation.
To determine whether rugose colony phenotypes contribute to enhanced survival, we performed in vitro experiments to investigate the possible role of these variants in ROS resistance. We found that a majority of these rugose ΔmutS* variants did not display more ROS resistance in liquid cultures (Fig 3B, squares) and did not display increased catalase production compared to wildtype (Fig 3C, squares). The rugose colony phenotype is often the result of the overproduction of exopolysaccharides, a major component of the biofilm matrix [38, 41]. To examine whether exopolysaccharide overproduction is the cause of rugose colony formation in ΔmutS* isolates, we measured the biofilm formation capacity of various isolates. We found that biofilm mass formed by smooth variants of ΔmutS* was similar to that of wildtype and ΔmutS parental strains, whereas rugose variants displayed an increased biofilm formation capacity (Fig 5A). We thus hypothesized that rugose variants are enriched in ROS-rich intestines due to their increased biofilm production and predict that biofilm-associated cells are more resistant to ROS exposure. To test this prediction, we assessed the viability of planktonic and biofilm-associated cells after exposure to organic and inorganic oxidants (Fig 5B). Biofilms were formed on glass test tubes at the air-broth interface through static culture. The majority of planktonic cells were killed after exposure to 1 mM H2O2 or 100 μM cumene hydroperoxide exposure for 60 mins. In contrast, biofilm-associated cells displayed more than a 30-fold increase in resistance to ROS than planktonic cells (Fig 5B). ROS resistance was mostly eliminated when biofilm structures were disrupted by vortexing with glass beads prior to ROS exposure (Fig 5B, grey bars). These results indicate that it is primarily the physical structure of the biofilm that confers protection against ROS, rather than increased ROS resistance in individual cells. Taken together, our results imply that biofilm formation in vivo may play a role in ROS resistance.
Rugose colonies are often caused by mutations in the quorum sensing master regulator HapR and many clinically-isolated rugose variants harbor loss-of-function hapR mutations [9, 19]. We examined possible disruptions of the quorum sensing pathway in the rugose ΔmutS* mutants we isolated and found that they were similar to wildtype (Fig F (A) in S1 Text). Sequencing analysis did not reveal any mutations in the hapR locus. Indeed, although ΔhapR mutants form thicker biofilms [9, 19], ΔhapR displayed colonization defects in both NAC- and NAC+ mice (Fig F (B) in S1 Text), suggesting that HapR may regulate other targets that are involved in adult mouse colonization. To test our hypothesis that biofilm formation is important for in vivo ROS resistance, we first performed in vivo competition experiments using ΔmutS* rugose variants and their corresponding smooth revertants. We selected spontaneous rifampicin resistant smooth revertants in order to distinguish them with their parental rugose strains. Fig 5C shows that all three smooth revertants displayed different degrees of colonization disadvantage over their parental rugose ΔmutS* in NAC- mice. The competitive indexes of these rugose/smooth variants were comparable to those indexes when these rugose ΔmutS* isolates competed with wildtype (Fig 3A), suggesting that increasing biofilm formation is the main factor in the rugose isolates that promotes ROS resistance in vivo. To further confirm this, we then deleted vpsA, which encodes the major component of the Vibrio polysaccharide biosynthesis pathway [42], in ΔmutS and the resulting strain was competed with wildtype in NAC- mice. We found that abolishing biofilm formation capacity in ΔmutS mutants reduced the colonization advantage of ΔmutS mutants significantly (Fig 5D). These data again suggest that biofilm formation in vivo may play a role in ROS resistance. Of note, ΔmutS/ΔvpsA still outcompeted wildtype. It is possible that accumulation of other beneficial mutations, such as those enhancing catalase production, may elevate ROS resistance in vivo for ΔmutS/ΔvpsA mutants.
Mutations in DNA repair systems greatly increase mutation rates in bacteria, as shown by this and other studies, and it has also been reported that ROS enhances mutation frequency in bacteria [43–45]. We then sought to examine whether V. cholerae may display distinct mutation frequencies as a function of in vivo ROS exposure. Both wildtype and ΔmutS mutants were inoculated into mice with and without NAC treatment as done in previous experiments. After 3 days of colonization, we collected fecal pellets and outgrew V. cholerae in LB medium for 12 hrs. We then plated these bacteria on rifampicin to determine mutation rate through a gain of function mutation in rpoB that confers resistance to rifampicin. We determined that for wildtype V. cholerae colonized in NAC- mice, the mutation frequency was over 30-fold higher than those in NAC+ mice (Fig 6A). For ΔmutS mutants, as expected, the mutation frequency in vivo was high, but there was no significant difference between colonizers in NAC- and NAC+ mice (Fig 6A), suggesting a theoretical limit of in vivo mutagenesis or that the observed elevation in mutation frequency caused by ROS is mediated by a reduction in MMR activity. We also determined the in vitro mutation rate in the presence of ROS. Upon exposure to higher levels of H2O2, elevated mutation frequency was detected in wildtype, whereas changes in mutation rate in ΔmutS mutants had no statistical significance (Fig 6B). These data suggest that ROS enhances mutation rate for V. cholerae in both in vitro and in vivo environments. This stress-induced mutagenesis and resulting increased genetic variability may provide additional means for V. cholerae to adapt to ROS-rich environments.
Bacterial pathogens are constantly confronted with changing and aggravating environments and have been known to leverage genetic adaptation as a means to overcome challenges faced in these environments. In this study, we used a streptomycin-treated mouse model to study V. cholerae ROS resistance in vivo. Bacteria experience host-generated oxidative stress in the streptomycin-treated adult mouse model [15, 27, 28]. Inclusion of antioxidant N-acetyl cysteine (NAC) significantly reduced ROS levels [15](Fig G in S1 Text). For mice without the streptomycin treatment, ROS levels were lower than streptomycin-treated mice, but remained detectable (Fig G (A) in S1 Text). In addition, it has been reported that during choleric diarrhea, ROS levels were increased in the host [12, 13]. Taken together, it is suggestive that ROS stress encountered by V. cholerae in the streptomycin-treated mouse model may be physiologically relevant. By the Tn-seq screen, we discovered that hypermutation rates resulting from the impairment of the V. cholerae mismatch repair system (ΔmutS) led to a colonization advantage in mice, which was not observed in NAC-treated mice. E. coli colonization studies of mouse intestines have shown that hypermutation is initially beneficial because it allows for a rapid adaptation to the mouse gut environment [26]. However, such strains then experience a loss of fitness due to the constant accumulation of detrimental mutations. To prevent additional detrimental mutations and to be able to study those mutations that conferred a colonization advantage in vivo, we complemented ΔmutS isolates from NAC- mice immediately after isolation. Further study shows that passage of ΔmutS through NAC- mice resulted in the enrichment of catalase-overproducing isolates and a high frequency rugose phenotype. These ΔmutS* isolates remained super-colonizers in NAC- mice but did not gain advantages in NAC+ mice (Fig 3A). We also examined infant mouse colonization (Fig H (A) in S1 Text) as well as virulence gene expression (Fig H (B&C) in S1 Text) and found that compared to wildtype, some ΔmutS* isolates were defective in infant mouse colonization and virulence factor production. These results suggest that mutations are specifically selected to overcome ROS stress in the NAC- mice. Indeed, in a previous report [45] by Davies, et al., it was observed that V. cholerae ΔmutS mutants displayed an approximately 5-fold defect in infant mouse colonization. Considering the short incubation time in infant mouse colonization (18 hrs) and the speculated lack of inflammation in infant mouse gut, it is possible that ΔmutS mutants do not experience the same selective pressures as in ROS-rich adult mice. Similarly, in P. aeruginosa, ΔmutS mutants are attenuated in a mouse model of acute infections but are favored in long term persistence of oropharyngeal colonization in cystic fibrosis mice [25].
Many hyper-mutational bacterial pathogens are frequently identified from clinical and environmental isolates, including E. coli, Salmonella, P. aeruginosa, Haemophilus influenzae, Neisseria meningitidis, and Streptococcus pneumoniae [46]. This is often the case when bacteria need to adapt a new stressful environment. For example, a high percentage of mutators of P. aeruginosa, H. influenzae, and S. aureus were isolated from cystic fibrosis patients who received antibiotic treatments [47]. Infection of a mammalian host is certainly another new environment to adapt to and an increase in genetic variability can help to cope with host defense systems [48]. V. cholerae hypermutators have also been found in clinical isolates. In a recent study [49], Didelot et al. reported that among 260 V. cholerae genomes they sequenced and analyzed, 17 isolates have unusually high number of SNPs that are evenly spread throughout their genomes. Further analysis shows that 14 of these 17 genomes possess genetic variations in one or more of four genes in the MMR system and the mutation rate of these strains are significantly increased compared to the others. Interestingly, the majority of these hypermutator strains were isolated between 1961 and 1965, relatively soon after the beginning of the seventh pandemic. The authors cautiously speculated that hypermutators might be causally associated with the rapid spread of the seventh pandemic. In addition, a mobile element is found to insert into the mutS gene of a marine Vibrio species, providing a new mechanism for altering the mutation rate [50].
Hypermutation may promote adaptive evolution for bacteria, but the high mutation rate comes at a cost in fitness in the long term [26]. It has been proposed that bacteria may transiently modulate their mutation rates to balance the trade-off between adaption and the accumulation of detrimental mutations [51]. For example, the expression of mutS is downregulated by RpoS in response to antibiotic stress, which increases the mutation rate in several bacterial species including V. cholerae [52]. In Streptococcus pyogenes, the integration and excision of a prophage inserted between mutS and mutL causes a reversible increase in mutation rate in response to the environmental stress [53]. We found that in wildtype V. cholerae, mutation rate was significantly increased when colonizing NAC- mice compared to NAC+ mice (Fig 6A). This finding suggests that V. cholerae might utilize increased mutation rates as a temporal strategy for adopting advantageous phenotypes during infection of a ROS-rich host. Interestingly, our Tn-seq screen (S1 Data) revealed that rpoS mutants failed to colonize NAC- mice but not NAC+ mice. Further testing is required to investigate whether the observed increase in mutation rate in NAC- mice is RpoS-mediated. Moreover, the rugose variants isolated from NAC- mice could reverse to smooth colonies in vitro (Fig 4C), indicative of use of temporal adaptive strategies by V. cholerae to combat ROS during infection. Of note, the mechanism of V. cholerae smooth-rugose phase variation is not clear, but DNA repair pathways have been implicated in phase variation in several species, including Neisseria gonorrhoeae, V. parahaemolyticus, and Pseudomonas sp.[54–57]. Interestingly, the high reversion rate from rugose to smooth colonies under aerobic growth (Fig 4C) occurred in mutS complemented background (ΔmutS*), suggesting that reversion is not due to DNA mutation. Cell variants in some bacterial species are generated without the burden of mutation, but rather from reverse biostability, which can be controlled by genetic mechanisms such as DNA rearrangement or epigenetic mechanisms such as DNA methylation [58, 59]. Alternatively, the rapid reversion from rugose to smooth in vitro even though these cells have been repaired for mutS may simply because that the selective pressure for reversion to the smooth variant is remarkably strong during aerobic growth and therefore the reverting mutations arise rapidly. The exact mechanisms of O2-dependent rugose-to-smooth phenotypic switch is currently under investigation.
Hypermutable strains are often associated with higher incidences of antibiotic resistance than strains with lower mutations rates. This study proposes a model of in vivo temporal hypermutation by mutating MMR and complementing mutants with functional MMR after isolation. This approach allowed for the identification of ROS resistance mechanisms that could be genetically upregulated under ROS stress. It is likely that this approach could be utilized in the context of distinct stressors such as low pH, desiccation, nitrosative stress, etc., revealing likely mechanisms used to overcome those specific environments by comparing mutation spectra or phenotypic changes between experimental groups. In reverse order, MMR mutants could also be used to shed light on stresses experienced in undefined environments by bacteria by associating enriched pathways with stressors. Insight into the mechanisms used to overcome specific stressors could be used to refine antibacterial strategies. This insight would allow for the proactive targeting of arising mutators under treatment, preventing resistant lineages. This application could improve the efficacy of antibacterial agents and reduce the incidence of resistant mutators.
All animal experiments were carried out in strict accordance with the animal protocols that were approved by the Ethical Committee of Animal Experiments of Nanjing Agricultural University (Permit Number: SYXK (Su) 2017–0007). All efforts were made to minimize animal suffering. Euthanasia was performed by CO2 inhale.
V. cholerae El Tor C6706 [60] was used as a parental strain in this study, and was propagated in LB media containing appropriate antibiotics at 37°C, unless otherwise noted. The mutS and dinB in-frame deletions were constructed by cloning the regions flanking mutS or dinB into the suicide vector pWM91 containing a sacB counter-selectable marker [61]. The resulting plasmids were introduced into V. cholerae by conjugation and deletion mutants were selected for double homologous recombination events. The construction of hapR, katG, katB, oxyR, and vpsA mutants has been described previously [14, 19, 62]. The mutS overexpression plasmid was constructed by cloning mutS coding sequences downstream of the lac promoter in pBBR-MCS-3 [63]. Chromosomal complementation of mutS was constructed by inserting mutS into the lacZ locus using pJL1 [64]. AKI medium was used to induce virulence gene expression [37]. Transcriptional lux reporters of promoter regions of tcpA have been described previously [64]. For growth of oxyR mutants on LB plates, 10 μg/ml catalase from bovine liver was included in the medium. When necessary, rugose variants were propagated in LB media without shaking to avoid smooth revertants.
The streptomycin-treated adult mouse model was used to examine V. cholerae ROS resistance in vivo as previously described [15, 27] with the following modifications. Six-week-old CD-1 mice were provided with drinking water or drinking water containing the antioxidant N-acetyl cysteine (NAC) [1% (wt/vol)] for one week. 0.5% (wt/vol) streptomycin and 0.5% aspartame were then added to the drinking water for the remainder of the experiment. Two days after streptomycin treatment, approximately 108 CFU of each of the two differentially-labeled strains (wildtype and mutant) were mixed at a 1:1 ratio and intragastrically administered to each mouse. Fecal pellets were collected from each mouse at the indicated time points, resuspended in LB, serially diluted, and then plated on plates containing 5-bromo-4-chloro-3-indolyl-β-D-galactopyranoside (X-gal) and appropriate antibiotics. The competitive index was calculated as the ratio of mutant to wildtype colonies normalized to the input ratio.
The infant mouse colonization assays were performed as previously described [65] with the following modifications. Briefly, mid-log phase cultures of WT (lacZ +) and mutants (lacZ -) were mixed in a 1:1 ratio and approximately 105 cells were intragastrically inoculated into 5-day-old CD-1 suckling mice. After a 20-hr period of incubation, mice were sacrificed. Small intestines were harvested and homogenized, the ratio of mutants to WT bacteria was determined by plating on LB agar containing antibiotics and X-Gal.
Approximately 108 CFU from overnight culture of a saturated Tn5 insertion C6706 library using pRL27 [66] were then intragastically inoculated into six-week-old CD-1 mice +/- N-acetyl cysteine (NAC) treatment (5 mice/group). 3 days PI, freshly-collected fecal pellets from each group were pooled and homogenized, the samples were then filtered through a 40 μm membrane. The filtrates were centrifuged, bacterial pellets were resuspended into 20 ml LB medium with appropriate antibiotics and were grown to saturation for DNA extraction (output library). The transposon junctions were amplified from sheared gDNA samples and subjected to massive parallel sequencing using Illumina MiSeq as described previously [6]. All read mapping and data analysis were performed using previously described methods [67].
Overnight cultures of wildtype, ΔmutS, and in vivo-isolated mutS (lacZ::mutS)(designated ΔmutS *) strains were inoculated at 1:100 into fresh LB containing appropriate antibiotics and shaken at 37°C until mid-log phase. Cultures were then diluted into saline and into saline containing 300 μM H2O2 and were further incubated for 1 hr. Viable cells were then enumerated by serial dilution and plating. Survival rate was calculated by normalizing CFU to the H2O2-treated group. Catalase production assays used mid-log cultures that were induced with 500 μM H2O2 for 1 hr. 1 ml of culture samples was withdrawn. Rinsed cells were collected and lysed using sonication. The lysates were then subjected to catalase activity assays using the Fluorometric Catalase Activity Assay Kit (Enco Scientific) per the manufacturer’s instructions. Mid-log cultures were induced with 500 μM H2O2 for 1 hr for measuring catalase expression. Bacterial cells were then collected and total RNA was extracted using TRIzol (Invitrogen). Single-stranded cDNA was synthesized using SuperScript III reverse transcriptase (Invitrogen) with hexadeoxyribonucleotide mixture as primers. Reverse transcription-quantitative PCR (qRT-PCR) was carried out by using the CFX96 real-time PCR system (Bio-Rad) and a two-step RT-qPCR kit with SYBR green detection (TaKaRa). To standardize results, the relative abundance of 16S rRNA was used as the internal standard.
Overnight cultures of wildtype, ΔmutS, and in vivo-isolated ΔmutS* strains were inoculated at 1:100 into fresh LB containing appropriate antibiotics and incubated without shaking at 37°C for 16 hrs. Culture supernatants were removed, and biofilms were washed with PBS. Biofilm formation was quantified by crystal violet staining as previously described [9].
To compare the ROS resistance of planktonic and biofilm associated cells, overnight cultures were inoculated at 1:100 into LB and incubated for 16 hrs at 37°C without shaking. Planktonic cells were removed and pelleted, while the remaining biofilms were rinsed with PBS. Fresh LB containing 1 mM H2O2 or 100 μM cumene hydroperoxide (CHP) was then added into tubes containing either rinsed biofilms or pelleted planktonic cells and further incubated for 1 hr. To disrupt biofilm structures, cultures were vortexed for 1 minute in the presence of glass beads. The surviving cells were then enumerated by serial dilution and plated onto LB agar.
Overnight cultures of wildtype, ΔmutS, and ΔmutS* strains were inoculated into fresh LB containing different concentrations of H2O2 and grown at 37°C shaking for 12 hrs. The cultures were then plated onto LB agar +/- 50 μg/ml rifampicin. After overnight growth at 37°C, rifampicin resistant colonies were scored. The in vivo mutation frequency was determined using the protocol described previously [45] with modifications. Briefly, fecal pellets from V. cholerae colonized mice were collected and homogenized in 10 ml LB containing 500 μg/ml streptomycin. After brief centrifugation, the supernatants were incubated at 37°C shaking for 12 hrs. The cultures were then serially diluted and plated onto LB agar containing streptomycin (500 μg/ml) and LB agar containing rifampicin (50 μg/ml) and streptomycin (500 μg/ml). After overnight growth at 37°C, rifampicin resistant colonies were scored.
Overnight cultures of V. cholerae strains containing PtcpA-luxCDABE transcriptional fusion plasmids were inoculated 1:10,000 into AKI medium [37] and incubated without shaking at 37°C for 4 hrs, followed by shaking at 37°C for an additional 3 hrs. Luminescence was then measured at the indicated time points and normalized to OD600. At the final time point, 109 cells were subjected to sodium dodecyl sulfate-polyacrylamide gel electrophoresis (SDS-PAGE) and immunoblotting using anti-TcpA antiserum.
To determine the functionality of HapR-regulated quorum sensing, the cosmid pBB1, carrying the V. harveyi lux operon [68] was introduced into V. cholerae strains by conjugation. The resulting strains were grown in LB with appropriate antibiotics at 30°C overnight, diluted to a concentration of 1:100 in fresh LB and transferred to white opaque 96 well plates and incubated at 30°C shaking. Luminescence was read at OD600 = 1.
|
10.1371/journal.pbio.0060103 | Bergmann Glia and the Recognition Molecule CHL1 Organize GABAergic Axons and Direct Innervation of Purkinje Cell Dendrites | The geometric and subcellular organization of axon arbors distributes and regulates electrical signaling in neurons and networks, but the underlying mechanisms have remained elusive. In rodent cerebellar cortex, stellate interneurons elaborate characteristic axon arbors that selectively innervate Purkinje cell dendrites and likely regulate dendritic integration. We used GFP BAC transgenic reporter mice to examine the cellular processes and molecular mechanisms underlying the development of stellate cell axons and their innervation pattern. We show that stellate axons are organized and guided towards Purkinje cell dendrites by an intermediate scaffold of Bergmann glial (BG) fibers. The L1 family immunoglobulin protein Close Homologue of L1 (CHL1) is localized to apical BG fibers and stellate cells during the development of stellate axon arbors. In the absence of CHL1, stellate axons deviate from BG fibers and show aberrant branching and orientation. Furthermore, synapse formation between aberrant stellate axons and Purkinje dendrites is reduced and cannot be maintained, leading to progressive atrophy of axon terminals. These results establish BG fibers as a guiding scaffold and CHL1 a molecular signal in the organization of stellate axon arbors and in directing their dendritic innervation.
| Large principal neurons in vertebrate neural circuits often consist of distinct anatomical and physiological compartments, which allow distributed and compartmentalized signaling and greatly increase the computational power of single neurons. Superimposed upon this intrinsic compartmental architecture is the subcellular organization of synaptic inputs, which exert local control over the biophysical properties and differentially regulate the input, integration, and output of principal neurons. In the cerebellar cortex, Purkinje neurons are innervated by GABA inhibitory synapses from the stellate and basket cells at dendrites and soma-axon initial (AIS) segments, respectively. Previous studies have shown that an L1 family immunoglobulin cell adhesion molecule (neurofascin186) is distributed as a subcellular gradient and directs basket cell axons to innervate Purkinje cell AIS. Here, we examine the mechanisms underlying the innervation of Purkinje cell dendrites by stellate axons. We found that stellate axons are organized into characteristic trajectories and guided towards Purkinje dendrites by an intermediate scaffold of astroglia—the Bergmann glial (BG) fibers. Another member of L1 family, Close Homologue of L1 (CHL1), is localized to BG fibers and stellate cells, and contributes to the organization of stellate axons along BG fibers and to the innervation of Purkinje cell dendrites.
| Neurons are often characterized by striking polarity and extensive subcellular specialization. For example, large principal neurons in many vertebrate neural circuits consist of distinct anatomical and physiological compartments [1], which allow distributed and compartmentalized signaling [2–4], and may greatly increase the computation power of single neurons [5]. Indeed, the biophysical and signaling machineries of principal neurons are organized into discrete subcellular domains [6], best exemplified by the highly restricted distribution of all major classes of ion channels along the axon–dendritic surface [7]. Superimposed upon the intrinsic compartmental architecture of principal neurons is the subcellular organization of synaptic inputs [8,9], which exert further control over the biophysical properties, not only within a neuron, but also within a neural ensemble [10]. Subcellular synapse organization is a prominent feature of neuronal wiring specificity, but the underlying cellular and molecular mechanisms are not well understood.
A prime example of subcellular synapse organization is the Purkinje neurons of the cerebellum. The cerebellar cortex is organized as a near lattice-like circuit architecture along the two axes of the cerebellar lobules, the translobular and parlobular planes [11]. At the focal position in the cerebellar cortex and as its sole output, Purkinje neurons are restricted in the translobular plane and receive at least four sets of subcellularly targeted excitatory and inhibitory inputs [11]. The glutamatergic parallel fibers synapse onto the slender spines of the more distal dendrites, whereas the climbing fibers prefer the stubby spines of the proximal dendrite. In addition, the GABAergic basket interneurons target Purkinje cell soma and axon initial segment (AIS), whereas the stellate interneurons innervate the dendritic shafts. The mechanisms underlying subcellular synapse organization along Purkinje neurons are only beginning to be understood [12]. There is evidence that the innervation of Purkinje AIS by basket interneurons is guided by a subcellular gradient of neurofascin186, an L1 family immunoglobulin cell adhesion molecule, recruited by the ankyrinG membrane adaptor protein [13]. On the other hand, the mechanisms that direct the innervation of Purkinje dendrites by stellate interneurons are unknown.
Stellate cells mainly occupy the upper half of the molecular layer (ML) and are the only cell type of the upper third of the ML. Like the basket cells, stellate cells extend their axons within the translobular (e.g., parasagittal) plane of the cerebellar cortex [11]. Although these axon arbors range from relatively simple to fairly complex, the most characteristic feature is their largely vertically oriented ascending and descending collaterals, which innervate multiple Purkinje dendrites along their path [11]. Unlike climbing fibers, which have been well documented to grow along and eventually innervate single Purkinje dendrites, the cellular and developmental process by which stellate axon approach and innervate Purkinje dendrites have not been described.
Bergmann glia (BG) cells are highly polarized astrocytes, whose radial fibers dominate the cerebellar cortex [11,14,15]. During postnatal cerebellar development, the apical BG fibers form the earliest radial structures across the cerebellar cortex [14]; BG fibers subsequently undergo dramatic differentiation and are transformed into a highly elaborate meshwork, dominated by a scaffold of radial fibers [16–18]. The Bergmann glia cells are positioned to interact with multiple neuronal components and likely contribute to multiple aspects of cerebellar circuit assembly at different developmental stages. Whether BG fibers play a role in axon guidance and organization at later stages, in addition to guiding granule cell migration [19], has not been explored. By using a genetic strategy to simultaneously label stellate axons and BG fibers at high resolution, here we provide evidence that BG fibers constitute an intermediate template in the organization of stellate axon arbors into characteristic trajectories, and in their guidance to innervate Purkinje dendrites.
The L1 family immunoglobulin cell adhesion molecules (L1CAMs) have been implicated in axon growth, guidance [20], and the subcellular organization of GABAergic synapses [12]. Given the role of NF186 in the targeting of basket axon and pinceau synapses to Purkinje AIS [13], here we explored whether other members of the L1CAMs might contribute to the organization of stellate axons and innervation of Purkinje dendrites. We found that each member of the L1 family is localized to subcellular domains in neurons and glia in the developing and mature cerebellar cortex. In particular, CHL1 (Close Homologue of L1) is prominently expressed on BG fibers during the development of stellate axons and their innervation. In addition, we demonstrate a crucial and highly specific role of CHL1 in the patterning of stellate axons and in targeting their innervation to Purkinje cell dendrites.
To investigate the cellular mechanisms underlying development of the stellate axon arbor and innervation pattern, it is necessary to visualize stellate axons together with their postsynaptic targets at high resolution and during the developmental process. We generated several lines of bacterial artificial chromosome (BAC) transgenic reporter mice to achieve such visualization. The calcium binding protein parvalbumin (Pv) is normally expressed in all Purkinje, basket, and stellate neurons in the postnatal cerebellar cortex. In our Pv-GFP BAC transgenic mice (Figure 1A), different fractions of Purkinje, basket, and stellate neurons express green fluorescent protein (GFP) among different stable lines (from a few percent to near 100%; unpublished data), likely due to different genomic integration sites of the transgene. In the B20 line, sparse GFP expression allowed visualization of single Purkinje, basket, and stellate cells with synaptic resolution (Figure 1B–1E). Combined with generic Purkinje cell markers (e.g., calbindin), we were able to examine the precise trajectory of stellate and basket axons, and compare how they approach and innervate postsynaptic Purkinje neurons.
Stellate and basket cells occupy mainly the upper or lower half of the ML, respectively. Consistent with previous Golgi studies [11], GFP labeling revealed distinct features of stellate and basket axons in their morphology and subcellular innervation pattern, even when they are found at the same location in the mid-lower ML (Figure 1F–1H). For example, basket axons were smooth, whereas stellate axons were beaded. Although basket axons extended terminal branches along Purkinje cell proximal dendrites, soma, and towards AISs (Figure 1H), stellate axons clearly did not extend along the distinct contours of Purkinje dendrites. Instead, stellate axon arbors were often characterized by rather straight ascending and descending collaterals (also described in [11]) that crossed Purkinje cell dendrites at rather sharp angles (Figure 1F and 1G). Interestingly, the axon tips of stellate cells in the lower ML often reached the Purkinje cell layer (PCL), as the basket axons. However, these descending collaterals never extended along Purkinje dendrite-soma-AIS, but always had rather straight paths that terminated abruptly at the PCL (Figure 1G). These contrasting features suggest that basket and stellate axons approach and innervate Purkinje neurons via profoundly different cellular and developmental mechanisms. Furthermore, the arborization and innervation patterns of stellate axons also contrast those of a dendrite-targeting climbing fiber, which grows along and “monopolizes” an entire Purkinje dendrite [11]. These comparisons raise an obvious question: how do stellate axons approach and innervate segments of multiple Purkinje dendrites without growing along any single target?
To better understand the developmental process by which stellate axons innervate Purkinje dendrites, we used our reporter mice expressing GFP from the GAD67 promoter elements (GAD67-GFP BAC reporter mice, Figure 2A and [13]), which label Purkinje, basket, and stellate cells from embryonic stage to adulthood. In the G42 line, both interneurons and Purkinje cells were labeled; this line was mainly used to characterize the migration of stellate cells in the first two postnatal weeks (summarized in Figure 2C). In the G1 line, GFP was mainly expressed by basket and stellate cells; this line was used to characterize the development of stellate axons.
Basket and stellate cells are derived from dividing progenitors in the postnatal cerebellar white matter (WM; [21]). These progenitors migrate into the cerebellar cortex in the first two postnatal weeks as simple unipolar cells until arriving in the ML. They then undergo a series of morphological transformations that culminate in formation of mature interneurons during the third and fourth postnatal week. However, the morphological maturation of stellate axons has not been well described. Compared to basket cells, stellate cell precursors migrate into the ML a few days later, peaking between postnatal day 8 and 11 (P8–P11) but continue to arrive as late as P14 (Figure 2B–2D; also see [22]). Using our G1 reporter mice, we found that upon reaching their positions in the ML, stellate cells first appeared bipolar and extended largely horizontally oriented neurites (Figure 2B; also see [21]). Between P16–18, stellate axons sent ascending and descending collaterals (Figure 2E), which further gave rise to plexus of more elaborate branches in the subsequent 2 wk (Figures 2F). Mature stellate axon arbors range from relatively simple to fairly complex; the most characteristic feature is their largely vertically oriented ascending and descending collaterals [11]. Our GFP labeling is highly consistent with these previous descriptions using Golgi methods (Figure 2E and 2F, and unpublished data). The stereotyped morphology and development of stellate axons pose an obvious question: how are they organized into characteristic trajectories, presumably by mechanisms other than Purkinje cell dendrites?
Besides Purkinje dendrites, an equally prominent cellular component of the cerebellar cortex are the BG fibers [15] (Figures 2D–2F and S1). In rodents, BG are present during embryonic stages [23]; they migrate to the cerebellar cortex before birth, and their radial fibers reach the pia to form characteristic endfeet by late embryonic stages [14]. BG fibers thus represent the earliest radial structures across the cerebellar cortex, before the arrival of Purkinje neurons [14]. During the first postnatal week, BG fibers are thin, smooth, and unbranched. The glia-specific cytoskeleton protein GFAP can be detected by P4 [24]. The simple BG fibers subsequently undergo profound morphological differentiation and maturation [16,18,25]. During the second week, when Purkinje dendrites extend, BG fibers differentiate in a deep to superficial gradient: whereas BG fibers transversing the external granule layer (EGL) remain smooth, they extend coarse lateral appendages in the underlying ML [26]. During the third and fourth week, BG fibers further branch, extend lateral varicoses and fine processes, eventually forming an extensive reticular meshwork [16,18,25]. Consistent with these results, using single-cell electroporation to label BG with GFP, we found that BG fibers project highly irregular lateral branches during the third postnatal week (Figure S4A). Furthermore, using transgenic mice expressing GFP under the control of a mouse GFAP promoter [27], we were able to visualize the extensive meshwork of the BG system in the ML, and found that GFAP was largely concentrated in the radial BG fibers, but not the finer lateral appendages and processes (Figure S4B). In addition to the apical radial fibers in the ML, BG cell bodies also give off numerous lamellar processes that enwrap Purkinje cell soma and AIS after the third postnatal week [11,14], although the more precise timing of this process is unclear. Mature BG cells are thus highly polarized astrocytes with distinct subcellular specializations.
The vertical bias of the orientation of stellate axon collaterals prompted us to examine their relationship with BG fibers during the development of dendritic innervation. As expected, when stellate cell precursors were migrating across the PCL in the second postnatal week, GFAP-positive BG fibers were prominent throughout the ML (e.g., P8, Figure 2D). Upon reaching their destination in the ML, stellate cells began to extend neurites. Although their axons extended in different directions, many of their descending/ascending branches were strictly associated with GFAP-labeled BG fibers (Figure 2E). This association was particularly prominent in the upper ML, where stellate axons often perfectly followed the curving BG fibers for tens of microns, and remained so in subsequence weeks (Figure 2F). Such extensive association with BG fibers contrasts the rather patchy and “en passant–type” interaction between Purkinje dendrites and BG fibers [14,16,28]. Our detection of association between stellate axons and Bergmann glia was probably an underestimate, since the finer lateral BG appendages were not well labeled by GFAP. Importantly, there was no association between basket axons and BG fibers (Figure 2G), consistent with the finding that basket terminals grow along the proximal dendrite-soma-AIS of Purkinje cells (Figure 1H).
To substantiate this finding, we further examined the association of GABAergic synaptic markers with BG fibers. GAD65, an isoform of glutamic acid decarboxylase, is localized to GABAergic presynaptic terminals and physically coupled to synaptic vesicles ([29]). The onset of GAD65 expression has been shown to coincide with GABAergic synaptogenesis in the cerebellum [30]. We focused our analysis to the upper ML, where most, if not all, GAD65 signals are derived from stellate cell axons. At P16, shortly after stellate neurons begin to send their axons and made synaptic contacts, double labeling revealed a 52% colocalization between GAD65 and GFAP (Figure 3A). This colocalization increased to 65% in the more mature ML (Figure 3B and 3C). To rule out the possibility that these levels of colocalization can be reached by chance, we artificially shifted the confocal image stacks of GFAP horizontally relative to that of GAD65 by 5 μm, and then reanalyzed GAD65 and GFAP colocalization (since BG fibers were arranged vertically with an average gap of approximately 5 μm between neighboring fibers; see Materials and Methods). This shift analysis revealed a highly significant 30% decrease in GAD65-GFAP association (p ≤ 0.01, n = 20 different sections in 3 different mice), indicating that the organization of GAD65 along BG fibers was not due to chance. This analysis is likely an underestimate of GAD65–BG association since GFAP antibodies did not label well the finer BG processes (Figure S4B and S4E). In addition to the statistical association, strings of GAD65 puncta, indicative of an underlying stellate axon branch, were frequently seen to perfectly align with GFAP fibers (Figure 3B and 3C). The combined observations suggest that BG fibers in the cerebellar cortex provide a growth template, which may organize stellate axons into characteristic orientations and trajectories. These results are consistent with ultrastructural observations that stellate cell axons and presynaptic terminals are surrounded by glial processes during the third postnatal week [14].
Since Purkinje dendrites are the major postsynaptic targets of stellate axons, our results raise the question of whether and how BG fibers guide stellate axons to Purkinje dendrites. Mature Purkinje dendrites bear large numbers of synaptic boutons, but much of their surface is ensheathed by a thin BG process [14]. The glial sheath of a dendritic segment is thought to consist of processes derived from several neighboring BG cells [14]; and it has been recognized since Cajal that the BG fibers are intercalated between the dendritic trees of successive Purkinje cells [11]. Using our Pv-GFP reporter mice, which label individual Purkinje dendrites, and GFAP antibody, we found that BG fibers most often intersected dendritic shafts at sharp angles and did not extend along dendrite at significant length (Figures 3D and S1A). Furthermore, in GFAP-GFP transgenic mice, which occasionally gave sparse labeling of BG cells, double labeling with calbindin antibody showed that a single BG fiber most likely encounters several intercalated Purkinje dendrites (Figure S1B). Therefore, BG fibers impinge upon and enwrap multiple Purkinje dendritic segments in a patchy, en passant pattern.
To examine the precise relationship among stellate cell presynaptic terminals, BG fibers, and Purkinje dendrites, we performed triple labeling with GAD65 and GFAP antibodies in our Pv-GFP mice. As expected, GAD65 puncta colocalized with the shafts of Purkinje dendrites, occasionally aligned in a “beads along a string” pattern, indicative of a stellate axon branch (Figure 3D1). Importantly, the same GAD65 puncta and clusters were also precisely aligned along a GFAP fiber (Figure 3D2), indicating that stellate axon boutons are formed at the intersection between BG fibers and Purkinje dendrites (Figure 3D3). Together, these results suggest that BG fibers in the ML represent an “intermediate scaffold,” which may guide stellate axons to approach Purkinje dendrites in defined orientation and trajectories, and form synaptic contacts at the intersection between BG fibers and Purkinje dendrites (Figure 3G).
To explore the molecular mechanisms underlying the GABAergic innervation of Purkinje dendrites, we took a candidate gene approach and focused on the L1CAMs. The L1CAM subfamily consists of L1, CHL1, NrCAM, and neurofascin [20]. We have previously shown that a Purkinje cell–specific splice variant of neurofascin (NF186) directs the innervation of axon initial segment by basket cell axons [13]. We therefore systematically examined the expression pattern of every L1CAM during the postnatal development of the cerebellar cortex. Interestingly, each member was localized to distinct subcellular compartments in neurons and glia cells (Figure S2). During the third postnatal week (e.g., P16), whereas NF186 was highly restricted to AIS-soma of Purkinje cells [13], L1 was abundantly expressed in parallel fibers and other unmyelinated and premyelinated axons (Figure S2D and S2G). NrCAM was more diffusely (but certainly not ubiquitously) expressed in the ML, although the precise cellular and subcellular locations could not be ascertained (Figure S2E). Interestingly, in the PCL, NrCAM appeared to localize to the basal lamellae of BG that wrapped around Purkinje soma and AIS (Figure S2E). Finally, using an antibody to a peptide epitope in the FNIII domain of CHL1 (Figure 3), we found that CHL1 was distributed in a prominent radial stripe pattern that resembled BG fibers along with diffuse labeling in the ML (Figure 4A–4D). Indeed, CHL1 closely colocalized with GFAP (Figure 4E), but not the Purkinje dendrite marker calbindin (Figure S2F). Such colocalization with GFAP was detected throughout postnatal development (unpublished data).
We further characterized the postnatal developmental expression of CHL1. At P8, when stellate cells were just migrating across the PCL, CHL1 was already prominent along BG fibers (Figure 4A1; colocalization with GFAP not shown). CHL1 was subsequently also detected along the lateral appendages during the second and third week (P14–20, Figure 4C1 and 4E). Importantly, along the polarized BG cells, CHL1 was mainly localized to the apical radial fibers and processes, but not to the basal lamellae that extend towards Purkinje cell AIS (Figure 4F). This pattern in the PCL was clearly distinct from that of NrCAM (Figure S2E). CHL1 expression subsequently diminished in the BG fibers and became more diffuse, yet prominent, in the molecular layer (Figure 4D1). In situ hybridization indicates that CHL1 is also expressed in stellate interneurons and granule cells at P14, but not in mature Purkinje neurons [31]. Consistent with these data, CHL1 immunofluorescence appeared in stellate cell somata as early as P14, and remained at P18 and P40 (Figure 4A2–D2). It was difficult to determine whether CHL1 was also distributed along stellate axons and/or dendrites because of the more diffuse labeling in the ML. Lower levels of CHL1 expression in the ML remained in adulthood (in 1-y-old mice, unpublished data).
To investigate whether CHL1 plays a role in the GABAergic innervation of Purkinje dendrites, we first examined the expression of the presynaptic marker GAD65 in the ML of CHL1 knockout mice (Figure 5A and 5C). As a control, we also surveyed all the viable L1CAM mutant mice using the same assay (Figure 5E–5H). The vast majority of GABAergic terminals in the ML are derived from stellate axons; Purkinje collaterals and basket axons only contribute to a small minority near the PCL [11]. Purkinje dendrites are the predominant targets of stellate axons, although the dendrites of stellate, basket, and Golgi cells are also innervated [11]. In the adult cerebellar cortex (>P40), we found a profound reduction of GAD65 labeling in CHL1−/− mice, but not in L1−/− and NrCAM−/− mice (Figure 5E–5H). This reduction was specific to the ML layer: GAD65 labeling at Purkinje AIS in CHL1−/− mice was identical to that of wild-type (WT) littermates, L1−/− mice, and NrCAM−/− mice (Figure 5A and 5C, and unpublished data). We took advantage of this result to quantify GAD65 signals in the upper ML as a ratio to those at the Purkinje cell AIS. Such quantification revealed an approximately 60% reduction (p ≤ 0.01, n = 4 mice) of GAD65 in CHL1−/− mice compared to their WT littermates (Figure 5I). This significant reduction was not due to a defect in the migration of stellate cells, since stellate cell density and distribution in the ML assayed by Pv immunofluorescence were the same as those in WT mice (Figure 5B and 5D). Furthermore, calbindin staining did not reveal any discernable defects of Purkinje dendrites. We also examined glutamatergic innervation of Purkinje dendrites. The density of parallel fiber synapses and climbing fiber synapses detected by vGluT1 and vGluT2 [32] immunofluorescence, respectively, showed no differences between CHL1−/− and WT mice (Figure S5A–S5F). The ultrastructures of parallel fiber and climbing synapses also appeared normal (Figure S5G–S5K). These results suggest that, among the L1CAMs, CHL1 appears to play a highly specific role in the GABAergic innervation of Purkinje dendrites. The GAD65 assay itself does not rule out the possibility that stellate innervation of other cell types may also be affected.
Although CHL1 has been shown to modulate radial migration of certain populations of pyramidal neurons in sensory areas of developing neocortex [33], we did not find notable defects in density and position of Purkinje neurons in the cerebellar cortex, although subtle defects cannot be ruled out. We did notice occasional mispositioning of BG cell soma in the granule cell layer (Figure S1C). BG fibers labeled by GFAP were also largely normal, except that they occasionally appeared somewhat less well organized (Figure S1). It is not clear whether these are due to a direct effect of CHL1 deficiency in BG or an indirect consequence of their disrupted association with stellate axons.
To investigate the role of CHL1 in the development of stellate axons, we examined the morphology of single stellate axon arbors using our Pv-GFP (B20) mice. In the ML of mature WT B20 mice (P44), stellate axons display complex arbors with characteristic orientations (Figure 6A); a majority of these axon branches displayed a predominantly vertical orientation and were associated with GFAP-labeled BG fibers (Figure 6C–6E). Quantification of the orientation of stellate axon branches relative to the pia surface revealed that they followed a Gaussian distribution, with a peak between 80° and 100° (Figure 6C). In addition, 70% of these vertically oriented axon branches were associated with GFAP-positive fibers (Figure 6D and 6E). Even when axons branched and turned, they often switched between neighboring BG fibers (Figure 6A3; indicated by arrowheads and stars). Mature stellate axons bore distinct boutons, and more than 90% of these boutons contained the synaptic marker GAD65 (Figure 7A and 7B).
In PV-GFP(B20)::CHL1−/− littermates, most stellate axons still were able to develop fairly complex arbors at this age, but appeared thinner, more wavy, with significantly altered orientation and trajectories (Figure 6B). When double labeled with GFAP, the notable defects were their reduced association with BG fibers and the reduction of vertically oriented branches. Indeed, the orientation of axon branches was much more evenly distributed (Figure 6C), and many of these more horizontally oriented axons often simply crossed over the BG fibers (Figure 6B3). Quantification revealed that less than 30% of stellate axon branches were associated with GFAP fibers, regardless of their orientation, indicating a significant reduction compared to that in WT mice (Figure 6D and 6E). The altered arbor morphology of stellate axons and their reduced association with GFAP fibers was apparent at P16 and P20 (compare Figure S6A and S6B with Figure 2). In several extreme cases in P44 CHL1−/− mice, stellate axons were grossly abnormal, with much-reduced branching and simpler arbors. These axons extended rather randomly, twisted, tangled, and even circled around (Figure S6C), with apparently a complete loss of orientation preference. The failure to interact with GFAP fibers may have profoundly altered stellate axonal organization and trajectory in CHL1−/− mice. These axons also bore smaller boutons (Figure 7C and 7D), and only 50% of these boutons contained detectable GAD65, a 43% reduction compared to WT mice (p ≤ 0.001; Figure 7E and 7F). Importantly, these defects were highly specific to stellate axon, basket axons and their innervation of Purkinje AISs appeared entirely normal in CHL1−/− mice both at single-cell resolution (Figures 7C2 and S7) and when assayed with GAD65 (Figure 5C).
We used electron microscopy to directly examine stellate synapses on Purkinje dendrites. We restricted our analysis on the upper third of the ML, where all symmetric synapses are derived from stellate axons. In WT mice at P44, stellate terminals exhibiting symmetric synapses were identified along the Purkinje dendritic shafts as clear varicosities containing densely studded, flattened vesicles (Figure 7G). The density of stellate terminal boutons with symmetric synapses was quantified against Purkinje dendritic surface area from serial ultrathin sections to avoid overlooking stellate terminal profiles. In CHL1−/− littermates, morphologically normal terminal boutons with symmetric synapses were clearly present along Purkinje dendrites (Figure 7H), with diameters ranging from 0.4–0.7 μm, and an active zone length of 0.15–0.26 μm. However, the density of symmetric synapses was reduced by 60% (p ≤ 0.001). At P30, there was also a significant reduction in the density of symmetric synapses by approximately 40% (p < 0.03). On the other hand, basket axon synapses on Purkinje somata, parallel fiber synapses on dendritic spines, and climbing fiber synapses on dendritic shafts were all indistinguishable between P44 WT and CHL1−/− mice (Figures S5G–S5K, S7C, and S7D).
Much more severe defects of stellate axon terminals in CHL1−/− mice were detected at older ages. In 3-mo-old mutants, degenerating axon profiles were frequently seen in the upper ML, exhibiting electron-dense membrane accumulations and electron-lucent empty spaces (Figure 7J). On the other hand, nearby climbing fiber terminal profiles along the same Purkinje dendrites were perfectly normal. Together, these ultrastructural results suggest that in the absence of CHL1, aberrantly organized and oriented stellate axons can still manage to contact Purkinje dendrites and form synapses, but at significantly reduced efficiency and density. In addition, these synapses cannot be maintained, leading to atrophy of stellate axon terminals.
Besides BG, CHL1 is also expressed in other cell types, such as stellate cells, granule cells, and their parallel fibers in the developing cerebellum [31]. To further investigate the role of CHL1 in BG, we bred a conditional CHL1 mutant strain (CHL1flx) (see Materials and Methods) with a transgenic lines expressing CRE recombinase under the control of GFAP [34]. At P14, P20, and P40 in GFAP-Cre::CHL1flx/flx mice, CHL1 expression was undectable along BG fibers but was clearly present in stellate cells (Figure 8A–8C). At P40, there was a significant reduction of GAD65 density in GFAP-Cre::CHL1flx/flx mice compared to CHL1flx/flx controls (Figure 8D, 8F, and 8H; 27 ± 7%; p ≤ 0.05). We also deleted CHL1 in Purkinje cells by breeding CHL1flx mice with the L7-Cre transgenic mice [35]; there was no reduction of GAD65 density at P40 (Figure 8G and 8H), suggesting that CHL1 in Purkinje cells, if any, was not involved in the development of stellate synapses. These results suggest that CHL1 expression in BG contributes to the development of stellate cell synaptic innervation in the ML. Compared with germline CHL1−/− mice (Figure 5E and 5H), the intermediate reduction of GAD65 in the ML of GFAP-Cre::CHL1flx/flx may be due to two reasons. First, the association between stellate axon and BG fibers may be mediated by CHL1 homophilic as well as heterophilic interactions; absence of CHL1 in BG fibers thus partially impairs the association of stellate axons with BG fibers and the innervation of Purkinje dendrites. CHL1 expression in stellate cells likely also plays a significant role. Second, it is possible that only the GABAergic innervation of Purkinje dendrites (but not stellate dendrites, for example) is guided by CHL1 expression on BG fibers; absence of CHL1 in BG fibers thus only partially reduced the GAD65 signal in the ML.
The spatial distribution of a neuron's output is determined by the geometry of its axon arbor and the pattern of its innervation. Different classes of neurons often display characteristic axon arbors which target restricted spatial locations, cell types, and subcellular compartments in neural circuits. Some of the best examples of neuronal class-specific innervation patterns are found along Purkinje neurons, which reside in the translobular plane of the cerebellar cortex and receive four sets of synaptic inputs (Figure 1A) [11]. Among the glutamatergic inputs, the parallel fibers run along the parlobular axis and impinge upon Purkinje dendrites perpendicularly; each granule cell axon often contacts a single spine from one entire Purkinje dendrite, but may innervate hundreds of dendrites along its path. In sharp contrast, the climbing fibers restrict their arbors in the translobular plane; each eventually innervates only one Purkinje cell but with hundreds of synapses along its dendritic shaft. The two types of GABAergic interneurons both extend their axons within a rather narrow translobular plane and innervate multiple targets within a few rows of Purkinje cells [11,14]. Whereas a basket axon typically grows along and innervates seven to ten Purkinje somata and AISs [11], the descending and ascending collaterals of a stellate axon likely innervate multiple Purkinje dendrites [11] but do not grow along Purkinje cells (Figure 1). The geometric and subcellular organization of dendritic-targeting GABAergic axons and innervation patterns are crucial in the physiological control of synaptic integration in postsynaptic neurons, yet the underlying mechanisms are largely unknown. Here, we present evidence that stellate axons are organized in characteristic trajectories and guided to Purkinje dendrites by an intermediate scaffold—the BG fibers; in addition, the L1CAM CHL1 is a molecular signal that contributes to the patterning and subcellular organization of stellate axons and innervation (Figure 9).
In mature cerebellar cortex, each BG cell gives rise to several ascending BG fibers, which extend approximately 40–50 μm in the translobular plane and 15–20 μm in the parlobular plane [14,15]. Interestingly, these largely radial fibers from neighboring BG cells are further aligned as thin walls, or palisades, in the parlobular plane perpendicular to a Purkinje dendrite, which extends approximately 300–400 μm in the translobular plane and 15–20 μm in the longitudinal plane [14,15]. The consequence of these arrangements is that several BG palisades cut across a single Purkinje dendrite [14,15]. Although this striking spatial organization of BG fibers has long been recognized and postulated to contribute to the architecture of the cerebellum, no specific neuronal elements and developmental process have been identified that rely on such fine arrangement. By high-resolution labeling of stellate axons superimposed upon BG and Purkinje cells, we realized that BG fibers may be an ideal intermediate scaffold to “presort” a stellate axon into characteristic trajectories and distribute them towards multiple Purkinje dendrites.
During cerebellar development, BG fibers represent the earliest radial structures across the cerebellar cortex, even before the arrival of Purkinje neurons [14,23]. The initially simple BG fibers undergo dramatic differentiation and maturation in the second to fourth postnatal week and are transformed into a highly elaborate meshwork, dominated by the vertical palisades [14,16,18,25,26]. Although the elaborate BG fibers appear to be positioned to interact with multiple neuronal components, such as migrating granule cells [19,36], and likely contribute to multiple aspects of cerebellar circuit assembly at different developmental stages, our discovery of their close association with the developing stellate axons is particularly compelling. First, the association was apparent as soon as stellate cells begin to extend axons during the second postnatal week. Second, stellate axons often strictly followed the curving contours of BG fibers for tens of microns, as well as the lateral appendages of BG fiber. Finally, the association between BG fibers and stellate axons was specifically disrupted by the loss of an immunoglobulin family cell adhesion molecule expressed in both BG fibers and stellate cells. It is thus likely that BG fibers mainly serve as a growth template for stellate axons, and additional molecular and/or activity-dependent mechanisms may regulate the size and exuberance of axon arbors. Interestingly, BG processes also express GABAA receptors that enwrap inhibitory synapses [37]; it is thus possible that BG fibers may respond to GABA signaling from developing stellate cell axons. Mature stellate axons extend characteristic ascending and descending collaterals as well as plexus of finer branches and terminals [11]. Our GFAP labeling of BG likely underestimated their association with stellate axons. It is possible that the GFAP-positive BG fibers may represent “highways” for stellate axon collaterals, and that the lateral appendages and processes may serve as “local roads” for axon terminals to approach and innervate Purkinje dendrites.
In both invertebrates and vertebrates, the crucial role of glia cells in axon guidance has been well recognized [38,39]. Glial cells can function as guideposts to attract [40–42], repel [43–45], or stop [46] growth cones of projection neurons [38,47], and can also serve as intermediate targets to coordinate pre- and postsynaptic interactions [46,48]. In the developing rodent olfactory bulb, radial glial cells interact with olfactory receptor neuron axons [49] and have been postulated to contribute to the formation of glomeruli [50]. At hippocampal excitatory synapses, astrocytes form tripartite complexes with pre- and postsynaptic structures, and regulate synapse morphogenesis and maturation [51,52]. Here, we provide the first evidence to our knowledge that the characteristic astroglial processes organize the axon trajectory of GABAergic interneurons and contribute to the establishment of precise patterns of connectivity in complex local circuits, including subcellular synapse targeting. In many areas of the vertebrate brain (e.g., neocortex and hippocampus), highly abundant and morphologically elaborate astrocytes mature during postnatal development along with the assembly of local circuits. It is thus possible that an astroglial intermediate scaffold might be a more general mechanism for directing the trajectory of axon extension, pre- and postsynaptic target interaction, and complex patterns of innervation.
Like other members of the L1CAM [20], CHL1 is expressed in both neurons and glia [31,53]. Although there is evidence that CHL1 promotes and inhibits neurite outgrowth in vitro through both heterophilic and homophilic interactions, respectively [54,55], and may regulate hippocampal axon projection and organization [56,57], the cellular interactions involved and the logic of CHL1′s action have been unclear. The well-defined architecture and connectivity in cerebellar cortex present an advantage in defining the cellular and subcellular expression of CHL1 and in dissecting its role in axonal and synapse development. CHL1 is prominently localized to apical BG fibers since the first postnatal week, and subsequently extend to the lateral appendages during the second and third postnatal week. We cannot ascertain whether CHL1 is also expressed in the fine BG fiber processes (due to the presence of CHL1 in parallel fibers and possibly other neural elements in the ML). Importantly, CHL1 is not localized to the basal lamellae of BG cells, which extend to Purkinje soma and AIS. Such polarized distribution of CHL1 in BG cells may present a permissive substrate for the growth and patterning of stellate axons and for their restriction to the ML to innervate Purkinje dendrites. In addition, CHL1 is expressed in stellate cells, but not in Purkinje neurons [31]. CHL1 immunoreactivity could be clearly detected in stellate cell somata by P14 (Figure 4), although its subcellular distribution (on axons and dendrites) was difficult to discern.
In our analysis of CHL1−/− mice, the trajectory and orientation of stellate axons and their innervation of Purkinje dendrites were profoundly aberrant. In contrast, basket axons and their innervation of the Purkinje soma-AIS were entirely normal. In addition, Purkinje dendrites and their glutamatergic innervation by climbing fibers and parallel fibers also appeared intact, even though CHL1 is known to be expressed in granule cells and parallel fibers [31,53]. These results reveal a highly specific role for CHL1 in the patterning of stellate cell axon arbors. The significant reduction of GAD65 puncta in the ML may result from a reduction in the number of stellate synapses, deficient synapses, or both. Whereas double labeling and confocal microscopy detected a reduced localization of GAD65 to stellate boutons, ultrastructural analysis confirmed a significant reduction of stellate synapses along Purkinje dendrites. Interestingly, a recent study shows that CHL1 is localized at presynaptic terminals of glutamatergic and GABAergic axons in dissociated hippocampal cultures [58]; CHL1 appears to be targeted to synaptic vesicles by endocytosis in response to synapse activation and regulates the uncoating of clathrin-coated synaptic vesicles [58]. It is thus conceivable that the absence of CHL1 in stellate cell axons may impair GABAergic vesicle endocytosis and GAD65 synaptic localization. We suggest that CHL1 deficiency results in dissociation of stellate axons from their normal BG fiber “tracks,” aberrant axon orientation and trajectory, which contribute to subsequent deficiency in synapse formation and stability. Our current analysis cannot distinguish whether the decreased number of GABAergic synapses from stellate onto Purkinje cells in CHL1−/− mice results from the inability of the stellate axon to engage in synapse assembly, deficient cell adhesion prior to synapse assembly, or deficient synapse maintenance.
The reduction of GAD65 signals in the ML of BG-restricted CHL1 knockouts further pinpoints a specific role of CHL1 in BG fibers. On the other hand, the intermediate phenotype in these mice compared to that in germline CHL1−/− mice implies that CHL1 in other cell types, e.g., stellate and granule cells, may also contribute to their axon and synapse development. It is possible that CHL1 may localize to stellate axons and contribute to arbor patterning through homophilic interaction with CHL1 distributed on BG fibers. On the other hand, unknown CHL1 ligands in stellate cells and BG fibers may mediate heterophilic interactions during stellate axon development. Indeed, CHL1 can act as a coreceptor for neuropilin-1 to mediate axon guidance by semaphorin3A during development of the thalamocortical projection [59]. During the third postnatal week, CHL1 expression in stellate cells might also promote the maturation and stability of synaptic innervation through heterophilic interactions with Purkinje dendrites and hetero- or homophilic interaction with BG fibers. Finally, CHL1 might also be localized to stellate dendrites, which are innervated by other stellate axons. Deficiencies in stellate axon arbor and synaptic innervation in CHL1−/− mice may contribute to the impairment in their motor behaviors, such as the ability to maintain balance on an accelerated Rota-Rod [60].
Although it was once debated whether basket cells and stellate cells were variants of the same class of cerebellar interneurons, it is now established that they constitute distinct cell types, likely with distinct genetic origins [61], and a fundamental difference is their subcellular target innervation. In both cell types, the final axon arbor and innervation pattern is achieved through sequential developmental processes, which may involve: the pattern and order of their migration into the ML, the elaboration of axon arbors along defined cellular substrates and adhesion mechanisms, and the formation–stabilization of synaptic contacts along different compartments of Purkinje neurons. Here, we demonstrate that stellate and basket cells deploy different cellular and molecular mechanisms to achieve their distinct axon arborization and innervation patterns.
The basket cells make synaptic contacts along the soma-AIS of a Purkinje cell, a highly restricted synaptic target area. It is perhaps not surprising that basket axons arrive at their destination, in part, by growing along the Purkinje proximal dendrite-soma-AIS, guided by a subcellular gradient of neurofascin [13]. The stellate cells, on the other hand, face a rather different task when innervating Purkinje dendrites: even though each Purkinje dendrite is a largely 2-dimensional, flat target, it expands hundreds of microns in the translobular plane. Unlike a climbing fiber, which adheres to and monopolizes a single Purkinje dendrite, a stellate axon innervates segments of multiple dendrites, often with characteristic descending and ascending collaterals. It is thus not obvious how stellate axons can ever achieve such a distinct innervation pattern by direct and strong adhesion to Purkinje dendrites. The BG fibers seem to provide a useful solution to this problem. As an extensive and largely radial scaffold in the ML, the BG fibers are well suited to organize and deliver stellate axons to Purkinje dendrites, with defined orientations and trajectories. In addition, by relying on a glial instead of neuronal substrate, stellate axons may reduce the risk of making ectopic and unnecessary synaptic contacts. Furthermore, the BG fibers appear to direct both the patterning of axon arbor and subcellular innervation. It remains to be investigated whether such an intermediate glial scaffold is a more general strategy to sculpt precise neuronal connections in other brain areas.
We present evidence that, CHL1, a close homolog of neurofascin186, is involved in the development of stellate axons and their dendritic innervation. Our results suggest that different members of the L1 family may contribute to axon patterning and subcellular synapse organization in different cell types, and may act in glia as well as in neurons. The subcellular recruitment of NF186 is achieved by the ankyrinG membrane adaptor protein at the Purkinje AIS [13]. It is tempting to speculate that another form of ankyrin in BG cells may organize CHL1 subcellular localization. In addition to permissive/attractive signals, such as NF186 to basket axons and CHL1 to stellate axons, repulsive or bifunctional signals (depending on different axon types) at distinct subcellular sites may also contribute to topographically precise synapse organization. The identification of physiological ligands for NF186 and CHL1 in basket and stellate axons will further our understanding of the underlying molecular mechanisms.
BAC clones containing the mouse parvalbumin (PV) genes were identified from the RPCI-23 library (CHORI). A BAC clone containing the entire PV gene and approximately 150 kb of upstream and 25 kb downstream regions was used for BAC modifications. A GFP expression cassette was inserted in the first coding exon at the translation initiation site using a procedure developed by Yang et al. [62]. Circular BAC DNAs were injected into the fertilized eggs of the C57BL/6 strain at a concentration of 0.5 ng/μl in microinjection buffer (10 mM Tris [pH 7.4], 0.15 mM EDTA [pH 8.0]) using standard procedures and as described previously (Ango et al., 2004 [13]). Five transgenic founders were identified by PCR and confirmed by southern blotting. All founder lines resulted in germline transmission. GFP expression was first analyzed in fixed brain sections immunolabeled with antibodies to various GABAergic interneuron markers: Pv, somatostatin, calretinin, and VIP. In the cerebellum, different fractions of Purkinje, basket, and stellate neurons expressed GFP among different transgenic lines, from a few percent (the B20 line) to near 100% (the B13 line, and unpublished data), likely due to different genomic integration sites of the transgene. The onset of GFP usually started in the late second postnatal week and increased to higher levels by the fourth week. The GAD67-GFP reporter mice were described in Ango et al., 2004 [13].
The CHL1−/− mice were described in [56]. L1−/− and NrCAM−/− mice were provided by Drs. Dan Felsonfeld and Dr. Martin Grumet, respectively. The CHL1 conditional mutant will be published elsewhere (Kolata et al., unpublished data). The L7-cre mice [34] were obtained from Mutant Mouse Regional Resource Centers (MMRRC) and the GFAP-cre mice [33] from JAX Mice.
Mice were anesthetized (sodium pentobarbitone, 6 mg/100 g of body weight) and transcardially perfused with 4% paraformaldehyde in phosphate buffer (pH 7.4). Sagittal sections (80-μm thick) were cut from the cerebellum using a vibratome (Leica VT100). Brain sections were blocked in 5% NGS and 0.1% Triton X-100, and immunostained with antibodies against GAD65 (monoclonal antibody, 1:1,000; Boehringer), GFP (rabbit or chicken polyclonal antibody, 1:500; Chemicon), Pv (monoclonal antibody, 1:1,000; Sigma), CHL1 (chicken polyclonal antibody, 1:500), calbindin (rabbit polyclonal antibody, 1:1,000; Swant), and GFAP (rabbit polyclonal; Geko). Sections were incubated with either Alexa594-conjugated goat anti-mouse or anti-rabbit IgG and Alexa488-conjugated goat anti-rabbit or anti-chicken IgG (1:500; Molecular Probes) and mounted. Sections were imaged using a 63× water immersion objective (Zeiss) using a confocal microscope (Zeiss LSM510) under the same conditions. Scans from each channel were collected in multiple-tracks mode and subsequently merged. Care was taken to use the lowest laser power, and no bleedthrough was visible between the Alexa594 and Alexa488 channels.
SHIFT analysis. All confocal images were acquired using the same microscope setting. Confocal stacks were first merged using maximum transparency setting. The maximum Z stack used was 2 μm. Using the ImageJ software, the green (GFAP) and the red (GAD65) channels were then separated and transformed into grey-level 8-bit images before being thresholded. The minimum size of GAD65 puncta was set to between 12 to 750 pixels (signals smaller or larger would not count it). The total number of GAD65 puncta (X) was measured using the dot counting function of ImageJ. The grey color images of GFAP and GAD65 were then remerged. Since both images were grey, those GAD65 puncta that colocalized with GFAP were fused into the GFAP signals (as “bubbles along fiber”-like patterns) and would be excluded from the counting procedure set above. Thus in the remerged image, only the GAD65 puncta that were not colocalized with GFAP (Y) were counted. With this approach, we were able to count the number of GAD65 puncta in a nonbiased way using the counting function of ImageJ. We then obtained the percentage of GAD65 puncta that colocalized with GFAP as (X − Y) / X * 100. The same procedure was used before and after shifting the GFAP image relative to that of GAD65 by ±5 μm. The Wilcoxon signed rank test was used for paired comparisons of GAD65 puncta density and colocalization with GFAP. Significance was set at p < 0.05, and values are means ± standard deviation (s.d.)
Analysis of stellate axon orientation and their association with GFAP. All analyses were done blind to the genotype. All ascending and descending axonal branches with a length greater that 4 μm was included in our analysis. Axonal branches of individual neurons were visualized with GFP in our PV-GFP (B20) mice (eight neurons in at least five different mice in each genotype). From each selected axonal branch, its length (X) was first measured with the LSM confocal software (Zeiss). The axonal length that colocalized with GFAP (Y) was then measured; and the proportion of the branch that colocolized with GFAP was obtained as Y / X * 100. We measured the angle of each axon branch in relation to the pia surface, which was defined as a horizontal line in our projected image at the 0° angle. An axon branch that was perfectly perpendicular (ascending or descending axonal branches) to the pia would be at the 90° angle. We set a virtual horizontal line (the closest to the pia surface orientation) for each branch and measured its angle in relation to the virtual horizontal line. We took into account only branches with angles between 50° to 130°; and these angles of axon branches were grouped into 10° bins. Values in each bin were pooled together and analyzed with Kaleidagraph (Synergy Software) or Excel (Microsoft) software.
Paired comparisons of GAD65 density and colocalization with GFAP used the Wilcoxon signed rank test. Significance was set at p < 0.05. Parameter values are means ± s.d.
Antibodies against CHL1 (CS1123) were raised in chicken against a peptide with the sequence SLLDGRTHPKEVNILR corresponding to the region within the third FNIII domain of the protein plus an N-terminal cysteine for coupling. The specificity of the CHL1 antibodies was confirmed using CHL1−/− mice, and COS cells transfected a CHL1 expression construct (Figure S2). Production and IGY purification was done by Covance Immunology Service. Similar staining patterns, but higher intensity, were seen with a polyclonal antibody from R&D Systems.
Brains were perfusion fixed according to routine procedures as described earlier [63]. Briefly, deeply anaesthetized mice were transcardially perfused with a brief rinse in phosphate buffer, 0.1 M (pH 7.4), followed by a solution of 4% freshly depolymerized paraformaldehyde and 0.1% glutaraldehyde in phosphate buffer, supplemented with 2% PVP and 0.4% NaNO2. The brains were removed from the skull and left in the same fixative for at least several days. Sagittal vibratome sections of 50–60 μm were postfixed in 1% osmium tetroxide with 1% sodium ferricyanide in 0.1 M cacodylate buffer for 20 min, dehydrated in series of ethanol, and then flat embedded in epoxy resin. Semithin sections (1 μm) were cut and stained with toluidine blue and used for orientation purposes. Ultrathin sections of selected areas of the cerebellar cortex with reference to the ML were cut, using an ultratome LKB IV (Reichert-Jung). and collected on single-slot grids or 75-mesh grids coated with Formvar (Electron Microscopy Sciences) Ultrathin sections were contrasted with uranyl acetate and lead citrate, and analyzed in a Philips CM 100 transmission electron microscope (FEI Electron Optics).
Morphological analysis. Synapses were defined by the presence of a clear postsynaptic density facing a number of synaptic vesicles. By means of a goniometer, sections could be tilted in the beam, thereby determining the symmetry or asymmetry of the synaptic profiles. Measurements of profile length and diameter were made using a morphometric program (Soft imaging system SIS; Olympus).
P3–P5 pups were anesthetized with ketamine (0.56 mg/g; xylazine, 0.03 mg/gm body weight). After incision of the skin overlying the skull, a small hole was made directly over the left hemisphere of the cerebellum. A patch pipette filled with 1–2 μl of GFP DNA construct (endotoxin-free preparation) were injected directly into the tissue (1 μg/μl DNA), and mouse pups were subjected to electric pulses (four to six pulses at 200 mv/cm for 50 ms with intervals of 950 ms) by gold-plated electrode (BTX) placed directly on the skull. The skin was then sutured. After recovery from anesthesia, pups were returned to mother under standard housing. Mice were then sacrifice at P16 and analyzed.
|
10.1371/journal.pcbi.1002297 | Information Routing Driven by Background Chatter in a Signaling Network | Living systems are capable of processing multiple sources of information simultaneously. This is true even at the cellular level, where not only coexisting signals stimulate the cell, but also the presence of fluctuating conditions is significant. When information is received by a cell signaling network via one specific input, the existence of other stimuli can provide a background activity –or chatter– that may affect signal transmission through the network and, therefore, the response of the cell. Here we study the modulation of information processing by chatter in the signaling network of a human cell, specifically, in a Boolean model of the signal transduction network of a fibroblast. We observe that the level of external chatter shapes the response of the system to information carrying signals in a nontrivial manner, modulates the activity levels of the network outputs, and effectively determines the paths of information flow. Our results show that the interactions and node dynamics, far from being random, confer versatility to the signaling network and allow transitions between different information-processing scenarios.
| Far from being silent and static, the habitat of a cell is usually composed by multiple and simultaneous signals. We can consider nutrients, hormones, temperature, light, and other stimuli as elements building a default environment in which cells grow, divide and die. This environment, which has an intrinsically fluctuating nature, is the setting in which cells process all incoming stimuli. Here we examine the role that this background activity –or signaling chatter– plays in the transmission of information in a typical human cell. We address this question using a cellular model of signal transduction that we simulate using both random and periodic stimuli. We find that the level of background chatter determines the response of the whole signaling network to external stimuli. Different areas of the network are activated by specific levels of background activity, routing the information through chatter-dependent paths. In this way, different levels of chatter allow the network to select between different responses, given the same stimulus. These features depend on the architecture and functional connectivity of a truly biological network, since we find that randomized versions of the model are incapable of showing this behavior.
| Signal transduction, the process through which information about the extracellular environment is conveyed to the cell's interior, is a property of all living organisms. Signaling molecules stimulate their receptors, which transmit the signal downstream through a series of protein-protein interactions that ultimately modify DNA expression and protein levels [1], [2]. In this manner, external information affects cell behavior. This description of signal transduction has traditionally involved independent signaling cascades –or pathways–, in which information is linearly transmitted from membrane to nucleus. Correspondingly, experimental studies have usually analyzed pathway stimulation by single inputs, such as variations in a chemical signal (e.g. a nutrient or hormone) or a physical property (e.g. illumination or mechanical pressure). However, extracellular media often contain a complex mix of molecules that have the potential to feed the signaling network with multiple inputs simultaneously [3]. In addition, it is now known that proteins in one signaling cascade often interact with proteins of other pathways, forming a dense web of connections 4–6. Thus cells must be able to perform such complex information-processing tasks as signal integration [7]–[9] and multiplexing [10] while dealing with cross-talk [11], [12].
Adding to these information-processing requirements, the signaling machinery has to cope with the fact that a cell's environment is not stationary, but subject to fluctuations [13], [14]. Here we explore the impact of these environmental fluctuations on the information processing capabilities of the signaling network as a whole. In particular, we study how transmission of information from one single input node is affected by the fluctuating background activity, or chatter, provided by other network inputs. To address this issue in a way that explicitly accounts for the complexity of the system under consideration, we use one of the most comprehensive dynamic models of cell signaling currently available in the literature, a recently published Boolean network for the human fibroblast that involves over 130 protein species [15] (see Figure 1a; a fully annotated version can be found in Figure S1). The dynamics of this network are implemented as a set of logic rules, an approach that, despite its simplicity, represents a good choice when building a detailed kinetic model is unfeasible. Indeed, Boolean networks have successfully been applied to modeling numerous biological processes, including gene regulation [16]–[19], cellular differentiation [20], [21], developmental patterning [22], and signal transduction [23]–[26], evidencing that sequences of cellular events can be reproduced by this type of discrete dynamic models.
Given the ubiquity of periodic oscillations in cellular processes [27]–[29], we assume that the input signal is either periodic in time (structured signal) or erratic (unstructured). By performing extensive numerical simulations, we characterize the response of the fibroblast's signaling network to such periodic input signals under different chatter levels. Our findings suggest that the level of background activity shapes the response of the entire network to the input signal, thus providing a mechanism for context-dependent signaling [11] in dynamic situations.
The Boolean Network (BN) model used in this work was built by Helikar and coworkers [15] to describe the signaling pathways in a prototypical human fibroblast (see Supporting Text S1 for additional details). The network, which was created by careful inspection of a large body of experimental literature, contains 9 input nodes and 130 internal nodes (see Figure 1a). The input nodes represent signals of varying nature, namely stress signals, a growth factor, a calcium channel, signaling by extracellular matrix components and by ligands that use G-protein coupled receptors. Following the original work, we consider six of the 130 internal nodes to be outputs of the network, even though they also signal to other nodes. The choice of these six species (the proteins Akt, Erk, Rac, Cdc42, SAPK and p38) as network outputs was motivated by their role in regulating well-defined cellular processes: programmed cell death (apoptosis) in the case of Akt, gene transcription for Erk, cytoskeletal regulation for Rac and Cdc42, and, finally, cellular stress for SAPK and p38.
Mathematically, the BN used here (termed original network hereafter) consists of 139 elements, or nodes, connected by 542 links. Nodes represent chemical species, which are assumed to be either active or inactive, and edges represent their interactions. The states of all species are updated synchronously at each iteration according to a set of node-specific deterministic rules. Hence, the state of node (equal to 0 if the node is inactive and 1 if it is active) is completely determined by the states of its inputs at iteration and by its logic rule (see full details in the Supporting Text S1). Therefore, for given inputs and initial conditions of the network, the states of all nodes evolve in a deterministic and reproducible manner. To introduce the unpredictable evolution of the cell environment into the model, here we allow random fluctuations in the activity of the input nodes. We define a probability for an input to be active, and at each iteration of the network dynamics we draw the state of this input node from a Bernoulli distribution with probability of success equal to (i.e. for input node ). The parameter therefore determines the average level of background chatter in the network, which is lowest for and highest for . Note that additionally controls the degree of variability in the input sequence, which is maximal at , and disappears for both and .
In this context, a single realization of the network dynamics may be assumed to correspond to the behavior of a single cell, and the ensemble average of the activity over cell realizations at a fixed level can be regarded as the average cell population activity, . Note that is a continuous variable representing the proportion of cells in the population with species active at iteration . This ensemble representation, together with the chatter model introduced above, allows for a realistic description of stochastic fluctuations, as it dilutes the effect of flipping input-node states at the macroscopic level [30], provided the number of cell realizations is large enough. In this representation, the chatter levels correspond to the population average activities of the input nodes.
The random sequences of activity states obtained for inputs with a constant chatter level provide the network with an unstructured signal that lacks temporal information. Figure 1b shows the behavior of the output node p38 (red) for a given realization of the unstructured sequences for all inputs set at chatter level (for clarity, only the input stress is shown, in blue). All simulations below have been made for a duration of 1600 iterations, well beyond the span of any transient behavior [26] (which has been eliminated by removing the first 160 iterations before data analysis). Notice that both nodes show constant activity at the population level (see bottom plot in Figure 1b). In this case, averages over cell realizations are effectively equivalent to temporal averages. In this paper we consider an additional type of input sequence that does introduce temporal information: oscillating (structured) inputs whose states turn on and off periodically. We illustrate this case in Figure 1c, in which the input stress oscillates and the rest of inputs fluctuate with a fixed level of chatter, (only the input stress is shown). In this case, the average population activity of the output node p38 also oscillates at the frequency of the input, therefore recovering the temporal information supplied by the external stimulation. In the following we describe the conditions in which these structured and unstructured signals are transmitted, and the role played by background chatter during this process.
To study the contribution of chatter to the network dynamics, we first consider the response of the network to unstructured inputs. These have been implemented with a constant chatter level for all the input nodes. Under these conditions, the population activities of the output nodes fluctuate around a constant value that depends on the chatter level (see bottom plot in Figure 1b). In Figure 2a we show the temporal average of the population signal for all the output nodes, for increasing levels of chatter. The average population activity increases monotonously for three of the outputs (Akt, Erk and Rac) as the chatter level increases. In particular, we observe that the average activity of Erk is approximately proportional to the chatter level. On the other hand, the average population activities of the other three outputs (Cdc42, SAPK and p38), depends non-monotonically on the chatter, becoming maximal for an intermediate value of . Thus, the original network responds to background chatter in a nontrivial manner.
We now ask to what extent the effects of constant chatter described above can be attributed to the specific connectivity architecture of the fibroblast network being used here. In order to address this issue, we generate a family of random networks that maintain the topology (i.e. the setting of nodes and links) of the original network, while permuting randomly the update rules, thus changing the logic of each node (see Figures S2 and S3, and accompanying Supporting Text S1 for the full details). The results obtained from multiple realizations of this randomized altered-logic (AL) version of the network show that its response is, in general, not sensitive to the chatter level (Figure 2b). Therefore, the responsiveness to chatter is not guaranteed solely by the network topology, but seems to require a particular type of logic rules governing the dynamics of the nodes. To check whether this is indeed the case, we generate a second family of randomized networks, keeping now unaltered both the topology of the original network and the logic rules of the nodes, but randomly reassigning the inputs of each of the update rules (see Figure S2). This randomization is less severe than the previous one, since it maintains the type of logic rules in the network. We observe that networks of this altered-input (AI) family are sensitive to chatter levels in a similar way to the original experimentally-based network (Figures S4 and S5). Taken together, these results reveal that the biologically realistic network studied here responds in a nontrivial manner to a constant level of background chatter, and it is the distribution of logic rules of the network nodes, which is far from random (see network properties in the Supporting Text S1), that determines this responsiveness.
Here we study the ability of the network to process and transmit structured information under different levels of background chatter. In order to do so, we examine the response of the network to the periodic stimulation of a specific input node, maintaining the rest of inputs at a constant chatter level (see examples of realizations and population activity for this type of input in Figure 1c). Contrary to what is observed for unstructured inputs (Figure 1b), the output signals obtained in these settings do have temporal structure. This is illustrated in Figure 3a, which shows the population average, , for all the output nodes upon periodic stimulation of the input node stress. Some outputs (Erk and Cdc42) do not show a significant response to the periodic modulation of stress, while others (Akt, Rac, SAPK, and p38) do oscillate at the period of this input.
In order to quantify the amount of periodicity in the network's response, we calculate the power spectral density (PSD) of each output node (see Figure S6). Figure 3b shows the value of this quantity at the frequency of the stress modulation, , as a function of increasing levels. This plot shows that the response at the input frequency changes in a nontrivial manner as a function of the chatter level (specially for Akt, Rac, SAPK, and p38). The different outputs reproduce the input periodicity in distinct ways, in some cases displaying their maximum response at intermediate values. The stress-activated protein kinase (SAPK), for instance, seems to respond better to periodic stimulation by input stress at chatter levels close to . Another stress-activated protein, p38, is also an interesting example that presents two ranges of high response for chatters around and . This behavior implies that the most responsive output to stress varies as the chatter level changes. In the particular example shown in Figure 3b, the most responsive output under increasing chatter values follows the sequence SAPK, Akt, p38 and Rac. In this sense, the network acts as a system capable of selecting its dominant output depending on the degree of background activity.
We have studied in detail the periodic stimulation of every input node of the fibroblast network, and have found that they all show the same qualitative phenomenology (see Figure S7), with the exception of the ligands. These ligands differ from the rest of inputs in that they represent generic pathways. For example, ligands correspond to the signals that use the subunit of the G protein (which include epinephrin, glucagon, TSH, and more), while ligands (like acetylcholine, serotonin and angiotensin) use the subunit, etc. They possibly fail to respond because their logic has somehow been altered during the generalization process. For clarity of presentation, we continue focusing hereafter on the periodic perturbation of stress only.
We now examine the extent to which the chatter-dependent ability of the network to select its response depends on its connectivity. To address this question, we now perform the same numerical experiment for the two randomizations of the network described previously (and shown in Figure S2). As in the case of unstructured inputs discussed in the previous section, AL networks in which the logic at all nodes is randomized (right column in Figure S2) are again insensitive to chatter, and in fact they do not respond to the periodic input at all (Figures S7 and S8). In contrast to the case of unstructured inputs, where AI networks did respond to the levels of chatter (Figure S5), we observe here that this second family of randomized networks are barely able to sense the periodic input, and are thus unable to show a sensitivity of the output to background chatter. A particular example of the response of such weakly randomized network to a periodic modulation of the stress input is shown in Figure 3c. Only p38 responds at all in this case. Other realizations of the randomization and the responses to other inputs are displayed in Figure S7, showing a similar behavior. This suggests that the topological structure of the network and the distribution of logic rules of the nodes are not sufficient for a successful information processing, but the original –specific– logic rules for each node are needed.
Next we study how the information is transmitted from the stimulated input to the dominant output, and which nodes participate in this transmission. To that end, we calculate the power spectral density at the stimulation period for all network nodes, and the maximum cross-correlation (in absolute value) between the average signals of all pairs of nodes (see Supporting Text S1 for additional details). Figure 4 shows this information in the case of a periodic modulation of the input stress, and for two different values of the chatter level, and , which correspond to the conditions for which SAPK and p38, respectively, are dominant outputs (Figure 3b). A common feature of both panels in Figure 4 is that there are several internal nodes that reproduce quite well the periodic input signal (i.e. they have high values of the power spectral density at the input frequency), and which are usually connected to each other by high cross-correlation values. However, there are also important differences between the two chatter levels. For instance, when (Figure 4a), most of the nodes that transmit the signal from the stimulated node to the dominant output node (in this case, SAPK) are not so active for (Figure 4b), and a different set of nodes transmit the information from the stress input to p38, Rac and Akt, which now become dominant outputs. Together, these results show that chatter is able to select which output responds dominantly to a given input by determining the set of internal nodes that are most affected by the input. These nodes in turn signal downstream until a given output node is reached.
Figure 4 shows that the chatter level sets which groups of internal interconnected nodes convey information from the inputs to the outputs. These groups of nodes and the links between them constitute preferred paths of information transmission. In order to characterize which of these paths are dominant in transmitting information, we resort to optimization algorithms of graph theory. For each stimulated input and chatter level, we assign a weight to each edge of the network equal to the inverse of maximum cross-correlation (see Supporting Text S1 for additional details). Then, for each of the network outputs we use a shortest path algorithm [31] to identify those paths going from the stimulated input to the considered output with the minimal sum of weights. This approach is well suited for our problem, as it penalizes large paths, and paths where at least one edge has a low cross-correlation. Each of the paths found using this method is assigned a score equal to the inverse of the sum of weights. Thus, the higher the score of a path, the higher the correlations of its constituent interactions. Those paths with highest score in terms of sums of these weights are what we define as dominant paths. In Figure 5a, we show the score of the best paths found going from input stress to p38 as a function of the chatter level (see Figure S9 for the results corresponding to other input-output combinations). In Figure 5b, we show the nodes and interactions forming these paths. They are relatively short, as they usually involve between 3 and 7 intermediate species (structurally the network has 7 paths with 3 or less intermediate species from stress to p38, and 570 paths with 7 or less intermediate species). For low chatter levels, a group of paths emerges involving the MKK3 and MKK6 activation of p38. This group of paths is responsible for the first peak in the power spectral density of p38 shown in Figure 3c. At high levels of chatter, these paths fade out, and the oscillatory behavior of p38 (second peak in Figure 3c) becomes then due to inhibition by the MAP phosphatases (MKP), which in turn are activated through the adenyl cyclase (AC-cAMP) pathway. This is a specific prediction of our model, according to which the preferential pathway through which p38 is activated by stress changes with chatter. Since chatter can be varied by controlling the concentrations in the culture medium [32] of all input signals other than stress, it would be interesting to vary the medium composition of a fibroblast culture in a controlled way. The goal would be to measure the correlation between p38 activity and that of the main players of the two alternative pathways, e.g. MKK3 and AC, to check whether this correlation changes with medium composition. Similar predictions can be extracted for other input-output pairs.
From the results of the previous section it is clear that different dominant paths emerge as a consequence of varying chatter levels. The remaining question is what happens to those internal nodes of the network not involved in the aforementioned paths when chatter level varies. To address whether they significantly change their processing capacity, we analyze the sensitivity to chatter variation of the power spectral density at the input frequency, , and of the maximum cross-correlation (in absolute value) between edges, . We call these two magnitudes, respectively, node sensitivity [ for node ] and edge sensitivity [ for the interaction pair ] (see Supporting Text S1 for additional details).
Figure 6 summarizes the results obtained in the case when the network is driven by an oscillating stress signal (see Figure S10 for other inputs). Both nodes and edges are colored according to the maximum (in absolute value) sensitivity for all chatter levels. Blue (red) color indicates a positive (negative) variation in the direction of increasing chatter. Color intensity indicates the magnitude of this maximum variation. In this figure, it can be observed that just a few nodes and a few of the 542 edges of the network have a significant variation of power and correlation when varying chatter levels. Note that most of the sensitive nodes and edges are involved in one or more dominant paths at a given chatter level (see Figures 4 and 5). Thus, while species belonging to paths involved in information transmission are sensitive to variations in the chatter level, the rest of the network nodes seem to be robust against these variations.
Cells live in environments whose composition affects the way in which they function. An example is the interstitial fluid (IF) surrounding cells in higher organisms, which affects processes as important as embryogenesis, tissue morphogenesis, remodeling and cancer progression [33]. The composition of the IF changes over time as a function of tissue irrigation rate, inflammation, and organ motility, for example. Modifications of the IF are known to affect fibroblasts [34], supporting the view that these cells are exposed to varying environments. While being subject to purely external sources of variation, cells also contribute to modifying their surroundings by secreting multiple signaling molecules themselves [35]. In a given physiological situation only a small subset of those signals will carry information relevant to the cell [36].
Noteworthily, the information-carrying signals are frequently dynamical, since oscillations in cell physiology are ubiquitous [37], and in many cases clearly periodic, driven for instance by regular biological rhythms such as those generated by circadian clocks [29], [38]. The remaining signals may constitute a source of background activity, or chatter, that is bound to affect the cellÕs response to the relevant inputs. Experimental evidence hints at the existence of signaling fluctuations in different cell types. Transient fluctuations in phosphate signaling, for instance, exist in yeast [39]. T cells, on the other hand, are known to be activatable by small numbers of T-cell-receptor (TCR) ligands [40], [41], and can therefore be expected to undergo strong fluctuations in TCR signaling [32]. In the particular case of fibroblasts, considered in this paper, fluctuations in extracellular pH are known to exist [42], and transient deactivation of ERK signaling (a pathway specifically considered in the model above) has been associated with cell cycle control [43]. The level of chatter will depend on many variables, including cell type, tissue, developmental stage, health status, etc. In this paper we numerically examine how the information transmission capabilities of a periodically stimulated human cell depend on the amount of background chatter.
At the experimental level, context-dependent signaling is beginning to be unraveled. For example, the Alliance for Cell Signaling (AfCS) recently compared the effect of 22 individually applied inputs (cytokines, GPCR ligands, TLR ligands, and tyrosine-kinase receptor ligands) upon 42 cell outputs (cytokine production, protein phosphorylation, calcium, and cAMP levels), to the effect of all possible pairwise combinations of those inputs [11]. According to the results of that study, only a few ligands are able to control cellular outputs independently from the other inputs. In contrast, most inputs act as modulators of signal transduction, providing the cell with the ability to perform context-dependent signaling. Our results fit well with these findings, as we see that the level of background activity of the input nodes determines the capability of the cell to respond to other inputs (in particular, to follow both unstructured and structured signals). In their work, Natajaran et al. [11] coin the term interaction agent to refer to the network circuits that couple different signaling pathways. They claim that such circuits would be silent in single ligand experiments and become active upon multiple input signaling, causing the non-additive effects observed for certain pairs of inputs. In our theoretical study we effectively observe different areas of the network being used at specific chatter levels, thus supporting the existence of these circuits.
Our results show specifically that a detailed signaling network, carefully compiled from published experimental data [15], responds in a nontrivial manner to background chatter, both intrinsically and in the presence of a periodic modulation of one of the inputs. This work extends the findings of Helikar and colleagues, who created the network in the first place and studied its stationary response to different input levels for increasing intensities of noise [15]. They concluded the network divides biological stimuli into categories, since it reduces the full range of possible external inputs to a limited number of cell responses, in a manner that is robust to noise. We divert from the work of Helikar et al. in that we focus on the dynamics of the system. Having recently explored the relaxation time and frequency response of the network [26], we now show that chatter is able to enhance the response of certain outputs to a given input when tuned to optimal levels. Given that chatter controls the amount of stochasticity acting upon the network, this is a situation reminiscent of stochastic resonance, a phenomenon in many physical and neurosensory systems by which the detection of a weak signal is enhanced by noise [44]. Our simulations have been performed using synchronous updating but, as shown in Supporting Text S1 and in Figure S11, our results are qualitatively unaltered when the updating is asynchronous (whose main effect is the destruction of deterministic attractors, which is also caused by chatter). We also note that the temporal character of the chatter is relevant for the phenomena reported here, as discussed in Supporting Text S1 and in Figure S12.
Recent studies have shown that signaling networks prioritize dynamic range over signal strength [45]. This entails a linear relationship between the input signal and the output response of the network, which ensures that the reaction of the network to an oscillatory input will also be oscillatory with the same main frequency, for a large wide range of input amplitudes. Our results fit well with this finding, and extend it by assigning a relevant role to the background chatter coming from other input nodes, which enhances the frequency response. It would be interesting to extend these studies to the situation in which more than one information-carrying signal act upon the system, following the recent experimental studies of Natajaran et al. [11] and Hsueh et al. [36] discussed above, which have revealed synergistic effects in signal integration.
Our results also show that varying chatter levels allow the network to select which output nodes respond preferentially to a given input. Indeed, output switching is achieved via a mechanism that places few requirements on the temporal structure of contextual, non-specific signals. Thus, we conjecture that cells could use environmental noise (to which they are unavoidably subject) to choose among alternative information routes, and eventually among different cellular responses. Randomized versions of the original network in which the topology of the connections is maintained -and only the integration rules at the nodes are altered- fail to reproduce this property, indicating that the chatter-driven selectivity reported here is fine-tuned to the specific architecture and logic of the experimentally-supported network.
Concurrent with the ability of background chatter to select the dominant output for a given input, chatter also selects the network path through which information is transmitted. The nodes belonging to these preferred paths can be expected to form the classifier hyperspace proposed by Oda and Kitano in their study of the Toll-like receptor signaling pathway [5]. These preferred paths are sensitive to chatter and allow transitions between different information processing scenarios that underlie different output responses. In that way, a given signaling network can have multiple working states that are selected by the background chatter. The rest of the network nodes not belonging to the preferred paths, on the other hand, remains insensitive to chatter. In that way, we can conjecture that signaling networks have a built-in balance between responsiveness and robustness within their coupling architecture, and this balance is modulated by background chatter. Taken together, the results presented here indicate that background activity levels are key for determining the response of the cell to a given input, by allowing the emergence of novel system-level properties such as information routing, output switching, and context-dependent signaling.
|
10.1371/journal.pgen.1000316 | Gata3 Acts Downstream of β-Catenin Signaling to Prevent Ectopic Metanephric Kidney Induction | Metanephric kidney induction critically depends on mesenchymal–epithelial interactions in the caudal region of the nephric (or Wolffian) duct. Central to this process, GDNF secreted from the metanephric mesenchyme induces ureter budding by activating the Ret receptor expressed in the nephric duct epithelium. A failure to regulate this pathway is believed to be responsible for a large proportion of the developmental anomalies affecting the urogenital system. Here, we show that the nephric duct-specific inactivation of the transcription factor gene Gata3 leads to massive ectopic ureter budding. This results in a spectrum of urogenital malformations including kidney adysplasia, duplex systems, and hydroureter, as well as vas deferens hyperplasia and uterine agenesis. The variability of developmental defects is reminiscent of the congenital anomalies of the kidney and urinary tract (CAKUT) observed in human. We show that Gata3 inactivation causes premature nephric duct cell differentiation and loss of Ret receptor gene expression. These changes ultimately affect nephric duct epithelium homeostasis, leading to ectopic budding of interspersed cells still expressing the Ret receptor. Importantly, the formation of these ectopic buds requires both GDNF/Ret and Fgf signaling activities. We further identify Gata3 as a central mediator of β-catenin function in the nephric duct and demonstrate that the β-catenin/Gata3 pathway prevents premature cell differentiation independently of its role in regulating Ret expression. Together, these results establish a genetic cascade in which Gata3 acts downstream of β-catenin, but upstream of Ret, to prevent ectopic ureter budding and premature cell differentiation in the nephric duct.
| In humans, kidney development originates during embryonic development by the sprouting of an epithelial bud—called the ureteric bud—from a simple epithelial structure—the nephric duct. The ureteric bud quickly grows and branches in a treelike fashion to form the kidney collecting duct system, while the emerging ureteric tips induce nephron differentiation. One of the most important steps during kidney development is the positioning of a single ureteric bud along the nephric duct, since mutations of genes implicated in this process lead to severe urogenital malformations. In this study, we identified the Gata3 protein as a crucial regulator of ureteric bud positioning by using genetically modified mice. Deleting the Gata3 gene in the mouse resulted in the development of multiple kidneys emerging at improper positions. We show that this defect was caused by a hypersensitivity of nephric duct cells in their response to local growth signals. Interestingly, this phenomenon was partly triggered by premature differentiation of a subset of nephric duct cells. Furthermore, we report a genetic pathway in which Wnt/β-catenin signaling activates the Gata3 gene, which in turn positively regulates the Ret gene. In summary, we introduce a mouse model system that can be used to study human birth defects affecting the urogenital system.
| In human, urinary tract anomalies rank among the most common birth defects, with an estimated occurrence of 1 in 250 live births [1]. Most of these ontogenic malformations are classified as Congenital Anomalies of the Kidney and Urinary Tract (CAKUT) [2], which is a highly heterogenous condition frequently diagnosed in combination with genital tract anomalies [3]. The most relevant clinical manifestations include absent, dysplastic or obstructed renal systems in infants as well as infertility, pregnancy complications, hypertension and chronic renal failure in adults [4].
The development of the urogenital system (UGS) begins with the formation of the nephric duct (or Wolffian duct) [5],[6]. This epithelial duct is a central UGS component among all vertebrates and serves as the primordium for the ureter, kidney collecting duct system and male genital tract [7]. Upon its induction in the intermediate mesoderm at embryonic day (E) 8.5 in the mouse, the nephric duct rapidly elongates caudally until it reaches the cloaca, a pouch from which the bladder and urethra later develop. At E10.5, the formation of the definitive (metanephric) kidney is initiated by sprouting of the ureteric bud from the nephric duct into the adjacent metanephric mesenchyme. The ureteric bud subsequently undergoes several branching cycles to form the collecting duct system, whereas the ureter tips induce nephron formation in the surrounding mesenchyme [8].
Ureteric bud outgrowth and positioning are among the most crucial steps of UGS development, since anomalies at the budding stage account for the majority of kidney and urinary tract developmental defects [9],[10]. Extensive evidence has identified GDNF/Ret signaling as a central regulator of ureteric bud induction [11]–[16]. In this system, GDNF secretion in the metanephric mesenchyme activates the Ret receptor tyrosine kinase via its ligand binding GFRα1 co-receptor. In turn, Ret activation results in the initiation of intracellular signaling cascades, which mediate bud outgrowth, proliferation and subsequent ureter branching. Understandably, the activity of GDNF/Ret signaling is tightly regulated by various mechanisms to allow for the generation of a single ureter at the appropriate position [9]. In the mesenchyme surrounding the nephric duct, the forkhead transcription factor FoxCI and the Slit2/Robo2 ligand-receptor pair repress the rostral expression of GDNF [17],[18], while Bmp4 antagonizes its activity [19]. At the budding site, Gremlin releases GDNF inhibition by Bmp4 [20], thereby allowing ureteric bud outgrowth. In the nephric duct epithelium, Sprouty1 protein function is crucial to negatively modulate Ret signaling levels [21]. Of notice, gene mutations in any of these regulators of GDNF/Ret signaling result in CAKUT-like phenotypes in mice. In support of the clinical importance of these genes, a growing number of them are found mutated in human developmental diseases affecting the UGS [22]–[26].
Gata3 is a transcription factor of the Gata Zn-finger family, which perform important functions during organogenesis [27]. In humans, GATA3 haploinsufficiency causes hypoparathyroidism, sensorineural deafness and renal anomalies (HDR) syndrome [28]. The urogenital defects of HDR patients closely resemble CAKUT in combination with genital tract anomalies and include renal aplasia, dysplasia, hypoplasia and vesicoureteral reflux [29],[30]. Gene ablation studies in mice further revealed a critical role for Gata3 in the development of several tissues [31],[32]. In the urogenital system, Gata3 is necessary for proliferation control and guidance of the nephric duct [33]. Accordingly, it is the only Gata factor expressed in this tissue prior to E12.5 [34].
Here we report the conditional inactivation of Gata3 specifically in the nephric duct, at a stage past the developmental defects observed in germline knockout embryos. These mice display multiple UGS malformations affecting the kidney, ureter and genital tracts. The detailed analysis of this phenotype reveals a genetic cascade whereby β-catenin promotes Gata3 expression in the nephric duct, which in turn activates Ret expression, maintains an undiffererentiated epithelial cell state and prevents the inappropriate response to signaling pathways promoting ureter budding.
The strong mesonephric phenotype observed in Gata3−/− embryos [33] precludes the study of Gata3 function later during urogenital system development. To investigate the role of Gata3 in ureteric bud formation, we first generated a Gata3 conditional loss of function allele (Figure 1A,B). For this, the parental Gata3 allele (Gata3ex4GFP) [33] was crossed with a transgenic strain expressing FLPe in the germline [35] to excise the GFP-neo reporter cassette, thereby generating a conditional Gata3 allele with loxP sites flanking exon 4 (Gata3flox). These mice were subsequently bred with the More-Cre germline deleter strain [36] to generate the Gata3Δ allele in which exon 4 is removed. The splicing from exon 3 to exon 5, expected from this modification, would generate a frameshift leading to protein truncation just upstream of the first zinc-finger DNA-binding domain. This gene mutation is therefore predicted to be null, like other previously reported Gata3 mutant alleles [32],[33],[37]. To delete Gata3 specifically in the maturing nephric duct and derived collecting duct system, we crossed Gata3flox mice with the HoxB7-Cre mouse strain [38] to generate HoxB7-Cre; Gata3flox/Δ and HoxB7-Cre; Gata3flox/flox embryos. Both genotypes had the same phenotype and are subsequently referred to as Gata3ND−/− embryos. Gata3flox/flox, Gata3flox/+ and HoxB7-Cre; Gata3flox/+ embryos failed to show any overt phenotype and were used as controls.
We initially characterized whole dissected urogenital systems (UGS) of Gata3ND−/− embryos at embryonic day (E)18.5. This gross analysis revealed a broad variety of malformations affecting the kidneys and genital tracts. Kidney defects included agenesis (15%), aplasia (20%) and severe dysplasia (65%) (Figure 2A–G). Moreover, one third of the Gata3ND−/− embryos displayed duplex kidneys (arrows in 2F). Histological analysis further revealed that the dysplastic kidneys were associated with hydronephrosis and hydroureter (Figure 2H,I). As expected from embryos with such poor kidney endowment, no Gata3ND−/− pups could be recovered after birth.
In addition to these renal defects, over 80% of male mutant genital tracts displayed a massive enlargement of the vas deferens in comparison to control embryos (Figure 2A,B,J,K). In female embryos, Gata3 inactivation in the nephric duct resulted in a complete loss of uterus in over 85% of UGSs examined. The oviduct, however, was still present in these embryos (Figure 2C,D,L,M). Hence the conditional inactivation of Gata3 leads to a broad spectrum of urogenital defects, including a high incidence of hydronephrotic kidneys and hydroureters.
The combination of kidney hydronephrosis, hydroureter and duplex kidneys seen in Gata3ND−/− embryos pointed to a primary defect at the level of ureter budding. In order to easily visualize nephric duct cells, we crossed the Rosa26STOPLacZ allele [38] into control and Gata3ND−/− genetic backgrounds and stained the UGSs for β-Galactosidase activity.
During normal development, a localized swelling of the caudal portion of the nephric duct indicates the site of ureteric bud outgrowth at E10.5 (Figure 3A). The bud quickly emerges and undergoes the first dichotomous branching event at E11.5, forming the T-stage kidney (Figure 3C). Subsequently, the ureter lengthens and multiple branching cycles mark the development of the metanephric kidney (Figure 3E). In Gata3ND−/− mutant embryos at E10.5, nephric duct swelling was sometimes observed (Figure 3B), but normal ureteric bud formation failed in most embryos analyzed. In addition, the nephric duct looked more sinuous in appearance (Figure 3B). Strikingly, at E11.5, ectopic epithelial buds formed along the entire length of the nephric duct, with a preferential accumulation in the middle segments of the duct (Figure 3D). By E12.5, the majority of the ectopic buds started to regress, while some buds expanded to form ectopic kidneys at a position far more rostral than the normal kidney induction site (Figure 3F). Following ectopic bud regression, the male nephric duct started to enlarge (Figure 4A,B) This was accompanied by a slight increase of about 50% in cell proliferation index, as determined by phospho-histone H3 immunolabelling at E13.5 (Figure 4 C,D). In females, in situ hybridization with cRNA probes against Emx2 and Wnt4, staining the Müllerian duct epithelium and mesenchyme, respectively, revealed a block in female genital tract elongation in the region where most ectopic ureteric buds occurred (Figure 4E–H). Since the expression of the key Müllerian duct regulators Wnt4, Emx2, Wnt9b and Lim1 was normal in these embryos (Figure 4E–H and data not shown), it is possible that this elongation defect is simply caused by a physical obstruction by ectopic ureteric buds. Together, these results indicate that the urogenital defects observed in Gata3ND−/− embryos are largely caused by the emergence of ureteric buds at aberrant positions along the nephric duct.
To determine the cause of ectopic ureter budding in Gata3ND−/− embryos, we first looked at Ret expression by in situ hybridization. Consistent with the previous analysis of Gata3 germline knockout embryos [33], most Gata3ND−/− nephric duct cells lost Ret expression. Curiously, however, the ectopic buds of Gata3ND−/− embryos remained positive for Ret (Figure 5A,B). In situ hybridization with a cRNA probe against Gata3 exon 4 (which is excised in the Gata3 conditional allele) confirmed that the bud cells still expressed Gata3 (Figure 5C,D), whereas nephric duct cells had already lost Gata3 expression (Figure 5C–F). The remaining Ret expression therefore resulted from the incomplete action of Cre in the nephric duct at this stage. Interestingly, these ectopic buds consisted of Gata3+/Ret+ and Gata3−/Ret− cells segregated from each other (Figure 5A–D). The continuous inactivation of Gata3 as well as the downregulation of Ret expression was confirmed in dysplastic Gata3ND−/− metanephric kidneys derived from such ectopic buds (Figure 5G–J). The severity of kidney dysplasia in these embryos correlated well with the amount of remaining Ret expression (Figure 5J insert, data not shown). As expected, these dysplastic kidneys additionally showed severely impaired nephron differentiation, as evidenced by the strong reduction in Fgf8-positive nephron precursors in comparison to control embryos (Figure 5K,L).
Since Ret expression was specifically localized in the ectopic ureteric buds of Gata3ND−/− embryos (Figure 5A,B), we hypothesized that Ret signaling might be causally involved in the budding process. To investigate whether the ectopic buds required GDNF/Ret signaling, we performed organ culture experiments with dissected Gata3ND−/− and control UGSs starting at E10.5. After 42 hours in culture, their development progressed, forming either T-stage metanephric kidneys in control UGSs or multiple ectopic buds in Gata3ND−/− UGSs (Figure 6A,B). Interestingly, addition of the Ret tyrosine kinase inhibitor SU5416 [39] efficiently suppressed ectopic budding in Gata3ND−/− UGSs in culture (Figure 6D) as well as primary ureteric bud formation in control UGSs (Figure 6C,D). To further assess the role of GDNF/Ret signaling in ectopic bud formation, we treated cultured UGSs with recombinant GDNF [40]. This treatment, sufficient to induce ectopic budding in control UGS cultures (Figure 6E), significantly increased the size of ectopic buds in Gata3-deficient cultures (Figure 6F, compare with 6B). In addition, an antibody blocking GDNF activity successfully inhibited both primary and ectopic bud formation in control and Gata3ND−/− UGS cultures, respectively (Figure G,H), thereby demonstrating the direct implication of GDNF/Ret signaling in ectopic ureteric bud formation. Interestingly, a 50% lower concentration of GDNF blocking antibody still inhibited primary bud induction but failed to inhibit the formation of ectopic buds (data not shown), suggesting the presence of additional factors involved in the budding process. One candidate is the Fgf signaling pathway, which has the potential to induce ectopic ureter budding [41],[42]. Supplementing the culture medium with a soluble FgfR2-Fc chimeric protein [43] did not perturb primary ureteric bud outgrowth in control cultures (Figure 6I). In striking contrast, however, it could efficiently suppress ectopic ureter budding in Gata3ND−/− cultures (Figure 6J). In conclusion, these results demonstrate that the generation of the ectopic buds in Gata3ND−/− embryos is mediated through the combined action of GDNF and Fgf signaling activities.
To study ectopic bud formation at the molecular level, we looked at the expression of key components of the GDNF/Ret signaling cascade in E10.5 and E11.5 embryos. Already at E10.5, prior to budding, Ret expression was lost in a subset of nephric duct cells in Gata3ND−/− embryos (Figure 7A,B). Staining of adjacent sections with a GDNF in situ probe revealed that the remaining Ret-expressing cells were concentrated near the mesenchymal source of GDNF (Figure 7C,D). At E11.5, only the caudal ectopic buds that could maintain GDNF and Pax2 expression in the surrounding mesenchyme were able to develop into ectopic kidneys (Figure 7E,F and data not shown). In some Gata3ND−/− embryos, the forming buds induced more Wnt11 expression than the presumptive ureteric bud induction site in control embryos, pointing to an elevated GDNF/Ret signaling response in those cells (Figure 7G,H). By E12.5, the expression of Wnt11 and Ret was lost in the regressing ectopic buds of Gata3ND−/− embryos (data not shown). Surprisingly, assessing the expression of the GDNF/Ret modulators Spry1, Spry2, Spry4, Slit2, Robo2, FoxcI, FoxcII, Bmp4 and Grem1, only revealed a modest downregulation of Spry1 and Slit2 expression in Gata3-mutant nephric duct cells at E10.5, while the other markers remained unaffected (Figure 7I–L and data not shown). In order to further support the involvement if Fgfs in ectopic bud formation, we additionally probed control and Gata3ND−/− embryos with the soluble Fgfr2-Fc fusion protein, which recognizes several Fgfs [44]. This experiment revealed strong and similar Fgf expression levels surrounding the nephric duct in both Gata3-mutant and control embryos at E11.5 (Figure 7M,N). This protein localization was consistent with the expression of one of the known FgfR2 ligands, Fgf10 (Figure 7O,P). Hence, the molecular marker analysis reveals no obvious mesenchymal defects, pointing to a primary defect in the response of the nephric duct to GDNF and Fgf signals.
The developmental defects of Gata3ND−/− embryos described above show striking similarities with the phenotypes resulting from the conditional inactivation of β-catenin (ctnnb1) in the nephric duct [45]. The fact that both animal models have the same spectrum of genital tract and kidney defects, prompted us to verify whether Gata3 and β-catenin act in the same genetic pathway. For this, we first performed in situ hybridization against Gata3 in E11.5 HoxB7-Cre; Ctnnb1flox/− embryos (Ctnnb1ND−/−) and found a strong downregulation of Gata3 expression in the nephric duct in comparison to control embryos (Figure 8A,B). Accordingly, in situ hybridizations with a Ret cRNA probe revealed a loss of Ret expression in Ctnnb1ND−/− embryos (Figure 8C,D), which mimics the loss of Ret expression in Gata3ND−/− embryos (Figure 5). Staining of adjacent tissue sections with in situ probes for the canonical Wnt target genes Axin2, Sp5 and Daple [46],[47] confirmed the loss of β-catenin transcriptional response in Ctnnb1ND−/− embryos (Figure 8E,F and data not shown). In order to clarify the genetic hierarchy between Gata3 and β-catenin, we next stained Gata3ND−/− embryos with antibodies against β-catenin and phospho-β-catenin. These experiments failed to show any modification of β-catenin expression levels or activity following Gata3 inactivation (Figure 8G,H and data not shown). In support of this, the expression of the canonical Wnt-signaling target genes Axin2, Daple and Sp5 was unchanged in Gata3ND−/− embryos (Figure 8I–L and data not shown). From these data, we conclude that Gata3 acts genetically downstream of β-catenin but upstream of Ret in the nephric duct. To further characterize the molecular basis of the β-catenin-Gata3-Ret pathway, we first performed a bioinformatics analysis of the 50 kb genomic region upstream of the mouse and human Gata3 genes. Highly conserved sequences shared by the two species were further analyzed for putative binding sites for TCF/Lef. In total, 10 putative TCF/Lef binding sites, located in 5 of the 8 conserved regions, were identified (Figure 9A). In this analysis, we additionally included a Gata3 urogenital enhancer located at −110 kb [48], but failed to detect any conserved TCF/Lef binding sites in this element. A similar bioinformatics analysis of the Ret regulatory region revealed putative Gata3 binding sites in 4 of the 11 conserved regions 50 kb upstream of the Ret-ATG. (Figure 9B). The potential of β-catenin to regulate endogenous Gata3 expression was further evaluated in mouse IMCD3 collecting duct-derived cells. Using the GSK3β inhibitor BIO to stabilize the β-catenin protein [49], we observed a significant increase of Gata3 expression levels in those cells (Figure 9C). In order to assess the activity of Gata3 on the Ret regulatory region, we took advantage of the fact that one of the Gata3 binding sites mapped to a region previously reported to drive reporter gene expression in the zebrafish pronephros (Figure 9B, large asterisk) [50]. To test whether Gata3 acts on this site, we isolated the 1.2 kb conserved fragment and introduced a specific point mutation in the Gata3 binding site (Figure 9B). The wild-type and mutated fragments cloned upstream of a β-Gal reporter construct were transfected in IMCD3 cells stably expressing Gata3. The inactivation of the Gata3 binding site led to a significant downregulation of β-Gal expression relative to the wild-type control (Figure 9D), thereby suggesting that Gata3 may regulate Ret expression directly from this binding site. Together these data are consistent with a β-catenin-Gata3-Ret genetic cascade in the nephric duct.
Another aspect of β-catenin-loss in the nephric duct is a premature differentiation of the affected cells [45]. To test whether this phenotype was also present in Gata3ND−/− embryos, we stained mutant and control embryos with the differentiation markers DBA and Zo1+. In the developing metanephric kidney, DBA and Zo1+ are strongly expressed in the distal collecting duct and downregulated in Ret+ ureter tip cells [45],[51]. At E11.5, DBA and Zo1+ expression were almost undetectable in the nephric duct of control embryos (Figure 10A,B). However, Gata3-deficient cells upregulated both differentiation markers already at this stage, suggesting a premature acquisition of a collecting duct identity in these cells (Figure 10C,D). To clarify whether this premature differentiation was a result of Gata3 inactivation only or secondary to the loss of Ret expression in Gata3-mutant cells, we next tested the expression of the two differentiation markers in Ret mutant embryos. Interestingly, no difference in DBA and Zo1+ expression levels could be observed between wild-type and Ret deficient embryos (Figure 10A,B and E,F). To quantifiy the differentiation defect in Gata3ND−/− embryos, we counted the total amount of DBA-positive nephric duct cells in the three different genotypes. Strikingly, we observed a ten-fold increase in differentiated nephric duct cells in Gata3-deficient embryos when compared to control or Ret mutant embryos (Figure 10G). From these data, we conclude that Gata3 acts independently of Ret to maintain a precursor state in the nephric duct epithelium, downstream of β-catenin.
We previously reported the critical role played by Gata3 in proliferation control and nephric duct guidance in the pro/mesonephros [33]. To circumvent this early renal phenotype and the embryonic lethality of Gata3−/− embryos at midgestation [31]–[33], we generated a conditional knockout allele of Gata3. In order to specifically address the later role of Gata3 in urogenital system morphogenesis, we inactivated Gata3 in the nephric duct using the HoxB7-Cre transgenic line (Gata3ND−/−). This resulted in severe malformations of the urogenital system including kidney agenesis, aplasia, dysplasia, duplex systems, uterine agenesis and vas deferens hyperplasia. This spectrum of malformations overlaps with the urogenital phenotypes observed in HDR syndrome patients (heterozygous for GATA3) and is generally reminiscent of human congenital anomalies of the kidney and urinary tract (CAKUT). Interestingly, most of these phenotypes can be either directly or indirectly attributed to ectopic ureteric budding observed in Gata3ND−/− embryos. Furthermore, our results identify Gata3 as a critical mediator of β-catenin signaling, which regulates both cell differentiation and Ret expression in the nephric duct.
The UGS malformations observed in Gata3ND−/− embryos were strikingly similar to the ones reported recently for Ctnnb1ND−/− embryos [45]. Gene expression analyses in both genotypes revealed a genetic cascade whereby Gata3 acts downstream of β-catenin in the nephric duct to maintain Ret expression and prevent premature epithelial differentiation. These genetic interactions were further supported by promoter analyses and cell culture assays. Hence, our results identify Gata3 as a crucial mediator of β-catenin activity in the nephric duct. Using Ret−/− embryos, we could further determine that the premature differentiation phenotype observed in Ctnnb1 and Gata3 mutant embryos is not mediated by Ret, thereby establishing at least two distinct cellular functions regulated by the β-catenin/Gata3 pathway (Figure 11A). The possibility remains that β-catenin also has a Gata3-independent effect on Ret expression. However, our data indicate that, if present, this effect is not sufficient for Ret expression in the absence of Gata3. The identification of β-catenin acting upstream of Gata3 also raises the question of the transcriptional control of Gata3 expression in the nephric duct. The only other regulators of Gata3 identified in this tissue are the transcription factors Pax2 and Pax8 [33]. It is thus possible that Pax2/8-mediated activation of Gata3 acts through the β-catenin pathway. Alternatively, Pax2/8 and β-catenin may act independently to regulate Gata3 expression in the nephric duct either together or as activation and maintenance factors, respectively. We favor the latter model based on the observation that the canonical Wnt signaling readouts Axin2 and Daple are not yet expressed in 18-somite stage embryos (E9.0; D.G., M.B. unpublished results), whereas Gata3 is already under the control of Pax genes at this stage [33]. The recent identification of a Gata3 kidney enhancer active in the nephric duct may help clarify some aspects of Gata3 regulation in this tissue [48].
Defects in ureteric budding account for most UGS defects in humans and mice. The underlying cause is typically a deregulation of GDNF/Ret signaling [9]. Accordingly, our data show a direct implication of GDNF/Ret signaling in the formation of Gata3ND−/− ectopic buds. Recombinant GDNF treatment indeed sustained ectopic bud growth in organ culture, while blocking the signaling pathway with a GDNF blocking antibody or a chemical inhibitor against Ret was sufficient to prevent ectopic ureter budding. We also demonstrated that the emerging buds consist of Ret+ cells, which respond to mesenchymal GDNF by upregulating Wnt11 expression. Importantly, however, we show that Ret expression was lost in Gata3-mutant nephric duct cells and that these Gata3−/Ret− cells failed to contribute to ectopic ureteric bud formation. This suggests that ectopic ureteric buds form as a consequence of Gata3 loss in neighboring nephric duct cells. It further implies that Ret-mediated signaling played a role in the segregation of Gata3+/Ret+ from Gata3−/Ret− cells. Interestingly, a biased contribution of Ret+ and Ret− cells was also observed in the ureteric bud and metanephric kidney of embryos chimeric for the Ret gene [40]. However, the precise mechanism leading to Gata3+/Ret+ and Gata3−/Ret− cell segregation remains unclear.
Importantly, a number of observations argue against a simple GDNF/Ret-based budding mechanism. Among them are 1) the different sensitivity of ectopic buds towards GDNF-blocking antibody concentration and 2) the fact that ectopic buds emerge in all directions whereas GDNF is expressed in the inner intermediate mesoderm only. Together, these observations point to the existence of a parallel pathway to GDNF/Ret promoting ectopic ureter budding in Gata3ND−/− embryos. Using a soluble recombinant Fgf receptor 2, we identified Fgf signaling as a crucial component of this alternative pathway. In support of this, Fgfs are able to induce ectopic budding in organ culture [41],[42]. Of notice, the RTK inhibitor (SU5416) we used to inhibit Ret signaling in culture was also reported to have an effect on FgfRs [39], which may have enhanced its activity against ectopic ureteric bud formation.
The genetic regulation of Ret by Gata3 corroborates our previous observations in germline Gata3 mutant embryos [33], but raises the intriguing question of the relationship between Gata3+/Ret+ budding cells and their surrounding Gata3−/Ret− cells in ectopic ureteric bud formation. Remarkably, the expression of the GDNF/Ret signaling regulators, Foxc1, Foxc2, Robo2, Bmp4, Grem1 and GDNF was not significantly affected in Gata3ND−/− embryos. The modest dowregulation of Spry1 and Slit2 expression was additionally restricted to non-budding Gata3-mutant cells and might be secondary to the loss of Ret expression. Hence, none of the major known mechanisms of GDNF/Ret modulation are likely to cause ectopic ureter budding in Gata3ND−/− embryos. Instead, this suggests that the juxtaposition of nephric duct cells with different Gata3/Ret status is central to the ectopic budding phenotype. Hence, the simplest model to explain ectopic ureteric bud formation would be that the combined expression of GDNF and Fgf along the defective Gata3ND−/− nephric duct at E10.5 is sufficient to initiate ectopic bud formation by wild-type nephric duct cells. At E11.5, the source of GDNF becomes restricted to the metanephric mesenchyme adjacent to the caudal region of the nephric duct. Thus, only the ectopic buds which are close enough to this mesenchymal source of GDNF and still harbor a sufficient number of Ret+/Gata3+ cells will be able to sustain the GDNF/Ret/Wnt11 positive feed back loop [52] and continue to grow to form ectopic kidneys. At this stage, most of the rostral buds are devoid of local GDNF expression and will therefore regress.
Importantly, the trigger for the aberrant responses to GDNF and Fgf inductive signals necessarily results from the loss of Gata3 in the surrounding nephric duct cells (Figure 11B). Ectopic budding may thus be a consequence of the premature differentiation and changes in adhesive properties observed in Gata3-mutant cells, as reported for Ctnnb1ND−/− embryos [45]. Alternatively, the loss of Gata3 may result in the activation of yet unidentified signals affecting the behavior of neighboring cells (Figure 11B). By avoiding these aberrant cellular responses in wild-type embryos, Gata3 maintains epithelial homeostasis and thus prevents the formation of ectopic buds by GDNF and Fgf in nephric duct regions that normally express their respective receptor.
In humans, heterozygous inactivation of the GATA3 gene is responsible for HDR syndrome [28],[30]. As many as 35 of the 38 mutations identified to date result in a loss of GATA3 DNA binding activity, mostly as a consequence of complete gene deletion or truncation of GATA3 DNA-binding domains [53]. As in Gata3ND−/− embryos, this GATA3 haploinsufficiency in HDR patients leads to a variety of UGS defects such as kidney agenesis, dysplasia, hypoplasia, sometimes associated with female genital tract malformations [29],[30]. This suggests that a critical threshold level of GATA3 is necessary to perform its normal function in the UGS and that cells or tissues expressing GATA3 below this level behave aberrantly. The Gata3 threshold is apparently reached at 50% gene dosage in human. Such a threshold, however, is not observed in Gata3 heterozygous mice [32]. Instead, UGS anomalies in mice are only observed at Gata3 expression levels below 50% of the wild-type levels [48], complicating the use of Gata3 mutant mice as disease models. The mosaic inactivation we observe in Gata3ND−/− embryos may, in fact, better mimic the threshold mechanism of HDR malformations whereby some cells respond normally to Gata3 while others behave aberrantly. In this respect, it is possible that the urogenital anomalies of HDR patients are caused by premature nephric duct progenitor cell differentiation and involve an aberrant response to GDNF and Fgf signaling. Hence, the Gata3ND−/− embryos may provide a suitable model to understand the molecular mechanisms and morphogenetic defects underlying HDR syndrome and other CAKUT-related UGS diseases.
Conditional and germ-line Gata3 knockout mice were generated by crossing Gata3ex4GFP mice [33] with ACTB-FLPe [35] and More-Cre [36] transgenic mice. The Gata3flox and Gata3Δ alleles were genotyped using the primers 5′-TATCAGCGGTTCATCTACAGC and 5′-TGGTAGAGTCCGCAGGCATT. HoxB7-Cre and Rosa26STOPlacZ [38],[54] mice were purchased from The Jackson Laboratory. All mice were kept in a pure C57BL/6 genetic background. Mutant mice for Ctnnb1 [45] and Ret [16] were generated as described. The HoxB7-Venus line will be reported elsewhere.
Embryo dissections and processing as well as in situ hybridization on whole mount embryos or tissue sections were performed as described [55],[56]. The Gata3-exon 4 probe was generated by PCR amplification of the coding sequence of Gata3 exon 4 and subsequent cloning into the pGEM-T-easy vector. The Daple (Ccdc88c) in situ probe was transcribed from image clone 6408630 linearized at an internal Asp718 restriction site. The Emx2 [57] Wnt4 [58], Ret [59], Gata3 [60], Axin2 [61], Fgf8 [62], GDNF [63], Wnt11 [52], Slit2 [64], Spry1 [65] and Fgf10 [66] in situ probes have been reported previously. Hematoxylin and eosin stainings were performed on 6 µm-thick paraffin sections using standard procedures.
Frozen sections have been prepared for immunohistochemistry as described [67]. For immunostainings against β-catenin and Zo1+, an antigen retrieval step has been included [45]. The following antibodies and conjugates were used: rabbit anti-phospho-H3 (1∶200, Upstate Biotechnology), rat anti-E-cadherin (1∶400, Zymed Laboratories), mouse anti-β-catenin (1∶400, Sigma), rabbit anti-phospho-β-catenin (1∶100, Cell signaling), rat anti-Zo1+ (1∶400, Chemicon), rabbit anti-GFP (1∶1000, Abcam) and biotinylated dolichos biflorus agglutinin (DBA) (1∶500, Vector Laboratories). Secondary detection was performed, using Alexa488 or Alexa568 labeled anti-mouse, anti-rabbit or anti-rat antibodies (1∶200, Invitrogen). DBA-lectin staining was visualized with FITC or Cy5 conjugated streptavidin (1∶200, Zymed Laboratories). FgfR2-ligand detection was performed on 12 µm thick cryosections cut from freshly frozen unfixed embryos. After 3 washes in cold PBS, the slides were incubated for 1 hr at 4°C in blocking solution (10% normal goat serum, 2% BSA in PBS supplemented with 0.5 mM MgCl2 and 1 mM CaCl2). The slides were treated with 2.5 µg/ml mouse FgfR2-Fc (R&D systems) in 0.5× blocking solution for 1 hr at 4°C. Following several washes in PBS, the slides were fixed for 10′ in 4%PFA. Subsequently, the samples were handled as for standard immunohistochemistry. Secondary detection was performed using a successive combination of biotinylated anti-human IgG (1∶500, Vector Laboratories) and streptavidin-FITC conjugate (1∶200, Zymed Laboratories). All slides were counterstained with 50 µg/ml DAPI and mounted with Slow Fade Gold mounting medium (Invitrogen).
Mouse urogenital ridges were micro-dissected in cold PBS supplemented with 1% FBS, 1 mM CaCl2 and 0.5 mM MgCl2. The ridges were collected into 10% FBS/DMEM medium (Wisent) containing Penicillin/Streptomycin (Gibco) and L-Glutamin (Gibco), preincubated at 37°C in 5% CO2 and kept under these conditions until all ridges were dissected. Subsequently, the urogenital ridges were cultured in the presence of the appropriate compound on 0.4 µm Transwell filters (Corning) at 37°C in 5% CO2 for 42 hrs. The following compounds were used: rat recombinant GDNF (300 ng/ml, R&D systems) anti-GDNF blocking antibody (10 or 20 µg/ml, R&D systems), SU5416 (20 µM, Calbiochem), IgG (20 µg/ml), FgfR2-Fc fusion protein (20 mg/ml, R&D systems) or DMSO (0.01%, Fisher Scientific). β-Galactosidase activity was detected by X-gal staining of dissected or cultured urogenital ridges as described [68].
Murine inner medullary collecting duct cells (mIMCD3, kindly provided by Dr Paul Goodyer) were cultured in a 1∶1 mix of DMEM and HAM's F12 media (Wisent) supplemented with 10% fetal bovine serum. To induce the canonical Wnt-pathway, mIMCD3 cells were stimulated with 3 µM BIO ((2′Z,3′E)-6-Bromoindirubin-3′-oxime, EMD) or 0.3% DMSO for 4 hours. RNA was extracted (Rneasy Mini kit, Qiagen) and reverse transcribed (Superscript III, Invitrogen) according to the manufacturer's instructions. Quantitative PCR was performed using iQ Sybr Green Supermix (BioRad) in a RealPlex2 cycler (Eppendorf) using the following primers: Gata3 sense 5′-CTCTGGAGGAGGAACGCTAA-3′ and antisense 5′-TTTGCACTTTTTCGATTTGC-3′, S16 sense 5′-GTACAAGTTACTGGAGCCTGTTTTG-3′ and antisense 5′-GCCTTTGAGATGGACTGTCGGATGG-3′. For the production of mIMCD3 cells, stably overexpressing Gata3, mouse Gata3 cDNA was cloned into pMSCV-HA3-IRES-GFP vector (kindly provided by Dr. Jerry Pelletier). The Gata3-pMSCV vector was then co-transfected with pVPack-GP and pVPack-VSV-G vectors (Stratagene) in HEK-293T cells, for virus production. The virus containing supernatant was harvested after 48 hrs by filtration and added to mIMCD3 cells. 48 hours post-infection, the cells were sorted according to GFP expression on an FACSAria-sorter (BD-Bioscience). Gata3 expression levels were assessed by western blotting and the medium level expressing cells were used for the subsequent β-Galactosidase reporter analysis. The conserved RET regulatory element was isolated from human RCC cells (ACHN) using published primers [50] and subcloned into the pGEM-T-easy vector (Promega). The point mutation in the Gata3 binding site was introduced via site-specific mutagenesis using the overlap extension method. Subsequently, the wild-type and mutated element were cloned into the β-Galactosidase expression vector pTrap [68] using the unique SalI, SphI restriction sites. The pGAL4-SV40-Luc vector (kindly provided by Dr. Xiang-Jiao Yang), constitutively expressing firefly luciferase, was used as a transfection control. The Gata3 expressing mIMCD3 were transiently co-transfected with either the wild-type or mutated pTrap vector in combination with pGAL4-SV40-Luc using Lipofectamine 2000 (Invitrogen) according to manufacturer's instructions. The analysis was carried out 24 hours post-transfection with the DualLight Combined Reporter Gene Assay System (Applied Biosystems) according to manufacturer's instructions. All experiments were done in triplicates.
The sequence alignments were performed, using blast2seq (NCBI). The transcription factor binding sites were identified with Mac Vector.
|
10.1371/journal.pbio.1001513 | An Abscisic Acid-Independent Oxylipin Pathway Controls Stomatal Closure and Immune Defense in Arabidopsis | Plant stomata function in innate immunity against bacterial invasion and abscisic acid (ABA) has been suggested to regulate this process. Using genetic, biochemical, and pharmacological approaches, we demonstrate that (i) the Arabidopsis thaliana nine-specific-lipoxygenase encoding gene, LOX1, which is expressed in guard cells, is required to trigger stomatal closure in response to both bacteria and the pathogen-associated molecular pattern flagellin peptide flg22; (ii) LOX1 participates in stomatal defense; (iii) polyunsaturated fatty acids, the LOX substrates, trigger stomatal closure; (iv) the LOX products, fatty acid hydroperoxides, or reactive electrophile oxylipins induce stomatal closure; and (v) the flg22-mediated stomatal closure is conveyed by both LOX1 and the mitogen-activated protein kinases MPK3 and MPK6 and involves salicylic acid whereas the ABA-induced process depends on the protein kinases OST1, MPK9, or MPK12. Finally, we show that the oxylipin and the ABA pathways converge at the level of the anion channel SLAC1 to regulate stomatal closure. Collectively, our results demonstrate that early biotic signaling in guard cells is an ABA-independent process revealing a novel function of LOX1-dependent stomatal pathway in plant immunity.
| Stomata are microscopic pores that are present in the epidermis of the aerial parts of higher plants, such as the leaves. These pores, which are flanked by a pair of cells called guard cells, regulate transpiration and the exchange of gas between leaves and the atmosphere. It is well documented that the phytohormone abscisic acid (ABA) is a key regulator that controls the osmotic pressure in guard cells, allowing pore size to be adjusted in response to environmental conditions. Recently, stomata have also been shown to play an important role in the innate immune response. Indeed, upon contact with microbes, plants actively close stomata to prevent the entry of microbes and the consequent colonization of host tissue. This response is known as the stomatal defense response. However, the molecular mechanisms that regulate this defense response are not well understood. Using a variety of approaches, we show in this study that LOX1, a gene that encodes lipoxygenase (LOX) in guard cells, plays a major role in stomatal defense in the model plant Arabidopsis thaliana. Mutations in LOX1 impair stomatal closure and make plants more susceptible to the bacterium Pseudomonas syringae pv. tomato. We also show that several LOX-derived metabolites, the oxylipins, are potent inducers of stomatal closure. Finally, we provide evidence to show that ABA plays only a minor role in stomatal defense response, specifically by modulating this response.
| Due to their sessile nature and the lack of an adaptive immune system, plants have evolved processes to synthesize a vast array of secondary metabolites devoted to their protection against pathogens. Some of these compounds, known as oxylipins, originate from the incorporation of one or several oxygen atoms in the carbon chain of polyunsaturated fatty acids (PUFAs), mainly linoleic acid (18∶2), linolenic acid (18∶3), and roughanic acid (16∶3). This first step leading to the formation of oxygenated fatty acids can be initiated either by reactive oxygen species (ROS) or enzymes such as lipoxygenases (LOXs), α-dioxygenases, or cytochromes P450. The primary hydroperoxy or hydroxy fatty acids can subsequently be converted into a plethora of metabolites, including aldehydes and oxo-acids, epoxydes, epoxy alcohols, hydroxides, divinyl ethers, and cyclized compounds such as jasmonates or phytoprostanes [1]–[4], which are involved in different aspects of plant physiology such as development, fertility, senescence, programmed cell death, and defense [3],[5]–[7],[8]–[10].
The 9- and 13-specific LOXs that introduce two atoms of oxygen on carbon 9 or 13 of octadecanoic fatty acids, respectively, contribute to the synthesis of a great number of oxylipins. LOX genes are up-regulated and LOX-derived oxylipins are synthesized upon stress and pathogen challenges, suggesting that LOXs play a role in stress adaptation and pathogen defense [11],[12]. The 13-LOX product JA has received much attention as it has been demonstrated to be involved in almost all above-mentioned physiological processes [10],[13]–[15]. The JA signaling pathway is now well documented involving the conjugate jasmonate-Isoleucine (JA-Ile) as the active form of jasmonic acid (JA), the F-box protein Coronatine Insensitive 1 (COI1) as receptor, and Jasmonate Zim Domain (JAZ) transcription factors as repressors whose degradation through the 26S proteasome pathway leads to the transcription of JA-responsive genes [16]. The 9-LOX-derived products have also been identified as signaling molecules regulating plant defense, cell death, as well as lateral root development and singlet oxygen-mediated stress [7],[17]–[21]. In contrast to JA, the mechanisms whereby 9-LOX products exert their biological functions are still unknown. Due to their enzymatic activity LOX can produce reactive electrophile species oxylipin (RES oxylipin) from fatty acid hydroperoxides. Recently, the mechanisms underlying RES oxylipin signaling has been uncovered. RES oxylipins possess reactive α,β-unsaturated carbonyl or epoxide groups and can add to thiol- or amine-containing compounds, such as glutathione (GSH), proteins, or nucleic acids [22]–[27]. The chemical reactivity of RES oxylipins toward the antioxidant GSH can modify the cellular redox state. RES oxylipins can also target transcription factors and the TGA transcription factors (TGA2, TGA5, and TGA6) were demonstrated to regulate the RES oxylipin-mediated gene transcription [26].
The development of transcriptomic studies and the possibility to access analyses of cell-type-specific gene expressions have enabled us to refine our understanding on the role of genes in plant-specific organs or cells. In this respect, the analyses of cell-type-specific leaf transcriptome [28] allowed much progress in the knowledge of guard cell signaling pathways that contribute to stomatal movements in response to various environmental stresses [29]–[31].
Recently, a role of guard cells as an active plant defense mechanism against pathogens has been highlighted by the finding that stomata also function in innate immunity against bacterial invasion [32]–[34]. Soon after pathogen recognition, the two guard cells that flank the stomata actively restrict the opening size preventing entry of microbes and host tissue colonization. This response is referred to as the stomatal defense [35], and it has been suggested that the abscisic acid (ABA) was a key regulator of the pathogen-mediated stomatal closure [32].
The analysis of the cell-type-specific leaf transcriptome of Arabidopsis [28] allowed us to identify the guard cell specifically expressed LOX encoding genes LOX1 and LOX6. By multiple approaches we demonstrate that, in contrast to LOX6, LOX1 plays a major role in the control of stomatal defense and plant innate immunity. We provide evidence that the MAPKs MPK3 and MPK6 and salicylate participate in a LOX1-specific stomatal pathway to respond to pathogens and PAMPs. Moreover, we show that the ABA pathway including the protein kinases OST1, MPK9, and MPK12 only contribute to a small extent to the stomatal response to pathogens. Overall, our data reveal the functioning of an oxylipin- and an ABA-dependent pathway that converge at the level of the anion channel SLAC1 to regulate stomatal closure.
Cell-type-specific transcriptome analysis of Arabidopsis leaves [28] revealed striking differences in the expression of LOX genes (Figure S1A). Genes encoding 13-LOXs, LOX2, LOX3, and LOX4 [36] were expressed in mesophyll cells, while LOX6 expression was restricted to guard cells. Conversely, expression of the 9-LOX encoding gene LOX1 was detected in guard cells but not in mesophyll cells. Microarray analyses were confirmed by RT-PCR for LOX1, expressed in guard cells and LOX2, LOX3, and LOX4 whose expression is restricted to mesophyll cells (Figure S1B).
In order to determine whether LOX1 contributes to stomatal defense, two lox1 knockout lines (lox1-1 and lox1-2) along with a complemented lox1-2 (lox1-2 35S:LOX1) line and Col-0 as the wild-type (WT) control line were assessed for their resistance to virulent Pst DC3000 upon spray inoculations (Figure 1A). The two lox1 mutant lines displayed more than 10-fold enhanced growth of Pst DC3000 as compared to the complemented and WT lines, suggesting that LOX1 participates in the control of bacterial leaf colonization.
To test whether LOX1 exerts its effect on plant immunity at the guard cell level, we compared the behavior of stomata of lox1 mutants with WT plants in response to Pst. In contrast to WT and lox1-2 35S:LOX1, lox1 mutants were significantly compromised in their ability to close stomata in response to virulent Pst DC3000 and avirulent Pst DC3000 AvrRpm1 strains (Figure 1B). Similarly, the flg22-induced stomatal closure was strongly impaired in both lox1 mutants (Figure 1B). On the contrary, the lox6-1 mutant line (Methods S1), was not compromised in its ability to close stomata in response to flg22 (Figure S2).
Considering that ABA has been identified as a positive regulator of pathogen and PAMP-induced stomatal closure [33], we investigated also whether lox1 mutants showed altered sensitivity to ABA. We observed that both lox1 mutant lines were as sensitive to exogenously applied ABA as WT and lox1-2 35S:LOX1 (Figure 1C). Together, these results suggest that LOX1 activity mediates pathogen- and flg22-induced stomatal closure independently of ABA.
Since PUFAs are known LOX substrates, we first tested linoleic acid (18∶2) for its ability to induce stomatal closure. We observed that linoleic acid induced significant stomatal closure at 10 nM (Figure 2A). At 100 nM both linoleic (18∶2) and linolenic acid (18∶3) induced stomatal closure in WT and lox1-2 35S:LOX1 plants but not in both lox1 mutants (Figure 2B). In contrast, oleic acid (18∶1), which is not a substrate of LOXs, was unable to trigger stomatal closure in any plant line (Figure 2B). These results strongly suggest that products of LOX activity are able to promote stomatal closure.
LOXs directly generate FAHs by incorporating molecular oxygen into the backbone of PUFAs (Figure S3). We therefore tested the ability of 9- and 13-HPODE to trigger stomatal closure and compared their activity to H2O2, which was previously shown to mediate ABA-induced stomatal closure [37] and to tBuOOH as a model organic hydroperoxide. We found that 9- and 13-HPODE promoted stomatal closure at considerably lower concentrations than H2O2 or tBuOOH (Figure 2C), demonstrating that LOX-derived FAHs are potent inducers of stomatal closure.
Oxidative processes arising from increased cellular hydroperoxide concentrations can be effectively prevented by the addition of free radical scavengers including thiol-containing compounds. Among them, N-acetyl cysteine (NAC) is one of the less potent free radical scavengers [38]. Alternatively, thiols can act as nucleophiles to conjugate with reactive electrophiles through the Michael addition [25],[27],[39]. In order to get further insights into the mechanisms at the origin of the biological activities of hydroperoxides, epidermal peels were pre-incubated with 1 mM NAC before treatment with 10 µM 9- or 13-HPODE, 40 µM H2O2, or tBuOOH (Figure 2D). The inability of NAC to prevent H2O2- or tBuOOH-induced stomatal closure suggested that NAC did not function as an antioxidant. On the other hand, its specific inhibitory effect on FAHs rather suggested that it could act as a nucleophilic compound.
RES oxylipins containing α,β-unsaturated carbonyl groups can result from the metabolization of FAHs by LOX activities (Figure S3) [40],[41]. RES oxylipins are able to add to thiol-containing compounds such as cysteines of proteins or glutathione according to the Michael addition (for review, see [27]) and are produced during plant microbe interactions [22],[26],[42]. We investigated whether the two RES oxylipins, 9-KODE and 13-KODE, show a dose-dependent effect on stomata, finding that both compounds induce stomatal closure already at nanomolar concentrations (Figure 3A).
To test whether the presence of thiol reactive structures was required for the biological activity of 9- and 13-KODE, the ketone group was reduced into nonreactive alcohol. The alcohols, 9- and 13-HODE, were both unable to promote stomatal closure (Figure 3B). Moreover, the effect of different Michael acceptors along with halogenated organic compounds and isothiocyanate was tested. All these molecules are known as thiol reagents and were potent inducers of stomatal closure (Figures 3C, D and S4). Additionally, the thiol-containing compound NAC that inhibited the action of FAHs also blocked the activity of 13-KODE and the other thiol reagents PCMB and sulforaphane, suggesting that reactivity toward thiols was critical for biological activity of these compounds. Conversely, NAC did not prevent ABA-induced stomatal closure (Figure 3D), suggesting that RES oxylipins and ABA do not share a common signaling pathway.
The biologically inactive alcohols, 9- and 13-HODE, displaying chemical structures close to the corresponding ketones prompted us to verify whether these alcohols could counteract the biological activity of 9- and13-KODE and different thiol reagents. As shown in Figure 4A, pre-incubation with 1 nM 13-HODE abolished the activity of the three thiol reagents (13-KODE, PCMB, and sulforaphane). Similar results were obtained with 9-HODE (unpublished data). However, the inhibitory effect of 13-HODE was overcome by using higher concentrations (10 or 100 nM) of 13-KODE (Figure 4B). On the other hand, stoichiometric concentrations of 13-HODE did not alter the ABA-induced stomatal closure (Figure 4A). Finally, the effects of NAC and 13-HODE were investigated on epidermal peels either inoculated with Pst DC3000 or treated with flg22 (Figure 4C). In plants containing a functional LOX1, the two compounds abolished stomatal closure induced by both Pst and flg22. Together, these data further support the existence of a biotic stress-mediated stomatal response independent of ABA.
COR secreted by virulent strains of Pseudomonas disables stomatal defense in a COI1-dependent manner and notably inhibits PAMP- and ABA-mediated stomatal closure, subsequently triggering stomatal reopening required for pathogen entry and host colonization [32]. To assess whether COR affects the activity of RES-oxylipin, we pre-treated epidermal peels with the toxin prior to treatment with RES oxylipins. As shown in Figure 5A, COR at 4 µM blocked the effects of both 1 nM 13- and 9-KODE. The COR chemical structure resembles that of JA-Ile, the biologically active form of JA, and enables this phytotoxin to interact with the JA-Ile receptor COI1 to hijack the JA-Ile hormone functions. In an attempt to better understand the biological activity of these chemically related compounds, we examined the action of methyl jasmonate (MJ) on guard cells. In our hands, MJ was unable to promote stomatal closure in a range of concentrations between 1 nM and 100 µM (Figure S4). However, competition experiments of MJ with RES oxylipins (9/13-KODE and 12-OPDA) indicated that pretreatment of epidermal peels with 1 nM MJ reversed the activity of RES oxylipins applied at the same concentration (Figure 5B). At a 100-fold higher concentration of RES oxylipins, this inhibition was no longer observed, suggesting that MJ mediates the inhibition of RES oxylipins activity through a similar mechanism as the alcohols 9/13-HODE (Figure 4). However, in contrast to COR, MJ did not counteract the action of ABA (Figure 5C). In addition, our data establish that although COR and MJ are both able to inhibit the RES oxylipin-mediated stomatal closure, in contrast to COR, the MJ-mediated activity is COI1-independent (Figure 5D).
Bacterial- and PAMP-mediated stomatal closures have been demonstrated to be SA-dependent [32] and SA induces stomatal closure in Arabidopsis [35]. In an attempt to clarify the role of SA in the RES oxylipin- and ABA-signaling cascades, we conducted experiments with the transgenic SA-deficient NahG line, along with the two SA biosynthesis mutant lines, sid1-1 and sid2-1 [43],[44]. Results shown in Figure 5E demonstrate that these three lines normally responded to ABA, whereas they were all impaired in their ability to respond to the RES oxylipins 9/13-KODE. In addition, our data also indicate that lox1 mutant lines were as sensitive to SA (100 µM) as WT and lox1-2 35S:LOX1 (Figure 5F), demonstrating that exogenously applied SA can complement for LOX1 deficiency. Together these results support that SA is required to convey the RES oxylipin- but not the ABA-mediated signal leading to stomatal closure.
Many steps involved in the ABA-dependent signaling pathway are now well documented [45],[46]. More recently, ABA has also been described as a key player in pathogen- and PAMP-induced stomatal responses [33]. It has been demonstrated that the Arabidopsis mutants aba3-1 and ost1-2, which are defective in the hormone biosynthesis and signaling, respectively, were both no longer able to close stomata in response to either Pst or flg22. In order to clarify the function of ABA signaling in pathogen-induced stomatal closure, we first compared the effects of ABA, flg22, and 13-KODE on stomata of the mutant ost1-2. As previously reported, the ost1-2 mutant does not close stomata in response to ABA even at very high concentrations (Figure 6A). However, ost1-2 plants remained responsive to both flg22 and 13-KODE, albeit at higher doses than WT plants. The stomatal behavior of the aba2-1 mutant, which is defective in ABA production, was also assessed and showed a similar response profile as the ost1-2 mutant (Figure S5).
In order to get further insights into the signaling steps triggered by the three compounds, we have assessed the activation of OST1 on Arabidopsis cell suspensions. In-gel kinase assays (Figure 6B) of immunoprecipitated OST1 showed that in accordance with previous work [47]–[49], ABA clearly activated OST1 kinase. In contrast, flg22 and both RES oxylipins 9- and 13-KODE were unable to activate OST1. This is consistent with the hypothesis that different signaling pathways can lead to stomatal closure, but that the optimization of the guard cell response to biotic stress requires ABA-dependent mechanisms.
Apart from OST1, a number of other protein kinases have been identified to function in ABA signal transduction, including mitogen-activated protein kinases (MAPKs), which play a prominent role in both biotic and abiotic stress signaling [50],[51]. MPK9 and MPK12 were shown to play an important role in ABA-induced closure of stomata [29]. To clarify the role of these MAPKs in ABA and flg22 signaling, we tested mpk9-1/12-1 double mutant plants, which were reported to be impaired in ROS-mediated ABA signaling [29]. Double mutant plants displayed a clear defect in ABA-induced stomatal closure, but were as responsive to flg22 as wild-type plants (Figure 6C).
To further investigate the role of protein kinases in response to ABA and biotic stress, we carried out kinase assays on cells treated with ABA, flg22, and 9- or 13-KODE. In-gel kinase assays showed flg22-mediated activation of three protein kinases (Figure 6D) that were previously identified as MPK3, MPK4, and MPK6 [52]. Immunoblotting using the anti-phospho-p44/42 MAPK antibody confirmed the presence of these activated MAPKs in seedlings exposed to flg22 (Figure 6E). However, none of these MAPKs were activated in response to ABA (Figures 6E), the kinase activities stimulated in ABA-treated cells corresponding to OST1 and other SnRK2s [48]. In order to verify the role of these kinases in the ABA- and PAMP-mediated signaling cascades, the behavior of mpk3 and mpk6 knockout mutants was investigated. As shown in Figure 6F, stomata of mpk3 and mpk6 responded normally to ABA, 18∶2 and 9-KODE, whereas their response to flg22 was strongly impaired demonstrating that different sets of MAPKs are required to convey signals generated by the action of ABA and biotic stress and suggesting that RES oxylipin production and action occur downstream of MPK3/6.
To further assess whether early pathogen-mediated signaling cascade depended on ABA, plantlets were submitted either to ABA or to Pst DC3000 or to Pst followed by ABA sprays and sampled over a time scale corresponding to the pathogen-induced stomatal closure. As shown in Figure S6, transcript levels of the three ABA-specific genes, RD29b, ABI1, and ABI2, (Methods S1), significantly increased as soon as 30 min after application of the hormone, whereas transcript levels of these genes did not increase in plants treated with Pst DC3000. In addition, pretreatment with Pst did not modify the response of plants to ABA. Together, these results suggest that during the early phase of the plant–pathogen interaction when stomatal closure takes place, ABA is not produced.
In the present work, we show that a defect in LOX1 compromises the ability of plants to close stomata in response to both virulent and avirulent strains of Pst or the PAMP flg22 (Figure 1). Impairment in ABA responsiveness, which could have explained the lox1 mutant phenotype, was ruled out since mutant and WT lines were similarly able to close stomata upon ABA treatment (Figure 1). On the other hand, involvement of oxylipins arising from LOX1 activity in controlling stomatal access of bacteria into plants was clearly established (Figure 1). Induction of stomatal closure by LOX substrates (PUFAs) or LOX reaction products (FAHs and KOD(T)E) was demonstrated (Figures 2, 3, and S4). On the other hand, Arabidopsis mutants lacking LOX1 were no longer able to respond to PUFAs (Figure 2), while they still remained responsive to FAHs (unpublished data) and 9/13-KODE (Figure S7A), indicating that absence of LOX1 activity can be rescued not only by FAHs but also by their metabolites. In line with this result, it is noteworthy that pretreatment of the lox1-2 mutant line with the 9-specific LOX compound 9-KODE (Methods S1), partially restores resistance of LOX1-deficient lines against Pst DC3000 (Figure S7B), confirming that LOX1 controls stomatal defense.
Dose-response curves of linoleic acid and the two FAHs (9- and 13-HPODE) for stomatal closure showed that these compounds close stomata at concentrations several orders of magnitude lower than ABA (Figures 2 and S4). Our data also demonstrate that the efficiency of oxylipins to close stomata is not position-specific, since 9- and 13-HPODE had the same biological activity and that trienic fatty acids were as active as dienic ones (Figure 2). To better understand the mechanisms responsible for the biological activity of FAHs, their efficiency was compared to that of H2O2. FAHs were remarkably more active than H2O2 and tBuOOH (Figure 2). To explain this discrepancy, we postulate that oxidative stress potentially generated by the peroxides might target different signaling components. Unlike H2O2 or tBuOOH, the more hydrophobic FAHs might access and oxidize critical residues in hydrophobic sites of key signaling proteins.
Thiol-containing compounds can protect cells against oxidative injury by scavenging reactive oxygen species (ROS) or by conjugating to reactive electrophile species (RES) often generated from FAHs [27]. NAC, used in the present work for its rather weak antioxidant property [38], did not inhibit the H2O2- or tBuOOH-induced responses, but it strongly compromised FAH responses (Figure 2). Hence, it is likely that inhibition of FAH activity by NAC was rather due to its nucleophilic properties. Downstream of FAHs, those oxylipins containing α,β-unsaturated carbonyl structures are powerful RES able to add to thiol-containing substances [25]–[27],[39]. Consequently, the activity of FAHs on stomata can rather be explained by their metabolization into RES-oxylipins (Figure S3). In good agreement with this hypothesis, the two RES oxylipins 9- and 13-KODE displayed equivalent and very strong activity on guard cells (Figure 3). Five additional RES oxylipins including trienic fatty acid ketones (9- and 13-KOTE), two cyclopentenone derivatives (12-OPDA and ProstaglandinA2), and the aldehyde 4-HNE that can be generated by decomposition of FAHs all displayed strong biological activity (Figure S4).
Among these oxylipins, the JA precursor 12-OPDA displayed a strong activity on stomata, whereas MJ was inactive in a range of concentrations between 1 nM and 100 µM (Figure S4). The ability of MJ to induce stomatal closure is still controversial ([53]–[55] and this work), but very recent results show that MJ is active only when ABA endogenous levels reach a critical threshold [56]. In the ABA-deficient line aba2-2 or in plants treated by fluridon, an inhibitor of ABA biosynthesis, MJ was no longer able to promote stomatal closure, indicating that MJ activity on guard cells is intimately dependent on the physiological status of plants. Hence, culture conditions used to produce the biological material can strongly influence this status and modify the ability of plant to respond to MJ.
The α,β-unsaturated carbonyl structure in oxylipins, as soft electrophiles, preferentially add to soft nucleophiles, which are cysteine residues in GSH or proteins [27]. Reducing the RES oxylipins, 9-, and 13-KODE into their corresponding alcohols 9- and 13-HODE, the latter being not electrophilic, has allowed us to assess the involvement of this putative mechanism. Given that both alcohols were inactive (Figure 3), reactivity of 9- and 13-KODE toward thiols could explain their biological activity. This hypothesis was confirmed by the fact that a large set of known thiol-reagents, including different natural and artificial Michael acceptors, several halogenated organic compounds, and one isothiocyanate, all promoted stomatal closure at nanomolar concentrations (Figures 3C, D and S4). Altogether, these results strongly suggest the involvement of a thiol-containing target in the RES oxylipin-mediated cascade.
At stoichiometric concentrations, both the fatty acid alcohols 9/13-HODE and MJ were found to block the biological activity of RES oxylipins but not that of ABA (Figures 4 and 5). On the other hand, at concentrations above 1 µM, the Pseudomonas phytotoxin COR disabled both RES oxylipin and ABA activities ([32] and Figure 5). These results suggest that the COR-mediated inhibition of RES oxylipin activity is COI1-dependent, whereas fatty acid alcohols and MJ inhibit this activity by means of another mechanism. This hypothesis is further supported by results shown in Figure 5D, which indicate that, on the mutant coi1-17, COR no longer exerts inhibition of the ketone effects, whereas MJ remains active. Neither NAC, 13-HODE, nor MJ inhibited the activity of ABA (Figures 4 and 5), suggesting that RES oxylipins do not share a common signaling mechanism with ABA. Considering that stomatal closure induced by inoculation with Pst or treatment with flg22 were inhibited by NAC or 13-HODE (Figure 4), it can be concluded that the early stomatal response to pathogens is not under the control of ABA but depends on LOX1-derived RES oxylipins. According to all these data, it is tempting to speculate that thiol-reactive compounds such as RES-oxylipins generated in response to biotic stress would react on guard-cell-specific target(s) bearing a critical cysteine residue buried in a hydrophobic cavity. This covalent binding (alkylation) would then trigger stomatal closure. Competition with thiol-containing compounds such as NAC or the presence of FAH or RES oxylipin metabolites such as 9/13-HODE and jasmonates that could fit into the target cavity without reacting with the critical cysteine residues would prevent the target alkylation and thereby inhibit the biological process. The equal activity of 9- and 13-oxylipins on stomata suggests that these compounds are perceived by the same guard cell target(s). The 9- and 13-LOXs could both contribute to production of metabolites regulating stomatal movements, however in mutants lacking LOX1 activity, (i) PUFAs were not able to rescue stomatal function and (ii) bacteria were unable to trigger stomatal closure. Hence, LOX1 exhibits a specific function for which no other LOXs can be substitutes. Consistent with this assumption, it is clearly established that mutation in LOX6, the 13-LOX encoding gene expressed in guard cells, did not impair stomatal responses to flg22, linoleic acid, or 9-KODE (Figure S2).
Our examination of the SA function in the stomatal response indicated that the synthesis of SA is required to convey the RES oxylipin (9/13-KODE) signal but not that of ABA and SA acts downstream of these RES oxylipins (Figure 5). Hence it can be concluded that SA participates in the RES oxylipin-mediated signaling cascade in accordance with previous results indicating that PAMP- or Pst-regulated stomatal closure was SA-dependent [32] but SA is not an intermediate of the ABA-dependent pathway (Figure 5). Again, these results further support the conclusion that ABA does not act as a regulator of the biotic stress-dependent signal in guard cells.
A major breakthrough in the understanding of the plant innate immune response against bacterial pathogens was the finding that stomatal closure was a key part of this process preventing bacterial entry in leaves [33]. This work also suggested that PAMP-induced stomatal closure required ABA biosynthesis and components of the ABA signal transduction pathway such as the protein kinase OST1 and nitric oxide biosynthesis.
Our data argue in favor of the existence of an ABA-independent guard cell signaling cascade controlling stomatal closure during the early steps of plant-microbe interaction. First, the ability of lox1 mutant lines to respond to ABA strongly suggested that LOX1 did not act downstream of the hormone (Figure 1). Second, in contrast to ABA, neither flg22 nor the RES oxylipins 9- and 13-KODE could activate OST1 kinase in Arabidopsis cells (Figure 6), suggesting that different early signaling pathways are employed for ABA and flg22 or RES oxylipins. Consistent with these results, the mutant ost1-2, which is insensitive to ABA, still responded to flg22 or 13-KODE (Figure 6), albeit responses were less intense than in WT. Similarly, the stomatal responses of the aba2-1 mutant, which is impaired in ABA synthesis, showed a marked decrease in sensitivity not only to ABA but also to flg22 and 13-KODE (Figure S5). Overall, our results strongly suggest that although ABA is not the major regulator of bacterial and PAMP-induced stomatal closure, it contributes to modulate stomatal responses to biotic stress. In a recent genetic screen, Zeng et al. [35] identified Arabidopsis mutants named scord (susceptible to coronatine-deficient Pst DC3000) that were compromised in their ability to close stomata in response to bacteria. Interestingly, a majority of these mutants showed normal closure in response to ABA, confirming our results that ABA does not directly control bacterial-induced stomatal closure. Stomatal closure requires the presence in guard cell membranes of effector proteins that control osmotic pressure. Their transcription and/or posttranslational regulation are known to mainly depend on ABA in an OST1-mediated pathway [57]. To explain that mutants affected in ABA synthesis or regulation remained responsive to biotic stress, it must be assumed that still unidentified components also regulate this ultimate step. However, in these mutants, the stomatal closure machinery is likely less efficient, explaining the lower sensitivity of such mutants to biotic stress. Interestingly, the activation of one of these effectors, the slow-type anion channel SLAC1 [58], is required in both ABA- and biotic stress-mediated responses (Figure S9), suggesting that these pathways converge at the level of the ion channels controlling the osmotic pressure of the guard cell necessary to regulate stomatal aperture.
Numerous MAPK cascades have been reported to convey diverse environmental signals [50],[51]. Both MAP kinases MPK9 and MPK12 were identified as critical steps of the guard cell response to ABA, and the mpk9-1/12-1 double mutant is compromised in ABA-induced stomatal closure [29]. However, this mutant responded normally to flg22 (Figure 6), further supporting that distinct ABA- and flg22-mediated pathways exist in guard cells. The two well-characterized MAPKs, MPK3 and MPK6, that are key response elements in both abiotic and biotic stress signaling have also been implicated in ABA signal transduction. MPK3 antisense plants were shown to be partially impaired in ABA-promoted inhibition of stomatal opening, while they responded normally to the hormone in stomatal closure [59]. However, these plants were compromised in their ability to close stomata in response to Xanthomonas campestris inoculation [60]. MPK6 was also reported to mediate ABA- and sugar-regulated seed germination [61] but also responses to other stresses such as pathogen infection, cold and salt stress, and wounding [62]–[65]. Our work demonstrated that flg22 but not ABA activated MPK3, MPK4, and MPK6 early after treatment (Figure 6). A role of MPK3 and MPK6 in flg22-induced stomatal closure was subsequently confirmed by testing mpk3 and mpk6 knockout lines. The fact that both mutants were compromised in the flg22-mediated stomatal closure (Figure 6) suggests that the two MAP kinases do not play a redundant role in this process and agrees with recent findings that these two kinases both play an important but distinct role in plant immunity and signaling [66],[67]. Overall, our results demonstrate that MPK3 and MPK6 are involved in the flagellin, but not in the ABA guard cell signaling cascade. The fact that linoleic acid (18∶2) and 9-KODE were as active on WT as on mpk3 or mpk6 mutants suggests that LOX1-mediated synthesis of oxylipins is downstream of MPK3/6, although it cannot be ruled out that these two steps are independent. The involvement of these MAPK activities as early specific steps of the flg22-dependent signaling cascade provides another argument supporting the existence of an ABA-independent mechanism leading to stomatal closure.
Finally, gene expression studies of plantlets that were treated with ABA or Pst DC3000 showed that RD29b, ABI1, and ABI2 transcript levels rapidly increased only in response to ABA but not after Pst inoculation (Figure S6). Additionally, given that plants first treated with Pst and then sprayed with ABA displayed identical responses as those observed in plants treated by ABA only, it can be concluded that ABA is not produced during the very early stage of the interaction with bacteria.
Overall, our data suggest that early signals produced upon challenging of plants with ABA and pathogens are different. As depicted in the proposed model (Figure 7), whereas the response to ABA is transduced principally by OST1 and MPK9/12, microbial signals are conveyed by MPK3/6, LOX1-induced RES oxylipin derivatives, and SA. Our work also provides evidence that ABA is essential to potentiate guard cells to enable them to properly respond to biotic stimuli, and further studies are necessary to characterize these steps and their role in the stomatal responses to other signals. Finally, the present results shed new light on the function of oxylipins in the plant immune response.
All experiments were performed in triplicates, and statistical differences of means were analyzed with the Student's t test with the following symbols for the classes of probability: * p<0.05, ** p<0.01, and *** p<0.001. To compare more than two means, the ANOVA test was performed and letters have been used to mark statistically identical groups of means. The number of degrees of freedom (dfs), the value of the F statistic and the p value are given in the legend of figures for each group of means.
The following Arabidopsis thaliana transgenic and mutant lines used in this study were derived from ecotype Columbia (Col-0), lox1-2 35S:LOX1 (see below), lox1-1, lox1-2 [7], lox6-1 (see below), aba2-1, mpk3, mpk6 [67], mpk9-1/12-1 [29], nahG [68], slac1-3 [58], sid1-1 [44], sid2-1 [43], and coi1-17 [69], whereas the ost1-2 line derived from Landsberg erecta (Ler). Specific growth conditions are specified below for each type of experiment. coi1-17 mutant plants were selected from an heterozygote population using root sensitivity to 10 µM MJ according to [69].
The Escherichia coli strain TOP10 (Invitrogen) was used for cloning and propagation of the different vectors. Agrobacterium tumefaciens C58 strain transformation was performed by electroporation and was used in all plant transformation experiments. The full-length cDNA clone of LOX1 gene (accession number BT010358) provided by the Nottingham Arabidopsis Stock Center (http://arabidopsis.info) was used for this study. PCR amplification was performed with the cDNA containing vector as template and the primer couple: forward, 5′-GGGGACAAGTTTGTACAAAAAGCAGGCTTCGAAAACCTGTATTTTCAGGGAATGTTCGGAGAACTTAGGGATCTG-3′ and reverse, 5′-GGGGACCACTTTGTACAAGAAAGCTGGGTTTCAGATAGAGACGCTATTTGGAAT-3′. The amplicon was subcloned into the GATEWAY donor vector pDONRZeo (Invitrogen) using the BP clonase II (Invitrogen). The full-length cDNA was subcloned in the destination vector pMDC32 kindly provided by Mark Curtis (Institute of Plant Biology and Zürich-Basel Plant Science Centre, University of Zürich, Switzerland) using the LR clonase (Invitrogen). The homozygous T-DNA line lox1-2 was transformed with empty vector pMDC32 or pMDC32:LOX1 using flower dipping method. T1 plants were screened on germination medium supplemented with Hygromycine (30 mg/L). The T-DNA presence was monitored by PCR analysis using the primer couple: forward, 5′-GCGCGATTGCTGATCCCC-3′ and reverse, 5′- GCCCTCGGACGAGTGCTG-3′. The expression of LOX1 in the transgenic lines was measured by semi-quantitative RT-PCR using the primer couples: (LOX1) forward, 5′-AGACTATCCTTACGCAGTGGA-3′ and reverse 5′-TGCCGGTGACTCCGCCTTC-3′, and (EF1) forward, 5′-TACCTCCCAGGCTGATTGTG-3′ and reverse, 5′-TCTGACCAGGGTGGTTCATG-3′. Seven lox1-2:pMDC32:LOX1 independent lines were retrieved after the screening, T2. Progeny analysis was performed and lines with a single T-DNA locus were selected for further experiments.
Oxylipins and other chemicals used in this study are given in Table S2. Flg22 was kindly provided by Dr. Laurent Noël (Laboratoire des Interactions Plantes-Microorganismes CNRS/INRA Toulouse, France). The alcohols, 9-HODE, (10E,12Z)-9-hydroxy-10,12-octadecadienoic acid and 13-HODE, (9Z,11E)-13-hydroxy-9,11-octadecadienoic acid were produced by sodium borohydride reduction of the corresponding ketones according to the following procedure: 10 µL of 10 mM 9- or 13-KODE ethanolic solution was added to 20 µL of 5% (w/v) NaBH4 in 0.2 N NaOH and incubated 5 min at room temperature. After addition of 100 µL of 0.5 M potassium acetate pH 4.0, products were extracted 3 times with 200 µL of the solvent mixture hexane/diethyl ether, 70/30 (v/v). The organic phase was then evaporated to dryness under a nitrogen stream, and the residue was dissolved in absolute ethanol. Purity of compounds was controlled by HPLC chromatography according to Montillet et al. [70]. The chromatographic profiles were recorded at 234 and 280 nm for alcohol and ketone detections, respectively. None of the alcohol preparations were contaminated with a detectable amount of ketone.
Plants were grown on soil in a controlled environment chamber under an 8 h light regime (150–200 µE·m−2·s−1) at 22°C and 65% relative humidity. Pst DC3000 was grown for 24 h at 28°C on NYGA solid medium supplemented with 50 µg/mL Rifampicin. For in planta bacterial growth assays, 4-wk-old plants were spray-inoculated with bacterial suspensions at 5·107 cfu/mL in 10 mM MgCl2 with 0.03% (v/v) Silwet L-77 (Lehle Seeds). In planta bacterial titers were determined at day 3 by shaking leaf discs in 10 mM MgCl2 with 0.01% (v/v) Silwet L-77 at 28°C for 1 h as previously described [71]. At least five plants per genotype were used for each sampling.
After 3 d of stratification of seeds at 4°C, wild-type, mutant, and transgenic lines of A. thaliana were grown on soil in a plant growth chamber with an 8 h light period (250 µE·m−2·s−1) at 23°C and a 16 h dark period at 19°C, and relative humidity of 75%. The abaxial side of leaves of 4- to 5-wk-old plants was stuck on cover slips and peeled. Samples were then placed in Petri dishes containing 10 mM MES/Tris pH 6.0, 30 mM KCl, and 1 mM CaCl2 (working buffer). After 30 min in darkness, epidermal peels (except dark controls) were transferred for 2 h under light (300 µE·m−2·s−1) at 22°C in order to assure that most stomata were open before treatments. For promotion of closure assays, stock solutions of compounds (in water or ethanol as indicated in the figures) were directly diluted in the working buffer in contact with epidermal peels. Stock solutions prepared in ethanol were diluted so that the final solvent concentration did not exceed 0.1% (v/v). Pst DC3000 and Pst DC3000 AvrRpm1 were grown for 18 h in a rotating incubator at 30°C and 200 rpm in liquid LB medium supplemented with 50 µg/mL of Rifampicin and 50 µg/mL of Rifampicin plus 10 µg/mL of tetracycline, respectively. For treatments, bacterial suspensions were centrifuged and resuspended in the working buffer at a final concentration of 108 cfu/mL. After treatments, samples were further incubated under light for 2.5 h (or 1 h for treatments with bacteria) before stomatal aperture measurements according to Merlot et al. [72]. To avoid bias results, blind experiments were performed so that experimenters in charge of stomatal measurements were not informed of the evaluated genotypes or treatments. Values reported are the means of at least three independent experiments, for which 60 aperture widths were measured each time. Error bars represent standard deviations of the means.
Arabidopsis thaliana Col-0 seedlings were grown in liquid MS medium (SIGMA) for 14 d in a growth chamber (24°C, 16 h photoperiod). Seedlings were treated with water (mock) or 1 µM flg22 for 10 min or with 0.1% ethanol (mock) or 100 µM ABA for 10, 20, 30, or 60 min. For protein extraction, ground samples were resuspended in SDS-PAGE loading buffer, boiled, and centrifuged at 8,000 g for 2 min. Supernatants, containing total protein extracts, were separated on 10% SDS-PAGE gels and transferred to PVDF membranes. After blocking with 5% BSA in TBST, membranes were probed with 1/5,000 anti-phospho-p44/42 MAPK (Erk1/2) (Thr202/Tyr204) rabbit monoclonal antibody (Cell Signaling) followed with 1/10,000 horseradish peroxidase (HRP)-conjugated anti-rabbit IgG antibody (SIGMA). HRP activity was detected using the ECL Plus reagent kit (GE Healthcare) and the GeneGnome imaging system (Syngene). Coomassie blue staining of blots was carried out for protein visualization.
Arabidopsis thaliana cell suspension (T87 line, Columbia ecotype) was cultured as previously described [73]. Cells were treated with 30 µM ABA, 100 nM flg22, 100 nM 13-KODE or 9-KODE, or 0.1% ethanol (mock) for 10 min. Protein extraction, immunoprecipitation with anti-SnRK2.6 (OST1) specific antibody [45] followed by in-gel kinase assay were performed as previously described [42].
The accession numbers for the genes discussed in this article are as follows: LOX1 (At1g55020), LOX2 (At3g45140), LOX3 (At1g17420), LOX4 (At1g72520), LOX5 (At3g22400), LOX6 (At1g67560), OST1 (At4g33950), MPK3 (At3g45640), MPK6 (At2g43790), MPK4 (At4g01370), (At3g18040), MPK9 (At3g18040), MPK12 (At2g46070), ABA2 (At1g52340), ABI1 (At4g26080), ABI2 (At5g57050), RD29b (At5g52300), SLAC1 (At1g12480), COI1 (At2g39940), SID1 (At1g74710), and SID2 (At4g39030).
|
10.1371/journal.ppat.1000192 | Direct Identification of the Meloidogyne incognita Secretome Reveals Proteins with Host Cell Reprogramming Potential | The root knot nematode, Meloidogyne incognita, is an obligate parasite that causes significant damage to a broad range of host plants. Infection is associated with secretion of proteins surrounded by proliferating cells. Many parasites are known to secrete effectors that interfere with plant innate immunity, enabling infection to occur; they can also release pathogen-associated molecular patterns (PAMPs, e.g., flagellin) that trigger basal immunity through the nematode stylet into the plant cell. This leads to suppression of innate immunity and reprogramming of plant cells to form a feeding structure containing multinucleate giant cells. Effectors have generally been discovered using genetics or bioinformatics, but M. incognita is non-sexual and its genome sequence has not yet been reported. To partially overcome these limitations, we have used mass spectrometry to directly identify 486 proteins secreted by M. incognita. These proteins contain at least segmental sequence identity to those found in our 3 reference databases (published nematode proteins; unpublished M. incognita ESTs; published plant proteins). Several secreted proteins are homologous to plant proteins, which they may mimic, and they contain domains that suggest known effector functions (e.g., regulating the plant cell cycle or growth). Others have regulatory domains that could reprogram cells. Using in situ hybridization we observed that most secreted proteins were produced by the subventral glands, but we found that phasmids also secreted proteins. We annotated the functions of the secreted proteins and classified them according to roles they may play in the development of root knot disease. Our results show that parasite secretomes can be partially characterized without cognate genomic DNA sequence. We observed that the M. incognita secretome overlaps the reported secretome of mammalian parasitic nematodes (e.g., Brugia malayi), suggesting a common parasitic behavior and a possible conservation of function between metazoan parasites of plants and animals.
| Parasitic nematodes are microscopic worms that cause major diseases of plants, animals, and humans. Infection is associated with secretion of proteins by the parasite; these proteins suppress the immune system and cause other changes to host cells that are required for infection. Identification of secreted proteins has been difficult because they are released only in trace amounts. We have developed very sensitive methods that enabled the discovery of 486 proteins secreted by the root knot nematode, Meloidogyne incognita; prior to this, only a handful of secreted proteins were known. Several secreted proteins appear to mimic normal plant proteins, and they may participate in the process by which the nematode hijacks the plant cell for its own purposes. Meloidogyne species infect many crops, including corn, soybean, cotton, rice, tomato, carrots, alfalfa, and tobacco. The discovery of these secreted proteins could lead to new methods for protecting these important crops from nematode damage. We observed that the secretome of the human pathogen, Brugia malayi, overlaps that of M. incognita, suggesting a common parasitic behavior between pathogens of plants and animals.
| M. incognita can infect 1,700 plant species [1]. At the infective juvenile (J2) stage of development, M. incognita enters the elongation zone of the root and burrows through the apoplast to the root tip where it enters the vascular cylinder, moving up to the zone of root differentiation. The nematode then inserts its stylet into the plant cell cytoplasm and induces nuclear division without cytokinesis, creating multinucleate giant cells that nurture the developing worm. Infection is associated with the reprogramming of plant cell development rather than host cell death [2]. M. incognita infection causes plant defense genes to become either promptly suppressed or transiently induced, in contrast to incompatible interactions, which immediately induce and sustain expression of defense genes [3].
The proteins and metabolites secreted from the esophageal glands (subventral and dorsal glands) of plant-parasitic nematodes are thought to be responsible for compatibility [4]. The two subventral gland (SvG) cells are biologically active during the J2 stage, while the dorsal gland cell is predominantly active on the second day post-infection through to the end of the nematode's life. In vivo observations of the root cyst nematode, Heterodera schachtii, revealed that the dorsal gland secretions are released through the stylet into the plant cell [5]. Other nematode tissues also secrete proteins that may be important for plant-pathogen interaction: two amphids localized in the anterior part of the worm, around the lip region, and two phasmids at the posterior part could be receptors for chemotaxis [6]. These two kind of organs contain socket cells that are highly secretory but functions of these secretions remain mostly unknown [6],[7].
Following the establishment of compatibility, pathogen-produced effector molecules are the key to infection. These molecules have been found in well-characterized pathosystems where they modulate host signaling pathways to prevent defense responses [8], but little is known about effectors that mediate plant-metazoan pathogenesis. Bird and Saurer (1967) characterized secreted molecules from the esophageal gland cells of Meloidogyne javanica [9]. They showed that the secretions were mainly proteins; no nucleic acids were detected. Antibodies have been used to monitor the expression of esophageal antigens from several plant nematode species [10]. In Globodera rostochiensis, antibodies recognized proteins present both in the subventral gland cells and on the surface of the nematode [11]. Other studies of Meloidogyne spp. showed that silencing of genes expressed in the SvG reduced pathogenicity [12],[13]. Secretions of the animal-parasitic nematode, Trichinella spiralis, appear to reprogram the host cell into a nurse cell, and in vitro injection of collected secretions from T. spriralis into rat muscles mimicked cellular changes that occur in vivo [14].
Pathogen associated molecular patterns (PAMPs) are typically proteins or nucleic acids that are wide-spread in microbes and are shed during infection. Host receptors are activated by PAMPs. For example, flagellin from bacteria stimulates innate immunity from both plant and mammalian cells [15],[16]. No PAMPs from metazoans have been reported.
The identification of secreted proteins from M. incognita may facilitate the discovery of effectors and PAMPs. Effectors might account for how root knot nematodes reprogram plant cells to become giant cells and to form root knots. Several hypotheses have been proposed to explain how M. incognita establishes compatibility with its plant hosts. It may invade root tissues by first producing cell wall-degrading enzymes. Once established in the root it could produce detoxifying enzymes, followed by additional effectors that induce giant cell formation [2].
Discovery of secreted proteins by bioinformatics is possible for organisms with known genomic DNA sequences. Recently a secretome of Plasmodium falciparum comprising 200 proteins was predicted using bioinformatics [17]. A similar approach cannot be applied to M. incognita since its genomic sequence is not yet known. Instead, experimental approaches have been used. A transcript profile of the esophageal gland cells of M. incognita has been reported [18]. Based on bioinformatic analyses of cDNA sequences, secreted proteins were predicted to include cellulases, chitinases, extensins, proteases, and a superoxide dismutase (SOD). In a recent study, Roze et al. (2008) analyzed the cDNA sequences of proteins putatively secreted by Meloidogyne chitwoodi [19]. They identified cDNAs corresponding to 398 putative proteins and confirmed by in situ hybridization seven that are specifically transcribed in the SvG, one in the dorsal gland, and one in the phasmids.
We chose to directly identify secreted proteins based on the pioneering work of Jaubert et al. (2002) who used resorcinol to induce esophageal gland secretion by M. incognita. It was clear from their work that many more proteins were secreted than were identified [20]. To explore the M. incognita secretome in greater depth, we developed sensitive methods for high-throughput proteomics based liquid chromatography, nano-electrospray ionization and tandem mass spectrometry (nanoLC ESI MS/MS) [21]. This method requires both a protein database as well as algorithms to assign peptide sequences to mass spectra. The conservation of protein sequences between species enables a protein database from heterologous species to partially substitute for a database from the cognate species. To control for false positive identifications we reversed the amino acid sequence of the protein databases and filtered the search result so that our protein false discovery rate (FDR) was 0.4%. While use of heterologous databases precludes discovery of peptides that are unique to the organism, thereby reducing the number of proteins identified, it nevertheless opens a window on the proteome. In this study, we identified 486 proteins from the M. incognita secretome, including proteins that could play a role in root knot formation by regulating the plant cell cycle and plant growth.
We induced protein secretion by J2 stage M. incognita nematodes by treating first with filtered, low-molecular weight (<3,500 Da) tomato root exudates followed with resorcinol. A sample of nematodes was removed and stained with Coomassie Blue, which confirmed that treatment caused proteins to be secreted from the stylet region. Secreted proteins were extracted from the solution bathing unstained nematodes and were identified by nanoLC ESI MS/MS (Figure 1). To ensure the accuracy of protein identifications, the threshold for mass spectral quality was set at high stringency using very low peptide and protein false discovery rates (FDRs). FDRs were determined by searching the MS/MS spectra against a concatenated 1∶1 forward-reverse database [22]. MS/MS spectra with peptide FDR less than 0.1% were considered valid. We set the protein FDR at 0.4%. Due to the multiple protein databases used in this study and the natural sequence redundancy in the protein databases, the same peptide sometimes appeared in multiple protein sequences. In order to address this protein redundancy issue, protein sequences containing the same set or subset of valid peptides were grouped together into protein groups with the best match listed first [23]. The numbers of proteins we report in this paper are protein group numbers. This is a conservative measure because more than one protein within a group may actually be detected. Only proteins with at least 2 valid MS/MS spectra were reported. Proteins with a single unique peptide but multiple spectra were manually validated.
Observations of M. incognita and Heterodera glycines stylet activity with and without stimulation by neuroregulators has been extensively studied [24]. The authors reported that neuroregulators induce a dramatic stimulation of stylet pulsing frequency but they pointed out that even without stimulation, stylet pumping occurred. We observed by both mass spectrometry and silver stained gels that J2 nematodes secrete low but detectable levels of proteins (less than 1% as much as after stimulation). Proteins identified in the absence of stimulation included 14-3-3b [listed as protein (4) in Table S1]; Hsp90 (9); SEC-2 (11); aldolase (20); glyceraldehyde-3-phosphate-dehydrogenase (14); protein with thioredoxin domain (52); and protein with glutathione S-transferase domain (43).
We identified 486 proteins from the M. incognita secretome after treatment using a protein FDR of 0.4% (Figure 2). These include all seven proteins reported by Jaubert et al. (2002), indicating that our results both confirm and extend previous studies. The majority of the proteins (311; 64%) were identified by the detection of 2 or more peptides. Of the 175 proteins identified by only one peptide, some were previously shown to be secreted. Proteins identified by several MS/MS spectra but only one peptide have been manually validated and the spectra are summarized in Table S2.
To serve as a control for potential contamination of the secretome by cellular debris, we examined proteins extracted from intact nematodes. Visual inspection of nematode preparations did not reveal any signs of damage or debris. We compared the relative abundance of proteins in the secretome to their abundance in extracts from intact nematodes. This revealed that many (19%) of the secreted proteins are highly abundant in intact nematodes; these were removed from consideration out of concern that they may be contaminants, even though they were not observed in the water control. The normalized spectrum count ratio of each protein (secretome/whole nematode proteome) was used to calculate secretome enrichment. Most of the proteins identified in the solution bathing treated nematodes (i.e., the secretome) were significantly less abundant or absent in the proteome of whole nematodes, providing further evidence that they are indeed secreted (Table S1, column 5). Approximately 81% (394) of the secreted proteins are enriched and 60% (288) are at least 2-fold more abundant in the M. incognita secretome (Figure 3 and Table S1, column 5). Due to the relatively large size of the SvGs and the number of dense granules in them, it would not be surprising to find secreted proteins in the whole nematode extract. The remaining 19% (92) of non-enriched proteins (e.g., actin) may in fact be secreted, but to be conservative we do not consider them further.
High-throughput nano-LC ESI MS/MS depends upon protein databases and is most useful when the entire annotated genome sequence of an organism is available. However, with the proliferation of genome projects, adequate sequence information has become available to enable protein identification using databases from other species. We used two M. incognita cDNA sequence databases with sequence databases from all nematodes and plants (Figure 1 and Table S3). Nearly all of the secreted proteins (481; 99%) were identified by reference to the nematode protein sequences (Figure 2). Approximately half of the proteins (235; 48%) were identified both by M. incognita sequence and by sequence from other nematodes. Only 20% (95) were identified by orthologous nematode sequence alone and 31% (151) from M. incognita sequence alone. A total 69% of the M. incognita secretome could have been identified without reference to the M. incognita DNA sequences (Figure 2). Table S4 shows full-length proteins with identified peptides derived from searching the M. incognita sequence database.
Comparison of our observations with published reports of proteins secreted by M. incognita revealed extensive overlap and, in addition, we identified orthologs of proteins that are secreted by other parasitic nematodes (Table S5). Among the 10 most abundant proteins in our data (Table S1), 14-3-3b protein and calreticulin were previously shown to be produced and secreted by the SvG of M. incognita [25],[26].
Comparison of our M. incognita secretome with that from the parasitic helminth, Brugia malayi, reveals significant overlap [27]. Of the 80 proteins known to be secreted by B. malayi, 26 are also secreted by M. incognita (Table S5). This conserved group includes proteins involved in detoxification (e.g. SODs), cytosolic stress response (e.g. 14-3-3-like proteins), cytosolic energy metabolism (e.g. a triose phosphate isomerase), structure (e.g. actin), protein turnover or folding (e.g. ubiquitin-like proteins SMT3 and protein disulfide isomerases PDI), protease inhibitors (e.g. Cystatin-type Cysteine Protease Inhibitor CPI-2), and two transthyretin-like family proteins (TTLs).
The discovery of effectors from nematodes has lagged behind progress made with bacterial and oomycete pathogens, but recently phytopathogenic nematode effectors have been reported. We re-examined our mass spectra using sequence from members of the SPRYSEC protein family, which includes effectors from G. rostochiensis [28],[29]. We also searched for Cg-1, an M. incognita candidate effector gene acting in the Mi1.2 resistance pathway [30], and for MAP-1, a putative avirulence protein produced by amphids [31]. We did not identify peptides corresponding to any of these proteins in the M. incognita secretome nor in the extract of intact nematodes.
We searched our mass spectra for peptides from proteins secreted by M. chitwoodi but could find none [19]. However, by doing a BLASTP search using proteins identified in our study, we were able to show that 4 proteins secreted by M. chitwoodi are also in the M. incognita secretome (cysteine protease, beta-1,4-endoglucanase, VAP-1 and pectate lyase). The reason we initially missed them is because our search algorithms require exact amino acid sequence matches but the peptides identified in the M. incognita secretome have at least one amino acid difference compared to those deduced from M. chitwoodi ESTs.
Using the euKaryotic Orthologous Groups (KOGs) classification scheme to annotate the secreted proteins [32] we found that 103 proteins catalyze post-translational modifications, protein turn-over or chaperone functions; 93 participate in protein synthesis or secretion; 88 trigger metabolic reactions; 48 interact with nucleic acid (DNA or RNA); 25 are involved in signal transduction and 33 interact with actin or microtubules. We performed a BLASTP search for each protein to refine their annotations (Table S1, column 8 and 9). We combined the KOGs and BLASTP results to classify the M. incognita secretome into 9 subfamilies (Tables S1 and S6): Proteins interacting with actin or microtubules (33 proteins, family 1); Proteins interacting with nucleic acids (48 proteins, family 2); Post-translational modification, protein turnover, and chaperone functions (103 proteins, family 3); Metabolism (88 proteins, family 4); Signal transduction (25 proteins, family 5); Protein synthesis and secretion (93 proteins, family 6); Detoxification (17 proteins, family 7); Cell wall modification enzymes (8 proteins, family 8); and Other (94 proteins, family 9).
Nematode infection causes gene expression changes in the plant cell [33]. These changes could be due to indirect effects, but there is evidence for secreted nematode proteins interacting directly with plant transcription factors (reviewed in references [2],[34]). This was first suggested when putative secreted factors were observed to have nuclear localization signals (NLSs) [18]. Later, an mRNA was identified from the esophageal gland of H. schachtii and the capacity of its expressed protein to interact in planta with two putative plant SCARECROW-like transcription factors was reported [35]. To determine whether the secreted proteins we observed could be targeted to the plant nucleus and could potentially modify plant gene expression, we searched for NLSs and DNA or chromatin interaction motifs. We found 66 proteins that meet one or both criteria: 26 proteins with an NLS motif and 40 additional proteins with putative nucleotide binding activity. Of these, 8 proteins are predicted to have both an NLS and a nucleotide binding activity (Table S7).
We identified 5 secreted proteins present only in the plant protein sequence database (Table S1). Among them was LeMir, a protease inhibitor known to be upregulated in plants during nematode infection. Low molecular weight tomato root exudates were used to induce nematode secretion so we examined as a control the water medium without nematodes for proteins and peptides that could potentially diffuse across the membrane that separated root exudates from nematodes. No proteins were detected in gels by silver staining (data not show) but, using mass spectrometry, we identified 4 peptides derived from 3 plant proteins (Table S8). Only one protein overlapped with the nematode secretome (remorin 1); we could not detect LeMir or any of the other plant homologs in the nematode secretome indicating that they were not contaminants. Earlier reports identified other proteins with putative horizontal gene transfer (HGT) origins. We confirmed that several of these are in the secretome, including two pectate lyases [36], a cellulose binding protein [37], and two beta-1,4-endoglucanases [38],[39]. McCarter et al., (2003) reported cases of potential HGT from microbes; we confirmed their existence in the secretome, including a Rhizobacterial homolog of nodL (CL221Contig1_1) and a polygalacturonase (221104r1.1_1) [40]. Two other putative HGT candidates were identified: a conserved hypothetical protein from Trichomonas vaginalis (MI00116) and a putative Type IV secretory pathway VirB6 component from Ehrlichia canis str. Jake (CL1842Contig1_1) (Table 1).
We localized mRNA corresponding to a subset of secreted proteins using in situ hybridization to J2 stage nematodes (Figure 4). As a positive control, we localized transcripts for two previously characterized secreted proteins from the SvG: beta-1,4 endoglucanase and calreticulin (Figure 4D and G respectively). We tested and confirmed that the following members of the M. incognita secretome are also expressed specifically in the SvG: CL312Contig1_1 (protein with unknown function); CL5Contig2_1 (SEC2); CL2552Contig1_1 (Transthyretin-like family protein homolog); CL321Contig1_1 (Translationally-controlled tumor protein homolog); CL480Contig2_1 (triosephosphate isomerase homolog). A BLASTX search revealed that CL312Contig1_1 encodes for a C. elegans homolog (E value 9E-06) that is predicted to be a membrane protein with unknown function. We also found a transcript that encodes a putative CDC48 protein (contig CL1191Contig1_1) that is enriched in phasmid organs.
We identified several proteins at low levels in the water control (untreated nematodes). This is not surprising since, even in the absence of stimulation, M. incognita could secrete proteins [24]. The identified proteins could play a role in plant ROS signaling or in suppression of plant cell death. Protection from ROS could be provided by glyceraldehyde-3-phosphate-dehydrogenase [41], the transcript product from CL2662Contig1_1_AA (a protein with thioredoxin domain) or from CL2084Contig1_1_AA (a protein with a glutathione S-transferase domain). Additional proteins found in the water control include SEC-2 and 14-3-3b. We showed in this study an enrichment of SEC-2 transcript in the SvG and it was previously reported that 14-3-3b transcript is enriched in the dorsal oesophageal gland and in an undetermined tissue anterior to the metacarpus of M. incognita [25].
Following stimulation, we identified 486 proteins in the M. incognita secretome, representing functions that are potentially required for invasion, immune suppression, and host cell reprogramming. A published scheme classified most secreted proteins into four categories: cell wall degrading enzymes; detoxification enzymes; plant nuclear localized proteins; and giant cell formation [2]. To accommodate for the large increase in protein diversity reported here, we propose expanding the M. incognita secretome classification into 8 categories plus some proteins that were not classified (Table 1 and Table S6); selected examples are described below.
We identified several chaperones that may be involved in protein secretion: thioredoxin, glutathione peroxidases, cyclophilins, and protein disulfide isomerases (PDIs). PDIs have also been found in the secretion of the nematode, Ostertagia ostertagi, where their overexpression increases the yield of secreted proteins [42]. PDIs participate in actin filament polymerization, gene expression, cell-to-cell interactions and in the regulation of receptor functions [43],[44]. Cyclophilins are associated with protein trafficking, protein folding, chromatin remodeling, and chaperone activity [45]. Coaker et al. (2005) showed that the Pseudomonas syringae cysteine protease, AvrRpt2, requires activation by a plant cyclophilin before it can cleave RIN4 [46]. It is possible that M. incognita secretes cyclophilins to activate its effectors.
The correct folding of secreted nematode proteins may be necessary for infection. It has been shown previously that the AVR9 peptide elicitor of Cladosporium fulvum contains three disulfide bridges and that its correct folding depends on the redox state of the environment, with folding rates greatly increased in the presence of PDI [47]. If AVR9 is even partially reduced, it loses all activity, illustrating the importance of disulfide bridges.
SvGs are known to secrete a beta-1, 4-endoglucanase in planta [48],[38],[49], as well as a pectinase [50] and an expansin [51]. We observed these and other cell wall degrading enzymes in the M. incognita secretome indicating that the nematode may use these enzymes for moving through the root or for assisting with plant cell wall remodeling during root knot formation.
One of the earliest plant responses to infection is the production of reactive oxygen species (ROS) [52]. Based on our study, the M. incognita secretome contains detoxification enzymes that may be able to degrade ROSs. This could assist the nematode to establish a successful feeding site. It was previously reported that M. incognita secretes proteins which protect it from ROSs [53]. In plant tissues, SODs exist in three main families containing Cu and Zn, Mn, or Fe in their active site. We found two putative cytosolic CuZnSODs in the M. incognita secretome. A CuZnSOD was highly expressed and active in emergent symbiotic Rhizobium nodules of Lotus japonicus suggesting that this enzyme could play an important role in the early stages of symbiosis [54]. Taking this into consideration it is possible that the nematode enzyme may play a role in establishing compatibility with the plant cell.
ROSs are also scavenged by ascorbate peroxidases, cytochrome C-peroxidases, catalases, thioredoxins and glutathione peroxidases [55]. Two glutathione peroxidases and one thioredoxin were observed in the M. incognita secretome, as were several glutathione S-transferases. Normally these enzymes are induced in plants by H2O2, where they act as calcium-dependent cellular protectants [56], so perhaps the nematode enzymes also provide protection from ROS-catalyzed damage. A similar mechanism has been observed in the maize pathogen, Ustilago maydis, which overcomes host redox defenses by sensing peroxide with Yap1. Once activated, Yap1 induces U. maydis peroxidase gene expression, leading to the successful establishment of infection [57]. Therefore, it is possible that M. incognita may have evolved enzymes to control the global oxidative status of the plant cell as a way to increase its virulence.
Two of the most obvious consequences of nematode infection are distortion of the plant cell-cycle and cytoskeleton, leading to the formation of giant cells and the characteristic root knot [58],[59]. We identified several secreted proteins that could be targeted to the plant cell nucleus, where they could regulate gene expression resulting in some of the morphological changes observed. The target of nematode effectors to the plant nucleus was first suggested by the presence of putative secreted proteins with nuclear localization signals [18]. Later, a small, secreted peptide was identified that interacts in planta with two plant SCARECROW-like transcription factors [35]. We identified 66 secreted proteins with putative nuclear localization, DNA binding, or chromatin modification domains. These include two helicases, several RNA and DNA binding proteins, histones and the Nucleosome Assembly Protein, NAP-1 (Table S7). NAP proteins move histones into the nucleus, assist with nucleosome assembly, and modulate transcription [60].
Several secreted proteins were identified that could potentially regulate plant cell proliferation including a CDC48-like protein (VCP/CDC48), SKP1, TCTP, NAC protein, and a CDPK. We confirmed by in situ hybridization that the corresponding mRNA of the CDC48-like protein is specifically expressed in the nematode phasmid (Figure 4A). A previous study using Coomassie Brillant Blue G-250 revealed that phasmids secrete proteins that take up the stain [61]. Phasmids are specialized pairs of sensory organs found in the posterior lateral field of most nematodes. The function of phasmids remains unclear although a role as receptors for female sex pheromone was proposed for Scutellonema brachyurum [62]. Most plant parasitic nematodes have phasmids [6]. Baldwin (1985) identified two types of phasmids in the J2 stage of H. schachtii: a larger type that secretes and a smaller one that does not [63]. In proliferating cells of Arabidopsis, AtCDC48 is highly expressed, but it's down-regulated in most differentiated cell types [64]. CDC48/VCP/p97 in Zebrafish has been shown to induce cell proliferation [65]. Based on this discovered we can add phasmids to the set of organs that may play a role in nematode parasitism.
S-phase kinase-associated protein 1 (SKP1) is a key component of the SCF complex that provides ubiquitin-protein ligase activity required for cell cycle progression. Gao et al., (2003) identified a SKP1 homolog in the dorsal gland of Heterodera glycines [66]. The SKP1 homologue identified in our study has a nuclear localization signal, and therefore could be potentially targeted to plant nuclei. Translationally-controlled tumor proteins (TCTPs) are highly conserved and are implicated in several different cellular processes including growth, cell cycle progression, malignant transformation, and protection of cells against stress and apoptosis [67]. TCTP proteins are expressed in rapidly growing plant organs, such as the apical meristem, suggesting a role in cell proliferation [68]. Overexpression of TCTP in cultured tobacco cells resulted in faster regeneration and the induction of more calli following Agrobacterium infection [69]. We found that the mRNA for secreted TCTP is enriched in the SvG of M. incognita (Figure 4F) suggesting that TCTP could be actively secreted into the host plant cell.
We observed one Calcium-Dependent Protein Kinase (CDPK) and several CaM proteins in the M. incognita secretome. Using RNA interference, Ivashuta et al. (2005) showed that in Medicago truncatula, CDPK1 is essential for root hair formation and cell elongation [70]. Inactivation of CDPK1 results in significant diminution of Rhizobial and mycorrhizal symbiotic colonization [70]. The CDPK family and signaling pathways are conserved across the plant kingdom [71], so nematodes may have developed the ability to control this central and ubiquitous element of plant development.
We identified several secreted proteins with established or suggested roles in the virulence of parasites. Anand et al. (2007) [72] used virus-induced gene silencing and an in planta tumorigenesis assay to identify plant genes involved in Agrobacterium-mediated plant transformation. They identified several genes that were required to produce the crown gall phenotype; we identified homologs in the M. incognita secretome. Among them were SKP1, actin or actin-binding proteins, and histones H3, H2a, and H2b. Histone H2a is required for T-DNA integration [73] and histone H3 has also been implicated [72]. We found a homolog of the Nodulin protein, NodL, in the M. incognita secretome, which is similar to the nodulin-like proteins (NLP) required for Agrobacterium-mediated transformation [73]. Root knot nematodes induce cytoskeletal changes that closely resemble those induced by Nod proteins [74]. MtENOD11 is expressed early following both arbuscular mycorrhizal infection and Meloidogyne infection of Medicago [75]. Therefore, it is possible that root knot nematodes use a Nod-like pathway to initiate giant cell formation.
We were surprised to observe that plant and animal metazoan parasites secrete a common set of proteins. For example, B. malayi and M. incognita both secrete transthyretin-like protein (TLP or TTL), which is a member of a growing family of transthyretin (TTR)-related proteins (TRPs). TRPs are related to the vertebrate transthyretin, an extracellular thyroid hormone carrier protein [76]. TRPs may represent the ancestor of the vertebrate thyroid hormone carriers [77]. We found in the M. incognita secretome a TTL and confirmed that its corresponding transcript is specifically expressed in the SvG of J2 stage nematodes (Figure 4E). Therefore, we reason that this TTL homolog is secreted into the plant cell where it regulates growth. A plant TTL is known to interact with the brassinosteroid receptor kinase to control plant growth [78].
Both cysteine (CPI-2) and aspartyl (API-2) protease inhibitor (PI) family members were observed in the M. incognita secretome. The function of PIs in nematodes is to protect their intestine from dietary proteases [79]. In plants, endogenous PIs are active against all four classes of proteinase (cysteine-, serine-, aspartyl-, and metallo-). PIs accumulate following wounding or herbivory and they may provide protection [80]. PIs have also been shown to regulate programmed cell death (PCD). For example, synthetic peptide inhibitors of caspases could suppress PCD induced by a Ps. syringae infection of tobacco [81]. Recently a cystatin CPI-2 protease inhibitor was identified in B. malayi secretions and it was proposed to inhibit host proteases required for antigen processing and presentation [27].
The M. incognita secretome contained metallopeptidases, aminopeptidases, a cysteine proteinase, proteasome components, and proteins involved in ubiquitination. Secreted proteases could have two obvious functions: either the destruction of plant defense proteins or nutritional pre-digestion. Cysteine proteinases are involved in both the initiation and execution of the cell death program [82] and intriguingly we found two kinds of cysteine proteinase inhibitors.
G. rostochiensis has been shown to secrete metalloproteases [11], as have other phytopathogenic nematodes and free living nematodes; a role in the hatching process was proposed for the latter [83],[11]. Nematode metalloproteases could catalyze protein degradation in planta to enable uptake of proteins that are otherwise too large [84].
We identified several ubiquitin proteins in the M. incognita secretome. Tytgat et al., (2004) [85] identified a ubiquitin extension protein secreted from the dorsal pharyngeal gland of root cyst nematodes. The ability of pathogens to manipulate the ubiquitination-proteasome system of animal immune systems is known (for a review see Loureiro and Ploegh., 2006) [86]. The ubiquitin pathway is required for innate immunity in Arabidopsis [87].
We identified 94 proteins that we were unable to classify (Table S6). Among them we had shown that the transcripts of two genes, SEC2 and CL312Contig1_1, are enriched in the SvG of J2 stage M. incognita (Figure 4C and 4B respectively). We identified several proteins with a putative function but we were unable to discern a role for them in pathogenicity. One example is a triosephosphate isomerase (TPI) homolog that is highly secreted (ratio secreted/whole = 10.35); the corresponding transcript (CL480Contig2_1; Figure 4H) is enriched in the SvG of J2 stage M. incognita. BLASP and KOG annotation revealed a putative function of this protein in metabolism. However, a similar TPI was also found in the fungal mammalian pathogen Paracoccidioides brasiliensis where TPI localized to the cell wall and cytoplasmic compartments [88]. The authors suggested that TPI is required for interaction between P. brasiliensis and the extracellular matrix and could be important for fungal adherence to and invasion of host cells. A similar function could be postulated for the M. incognita TPI since after the mobile J2 stage, the parasitic nematode is sedentary and is in close contact with plant tissue.
The development of sensitive proteomics methods has allowed us to significantly expand the known secretome of M. incognita. A rich set of candidates has been found that can now be functionally evaluated. Conservation of protein sequence allowed us to search our mass spectra using sequence databases from other nematode species and plants. Nearly half (48%) of our identifications from heterologous sequence databases were confirmed by matches to the limited M. incognita sequence that is publicly available, suggesting that proteomics can be useful even with nematodes for which no sequence information is available. As more M. incognita DNA sequence becomes available, we can probably identify additional proteins by re-searching our mass spectra. We confirmed that most secreted proteins are produced by esophageal glands and we found direct evidence for one secreted by phasmids [2],[34]. Twenty-six proteins overlap between the M. incognita and B. malayi secretomes (Table S5). These include proteins with potential functions in parasitic behavior (e.g., TCTP; Cystatin CPI-2). This remarkable conservation of sequence raises the possibility that plant and animal parasitic nematodes share conserved mechanisms of infection.
Meloidogyne incognita was propagated from greenhouse-grown tomato plants (Solanum esculentum cv. Rutgers). After 8 weeks of infection, eggs were recovered from tomato plants by shaking M. incognita-infected roots in 1∶9 dilution of bleach for 3 min in a flask. Eggs were collected onto a 25 µm mesh and were then bleached twice for 10 minutes with a 1∶5.7 dilution of bleach supplemented with 0.02% Tween 20. Eggs were rinsed four times with sterile ddH2O. Twenty million eggs were hatched at room temperature for 3 days in 10 mM Tris pH 7.0 with 300 mg/l carbenicillin (hatching buffer), and juvenile 2 stage (J2) worms were allowed to crawl though five Kimwipe tissue layers into the same hatching buffer. Freshly hatched J2s were washed several times in sterile water and then collected on 8 µm sieves.
Tomato seeds (Solanum esculentum cv. Rutgers) were placed above a plastic cylinder filled with cotton fiber and placed into an aerated hydroponics vessel constructed from a 2-liter flask. Hydroponic vessels were supplied with 250 ml sterilized solution of 0.5× Gamborg media basal salts medium complemented with 1× Gamborg vitamins, 0.5% sucrose and 200 mg/l carbenicillin. Tomato plants were maintained under a 16-h photoperiod for 6 weeks and root media was collected and filtered through a 0.22 µm syringe filter to give the “hydroponic tomato root culture solution”.
Hatched J2s were stimulated for 4 hours by hydroponic tomato root culture solution separated from the nematodes by a 3,500 MW cutoff mini dialysis membrane (Pierce, Rockford, USA). Then they were treated for 4 h with 0.4% resorcinol (Sigma-Aldrich Chimie, St Quentin, France). Stylet secretions were filtered through a 0.22 µm syringe filter to remove nematodes.
Secreted proteins were concentrated to ∼1 ml in a vacuum centrifuge at room temperature. Tris buffer was added to a final concentration of 20 mM (pH 7.2). Proteins were reduced and alkylated using 1 mM Tris (2-carboxyethyl) phosphine (Fisher, AC36383) at 65°C for 30 minutes and 2.5 mM iodoacetamide (Fisher, AC12227) at 37°C in dark for 30 minutes, respectively. Proteins were then digested with 1 µg trypsin (Roche, 03 708 969 001) at 37°C overnight.
Whole M. incognita worms were lysed in 100 µL 2% (w/v) RapiGest (Waters) by sonicating in a Branson Sonifier 450 fitted with a high intensity cup horn (Part No. 101-147-046, Branson) at 4°C for 2 minutes. Crude lysate was spun down at 16,100 g at 4°C for 5 min. Supernatant was collected and the pellet was discarded. RapiGest was diluted to 0.5% (w/v) by adding 300 µL of 20 mM Tris. Proteins were reduced and alkylated as described above. Protein concentration was measured using a Bradford assay. Protein (400 µg) was digested with 10 µg trypsin (Roche, 03 708 969 001) at 37°C overnight.
TFA (0.5% v/v) was added to each sample to a final pH of 1.8 to precipitate RapiGest after digestion. Samples were incubated at 4°C overnight and then centrifuged at 16,100 g at 4°C for 15 minutes. Supernatants were collected and centrifuged through a 0.22 µM filter to clear any solid particles.
An Agilent 1100 HPLC system (Agilent Technologies, Wilmington, DE) delivered a flow rate of 300 nL min−1 to a 3-phase capillary chromatography column through a splitter. Using a custom pressure cell, 5 µm Zorbax SB-C18 (Agilent) was packed into fused silica capillary tubing (200 µm ID, 360 µm OD, 20 cm long) to form the first dimension reverse phase column (RP1). A 5 cm-long strong cation exchange (SCX) column packed with 5 µm PolySulfoethyl (PolyLC) was connected to RP1 using a zero dead volume 1 µm filter (Upchurch, M548) attached to the exit of the RP1 column. A fused silica capillary (100 µm ID, 360 µm OD, 20 cm long) packed with 5 µm Zorbax SB-C18 (Agilent) was connected to SCX as the analytical column (RP2). The electro-spray tip of the fused silica tubing was pulled to a sharp tip with the inner diameter smaller than 1 µm using a laser puller (Sutter P-2000). The peptide mixtures were loaded onto the RP1 column using the custom pressure cell. Columns were not re-used. Peptides were first eluted from the RP1 column to the SCX column using a 0 to 80% acetonitrile gradient for 150 minutes. The peptides were then fractionated by the SCX column using a series of salt gradients (from 10 mM to 1 M ammonium acetate for 20 minutes), followed by high resolution reverse phase separation using an acetonitrile gradient of 0 to 80% for 120 minutes. To avoid sample carry-over and keep good reproducibility, a new set of three columns with the same length was used for each sample.
Spectra were acquired on LTQ linear ion trap tandem mass spectrometers (Thermo Electron Corporation, San Jose, CA) employing automated, data-dependent acquisition. The mass spectrometer was operated in positive ion mode with a source temperature of 150°C. As a final fractionation step, gas phase separation in the ion trap was employed to separate the peptides into 3 mass classes prior to scanning; the full MS scan range was divided into 3 smaller scan ranges (300–800, 800–1100, and 1100–2000 m/z) to improve dynamic range. Each MS scan was followed by 4 MS/MS scans of the most intense ions from the parent MS scan. A dynamic exclusion of 1 minute was used to improve the duty cycle.
MS/MS spectra were collected for secreted and whole M. incognita proteins (2,904,233 and 947,474 spectra, respectively). Raw data were extracted and searched using Spectrum Mill (Agilent, version A.03.02). MS/MS spectra with a sequence tag length of 1 or less were discarded. MS/MS spectra were searched against the protein databases (Table S3). The enzyme parameter was limited to full tryptic peptides with a maximum mis-cleavage of 1. All other search parameters were set to Spectrum Mill's default settings (carbamidomethylation of cysteines, +/−2.5 Da for precursor ions, +/−0.7 Da for fragment ions, and a minimum matched peak intensity of 50%).
To eliminate redundant protein identifications, proteins with one or more shared peptides were grouped. The numbers of proteins we report in this paper are protein group numbers. A concatenated forward-reverse database was constructed to calculate the in situ false discovery rate (FDR) [22]. We used an identification filtering criteria of 0.1% FDR at the peptide level for every search. A total of 486 secreted proteins from the forward database were identified, while 2 proteins (0.4% protein FDR) from the reverse database were identified.
Spectrum counting was used to determine the relative protein amounts in the secretome and the extract from intact nematodes. The number of valid MS/MS spectra from each protein was normalized to the total MS/MS spectra number of each dataset. The normalized spectrum count ratio of each protein (secretome/intact nematode proteome) was used to evaluate whether the protein was enriched in the secretome. The data associated with this manuscript may be downloaded from ProteomeCommons.org Tranche, http://tranche.proteomecommons.org, using the following hash (without the quotes): “FXMi2Tyve1I0DfzhT9FN17TmpNTTDiggs7Njjoh7MYMouHYIx+xUoDILMXFl17RZrVjueXuCZc5c3005l9fdKISeVUEAAAAAAAB0ug = = ”.
While this paper was under review, Abad et al. [89] reported the draft genome sequence of M. incognita. The 9,538 contigs resulting from the M. incognita genome assembly and annotation were deposited in the EMBL/Genbank/DDBJ databases under accession numbers CABB01000001–CABB01009538 for release at a future date. When these contigs become publicly available, further bioinformatics analysis of our mass spectra can be conducted to search for additional secreted proteins.
Identified protein sequences were BLASTed against the non-redundant database at NCBI (http://www.ncbi.nlm.nih.gov/). euKaryotic Orthologous Group (KOG) annotations were assigned based on sequence similarity searches against the KOG annotated proteins (http://www.ncbi.nlm.nih.gov/COG/grace/kognitor.html).
Putative nuclear function was assigned based on homologous proteins found using BLASTP or on the identification of a Nuclear Localization Site (NLS). The NLS search was performed using the PredictNLS search engine available at http://cubic.bioc.columbia.edu/predictNLS/ [90].
In situ hybridizations were performed on freshly hatched J2s as described in Rosso et al. [38]. Briefly, freshly hatched J2s were fixed in 2% paraformaldehyde for 16 h at 4°C and 4 h at room temperature. Nematodes were cut into sections and permeabilized with proteinase K, acetone, and methanol. The sections were then hybridized at 45°C with the sense or antisense probe. Clone and primers are listed in Table S9.
|
10.1371/journal.ppat.1000236 | Immune Mechanisms Responsible for Vaccination against and Clearance of Mucosal and Lymphatic Norovirus Infection | Two cardinal manifestations of viral immunity are efficient clearance of acute infection and the capacity to vaccinate against secondary viral exposure. For noroviruses, the contributions of T cells to viral clearance and vaccination have not been elucidated. We report here that both CD4 and CD8 T cells are required for efficient clearance of primary murine norovirus (MNV) infection from the intestine and intestinal lymph nodes. Further, long-lasting protective immunity was generated by oral live virus vaccination. Systemic vaccination with the MNV capsid protein also effectively protected against mucosal challenge, while vaccination with the capsid protein of the distantly related human Lordsdale virus provided partial protection. Fully effective vaccination required a broad immune response including CD4 T cells, CD8 T cells, and B cells, but the importance of specific immune cell types varied between the intestine and intestinal lymph nodes. Perforin, but not interferon gamma, was required for clearance of MNV infection by adoptively transferred T lymphocytes from vaccinated hosts. These studies prove the feasibility of both mucosal and systemic vaccination against mucosal norovirus infection, demonstrate tissue specificity of norovirus immune cells, and indicate that efficient vaccination strategies should induce potent CD4 and CD8 T cell responses.
| Human noroviruses are the most common cause of epidemic nonbacterial gastroenteritis in the world. Despite their importance as human pathogens, little is known about how the immune system controls and clears norovirus infection, and the potential and mechanisms of vaccination remain unclear. Here, we used norovirus infection of mice to show that vaccination can provide long-lasting immunity against mucosal norovirus challenge and to identify the types of immune cells that are important in vaccination against norovirus infection. Similarly, we identified the types of immune T cells that are important for clearance of acute infection. Efficient vaccination required all three major arms of adaptive immunity: CD4 T cells, CD8 T cell, and B cells. Importantly, protective vaccination against mucosal challenge was observed after either mucosal or systemic norovirus antigen exposure. The pore-forming molecule perforin was important for T cell-mediated control of norovirus infection. Our study has important implications for understanding adaptive immunity to norovirus infection, and may provide insight into the directions to take in developing a human norovirus vaccine.
| More than 90% of epidemic nonbacterial gastroenteritis worldwide can be attributed to human noroviruses (HuNV) [1]–[3]. Infection is transmitted fecal-orally, and symptomatic infection is characterized by nausea, vomiting and/or diarrhea lasting 24–48 hours within 24 hours of exposure [4]. Despite the significant costs and morbidity of HuNV infections, no vaccine is currently available. The elderly and individuals in long-term care facilities may be more susceptible to either norovirus infection or norovirus-induced disease [5] and would be an important target population for a norovirus vaccine. The reasons for increased incidence and/or susceptibility to HuNV disease are unknown. This is due in part to our incomplete understanding of norovirus immunity. The potential to vaccinate against these and related viruses has been demonstrated in gnotobiotic piglets, cats and rabbits [6]–[8], but the immune mechanisms responsible have not been identified. The challenges for vaccine efficacy may be very different between different caliciviruses. For example, variation in MNV strains is significantly less than between HuNV strains [9]. Human volunteer studies demonstrate short-term, but not long-term, protection against homologous, but not heterologous, viral challenge [10]–[12]. Since HuNV belong to 3 genogroups (GI, GII and GIV) with many strains in each genogroup [4], this lack of cross-protection is a challenge for vaccine development. Frequent exposure to noroviruses within short time periods stimulates sustained immunity and resistance to norovirus induced illness [13],[14]. Serum antibody levels in adults reflect susceptibility to infection and do not always correlate with protection [13],[14]. In children, however, serum antibody levels correlate with protection, likely reflecting short-term immunity and recent exposure [15]–[17]. A nonfunctional fucosyl transferase gene (FUT2) accounts for a significant proportion, though not all, of resistance to Norwalk virus infection, suggesting that other factors, yet undiscovered, may contribute to norovirus resistance [18],[19].
In the absence of a cell culture system for HuNV, virus like particles (VLPs) that assemble when the viral capsid protein is expressed have been important for evaluating norovirus immune responses [20]–[23]. Studies using Norwalk Virus (GI), Snow Mountain Virus (GII) and HuNoV-HS66 (GII) VLPs to evaluate immunity after infection with live virus or immunization with VLPs orally show production of T cell effector cytokines such as IL-2 and interferon γ (IFN-γ) and proliferation of norovirus specific T cells after in vitro restimulation with VLPs [24]–[26]. These studies show that T cell responses develop, but do not define their role in either clearance of primary infection or resistance to re-challenge. Together, they suggest the potential for vaccination, but leave open important questions about the effectiveness and longevity of vaccine immune responses, mechanisms of vaccination, the viral protein targets for protective responses, and the potential for cross-protection between distantly related noroviruses.
The identification of the first murine norovirus, MNV, and its propagation in cultured cells provides a facile animal model for studies of norovirus immunity and pathogenesis [27],[28]. MNV, an enteric virus that infects tissues of the gastrointestinal tract, is spread by the fecal-oral route ([27] and unpublished studies). The MNV genome encodes four open reading frames. ORF1 encodes a polyprotein that is cleaved into individual non-structural proteins similar to the polyprotein of HuNV [29]. ORF2 encodes the major capsid protein VP1 and ORF3 encodes a minor capsid protein. The existence of a protein product for ORF4 has not been confirmed. In the MNV virion structure, the capsid, like that of human noroviruses, consists of 90 dimers of VP1 [30]. There are differences between the MNV virion and previously reported VLP structures. The MNV protruding domain is lifted off the shell domain by approximately 16 Angstroms and rotated approximately 40 degrees in a clockwise fashion, forming interactions at the P1 base in an infectious virion that have not been observed previously. The existence of these novel aspects of the structure are consistent with the hypothesis that MNV may undergo a capsid maturation process [30].
Studies of MNV pathogenesis reveal an important role for interferon (IFN) and STAT-1 mediated innate immunity in resistance to infection and MNV induced lethality [27],[31]. The importance of adaptive immunity in control of MNV infection is indicated by the observation that RAG1-/- mice develop persistent MNV infection while wild type (WT) mice can clear infection with some strains of MNV [9],[27],[31].
While MNV is an efficient enteric virus that infects many mice in research mouse colonies around the world, diarrhea has not been reported after MNV infection. Thus, MNV provides an infection only model for HuNV infection. Viral titers in tissues of infected mice have not been reported to exceed 106 PFU/ml, and this highest level of viral titer is obtained after infection of highly susceptible STAT1-/- mice [31]. In RAG1-/- mice and WT mice, viral titers of 102 to 104 PFU/ml are routinely observed [9],[31]. The availability of a plaque assay for MNV allows the analysis of MNV infection despite these low titers. Some MNV strains persist at a low level in WT mice, while others are cleared from intestine, spleen, liver, mesenteric lymph nodes (MLN) and feces within 7 days of infection [9],[27],[31]. Additionally, in wild type C57BL6/J mice MNV replicates maximally in the distal ileum [9], in comparison to wild type 129S6/SvEvTac mice where replication occurs in the proximal intestine [31]. The significance of these differences is not known.
Studies of norovirus infection in human volunteers have not specifically investigated whether the infection spreads beyond the intestine to the local lymph nodes, however, it is possible that systemic invasion occurs in humans with chronic conditions or immunosuppressed hosts [32]–[36]. Additionally, viremia has been reported in infections of gnotobiotic pigs and calves [26],[37],[38]. Thus, the ability of MNV to spread to tissues other than the intestine after oral infection may not be unique, but the relationship of this aspect of MNV pathogenesis to human infection is not clear. The availability of strains that can be cleared from WT mice, such as MNV1.CW3, provides an opportunity to define the mechanisms responsible for two cardinal aspects of viral immunity: the capacity to effectively clear acute infection and the immune mechanisms responsible for effective vaccination.
B cells and MNV specific antibody are important in the clearance of primary MNV infection [39], but the role of T cells in clearance and the potential and mechanisms of vaccination against mucosal norovirus challenge are unknown. We show here that vaccination with either live MNV or Venezuelan Equine Encephalitis replicon particles (VRPs) expressing the MNV capsid protein VP1 protect the intestine against re-challenge for at least six months. Live virus was more effective than VRP-mediated vaccination. There was partial cross protection against MNV infection after vaccination with a HuNV capsid protein. We found that both the clearance of primary infection and vaccination require the concerted efforts of CD4 T cells, CD8 T cells, B cells, and that T cells required the effector molecule perforin for maximal impact on MNV infection. The effects of specific immune cell types were tissue specific, differing between ileum and mesenteric lymph nodes. These are the first studies to demonstrate immune mechanisms responsible for norovirus clearance and vaccination.
We first determined whether we could detect short-term immunity to homologous MNV challenge and whether proteins encoded by specific MNV ORFs could elicit effective immunity. VRPs expressing ORF1, ORF2 and ORF3 of MNV1.CW3 and ORF2 of the HuNV Lordsdale (genogroup GII.4) and Chiba (genogroup GI.4) were produced for vaccination experiments. Western blots of VRP-infected cell lysates revealed proteins of appropriate sizes [29],[40] and additionally showed that hyper-immune polyclonal rabbit antisera to MNV [28] cross-reacted at low levels with VLPs from Chiba virus and Lordsdale virus (Figure S1A).
WT mice were vaccinated and boosted three weeks later. Two weeks after boosting, mice were challenged with MNV1.CW3 and organs titered for MNV three days later (Figure 1A). In these WT mice, maximal MNV replication in the intestinal tract occurs in the distal ileum [9] and viral titers could not be detected in duodenum/jejunum (data not shown). After oral inoculation with MNV1.CW3, WT mice exhibit detectable viral titers in the distal ileum and the MLN three to five days post-infection [9],[31]. Prior infection with either MNV1.CW1 (p = 0.0002) or MNV1.CW3 (p = 0.0009) significantly decreased MNV1.CW3 replication in the distal ileum compared to control mice infected with reovirus (Figure 1B). Similar decreases were observed in the MLN after vaccination with MNV1.CW1 (p = 0.0001) or MNV1.CW3 (p = 0.0003) (Figure 1C) compared to the reovirus controls. Similar results were observed in the spleen (data not shown). There was no statistically significant difference between vaccination with MNV1.CW1 or MNV1.CW3. This demonstrates that a protective secondary immune response develops after clearance of primary MNV infection.
ORF2 VRPs protected against MNV1.CW3 in both distal ileum (p = 0.005) and MLN (p = 0.02) compared to control VRPs expressing hemagglutinin (HA) from a mouse adapted influenza A virus [41] (HA VRP control group). Controls for VRP vaccination also included PBS. HA VRP controls were not significantly different from PBS controls across all experiments and statistical comparisons for VRP vaccination are therefore shown to HA VRP controls. ORF1 VRPs alone in the distal ileum, or in both the distal ileum and MLN when combined with ORF3 VRPs, had a small but statistically significant effect on MNV1.CW3 levels (Figure 1B and 1C). ORF3 VRPs alone did not confer significant protection (Figure 1B and 1C). Together these data show that vaccination with either live virus or ORF2 VRPs can confer short-term protection against MNV challenge.
We next assessed vaccination with heterologous ORF2 proteins. Mice were vaccinated and boosted with VRPs expressing ORF2 from Chiba Virus or Lordsdale virus and challenged with MNV1.CW3. Vaccination with Lordsdale virus capsid led to statistically significant protection against MNV infection in the distal ileum, (p = 0.0007, Figure 1B) but not the MLN (Figure 1C). No significant reduction in MNV titers was seen after immunization with Chiba virus capsid (Figure 1B and 1C). Protection after Lordsdale ORF2 VRP vaccination did not correlate with generation of cross-reactive serum IgG in these mice, measured by ELISA, despite the potential for such cross-reactivity revealed by western blot (Figure S1B). Fecal extracts from immunized mice yielded no measurable homotypic or heterotypic IgG or IgA (data not shown). Taken together, these data show that there is measurable functional immunologic cross protection between Lordsdale virus and MNV in the distal ileum. The lack of a correlation between serum or fecal antibody responses and protection suggested that protection may be T cell mediated.
Since older adults may be more susceptible than younger adults to norovirus infection or disease [5], we determined whether increased age altered vaccine efficacy. Prior work has shown that mice older than 1 year of age have diminished vaccine responses to SARS virus antigens [42]. We therefore compared vaccine efficacy in adult (8 week old) and aged (14 month old) mice. Adult and aged mice were vaccinated and challenged as before. In contrast to studies using SARS virus antigens [42], aged mice responded as well as adult mice to MNV ORF2 vaccination in both the distal ileum and MLN (Figure 1D and 1E). Despite this protective effect, sera from vaccinated aged mice had significantly lower anti-MNV ORF2 IgG compared to adult mice (Figure S1C). These data indicated that protection against MNV infection occurred in the absence of robust serologic responses, again raising the possibility that T cells play a fundamentally important role in vaccination against MNV.
We next determined whether protection conferred by MNV1.CW3 or MNV ORF2 VRPs was long lived. WT mice were primed and boosted as shown in Figure 2A with MNV1.CW3 or MNV ORF2 VRPs. Mice were then challenged with MNV1.CW3 two, four, 14, or 24 weeks later and MNV titers measured three days post-challenge. Two weeks post-boost, we observed complete protection against ileal MNV1.CW3 infection after vaccination with either MNV1.CW3 (p = 0.0001) or ORF2 VRPs (p<0.0001) compared to reovirus or HA VRP controls (Figure 2B). At two weeks, while vaccination with either MNV1.CW3 or ORF2 VRPs limited MNV1.CW3 replication in MLN, live virus vaccination was more effective (p<0.0001) (Figure 2C). Live virus vaccination conferred full protection against MNV1.CW3 replication in both the distal ileum and the MLN at four, 14 and 24 weeks after vaccine boost. Vaccination with ORF2 VRPs was also protective, albeit less effective than vaccination with MNV1.CW3 (Figure 2B and 2C). Thus both live virus and subunit vaccine induce long-term protection against MNV infection, with live virus vaccination providing more complete protection.
We next determined the mechanism(s) responsible for effective vaccination. We vaccinated mice lacking both major histocompatibility complex (MHC) Class I and β2 microglobulin (β2M) [43] (CD8 T cells deficiency [44]), MHC Class II (CD4 T cells deficiency [45]), or B cell deficient mice [46] (Figure 3A). These experiments were conducted concurrently with the experiments in Figure 2 above, as such the data from WT mice are repeated in the figure for comparison. Live MNV vaccination induced significant protection against MNV challenge in both the distal ileum and the MLN of B cell-/-, MHC Class II-/- and MHC Class I×β2M-/- mice (p<0.05 in all cases, Figure 3B and 3C). However, there was considerable variation in the efficacy of vaccination in distal ileum and MLN between different immunodeficient strains. In B cell-/- mice, after vaccination with live virus, only 2 out of 15 mice had any titer (and those two mice had less than 100 PFU of MNV) and in MHC Class I×β2M-/- mice, similar vaccination led to undetectable viral titers in the distal ileum (Figure 3B) but detectable titers in the MLN (Figure 3C). In MHC Class II-/- mice, there were detectable titers in both tissues (Figure 3B and 3C). Results for ORF2 vaccination showed that protection required the activity of all major aspects of the adaptive immune response (Figure 3B and 3C). Moreover, there was no protection elicited by ORF2 vaccination in either intestine or MLN tissue after vaccination of MHC Class I×β2M-/- mice with ORF2 VRPs (Figure 3B and 3C) indicating that protection by VRPs critically depends on CD8 T cells. These data demonstrated that complete protection in all tissues after vaccination with live virus required the concerted actions of B cells, MHC Class II, MHC Class I and β2M. Further, the results were consistent with tissue specific roles for B cells, CD4 T cells and CD8 T cells in the development of complete protection against MNV infection.
We next determined whether the same cell types that were required for vaccination were also required for efficient clearance of acute infection. We focused on the role of T cells in clearance since the role of B cells in clearance has already been demonstrated [39]. To determine the role of T cells in clearance of acute MNV infection we inoculated WT, MHC Class II-/-, and MHC Class I×β2M-/- mice orally with MNV1.CW3 and measured viral titers in the distal ileum and MLN three, five, seven and 21 days post-infection (Figure 4B–4E).
There was no significant difference in viral titer between MHC Class I×β2M-/-mice and WT mice at three and five days post-infection, indicating that MHC Class I and β2M were not required in MNV infection at early time points (Figure 4B and 4C). However, at seven days post-infection, MHC Class I×β2M-/- mice had significant levels of MNV titers in both the distal ileum (p = 0.0002) and the MLN (p<0.0001) compared to WT mice, which had completely cleared the infection (Figure 4B and 4C). MHC Class I×β2M-/- mice eventually cleared MNV infection, demonstrated by the lack of viral titers at 21 days post-infection. Thus, MHC Class I and β2M, and by inference CD8 T cells were important for efficient clearance of MNV, but were not required for eventual clearance of MNV infection.
In contrast to MHC I×β2M-/- mice, MHC Class II-/- mice had higher MNV titers in the ileum than WT mice both three (p = 0.0002) and five (p = 0.0058) days after infection (Figure 4D). At seven days post-infection, minimal viral titers remained and by eight days post-infection, both MHC Class II-/- and WT mice had cleared the infection from the distal ileum. In MLN, viral titers in WT and MHC Class II-/- were not significantly different at days three and five post-infection. However, there was a small, but statistically significant increase in titer in the MLN of MHC Class II-/- compared to WT mice at seven days post-infection (p = 0.04, Figure 4E). By eight days post-infection, MLN infection was cleared. Together these data indicated that MHC Class II, and by inference CD4 T cells, were necessary for control of acute MNV infection but are not required for eventual clearance of MNV infection.
To exclude the possibility that the phenotypes we observed in MHC Class I×β2M-/- and MHC Class II-/- mice were due to abnormal immune ontogeny in knockout mice, we determined the requirement for CD4 and CD8 T cells in the clearance of primary MNV infection in WT mice depleted of CD4 and CD8 T cells. Depletion of CD4 and CD8 T cells was at least 90% effective as assessed by flow cytometry of isolated splenocytes (Figure 5A) and this depletion protocol is effective at depleting T cells in secondary lymphoid organs and the intestine [47],[48]. In comparison to control antibody, depletion of CD4 T cells, led to a significant increase in MNV titers in the distal ileum (p = 0.0053, Figure 5B), but not the MLN (Figure 5C). In contrast, depletion of CD8 T cells led to an increase in MNV titers in both the distal ileum (p = 0.004, Figure 5B), and the MLN (p = 0.0025, Figure 5C).
Together, these data from primary challenges of non-immune mice lacking antigen presenting molecules or depleted of specific T cell subsets demonstrated that CD4 T cells are important for efficient MNV clearance in the distal ileum especially at days three and five, while their role in the MLN is small. CD8 T cells are important for efficient clearance of MNV infection in both the MLN and distal ileum, and they function later in infection than CD4 T cells, being most important at days six and seven.
We next determined whether CD4 and CD8 T cells from vaccinated mice can, alone or in combination, clear MNV infection from mucosal sites. We have previously shown that MNV infected RAG1-/- mice have high levels of viral RNA present in multiple tissues up to 90 days post-infection [27]. We therefore determined MNV viral titers in RAG1-/- mice. By 42 days post-infection, all RAG1-/- mice had consistent, high levels of MNV in both duodenum/jejunum and distal ileum (Figure 6A and 6B), as well as several other tissues (data not shown). These data confirmed that mice lacking adaptive immunity fail to clear MNV infection [27].
The availability of persistently infected RAG1-/- mice allowed us to determine the role of CD4 and CD8 T cells in clearance of MNV infection using adoptive transfer of splenocytes from MNV immune WT mice into persistently infected RAG1-/- mice. Transfer of immune, but not non-immune, splenocytes significantly reduced MNV titer in the duodenum/jejunum (p<0.0001) and distal ileum (p<0.0001) six days post-transfer (Figure 7B and 7C). Thus, adoptively transferred immune splenocytes were sufficient to clear persistent MNV infection in the intestine of RAG1-/- mice.
To define which cells were required for MNV clearance, CD4 or CD8 T cells were depleted from splenocytes transferred into RAG1-/- recipients. Anti-T cell antibodies effectively depleted the appropriate T cell populations, as measured six days post-transfer by flow cytometry (Figure 7A). Depletion of either CD4 or CD8 T cells individually led to a significant increase in MNV titers in duodenum/jejunum compared to control depletion (Figure 7B, CD4 depletion p = 0.0042; CD8 depletion p = 0.0002). Depletion of both CD4 and CD8 T cells from transferred immune splenocytes caused a significant additional increase in MNV titers when compared to either CD4 depletion alone (p = 0.02) or CD8 depletion alone (p = 0.03). In the distal ileum, depletion of either CD4 T cells (p = 0.0003) or CD8 T cells (p<0.0001) led to a significant increase in MNV titers (Figure 7C). These data demonstrated that both immune CD4 and CD8 T cells were necessary for clearance of persistent MNV infection from the intestine.
Two major effector mechanisms for the antiviral effects of T cells are the production of IFNγ and perforin mediated cytolysis [49]. We therefore adoptively transferred immune splenocytes from IFNγ-/- or perforin-/- mice into persistently infected RAG1-/- mice and determined their capacity to clear intestinal MNV infection. Immune splenocytes from IFNγ-/- mice were as effective as those from WT mice (Figure 7B and 7C). However, immune splenocytes from perforin-/- mice were less effective at clearing MNV infection from the duodenum/jejunum (p = 0.0003, Figure 7C) or distal ileum (p = 0.0075, Figure 7B) than cells from either WT or IFNγ-/- mice, but more effective compared to transfer of non-immune cells in duodenum/jejunum (p = 0.0086) or distal ileum (p = 0.0001). Thus, while perforin was critical for efficient clearance of MNV infection from the intestine, it was not the only relevant effector mechanism.
In this paper we define the mechanisms of immunity to norovirus infection as measured by both vaccination against, and clearance of, mucosal infection. We found that it is possible to generate highly effective, and remarkably long lasting, immunity to norovirus infection by oral exposure to live virus. Further, systemic exposure to the viral capsid protein expressed in a vaccine vector resulted in effective immunity, albeit not as effective as that observed after live virus vaccination. Importantly, this shows that the MNV VP1 protein contains relevant B cell, CD4 T cell and CD8 T cell epitopes. Vaccination was effective in aged mice. Additionally, vaccination in adult mice required the concerted action of CD4 T cells, CD8 T cell, and B cells to be completely protective in the tissues surveyed. Interestingly, the activities of different components of the adaptive immune system in clearance and vaccination were tissue specific, with different cells playing roles in the intestine itself compared to the draining lymph nodes. Perforin was an important effector molecule. These data have important implications for understanding adaptive immunity to an animal norovirus, representative of a genus that causes significant disease in humans.
HuNV infection and disease is rapid, with symptoms developing within 24–48 hours of infection and lasting for a few days. Thus, we selected three days after challenge as a readout for infection in our studies, since relevant vaccine-generated immune responses would have to act very early after challenge. Lack of any of the three components of the adaptive response: B cells, CD4 T cells, or CD8 T cells significantly diminished vaccine effects generated by either live virus or VP1 capsid protein immunization, and delayed viral clearance during primary infection. This indicates that VP1 has antibody epitopes as well as MHC H-2b restricted CD4 and CD8 T cell epitopes. These data suggest that it may be necessary to engage the concerted actions of an intact immune response including both MHC Class I and MHC Class II restricted T cells and antibody responses to efficiently vaccinate against HuNV infection.
The protection against MNV infection in aged mice in the absence of robust generation of anti-MNV antibodies raised the possibility that an important component of the vaccine response is T cell dependent, a hypothesis borne out in adoptive transfer studies. Importantly, the antiviral effector perforin is important in the clearance of MNV from the intestine, suggesting that the cytotoxic T cell response is a key effector mechanism. It is possible that other cell types such as NK cells might also use perforin as a mechanism to control MNV infection. Our data do not rule out a role for IFNγ in clearance of MNV infection since NK cells in recipient RAG1-/- mice can make IFNγ, but do suggest that T cell derived IFNγ plays at most a minor role in effector T cell function in the ileum. This argues that classical CTL assays may be a good surrogate for the development of effective vaccine-generated immune responses to HuNV.
Live virus vaccination was more effective than VRP based vaccination. The lower level of protection that we observed with ORF2 VRPs in contrast to MNV1.CW3 may be due to many factors, and this study does not provide mechanistic insights into this difference. In comparison to VLPs, VRPs may have advantages in systemic vaccination including targeting dendritic cells and intrinsic adjuvant activities [50]. These properties of VRPs may be responsible for the effectiveness of systemic single protein subunit vaccination against mucosal viral challenge in this case. However, it may be that because VRPs undergo a single round of replication at the site of inoculation they cannot generate the same breadth of immunity that is generated by live replicating virus. While VRP vaccination clearly induces some relevant effector and memory cell responses, vaccination with capsid alone may not sufficient to generate the complete antigenic repertoire required for effective immunity. Interestingly, we found some protection with the non-structural ORF1 polyprotein, suggesting that protective epitopes exist outside of the capsid protein. As the ORF1 polyprotein is expressed early after infection, it may be that these epitopes would be valuable targets for generating an efficient immune response.
Of note, vaccination with VP1 via the subcutaneous route provided significant protection despite the fact that the vaccination occurred systemically, while protection was read out at a mucosal site. This indicates that an active systemic immune response can provide protection against norovirus infection, and a mucosal vaccine may not be necessary to vaccinate against norovirus infection. Importantly, systemic vaccination was dependent on T cells, indicating that the relevant cells can traffic to the intestine after peripheral VRP-based vaccination.
These studies leave several important questions unanswered. Firstly, we used a homologous virus challenge. In nature, it is likely that hosts are repeatedly challenged with antigenically distinct noroviruses. However, the mouse norovirus strains identified so far fall into a single genogroup, GV, which likely represents a single serotype [9]. In this way murine noroviruses identified to date may present less of a challenge for the immune system than HuNV, which are distributed across 3 genogroups and appear to evolve under antibody selection [51]. In addition, we selected a strain of MNV that is cleared by WT mice. Other strains persist for prolonged periods of up to 35 days [9]. It remains to be determined whether vaccination will be effective against persistent MNV strains. It is interesting that human noroviruses can persist beyond the time frame of usual clinical symptoms [52]–[55]. Long-term persistence might contribute to explaining the sporadic epidemics of infection in the absence of an animal reservoir. Antigenic and pathogenetic complexity will likely be a major issue for the development of norovirus vaccines. The lack of comparable variation in MNV strains limits the utility of the MNV model for assessing immunity to antigenically distinct strains. Perhaps this limitation will be overcome as additional strains of MNV are identified, sequenced, and studied. However, the fact that we observed partial cross protection between MNV and one HuNV, and the demonstration that vaccination with many different VLPs can enhance generation of cross reactive antibodies [56] provide some encouragement.
There are two ways in which murine norovirus infection may not represent the same biology as HuNV infection. The first is the lack of diarrhea in mice infected with the strains of MNV used here. It is possible that the adaptive responses that clear MNV from the intestine demonstrated here are irrelevant to the responses that may prevent human disease. In this regard, it is important to note that studies of adult mice with rotaviruses (also an infection only model), have been important to our considerations of rotavirus vaccines [57]. Importantly, human studies may not reveal the mechanisms of effective immunity and are based on surrogate assays of immunity, since invasive sampling of tissues may be technically difficult. Studies in piglets may be revealing since piglets develop diarrhea when infected with the HuNV strain, HuNoV-HS66 [37]. However, it is more difficult to study immune mechanisms in this system. Thus, we are left with several imperfect systems for considering what one should seek in a HuNV vaccine. Our studies in mice argue for a vaccine that induces all aspects of the adaptive immune response, and that assays for cytotoxic lymphocyte responses to HuNV infection may be an important surrogate assay for protection.
The second aspect of murine norovirus infection that is of unknown relevance to human infection is the impressive capacity of MNV to infect lymph nodes draining the intestine (this paper and [31],[58],[59]). This may be related to the tropism of MNV for dendritic cells and macrophages [28],[59] and likely reflects spread of MNV directly from the intestine, but may also reflect seeding of the MLN from systemic sites. Considering the distal ileum alone, B cells and MHC Class I and β2M were not required for live virus vaccination, and there was significant, but incomplete, protection in MHC Class II-/- mice (Figure 3B and 3C). Consistent with this, studies of primary clearance showed that any single arm of the adaptive response was dispensable for ultimate control of primary infection in the intestine. However, vaccination-mediated control of infection in the MLN, and clearance of primary infection from the MLN [39], required B cells. This differential requirement for components of the immune response in different organs raises an important question about norovirus pathogenesis and lymphoid infection: are the cells infected in intestine and MLN the same? Differences in viral tropism in the two tissues might explain the differential requirement for B cells between ileum and MLN, indicating the importance of future studies on the role of immunity in norovirus cell and organ tropism.
MNV strains MNV1.CW3 or MNV1.CW1 were used in all virus infections [9],[28],[31]. Two mutations (that result in changes in the encoded amino acids) distinguish the genomes of MNV1.CW3 and MNV1.CW1 [28]. To generate a concentrated virus stock, RAW 264.7 cells (ATCC, Manassas, VA) were infected in VP-SFM media (Gibco, Carlsbad, CA) for 2 days at a multiplicity of infection (MOI) of 0.05. Supernatants were clarified by low-speed centrifugation for 20 min at 3,000 rpm. Virus was concentrated by centrifugation at 4°C for 3 h at 27,000 rpm (90,000 g) in a SW32 rotor. Viral pellets were resuspended in PBS and titered on RAW 264.7 cells as previously described [28]. Type I Lang reovirus was kindly provided by Dr. Terrence S. Dermody (Vanderbilt University, Nashville, TN). Plaque assays were performed as previously described [28] with the following modifications. Tissues were harvested into sterile, screw-top 2-ml tubes containing 500 µl of 1-mm zirconia/silica beads (BioSpec Products, Bartlesville, OK) and stored at −80°C. To obtain viral titers in these tissues 1 ml of complete DMEM was added to each sample on ice and homogenized using a MagNA Lyser (Roche Applied Science, Hague Road, IN) prior to plaque assay. The limit of detection was 20 plaque forming units (PFU)/ml.
All VRPs were produced as previously described [60]. Briefly, ORFs 1, 2 and 3 from MNV1.CW3 and ORF2 from Lordsdale virus and Chiba virus were each cloned into VRP expression vectors. Following infection of BHK cells with VRPs for 24 h, culture supernatants were harvested and cells lysed. Proteins were separated by SDS-PAGE and analyzed by western blot with polyclonal rabbit anti-MNV serum [28]. VRP titers and efficient expression of recombinant protein were determined by immunofluorescence assay using mouse antisera generated from inoculation with respective antigens. Cell lysates from MNV ORF2, Chiba virus and Lordsdale virus VRP-infected cell cultures were further purified to obtain VLPs [56].
RAW 264.7 cells were maintained as previously described [28]. Monoclonal antibodies (MAbs) specific to CD4 (YTS191.1 [61]), CD8 (H35 [62]) and SFR3-DR5 (ATCC HB-151 [63]) were produced from hybridoma cell lines in INTEGRA Celline CL1000 flasks (Integra Biosciences, Ijamsville, MD) using CD Hybridoma media (Gibco, Carlsbad, CA) as previously described [64].
All mice were bred and housed at Washington University School of Medicine or the University of North Carolina at Chapel Hill in accordance with all federal and university policies. Wild type C57BL6/J (hereafter referred to as WT, Jackson # 000664), B6RAG1-/- (RAG1-/-, Jackson # 002216), IFNγ-/- (IFNγ-/- , Jackson # 002287), perforin-/- (perforin-/-, Jackson # 002407), MHC Class II-/- (MHC Class II-/-, Jackson #003584) and B-cell-deficient mice backcrossed onto a C57BL/6 background (B cell-/-, Jackson # 002288) mice were purchased from Jackson Laboratory (Bar Harbor, ME). MHC Class II deficient mice (MHC Class II-/-, Taconic #ABBN12-M) and their WT controls C57BL/6Ntac (WT, B6 Taconic) were purchased from Taconic (Germantown, NY). Kb−/−×Db−/−×β2M-/- [43] (MHC Class I×β2M-/-) mice were a generous gift of Dr. Ted Hansen (Washington University, St Louis, MO). For some studies, WT C57BL6/J mice were purchased from Harlan Sprague Dawley (Indianapolis, IN) and aged to 14 months. All mice (or cage sentinel mice for mice deficient in antibody production) were tested by ELISA for the presence of MNV antibody prior to experiments [27]. All mice used in these studies were seronegative at the initiation of experiments.
Mice used in vaccination studies were immunized with 3×107 PFU of MNV1.CW1 [28], MNV1.CW3 [31], or control Type I Lang reovirus per orally (p.o.) in 25 µl of DMEM containing 10% fetal bovine serum (Hyclone, Logan, UT) (cDMEM). VRP immunizations were with 2.5×106 infectious units (IU) of each VRP expressing MNV1.CW3 ORF1, ORF2, or ORF3 individually or in groups of 2–3 VRPs; Chiba virus ORF2 or Lordsdale virus ORF2 in 10 µl or 50 µl volume by footpad inoculation (into the subcutaneous space) [65] on day 0 and boosted on day 21. HA VRP and PBS immunizations in 10 µl or 50 µl volume by footpad inoculation [65] on day 0 and boosted on day 21 served as controls for all VRP immunizations. Mice were challenged with 3×107 PFU of MNV1.CW3 at specified times after boost and tissues harvested three days post-challenge. Controls for VRP vaccination included PBS or VRPs expressing hemagglutinin (HA) from a mouse adapted influenza A virus [41]. PBS served as a control for HA VRP in these experiments in the event that HA VRP had a significant effect on MNV replication. HA VRP controls were not significantly different from PBS controls in all experiments and both are presented in all figures for completeness.
RAG1-/- and all splenocyte donor mice were infected with 3×106 PFU of MNV1.CW3 p.o. in 25 µl of cDMEM. All other mice were infected with 3×107 PFU MNV1.CW3 p.o. In RAG1-/- mice two segments of the small intestine were harvested: a one inch section of the small intestine immediately distal to the pylorus of the stomach, (designated the duodenum/jejunum), and a one inch section of the small intestine immediately proximal to the cecum (designated the distal ileum). In all other mice the distal ileum and three mesenteric lymph nodes (MLN) were harvested. With the exception of RAG1-/- mice (inoculated at 4–6 weeks of age) and aged WT mice (inoculated at 14 months of age), all other mice were inoculated at 6–10 weeks of age.
Spleens were harvested from mice and single cell suspensions were generated [65]. Cells were counted and diluted in RPMI-1640 media (Sigma, Saint Louis, MO) supplemented with 10% fetal calf serum (HyClone, Logan, UT), 100 U penicillin/ml, 100 µg/ml streptomycin, 10 mM HEPES (N-2-hydroxyethylpiperazine-N9-2-ethanesulfonic acid), 1mM sodium pyruvate, 50 µM 2-mercaptoethanol and 2 mM L-glutamine (cRPMI). In all adoptive transfer experiments, 1×107 cells were injected into persistently infected RAG1-/- mice by intraperitoneal (i.p.) injection in 0.5ml cRPMI.
For depletions in WT mice, 500 µg of lymphocyte-depleting antibody or an isotype-matched control antibody [SFR3-DR5, IgG2b] was administered i.p. one day prior and one day after infection. For depletions in adoptive transfer experiments, depleting antibodies were administered to RAG1-/- recipients as described above with one dose one day prior to splenocyte transfer and a second dose on the day of trnsfer. The efficacy of lymphocyte depletion in both sets of depletion experiments was monitored by flow cytometric analysis of splenocytes at the end of the experiment.
ELISA to detect binding of polyclonal anti-serum or fecal extract-derived antibody to purified MNV virions or MNV VLPs was performed as previously described [27],[56].
All data were analyzed using GraphPad Prism software (GraphPad Software, San Diego, CA). Viral titer data were analyzed with the nonparametric Mann-Whitney test. All differences not specifically stated to be significant were insignificant (p>0.05).
|
10.1371/journal.ppat.1000054 | SARS-Coronavirus Replication/Transcription Complexes Are Membrane-Protected and Need a Host Factor for Activity In Vitro | SARS-coronavirus (SARS-CoV) replication and transcription are mediated by a replication/transcription complex (RTC) of which virus-encoded, non-structural proteins (nsps) are the primary constituents. The 16 SARS-CoV nsps are produced by autoprocessing of two large precursor polyproteins. The RTC is believed to be associated with characteristic virus-induced double-membrane structures in the cytoplasm of SARS-CoV-infected cells. To investigate the link between these structures and viral RNA synthesis, and to dissect RTC organization and function, we isolated active RTCs from infected cells and used them to develop the first robust assay for their in vitro activity. The synthesis of genomic RNA and all eight subgenomic mRNAs was faithfully reproduced by the RTC in this in vitro system. Mainly positive-strand RNAs were synthesized and protein synthesis was not required for RTC activity in vitro. All RTC activity, enzymatic and putative membrane-spanning nsps, and viral RNA cosedimented with heavy membrane structures. Furthermore, the pelleted RTC required the addition of a cytoplasmic host factor for reconstitution of its in vitro activity. Newly synthesized subgenomic RNA appeared to be released, while genomic RNA remained predominantly associated with the RTC-containing fraction. RTC activity was destroyed by detergent treatment, suggesting an important role for membranes. The RTC appeared to be protected by membranes, as newly synthesized viral RNA and several replicase/transcriptase subunits were protease- and nuclease-resistant and became susceptible to degradation only upon addition of a non-ionic detergent. Our data establish a vital functional dependence of SARS-CoV RNA synthesis on virus-induced membrane structures.
| The SARS-coronavirus (SARS-CoV), which causes the life-threatening severe acute respiratory syndrome, replicates in the cytoplasm of infected host cells. A critical early step in the SARS-CoV life cycle is the formation of a replication/transcription complex (RTC) that drives viral genome replication and subgenomic mRNA synthesis. Virus-encoded enzymes form the core of this RTC, which is believed to be associated with characteristic virus-induced membrane structures derived from modified host cell membranes. To investigate the connection between these membrane structures and SARS-CoV RNA synthesis, and to characterize RTC composition and function, we isolated these complexes and developed the first in vitro assay to study their activity. SARS-CoV genomic RNA and all eight subgenomic mRNAs were synthesized in this in vitro reaction. By centrifugation, RTC activity could be isolated from the cytoplasm, together with membrane structures, viral enzymes, and RNA. The activity of these isolated RTCs was dependent on a cytoplasmic host factor. RTC activity was destroyed by detergent treatment, suggesting a critical role for membranes that appeared to protect the complex against protease and nuclease digestion. Our data establish a functional connection between viral RNA synthesis and intracellular membranes and show that host factors play a crucial role in SARS-CoV RNA synthesis.
| Following infection and genome translation, positive-strand RNA (+RNA) viruses establish a cytoplasmic enzyme complex that directs the amplification and expression of their genome. The viral RNA-dependent RNA polymerase (RdRp) is the central enzyme of this ‘replication/transcription complex’ (RTC), but it also may include other viral non-structural proteins (nsps) and host factors that cooperate to synthesize viral RNA. Over the past decade, it has become clear that +RNA virus RTCs are invariably associated with virus-induced membrane structures, which are poorly characterized but presumably provide a framework for RNA synthesis by facilitating the concentration and cooperation of viral macromolecules on a dedicated membrane surface. They may also protect the viral RNA from nucleases in the cytoplasm of the host cell, aid in shielding the double-stranded RNA intermediates of virus replication from the host cell's innate immune system, or contribute to the coordination of the viral life cycle in time and space. These membrane-bound RTCs are the molecular machines that drive the RNA synthesis and evolution of +RNA viruses. Clearly, unraveling their structure and function will be critical to understand the biochemistry of +RNA virus replication and develop novel antiviral control strategies.
The RTC of coronaviruses, including that of SARS-coronavirus (SARS-CoV), the causative agent of the life-threatening severe acute respiratory syndrome (for a review, see reference [1]), stands out for a number of reasons. First, at 27–32 kb, the polycistronic coronavirus genome is by far the largest genome among currently known RNA viruses [2]. Second, the viral RNA-synthesizing machinery not only amplifies the genome, but also directs the synthesis of a set of subgenomic (sg) mRNAs (eight in the case of SARS-CoV; RNA2 to RNA9) to express the viral accessory and structural protein genes. The latter are produced from a corresponding set of subgenome-length negative strand RNAs, which derive from discontinuous negative-strand RNA synthesis [3],[4]. Third, the viral replicase/transcriptase (which will be referred to as “replicase” for brevity) is of unprecedented size and complexity [5],[6]. It is produced by translation of the partly overlapping open reading frames (ORF) 1a and 1ab, with expression of the latter requiring a -1 ribosomal frameshift near the end of ORF1a. In this manner, SARS-CoV genome translation yields the large replicase polyproteins pp1a (4,382 aa) and pp1ab (7,073 aa). Extensive autoproteolytic processing, mediated by two ORF1a-encoded protease domains [7]–[10], ultimately generates 16 nsps [5],[6],[11],[12]. These include key replicative enzymes (e.g. the nsp12-RdRp [13], and the nsp13-helicase [14]), a variety of subunits containing presumed accessory functions for viral RNA synthesis (e.g. the nsp8-primase [15],[16], nsp14-exoribonuclease [17],[18], and nsp15-endoribonuclease NendoU [19]–[22]) and several predicted multi-spanning membrane proteins (nsp3, nsp4 and nsp6; [23],[24]) that presumably modify cellular endomembranes and target the RTC to this scaffold.
Immunofluorescence microscopy previously revealed that newly synthesized SARS-CoV RNA and several nsps colocalize in perinuclear foci in SARS-CoV-infected cells [8], [14], [25]–[27]. Electron microscopy established the presence of typical paired membranes, membrane whorls, and double-membrane vesicles (DMVs), which labeled for nsps [26]–[29] and viral RNA [27] and were therefore proposed to carry the SARS-CoV RTC. The endoplasmic reticulum (ER) was identified as the most likely membrane donor [26] and recent electron tomography studies indeed revealed a network of SARS-CoV-induced membrane structures that is continuous with this organelle (Knoops et al., in preparation). In the past four years, substantial progress has been made in the characterization of individual replicase subunits using enzymatic assays, reverse and classical genetics, bioinformatics and structural studies. However, the composition and mechanistics of the native ribonucleoprotein complexes, in which these different components interact to drive coronavirus replication and transcription, have remained completely uncharacterized thus far. We therefore set out to isolate active RTCs from SARS-CoV-infected cells and used those to develop an in vitro system that faithfully reproduced the synthesis of both genomic and sg RNAs, mainly of positive polarity. RTC activity cosedimented with newly synthesized viral RNA and several replicase subunits in a dense membrane fraction containing structures that could be labeled for nsp3 and nsp4. The in vitro activity of the pelleted RTC depended on the presence of a cytoplasmic host factor. Furthermore, RTC activity was destroyed by addition of (non-ionic) detergents, which also released replicase subunits and (mainly) sg RNA from the membrane fraction. Protease and nuclease protection experiments indicated that viral RNA and nsps were protected by membranes, thus further substantiating the functional connection between SARS-CoV RNA synthesis and virus-induced membrane structures that appear to be essential for RTC activity.
In order to characterize isolated SARS-CoV RTCs, we developed an in vitro RNA synthesis assay (IVRA) to study their activity in vitro. In this reaction, the incorporation of [α-32P]CTP into viral RNA was analyzed in a mixture containing NTPs, Mg2+, an ATP-regenerating system, and an inhibitor of cellular transcription (Actinomycin D). The RTC activity in cytoplasmic extracts prepared from SARS-CoV-infected Vero-E6 cells produced a set of 32P-labeled RNA molecules with sizes corresponding to those of the SARS-CoV genome and all eight sg RNAs (Fig. 1). These products were not detected when using mock-infected cell lysates (Fig. 1A, mock), demonstrating that SARS-CoV RTC activity was responsible for their synthesis. Reaction conditions were optimized by varying several parameters, including the composition of the reaction mixture, incubation time, temperature, and the Mg2+ concentration (Fig. 1 and data not shown). In a time course experiment, in vitro synthesized viral RNA accumulated up to 100 min into the reaction (Fig. 1A), after which a decrease was observed, probably due to declining RTC activity in combination with continued RNA degradation by cellular nucleases. The optimal reaction temperature was 30°C (Fig. 1B). RTC activity was strongly dependent on the Mg2+ concentration and was maximal when 2 mM of Mg2+ was added to the reaction (Fig. 1C). Manganese could not replace Mg2+, as IVRAs containing Mn2+ only yielded a ladder of small radiolabeled RNA molecules with aberrant sizes (Fig. 1D), suggesting an effect on RdRp processivity. Addition of ionic (SDS and deoxycholate (DOC)) or non-ionic detergents (Nonidet P40 (NP-40) and Triton X-100 (TX-100)) to the post-nuclear supernatant (PNS) prior to the IVRA completely abolished the accumulation of radiolabeled viral RNA, suggesting that the integrity of membranes is an important factor for SARS-CoV RTC activity (Fig. 1D).
To determine the polarity of the in vitro produced RNAs, the 32P-labeled products of an IVRA were hybridized to a membrane containing immobilized RNA probes specific for SARS-CoV positive- or negative-stranded RNA (Fig. 1E). A strong hybridization with the positive strand-specific probe was observed, demonstrating that the RTC mainly synthesized RNA of positive polarity in vitro. After longer exposure times, a small quantity of radiolabeled material hybridizing to the negative strand-specific probe became visible, but a similar signal was observed with the negative control RNA (Fig. 1E). This indicated that the quantity of in vitro synthesized negative-stranded RNA was very small (less than 2% of the total RNA), which is in line with the large excess of positive over negative strands that is commonly observed in vivo.
To assess whether protein synthesis occurred during IVRAs, we determined whether 35S-labeled amino acids were incorporated into proteins during a 100-min reaction. When using the PNS of uninfected cells, SDS-PAGE revealed a smear of 35S-labeled material (Fig. 2A, lane 2). These products were absent when the PNS was heated to 96°C for 5 min prior to the assay (Fig. 2A, lane 1), suggesting they resulted from translation under IVRA conditions. When using the PNS of SARS-CoV-infected cells, we observed incorporation of radiolabel also into a set of discrete polypeptides (Fig. 2A, lane 4), including species with sizes matching those of the SARS-CoV membrane (M) and nucleocapsid (N) proteins. This was likely due to the fact that the lysate contained large amounts of the sg mRNAs encoding these proteins, possibly in combination with the virus-induced shut-off of host cell translation [30]. Protein synthesis was completely inhibited when the translation inhibitors cycloheximide or puromycin were present during the IVRA (Fig. 2A, lanes 5 and 6), but this did not affect in vitro RTC activity since the quantity of radiolabeled RNA products was unchanged (Fig. 2B).
To further characterize the active RTC, the PNS of SARS-CoV-infected cells was subjected to differential centrifugation. A 10,000×g supernatant fraction (S10) showed no RTC activity (Fig. 3, lane 2), but only a trace amount of the original activity was recovered in the 10,000×g pellet fraction P10 (Fig. 3, lane 4). Surprisingly, RTC activity in this P10 fraction could be largely restored by adding an aliquot of the cytoplasmic S10 fraction (Fig. 3, lane 5). An S10 fraction prepared from mock-infected cells was equally capable of restoring the RTC activity in P10, indicating that a cytoplasmic host factor was required (Fig. 3, lane 6). Routinely, about 50% of the RTC activity that was originally present in the PNS could be recovered in the P10 fraction (in assays supplemented with S10). Remarkably, virtually all replicative activity was lost, while transcription was only 2- to 3-fold decreased, in the P10 fraction depleted of the host factor (Fig. 3, lane 4). The sedimentation of the RTC activity at 10,000×g suggests that it is associated with heavy membrane structures.
The P10 fraction, which contained the bulk of RTC activity, was analyzed by electron microscopy (negative staining) in combination with an immunogold labeling for the (putative) transmembrane proteins nsp3, nsp4, and nsp6 (Fig. 4 and data not shown). Clusters of vesicles (with diameters between 100 and 350 nm) were observed, which appeared to be associated with more tubular or flattened membrane structures. A strong immunolabeling of these structures for SARS-CoV nsp3 (Fig. 4A) and nsp4 (Fig. 4B) was observed. Membrane structures immunoreactive for nsp3 (Fig. 4C) or nsp4 (data not shown) were not detected in a control P10 fraction prepared from mock-infected cells. Occasionally, double membranes could be distinguished (Fig. 4B, arrow). These observations are consistent with the notion that the P10 fraction is enriched for SARS-CoV-induced nsp-containing membrane structures that have been documented in infected cells.
The distribution of newly synthesized SARS-CoV RNAs between the RTC-containing P10 and cytoplasmic S10 fractions was analyzed by fractionation of PNS after an IVRA (Fig. 5A). The bulk (76%) of newly made genomic RNA was recovered from the P10 fraction, suggesting it remained associated with the heavy membrane structures. In contrast, newly synthesized sg RNAs were, depending on their size, progressively more abundant in S10, suggesting their release from the RTC. To further investigate the role of membranes in RNA localization, an IVRA was performed with PNS, after which 0.5% TX-100 was added and the distribution of viral RNAs between P10 and S10 was analyzed (Fig. 5B). The bulk of the smaller RNA species (RNA5-9) was now recovered from the S10 fraction. In contrast, one-half of the genomic RNA remained associated with the P10 fraction after detergent treatment, suggesting product-specific differences in RTC operation and organization, which appears to include partly detergent-resistant structures.
For selected nsps, for which suitable antisera that are reactive in Western blot experiments were available, the distribution between the cytoplasmic S10 fraction and the RTC-containing P10 fraction was analyzed. This revealed that these RTC subunits were enriched or mainly present in the P10 fraction (Fig. 6). The bulk of nsp3 was in the P10 fraction and nsp5 was found almost exclusively in the P10 fraction (Fig. 6, lane 3). Most of nsp8 was detected in the P10 fraction although also a substantial amount was found in the cytoplasmic fraction (Fig. 6, lane 2 & 3). Treatment of PNS with 0.5% TX-100 prior to P10-S10 fractionation, led to the redistribution of nsp3, nsp5, and nsp8, which were no longer found in the P10 fraction, but were recovered at increased levels in the S10 fraction (Fig. 6, lanes 4 & 5). This suggests that their direct or indirect association with membranes caused them to cosediment with the RTC activity in the P10 fraction.
To further assess the role of membranes in SARS-CoV RNA synthesis, it was investigated whether they protect the RTC. A standard 100-min IVRA was performed, followed by treatment with the nuclease Bal31, a non-specific nuclease that degrades both single- and double-stranded RNA, in the presence or absence of 0.5% TX-100. After fractionation of the samples into P10 and S10, the quantity of in vitro synthesized radiolabeled RNA in each fraction was analyzed (Fig. 7A). In untreated control samples, newly made viral RNA was found both associated with the RTC in the P10 fraction (predominantly genomic RNA) as well as released in the cytoplasmic S10 fraction (enriched in sg RNA; Fig 7A, lane 1 & 2). The newly made viral RNA in the cytoplasm was completely degraded upon nuclease treatment (Fig 7A, lane 3), while the RNA associated with the RTC was protected (Fig 7A, lane 4). The latter products only became susceptible to nuclease treatment upon addition of 0.5% TX-100, suggesting that the replicating RNA is enclosed by membranes (Fig 7A, compare lanes 4 and 6).
To determine whether also replicase subunits were protected by membranes, PNS was treated with proteinase K, either in the absence or presence of 0.5% TX-100. Protease-treated samples and untreated controls were subsequently fractionated into P10 and S10, after which the presence of nsp3, nsp5, and nsp8 was probed by Western blotting (Fig. 7B). Both cytoplasmic nsp3 in S10 and pelleted nsp3 in P10 were susceptible to protease treatment (Fig. 7B, top panel). The nsp5 subunit, which mainly cosedimented with the RTC in P10, was largely resistant to protease treatment (Fig 7B, middle panel, lane 5), but it was degraded in the presence of TX-100. The observed protease-resistance of nsp5 is not due to a lack of proteinase K activity, since both nsp3 and a host protein cross-reacting with the nsp5 antiserum were completely degraded in this same sample. Likewise, nsp8 in the P10 fraction was resistant to protease treatment, and, surprisingly, this was also true for the nsp8 that was present in the S10 fraction (Fig. 7B, lower panel). Both forms of nsp8 were susceptible to the protease in the presence of a non-ionic detergent (TX-100). These data suggest that nsp5 and nsp8 were enclosed by membranes. In agreement with the membrane topology predictions for the nsp3 domains used to raise our antiserum [5],[23], a major part of nsp3 was exposed on the surface of these membrane structures.
The SARS-CoV RTC, like the RTCs of other +RNA viruses [31]–[33], is believed to be associated with virus-induced structures derived from intracellular membranes. The coronavirus RTC is composed of an unusually large number of subunits, including several nsps with unique enzyme functions [2],[34]. Despite steady progress, the functional characterization of the 16 SARS-CoV nsps, including the RdRp and helicase enzymes that are central to replication, is still in an early stage. To investigate the details of the molecular interplay between these subunits, the viral RNA template, and host factors, in vitro assays for viral RNA synthesis will be indispensable. By now, the soluble expression and purification of several individual coronavirus nsps has proven to be problematic. In combination with the membrane-associated nature of the complex, this suggests that the reconstitution of the RTC from its purified components, remains a distant perspective. As a complementary approach, we therefore set out to isolate the active SARS-CoV RTC from the only currently available source: virus-infected cells. The newly developed IVRA described in this paper (Fig. 1) will allow us to obtain more insight into the architecture and function of the SARS-CoV RTC as a whole, and may aid to address the poorly defined role of cellular membranes.
Although RdRp activity in cell lysates was previously reported for the coronaviruses mouse hepatitis virus and transmissible gastroenteritis virus [35]–[40], this is to our knowledge the first description of a robust in vitro system for coronavirus RNA synthesis that produces the full spectrum of viral mRNAs (both genomic and sg RNAs) generated in infected cells. A similar in vitro system was recently developed for the distantly related arterivirus equine arteritis virus (manuscript in preparation), suggesting that our method may be generally applicable to nidovirus RTCs.
Protein synthesis occurred in our lysates under the IVRA conditions used, but its inhibition did not affect in vitro RTC activity (Fig. 2). This suggests that, in contrast to what was described for cells infected with mouse hepatitis virus [41]–[43] or SARS-CoV (our unpublished data), continued translation is not required for RTC activity in vitro. Likely, inhibition of protein synthesis does not influence the activity of the preformed, active RTCs present in our PNS, which are mainly synthesizing RNA of positive polarity (Fig. 1E).
Currently, suitable small SARS-CoV RNA replicons, which could be added to an IVRA as exogenous template and be distinguished from natural viral RNAs on the basis of size, are not available. Consequently, addressing the question whether de novo initiation of RNA synthesis occurs in our system must wait until further technical advances (in this area) have been made. Still, a potential complication may be the inability of such exogenous templates to enter the membrane-protected RTC, as also observed in this study for molecules like Bal31 nuclease and proteinase K (Fig. 7).
SARS-CoV RTC activity was recovered in a 10,000×g heavy membrane pellet (P10), but the isolated RTCs had to be supplemented with an S10 fraction from infected or uninfected cells to regain activity (Fig. 3). This indicates that, besides the host factors possibly associated with the RTC in the P10 fraction, also a cytoplasmic host factor is required for SARS-CoV RNA synthesis. The nature of this host factor is currently being analyzed. Replication appeared to be particularly dependent on the presence of this host factor. While transcription was only 2- to 3-fold reduced, replication was barely detectable in the P10 fraction depleted of the host factor (Fig. 3). Whether this difference is merely due to the larger size of the genomic RNA and/or reflects a higher demand or specific role for the host factor in replication remains to be investigated in more detail. The RTC activity cosedimented with newly synthesized viral RNA, several replicase subunits and nsp3-, nsp4- and nsp6-containing membrane structures. The latter proteins are (putative) multi-spanning transmembrane proteins [23],[24],[44] presumed to be important in the induction of the RTC-related membrane rearrangements that accompany SARS-CoV infection [26],[27]. Furthermore, the cosedimentation of nsp3 with the RTC (Fig. 6) may indicate that one or multiple of the enzymatic activities of this multidomain protein [5] are important for RNA synthesis. All of the nsp5 main proteinase copurified with RTC activity in the P10 fraction, although it remains to be investigated whether this finding is directly related to the site of replicase polyprotein processing. The RTC's core enzyme, the nsp12-RdRp, has been postulated to work in concert with a unique second RdRp activity that was recently identified in nsp8. Its proposed RNA primase activity [15],[45] would be consistent with the (partial) cosedimentation of nsp8 with the RTC that we observed in this study (Fig. 6). After TX-100 treatment, nsp3, nsp5, and nsp8 no longer cosedimented in P10, suggesting they had been released from the membrane structures. In addition, protease protection experiments in the absence and presence of detergents revealed that nsp5 and (part of) nsp8 were shielded by membranes, while the predicted cytoplasmic domains of nsp3 were not [23]. The experiments in Fig. 7 indicated that a cytoplasmic form of nsp8 (in S10) was also shielded from protease activity by membranes. This suggests the existence of membrane structures distinct from the RTC-containing complexes in P10. The cosedimentation of nsp3, nsp5, and nsp8 with RTC activity is in line with their colocalization in specific structures in the perinuclear region of SARS-CoV-infected cells, as observed by immunofluorescence microscopy [25],[26].
Free RNA of transmissible gastroenteritis virus was previously found to be susceptible to nuclease treatment, whereas most negative-stranded RNAs, and a small fraction of (probably nascent) positive-stranded RNAs, were present in membrane-protected complexes [46]. In our study, non-ionic detergents rendered SARS-CoV RNA susceptible to nuclease digestion (Fig. 7) and destroyed all RTC activity (Fig. 1). This again signifies the importance of intact membrane structures for viral RNA synthesis. Their disruption may have dissociated the active enzyme complex and/or changed the RTC's microenvironment, or may have provided access to cytoplasmic nucleases.
The bulk of newly synthesized SARS-CoV genome remained associated with the RTC-containing heavy membrane structures, while sg RNAs appeared to be more readily released from the structures in which they had been synthesized. In previous studies with transmissible gastroenteritis virus, it was also found that preferentially sg RNAs were no longer associated with the membrane-associated complexes [46]. The released RNA molecules might represent a pool of mRNAs destined for translation into structural and accessory proteins (sg RNAs) and additional replicase proteins (RNA1), while the RTC-associated RNAs might be engaged in replication and/or packaging. In this manner, the intracellular compartmentalization mediated by the formation of specialized membrane structures might also serve to coordinate different steps in the viral life cycle and/or enhance their specificity for viral RNA. Surprisingly, after treatment with 0.5% TX-100, a large fraction of genomic RNA remained in the 10,000×g pellet, suggesting it is associated with detergent-resistant structures. This might indicate that, as postulated for hepatitis C virus replication complexes [47]–[49], the SARS-CoV RTC is associated with lipid rafts or lipid droplets, a feature that could also explain the proposed role of lipid rafts during the early stages of SARS-CoV replication [50].
If SARS-CoV RTCs, as this study suggests, are enclosed by membranes that may provide an optimal environment for viral RNA synthesis, this raises the question of how newly synthesized RNA products are released from these structures. Moreover, the fact that RTC activity depends on a cytoplasmic host factor that does not cosediment with the complex is an additional indication that crosstalk between cytoplasm and RTC-containing membrane structures must occur, e.g. via channels that may facilitate transport across membranes. Taken together, our data support the existence of a functional link between SARS-CoV RNA synthesis and the unusual membrane structures induced upon coronavirus infection.
Vero-E6 cells were infected with SARS-CoV (strain Frankfurt 1) at a multiplicity of infection of 5 as described previously [26]. All procedures involving live SARS-CoV were performed in the biosafety level 3 facility at Leiden University Medical Center. Rabbit antisera recognizing nsp3, nsp5, and nsp8 were described previously [26]. Antisera against nsp4 and nsp6 were raised in New Zealand White rabbits using as antigens the bovine serum albumin-coupled synthetic peptides FSNSGADVLYQPPQTSITSAVLQ and LNIKLLGIGGKPCIKVATVQ, representing the C-terminal sequences of nsp4 and nsp6, respectively.
SARS-CoV- or mock-infected cells (eight 175 cm2 flasks) were harvested by trypsinization at 10 hours post infection. To inhibit cellular transcription, 2 µg/ml actinomycin D was present in all solutions used for harvesting and washing of the cells. After washing with PBS, cells were resuspended in 2 ml ice-cold hypotonic buffer (20 mM HEPES, 10 mM KCl, 1.5 mM MgOAc2, 1 mM DTT, 133 U/ml RNaseOUT (Invitrogen) and 2 µg/ml actinomycin D, pH 7.4) and incubated for 10 min at 4°C. Cells were disrupted using a Dounce homogenizer by giving 30 strokes with a tight fitting pestle. Isotonic conditions were restored by adding HEPES, sucrose, and DTT, which resulted in a final lysate containing 35 mM HEPES, pH 7.4, 250 mM sucrose, 8 mM KCl, 2.5 mM DTT, 1 mM MgOAc2, 2 µg/ml actinomycin D, and 130 U/ml RNaseOUT. Nuclei, large debris, and any remaining intact cells were removed by two successive 5-min centrifugations at 1,000×g, and the resulting PNS was either assayed immediately for RTC activity or stored at −80°C. The SARS-CoV titer present in PNS was approximately 108 plaque-forming units per ml. Plaque assays performed before and after IVRAs revealed that no measurable de novo virus production occurred during this assay (data not shown). A 10,000×g pellet (P10) and supernatant (S10) fraction were prepared from PNS by centrifugation at 10,000×g for 10 min. The pellet was resuspended in dilution buffer (35 mM HEPES, 250 mM sucrose, 8 mM KCl, 2.5 mM DTT, 1 mM MgOAc2, pH 7.4), in 1/10 of the original PNS volume from which the pellet had been prepared. In some experiments, PNS was incubated for 15 min at 4°C with 0.5% TX-100 prior to the preparation of P10 and S10 fractions.
Assays were performed using either 25 µl PNS, 20 µl S10, 5 µl P10, or 5 µl P10 supplemented with 20 µl S10. When required, the total volume was adjusted to 25 µl with dilution buffer. The subsequent addition of reaction components yielded a 28 µl final reaction volume, containing 30 mM HEPES pH 7.4, 220 mM sucrose, 7 mM KCl, 2.5 mM DTT, 2 mM MgOAc2, 2 µg/ml actinomycin D, 25 U RNaseOUT, 20 mM creatine phosphate (Sigma), 10 U/ml creatine phosphokinase (Sigma), 1 mM ATP, 0.25 mM GTP, 0.25 mM UTP, 0.6 µM CTP and 0.12 µM and 10 µCi [α-32P]CTP (GE Healthcare). Unless otherwise indicated, IVRAs were performed for 100 min at 30°C. Reactions were terminated by adding 60 µl of a mixture containing 5% lithium dodecyl sulfate, 0.1 M Tris-HCl, pH 8.0, 0.5 M LiCl, 10 mM EDTA, 5 mM DTT, and 0.1 mg/ml proteinase K, and incubating at 37°C for 10 min. When protein synthesis was tested, [α-32P]CTP was replaced with 14.3 µCi of Promix (GE Healthcare), containing a mixture of [35S]methionine and [35S]cysteine. To assess the effect of translation inhibition, 70 µg/ml of cycloheximide or 350 µg/ml of puromycin were added.
RNA was isolated from IVRA reaction mixtures by acid phenol extraction and isopropanol precipitation. Reaction products were analyzed by denaturing formaldehyde agarose gel electrophoresis essentially as described previously, except that a 1% agarose gel was used [51]. Radiolabeled in vitro synthesized RNA was detected by exposing a PhosphorImager screen directly to the dried gel, after which screens were scanned with a Personal Molecular Imager FX (Bio-Rad) and data were analyzed with Quantity One version 4.5.1 (Bio-Rad). Unlabeled endogenous SARS-CoV RNA was detected by hybridization with a 32P-labeled oligonucleotide SARSV002 (5′-CACATGGGGATAGCACTAC-3′), which is complementary to a sequence present in the 3′-end of all SARS-CoV RNAs [5].
In vitro transcribed RNAs (0.75 µg) corresponding to nt 29,364-29,727 of the 3′-terminal region (3′-TR(+)) or complementary to nt 1-378 (3′-TR(−)) of the SARS-CoV genome were immobilized to Hybond N+ membrane (GE Healthcare). As negative controls, RNAs corresponding to nt 12,313–12,660 (ctrl. a) of the equine arteritis virus genome or its complementary sequence (ctrl. b) were included. The membrane with the immobilized probes was prehybridized for 4 hours in a hybridization mixture containing 5×SSPE (750 mM NaCl, 50 mM NaH2PO4, 5 mM EDTA, pH 7.0), 0.05% SDS, 5x Denhardt and 100 µg/ml homomix I. Subsequently, the membrane was hybridized with half of the 32P-labeled RNA recovered from a 28-µl IVRA in 0.8 ml hybridization mix, which was first heat denatured at 70°C for 15 min. After hybridization for 16 h at 56°C, membranes were washed twice for 20 min at 56°C with 4 ml of 5x SSPE, 0.05% SDS, and the hybridization signal was quantified by PhosphorImager analysis.
Proteins were separated by SDS-PAGE and transferred to Hybond-P PVDF membrane (GE Healthcare) by semi-dry blotting. After blocking with 1% casein in PBS containing 0.1% Tween-20 (PBST), membranes were incubated with anti-nsp3, anti-nsp5 or anti-nsp8 rabbit antisera, diluted 1∶2000 in PBST with 0.5% casein and 0.1% BSA. Peroxidase-conjugated swine anti-rabbit IgG antibody (DAKO) and the ECL-plus kit (GE Healthcare) were used for detection.
Protease protection experiments were done by incubating PNS (50 µl) for 10 min at 20°C with 20 µg/ml of proteinase K either in the absence or presence of 0.5% TX-100. After inactivation of the protease by addition of 2 mM PMSF and fractionation into a 10,000×g pellet (P10) and supernatant (S10), samples were analyzed by Western blotting. For nuclease protection assays, a standard 100-min IVRA was performed with the PNS, after which 5U of Bal31 nuclease was added, either in the presence or in the absence of 0.5% TX-100. After a 10-min incubation, samples were fractionated into S10 and P10 fractions. Radiolabeled RNA was isolated from the fractions and analyzed as described above.
One volume of 6% paraformaldehyde in 60 mM PIPES, 25 mM HEPES, 2 mM MgCl2, 10 mM EGTA, pH 6.9 was added to P10 fractions. Formvar-coated grids were placed on 10-µl drops of these fixed P10 fractions and incubated at room temperature for 1 min. After blocking with 1% BSA in PBS, grids were incubated for 30 min with rabbit antisera directed against nsp3, nsp4 or nsp6 (1∶200) in PBS containing 1% BSA. Bound rabbit IgG was detected with protein A carrying 15-nm gold particles. After negative staining with 2% phosphotungstic acid, grids were viewed in a FEI T12 transmission electron microscope at 120 kV. |
10.1371/journal.ppat.1000299 | Dengue Virus Type 2 Infections of Aedes aegypti Are Modulated by the Mosquito's RNA Interference Pathway | A number of studies have shown that both innate and adaptive immune defense mechanisms greatly influence the course of human dengue virus (DENV) infections, but little is known about the innate immune response of the mosquito vector Aedes aegypti to arbovirus infection. We present evidence here that a major component of the mosquito innate immune response, RNA interference (RNAi), is an important modulator of mosquito infections. The RNAi response is triggered by double-stranded RNA (dsRNA), which occurs in the cytoplasm as a result of positive-sense RNA virus infection, leading to production of small interfering RNAs (siRNAs). These siRNAs are instrumental in degradation of viral mRNA with sequence homology to the dsRNA trigger and thereby inhibition of virus replication. We show that although dengue virus type 2 (DENV2) infection of Ae. aegypti cultured cells and oral infection of adult mosquitoes generated dsRNA and production of DENV2-specific siRNAs, virus replication and release of infectious virus persisted, suggesting viral circumvention of RNAi. We also show that DENV2 does not completely evade RNAi, since impairing the pathway by silencing expression of dcr2, r2d2, or ago2, genes encoding important sensor and effector proteins in the RNAi pathway, increased virus replication in the vector and decreased the extrinsic incubation period required for virus transmission. Our findings indicate a major role for RNAi as a determinant of DENV transmission by Ae. aegypti.
| Dengue viruses, globally the most prevalent arboviruses, are transmitted to humans by persistently infected Aedes aegypti mosquitoes. Understanding the mechanisms mosquitoes use to modulate infections by these agents of serious human diseases should give us critical insights into virus–vector interactions leading to transmission. RNA interference (RNAi) is an innate defense mechanism used by invertebrates to inhibit RNA virus infections; however, little is known about the antiviral role of RNAi in mosquitoes. RNAi is triggered by double-stranded RNA, leading to degradation of RNA with sequence homology to the dsRNA trigger. We show that dengue virus type 2 (DENV2) infection of Ae. aegypti by the natural route generates dsRNA and DENV2-specific small interfering RNAs, hallmarks of the RNAi response; nevertheless, persistent infection of mosquitoes occurs, suggesting that DENV2 circumvents RNAi. We also show that DENV2 infection is modulated by RNAi, since impairment by silencing expression of genes encoding important sensor and effector proteins in the RNAi pathway increases virus replication in the vector and decreases the incubation period before virus transmission. Our findings indicate a significant role for RNAi in determining the mosquito vector's potential for transmitting human diseases.
| Dengue virus serotypes 1–4 (DENV1-4; Flavivirus; Flaviviridae) are medically important, positive-sense RNA viruses transmitted to humans by Aedes aegypti mosquitoes during epidemic outbreaks [1],[2]. DEN fever and DEN hemorrhagic fever are major public health burdens in many parts of the world [3]; however, although DENVs can cause severe disease in humans, mosquito infections are non-pathogenic and persistent. We hypothesize that the difference in infection outcomes results from host defense (immune) responses. Ae. aegypti is an important vector because it feeds almost exclusively on humans and is well adapted to life in tropical urban environments [4]. We have only a rudimentary understanding of DENV molecular interactions with Ae. aegypti vectors, including the mosquito's innate defense pathways against arboviruses. DENVs infect the mosquito midgut following ingestion of a viremic blood meal from an acutely infected human, replicate, disseminate to the salivary glands where they are further amplified, and emerge into saliva at the time of transmission. Approximately 10 to 14 days are required for the extrinsic incubation period (EIP), the time between initial infection of the mosquito and transmission [5]. The recent release of the Ae. aegypti genome sequence [6] provides an important tool to begin understanding critical virus-vector interactions during the EIP. Identification of mosquito genes that are orthologs of genes known to be part of innate immune pathways in Drosophila [7],[8] is an important step in characterizing mosquito defense mechanisms and makes it possible to manipulate putative antiviral pathways during virus infection. Xi et al [8] have recently shown that the Ae. aegypti Toll pathway, which is also implicated in Drosophila defense against certain viruses, has a role in controlling DENV replication after establishment of a persistent infection.
Recent studies with Drosophila clearly show that RNA interference (RNAi) is a potent innate antiviral pathway that is presumably triggered by dsRNA formed in virus-infected cells and leads to degradation of the RNA virus genome. Several groups have shown that RNAi can inhibit infection of Drosophila with RNA viruses from the Dicistroviridae, Nodaviridae, and Togaviridae families [9]–[11]. Mutant Drosophila lacking functional key RNAi pathway genes such as dcr2 or ago2 are highly susceptible to some RNA virus infections [9]–[11]. In Drosophila, dcr2 encodes the RNAi sensor protein Dicer-2 (Dcr2) that recognizes and cleaves long dsRNA, producing 21–25 bp short interfering RNAs (siRNAs) [12],[13]. siRNAs are duplexes with 3′ overhangs of 2 nucleotides and 5′ phosphate and 3′ hydroxyl ends [14]. With the assistance of Dcr2 and the protein R2D2, one strand of siRNA is incorporated into a nuclease complex called the RNA-induced silencing complex (RISC), to start the effector phase of the pathway [15]–[18]. The siRNA strand associated with RISC acts as a guide sequence and anneals to target RNA having sequence complementarity (identity with one strand of the dsRNA trigger) [17],[19],[20]. The Argonaute-2 (Ago2) protein in RISC has sequence specific “slicer” activity and cleaves the target RNA, leading to its degradation [21],[22].
In silico searches using sequences of RNAi pathway genes from Drosophila show a number of putative RNAi gene orthologs in Anopheles gambiae and Ae. aegypti genome databases including dcr2, r2d2 and ago2. We have carried out limited functional assays to confirm the role of these genes in RNAi [23]–[27]. We have demonstrated that Ago2 expression is essential in mosquitoes for modulation of alphavirus (Togaviridae) infection of An. gambiae and Ae. aegypti [25],[27]. We also have demonstrated that RNAi that inhibits the replication of DENV can be triggered in the midgut of transgenic mosquitoes by expression of an inverted repeat RNA (IR-RNA or dsRNA) derived from a portion of the DENV genome prior to or at the same time as virus challenge in an infectious blood-meal [23]. The Carb 77 transgenic line of mosquitoes expressing DENV2-specific siRNAs generated from the trigger sequence failed to accumulate viral genomic RNA, and poorly disseminated and transmitted DENV2. Significantly, injection of dsRNA derived from either Ae. aegypti ago2 or dcr2 sequences prior to ingestion of an infectious blood-meal reversed the resistance phenotype [23]; (Franz and Olson, unpublished data). These studies demonstrated that the RNAi pathway is functional in the mosquito midgut if artificially/exogenously triggered. However, during natural, persistent infections of DENV2 competent mosquitoes, viral RNA is readily detected in midguts from 2–3 days post infection (dpi) to 14 days dpi [5],[28], suggesting that RNAi is not totally effective in preventing replication. An obvious viral strategy to circumvent RNAi is to encode a suppressor of RNAi and a number of examples of suppressors have been described in plant and insect RNA viruses [29],[30], although to date no RNAi suppressors have been associated with arboviruses. Alternatively, DENVs may simply evade RNAi by sequestering their dsRNA replicative forms within virus-induced double membrane structures associated with the endoplasmic reticulum, preventing dsRNA recognition by an RNAi sensor [31],[32]. A focused study of DENV2 infection of competent Ae. aegypti should reveal what role RNAi plays in this natural vector-virus interaction. In this paper, we show that persistent DENV2 infections of Ae. aegypti cells and mosquitoes generate dsRNA triggers and small DENV2-specific RNAs consistent in size and sequence with siRNAs. Impairment of the RNAi pathway increases the titer of infectious virus in the vector and shortens the EIP, which could increase transmission efficiency.
We initially used cultured Ae. aegypti (Aag2) cells to examine RNAi activity resulting from DENV infection. Triplicate flasks of confluent Aag2 cells were infected at a multiplicity of infection (MOI) of 0.01 with DENV2 (Jamaica 1409, American-Asian genotype) and cell culture medium was collected daily for titration of infectious virus. Virus titer increased daily and peaked at 7 days post infection (dpi) at 6.5×105±2.3×105 pfu/ml (Fig. 1A). DENV2 genomic RNA was detected in cells infected at a MOI of 0.005 by northern blot analysis at relatively high levels by 3 dpi and continuing to at least 7 dpi (Fig. 1B). To determine if the infected cells contained dsRNA, the RNAi trigger, we used the dsRNA-specific J2 antibody [33] in an immunofluorescence assay (IFA) at 3 dpi. Cells that had been mock infected displayed no fluorescence (Fig.1C.a), but strong signal was detected in the cytoplasm of infected cells (Fig. 1C.b), presumably at sites associated with viral replication. We looked for siRNA production by northern blot analysis using total RNA isolated from Aag2 cells. Small RNAs approximately 21 nt in length were detected with both genome-sense (Fig. 1D.a) and antisense (Fig 1D.b) DENV2 prM gene probes beginning at 3 dpi and continuing to 7 dpi. We carried out limited cloning and sequencing of small RNA from Aag2 cells to verify the presence of DENV2 genome sequences and detected a putative siRNA derived from the positive-sense strand of the prM gene (nt 714–733; agtggcactcgttccacatg). Several groups of overlapping 18–22 nt siRNAs appeared to be derived from genome “hot spots”; for example, a total of 9 putative siRNAs were from the nt 4659–5396 region of the NS3 gene, 5 from the positive strand and 4 from the negative strand. Three of these formed an overlapping cluster at nt 4962–4995 (J. Scott and J. Wilusz, unpublished results). The most intense hybridization signal came from the negative sense probe, suggesting that the majority of small RNAs were of the same polarity as genomic RNA.
Extensive spatial and temporal analyses were performed to understand the DENV2 infection pattern in our model Higgs white-eye (HWE) Rexville D strain of Ae. aegypti mosquitoes following ingestion of an infectious artificial blood meal containing approximately 107 pfu/ml of DENV2. Landmark events in the infection pattern of Ae. aegypti HWE mosquitoes are summarized in Figure 2. Infectious virus titers followed a similar pattern as described previously using the same virus strain and various Ae. aegypti strains [5],[34] (Fig. 2A). Virus titers increased in the midgut beginning at 3–5 dpi, peaked at 10 dpi (9×103±3.6×103 pfu/ml), and began to decline by 12 dpi (7.4×102±2.9×102 pfu/ml). For HWE mosquitoes, DENV2 envelope (E) antigen was readily detected by IFA in the midgut at 7 and 14 dpi (Fig. 2B.a; Fig. 2B.b), virus dissemination to fat body was detected at about 7 dpi (Fig. 2B.c), and antigen was found abundantly in salivary glands at 14 dpi (Fig. 2B.d). Viral E antigen and virus titer declined in the midgut by 14 dpi but increased in the salivary glands from 14–21 dpi (9.4×104±3.6×104 pfu/ml virus titer in carcass at 16 dpi). Viral RNA could be detected from 5 to 14 dpi by northern blot analysis of total RNA from midguts (Figure 2C). Analysis of remaining tissues (carcass) showed that viral RNA could be detected outside the midgut from 9 to 14 dpi, when the experiment was terminated. In previous studies we have shown that by allowing infected mosquitoes to probe an artificial feeding membrane, we can most consistently detect DENV2 in saliva of HWE mosquitoes at 14 dpi [23].
The J2 antibody was used to detect dsRNA in midguts, the first mosquito tissue infected after acquisition of a DENV2-containing blood meal. Midguts from mosquitoes that received a blood meal with no virus showed no fluorescent signal (Fig. 3A.a); however, midguts from mosquitoes that had ingested the virus in a blood meal 7 days earlier (Fig. 3A.b) produced a signal reminiscent of DENV2 infection patterns seen by IFA for viral E antigen at the same time point (Fig. 2B.a; [5]. This indicated that DENV2 infection of mosquito tissues also generates the RNAi trigger.
To detect virus-specific siRNAs in infected midguts, total RNA was enriched for small RNAs as previously described [23]. An antisense RNA oligonucleotide from the prM gene that had been previously characterized by cloning and sequencing siRNA was used as an RNase protection probe to detect DENV RNA-derived siRNA in infected midguts [23]. We detected 22 nt RNA at 8 and 14 dpi (Figure 3B). To further verify the existence of siRNAs in mosquitoes following infection, we performed northern blot analysis on total mosquito RNA extracted 14 days after DENV2 infection using an antisense RNA prM probe and detected viral-specific RNAs 22–24 nt in size (Fig. 3C). Thus, in both infected mosquitoes (Fig. 2C and 3B and C) and cultured cells (Fig. 1B and 1D) we demonstrated that viral genomic RNA accumulated concurrently with increasing levels of virus-specific small RNAs, suggesting that the virus employs a mechanism to circumvent the RNAi defense.
We and other investigators have demonstrated that it is possible to transiently silence (or impair) expression of key components of the RNAi pathway by RNAi knock-down [23]–[25],[29]. We found that injection of dsRNAs derived from RNAi pathway genes into Carb77 transgenic mosquitoes caused impairment of antiviral RNAi, reversing the DENV2-resistant phenotype [23]. To determine if modulation of the RNAi pathway in normal mosquito vectors had an effect on virus replication we injected dsRNA to knock-down expression of key pathway components Aa-dcr2, Aa-r2d2, and Aa-ago2. Primers targeting specific regions of each mRNA and used to generate dsRNAs are listed in Table 1. We intrathoracically injected groups of 200 HWE mosquitoes with 500 ng of dsRNA derived from Aa-dcr2, Aa-r2d2, Aa-ago2, E. coli βgal or with PBS. Five mosquitoes from each group were analyzed by northern blot at 2 days after injection to demonstrate silencing of Aa-ago2 mRNA with dsRNA.ago2 (Fig 4A.a), Aa-r2d2 mRNA with dsRNA.r2d2, (Fig. 4A.b), and Aa-dcr2 mRNA with dsRNA.dcr2 (Fig.4A.c) at the time of oral infection with DENV2, which is most crucial in innate defense. Injection of dsRNA.βGAL slightly enhanced, in the cases of Aa-r2d2 and Aa-dcr2, and slightly reduced, in the case of Aa-ago2, but did not silence any of the RNAi component mRNAs. Although normal levels of mRNA were detectable in some mosquitoes after injection of dsRNA.ago2, injection of dsRNA.dcr2 and dsRNA.r2d2 caused almost complete silencing of cognate mRNA in all mosquitoes tested. To assess the effect of silencing RNAi pathway genes on DENV2 replication, dsRNA-injected mosquitoes were allowed to recover for 2 days, then given an infectious blood meal. Mosquitoes were assayed for infectious virus 7 days later (Fig 4B). The titers in both groups of control mosquitoes, non-injected and injected with dsRNA.βGAL (1.6×103±7×102 pfu/ml and 2.7×103±6.5×102 pfu/ml, respectively) were similar, as were the percentages of mosquitoes infected [59% (30/51), and 52% (36/71), respectively]. In mosquitoes injected with dsRNA.ago2 the titer (1.2×104±6.5×103 pfu/ml) and the number of mosquitoes infected (47%; 36/68) was similar to the non-injected group. In contrast, in mosquitoes injected with dsRNA.r2d2 the infectious virus titer was higher (7.0×103±1.5×103 pfu/ml; P<0.05) than in the non-injected group; however, the percentage of mosquitoes infected (55%; 44/80) was not different from the non-injected group. The most profound effect on DENV2 infection was observed in mosquitoes injected with dsRNA.dcr2 where both the infectious titer (2.5×104±1.0×104 pfu/ml) and the proportion of mosquitoes infected (75%; 63/84) were higher (P<0.05) than the non-injected group.
DENV2 genomic RNA in whole mosquitoes was examined by northern blot analysis 16 days after injection with dsRNA and 14 days after virus challenge. In general, we observed a slight increase in levels of viral RNA in mosquitoes injected with dsRNA.ago2 and dsRNA.r2d2 as compared to control mosquitoes (Fig. 4C). A greater than 4-fold increase in DENV2 genomic RNA signal over controls was seen in mosquitoes injected with dsRNA.dcr2, suggesting that impairment of RNAi early in infection leads to more robust replication of viral RNA in the mosquito.
To determine if transient impairment of RNAi pathway gene expression affected the length of the extrinsic incubation period (EIP) of DENV2 in Ae. aegypti HWE mosquitoes we injected groups of 200 mosquitoes with dsRNA.ago2, dsRNA.r2d2, dsRNA.dcr2, dsRNA.βGAL, or PBS and included one non-injected group. Two days later all mosquito groups received an infectious blood meal. Mosquitoes from each group that were engorged with blood were placed into cartons (10 mosquitoes/carton). At 7, 10, and 12 days after infection, mosquitoes in four cartons from each group were allowed to probe an artificial feeding solution for collection of saliva. The feeding solutions were assayed for infectious virus and their mean titers are shown in Figure 5A–C. Virus titers in individual mosquitoes were also determined (data not shown) and indicated that all mosquitoes were infected; therefore lack of virus titer in the feeding solutions was not due to absence of infection.
Impairment of certain components of the RNAi pathway by injection of dsRNA resulted in significant effects (P<0.05) on time of appearance and titer of infectious virus in mosquito saliva. On days 10 and 12 post infection (Fig 5B and 5C), the infectious virus titers in feeding solutions from at least one of the mosquito groups with impaired RNAi (dsRNA.dcr2-, dsRNA.ago2-, and dsRNA.r2d2-injected) were higher (P<0.05) than in control groups (dsRNA.βGAL-injected, PBS-injected, and non-injected). At day 7, the titers in feeding solutions were not significantly different (Fig. 5A). At day 10, feeding solutions collected from dsRNA.dcr2-injected mosquitoes had significantly higher titers (P<0.05) than control groups (dsRNA.βGAL-injected, PBS-injected, and non-injected) (Fig. 5B). At day 12 the feeding solution titers from dsRNA.dcr2- and dsRNA.ago2-injected mosquitoes were significantly higher (P<0.05) than from all control groups, and dsRNA.r2d2-injected mosquito saliva titer was higher (P<0.05) than the non-injected group (Fig. 5C). Analysis by Chi-square test revealed that dsRNA.dcr2-injected mosquitoes had a significantly higher (P = 0.053) number of infected feeding solutions compared to the non-injected group at 7 dpi (Fig. 5A). At 10 dpi dsRNA.r2d2- and dsRNA.dcr2-injected groups had increased proportions (P<0.05) of feeding solutions infected compared to the non-injected group (Fig. 5B). At 12 dpi, dsRNA.ago2-, dsRNA.r2d2- and dsRNA.dcr2-injected groups had higher proportions (P<0.05) of feeding solutions infected (Fig. 5C).
We previously have identified a number of putative RNAi genes in An. gambiae and Ae. aegypti genome databases, including dcr2, r2d2 and ago2, and by limited functional assays have confirmed the role of dcr2 and ago2 in RNAi [23]–[27]; (Franz and Olson, unpublished data). We also have demonstrated that the RNAi pathway in Ae. aegypti inhibits the replication of alphaviruses and flaviviruses if first triggered by introducing dsRNA derived from virus genome sequences [23],[26],[27]. We show here that RNAi is implemented naturally after acquisition of DENV2 by mosquitoes via the normal, oral route of infection. DENV2 replication generated virus-specific small RNAs, consistent in size with siRNAs, in mosquito cell cultures as well as tissues of female mosquitoes. However, despite the presence of siRNAs in their tissues, DENV2-competent mosquitoes accumulated viral genome RNA and infectious virus in midguts and other tissues and transmitted virus in saliva. We found that RNAi-mediated knock-down of Aa-r2d2, as well as Aa-ago2 and Aa-dcr2, resulted in increased virus replication and shortened EIP in Ae. aegypti.
Cultured Ae. aegypti cells became persistently infected and released increasing titers of infectious virus until at least 10 dpi. DENV2-competent mosquitoes accumulated viral RNA and infectious virus in the midgut up to 10 dpi. The decline in virus titer and E antigen in the midgut after 10 dpi may have been a result of clearance of virus by defense mechanisms mediated by RNAi as well as the Toll pathway [8] or due to the nutritional status of the mosquito, since these mosquitoes were not given multiple blood meals; however, we found that the DENV2 infection persisted in tissues outside the midgut, including salivary glands, to at least 21 dpi, which is consistent with previous findings [5].
DENV2 infections generated substantial amounts of dsRNA in infected Aag2 cells and midguts of Ae. aegypti mosquitoes. Others have shown by IFA that dsRNA generated in DENV2-infected mammalian cells co-localizes with non-structural viral proteins known to be part of the replication complex within double-membrane vesicles [35]. The fact that we detected siRNA-like RNAs by northern blot hybridization using both sense and antisense probes suggests that dsRNAs associated with DENV2 replication are vulnerable to cleavage by Dcr-2 protein. This was particularly true in Aag2 cells, where we detected antisense siRNAs that undoubtedly resulted from processing of dsRNA that was a part of a replication intermediate. However, in contrast to mosquito cells where RNAi is induced by expression of a dsRNA transcript and siRNA polarity appears to be symmetrical [23],[26], in infected Aag2 cells we observed a preponderance of DENV2-specific siRNAs arising from the genome strand, suggesting that intramolecular secondary structure of viral genomic RNA may account for much of the dsRNA cleaved by Dcr2 to produce siRNAs. Others have shown similar asymmetry in siRNA polarity and have hypothesized that secondary structure of genomic RNA for some positive strand RNA viruses of plants and animals contributes substantially to siRNA formation [36],[37]. The sequences of the DENV2-specific siRNAs will be analyzed to accurately describe the population of 20–26 nt RNAs in DENV2 infected mosquito cells.
To show that the RNAi pathway affects virus replication in a competent DENV2 vector, we transiently silenced key RNAi pathway genes and examined resulting changes in virus titer. Previously, we have shown that injection of dsRNA derived from An. gambiae ago2 gene sequence transiently impaired the RNAi pathway and significantly increased replication of O'nyong-nyong virus (genus, Alphavirus; family Togaviridae) in the vector [25]. We suggested that a robust RNAi response normally prevented widespread infection by ONNV in An. gambiae. Injection of dsRNAs targeting Aa-dcr2 and Aa-ago2 mRNAs reversed RNAi-based DENV2 resistance in the Carb77 transgenic line [23] (Franz and Olson, unpublished data). In the current study, we intrathoracically injected HWE mosquitoes with dsRNA.dcr2, dsRNA.ago2 or dsRNA.r2d2 prior to DENV2 infection and showed reduced corresponding mRNA 2 days later, at the time of infection. We observed that reduced expression of key RNAi genes resulted in increased mean virus titers in mosquitoes at 7 dpi. Injection of dsRNA.dcr2 was most effective for mRNA knock-down, and targeting Aa-dcr2 by dsRNA.dcr2 injection caused the largest increase in infection rate and virus titers. The average titer for mosquitoes injected with dsRNA.dcr2 was >10-fold higher than the average titer of control dsRNA-treated mosquitoes and a small but significant number of dsRNA.dcr2- and dsRNA.ago2-injected mosquitoes had virus titers 100-fold higher than in control infections. If expression of RNAi component genes and activity of the RNAi pathway is variable among vector populations this may explain in part observed differences in vector competence [38]. We are currently determining whether there are genetic associations between RNAi genes such as Aa-dcr2 and vector competence [39].
Finally, given the increase in DENV2 titers in dsRNA.dcr2-, dsRNA.r2d2- and dsRNA.ago2-treated mosquitoes, we asked whether the EIP is affected in mosquitoes with an impaired RNAi pathway. In mosquitoes pre-treated with dsRNA.dcr2, DENV2 transmission was detected as early as 7 dpi, indicating a shorter EIP when this defense mechanism was diminished. This could have significant epidemiological implications for DENV2 transmission, affecting vectorial capacity, which includes vector competence, but also incorporates all the other biological, behavioral, and ecological attributes of the vector that function in virus transmission. If the adult female mosquito has a shorter EIP and transmits virus more quickly after infection, the mosquito will have more opportunities to transmit virus during its lifetime.
An important function of RNAi in the mosquito may be to place limits on RNA virus replication and in the absence of RNAi, DENV2 and other arboviruses may increase their amplification to levels that begin to compromise vector fitness. This situation has been observed in Drosophila with dcr2 or ago2 loss-of-function mutations that are highly vulnerable to RNA virus infections due to disabled RNAi responses [10]. Although we believe that the early innate immune response is important in limiting virus replication in mosquitoes, injection of dsRNA results in a transient impairment of the RNAi defense system and longer-term suppression induced by a genomic mutation might cause a more pronounced and effective response. Future work should lead toward understanding natural variations in expression of RNAi pathway genes in vector populations and resulting effects on DENV infection and transmission. We are currently developing null-mutants for dcr2 in Ae. aegypti to more fully understand the role of this RNAi pathway gene in vector competence and vectorial capacity, and determine the possible effect of RNAi on infection of the vector by various DENV2 genotypes.
Variations in interaction of DENVs with the RNAi pathway could also have implications for determining why certain genotypes of DENVs are more efficiently transmitted than other genotypes and help explain the rapid spread of potentially virulent genotypes in the Americas [40],[41]. We hypothesize that DENV2 genotypes that most effectively circumvent RNAi would more efficiently replicate and be transmitted. An important question is how do arboviruses like DENV2 evade RNAi? Pathogenic RNA viruses of insects such as drosophila C virus express RNAi suppressors that increase viral pathogenicity, thus facilitating virus transmission to other insects [10],[42]. However, for efficient transmission of arboviruses to vertebrate hosts during acquisition of blood meals by the mosquito, it may be advantageous for the virus not to harm the vector. Indeed, the balance between RNAi activity in the vector and viral circumvention mechanisms might be a determinant in maintenance of the persistent virus infection. The La Crosse virus (LACV; Bunyaviridae) nonstructural protein NSs was reported to suppress RNAi in mammalian cells [43], but other studies have demonstrated that in mosquito cells persistently infected with recombinant LACV, expression of functional NSs protein has no effect on specific RNAi responses [44]. Another mechanism arboviruses may use to evade RNAi is minimizing early detection of the dsRNA trigger by the RNAi machinery. Geiss et al. [31] showed a significant reduction in WNV RNA in mammalian cells that were pre-treated with virus-specific siRNAs; however, cells that were treated subsequent to the establishment of viral replication did not have the same reduction in viral mRNA, suggesting that replicating viral RNA may be sequestered from the RNAi machinery in the cell. Identifying possible mechanisms of RNAi suppression or evasion associated with DENV2 infections is now being pursued.
Others have recently suggested that West Nile virus (WNV; Flaviviridae) fails to trigger RNAi in cultured C6/36 (Ae. albopictus) mosquito cells; however, the RNAi response in Drosophila appears to be protective against WNV infection [45]. We show here that the RNAi pathway in Ae. aegypti is not completely protective against DENV2 infection but modulates replication, suggesting that DENV2 has co-adapted to the major vector's RNAi response in ways as yet undefined. We hypothesize that RNAi is critical to maintaining a persistent virus infection in the vector, leading to long-term survival of the infected mosquito and efficient transmission of the virus with each successive blood meal. Important goals of future studies will be to determine how the virus mitigates the effects of RNAi in the vector, whether RNAi interacts with other innate defenses such as Toll pathway responses, and the consequences of impaired RNAi on vector fitness following infection with arboviruses.
LLC-MK2 monkey kidney cells and C6/36 (Ae. albopictus) cells were cultured in modified Eagle's medium (MEM) supplemented with 8% fetal bovine serum, L-glutamine, non-essential amino acids and penicillin/streptomycin and maintained at 37°C and 28°C, respectively. Aag2 (Ae. aegypti) cells were cultured in Schneider's Drosophila medium with 10% fetal bovine serum, L-glutamine, non-essential amino acids and penicillin/streptomycin and maintained at 28°C. High passage DENV2 (Jamaica 1409) was used to infect fresh cultures of C6/36 cells to prepare infectious blood meals. Briefly, monolayers of C6/36 cells were inoculated with DENV2 at a MOI of 0.01 and held at 28°C; 6 days later medium was replaced and infected cells and medium were harvested at 12–14 days.
LLC-MK2 cells were grown to confluent monolayers in 24-well plates, infected with 10-fold serial dilutions of virus for 1 hour, then overlaid with an agarose-nutrient mixture. After 7 days incubation at 37°C cells were stained with 5 mg/ml MTT (3-[4,5-dimethylthiazol-2-yl]-2,5-diphenyltetrazolium bromide) solution and incubated for 4 hours [46],[47]. Viral titers were determined by counting plaques. Individual mosquitoes were triturated in 1.0 ml of L15 medium to release infectious virus as previously described [23]. Individual mosquito titers are reported as plaque forming units (pfu) per ml (values are expressed as the means±SEM).
Ae. aegypti HWE mosquito eggs were hatched from an egg liner (containing 10,000–100,000 eggs) in 150 ml of deionized, autoclaved water. Larvae were transferred to a large rearing pan, collected as pupae 7–9 days later, transferred to an emergence container within a cage and maintained in the insectary at 28°C, 82% relative humidity until adult mosquitoes were harvested. Groups of 200 one-week-old adult females were placed in 2.5 liter cartons, deprived of sugar source overnight and allowed to feed on artificial blood meals consisting of virus-infected C6/36 cell suspension (60% vol/vol), 40% (vol/vol) defibrinated sheep blood (Colorado Serum Co., Boulder, CO) and 1 mM ATP. Virus titers were usually 106–107 pfu/ml for DENV2 [48]. The artificial blood meal was prewarmed to 37°C and then pipetted into water-jacketed glass feeders covered with a hog gut membrane and maintained at a constant temperature of 37°C. Feeders were placed onto the net covering the cartons to allow females to feed through the hog gut membrane for 1 h. Fed females were selected, put into new cartons, provided with water and sugar and maintained in the insectary for analysis.
Viral RNA and mRNA originating from endogenous mosquito genes were analyzed by northern blot hybridization performed as described [49]. Briefly, total RNA was extracted from mock or DENV2 infected Aag2 cells (MOI = 0.05) or whole mosquitoes using Trizol reagent (Invitrogen) following the manufacturer's instructions; 3–5 µg of total RNA were electrophoresed in a 1.2% agarose gel and blotted onto a positively charged nylon membrane (Ambion). Blots were hybridized with antisense 32P-UTP-labeled RNA probes that were transcribed in vitro from linearized pBluescript II SK (Stratagene) containing a cDNA insert derived from the specific sequence of DENV-2 (prM), Aa-dcr2, Aa-r2d2 or Aa-ago2 RNA., and hybridization was visualized using a phosphorimager. Alternatively, random-primed 32P-dCTP-labeled DNA probes were generated from the same template by using the Megaprime DNA Labeling Kit (Amersham Pharmacia Biosciences).
Approximately 50 µg of total RNA obtained from mock or DENV2 infected Aag2 cells (Fig 1D) or 50 µg of total RNA from 60 mosquitoes (Fig 3C) were loaded per lane onto a 15% polyacrylamide-TBE-urea denaturing gel and separated by electrophoresis. RNA was electrophoretically transferred to a non-charged nylon membrane and chemically cross-linked to the membrane using carbodiimide [50]. The membrane was incubated in UltraHyb (Ambion) hybridization buffer at 42°C for 30 minutes. For detecting siRNAs from infected Aag2 cells or mosquitoes, single-stranded RNA probes were transcribed in sense or antisense orientation from the 498 bp cDNA encoding the DENV2 prM gene using the MEGAscript kit (Ambion), with approximately 9% of the UTP in the transcription reaction conjugated to biotin for detecting siRNA from Aag2 or with 32P-UTP for detecting siRNA from mosquitoes. Five micrograms of labeled sense or antisense RNA were hydrolyzed to 50–100 nt fragments in 200 mM carbonate buffer at 60°C for approximately 2.5 hours. Probe was added to the UltraHyb buffer after the pre-hybridization and incubated for 16 hr at 42°C. The membrane was washed twice with 2× SSC-0.1% SDS for 5 minutes each, followed by two washes in 0.1× SSC-0.1% SDS for 15 minutes each. Biotin-labeled probes were detected with the BrightStar BioDetect Kit (Ambion) and membranes were exposed to X-ray film for 18 hours (Fig. 1D), and 32P-labeled probes were detected using a phosphorimager (Fig. 3C).
siRNAs were enriched from 200 µg total midgut RNA extracted from female mosquitoes 1–14 days post bloodmeal according to a previously described protocol [23]. High molecular-weight RNA was precipitated by addition of 5% (wt/vol) polyethylene glycol (Mr 8,000) and 0.5 M NaCl. Supernatants containing siRNAs were analyzed in a ribonuclease protection assay by using the mirVana miRNA Detection Kit (Ambion). For nuclease protection, siRNAs were hybridized with an antisense RNA probe 30 nt in length containing 22 nt of sequence complementary to a portion of the prM-encoding region of DENV2 RNA (Integrated DNA Technologies, Colorado State University). RNA probes were end-labeled with [γ-32P] ATP. Hybridizations were performed at 42°C. After RNase digestion, hybridized RNA samples were electrophoretically separated on a 16% polyacrylamide gel containing 7 M urea.
Aag2 cells were infected with DENV2 at a MOI of 0.1 and Ae. aegypti mosquitoes were given a non-infectious or infectious blood meal containing ∼106 pfu/ml of virus. Both cells and midguts were fixed in 4% paraformaldehyde in PBS at 4–5 dpi and 7 dpi, respectively. dsRNA was detected using J2 antibody (Scicons, Hungary) as previously described with minor modifications [33]. All images were taken with a Leica DM 4500B microscope.
Oligonucleotide primers incorporating T7 RNA polymerase promoter sequences at the 5′ ends were designed to amplify ∼500-bp regions of RNAi pathway genes (Table 1) to serve as transcription templates for dsRNAs for Aa-dcr2, Aa-r2d2 and Aa-ago2 according to the method previously described [25]. One week old adult female mosquitoes were intrathoracically injected with 500 ng of each dsRNA prior to infection.
Groups of 200 Ae. aegypti HWE mosquitoes were intrathoracically injected with dsRNA.r2d2, dsRNA.ago2, dsRNA.dcr2, dsRNA.βGAL (500 ng dsRNA/mosquito), or PBS. A separate group of mosquitoes was not injected. Two days later all mosquito groups were challenged with an infectious blood meal containing 6.6×106 pfu/ml of DENV2. At pre-determined times, groups of 10 mosquitoes were allowed to probe and feed on 350 µl of a feeding solution (50% FBS/164 mM NaCl/100 mM NaHCO3/0.2 mM ATP/∼50 µg sucrose/phenol red, pH 7) that was placed between two parafilm membranes stretched over a glass feeder. After probing, mosquitoes and feeding solutions were collected and prepared for plaque assays [23].
To determine if impairment of the RNAi pathway in mosquitoes has an effect on virus replication and EIP, data were subjected to analysis of variance (ANOVA) using the general linear model of SAS (SAS User's Guide, Cary, NC: Statistical Analysis System Institute, Inc., 1987). Sources of variation for virus titer in mosquitoes were treatment (non-injected, dsRNA.βGAL-, dsRNA.Aa.ago2-, dsRNA.Aa.r2d2-, and dsRNA.Aa.dcr2-injected) and replicate (4); and for feeding solution titer were treatment (non-injected, PBS-, dsRNA.βGAL-, dsRNA.Aa.ago2-, dsRNA.Aa.r2d2-, and dsRNA.Aa.dcr2-injected), replicate (2), days post-infection (7, 10, and 12), and the effect of treatment by day post infection. Since variation between replicates was not significant, it was removed from the model. If variances were not homogeneous, data were subjected to log 10 transformations. When differences among treatment means were detected, they were separated using least significant difference (LSD) procedure. The effect of treatments on the number of mosquitoes infected and on the number of feeding solutions infected was analyzed by chi-square test (SAS). For the number of mosquitoes infected each experimental group was compared to the non-injected group; for the number of feeding solutions infected each experimental group was compared to the average of the three control groups.
|
10.1371/journal.pcbi.1006084 | Toward a theory of coactivation patterns in excitable neural networks | The relationship between the structural connectivity (SC) and functional connectivity (FC) of neural systems is of central importance in brain network science. It is an open question, however, how the SC-FC relationship depends on specific topological features of brain networks or the models used for describing neural dynamics. Using a basic but general model of discrete excitable units that follow a susceptible—excited—refractory activity cycle (SER model), we here analyze how the network activity patterns underlying functional connectivity are shaped by the characteristic topological features of the network. We develop an analytical framework for describing the contribution of essential topological elements, such as common inputs and pacemakers, to the coactivation of nodes, and demonstrate the validity of the approach by comparison of the analytical predictions with numerical simulations of various exemplar networks. The present analytic framework may serve as an initial step for the mechanistic understanding of the contributions of brain network topology to brain dynamics.
| Functional connectivity, as reflected in the statistical dependencies of distributed activity, is widely used to probe the organization of complex systems such as the brain. While this measure has been helpful for characterizing brain states and highlighting alterations of brain dynamics in various diseases, the mechanisms underlying the generation of FC patterns remain poorly understood. One prominent factor shaping FC is the underlying neural network structure. Using a minimalist model of excitation, we investigate how the topology of excitable neural networks contributes to FC. Specifically, we show that FC can be analytically predicted from the way in which the nodes are embedded in the network and how they are related to basic self-organizing units of excitable dynamics, particularly, short pacemaker cycles. These insights are a step towards a mechanistic understanding of the activation patterns of complex neural networks.
| The network perspective has become a powerful and central approach for representing and analyzing complex biological systems, spanning from the study of interacting genes to neuronal assemblies [1]. Particularly for brain networks, a fundamental challenge is to understand the relationship between the network organization (or topology) of the structural connectivity and network activity or dynamics, as reflected by the network’s functional connectivity [2].
For more than a decade, increasingly sophisticated computational neuroscience approaches have been used to model the activity patterns of brain networks based on their characteristic network connectivity [3–7]. These approaches have used a variety of highly diverse node models, ranging from detailed biophysical models, eg, [8], neural mass models summarizing the properties of neuronal populations [9], to more abstract and phenomenological models, such as the Kuramoto model [10, 11], as well as stylized simple discrete models of neural excitation [12]. Intriguingly, the choice of the specific computational model does not seem to crucially affect the resulting global patterns of functional connectivity, as highly diverse models applied on the same network may result in very similar fits of the empirical FC [13, 14]. This finding suggests that the crucial aspect in producing functional connectivity may not be the specific local models, but the characteristic topology of the underlying SC. Ubiquitous topological features of brain networks are, for instance, modules (formed by nodes that are more frequently connected among each other than to the rest of the network) and hubs (central nodes that have more connections than average network nodes) [15, 16].
Our present goal is to systematically explore key topological contributions to SC-FC relationships, and develop a mechanistic understanding of the involved elementary processes: How do excitable dynamics translate specific topological patterns into systematic coactivations of nodes or functional connectivity? While previous studies have demonstrated that network topology matters in determining the network activity patterns, they have not yet provided a universal framework of the SC-FC relationship. Several studies have used sets of intuitive topological rules, including the topological overlap, paths distribution, as well as branching and convergence of projections, for predicting FC [17–19]. While these predictions were comparable to more sophisticated computational models in successfully predicting the empirical FC, the predictions were still far from perfect, leaving room for further improvements as well as an analytic understanding of the SC-FC relationship.
As we strive for a mechanistic understanding and an analytical description of how SC-FC relationships change with network architecture, we study this question in a minimal, deterministic model of excitable dynamics. Subsequent investigations will then analyze the relevance of our findings for more general dynamical regimes. Here, we used the stylized dynamics of the discrete excitable SER model [20]. The letters S-E-R denote the basic node behavior of susceptible (S) nodes becoming excited (E) by excited neighbors, then refractory (R), before turning once again susceptible, in discrete time steps. What the SER model offers is a detailed mechanistic understanding of how topology regulates particular contributions to the coactivation matrix and in this way determines FC. This minimal model of an excitable system has a rich history in many disciplines, ranging from the propagation of forest-fires [21, 22], the spread of epidemics [23, 24], to neuronal dynamics [25, 26]. Using a similar model setup we have shown that the distribution pattern of excitations is regulated by the connectivity as well as by the rate of spontaneous excitations [27]. An increase in each of these two quantities leads to a sudden increase in the excitation density accompanied by a drastic change in the distribution pattern from a collective, synchronous firing of a large number of nodes in the graph to more local, long-lasting and propagating excitation patterns. More recently, and focusing particularly on the topology of excitable networks, we have shown that SC-FC relationships strongly and non-trivially depend on neural network topology [28–30]. Qualitatively speaking, modularity is associated with high positive correlations between SC and FC, while a broad degree distribution, or low network density, may lead to negative SC-FC correlations. The mechanism behind this finding is as follows. High connectivity is associated with an elevated excitation density. Locally high connectivity (that is, within a module) results in a statistically higher number of excitations among nodes within the same module and, as a consequence, systematically higher coactivations of nodes in the same module. This positive contribution to the correlation between SC and FC tends to ‘overwrite’ the typically negative correlations arising from suppressed coactivations of linked nodes due to sequential excitation [28]. Moreover, in [31] a first evidence for a dependence of dynamical features on the initial conditions and the potential relevance of triangles was provided. The discrete SER model, thus, provides us with clear hypotheses of how SC is translated into FC.
In the present work, we present a computational framework for analytically predicting patterns of functional connectivity from excitable dynamics running on an arbitrary network architecture. Our theoretical framework should be capable of addressing two main questions: (1) Why do different network topologies show systematically different levels of SC-FC relationships? (2) How do specific subsets of initial conditions suppress or enhance different topological features and hence lead to systematically altered coactivation matrices and, as a consequence, alterations in SC-FC relationships? After illustrating the operation of the SER model with some introductory examples, we present potential analytical predictions with increasing realism. Subsequently, we explore to which extent these predictions can reproduce simulated patterns of FC across different topological network configurations.
The topological structure of a network is represented by the adjacency matrix A, where Aij = 1, if nodes i and j are connected, 0 otherwise. Here we consider only undirected networks. In a simplifying approach, we consider functional connectivity to be the coactivation of nodes derived from excitation patterns of the deterministic SER model on a given graph. In the following investigations, T denotes the probability of having a node in the T state in the initial conditions (or put differently, the proportion of nodes in the T state, at first), T = Pr(node state = T) with T ∈ {S, E, R}. The predictive value of the different analytical proposals, including SC itself, was investigated by computing the (Pearson) correlation as well as the mean signed difference between the simulated and predicted FCs. For illustrative purposes, Fig 1 shows a set of functional connectivity patterns resulting from the SER dynamics starting at different initial conditions. Strikingly, FC is not only shaped by the network topology, but also strongly by the initial conditions of the SER model. For example, negative SC-FC correlations as observed in scale-free graphs can turn into positive correlations, depending on the initial conditions (Fig 1). Throughout the manuscript, unless otherwise stated, we explore the effect of initial conditions such that E is varied between 0 and 1, while R = S = (1 − E)/2. See Materials and Methods section for further details.
Let us consider a pair of nodes (i, j). The most elementary contribution to coactivation is that a common neighbor jointly activates the two nodes i and j. This first simple consideration already points to the fact that the number of common neighbors of two nodes could predict functional connectivity. Thus, we obtain a first prediction for the coactivation of nodes i and j, denoted here TO:
T O i j = | N i j | = ∑ k A i k A j k , (1)
where N i j denotes the set of common neighbors of the pair (i, j). This quantity is better known by its normalized version, the matching index or topological overlap [32]. However, this quantity is not able to predict the potential effect of the initial conditions, as no dependence on the probabilities of states is incorporated. In the following section we describe how the exact pattern of triangles around a pair of nodes (i, j) can generate specific patterns of coactivations. Furthermore, we show that a striking dependence of coactivations on the statistics of initial conditions is instigated through the probability of initializing a triangle as a pacemaker.
Typically in the SER simulations, starting from random initial conditions, after a short transient all nodes settle into a period-3 oscillation (except for very sparse graphs). The dynamics, thus, typically partition the nodes of a graph into three cohorts: those jointly active at time t, at time t + 1 and at time t + 2, respectively. The contribution of a single run to the coactivation matrix is, therefore, almost completely determined by the initial conditions, rather than by the network architecture. When accumulating information over a large number of runs, the pattern of coactivations becomes governed by topological features of the graph. In this deterministic setting of the SER model, the drivers of the dynamics are autonomously oscillating triangles serving as pacemakers [28]. A pacemaker corresponds to an isolated triangle initialized with any permutation of the three states S, E and R, displaying stable oscillations that cannot be disrupted by noise or surrounding influences when embedded in larger networks. It is a decisive advantage of this deterministic cellular automaton model that it allows for enumerating the full state space of such network motifs.
As an illustration of this general approach, we first consider a small toy network model consisting of the pair of nodes under consideration, a number of common neighbors and independent triangles around each of these common neighbors (Fig 2A). We assume that, after a short transient, the system settles into a period-3 oscillation (which corresponds to assuming that the graph has a large enough number of triangles and the initial conditions contain non-zero numbers of nodes in the S, E and R states). For each initial condition, the pattern of coactivations in the deterministic SER model is then essentially a consequence of the two different possible triangle usages: active pacemakers (i.e., triangles initialized as some permutation of S, E, R and, thus, producing cyclic activity) and passive elements driven by other pacemakers (i.e., triangles initialized in any other way, which autonomously would settle into an all-susceptible state, but are typically excited by excitations coming from other parts of the network). Our hypothesis is that, in order to simultaneously activate two nodes i and j, they must (1) be not part of an active pacemaker and (2) have at least one of their common neighbors part of an active pacemaker.
The derivation of the predictions is inspired by mean-field approaches classically employed in epidemic models [33]. The dynamics of the nodes is characterized at the population level per state, which considerably reduces the dimensionality of the system. For example, the probability of finding a pair of nodes in the SR state in the initial conditions simply reduces to the product of the marginal probabilities, that is, SR. In the following section we formulate a first prediction for the case of non-overlapping triangles (among nodes i and j, as well as among their common neighbors). The probability that the pair (i, j) and one of their common neighbors form an active pacemaker is:
△ i j = 2 A i j [ S R ( 1 - ( 1 - E ) | N i j | ) + S E ( 1 - ( 1 - R ) | N i j | ) + R E ( 1 - ( 1 - S ) | N i j | ) ] , (2)
The set of common neighbors of nodes i and j is characterized by the number of triangles around each of these neighbors (not including the one formed with nodes i and j). Let cijk be the number of triangles around the kth common neighbor of the pair (i, j), this quantity corresponds to the unnormalized clustering coefficient of k minus one, if (i, j) are connected
c i j k = C k d k ( d k - 1 ) / 2 - A i j , (3)
where Ck (resp. dk) is the clustering coefficient (resp. the degree) of node k. Then the probability that this neighbor is not part of an active pacemaker is
▽ i j k = S ( 1 - 2 R E ) c i j k + E ( 1 - 2 S R ) c i j k + R ( 1 - 2 S E ) c i j k . (4)
Taking together Eqs 2 and 4, we obtain a prediction for the coactivation of nodes i and j, named FC1:
F C 1 i j = ( 1 - △ i j ) ( 1 - ∏ k ∈ N i j ▽ i j k ) / 3 , (5)
where the factor 1/3 is taking into account the maximum excitation frequency of nodes in the SER dynamics.
The prediction of coactivation probability, Eq 5, does not require the two nodes to be in the same state. When considered in isolation, the two nodes actually have the capacity to synchronize their respective phases. Indicative of such a synchronization are consecutive time steps of a node spent in the S state. In practice (i.e., when embedding such a substructure in a broader network context), the capacity of a pair of nodes to synchronize in such a way is strongly reduced due to other incoming excitations. A further difference between these toy model networks and more general topological situations is that a much wider range of triangle types needs to be taken into account. When considering more realistic (complex) networks, the previous prediction fails to match simulations (Fig 2B). In fact, given the deterministic nature of the model, any triangle used as a pacemaker in the neighborhood of a pair of nodes contributes (non-trivially) to the coactivations. Therefore, we have to enumerate all possible triangle configurations surrounding a pair of nodes.
There exists a variety of triangle motifs potentially surrounding a pair of nodes (Fig 3). In the following, we classify such triangles according to their distance to the nodes of interest. First, we have triangles adjacent to the pair which can be characterized as follows, triangles adjacent to:
One step further, we have a set of triangles not directly adjacent to (i, j), but adjacent to their common neighbors; as such we have the set of triangles adjacent to:
Such an arrangement defines an hierarchy of triangles, where the levels are defined by the distance from the pair of nodes. The first level corresponds to the set {t0k} where the average distance is zero, {t1k} represents the second level with an average distance of 1. In this way, we can iteratively define the successive levels of the hierarchy, for instance the third level (Fig 3). However, from a dynamical perspective, higher level triangles have a negligible contribution to FC. Moreover, taking into account such higher order triangles significantly complicates the analytical prediction (see below).
Once having enumerated all possible triangles surrounding a pair of nodes, we can now in a systematic way enumerate all possible contributions (or non-contributions) of these triangles to FC if used as pacemakers. We introduce here a new notation for easier reading. In particular, a circled arrow around a topological quantity crossed indicates that it is cancelled, meaning that no triangle is used as pacemaker. The orientation of the arrow represents the direction of propagation of the excitation within the triangle (double arrows represent the possibility of excitation running in both ways). Moreover, capital letters are used for indices i and/or j if a pacemaker is adjacent to them. For example, the above prediction reads as follows:
is the probability of having no triangle adjacent to both i and j (t0) used as pacemaker, and,
is the probability to have at least one triangle used as a pacemaker adjacent to a common neighbor of the pair (i, j) and to i and j at one step distance. Then, when we merge all conditional probabilities, the probability of coactivations, named FC2, reads:
(6)
where greylevel codes for the levels (1st gray and 2nd black). See Materials and Methods section for further details.
We investigated the relative predictive power of the different analytical formulae (including SC, TO, FC1 and FC2) across three typical network topologies (see Materials and Methods section). The formalism based on pacemakers (FC2—Eq 6) in general has the best predictive power, both in terms of correlation and mean difference (Fig 4). FC1 (resp. SC and TO) generally over- (resp. under-) estimated the empirical FC. The predictions appear robust across network realizations as well as across different level of network’s density (S1 Fig). Additionally, we also verified and validated our formalism with generic graphs with a given triangle motifs distribution (S2 Fig).
Next, we fully explore the space of possible initial conditions. We observed that the correlation between SC and FC is highly specific for the topological properties of the underlying structural network. As previously reported, random networks display no or slightly negative SC-FC correlations (depending on their density), scale-free graphs have moderate negative correlations between SC and FC, and modular networks show strong positive correlations (Fig 5). Additionally, the analytical framework based on pacemakers is able to predict the simulated FC for a relatively wide range and, thus, across almost the full space of possible initial conditions.
In the present study, we developed an analytical framework for predicting the patterns of functional connectivity (that is, the coactivations of nodes) in excitable dynamics on neural networks. As an initial step, we considered a minimal model of neural excitation, the deterministic SER model. The predictions were based on key dynamical ingredients in the deterministic regime of the model, that is, period-3 oscillations and the presence of triangles serving as pacemakers The proposed analytical formula generally predicts the simulated FC across a wide range of initial settings. The framework can be extended to achieve arbitrarily high precision, by including even higher order triangles motifs. However, there are diminishing returns due to the very small contributions of the higher order motifs, which are obtained at a high degree of computational complexity.
This study provides a fully analytic framework for predicting patterns of functional connectivity from the structural topology of networks. Previous work has already demonstrated that it is the network topology, rather than the specific computational model, that shapes FC [13, 14]. While some intuitive rules can be formulated regarding the features of the topology that may be relevant [17], a substantiated mechanistic explanation of the contributions of different network elements to non-stationary activity patterns and FC has so far been lacking. The present framework achieves this goal for the deterministic SER model and allows to predict FC for a wide range of initial conditions with high accuracy.
While the SER model at first glance may appear overly simplistic, it captures the essential steps of the susceptibility, excitation and recovery of neural dynamics and, due to this general nature, can be widely applied in neuroscience. Indeed, variants of this kind of discrete excitable model have been broadly used to simulate neural network dynamics ranging from the interactions of single cells [25, 26, 34] to multi-scale whole brain dynamics [35, 36]. When combined with an appropriate forward model, the discrete excitable model can also be brought to produce realistic approximations of observable neural dynamics, such as in fMRI BOLD signals [12, 30]. Thus, the present framework may also be used for the analytical prediction of empirical functional connectivity derived from large-scale neuroimaging.
At the same time, the discrete excitable model has some methodological advantages over more detailed and complex models of neural dynamics (such as biophysical and mass models). Due to the limited number of its parameters, involving only the statistics of the three states, as well as the absence of further free parameters, the model facilitates a systematic exploration of the initial conditions of excitable networks. Indeed, for suitably small graphs, the model even allows the exhaustive characterization of the complete operating range of a particular network architecture. Empirically, the dependence of dynamical patterns on initial conditions is still largely unexplored, due to the technical challenges of controling such factors. However, there is some evidence showing that neuronal excitability may modulate brain functional connectivity, for example in Alzheimer’s disease [37]. Moreover, the deterministic setting of the model allows the systematic identification of the topological contributions to the coactivation patterns of nodes. For instance, we observed that the predictions work generally fairly well for modular graphs, as in [30]. This is an crucial observation, as virtually all networks in systems biology are modular.
The insights gathered in the present study can be used as a basis for exploring more elaborate models of neural dynamics. In particular, it needs to be explored how the topological predictors of activation patterns identified in the present study can be transposed into more complicated scenarios, such as the stochastic version of the SER model [38] or alternative dynamical models, such as the Fitzhugh-Nagumo model [30]. These next steps will further deepen our mechanistic understanding of how the characteristic topological features of complex brain networks contribute to their global activity patterns and function.
To investigate the role of topology for the functional connectivity patterns of networks in the SER model, we considered three different types of undirected benchmark graphs: random, scale-free, and modular networks. The random graph was the classical Erdős-Rényi model [39], the scale-free graph was the Barabási-Albert model [40], and the modular graph was a composition of four small random graphs of identical size and with few links among them. The artificial networks had 60 nodes and about 800 links, and were generated with the software package NetworkX [41] as used in [28]. The layouts were generated using a force-directed algorithm [42].
Additionally, we also explored the robustness of the predictions across various network realizations and densities (from 0.1 to 0.6 by step of 0.1). For each density value, we generated 50 synthetic random graphs with 60 nodes, computed the simulated and predicted FCs, and quantified the predictive power of the analytical proposals. Synthetic graphs were generated using the Brain Connectivity Toolbox (https://sites.google.com/site/bctnet/) [43].
We used a simple three-state cellular automaton model of excitable dynamics, the SER model, representing a stylized biological neuron or neural population [20]. The SER model operates in discrete time and employs the following synchronous update rules, a node in the:
This deterministic version was investigated in detail in [28], where, for example, the role of cycles in storing excitations and supporting self-sustained activity was elucidated. The only remaining parameters are the underlying topology of the structural connectivity on which the model runs, and the pattern of initial states.
After appropriate initialization of the deterministic model, the network activity settles into a regular periodic behavior. Therefore, the nodes are divided into distinct groups; nodes are in the same dynamic group when they are simultaneously active. To analyze the pattern of joint excitations, we computed the number of simultaneous excitations for all pairs of nodes. The outcome matrix is the so-called coactivation matrix, a representation of the functional connectivity of the nodes:
C i j = ∑ t 1 E ( x i t ) 1 E ( x j t ) ,
where x i t ∈ {S, E, R} being the state of node i at time t, and 1 E the indicator function of state E
1 E ( x i t ) = { 1 , if x i t ∈ E 0 , otherwise
In the deterministic SER model, for each network and each initial condition setting, we simulated 5 000 runs of 50 time steps. Unless otherwise stated, the initial conditions were randomly generated, with a probability to set a node into the excited state E between 0 and 1; while the remaining nodes were equipartitioned into susceptible S and refractory R states. FC was summed over runs and normalized by dividing by the product of the number of runs and time steps, scaling FC values between 0 and 1/3. This normalized coactivation matrix was used for all subsequent analyses.
In order to probe the predictive power of the different analytical proposals, including SC itself, we computed the Pearson correlation as well as the mean signed difference between the simulated and predicted FCs. Diagonal elements of matrices were excluded to avoid spurious variations of the prediction. Additionally, SC and TO (the purely topological predictors) were normalized when computing their predictive values, to avoid spurious variations of mean differences with simulated FCs. The normalization was done by dividing them by three times their maximum, effectively scaling the values between 0 and 1/3, as for FC.
We here enumerate in a systematic way all possible (non-)contributions of the triangles motifs to FC if used as pacemakers. Given the periodic behavior of pacemakers, in the following treatment we only describe the cases where a pair of nodes (i, j) is in SS or SE states, all others configurations can be deduced in a similar way.
We have to consider two main components. The first one is the set of pacemakers which contribute (or not) systematically to FC, and the other one corresponds to the set of pacemakers which may lead to a systematic contribution to FC providing that the pair (i, j) synchronizes over time. The systematic (non-)contribution of each triangle motif is as follows:
Assuming synchronization of the pair (i, j), the contribution of the different triangles is as follows:
In order to validate our approach, we used synthetic toy graphs with a desired triangle motifs around a given pair of nodes. For each triangle motif, we then generated a random graph with 30 nodes. For triangle motifs involving common neighbors (ie, t01, t02, t11 and t12), we fixed the number of common neighbors to 6. Everything else was set randomly.
|
10.1371/journal.pcbi.1006886 | Evolution of interface binding strengths in simplified model of protein quaternary structure | The self-assembly of proteins into protein quaternary structures is of fundamental importance to many biological processes, and protein misassembly is responsible for a wide range of proteopathic diseases. In recent years, abstract lattice models of protein self-assembly have been used to simulate the evolution and assembly of protein quaternary structure, and to provide a tractable way to study the genotype-phenotype map of such systems. Here we generalize these models by representing the interfaces as mutable binary strings. This simple change enables us to model the evolution of interface strengths, interface symmetry, and deterministic assembly pathways. Using the generalized model we are able to reproduce two important results established for real protein complexes: The first is that protein assembly pathways are under evolutionary selection to minimize misassembly. The second is that the assembly pathway of a complex mirrors its evolutionary history, and that both can be derived from the relative strengths of interfaces. These results demonstrate that the generalized lattice model offers a powerful new idealized framework to facilitate the study of protein self-assembly processes and their evolution.
| Protein complexes assemble by joining individual proteins together through interacting binding sites. Because of the long time scales of biological evolution, it can be difficult to reconstruct how these interactions change over time. We use simplified representations of proteins to simulate the evolution of these complexes on a computer. In some cases the order in which the complex assembles is crucial. We show that biological evolution increases the strength of interactions that must occur earlier, and decreases the strength of later interactions. Similar knowledge of interactions being preferred to be stronger or weaker can also help to predict the evolutionary ancestry of a complex. While these simulations are too idealized to make exact predictions, this general link between ordered pathways in assembly and evolution matches well-established observations that have been made in real protein complexes. This means that our model provides a powerful framework to help study protein complex assembly and evolution.
| Many proteins self-assemble into protein quaternary structures, which fulfill a multitude of functions across a wide range of biological processes [1]. A general class of polyomino tile self-assembly models have strong analytic potential while maintaining semblance to protein quatenary structure and retaining qualitative realism.
The polyomino self-assembly model [2] combines lattice tile self-assembly with a quantification of biological complexity, examining the relationship between genetic description length and phenotypic complexity. The same model was developed and expanded with evolutionary dynamics by Johnston et al. [3], and used to probe general properties of genotype-phenotype maps by Greenbury et al. [4].
Here we develop a generalization of interactions using binary strings in these polyomino assembly models, in particular introducing variable binding strengths and relaxing the rejection of misassembly.
Binding affinity is difficult to assess experimentally but central to making predictions on assembly [5]. A dominant cause in altering the affinity is mutations to polar or charged groups [6]. While our binary interface polyomino self-assembly model does not account for the variety of amino acids and their particular properties, it provides a reasonable coarse-grained approach. Similar models of protein interactions using binary subunit interfaces have linked protein-protein interaction properties to experimental observations on protein family evolution [7, 8].
Adding these features into polyomino models enables preliminary explorations into the evolution of binding strengths and the implications binding strengths can have on preferred evolutionary pathways.
Several recent studies have revealed the deep relationship between evolutionary pathways and assembly properties like stoichiometry [9], symmetry [10], interaction topology [1], and binding strengths [11]. We aim to reproduce several of these observations in the framework of our generalized polyomino model in order to highlight its potential as a tool for the study of protein assembly and its evolution.
Although almost all proteins have geometric complexity beyond this model’s reach, the topology of interactions, and thus the assembly dynamics, can still be well approximated. Questions focusing on the general properties of shape evolution, interface strengthening, pathway preferences, etc. are hence within the model’s remit. Examining situations where the geometry, not just the topology, of interaction matter is much more limiting. Cooperative bindings, where one interaction may influence the affinity of another, and multi-site bindings are not presently in the model, and the model is not relevant to intrinsically disordered proteins or fuzzy complexes [12]. As such, this model is best applied to analyzing “structured” protein complexes consisting of one or more independent interactions between subunits.
Any self-assembling system requires two ingredients: assembly subunits with binding sites, and a method for determining the strength of an interaction between two such sites. The arrangement of the sites and their interactions can be described in the form of an assembly graph [13]. From these simple components, structures can be formed through the following stochastic assembly process:
If the subunits are square tiles on a lattice, connected sets of tiles are called polyominoes [14].
We can define a genotype that encodes a set of subunit interactions as a sequence, in which each sequence position represents the type of a particular binding site on a subunit. The assembly process maps a given genotype to a single polyomino (in the case of a deterministic genotype) or a statistical distribution of several different polyominoes (for a nondeterministic genotype). In either case these polyominoes can be thought of as abstract biological phenotypes.
The assembly process is independent of the order in which the subunits are represented in the genotype, and translations, rotations, or reflections of a given polyomino are not considered unique. The implementation of this invariance is outlined in S1 Text.
An example of the mapping from genotype to phenotype is shown in Fig 1, using the integer binding site conventions of existing polyomino models. Certain binding sites are noninteracting (labeled 0) while interactions of equal strength occur between fixed pairs of positive integers. The interacting pairs are 1 ↔ 2, 3 ↔ 4, etc.
Repeated assemblies of the same genotype do not necessarily produce the same polyomino, a property referred to as nondeterminism. There are many sources of nondeterminism, ranging from unbound aggregations of subunits to branching pathways in the course of the assembly process. A more general insight into nondeterminism in polyomino self-assembly is given by Tesoro, Ahnert, and Leonard [13].
Deterministic genotypes are significantly outnumbered by nondeterministic ones, and the addition of interactions typically increases the fraction of nondeterministic genotypes. In a biological context nondeterministic genotypes can be viewed as less desirable than deterministic ones, as the functions of many proteins strongly rely on the accuracy and reproducibility of their structures. We can therefore use nondeterminism in the polyomino self-assembly model to represent protein misassembly and thereby study the conditions under which proteins may evolve towards more stable and reliable assemblies.
In this paper we generalize the standard polyomino self-assembly model as outlined above by introducing interfaces that take the form of binary strings rather than integers. This definition of interfaces gives rise to further definitions of interface strength and symmetry. It also allows for non-transitive interactions between interfaces.
The assembly process outlined earlier is unchanged, with only the sites and thus how to determine interactions between them being redefined, as seen in Fig 2.
The number of bits per binding site is given by LI, providing 2 L I unique binding site configurations. Since the subunits are always encoded in a genotype following a common convention (e.g. clockwise around a tile), two adjoined sites have a “head to tail” alignment (see Fig 2).
The interaction strength between two sites relates to the Hamming distance dH between one site and the reversed alignment of the other, normalized by LI. As such, the interaction strength S ^ ∈ [ 0 , 1 ], and binding can occur if the strength is above some chosen critical strength S ^ ≥ S ^ c. The stochastic assembly process as outlined earlier is now extended to include a binding probability as a function of interaction strengths. Interacting subunits are no longer guaranteed to bind, but binding that does occur remains irreversible.
Binding probability can be linked to interaction strength via an abstract temperature T ∈ [0, ∞). More complex forms may have more physical justification, but a useful form of binding probability is
Pr b i n d i n g = H ( S ^ - S ^ c ) S ^ T
where H is the Heaviside function, taking H(0) = 1. The average number of attempts an interaction will take, effectively the binding time, is the reciprocal of the binding probability. With the choice T > 0, stronger bonds are expected to assemble more quickly than weaker bonds.
Using this model, even a small number of subunits can give rise to a large array of potential polyomino structures. We focused our attention on a subset of six assembly graphs that contained both deterministic and nondeterministic phenotypes and transitions, and in which each of the four more complex assembly graphs are in principle accessible from two other members of the set via point mutations. The assembly graphs and phenotypes are shown in Fig 3.
Within the vertical groupings of Fig 3, most phenotypes have comparable abundances, with the exception of the dimer being twice as common as the homotetramer. The abundances could be sampled directly from random sequences, or approximated through combinatorial arguments of permuting edges and noninteracting binding sites. Regardless, our focus centers on the likelihood of successful phenotype transitions, which is independent of abundance, and so is not explicitly treated here.
Evolution was modeled with a fixed-size haploid population undergoing discrete generations of selection and mutation. Reproduction was asexual, and mutations occurred with a fixed probability to flip each bit in a genotype. Non-negative fitnesses were assigned to every individual according to their phenotype properties, with more fit members proportionally more likely to reproduce into the next generation. Nondeterminism was punished by an individual only receiving a fraction of its potential fitness equal to the frequency of correct assembly exponentiated by a parameter γ ∈ [1, ∞).
Accessing information on the evolution of real protein binding strengths over sufficiently long time scales is effectively impossible. There are potential proxies, like looking at homologous proteins across an evolutionary tree [15]. Experimental work has suggested a link between ordered assembly pathways and the constraints they place on evolution [11], but focused on subunits fusing together rather than individual strengths evolving.
Here we show how the generalized polyomino model can simulate evolutionary selection for assembly order, such as observed in [11] for real protein complexes. The possibility of nondeterminism in our generalized model, combined with variable binding strengths, give rise to a space in which evolution can optimize binding strengths in order to maximize the probability that critical assembly steps occur in the right order for a desirable phenotype.
In the steady-state limit of the evolutionary simulations, mutation and selection effectively eliminate any trace of ancestry in the interface strengths. The steady state properties of interaction strengths depend only on the current phenotype. However, shortly after a new shape has evolved, it is possible to deduce ancestry from interface strengths. In the case of the 12-mer and the 16-mer, where we have one nondeterministic ancestor and one deterministic one, this is obvious as the interface strength distributions of the two ancestors differ considerably. As a result the two alternative ancestries for each of these two polyominoes can be clearly distinguished by bond strengths up to about 50 generations.
But even where we have deterministic ancestors, namely for the octomer and the heterotetramer, we notice that at the earliest time points the interface that is also present in the ancestor is stronger than the interface that is absent in the ancestor. This latter observation mirrors results found in real protein complexes, where the ordering of interface strengths often reflects the order of evolution, with the strongest interface as the oldest [10].
The time ordering of assembly steps in proteins is integral to the correct assembly of the protein structure. This holds true on many length scales of assembly, with cotranslational protein folding able to induce misassembly [16] all the way up to final quaternary structure as examined here. Experimental methods for devising binding strengths are still being developed [17], with an in silico approach recently introduced focusing on multimeric complexes [18].
One notable result was that given an equal rate of mutation, deterministic and nondeterministic assemblies adapted at different rates. The peak observed rate of binding strength increase in the 12-mer was approximately triple the rate in deterministic assemblies. Such an observation is fairly intuitive, as mutations which alter binding strength correctly or incorrectly are more strongly selected or purified respectively in the nondeterministic assemblies. This is in good agreement with the observation that unstable proteins adapt more quickly [19].
Binding strengths that deviate from neutral expectations do so to optimize determinism, assembling a core of the final structure as quickly as possible before adding further, peripheral elements. This evolutionary selection for a particular assembly pathway parallels real protein complexes, in which gene fusions are a way of cementing particular assembly order under evolutionary selection pressure in order to minimize the risk of misassembly [11].
Generalizing the binding sites from integers to binary strings provides a range of benefits. The number of binding site configurations is now fixed by a physically meaningful parameter and is exponentially large. Previous models frequently had identical binding sites at multiple locations, which is not observed in real proteins, whereas now repeated binding sites are vanishingly rare. Additionally, interaction rules in the integer model have trivial transitivity relations: Maintaining the notation of ↔ for interactions, that is to say for sites A, B, C that
( A ↔ B ) ∧ ( B ↔ C ) → ( A = C )
However, the generalized model does not require the above relation to be true, with knowledge of one interaction having little bearing on other interactions sharing a binding site. That it is to say for sites D, E, F, G that
( D ↔ E ) ∧ ( E ↔ F ) ∧ ( F ↔ G ) ↛ ( D = F ) ∨ ( D ↔ G )
This allows more complex interaction patterns to form, but also allows different binding sites to produce the same interaction behaviour, as seen in Fig 7. In addition, sites can self-interact, interact with another binding site, or both, like sites D and E supporting the interactions D ↔ E and E ↔ E.
Usefully, the generalized interactions are a superset of the integer model, and so any previous results could be trivially recovered by choosing S ^ c = 1 (up to relabeling binding sites). While the generalized model is still a strikingly abstract representation of biological self-assembly, the binary interfaces add physical realism and layered complexity to an already promising model.
Phenotype plasticity is another feature that is naturally introduced by the generalized model. By incorporating a dynamic fitness landscape, one that alternatively favors two (or more) phenotypes, the interaction strengths can continuously adapt to remain optimal, shown in Fig 8. The ability to modify a phenotype in a controllable manner, minimizing nondeterminism, is a huge advantage to survival. If a conformational change of a protein, in response to an environmental change or other external conditions, altered its binding strengths, it could quickly shift phenotypes.
Since changing interaction strengths can occur much quicker than creating new interactions, this plasticity allows adaptions that would otherwise be potentially too slow to survive. The relationship between conformational changes and their impact on evolution is uncertain, but it has been suggested that this behaviour can impose strong constraints on sequence evolution [20, 21]. Moreover, adding and removing interactions, rather than just reprioritizing them, exposes the assemblies to intermediate states and greater risk of negative outcomes [22].
Polyomino self-assembly models using integers as binding sites have demonstrated the value of abstract self-assembly models for the study of self-assembly phenomena and genotype-phenotype maps [2–4]. Generalizing the binding interfaces using binary subsites as outlined in this paper retains tractability while expanding applicability to more complex biological research questions. In particular, modeling the evolution of interaction strengths provides qualitative insights beyond the reach of previous polyomino studies.
With a few justifiable assumptions, analytic predictions of the interaction strengths in the absence of selection pressures can be found, which show strong agreement with simulations. Significant divergences from this prediction are observed in nondeterministic assemblies where time-ordering is important, and the interaction strengths are therefore under selection. This selection pressure drives these interactions to strengthen or weaken, and thus bind earlier or later in the assemble, to optimize the determinism. Certain interaction strength orderings are more suitable for transitioning to descendant phenotypes, and so can be used to statistically reconstruct evolutionary pathways.
Several observations from experimental studies have been recovered by this model, as well as suggesting that nondeterminism in the polyomino model provides an interesting framework for the study of protein misassembly. Many further avenues are imaginable that build on such investigations of nondeterminism, including gene duplication, phenotype plasticity, and more complex genotype-phenotype mappings.
A full implementation of the self-assembly algorithms, evolutionary dynamics, and phylogenetic analysis written by the authors can be found online through the Data Availability Statement.
Almost all of the parameters and function forms were chosen such that simulations remained within reasonable computation timescales and offered evident trends, but were otherwise arbitrary. The same observations were made while examining different binding site lengths, critical interaction strengths, fitness functions, etc.
As outlined earlier, evolution was modeled with asexual reproduction of haploids encoding two subunits (total of 8 binding sites per genotype). Binding site lengths were LI = 64 and the critical strength was taken as S ^ c = . 671875. Genotypes were initialized randomly, with the constraint that there were no interactions. Assembly could begin with either subunit as the seed, although monomers were ignored due to their trivial contribution.
A population of 250 individuals evolved for 1000 generations, with each genotype being assembled 25 times. Each binary subsite had a fixed probability to flip, such that the entire genotype had mutations that were binomially distributed with mean μ = 1. The temperature was set to T = 25, while the nondeterminism punishment was γ = 5.
An individuals fitness was calculated as ( F ) N I · ϕ γ, where F is the fitness jump between higher order assembly graphs, NI is the number of interactions in an assembly graph, and ϕ is the fraction of assemblies that built the correct phenotype. The fitness jump was F = 5 to balance the strong nondeterminism punishment. So, for example, the fitness for a heterotetramer correctly assembled 20 times out of 25 would be 52 ⋅ .85.
We restricted fitness allocation to the six stated phenotypes in Fig 3, assigning fitness 0 to all other phenotypes. The majority of transitions between these other phenotypes did not display any novel dynamics, and so were minimized to present the most concise results. The results presented here were initially observed in a full system, and this restriction was introduced to improve significantly improve the simulation fidelity and computation time required.
|
10.1371/journal.pcbi.1005644 | A phase transition induces chaos in a predator-prey ecosystem with a dynamic fitness landscape | In many ecosystems, natural selection can occur quickly enough to influence the population dynamics and thus future selection. This suggests the importance of extending classical population dynamics models to include such eco-evolutionary processes. Here, we describe a predator-prey model in which the prey population growth depends on a prey density-dependent fitness landscape. We show that this two-species ecosystem is capable of exhibiting chaos even in the absence of external environmental variation or noise, and that the onset of chaotic dynamics is the result of the fitness landscape reversibly alternating between epochs of stabilizing and disruptive selection. We draw an analogy between the fitness function and the free energy in statistical mechanics, allowing us to use the physical theory of first-order phase transitions to understand the onset of rapid cycling in the chaotic predator-prey dynamics. We use quantitative techniques to study the relevance of our model to observational studies of complex ecosystems, finding that the evolution-driven chaotic dynamics confer community stability at the “edge of chaos” while creating a wide distribution of opportunities for speciation during epochs of disruptive selection—a potential observable signature of chaotic eco-evolutionary dynamics in experimental studies.
| Evolution is usually thought to occur very gradually, taking millennia or longer in order to appreciably affect a species' survival mechanisms. Conversely, demographic shifts due to predator invasion or environmental change can occur relatively quickly, creating abrupt and lasting effects on a species survival. However, recent studies of ecosystems ranging from the microbiome to oceanic predators have suggested that evolutionary and ecological processes can often occur over comparable timescales—necessitating that the two be addressed within a single, unified theoretical framework. Here, we show that when evolutionary effects are added to a minimal model of two competing species, the resulting ecosystem displays erratic and chaotic dynamics not typically observed in such systems. We then show that these chaotic dynamics arise from a subtle analogy between the evolutionary concept of fitness, and the concept of the free energy in thermodynamical systems. This analogy proves useful for understanding quantitatively how the concept of a changing fitness landscape can confer robustness to an ecosystem, as well as how unusual effects such as history-dependence can be important in complex real-world ecosystems. Our results predict a potential signature of a chaotic past in the distribution of timescales over which new species can emerge during the competitive dynamics, a potential waypoint for future experimental work in closed ecosystems with controlled fitness landscapes.
| In many natural ecosystems, at least one constituent species evolves quickly enough relative to its population growth that the two effects become interdependent. This phenomenon can occur when selection forces are tied to such sudden environmental effects as algal blooms or flooding [1], or it can arise from more subtle, population-level effects such as overcrowding or resource depletion [2]. Analysis of such interactions within a unified theory of “eco-evolutionary dynamics” has been applied to a wide range of systems—from bacteria-phage interactions to bighorn sheep [3]—by describing population fluctuations in terms of the feedback between demographic change and natural selection [4].
The resulting theoretical models relate the fitness landscape (or fitness function) to population-level observables such as the population growth rate and the mean value of an adapting phenotypic trait (such as horn length, cell wall thickness, etc). The fitness landscape may have an arbitrarily complex topology, as it can depend on myriad factors ranging from environmental variability [5, 6], to inter- and intraspecific competition [7, 8], to resource depletion [9]. However, these complex landscapes can be broadly classified according to whether they result in stabilizing or disruptive selection. In the former, the landscape may possess a single, global maximum that causes the population of individuals to evolve towards a state in which most individuals have trait values at or near this maximum [10]. Conversely, in disruptive selection, the fitness landscape may contain multiple local maxima, in which case the population could have a wide distribution of trait values and occupy multiple distinct niches [11].
In eco-evolutionary models, the shape of the fitness landscape may itself depend on the population densities of the interacting species it describes. Specifically, the concept that the presence of competition can lead a single-peaked fitness landscape to spontaneously develop additional peaks originates in the context of “competitive speciation” first proposed by Rosenzweig [12]. This is formalized in genetic models in which sympatric speciation is driven by competitive pressures rather than geographic isolation [13]. Competition-induced disruptive selection has been observed in natural populations of stickleback fish [14], microbial communities [15], and fruit flies [16, 17].
Here, we model eco-evolutionary dynamics of a predator-prey system based on first-order “gradient dynamics” [10, 18], a class of models that explicitly define the fitness in terms of the population growth rate r, which is taken to depend only on the mean value of the trait across the entire population, c ¯ [19]. Despite this simplification, gradient dynamics models display rich behavior that can account for a wide range of effects observed in experimental systems—in particular, recent work by Cortez and colleagues has shown that these models can result in irregular cycles and dynamical bifurcations that depend on the standing genetic variation present in a population [20, 21].
In our model, gradient dynamics cause the prey fitness landscape to change as a result of predation, and we find that the resulting dynamical system exhibits chaotic dynamics. Chaos is only possible in systems in which three or more dependent dynamical variables vary in time [22], and previously it has been observed in predator-prey systems comprising three or more mutually interdependent species, or in which an external environmental variable (such as seasonal variation or generic noise) is included in the dynamics [23, 24]. Here we show that evolution of just one species in a two-species ecosystem is sufficient to drive the ecosystem into chaos. Moreover, we find that chaos is driven by a density-dependent change of the fitness landscape from a stabilizing to disruptive state, and that this transition has hysteretic behavior with mathematical properties that are strongly reminiscent of a first-order phase transition in a thermodynamical system. The resulting dynamics display intermittent properties typically associated with ecosystems poised at the “edge of chaos,” which we suggest has implications for the study of ecological stability and speciation.
Adapting the notation and formulation used by Cortez (2016) [21], we use a two-species competition model with an additional dynamical variable introduced to account for a prey trait on which natural selection may act. The most general fitness function for the prey, r, accounts for density-dependent selection on a prey trait c,
r ( x , y , c ¯ , c ) ≡ G ( x , c , c ¯ ) - D ( c , c ¯ ) - f ( x , y ) , (1)
where x = x(t) is the time-dependent prey density, y = y(t) is the time-dependent predator density, c is a trait value for an individual in the prey population, and c ¯ = c ¯ ( t ) is the mean value of the trait across the entire prey population at time t. r comprises a density-dependent birth rate G, a density-independent death rate D, and a predator-prey interaction term f, which for simplicity is assumed to depend on neither c nor c ¯. Thus the trait under selection in our model is not an explicit predator avoidance trait such as camouflage, but rather an endogenous advancement (i.e., improved fecundity, faster development, or reduced mortality) that affects the prey’s ability to exploit resources in its environment, even in the absence of predation.
The continuous-time “gradient dynamics” model that we study interprets the fitness r as the growth rate of the prey: [19, 25]
x ˙ = x r ( x , y , c ¯ , c ) | c → c ¯ (2)
y ˙ = y ( f ( x , y ) - D ˜ ( y ) ) (3)
c ¯ ˙ = V ∂ r ( x , y , c ¯ , c ) ∂ c | c → c ¯ . (4) Eq (2) is evaluated with all individual trait values c set to the mean value c ¯ because the total prey population density is assumed to change based on the fitness function, which in turn depends on the population-averaged value of the prey trait c ¯ [21]. The timescale of the dynamics in c ¯ are set by V, which is interpreted as the additive genetic variance of the trait [10]. While Eq (2) depends only on the mean trait value c ¯, the full distribution of individual trait values c present in a real-world population may change over time as the relative frequencies of various phenotypes change. In principle, additional differential equations of the form of Eq (4) could be added to account for higher moments of the distribution of c across an ensemble of individuals, allowing the gradient dynamics model to be straightforwardly extended to model a trait’s full distribution rather than just the population mean. However, here we focus on the case where the prey density dynamics x ˙ depend only on the mean trait value to first order, and we do not include differential equations for higher-order moments of the prey trait value distribution.
The use of a single Eq (4) to describe the full dynamics of the trait distribution represents an approximation that is exact only when the phenotypic trait distribution stays nearly symmetric and the prey population maintains a constant standing genetic variation V [10]. However, V may remain fixed even if the phenotypic variance changes, a property that is observed phenomenologically in experimental systems, and which may be explained by time-dependent heritability, breeding effects, mutation, or other transmission effects not explicitly modeled here [26–29]. More broadly, this assumption may imply that gene selection is weak compared to phenotype selection [30, 31]. S1D Appendix further describes the circumstances under which V remains fixed, and also provides a first-order estimate of the magnitude of error introduced by ignoring higher-order effects (such as skewness) in the trait distribution. The results suggest that these effects are small for the parameter values (and resulting range of x and y values) used here, due in part to limitations on the maximum skewness that a hypothetical trait distribution can achieve on the fitness landscapes studied here. In S1D Appendix, we also compare the results presented below to an equivalent model in which a full trait distribution is present, in which case Eq (2) becomes a full integro-differential equation involving averages of the trait value over the entire prey population. Detailed numerical study of this integro-differential equation is computationally prohibitive for the long timescales studied here, but direct comparison of the contributions of various terms in the velocity field suggests general accuracy of the gradient dynamics model for the fitness landscapes and conditions we study here. However, in general the appropriateness of the gradient dynamics model should be checked whenever using Eq (4) with an arbitrary fitness function. Fig 1A shows a schematic summarizing the gradient dynamics model, and noting the primary assumptions underlying this formulation.
Next, we choose functional forms for f, G, D, and D ˜ in Eqs (2) and (3). We start with the assumption that, for fixed values of the trait c an d its mean c ¯, the population dynamics should have the form of a typical predator-prey system in the absence of evolutionary effects. Because the predator dynamics are not directly affected by evolutionary dynamics, we choose a simple form for predator growth consisting of a fixed death rate and a standard Holling Type II birth rate, [32]
f ( x , y ) = a 2 x y 1 + b 2 x (5)
D ˜ ( y ) = d 2 (6)
The predator birth rate f saturates at large values of the prey density, which is more realistic than the standard Lotka-Volterra competition term xy in cases where the prey density is large or fluctuating [22]. A saturating interaction term ensures that solutions of the system remain bounded for a wider range of parameter values, a necessity for realistic models of long-term interactions [33].
For the prey net growth rate (Eq (1), the fitness) in the absence of the predator, we use the following functional forms,
G ( x , c ¯ , c ) = a 1 c ¯ 1 + b 1 c ¯ ( 1 - k 1 x ( c - c ¯ ) ) (7)
D ( c , c ¯ ) = d 1 ( 1 - k 2 ( c 2 - c ¯ 2 ) + k 4 ( c 4 - c ¯ 4 ) ) . (8)
The first term in Eq (7) specifies that the prey population density growth rate r | c → c ¯ depends only on a primary saturating contribution of the mean trait to the birth rate G. In other models a similar effect is achieved by modifying the mean trait evolution Eq (4), such that extremal values of the trait are disadvantaged [21]; alternative coupling methods based on exponential saturation would be expected to yield similar qualitative results [19]. However, the additional series terms in Eqs (7) and (8) ensure that the any individual’s fitness r may differ from the rest of the population depending on the difference between its trait value c and the population mean c ¯. Because the functional form of this difference is unknown, its contribution expressed as second-order truncation of the series difference of the form r ( c , c ¯ ) = r ˜ | c → 0 + ( r ˜ ( c ) - r ˜ ( c ) | c → c ¯ ) (where r ˜ represents an unscaled fitness function). This ensures that when c ˙ = 0 or c = c ¯, the system reduces to a standard prey model with a Holling Type II increase in birth rate in response to increasing mean trait value [25]. In the results reported below, we observe that all dynamical variables remain bounded as long as parameter values are chosen such that the predator density does not equilibrate to zero. This is a direct consequence of our use of saturating Holling Type II functional forms in Eqs (7) and (8), which prevent the fitness landscape from increasing without bound at large c, c ¯ and also ensure that the predator and prey densities do not jointly diverge. That the dynamics should stay bounded due to saturating terms is justified by empirical studies of predator-prey systems [34, 35]; moreover, other saturating functional forms are expected to yield similar results if equivalent parameter values are chosen [33, 36].
The nonlinear dependence of the mortality rate Eq (8) on the trait is based on mechanistic models of mortality with individual variation [19, 37, 38]. The specific choice of a quartic in Eq (8) allows the fitness function r to have a varying number of real roots and local maxima in the domain c, c ¯ > 0, affording the system dynamical freedom not typically possible in predator prey models with constant or linear prey fitness—in particular, for different values of k2, k4 the fitness landscape can have a single optimal phenotype, multiple optimal phenotypes, or no optimal intermediate values. Because any even, continuous form for the fitness landscape can be approximated using a finite number of terms in its Taylor series around c = 0, our choice of a quartic form simply constitutes truncation of this expansion at the quartic order in order to include the simplest case in which the fitness function admits multiple local maxima—for this reason, a quartic will always represent the leading-order series expansion of a fitness landscape with multiple local maxima. Below, we observe numerically that ∣ c - c ¯ ∣ < 1, ex post facto justifying truncation of the higher order terms in this series expansion. However, if the trait value c was strictly bounded to only take non-zero values on a finite interval (as opposed to the entire real line), then a second-order, quadratic fitness landscape would be sufficient to admit multiple local maxima (at the edges of the interval) [14]. However, the choice here of an unbounded trait value c avoids creating boundary effects, and it has little consequence due to the steep decay of the quartic function at large values of |c|, which effectively confines the possible values of c ¯ accessible by the system. In physics, similar reasons—unbounded domains, multiple local optima, and continuity—typically justify the use of quartic free energy functions in minimal models of systems exhibiting multiple energetic optima, such as the Ginzberg-Landau free energy used in models of superconducting phase transitions [39].
We note that the birth rate Eq (7) contributes a density-dependent term to the fitness function even in the absence of predation (y = 0) [21]. Unlike the death rate function, the effect of the individual trait value on this term is directional: the sign of c - c ¯ determines whether birth rates increase or decrease. As the population density x increases, the effect of these directional effects is amplified, consistent with the observed effect of intraspecific competition and crowding in experimental studies of evolution [40, 41]. The chaotic dynamics reported below arise from this density-dependent term because the term prevents the Jacobian of the system (2), (3) and (4) from having a row and column with all zeros away from the diagonal; in this case, the prey trait (and thus evolutionary dynamics) would be uncoupled from the rest of the system, and would thus relax to a stable equilibrium (as is necessary for a first-order single-variable equation). In that case, c ¯ would essentially remain fixed and the predator-prey dynamics would become two-dimensional in x and y, precluding chaos. For similar reasons, density-dependent selection has been found to be necessary for chaos in some discrete-time evolutionary models, for which chaotic dynamics require a certain minimum degree of association between the fitness and the trait frequencies [42].
Inserting Eqs (5), (7) and (8), into Eq (1) results in a final fitness function of the form
r ( x , y , c ¯ , c ) = a 1 c ¯ 1 + b 1 c ¯ ( 1 - k 1 x ( c - c ¯ ) ) - d 1 ( 1 - k 2 ( c 2 - c ¯ 2 ) + k 4 ( c 4 - c ¯ 4 ) ) - a 2 x y 1 + b 2 x . (9)
This fitness landscape is shown in Fig 1B, for typical parameter values and predator and prey densities used in the numerical results below. Depending on the current predator and prey densities, the local maximum of the system can appear in two different locations, which directly affects the dynamics described in the next section.
Inserting Eq (9) into Eqs (2), (3) and (4) results in a final form for the dynamical equations,
x ˙ = x ( a 1 c ¯ 1 + b 1 c ¯ - a 2 y 1 + b 2 x - d 1 ) (10)
y ˙ = y ( y a a 2 x 1 + b 2 x - d 2 ) (11)
c ¯ ˙ = c ¯ V ( ( 2 k 2 d 1 ) - ( 4 k 4 d 1 ) c ¯ 2 - ( a 1 k 1 ) x 1 + b 1 c ¯ ) . (12)
Due to the Holling coupling terms, the form of these equations qualitatively resembles models of vertical, tritrophic food webs—the mean trait value c ¯ affects the growth rate of the prey, which in turn affects the growth rate of the predator [24, 32, 43]. The coupling parameter ya introduces asymmetry into the competition when ya ≠ 1; however, it essentially acts as a scale factor that only affects the amplitude of the y cycles and equilibria rather than the dynamics. Additionally, because the predator-prey interaction term Eq (5) is unaffected by the trait, our model contains no triple-product x y c ¯ interaction terms, which typically stabilize the dynamics.
For our analysis of the system (10), (11) and (12), we first consider the case where evolution proceeds very slowly relative to population dynamics. In the case of both no evolution (V = 0) and no predation (y = 0), the prey growth Eq (10) advances along the one-dimensional nullcline y ˙, c ¯ ˙ = 0, y = 0. Depending on whether the fixed mean trait value c ¯ exceeds a critical value (c ¯ † ≡ d 1 / ( a 1 - b 1 d 1 )), the prey density will either grow exponentially (c ¯ > c ¯ †) or collapse exponentially (c ¯ < c ¯ †) because the constant c ¯ remains too low to sustain the prey population in the absence of evolutionary adaptation. The requirement that c ¯ > c ¯ † carries over to the case where a predator is added to the system but evolutionary dynamics remain fixed, corresponding to a two dimensional system advancing along the two-dimensional nullcline c ¯ ˙ = 0. In this case, as long as c ¯ > c ¯ †, the prey density can exhibit continuous growth or cycling depending in the relative magnitudes of the various parameters in Eqs (10) and (11). The appearance and disappearance of these cycles is determined by a series of bifurcations that depends on the values of c ¯ and b1, b2 relative to the remaining parameters a1, a2, d1, d2 (S1A Appendix).
In the full three-variable system (10), (11) and (12), c ¯ passes through a range of values as time progresses, resulting in more complex dynamics than those observed in the two-dimensional case. For very small values of V, the evolutionary dynamics c ¯ ˙ are slow enough that the system approaches the equilibrium predicted by the two-variable model with c ¯ constant. The predator and prey densities initially grow, but the prey trait value does not change fast enough for the prey population growth to sustain—eventually resulting in extinction of both the predator and prey. However, if V takes a slightly larger value, so that the mean trait value can gradually change with a growing prey population density (due to the density-dependent term in Eq (10)), then the population dynamics begin to display regular cycling with fixed frequencies and amplitudes (Fig 2A, top). This corresponds to a case where the evolutionary dynamics are slow compared to the ecological dynamics, but not so slow as to be completely negligible. Finally, when V is the same order of magnitude as the parameters governing the ecological dynamics, the irregular cycles become fully chaotic, with both amplitudes and frequencies that vary widely over even relatively short time intervals (Fig 2A, bottom). Typically, the large V case would correspond to circumstances in which the prey population develops a large standing genetic variation [10, 44].
That the dynamics are chaotic, rather than quasi-periodic, is suggested by the presence of multiple broad, unevenly-spaced peaks in the power spectrum [45] (Figure A in S1E Appendix), as well as by numerical tabulation of the Lyapunov spectrum (described further below). Due to the hierarchical coupling of Eqs (10), (11) and (12), when plotted in three-dimensions the chaotic dynamics settle onto a strange attractor that resembles the “teacup” attractor found in models of tritrophic food webs [24, 46] (Fig 2B). Poincare sections though various planes of the teacup all appear linear, suggesting that the strange attractor is effectively two-dimensional—consistent with pairings of timescales associated with different dynamical variables at different points in the process (Figure B in S1E Appendix). In the “rim” of the teacup, the predator density changes slowly relative to the prey density and mean trait value. This is visible in a projection of the attractor into the x - c ¯ plane (Fig 2B, bottom inset). However, in the “handle” of the teacup, the mean trait value varies slowly relative to the ecological dynamics (c ¯ ˙ ≈ 0), resulting in dynamics that qualitatively resemble the two-dimensional “reduced” system described above for various fixed values of c ¯ (Fig 2B, top inset).
The structure of the attractor suggests that the prey alternately enters periods of evolutionary change and periods of competition with the predator. A closer inspection of a typical transition reveals that this “two timescale” dynamical separation is responsible for the appearance of chaos in the system (Fig 3A). As the system explores configuration space, it reaches a metastable configuration corresponding to a high mean trait value c ¯, which causes the prey density to nearly equilibrate to a low density due to the negative density-dependent term in Eq (10). During this period (the “rim” of the teacup), the predator density gradually declines due to the lack of prey. However, once the predator density becomes sufficiently small, the prey population undergoes a sudden population increase, which triggers a period of rapid cycling in the system (the “handle” of the teacup attractor). During this time, the predator density continuously increases, causing an equivalent decrease in the prey density that resets the cycle to the metastable state.
The sudden increase in the prey population at low predator densities can be understood from how the fitness function r (from Eq (9)) changes over time. Fig 3B shows a kymograph of the log-scaled fitness Eq (9) as a function of individual trait values c, across each timepoint and corresponding set of (x, y, c ¯) values given in panel A. Overlaid on this time-dependent fitness landscape are curves indicating the instantaneous location of the local maximum (black) and minimum (white). By comparing panels A and B, it is apparent that the mean trait value during the “metastable” period of the dynamics stays near the local maximum of the fitness function, which barely varies as the predator density y changes. However, when y(t) ≈ 0.25, the fitness function changes so that the local minimum and local maximum merge and disappear from the system, leading to a new maximum spontaneously appearing at c = 0. Because V is large enough (for these parameters) that the gradient dynamics occur over timescales comparable to the competition dynamics, the system tends to move rapidly towards this new maximum in the fitness landscape, resulting in rapidly-changing dynamics in x and c ¯. Importantly, because of the symmetric coupling of the prey fitness landscape r to the prey density x, this rapid motion resets the fitness landscape so that the maximum once again occurs at the original value, resulting in a period of rapid cycling. The fitness landscape at two representative timepoints in the dynamics is shown in Fig 3C.
That the maxima in the fitness Function (9) suddenly change locations with continuous variation in x, y is a direct consequence of the use of a high-order (here, quartic) polynomial in c to describe the fitness landscape. The quartic represents the simplest analytic function that admits more than one local maxima in its domain, and the number of local maxima is governed by the relative signs of the coefficients of the ( c 2 - c ¯ 2 ) and ( c 4 - c ¯ 4 ) terms in Eq (9), which change when the system enters the rapid cycling portion of the chaotic dynamics at t = 500 in Fig 3A. This transition marks the mean prey trait switching from being drawn (via the gradient dynamics) to a single fitness peak at an intermediate value of the trait ceq ≈ 0.707 to being drawn instead to one of two peaks: the existing peak, or a new peak at the origin. Thus the metastable period of the dynamics corresponds to a period of stabilizing selection: if the fitness landscape were frozen in time during this period, then an ensemble of prey would all evolve to a single intermediate trait value corresponding to the location of the global maximum. Conversely, if the fitness landscape were held fixed in the multipeaked form it develops during a period of rapid cycling, given sufficient time an ensemble of prey would evolve towards subpopulations with trait values at the location of each local fitness maximum—representing disruptive selection. That the fitness landscape does not remain fixed for extended durations in either a stabilizing or disruptive state—but rather switches between the two states due to the prey density-dependent term in Eq (9)— underlies the onset of chaotic cycling in the model. Density-dependent feedback similarly served to induce chaos in many early discrete-time ecosystem models [23]. However, the “two timescale” form of the chaotic dynamics and strange attractor here is a direct result of reversible transitions between stabilizing and disruptive selection.
If the assumptions underlying the gradient dynamics model do not strictly hold—if the additive genetic variance V slowly varies via an additional dynamical equation, or if the initial conditions are such that significant skewness would be expected to persist in the phenotypic distribution, then the chaotic dynamics studied here would be transient rather than indefinite. While the general stability analysis shown above (and in the S1 Appendix) would still hold, additional dynamical equations for V or for high-order moments of the trait distribution would introduce additional constraints on the values of the parameters, which would (in general) increase the opportunities for the dynamics to become unstable and lead to diverging predator or prey densities. However, in some cases these additional effects may actually serve to stabilize the system against both chaos and divergence. For example, if additional series terms were included in Eq (8) such that the dependence of mortality rate on c ¯ and c had an upper asymptote [25], then c ¯ ˙ = 0 would be true for a larger range of parameter values—resulting in the dynamical system remaining planar for a larger range of initial conditions and parameter values, precluding chaos.
The transition between stabilizing and disruptive selection that occurs when the system enters a period of chaotic cycling is strongly reminiscent of a first-order phase transition. Many physical systems can be described in terms of a free energy landscape, the negative gradient of which determines the forces acting on the system. Minima of the free energy landscape correspond to equilibrium points of the system, which the dynamical variables will approach with first-order dynamics in an overdamped limit.
When a physical system undergoes a phase transition—a qualitative change in its properties as a single “control” parameter, an externally-manipulable variable such as temperature, is smoothly varied—the transition can be understood in terms of how the control parameter changes the shape of the free energy landscape. The Landau free energy model represents the simplest mathematical framework for studying such phase transitions: a one-dimensional free energy landscape is defined as a function of the control parameter and an additional independent variable, the “order parameter,” a derived quantity (such as particle density or net magnetization) with respect to which the free energy can have local minima or maxima. In a first-order phase transition in the Landau model, as the control parameter monotonically changes the relative depth of a local minimum at the origin decreases, until a new local minimum spontaneously appears at a fixed nonzero value of the order parameter—resulting in dynamics that suddenly move towards the new minimum, creating discontinuities in thermodynamic properties of the system such as the entropy [47]. First-order phase transitions are universal physical models, which have been used to describe a broad range of processes spanning from superconductor breakdown [48] to primordial black hole formation in the early universe [49].
In the predator-prey model with prey evolution, the fitness function is analogous to the free energy, with the individual trait value c serving as the “order parameter” for the system. The control parameter for the transition is the prey density, x, which directly couples into the dynamics via the density-dependent term in Eq (7). Because the fitness consists of a linear combination of this term in Eq (7) and a quartic landscape Eq (8), the changing prey density “tilts” the landscape and provokes the appearance of the additional, disruptive peak visible in Fig 3C. The appearance and disappearance of local maxima as the system switches between stabilizing and disruptive selection is thus analogous to a first-order phase transition, with chaotic dynamics being a consequence of repeated increases and decreases of the control parameter x above and below the critical prey densities x*, x** at which the phase transition occurs. Similar chaotic dynamics emerge from repeated first-order phase transitions in networks of coupled oscillators, which may alternate between synchronized and incoherent states that resemble the “metastable” and “rapid cycling” portions of the predator-prey dynamics [50].
The analogy between a first-order phase transition and the onset of disruptive selection can be used to study the chaotic dynamics in terms of dynamical hysteresis, a defining feature of such phase transitions [47]. For different values of x, the three equilibria corresponding to the locations of the local minima and maxima of the fitness landscape, ceq, can be calculated from the roots of the cubic in Eq (12). The resulting plots of ceq vs x in Fig 4 are generated by solving for the roots in the limit of fast prey equilibration, c ¯ → c e q, which holds in the vicinity of the equilibria (S1B Appendix).
The entry into the transient chaotic cycling occurs when x increases gradually and shifts ceq with it; x eventually attains a critical value x* (x* ≈ 0.45 for the parameters used in the figures), causing ceq to jump from its first critical value c* to the origin (the red “forward” branch in Fig 4). This jump causes rapid re-equilibration of c ¯ ( t ), resulting in the rapid entry into cycling observable in Fig 3A. However, x cannot increase indefinitely due to predation; rather, it decreases until it reaches a second critical value x**, at which point ceq jumps back from the origin to a positive value (the blue “return” branch in Fig 4; x** = 0.192 for these parameter values). This second critical point marks the return to the metastable dynamics in Fig 3A. This asymmetry in the forward and backwards dynamics of x lead to dynamical time-irreversibility (hysteresis) and the jagged, sawtooth-like cycles visible in the dynamics of the full system. Because the second jump in ceq is steeper, the parts of the trajectories associated with the “return” transition in Fig 3A appear steeper. Additionally, the maximum value obtained by c ¯ ( t ) anywhere on the attractor, c e q m a x, is determined by the limiting value of ceq as x → 0. Analytic values for c e q m a x, as well as (x*, c*) and (x**, c**), are derived in the S1 Appendix, and their corresponding numerical values are overlaid in each panel of Fig 3.
Comparing the values of c e q m a x, x*, c*, x**, c** to the dynamics of the system in Fig 3, it is apparent that calculation of critical points under the fast-evolution approximation correctly predicts key properties of the chaotic dynamics such as the maximum value attained by c ¯ ( t ), the quasi-static value c ¯ during the “metastable” period before chaotic cycling, and the approximate values at which x(t) enters and exits the rapid cycling portion of the dynamics. Thus the analogy between the fitness function and the Landau free energy provides insight into the dynamics of the chaotic ecosystem.
Moreover, for intermediate values of the prey density at which the two local maxima are equal heights, the relative fitnesses of the two trait values are equal (c * = c e q * *) and so both phenotypes would be equally favorable for the prey population. This is analogous to the coexistence of two phases during intermediate portions of a phase transition. As the prey density x approaches either critical value, the fitness landscape shallows and the dynamics begin to exhibit a form of “critical slowing down” associated with the onset of the phase transition—here represented by the relatively slow dynamics along the flattened handle of the teacup in Fig 2B.
The chaotic dynamics reported here are emergent; they result from predation reducing the fitness of intermediate trait values, which restructures the fitness landscape in a manner that later reverses as the predator density decreases. However, here, as in other models, the presence of chaos has other long-term implications for the ecosystem that would not be relevant in systems with only limit cycles or point equilibria.
The chaotic dynamics associated with fast evolutionary dynamics (large V, or high genetic variance [20, 21]) impose a statistical structure on the deterministic problem: given a sufficiently long observation time, a trajectory along the strange attractor will sample every point on the attractor [45]. For the predator-prey model studied here, ergodicity in the system is established by using a numerical scheme to estimate the spectrum of global Lyapunov exponents, which measure the rate at which two infinitesimally separated points in the configuration space move apart over time along the three dimensions present in the system. Simulations with varying timescales that start at various initial conditions on the attractor converge to the same estimates of the Lyapunov exponents, implying ergodicity [45] (S1C Appendix). A similar technique has been used to establish ergodicity in some models of chaotic multitropic food webs [51]. The Lyapunov spectrum can, in turn, be used to determine the Kaplan-Yorke fractal dimension of the attractor, DKY ≈ 2.01, which accounts for the two-dimensional shape of the full attractor (Fig 2B) and linear shapes of its Poincare sections (Figure B in S1E Appendix) discussed above.
Due to the ergodic property of chaotic attractors, one typical interpretation of their appearance in ecological dynamics is that they allow a sort of bet-hedging across timescales, conferring ecological stability against sudden external perturbations [23, 52, 53]. In the presence of external factors not explicitly included in the model, especially non-ergodic processes such as climate variation, a chaotic ecosystem will present a variety of different ratios of predator and prey concentrations at different times, ensuring robustness through biodiversity [54–56]. Moreover, in spatially-extended models in which different subpopulations may simultaneously exist at different points in the chaotic attractor, the chaotic attractor can allow one subgroup to recover from a sudden environmental catastrophe or to expand its range to a new location when favorable conditions spontaneously arise. In general, chaotic dynamics may present an adaptive benefit by making ecological networks robust, for example by preventing sudden exclusion of a keystone species [57].
Here, we suggest that chaos produces an additional effect when it arises due to eco-evolutionary dynamics: it creates a broad distribution of “windows” of time during which sympatric speciation may occur. The dynamics imposed in the predator-prey model in Eqs 10–12 do not explicitly include speciation, which represents an irreversible process in which the prey bifurcates into multiple co-evolving types (hence changing the number of distinct dependent variables present in the dynamics). This would violate the underlying conditions of the gradient dynamics model by creating a bimodal prey density vs trait distribution with substantial skew. However, this process would typically occur during periods of disruptive selection, during which speciation could occur either through assortative mating or through spatial isolation of phenotypically homogeneous subpopulations.
For this reason, the statistical property of the chaotic dynamics that is relevant to speciation is the distribution of time that the fitness function spends in the disruptive state, or the “epochs” of disruptive selection. This represents the distribution of opportunities for speciation to first occur in the system, at which point ergodicity would be broken and the dynamical equations would no longer remain valid. In many models of evolutionary processes, the distribution of epochs of dominance for certain phenotypes has rich statistical structure, including a heavy-tail distribution that some authors have taken to indicate the presence of self-organized criticality [58, 59]. These epochs can be detected by defining the “local” Lyapunov exponents, which represent the three eigenvalues of the Jacobian matrix for the Systems (10), (11) and (12) evaluated at each point along a trajectory in the chaotic attractor [60, 61],
λ i ( t ) ≡ eig ( ∂ x ˙ ( t ) ∂ x ( t ) )
where x = ( x ( t ) , y ( t ) , c ¯ ( t ) ) and i ∈ {1, 2, 3}. Plots of these local Lyapunov exponents during a typical period of metastable dynamics followed by cycling are shown in Fig 5A. Positive values suggest chaotic dynamics, while negative values suggest that nearby trajectories converge. The largest local Lyapunov exponent typically dominates the dynamics. Consistent with the destabilizing nature of disruptive selection, the largest local Lyapunov exponent increases dramatically during periods in which the fitness function has multiple local maxima. For this reason, the length of these long excursions in which the largest local Lyapunov exponent significantly exceeds zero can be used to estimate the distribution lengths of periods of disruptive selection (Fig 5B), based on a very long sample of the dynamics along the strange attractor. The broadness of this distribution suggests that speciation events could occur over a range of timescales in the system (for example, via hybrid breakdown), representing a potential signature of a chaotic past that could be observed in descendant populations with non-chaotic dynamics.
Despite the large fluctuations in the maximum value of the local Lyapunov exponents, the largest global Lyapunov exponent is only barely larger than zero, λmax ≈ 0.003. Similar behavior has been reported a real-world ecosystem consisting of competing species in a rocky intertidal environment, in which a small global Lyapunov exponent paired with a fluctuating largest local Lyapunov exponent was taken to suggest that the ecosystem had adapted to “the edge of chaos.” [51] A similar case has been reported experimentally in populations of voles in Northern Europe that appear to switch between chaotic and stable periods [62]. In that system, it was noted that occasional switches to chaotic dynamics serve to amplify the effect of environmental fluctuations, further suggesting that the irregular spacing of epochs resulting from chaotic dynamics may allow a range of timescales over which speciation may occur under temporally-varying external conditions.
If the underlying assumptions of the gradient dynamics model do not hold—such as V slowly varys in time or the trait distribution retains significant skewness—then the chaotic dynamics would be non-ergodic, causing the system to eventually exit the chaotic attractor and either diverge or settle to a fixed point or limit cycle. If the timescale of exit from the chaotic attractor is much longer than the average time between periods of rapid cycling (as determined, for example, by the peak in the power spectrum in Figure A of S1E Appendix), then the dynamics will demonstrate transient chaos, and the form of the distribution in Fig 5B will be roughly the same due to quasi-ergodicity. However, if the timescale of transience is much shorter, the dynamics may not fully sample the attractor, resulting in the distribution of epochs of disruptive selection being strongly dependent on the initial conditions.
We have shown that a simple two-species predator-prey ecosystem can display rich dynamical complexity when the prey evolves in response to predation, and that this complexity can be understood by analyzing the temporal variation of the fitness landscape. Future theoretical work will establish whether these dynamics qualitatively change when the predator also evolves over timescale comparable to the prey evolution [63]. Such predator-prey co-evolutionary systems have been shown to exhibit a distinct route to chaos, due a desynchronization of the predator and prey adaptation that comprises a form of the “Red Queen” effect [64, 65].
One limitation of our model arises from the form of the evolutionary dynamics Eq (4), which assume that the dynamics of the trait distribution can be adequately described by a mean trait evolution equation. S1D Appendix compares the results found here to those generated by a formulation of the problem in terms of a full integro-differential equation, and finds general agreement for the fitness landscape studied here. However, for more complicated fitness landscapes these conditions may not hold, requiring more advanced models that introduce additional dynamical equations to account for various effects such as non-constant additive genetic variance [30, 66, 67]. In such models, chaos may appear as a transient in the dynamics before the dynamical variables approach an equilibrium point or limit cycle.
Our findings for the minimal model studied here have implications for a wide variety of eco-evolutionary systems, because they suggest that even a minimal deterministic model can exhibit unstable cycling and chaos—effects that would typically become more pronounced when more species are added to the system [22, 23]. The mechanism by which chaos appears in our system is generic, resulting purely from changes in the number of local maxima in the fitness landscape, suggesting the applicability of our findings to observational systems (such as bacteria and viruses in microenvironments) in which the fitness landscape can be monitored, but not necessarily all of the underlying species interrelationships [68]. For these systems, recent advances in genetic barcoding of entire microbial communities [69, 70] may allow direct observation of the role of dynamic fitness landscapes in creating opportunities for sympatric speciation.
In addition to being an emergent property of the underlying species interactions, we suggest here that these chaotic properties may confer adaptive benefits via community robustness, either by enforcing phenotypic diversity or by preventing environmental variation from fully excluding a single species. The system described here also represents an example of a small ecosystem that adapts towards the “edge of chaos”, which can further adjust how the system responds to external perturbations [71, 72]. Potential experimental systems in which the adaptive role of eco-evolutionary chaos may be explored include phytoplanktonic ecosystems, which can be isolated in the laboratory and which are known to to maintain biodiversity using chaotic effects [54]. In particular, it would be interesting to determine whether non-synchronized replicates of experimentally-controlled chaotic ecosystems could recover from a synchronized perturbation (i.e. temporary salinity shock) more quickly than non-chaotic controls [74]—suggesting that the ability of chaotic systems to continuously sample a wide variety of dynamical conditions confers robustness. In these systems, reconstruction of of an experimental chaotic attractor derived from lagged coordinate embedding [43] could yield insight into whether chaos arises due to changes in the general topology of the fitness landscape, which would result in a nearly two-dimensional attractor due to distinct timescales associated with stabilizing and disruptive effects.
Moreover, the underlying cause of the chaotic dynamics—a reversible transition between stabilizing and disruptive selection—is mathematically analogous to the change in the shape of the free energy landscape during a first-order phase transition in thermodynamics. Our findings thus fit within more general extensions of mathematical theories of evolution that include formalism from statistical and condensed matter physics [8, 58, 72, 73], suggesting that universal mechanisms may underly subtle transient properties observed in many natural ecosystems, including hysteresis and dynamical robustness [75].
|
10.1371/journal.pntd.0003356 | Immunizing Adult Female Mice with a TcpA-A2-CTB Chimera Provides a High Level of Protection for Their Pups in the Infant Mouse Model of Cholera | Vibrio cholerae expresses two primary virulence factors, cholera toxin (CT) and the toxin-coregulated pilus (TCP). CT causes profuse watery diarrhea, and TCP (composed of repeating copies of the major pilin TcpA) is required for intestinal colonization by V. cholerae. Antibodies to CT or TcpA can protect against cholera in animal models. We developed a TcpA holotoxin-like chimera (TcpA-A2-CTB) to elicit both anti-TcpA and anti-CTB antibodies and evaluated its immunogenicity and protective efficacy in the infant mouse model of cholera. Adult female CD-1 mice were immunized intraperitoneally three times with the TcpA-A2-CTB chimera and compared with similar groups immunized with a TcpA+CTB mixture, TcpA alone, TcpA with Salmonella typhimurium flagellin subunit FliC as adjuvant, or CTB alone. Blood and fecal samples were analyzed for antigen-specific IgG or IgA, respectively, using quantitative ELISA. Immunized females were mated; their reared offspring were challenged orogastrically with 10 or 20 LD50 of V. cholerae El Tor N16961; and vaccine efficacy was assessed by survival of the challenged pups at 48 hrs. All pups from dams immunized with the TcpA-A2-CTB chimera or the TcpA+CTB mixture survived at both challenge doses. In contrast, no pups from dams immunized with TcpA+FliC or CTB alone survived at the 20 LD50 challenge dose, although the anti-TcpA or anti-CTB antibody level elicited by these immunizations was comparable to the corresponding antibody level achieved by immunization with TcpA-A2-CTB or TcpA+CTB. Taken together, these findings comprise strong preliminary evidence for synergistic action between anti-TcpA and anti-CTB antibodies in protecting mice against cholera. Weight loss analysis showed that only immunization of dams with TcpA-A2-CTB chimera or TcpA+CTB mixture protected their pups against excess weight loss from severe diarrhea. These data support the concept of including both TcpA and CTB as immunogens in development of an effective multivalent subunit vaccine against V. cholerae.
| Vibrio cholerae is the bacterium that causes cholera, a pandemic diarrheal disease transmitted by ingestion of contaminated food or water. We developed a novel vaccine containing two protective antigens of V. cholerae, TcpA and CTB, incorporated into a defined oligomeric protein chimera. CTB is the non-toxic binding domain of cholera toxin, the protein that causes profuse watery diarrhea in cholera patients. TcpA is the subunit of the toxin-coregulated pilus, a V. cholerae surface structure that is required for intestinal colonization and disease. Intraperitoneal immunization of adult female mice with this TcpA-A2-CTB chimera elicited stronger early anti-TcpA responses and equivalent anti-CTB responses compared to immunizing with a TcpA+CTB mixture. Furthermore, all reared infant mice from females immunized with the chimera or TcpA+CTB were protected against a large challenge dose of V. cholerae that was sufficient to kill all infant mice from non-immunized control and TcpA- or CTB-immunized adults. Our study supports the concept of including both TcpA and CTB as antigens in development of a safe and effective subunit vaccine against cholera.
| Cholera is an intestinal infection that is associated with acute watery diarrhea and is caused by the Gram-negative bacillus Vibrio cholerae. Cholera is spread by the ingestion of contaminated food and water. An estimated 3–5 million people are infected yearly with cholera, resulting in approximately 100,000 deaths [1]. Cholera is endemic in over 50 countries in the developing world where risk factors such as over-crowding, lack of clean food and water, and poor sanitation allow for its persistence in the environment [1], [2]. Cholera can cause severe life-threatening dehydration, and stool outputs as high as 500–1000 ml/hr can rapidly lead to death in untreated patients [2]. The most effective treatment for cholera is rehydration therapy, and if treatment is started early enough the case fatality rate (CFR) is below 1% [3]. However, it is often difficult for poor and impoverished patients to have access to medical treatment. Cholera can be prevented by vaccination. In 2011, a review of published studies on five variants of an oral whole-cell killed (WCK) cholera vaccine showed that their overall protective efficacy after two years was 62% in adults, they were less effective in children under 5 years of age, and they were unlikely to provide protection beyond three years [4]. In 2013, a study of a re-formulated WCK oral cholera vaccine in Kolkata, India, showed a 5-year cumulative protective efficacy of 65% in all individuals over 1 year of age, but a lower 5-year protective efficacy of 42% in children from 1 to 5 years of age [5]. Nevertheless, a recent critical analysis concludes that current WCK cholera vaccines are poorly suited to control endemic or epidemic cholera because of limited efficacy in young children, requirements for multiple doses, a cold chain, and complex delivery logistics, and costs that are high for resource-poor regions [6]. Finding solutions for such issues is an important goal for developing improved cholera vaccines.
We are investigating development of safe and effective subunit vaccines against cholera. Subunit vaccines can present important virulence determinants such as colonization factors and toxins that might not be present or highly immunogenic in a WCK or living attenuated vaccine. For Vibrio cholerae, two essential virulence factors are cholera toxin (CT) and the toxin-coregulated pilus (TCP). CT is an AB5 toxin that is primarily responsible for diarrhea in cholera. CT consists of a monomeric A subunit (CTA) and a homopentameric B subunit (CTB) [7], binds to monosialosyl ganglioside GM1 receptors on enterocytes [8], enters them by endocytosis, trafficks to the endoplasmic reticulum, and releases its CTA1 fragment for retrotranslocation into the cytosol [9]. In the cytosol, CTA1 ADP ribosylates the α subunit of Gs (Gsα), resulting in activation of adenylate cyclase, accumulation of intracellular adenosine-3,5-cyclic monophosphate (cAMP), and downstream events including an efflux of ions and water into the intestinal lumen that presents clinically as diarrhea [10], [11].
TCP is a type IV pilus composed of repeating subunits of the major pilin subunit TcpA [12]. TCP functions in vivo by mediating bacterium-bacterium interactions that are essential for the formation of microcolonies on the surface of enterocytes in the small intestine [13], [14]. In recent infant mouse experiments, TCP has also been demonstrated to mediate attachment of V. cholerae to epithelial cells and to form TCP matrices that engulf the bacteria and may help to protect them from antimicrobial agents [14].
The importance of CT and TCP for V. cholerae virulence has been demonstrated both in animals and in humans, as strains of V. cholerae that fail to produce either CT or TcpA are severely attenuated [12], [15]–[17]. Immunization with CT or non-toxic derivatives of CT has been shown to elicit protective immunity in animal models but not in humans [18]–[21]. Passive orogastric administration of anti-TCP antibodies can provide excellent protection in the infant mouse model of cholera [22], [23], but immunization of humans with intact TCP or with TcpA subunits has not yet been investigated.
In the study reported here, we tested recombinant forms of TcpA and CTB (either alone, in combination, or as a holotoxin-like chimera) as candidate cholera vaccines in the infant mouse model of cholera.
All procedures involving experimental animals were approved by the University of Colorado Denver (UCD) Animal Care and Use Committee. These procedures were done in compliance with all institutional and governmental requirements and regulations regarding the appropriate ethical use of animals in research. UCD is accredited by the Association for the Assessment and Accreditation of Laboratory Animal Care, International (file number 00235).
All genes were PCR amplified using genomic DNA from V. cholerae El Tor strain N16961 and for FliC from genomic DNA from Salmonella typhimurium strain 14028s. The TcpA-A2-CTB chimera dual promoter expression plasmid pGAP31-2XT7 was constructed in several steps. First, the a2 gene fragment encoding CTA2 was amplified by PCR using the forward primer oA2-Fnot and the reverse primer oA2-Rxho containing the NotI and XhoI restriction sites respectively (Table 1; restriction sites shown in bold). The primer oA2-Fnot contained a point mutation in the a2 coding sequence to change a cysteine to a serine (Table 1; point mutation underlined). Second, the tcpA gene fragment encoding residues 29–199 of the mature TcpA polypeptide was PCR amplified using the forward primer oTcpAn16961-Fmsc and the reverse primer oTcpAn16961-Rnot containing the MscI and NotI restriction sites respectively (Table 1). Previous studies demonstrated this polypeptide to be soluble, surface-exposed, and immunogenic [24], [25]. Third, the a2 and tcpA genes were subcloned into an altered pET22b(+)[EMB Biosciences, Gibbstown, NJ] expression plasmid in which the ampicillin resistance marker was replaced with a kanamycin resistance marker. The kanamycin resistance marker was obtained from pET28b(+) [EMD Biosciences, Gibbstown, NJ] which was cut with EcoRI and PpuMI, and the isolated restriction fragment was then ligated into pET22b(+). The insertion of the tcpA and a2 gene fragments was downstream and in frame with the pelB signal sequence. Fourth, a second T7 promoter containing the mature ctb gene in frame with the pelB signal sequence was PCR amplified from the expression plasmid pGAP20K [26] using the forward primer oT7-FppuMI and the reverse primer oCTB-RpshAI. Finally, the t7-pelB-ctb gene product was subcloned into the PshAI and PpuMI sites of the TcpA-A2 expression plasmid generated in step 3 above, thereby creating the dual T7 promoter expression plasmid pGAP31-2XT7 (Fig. 1).
The CTB expression plasmid pGAP20K, which encodes the ctxB allele from V. cholerae El Tor strain N16961, was constructed as previously described [26]. The N-terminal 6-histidine tagged-TcpA expression plasmid was created by PCR amplifying the tcpA gene fragment encoding residues 29–199 of the mature TcpA polypeptide (Fig. 1) using the forward primer oTcpAn16961-Fnde and the reverse primer oTcpAn16961-Rxho (Table 1). This was inserted into pET28b(+) using the NdeI and XhoI restriction sites, downstream and in frame of an N-terminal 6-his tag, creating the expression plasmid pGAP33.
The N-terminal 6-histidine-tagged-FliC expression plasmid was created by PCR amplifying the Salmonella typhimurium fliC gene using the forward primer oFliC-Fnde and the reverse primer oFliC-Rxho (Table 1). This was inserted into pET28b(+) using the NdeI and XhoI restriction sites, downstream and in frame with the N-terminal 6-his tag, creating the expression plasmid pGAP32.
The TcpA-A2-CTB chimera was produced in Escherichia coli BL-21(DE3) Star™ cells (Invitrogen, Grand Island, NY). Half-liter cultures were grown in NZTCYM medium pH 7.5 (1% N-Z-amine AS [Sigma, St. Louis, MO], tryptone 1%, NaCl 0.5%, yeast extract 0.5%, casamino acids 0.1%, MgSO4 0.2%) and 100 µg/ml kanamycin at 37°C, 250 rpm until cultures reached an OD600 of ∼3.0. The cultures were then placed at 16°C and 250 rpm for 30 minutes to acclimate to the new temperature then induced with 0.2 mM IPTG and grown overnight at 16°C for ∼16–18 hrs. 6His-TcpA(29–199) was produced in SHuffle™ T7 Express E. coli (NEB, Ipswich, MA) in half liter cultures of TCYM media pH 7.5 (tryptone 1%, NaCl 0.5%, yeast extract 0.5%, casamino acids 0.1%, MgSO4 0.2%) and 100 µg/ml kanamycin at 37°C, 250 rpm until cultures reached an OD600 of ∼3.0. The cultures were then placed at 16°C, 250 rpm for 30 minutes, and then induced with 0.1 mM IPTG and grown overnight as above. Cultures of E. coli BL-21(DE3) cells producing 6His-FliC were grown in half liter cultures of TCYM pH 7.5 at 37°C 250 rpm until they reached an OD600 of ∼2.0–3.0. After acclimating to 30°C with shaking at 250 rpm for 30 minutes, the cultures were then induced with 0.5 mM IPTG and incubated for 4 hrs. CTB was grown and induced as described previously [26].
Preparation of the bacterial extracts and primary metal affinity purification of all proteins was performed as described previously [26]. A secondary purification step on the strong cation-exchange resin POROS 20 HS (Applied Biosystems, Carlsbad, CA) was performed for TcpA-A2-CTB, 6His-TcpA(29–199), and CTB. Both the TcpA-A2-CTB chimera and 6His-TcpA(29–199) were purified under the same conditions. Each was dialyzed overnight at 4°C in 25 mM potassium phosphate buffer pH 6.8. Soluble and filtered protein was loaded onto a POROS 20 HS (Applied Biosystems, Carlsbad, CA) column, and eluted with a linear 0 to 0.5 M gradient of NaCl in 25 mM potassium phosphate buffer at pH 6.8. This purified, soluble, recombinant 6His-TcpA(29–199) protein is subsequently called TcpA.
For CTB, the protein was first desalted using Zeba™ Desalt Spin Columns (Thermo Fisher Scientific, Rockford, IL) following the manufacturer's protocol. CTB was desalted into potassium phosphate buffer pH 6.6 and then filtered through a 0.45 µM syringe filter to remove precipitated material. An ion-exchange purification step was then conducted using POROS 20 HS (Applied Biosystems, Carlsbad, CA) resin. The bound protein was eluted using a linear 0 to 0.5 M gradient of NaCl in 25 mM potassium phosphate buffer pH 6.6.
6His-FliC was dialyzed overnight against 20 mM Tris-Cl pH 8.0. An ion-exchange purification step was then conducted using the strong anion-exchange resin POROS 20 HQ (Applied Biosystems, Carlsbad, CA). The bound protein was eluted from the resin using a linear gradient of 0 to 0.5 M NaCl in 20 mM Tris-Cl buffer at pH 8.0. This purified, soluble, recombinant 6His-FliC protein is subsequently called FliC. Following purification TcpA, CTB, and FliC were dialyzed overnight at 4°C against 1× PBS and stored at −80°C. The TcpA-A2-CTB chimera was dialyzed overnight at 4°C against 50 mM Tris buffer containing 200 mM NaCl and 1 mM EDTA at pH 7.5.
Female CD-1 mice, 6–8 weeks old, were purchased from Charles River Labs and given food and water ad libitum. Groups of 7–10 mice were immunized three times IP at 14 day intervals. The group that was immunized with the TcpA-A2-CTB chimera received 50 µg/dose. All other groups received amounts of each antigen that were equimolar with the amount of the corresponding antigenic component in a 50-µg dose of the chimera. For the groups immunized with TcpA combined with FliC or FliC alone, the dose of FliC administered for the first or second immunization was 5 µg, and a 2 µg dose was administered for the final immunization. Blood and fecal samples were collected one day prior to the initial immunization (Day −1) and on days 21 and 42. Sample collection and processing was performed as previously described [26]. To obtain blood samples from infant mice, the pups were first asphyxiated with CO2, and then a scalpel was used to sever the cervical spinal cord. A heparinized capillary tube was used to collect blood that seeped from the incision. One or two pups were used per dam, and blood was collected on the same day that siblings were challenged with V. cholerae. Blood was pooled if two siblings were used to obtain sera.
To measure antigen-specific antibody amounts in serum and fecal extracts, we used quantitative ELISAs as described previously [26]. The concentration or amount of antigen-specific IgG or IgA antibodies in unknown samples was determined by interpolation from a standard curve using KC4 v3.4 software (Bio-Tek. Winooski, VT). GM1 ganglioside ELISAs were performed as previously described [26] using sera from rabbits immunized with recombinant TcpA or CTB.
The infant mouse challenges were performed as previously described [21], [26]. All pups were six days old at the time of inoculation. The pups were monitored for survival over the course of 48 hrs. Pup weights were recorded immediately prior to inoculation and at 24 and 48 hours post-infection. For pups that died prior to 24 hours, their carcass weights were measured and included with the group at 24 hours. For pups that died between 24 and 48 hours, their carcass weights were measured and included with the group at 48-hours.
All statistical comparisons were performed using GraphPad PRISM 4 (La Jolla, CA). ANOVA analysis using the Tukey-Kramer post-test was used to determine statistical significance between immunization groups for antigen-specific antibody concentration differences and for weight loss differences between immunization groups for the infant mouse challenge. Within-group statistical differences for antigen-specific antibody amounts at days 21 and 42 were analyzed using a paired two-tailed t-test. Survival curves were generated using Kaplan-Meier method, and statistical differences between experimental groups were determined using the log-rank (Mantel-Cox) test. P values less than 0.05 were considered significant.
Both gene and protein sequences for the antigens used in this study can be found in the National Center for Biotechnology Information (NCBI) database. The genes (accession numbers) are as follows: ctxB (NC_002505), tcpA (AF536868), and fliC (NC_016856). The protein sequences (accession numbers) are as follows: CTB (NP_231099), TcpA (AAN15109) and FliC (YP_005237927).
The TcpA-A2-CTB chimera was expressed in E. coli using the dual T7 promoter plasmid pGAP31-2XT7 (Fig. 1), and it was purified using sequential metal affinity chromatography and ion-exchange chromatography. Upon heating and denaturation, the purified chimera separated into the TcpA-A2 fusion protein (∼23 kDa) and monomeric CTB (∼11.5 kDa; Fig. 2). TcpA-A2 migrated more slowly than TcpA, reflecting the greater molecular mass of the fusion protein due to the presence of CTA2 at its carboxyl-terminus (Fig. 2). We used GM1 ganglioside ELISAs to demonstrate immunoreactivity of the TcpA-A2 moiety and both immunoreactivity and ganglioside GM1 receptor-binding activity of the pentameric CTB moiety of the purified TcpA-A2-CTB chimera (Fig. 3). Solutions containing equimolar amounts of the TcpA-A2-CTB chimera, or TcpA alone, or CTB alone, were serially diluted and added to ELISA plates that had previously been coated with ganglioside GM1 and then blocked to prevent nonspecific binding of the test antigens. Subsequently the plates were washed and then probed with either anti-CTB rabbit antiserum or anti-TcpA rabbit antiserum. The TcpA-A2-CTB chimera and CTB (but not TcpA) bound avidly to the plates coated with GM1 ganglioside. Bound TcpA-A2-CTB chimera and bound CTB were both detected with anti-CTB antiserum (Fig. 3, top), but only the bound chimera was detected with anti-TcpA antiserum (Fig. 3, bottom). None of these antigens bound above background levels to control plates that were blocked but not coated with ganglioside GM1. Taken together, these results showed that the TcpA-A2-CTB chimera is a bi-functional oligomeric complex that exhibits TcpA immunoreactivity associated with its TcpA-A2 fusion polypeptide and both CTB immunoreactivity and ganglioside GM1 binding activity associated with its pentameric CTB subunit.
To compare the immunogenicity of the TcpA-A2-CTB chimera with non-chimeric forms of TcpA and CTB, we immunized groups of 7 to 10 female CD-1 mice three times by the IP route according to the immunization timeline shown in figure 4. In an attempt to show whether inherent immunogenicity of recombinant TcpA protein could be enhanced by use of an adjuvant, we included separate groups of mice immunized either with TcpA alone or with TcpA plus the recombinant flagellin subunit protein FliC from Salmonella typhimurium. FliC is a toll receptor 5 (TLR5) agonist and has been demonstrated previously to act as an adjuvant for co-administered antigens [27]. Serum and fecal antigen-specific antibody responses were measured using quantitative ELISA for the samples collected on days 21 and 42 (Fig. 4). We found that immunization with the TcpA-A2-CTB chimera elicited a significantly higher mean concentration of serum anti-TcpA IgG on day 21 compared with all other immunization groups (P<0.001, Fig. 5A). However by day 42, 14 days following the third and final immunization, the mean serum anti-TcpA IgG concentrations for the groups immunized with the TcpA-A2-CTB chimera, the TcpA+CTB mixture, and the TcpA+FliC mixture were comparable (P>0.05), but all were significantly greater than the mean serum anti-TcpA IgG concentration after immunization with TcpA alone (P<0.01). These results demonstrated that either incorporating the TcpA-A2 fusion protein into the TcpA-A2-CTB chimera or administering TcpA together with CTB or FliC enhanced the immunogenicity of TcpA, and the results for the samples collected at day 21 suggest that the TcpA-A2-CTB chimera presented the TcpA moiety in its most immunogenic form. Serum anti-CTB IgG responses were robust in all groups that received CTB, either as CTB alone, as TcpA+CTB, or as the TcpA-A2-CTB chimera (Fig. 5B). There were no significant differences in mean anti-CTB IgG concentrations between these groups either at day 21 (P>0.05) or at day 42 (P>0.05), although the mean titers were significantly higher at day 42 than at day 21 (P<0.0001). As expected, control mice immunized with PBS did not develop any detectable anti-TcpA or anti-CTB antibodies.
The amounts of antigen-specific IgA antibody and total IgA immunoglobulin were measured in each fecal extract. In figure 6, the amount of antigen-specific IgA antibody is shown as a percentage of total IgA for each fecal extract. We normalized the data in this way to minimize differences that might result from mouse-to-mouse variations in production of fecal IgA or sample-to-sample variations in recovery of IgA from the fecal specimens. Immunization with either the TcpA-A2-CTB chimera or the TcpA+FliC mixture elicited a significantly greater mean fecal anti-TcpA IgA response on day 21 than immunization with any of the other antigens (Fig. 6A, P<0.05). Interestingly, at day 42 the mean fecal anti-TcpA IgA response to immunization with the TcpA-A2-CTB chimera was less than at day 21, but the mean fecal anti-TcpA IgA response to immunization with TcpA+FliC was greater than at day 21. However, neither of these pairwise differences in mean antibody amounts between days 21 and 42 was significant (P>0.05). On day 42 the mean fecal anti-TcpA IgA response to immunization with TcpA+CTB had increased dramatically to a level that was comparable to the TcpA+FliC immunization group (P>0.05), and both were significantly greater than mean values for all other immunization groups at day 42 (P<0.05) (Fig. 6A).
As with the serum CTB-specific IgG responses, the fecal anti-CTB IgA responses at day 42 were not significantly different in any of the groups that received CTB as an immunogen, either as CTB alone, as TcpA+CTB, or as the TcpA-A2-CTB chimera (Fig. 6B; P>0.05). In contrast to the increases in the serum CTB IgG concentrations that occurred from day 21 to day 42, however, the fecal CTB-specific IgA percentages on day 21 and day 42 were comparable (P>0.05). Finally, for each group that was immunized both with TcpA and CTB (e.g., immunized with either TcpA-A2-CTB or TcpA+CTB), the mean percentage of fecal anti-CTB IgA was greater than the mean percentage of fecal anti-TcpA IgA in the same group, both at day 21 and at day 42 (compare results and note differences in scales for the Y axes for Figs. 6A and 6B).
To compare the protective efficacies of selected vaccine regimens, immunized dams were mated (see timeline in Fig. 4) and groups of their reared 6-day old pups were challenged orogastrically with 10 LD50 of V. cholerae El Tor strain N16961 and monitored for survival at 24 and 48 hrs (Fig. 7A). All pups from dams immunized either with TcpA-A2-CTB chimera (n = 20) or TcpA+CTB (n = 20), and all sham-infected pups (n = 20) survived for 48 hrs. All pups from dams immunized with CTB alone (n = 20) survived for 24 hrs, and 70% survived for 48 hrs (P = 0.0087 vs. pups immunized with TcpA-A2-CTB or TcpA+CTB). In contrast, pups from dams immunized with TcpA+FliC (n = 20) or FliC alone (n = 10), like pups from PBS immunized dams (n = 20), experienced 77.5–80% mortality by 24 hrs and 100% mortality by 48 hrs (P<0.0001 vs. pups immunized with TcpA-A2-CTB or TcpA+CTB). We did not challenge pups from dams immunized with TcpA alone, because those dams were previously shown to have much lower serum and fecal anti-TcpA antibody levels than dams immunized with TcpA+FliC (Figs. 5A and 6A).
To investigate under more stringent conditions the contributions of anti-TcpA and anti-CTB antibodies in protecting infant mice against cholera, we challenged additional pups from the immunized dams with a higher 20 LD50 challenge dose of V. cholerae El Tor strain N16961 (Fig. 7B). All pups from dams immunized with TcpA-A2-CTB chimera (n = 19) or TcpA+CTB (n = 20), and all sham-infected pups (n = 20), survived for 48 hrs. In contrast, no pups from dams immunized with CTB alone (n = 20), TcpA+FliC (n = 20), or PBS (n = 20), survived for 48 hrs (P<0.001 vs each of the three previous groups). The mean concentrations of serum anti-TcpA IgG at day 42 did not differ significantly in dams immunized with TcpA-A2-CTB chimera, TcpA+CTB, or TcpA+FliC (Fig. 5A, P>0.05), and the mean percentages of fecal anti-TcpA IgA also did not differ significantly in the dams immunized with TcpA+CTB or TcpA+FliC (Fig. 6A, P>0.05). Similarly, the mean concentrations of serum anti-CTB IgG at day 42 did not differ significantly in dams immunized with TcpA-A2-CTB chimera, TcpA+CTB, or CTB alone (Fig. 5B), and the mean percentages of fecal anti-CTB IgA did not differ significantly among dams in these immunization groups (Fig. 6B). Published studies show that transfer of maternal antibodies to pups (which can occur in utero, by suckling, or by both pathways) is the primary mechanism by which immunization of dams confers immunologically specific protection to their pups [28]–[30]. Therefore, the complete protection achieved in pups from dams immunized with the TcpA-A2-CTB chimera or the TcpA+CTB mixture, versus the lack of any protection in pups from dams immunized with TcpA+FliC or CTB alone, cannot be explained either by poorer serum or fecal anti-TcpA or anti-CTB antibody responses, respectively, vs. the comparable antigen-specific antibody responses in the mice immunized with the TcpA-A2-CTB chimera or the TcpA+CTB mixture.
Sham-infected pups experienced about 2% weight loss at 24 hrs and 10% weight loss at 48 hrs because they were separated from their dams since 3 hrs before challenge (Fig. 8). In addition, mouse pups develop diarrhea if they are not fully protected against V. cholerae infection by active or passive immunity [31]. At the 10 LD50 challenge dose of V. cholerae N16961, pups from dams immunized with the TcpA-A2-CTB chimera or TcpA+CTB mixture did not lose significantly more weight at 24 or 48 hrs than sham-infected pups (Figs. 8A and 8B; P>0.05), and all survived for 48 hrs (Fig. 7A). In contrast, pups from PBS immunized dams or dams immunized with TcpA+FliC or FliC alone did experience much greater weight losses than sham-infected pups both at 24 hrs (Fig. 8A; P<0.001) and at 48 hrs (Fig. 8B; P<0.05), and all died by 48 hrs (Figs. 7A and 7B). The severity and timing of their excess weight losses and their death within 48 hrs reflected the onset of severe diarrhea. Pups from dams immunized with CTB experienced less-dramatic excess weight losses at 24 hours than pups from dams immunized with PBS, TcpA+FliC, or FliC (Fig. 8A; P<0.001), but they experienced greater weight losses than sham-infected pups at 24 and 48 hrs (Figs. 8A and 8B; P<0.001). This resulted in a 70% survival rate of CTB immunized pups at 48 hrs. These findings indicate that pups from dams immunized with CTB were partially protected against challenge with 10 LD50 of V. cholerae N16961, and they experienced less severe diarrhea than PBS immunized pups (Figs. 7A and 7B).
At the 20 LD50 challenge dose, pups from dams immunized with the TcpA-A2-CTB chimera exhibited significantly greater weight losses at 24 hrs than sham-infected control pups (Fig. 8C; P<0.05), and pups from dams immunized either with the TcpA-A2-CTB chimera or with TcpA+CTB exhibited significantly greater weight losses at 48 hrs than sham-infected control pups (Fig. 8D, P<0.05) although all pups in both groups survived for 48 hrs. The excess weight losses among pups in these two groups indicate that they experienced mild diarrhea at the 20 LD50 challenge dose. The challenged pups from dams immunized with TcpA+FliC or CTB alone experienced much greater weight losses than the sham-infected controls (P<0.001; Figs. 8C+8D), and all of them died before 48 hrs, indicating that they experienced severe diarrhea at the 20 LD50 challenge dose.
One or two pups from each dam in the immunization groups included in the suckling mouse challenge studies described above were sacrificed at the same time that their siblings were challenged to obtain serum samples for measurements of antigen-specific IgG concentrations by quantitative ELISA. As shown in Fig. 9, the mean anti-TcpA IgG serum antibody concentrations from the pups were statistically equivalent regardless of whether their dams were immunized with TcpA-A2-CTB chimera, TcpA+CTB, or Tcp+FliC (P>0.05). The mean anti-CTB IgG serum antibody concentrations from the pups were also statistically equivalent regardless of whether their dams were immunized with TcpA-A2-CTB chimera, TcpA+CTB, or CTB (P>0.05). Consistent with results shown previously for serum antibodies at day 42 in immunized dams, the mean concentrations of antigenic-specific serum IgG antibodies from these pups were much greater for the anti-CTB antibodies than for anti-TcpA antibodies. None of the sera from pups had detectable anti-TcpA-specific IgA antibodies, and only one serum from a pup born to a dam immunized with TcpA-A2-CTB chimera had detectable anti-CTB-specific IgA antibodies.
Early studies showed that pups from non-immunized dams survived large orogastric challenge doses of V. cholerae (500–2000 LD50) when the bacteria were pre-mixed with anti-CT or anti-TCP antiserum [23], [32], but pups from dams immunized against TcpA or CTB survived only when challenge doses were much smaller (1–15 LD50) [21], [26], [33]. Titers of serum anti-TcpA IgG1 and IgA antibodies in dams correlated with survival rates of their challenged pups [33]. Survival rates of challenged pups from immunized dams fell more rapidly as the dam's log10 anti-TcpA IgG1 titers decreased than did survival rates of pups from unimmunized dams with comparable decreases in anti-TCP antiserum doses [33]. Pups given anti-TCP antiserum intraperitoneally also survived V. cholerae challenges given 24 hrs later [23]. Taken together, these findings show that intestinal anti-TcpA or anti-CTB antibodies protect infant mice from potentially lethal V. cholerae challenges and indicate that maternal antibodies are delivered into the intestines of infant mice either actively by suckling or passively by transudation from internal body fluids.
In the studies reported here, we investigated whether immunizing dams with TcpA-A2-CTB chimera or TcpA+CTB protected their pups more effectively than immunizing dams with TcpA or CTB alone in the infant mouse model of cholera. We challenged separate groups of pups with 10 LD50 and 20 LD50 doses of V. cholerae El Tor N16961 to assess protection under stringent conditions.
Few previous studies compared protective efficacy of immunization with TcpA+CT (or CTB) vs. TcpA or CT (or CTB) alone in animal models of cholera. Transcutaneous immunization (TCI) of dams with CT+TcpA protected pups better against a 1 LD50 challenge with V. cholerae N16961 (69% survival) than did TCI with CT alone (36% survival), but TCI with TcpA alone induced no detectable anti-TcpA antibodies in dams and their pups were not challenged [25]. Although that study and our study used different methods to measure antigen-specific antibodies, the more robust protection of pups at higher challenge doses that we observed likely indicates higher serum anti-TcpA and anti-CTB antibody levels in our immunized dams. Other investigators used the ligated ileal segment model in adult rabbits to compare protection conferred by intranasal immunization with TcpA+CTB, TcpA alone, or CTB alone [34]. Fluid accumulation in ligated ileal segments decreased by 41.1% vs. unimmunized controls in rabbits immunized with TcpA and by 70.5% in rabbits immunized with CTB, but no fluid accumulated in ligated ileal segments of rabbits immunized with TcpA+CTB [34]. The immunized rabbits also developed intestinal sIgA antibodies against the TcpA and/or CTB antigens that they received [34].
In our studies (see Fig. 7), all pups from dams immunized with TcpA-A2-CTB chimera or TcpA+CTB survived 48 hrs after a 10 or 20 LD50 challenge dose of V. cholerae El Tor N16961; no pups from dams immunized with TcpA+FliC survived 48 hrs at either challenge dose; and pups from dams immunized with CTB had 70% 48-hr survival at the 10 LD50 challenge dose a 0% 48-hr survival at the 20 LD50 dose. At each challenge dose, the 100% survival rate for pups with both anti-TcpA and anti-CTB antibodies significantly exceeded the sum of the survival rates for pups with only anti-TcpA antibodies and for pups with only anti-CTB antibodies [e.g., (0%+70%) = 70% cumulative survival at 10 LD50 and (0%+0%) = 0% cumulative survival at 20 LD50 among pups with only anti-TcpA antibodies plus pups with only anti-CTB antibodies]. Because these differences in survival could not be explained by significant differences in mean values of TcpA-specific or CTB-specific serum IgG or fecal IgA antibodies among groups of pups from dams immunized with vaccine formulations that contained any form of TcpA or CTB, respectively (see Figs. 5 and 6), our results constitute strong preliminary evidence that anti-TcpA and anti-CTB antibodies act synergistically rather than additively to prevent death in the infant mouse model of cholera.
In humans, cholera is caused either by V. cholerae serogroup O1 (with classical and El Tor biotypes and Inaba and Ogawa serotypes) or V. cholerae serogroup O139 (first recognized in 1992–1993) [2], [3]. Early clinical isolates of V. cholerae O139 were closely related to V. cholerae O1 El Tor (but with different genes at the O antigen locus), but later V. cholerae O139 isolates belong to multiple lineages derived from different V. cholerae progenitors [35], [36]. The O1 and O139 lipopolysaccharides are essential for virulence of V. cholerae O1 and O139 in humans, are important protective antigens, and elicit serogroup-specific antibodies that do not cross-react with each other [36]–[38]. CT and TcpA are also essential for virulence of V. cholerae O1 and O139 and are immunogenic in humans [16], [17], [39]–[41], but analyzing their roles in protective immunity is complicated by the existence of multiple antigenically cross-reacting variants of each protein among classical, El Tor and “hybrid El Tor” isolates of V. cholerae O1 and V. cholerae O139 [22], [23], [26], [42]–[44].
Early studies in human volunteers suggested that immunity against cholera was mediated primarily by antibacterial rather than antitoxic mechanisms [20], and for decades the best serological (but non-mechanistic) correlate of protection among patients and volunteers who recovered from a previous episode of cholera was the titer of complement-dependent vibriocidal antibodies [45]. More recent studies in humans showed that serum IgA (but not IgG) antibodies against CTB, LPS, or TcpA also correlate with protection against cholera [3], [46], [47]. However, because recovery confers protection against a future episode of cholera that persists much longer than titers of vibriocidal, anti-CTB, anti-LPS, or anti-TcpA antibodies remain elevated, prompt anamnestic antibody responses following exposure to V. cholerae are believed to be important for long-term immunity against cholera. Consistent with this view, patients who recover from cholera have been found to develop IgG and IgA memory B (BM) cells specific for LPS, TcpA, and CTB as well as effector memory T (TEM) cells specific for CTB [45]–[48].
The mechanisms by which intestinal anti-CTB and anti-TcpA antibodies protect against cholera (e.g., blocking CT-mediated toxicity and TCP-mediated contributions to colonization) are believed to be similar in humans and in infant mice. Our results provide proof-of-principle that immunizing dams with TcpA-A2-CTB chimera or TcpA+CTB can protect 100% of pups against challenges with up to 20 LD50 of V. cholerae El Tor N16961, which produces CTB and TcpA variants homologous to those used for immunization. To the best of our knowledge, no other reported immunization regimen for dams protects pups so well against such a stringent challenge in the infant mouse model of cholera. Further studies will be needed: 1) to assess the relative protective efficacy of current TcpA-A2-CTB and TcpA+CTB vaccines against challenges with V. cholerae O1 classical or El Tor or V. cholerae O139 strains that produce homologous or heterologous variants of CTB and TcpA; and 2) to determine whether immunization with at least two variants each of TcpA and CTB can provide broader protection than immunization with one variant of each protein against V. cholerae challenge strains that produce several different CTB and/or TcpA variants. Using a TcpA-A2-CTB chimera instead of a TcpA+CTB mixture in a vaccine formulation has several potential advantages, since the chimera is a chemically defined, highly immunogenic, macromolecular complex that can be assembled spontaneously in E. coli and purified as a single entity. If necessary, different variants of TcpA and CTB could be incorporated into different chimeras, which could then be combined to create a vaccine formulation containing multiple variants of CTB and TcpA.
The WCK oral cholera vaccine that provided significant protection for 5 years in an endemic region does not contain CTB, and it is unclear whether the heat and formalin treatments used to inactivate V. cholerae during preparation of that vaccine cause any damage to the immunogenicity of TcpA or other protein protective antigens of the bacteria [5], [6]. Our results show clearly that TcpA and CTB can be used successfully as protective subunit immunogens against cholera in the infant mouse model. Extending to humans the potential value of incorporating TcpA and CTB into effective vaccines against cholera will require additional studies to address the need: 1) to elicit production of antigen-specific sIgA antibodies in the human intestine; 2) to achieve long-term memory for protective intestinal immune responses; and 3) to develop vaccine formulations, adjuvants, routes of delivery, and immunization regimens to accomplish these goals.
|
10.1371/journal.pgen.1006552 | SBDS-Deficient Cells Have an Altered Homeostatic Equilibrium due to Translational Inefficiency Which Explains their Reduced Fitness and Provides a Logical Framework for Intervention | Ribosomopathies are a family of inherited disorders caused by mutations in genes necessary for ribosomal function. Shwachman-Diamond Bodian Syndrome (SDS) is an autosomal recessive disease caused, in most patients, by mutations of the SBDS gene. SBDS is a protein required for the maturation of 60S ribosomes. SDS patients present exocrine pancreatic insufficiency, neutropenia, chronic infections, and skeletal abnormalities. Later in life, patients are prone to myelodisplastic syndrome and acute myeloid leukemia (AML). It is unknown why patients develop AML and which cellular alterations are directly due to the loss of the SBDS protein. Here we derived mouse embryonic fibroblast lines from an SbdsR126T/R126T mouse model. After their immortalization, we reconstituted them by adding wild type Sbds. We then performed a comprehensive analysis of cellular functions including colony formation, translational and transcriptional RNA-seq, stress and drug sensitivity. We show that: 1. Mutant Sbds causes a reduction in cellular clonogenic capability and oncogene-induced transformation. 2. Mutant Sbds causes a marked increase in immature 60S subunits, limited impact on mRNA specific initiation of translation, but reduced global protein synthesis capability. 3. Chronic loss of SBDS activity leads to a rewiring of gene expression with reduced ribosomal capability, but increased lysosomal and catabolic activity. 4. Consistently with the gene signature, we found that SBDS loss causes a reduction in ATP and lactate levels, and increased susceptibility to DNA damage. Combining our data, we conclude that a cell-specific fragile phenotype occurs when SBDS protein drops below a threshold level, and propose a new interpretation of the disease.
| Shwachman Diamond syndrome (SDS) is an inherited disease. SDS presents, as hallmarks, exocrine pancreatic insufficiency, increased rate of infections, and higher incidence of leukemia. Most cases are due to mutations in the SBDS gene. SBDS encodes for a ribosome maturation factor. In this study, we immortalized mouse fibroblasts carrying one of the most common mutation of SDS patients and performed a thorough analysis of their properties. We show that the loss of SBDS activity causes a rewiring of gene expression and cellular metabolism. Overall we find a reduction of protein synthesis capability, a lower energy status, and increased lysosomal capability. SBDS mutant cells have an increased susceptibility to various forms of stress, but are strikingly resistant to oncogene-induced transformation. We propose a model that explains the complex phenotype of SDS patients and suggests roads for a rationale treatment.
| Ribosomopathies are inherited diseases due to the haploinsufficiency of ribosomal proteins or ribosome processing factors [1, 2]. Patients affected by ribosomopathies present multiorgan phenotypes [2]. Some relatively common features are bone marrow and skeletal deficits, and cancer predisposition [3, 4]. At the cellular level, ribosomal haploinsufficiency may cause the induction of tumor suppressor p53 [2, 5]. Consequently, in some models the depletion of p53 reduces the deleterious effects of ribosomal haploinsufficiency [6–9] leading to the hypothesis that abnormal p53 is the pathogenic culprit. However, this is not true for all cases [10, 11].
Shwachman-Diamond Syndrome (SDS) is a recessive ribosomopathy affecting 1 in 76,000 births. SDS is a multisystem disorder presenting in the first year of life and characterized by the hallmark of exocrine pancreatic dysfunction. Another common symptom is the susceptibility to chronic infections accompanied by neutropenia [12, 13]. With variable penetrance, patients affected by SDS may have low stature, skeletal defects and cognitive impairment [14, 15]. Finally, high risk of acute myeloid leukemia (AML) is associated with older patients [16]. Loss-of-function mutations in the SBDS gene have been identified as the cause of the disease [17].
Several studies addressed the function of SBDS protein in mammals and of its yeast homolog. A concise survey will be presented. SBDS has a role in the maturation of 60S ribosomal subunits. Deletion of the yeast homolog sdo1 is quasi-lethal, leading to pre-60S nuclear export defects. Importantly, point mutations of tif6, a gene necessary for 60S biogenesis [18], revert the quasi-lethal phenotype [19]. These and other studies, have led to a general model in which SBDS is necessary for the maturation of 60S subunits, in order to remove eIF6 (mammalian Tif6) from mature 60S subunits [20–22]. Since eIF6 controls 60S availability [23] and full translational activation [24, 25], manipulation of the binding of eIF6 to the 60S may be critical for restoring SBDS-mutant cells. A recent report suggested that SBDS contributes to the efficient translation of C/EBPα and C/EBP-β mRNAs, uORF-containing mRNAs that are, among others, indispensable regulators of granulocytic differentiation [26]. A general analysis of translated mRNAs depending from SBDS is still lacking. We do not know in detail whether other steps of 60S maturation beside eIF6 release are affected.
Several studies have attempted to pinpoint other functions of the SBDS protein, in the obvious effort to explain the multiorgan phenotype of patients. At the cellular level, several phenotypes associated to SBDS loss have been described. These phenotypes have been largely observed either in primary cells from patients, or upon shRNA experiments in cell lines. Increased apoptosis driven by FAS was seen in HeLa upon SBDS shRNA [27], and increased ROS production [28]. Increased apoptosis was also seen in SBDS-depleted HEK293 cells upon DNA and chemically induced endoplasmic reticulum stress [29], and in hemopoietic cells [30]. SBDS association with mitotic spindle has been proposed [31] and the lack of SBDS has been associated with genetic instability [32]. In general, reduced clonogenic potential is observed in hematopoietic precursors upon SBDS depletion [9, 33]. Reduced respiratory capability has been observed in mammalian and yeast cells lacking SBDS or its homolog, Sdo1p [34, 35]. Overall, it is unclear whether the cellular phenotypes ascribed to the loss of SBDS activity are direct or indirect, general or cell-specific, and have a relationship with protein synthesis.
Recently, a mouse strain in which one of the most frequent missense mutation found in SDS patients is modeled, R126T, has been produced [36]. This mouse model reproduces the clinical symptoms of SDS patients. We exploited the availability of this model to address the cell autonomous effects due to hypomorphic SBDS alleles. We have derived from these mice embryonic fibroblast cell lines, we have reconstituted control MEFs with wild type SBDS and then performed a full characterization of their properties. We addressed in SBDS mutant cells their predisposition to oncogenic transformation, changes in eIF6 binding, transcriptional and translational changes, metabolic parameters, and sensitivity to drugs and stresses. We unveiled several features due to the loss of SBDS. We provide a pathogenic model of SBDS deficiency that focuses on a diminished anabolic and energetic status of Sbds mutant cells due to reduced protein synthesis, which may be useful to design rational therapeutic ameliorative strategies.
Swhachman-Diamond (SDS) patients have an higher incidence of blood tumors, mainly acute myeloid leukemia (AML) [37]. This observation raises the question whether SBDS mutant cells are intrinsically susceptible to oncogene-mediated transformation. SBDS point mutation R126T corresponds to a common mutation found in SDS patients (c.377G>C) and it has been modeled in mice [36]. We immortalized R126T mouse embryonic fibroblasts (MEFs), together with their matched wild type controls, and we evaluated their capability to form colonies. Two strategies can be employed for immortalization of MEFs, a) sequential subcloning [38] or b) immortalization and transformation with oncogenes and tumor suppressor inhibitors [39]. By sequential subcloning, we were unable to derive SbdsR126T/R126T MEFs due to early senescence (Fig 1A), whereas we normally derived wt cells. This result is in line with a recent work describing early senescence in the pancreas of SBDS mutants [8]. In contrast, immortalization of SbdsR126T/R126T MEFs by infection with a vector carrying the dominant negative p53 and Ras G12V was successful (Fig 1A; S1A and S1B Fig). The growth of immortalized SbdsR126T/R126T MEFs was virtually identical to the one of wt cells (S1C Fig). Transformation of immortalized cells is induced by long-term growth at confluency. Next, we analyzed the capability of immortalized SbdsR126T/R126T MEF cells to form transformed colonies respect to wild type cells, upon long term culture at 100% confluency. We observed that SbdsR126T/R126T MEFs formed less foci respect to the wild type cells (Fig 1B and 1C). We then plated wt MEFs and SbdsR126T/R126T MEFs in soft agar, another indicator of transformation efficiency, to test their capability to grow in anchorage independent condition. We found that surviving clones of SbdsR126T/R126T MEFs grew as well as wt cells (Fig 1D), but their overall number was lower than the one of wt MEFs (Fig 1E). In order to evaluate the capability of these transformed colonies to induce tumor in vivo, we injected 500.000 transformed cells in nude mice and monitored tumor mass growth. Surprisingly and strikingly (Fig 1F), formally transformed SbdsR126T/R126T cells were inefficient (n = 1) or unable (n = 6) to grow in vivo respect to the wild type cells. We demonstrate that SBDS deficiency induces in a cell-autonomous fashion a growth and clonogenic deficit that can be unveiled when cells are challenged by environmental conditions.
We established a further model for studying SBDS function by generating, from immortalized SbdsR126T/R126T MEFs, new clones retransduced with either wild type Sbds (SbdsRESCUE) or mock control (SbdsMOCK) vectors (S1D Fig). By comparing parental SbdsR126T/R126T clones to their wild type counterparts, and the SbdsMOCK with the SbdsRESCUE clones, we can discriminate direct events due to SBDS lack from indirect effects. We describe the most important observations and discuss later a model that explains the pathogenic effect of SBDS deficiency. Since SBDS deficiency leads to a ribosomal defect [19], we performed a complete analysis of translation. Polysomal profiles can be used to analyze defects in initiation as well as in ribosome maturation. We observed, in line with previous reports [20], a strong unbalance in 60S and 80S peaks in SbdsR126T/R126T MEFs respect to the wild type cells (Fig 2A). SbdsRESCUE cells showed a complete rescue of the profile (Fig 2B), confirming a direct action of SBDS on ribosome maturation, and validating our model for discriminating direct versus clonal or indirect effects. We analyzed rRNA precursors with a pulse-chase assay by monitoring the incorporation of 5,6 3H-uridine in the nascent ribosomal RNA (S2A Fig). We did not observe differences in rRNA maturation associated with the SBDS mutation, a result consistent with the idea that only the late maturation of 60S is affected by SBDS deficiency. The nucleolus is the nuclear compartment where both ribosome biogenesis and early maturation occur. A defect in ribosomal export can be assessed by measuring the number and the size of nucleoli. We did not observe differences in the number of nucleoli between SbdsR126T/R126T and wild type cells (Fig 2C) or in SbdsMOCK and SbdsRESCUE cells (Fig 2D). In addition, we did not observe differences in co-localization of SBDS and nucleolar marker nucleophosmin (NPM) between SbdsR126T/R126T and wild type cells (Fig 2E and 2F). eIF6 nucleolar localization was relatively similar in SbdsR126T/R126T and wild type cells (S2B Fig).
eIF6 binds 60S subunits, blocking 80S formation and increasing free 60S peak [24]. It has been proposed that SBDS deficiency blocks eIF6 release [20]. This model is consistent with the accumulation of free 60S that we observed (Fig 2A). We recently developed a Ribosomes Interaction Assay (Fig 2G), able to quantify eIF6 binding sites on the 60S [40]. We immobilized equal ribosomes from SbdsR126T/R126T and wild type cells, and measured eIF6 binding sites. We found a 25% reduction in eIF6 binding sites on the ribosomes of SbdsR126T/R126T, compared to wild type cells (Fig 2H). Thus, we conclude that SBDS deficiency leads to a late maturation deficit of 60S consistent with the generation of a reduced pool of functional 60S subunits. It is worth to note that 60S peak increases at least 2-fold (Fig 2A and 2B), whereas eIF6 binding sites drop only 25% (Fig 2H).
We asked the consequences of the maturation deficit on translation. We developed a fully reconstituted in vitro model, in which translation competent extracts are prepared from equal amount of cells, normalized to the number of ribosomes and transduced with defined amounts of exogenous mRNA. This experiment allows to measure the maximal translational capability per cell/per ribosome. Ribosomal extracts from SbdsR126T/R126T MEF showed around 70% reduction in the translational capability of a cap-dependent reporter (Fig 3A). Cells rescued with SBDS in vivo (SbdsRESCUE) recovered their translational capability (Fig 3B). Importantly, adding wild type SBDS in vitro did not rescue the translational capability of extracts prepared from SbdsR126T/R126T MEFs (S2C Fig). This result indicates an overall translational impairment. Next, we adapted the canonical SUnSET protocol to citofluorimetry analysis. We observed both in SbdsR126T/R126T MEFs respect to wild type (Fig 3C) and in SbdsMOCK MEFs compared to SbdsRESCUE MEFs (Fig 3D) about 10% reduction in the number of cells incorporating medium to high levels of puromycin. Taken together, our results demonstrate that R126T mutation leads to a strong reduction of the pool of 60S subunits competent for translation.
An obvious question is whether the impaired maturation of 60S ribosomes, associated with a reduced translational capability in SbdsR126T/R126T cells, results in a qualitative difference of translation. We decided to proceed with an RNASeq study on total RNA extracted from sucrose gradient collected fractions. We studied 1) RNAs associated to polysomes, 2) RNAs associated to the 80S and 3) steady-state mRNA levels (total RNA) (Fig 4). The isolated fractions are shown in Fig 4A and 4B. The combination of these parameters allowed us to define the overall translational and transcriptional status associated with SbdsR126T/R126T mutation, assuming that mRNAs differentially localized to either polysomes or 80S are controlled at the translational level. We have decided to use this strategy over ribosome profiling because a) altered 60S subunits may lead to abnormal RNA protection, b) the big change in the 80S monosomal peak found in SbdsR126T/R126T cells is difficult to be normalized, and c) to efficiently reach deepness of more than 2x107 reads. S1 File contains read counts of polysomes and total RNAs. S2 File contains read counts of 80S. By comparing the polysomes of SbdsRESCUE and SbdsMOCK, we identified 844 modulated genes (Fig 4C; S1 File), most of them enriched in presence of mutant SBDS. However, when we estimated the translational efficiency, i.e. bona fide polysomal enrichment, by normalizing each mRNA level on the polysome to the amount present on the total, we found only 74 mRNAs with a significant modulation. Of these 74, 31 had less than 10 normalized read counts average expression, suggesting that fluctuations of poorly expressed mRNAs may contribute to the observed effects. The analysis of translation efficiency of SbdsRESCUE and SbdsMOCK confirmed that they do not differ in qualitative translational regulation (Fig 4D). The analysis on RNAs enriched on 80S subunits unveiled 250 genes modulated in 80S SbdsMOCK compared to 80S from SbdsRESCUE cells (Fig 4E; S2 File). However, identical to what we observed in the polysomal fraction, also in this case the mRNAs modulation on 80S followed the steady state levels (S3A and S3B Fig) as well the polysomal. Taken together, the data suggest that SBDS loss induces a solid transcriptional rewiring due to a general impairment of translation rather than specific translational regulation.
By analyzing polysomal versus steady state mRNAs and 80S modulated in Sbds mutated cells we made two major observations: a) a 4-fold enrichment of snoRNAs ACA8 and ACA31 in the ribosomes of SbdsMOCK cells compared to SbdsRESCUE. ACA8 and ACA31 drive the pseudouridylation of 28S rRNA U3832 and U3713 on the 60S. Other snoRNAs were not changed (Fig 4F; S1 and S2 Files). Intriguingly, there was no overlap between the genes regulated at the polysomal level by Sbds mutation (S1 File), to the ones we previously found regulated by eIF6 deficiency [41].
In conclusion, our data (Figs 2G, 2H, 3 and 4) suggest that the lack of mature 60S leads to a general reduction of translational capability associated with transcriptional rewiring.
The limited specific translational changes seen at the polysomal level (Fig 4), and the strong reduction in the maximal translational capability (Fig 3) were mirrored by a complex rewiring of gene expression and metabolism of SBDS deficient cells (Fig 5). Hereafter, we describe the transcriptional and metabolic changes directly due to SBDS deficiency, i.e. fully rescued by SBDS readministration in the SbdsR126T/R126T background. To simplify, 527 genes were at least 2-fold altered at the transcriptional level in SbdsMOCK cells, all of them presenting concomitant changes in the polysomal pool. Functional analysis by classical Gene Ontology (GO) for the Molecular Function Domain was performed on these 527 genes. Twentyseven ontology terms grouped in ten emerging categories were found as significantly enriched. Gene Set Association Analysis (GSAA) on both polysomal and total fractions confirmed the findings of the classical (GO) gene ontology (S3 File). Globally, we found that SBDS mutant cells had a decrease in genes encoding for the ribosomal and respiratory chains, and an increase in the lysosomal capability. We will specifically describe some of them. We found upregulated (both at the polysomes and at the steady-state level) mRNAs with peptidase activity including lysosomal cathepsins such as Ctsb, Ctsd, Ctsk and Ctsl (Fig 5A), and lysosomal genes with vacuolar ATPase activity, including Atp6ap1, Atp6ap2 and Atpv0b (Fig 5A, S1 File). Validation was confirmed by quantitative PCR (Fig 5B). M6pr, Lamp1 and Atp6ap1 upregulation suggests increased lysosomal activity associated with Sbds mutation, whereas increased CtsB suggests higher degradation activity. Consistently, we found in SbdsMOCK cells an increase in cell acidification by the Lysotracker assay (Fig 5C) and an increase in Lamp1 positive cells by immunofluorescence (Fig 5D and 5E) and by western blot (S5A Fig). Other coordinated gene expression changes observed in SbdsMOCK cells were a puzzling upregulation of membrane transporters for aminoacids and other intermediates (S4A Fig), a downregulation in most ribosomal components (S4B Fig), and a drop of expression of several genes important for mitochondrial function (S1 and S2 Files). These findings, in line with a reduction in the global protein synthesis capability observed in Fig 2G and 2H and in Fig 3, further predicted that SbdsR126T/R126T cells might have a decrease in the level of high energy molecules like ATP. Therefore, we measured intracellular ATP levels and found a decrease in ATP levels both in SbdsR126T/R126T and SbdsMOCK cells respect to their controls (Fig 5F). The reduction of ATP levels was also confirmed in HEK cells carrying an shRNA for SBDS (S5B and S5C Fig). In addition, we did not observe any significant difference 1. AMPK and 2. phospoAMPK levels, whereas a mild difference was appreciate in 3. PhosphoAMPK substrates by western blots (S5D and S5E Fig). We then measured the levels of secreted lactate and pyruvate (S6A and S6B Fig), as index of the glycolytic flux, and we found a reduction in the Lactate/Pyruvate ratio both in SbdsR126T/R126T and SbdsMOCK respect to their controls (Fig 5G). To further understand the metabolism of these cells we measured the levels of a. glycolytic activity, b. respiration and c. ROS production. We found that both SbdsR126T/R126T and SbdsMOCK cells show a decrease in glycolysis (S6C Fig) and a reduction in respiration (S6D Fig), while ROS levels remained unchanged (S6E Fig). In conclusion, we demonstrate that the reduction of SBDS activity causes a reduction of a global translational capability and a cellular adaptation with less energy production and more compensatory catabolic capability.
We focused our attention on information deriving from LINCS/CMap project to evaluate if the gene expression profile found altered in the total fraction of SbdsR126T/R126T MEFs cells was similar to transcriptional changes induced by drugs, observed in other cell types and to eIF6 deficiency [41]. Analysis performed by Query web tool on the complete list of modulated genes, and on a subset of the most changed ones identifies three common PKC activators: ingenol, phorbol-12-myristate-13-acetate and prostratin (S3 File).
Overall the analysis on SBDS deficient cells suggests that they have an impaired translational capability that leads to a compensatory transcriptional and metabolic rewiring that favors a catabolic processes and a state of low energy versus high synthetic capability. We reasoned that the transcriptional signature generated by SBDS deficiency could lead to a differential sensitivity to drugs or stressors which may be exploited at a therapeutic level. In principle, inhibitors that preferentially repress the growth of SbdsR126T/R126T immortalized MEFs may be of use in treating SDS patients who are affected by Acute Myeloid Leukemia (AML). On the contrary, selective stimulators of growth could be potentially interesting for early-phase of the disease, for instance in the context of neutropenia or pancreatic insufficiency. We decided to proceed with two different screenings, the first based on 100 compounds selected from commercial oral drugs (S4 File). We evaluated the response signature to drugs associated to R126T mutation by measuring cell viability after 48h of treatment. We found that overall SbdsR126T/R126T MEFs were similar to SbdsMOCK, and wt to SbdsRESCUE. This said, the signature profile that we found was surprisingly similar between wt and SBDS deficient cells (Fig 6A), with only one exception. In particular, we found increased sensitivity of SbdsR126T/R126T and SbdsMOCK cells to chlorambucil (Fig 6B), a DNA alkylating agent. Microtubule-targeting drugs mebendazole and colchicine were more toxic to SbdsR126T/R126T cells (S7 Fig), instead niclosamide had a stimulatory effect on SbdsR126T/R126T cells (S7 Fig), but the effects were not rescued in the SbdsRESCUE, suggesting indirect effects. The second screening with a Prestwick Library including 1280 compounds aimed at finding growth stimulators identified clindamycin (S7B Fig), but the effect was not rescued in SbdsRESCUE cells. In conclusion, 1. Sbds mutation cause limited effects unless cells are challenged, 2. chlorambucil data predict that SbdsR126T/R126T cells might be more sensitive to stress causing DNA damage.
The increased toxicity of the alkylating agent chlorambucil in SbdsR126T7R126T cells hinted a possible increased susceptibility to cell death associated to other forms of DNA damage. We first measured the basal level of apoptosis in wild type and in SbdsR126T/R126T cells by measuring the AnnexinV/7AAD positive population, and we did not find any differences between the two genotypes (Fig 7A). Treatment of cells with UV pulses caused an increase in cell death (Fig 7A), that was significantly more marked in SbdsR126T/R126T and in SbdsMOCK respect to wild type or SbdsRESCUE cells (Fig 7B). We analyzed DNA damage by examining cells with the COMET assay (Fig 7C and 7D), where the tail moment revealed a slight but significant increase in DNA damage in SbdsMOCK respect to wild type or SbdsRESCUE cells. This effect, even if small, is consistent with increased sensitivity of SbdsMOCK to two external stresses acting on DNA like chlorambucil and UV rays. These data, together with a decrease in energy levels and the uncapability to grow in vivo, show that Sbds mutated cells have an intrinsic fragility, that makes them weaker respect to their wild type counterpart.
Our study suggests a model (Fig 8A) that may explain the pathogenic culprit of SDS and will be first described. The table in Fig 8B outlines the phenotypes that we have observed.
A simple model for SBDS-driven pathological symptoms is based on a threshold concept. Total null SBDS mutations are never observed in humans indicating that a zero threshold of SBDS is incompatible with life. Mutant SBDS has a residual activity. This activity is sufficient for, i.e., normal proliferation, but becomes limiting for efficient colony formation or escaping senescence. If senescence is bypassed by oncogene-driven immortalization and transformation, then, SBDS remains still limiting for survival in nude mice. We can therefore imagine that each individual cell, in each moment, may or may not meet the condition of insufficient SBDS activity, with subsequent effects. Indeed, by the SUnSET protocol we see that only some SBDS mutant cells do not reach high translation levels.
If the threshold is not reached, our study supports a model in which SBDS deficiency leads primarily to a delayed maturation of 60S ribosomal subunits. We confirm that the number of eIF6 binding sites on 60S ribosomes is reduced, consistent with reduced generation of free 60S and with a proposed model of SBDS-mediated eIF6 release in the late maturation of 60S, in cooperation with EFL1 [26]. In detail, we quantify a 25% decrease of eIF6 binding sites on 60S, similar to [42]. We also report that the increase of free 60S is 2-fold, the amount of accumulated ACA8/ACA31 snoRNAs on immature 60S is 4-fold, and ribosomal extracts from SBDS mutants have a 70% reduction in translational capability. Apparently, the ribosomal deficit due to SBDS mutation is more pervasive than expected and suggests the generation of a bulk of immature 60S that are poorly active in translation, even if eIF6 is released. This said, eIF6 release may be critical in the rescue of some mature ribosomes and the alleviation of pathological consequences.
A variety of proxyes due to reduced SBDS activity are evident. Indeed, adaptation is evident in the absence of overt signs of SBDS insufficiency. Even if SBDS mutant cells proliferate normally in rich media, they show a signature with reduced energy level, increased lysosomal capability, decreased ribosomal production and less oxygen consumption. The drug screening identified a subtle signature of fragility which may contribute to the pathology. Previously, it was reported excess protease secretion in SDS-derived iPSC [43], reduced respiration in yeasts and mammalian cell models [34], and less ATP [35] that together with our extended gene signature define a status of reduced anabolic capability. This may be a keyfactor for rescuing the phenotype. Stimulators of ATP production, for instance, may have an impact on the disease. Our initial drug screening suggests that efforts in this direction may be valuable: we screened drugs for survival activity and found that SBDS mutants are remarkably similar to wt cells. However, a screening for a closer proxy of SBDS deficiency such as ATP levels, may lead to successful compounds.
We think that increased AML in SDS patients is a defect due to impaired host-cancer interplay, rather than due to genetic instability. If anything, SBDS-deficient cells are very resistant to oncogenic transformation and unable to grow in unfavourable conditions like the ones generated in tumors (see xenografts). In the bone marrow of SDS models, the neutrophil lineage can be impaired by the absence of sufficient endogenous SBDS activity [9]. SDS patients have neutropenia [44]. It is known that severe congenital neutropenia is associated with the progression to acute myeloid leukemia [45]. Therefore, we speculate that neutropenia in the bone marrow niche [30, 42] may lead to expansion of tumor cells that cannot be eliminated by proper immunosurveillance. Similar suggestions came from mouse work [46]. In summary, for therapy, we need to address which cell is directly affected by SBDS deficiency.
A divergence between our studies and previously published works relates to ROS [28] production. In our hands we did not find evidence for increased ROS production. It is possible that impaired ROS balance is observed as an indirect effect driven by mitochondrial alterations in specific cells, but is not a direct proxy of SBDS deficiency.
Finally, we would like to comment on translational control driven by SBDS and eIF6. SBDS insufficiency generates a unique phenotype. Similarly to SBDS mutants, eIF6 depletion protects from oncogene-induced transformation [25, 47], whereas its amplification is oncogenic [48, 49]. eIF6 depletion does not induce senescence [25], but induces a different transcriptional rewiring [41], a divergent bone marrow phenotype affecting platelets [50], and a generally improved metabolic status [41]. SBDS depletion reduces the translation of mRNAs undergoing reinitiation, as CEBP/β derived LIP peptide [26]. In our study the decrease of CEBP/β mRNA on polysomes driven by SBDS deficiency is around 10% (S1 Table). eIF6 depletion reduces both reinitiation and LIP expression as well as the translation of G/C rich mRNAs [41]. To summarize, the specific effects of eIF6 loss and SBDS loss are different, but overall they suggest that both eIF6 and SBDS have a stimulatory role in protein synthesis. In yeast and Dictyostelium eIF6 mutants revert the SBDS loss [19, 20, 22]. We speculate that these mutants are gain-of-function genes and that eIF6 agonists may be beneficial to the disease.
In conclusion, our results support the idea that cells with reduced SBDS activity are able to grow and cycle as well as wild type cells in proper conditions, but if they drop below a SBDS threshold level they show a phenotype. However, the adaptation signature suggests logical roads for improving their fitness. In the long run a screening for compounds which may accelerate the release of eIF6 from 60S subunits can be helpful to discover specific drugs for treating SDS (assuming that a modest increase in active 60S can strongly ameliorate the phenotype). Alternatively, screening on proxyes such as ATP levels may lead to faster therapeutic options.
Primary wild type and SbdsR126T/R126T MEFs (E12.5) were grown in DMEM (Lonza), supplemented with 10% Fetal Bovine Serum (FBS) and 1% penicillin, streptomycin, L-glutamine, and maintained at 37°C and 9% CO2. Mycoplasma testing was performed before experiments. These cells were infected at early passages through retroviral vectors carrying DNp53 + oncogenic H-rasV12 [25]. After immortalization, cells were maintained at 37°C and 5% CO2. Immortalized SbdsR126T/R126T MEFs were infected with lentiviral vectors carrying the wild type Sbds to generate the SbdsRESCUE line or with the corresponding empty vector to obtain the SbdsMOCK line. The lentiviral vectors (pHAGE-CMV-dsREDexpressIRESzsGreen backbone vectors) used to reconstitute SBDS wild type protein were kindly provided by A. Shimamura (Fred Hutchinson Cancer Research Center, Seattle). HEK-293T and HeLa cells from ATCC were cultured in DMEM (Euroclone) supplemented with 10% FBS and penicillin/streptomycin/glutamine solution (GIBCO) at 37°C and 5% CO2. Mycoplasma testing was performed before experiments. 293T cells were transfected at 60–70% confluence with pFCY, SBDS shRNA and pFCY scramble lentiviral vector (a generous gift from the lab of DC. Link, Washington University School of Medicine) to produce lentiviral particles. HeLa cells were infected with lentivirus at a confluence of 50%. Experiments were performed one week after infection. Silencing of protein was measured by western blot analysis. All animal experiments were carried out under the guidelines of the Canadian Council on Animal Care, with approval of procedures by The Animal Care Committee of the Toronto Centre for Phenogenomics, Toronto, AUP #0093.
For the transformation analysis, primary fibroblasts were infected at early passage with DNp53 + H-rasV12 retroviral vectors and left to grow at overconfluency. Foci were counted 3 weeks after infection and cells were recovered for in vivo experiments. Eight-weeks old CD1 athymic nude mice were used for detecting tumor growth after a subcutaneous injection of in vitro transformed MEFs (500,000 cells/mouse). Tumor growth was monitored and animals were euthanized when the tumor reached the size of 400 mm3. All experiments involving mice were performed in accordance with Italian National Regulations. Experimental protocols were reviewed by local Institutional Animal Care and Use Committees. The soft agar formation assay and the focus formation assay were performed as described previously [51]. In vitro assays were performed at least three times, each in triplicate.
Growing cells were lysed in 50 mM Tris-HCl, pH 7.5, 100 mM NaCl, 30 mM MgCl2, 0.1% Nonidet P-40, 100 μg/ml cycloheximide and 40 U/mL RNasin. After centrifugation at 12,000 r.p.m. for 10 min at 4°C, cytoplasmic extracts with equal amounts of RNA (10 OD260) were loaded on a 15–50% (or 10–30%) sucrose gradient dissolved in 50 mM NH4Cl, 50 mM Tris-Acetate, 12 mM MgCl2, 1 mM DTT and centrifuged at 4°C in a SW41Ti Beckman rotor for 3 h 30 min at 39,000 r.p.m. Absorbance at 254 nm was recorded by BioLogic LP software (BioRad) and fractions (1.5 mL each) were collected for subsequent proteins or RNA extraction. Each experiment has been performed at least six times, for each condition.
Total RNA was extracted from sucrose gradient aliquots. For the 15–50% gradient, we pulled fractions corresponding to polysomes in one fraction, and we pulled 100 μL from each fraction from the whole gradient in one fraction (total). For the 10–30% gradient, we pulled fractions corresponding only to the 80S peak in one fraction, named 80S, and we used as total RNA aliquots from the pre-gradient loading extract (pre-load). Afterward, we added to samples proteinase K (to a final concentration 100μg/mL) and SDS (to a final concentration of 1%) and we incubated them for 1 h at 37°C. Total RNA was then extracted by phenol/chloroform/isoamyilic acid method (https://tools.thermofisher.com/content/sfs/manuals/trizol_reagent.pdf). Libraries for Illumina sequencing were constructed from 100 ng of total RNA with the Illumina TruSeq RNA Sample Preparation Kit. The generated libraries were loaded on to the cBot (Illumina) for clustering on a HiSeq Flow Cell v3. The flow cell was then sequenced using a HiScanSQ (Illumina). A paired-end (2×101) run was performed using the SBS Kit v3 (Illumina). Experiments were performed in biological triplicates.
For quantitative PCR, 150 ng of RNA was retrotranscribed according to SuperScriptTM III First-Strand Synthesis SuperMix manufacturer protocol (18080400, Life Technologies). For RNASeq validation, Taqman probes specific for Atp6ap1 (Mm01187488_g1), Lamp1 (Mm00495262_m1), Ctsb (Mm01310506_m1) and M6pr (Mm04208409_gH) were used. Target mRNA quantification was performed by using ΔCt-method with 18S rRNA as an internal standard, performed on a StepOne Puls System (Applied Biosystems). Results are represented as means + s.d. of three independent experiments.
Gene set association analysis for polysomal and total fractions was performed by GSAA software (version 1.2) [57]. Raw reads for about ~ 22000 genes identified by Entrez Gene ID were analyzed by GSAASeqSP, using gene set C5 (mouse version retrieved from http://bioinf.wehi.edu.au/software/MSigDB) and specifying as permutation type ‘gene set’ and as gene set size filtering min 15 and max 800.
Analysis on CMap/LINCS gene signatures was performed using the ‘Query’ web tool (http://apps.lincscloud.org/query), that, taking as input a list of genes, computes the connectivity between this set and the gene expression signatures of the LINCS database. Human orthologs of mouse genes were retrieved relying on Biomart web tool (http://www.ensembl.org/biomart); only genes with a one-to-one orthologs relationship were maintained for downstream computations. Query tool was first applied on the list of 726 mouse genes significantly modulated in the total fraction; chemical compounds showing a mean rank value of at least 90 on 6 cell lines were selected as interesting. As second strategy, the same procedure was applied on the genes showing a robust modulation (143 genes with a fold-change higher than 3 as absolute value) and then selecting chemical compounds with a mean rank value of at least 90 on 4 cell lines. To find perturbations that are consistently retrieved, the overlap between the two analyses was considered as final result.
SDS-PAGEs were performed on protein extracts obtained with RIPA buffer (10 mM Tris-HCl, pH 7.4, 1% sodium deoxycholate, 1% TritonX-100, 0.1% SDS, 150 mM NaCl and 1 mM EDTA, pH 8.0). Protein concentration was determined with BCA analysis (Thermo Fisher Scientific). Equal amounts of proteins were loaded on each lane and separated on a 10% SDS-PAGE, then transferred on a PVDF membrane. The membranes were blocked in 10% Bovine Serum Albumin (BSA) in Phosphate Buffer Saline (Na2HPO4 10 mM, KH2PO4 1.8 mM, NaCl 137 mM, KCl 2.7 mM, pH 7.4) (PBS) with Tween (0,01%) for 30 minutes at 37°C. The following primary antibodies were used: β-actin (CST 4967L 1:4000), SBDS (Santa Cruz S15 SC49257 1:500) H-RAS (Santa Cruz SC520, 1:1000), Lamp1 (Santa Cruz SC20011, 1:200), AMPK (CST 5831 1:1000), phospho-AMPK (CST 2535 1:1000), phospho-AMPK Substrates (CST 5759, 1:1000). The following secondary antibodies were used: donkey anti-goat IgG HRP (Santa Cruz SC2020, 1:2000), donkey anti-rabbit IgG HRP (Amersham NA934 1:5000) and donkey anti-mouse IgG HRP (Santa Cruz, SC2005 1:5000). Each experiment was performed at least three times, each time in triplicate.
Cells were seeded the day before the staining. The next day, cells were rinsed three times with PBS then fixed with ice cold methanol 100% for 10 minutes at -20°C. After three washes with PBS, cells were blocked in Normal Goat Serum 5%, for 1 hour at room temperature. Primary antibodies were incubated overnight at 4°C, and after three washes with PBS, secondary antibodies were incubated 3h at room temperature in the dark. After three washes with PBS, cells were incubated with DAPI (Molecular Probes NucBlue Live ReadyProbes Reagent R37605) as manufacturer protocol, then washed and mounted on slides with Mowiol 20%. All the antibodies were diluted in blocking solution. The following primary antibodies were used: NPM [25], SBDS (Santa Cruz S15 SC49257 1:25), eIF6 [58] (1:100), Lamp1 (Santa Cruz sc-20011 1:100). The following secondary antibodies were used: donkey anti-goat, donkey anti-mouse, donkey anti-rabbit (Alexa Fluor secondary antibodies, Molecular Probes 1:500). The cells were examined by confocal microscopy (Leica SP5) and analyzed with Volocity 6.3 software (Perkin Elmer). Immunofluorescence experiments were performed at least three times, in triplicate.
Comet assay was performed following the manufacturer protocol (Trevigen, 4250-050-K). The stained nuclei were then examined by confocal microscopy (Leica SP-5), and analyzed (n≥30 for each condition) with Comet Assay IV software (Perceptive Instruments). The experiment was performed three times in triplicate.
Ribosome biogenesis was analyzed by pulse-chase experiments by adding 5,6-3H-Uridine to the medium (final concentration 3uCi/mL, NET367001MC PerkinElmer). 5,6-3H -Uridine was removed after 0, 10, 20 and 40 minutes of incubation and RNA was extracted with Trizol Reagent and hybridized on nitrocellulose membrane. This experiment was reproduced twice.
For in vitro translation assays, we used cell extracts prepared as described [59] with some optimizations. 80% confluent cells were trypsinized and lysed for 45 minutes at 4°C in 10 mM HEPES pH 7.6, 10 mM potassium acetate, 0.5 mM magnesium acetate, 5 mM DTT and protease inhibitor (Promega). Lysates were homogenized by syringing through a 27G, 3/4-inch needle. Lysates were clarified by centrifugation at 18.000 g for 1 minute and protein concentration was determined by BCA quantification assay. To make translation fully dependent on exogenously added mRNA, lysates were treated with 15 U/ml Micrococcal nuclease (MN Boehringer) and 0.75 mM CaCl2 and incubated at 25°C for 7 minutes. EGTA was added to terminate the reaction.
For the translation assay, 6 μl of cell extract were mixed to 1.2 μl of Master Mix (125 mM HEPES, 10 mM ATP, 2 mM GTP, 200 mM creatine phosphate, 0.2 mM aminoacid mix without methionine (Promega), 0.25 mM spermidine, 20 mM of L-methionine, 50 mM potassium acetate, 2.5 mM magnesium acetate, 20 U of RNAsin (from Promega) and 0.5 μg of purified reporter-encoding mRNA). The mRNA was obtained by using the Megascript T7 kit (Ambion) to perform the in vitro transcription reaction supplemented with 2 mM cap analog [M7G(5')PPP(5')G] (Ambion). The mix for the in vitro translation reaction was then incubated for 90 minutes at 30°C. Dual-Glo Luciferase Assay kit (Promega) was used to read Firefly and Renilla luciferase output with GloMax Luminometer (Promega). In vitro translation experiments were performed at least three times, in triplicate.
For protein synthesis measurements, we adapted SUnSET protocol [60]. 70% confluent cells were treated with 5μg/ml puromycin for 10 minutes, then trypsinized and centrifugated 5 minutes at 500 g. Then, cell were fixed and permeabilized using reagents from the Foxp3/Transcription Factor Staining Buffer Set (Affymetrix, 5523–00), incubated with anti-puromycin 12D10 (1:2000, Millipore, MABE343) and finally with secondary antibody Alexa Fluor 488 Goat anti-Mouse (Invitrogen) and DAPI, according to manufacturer protocol. Cells were then analyzed with FACSCanto II (BD Bioscience) and analyzed with FlowJo8.8.7 software. SUnSET experiments were performed at least three times, in triplicate.
1280 approved drugs were analyzed for their impact on SbdsR126T/R126T cells growth. Cells were seeded using a Multidrop Combi (Thermo Scientific) in 384 well plates (Perkin Elmer Cell Carrier, collagen coated) at a density of 4000 cells/well. After 24 hours the drugs were supplemented using a Hamilton Starlet liquid handler at a final concentration of 10 μM and cells were treated for 30 hours. The reference compound for toxicity was 5 μM Trichostatin A. Cytotoxicity was than evaluated by measuring in vivo nuclear shrinkage and loss of cells using Hoechst33342, and cell membrane disruption using the cell-impermeant nuclear dye BOBO-3. Briefly, following incubation with the drugs diluted in 50 μl growth medium, 25 μl of a dye cocktail containing 3 μM Hoechst33342 and 2.25 μM BOBO-3 were added and incubation was continued for further 45 min at 37°C, 5% CO2. Images were then immediately acquired and analyzed using the Operetta microscope (Perkin Elmer). For the validation of hits, 6-point dose-responses for each compound in duplicate were used (ranging from 30uM to 0.12uM). The number of dead cells/total number of cells was the readout of the assay, where dead cells correspond to nuclei stained with BOBO-3 and the total number of cells correspond to the nuclei stained with Hoechst33342.
All analysis were performed using the FACSCanto II Flow cytometer (BD) and analyzed with FlowJo software (BD). All experiments were performed when cells reached the confluence of 70%. For the apoptosis analysis, cells were UV irradiated with a cross-linker (9999 μJ/cm2, three times) and cell death was detected by using the 7AAD-Annexin V kit (640926, BioLegend). For ROS measurement, cells were incubated with 50 μM H2DCF-DA (2’, 7’, -dichlorodihydrofluorescein diacetate, AbCam Ab113851) for 30 min at 37°C and 5% CO2 and immediately analyzed by flow citometry. All experiments were performed at least three times, in triplicate.
iRIA assay was performed as described in [40]. Briefly, 96-well plates were coated with a cellular extract diluted in 50 μL of PBS, 0,01% Tween-20, O/N at 4°C in humid chamber. Coating solution was removed and aspecific sites were blocked with 10% BSA, dissolved in PBS, 0,01% Tween-20 for 30 minutes at 37°C. Plates were washed with 100 μl /well with PBS-Tween. 0,5 μg of recombinant biotynilated eIF6 was resuspended in a reaction mix: 2,5 ;mM MgCl2, 2% DMSO and PBS-0.01% Tween, to reach 50 ;μl of final volume/well, added to the well and incubated with coated ribosomes for 1 hour at room temperature. To remove unbound proteins, each well was washed 3 times with PBS, 0,01% Tween-20. HRP-conjugated streptavidine was diluted 1:7000 in PBS, 0,01% Tween-20 and incubated in the well, 30 minutes at room temperature, in a final volume of 50 μl. Excess of streptavidine was removed through three washes with PBS-Tween. OPD (o-phenylenediamine dihydrochloride) was used according to the manufacturer’s protocol (Sigma-Aldrich) as a soluble substrate for the detection of streptavidine peroxidase activity. The signal was detected after the incubation, plates were read at 450 nm on a multiwell plate reader (Microplate Bio-Rad model 680). This experiment was performed at least three times, in triplicate.
Cells were seeded the day before the assay. The next day, cells were incubated with Lysotracker DeepRed (Thermo Fisher Scientific L12492) according to the manufacturer protocol. The cells were examined with Nikon Ti-Eclipse microscope and analyzed with Volocity 6.3 software (Perkin Elmer). For subsequent analysis, cells were then fixed with paraformaldehyde 3% for 10 minutes at room temperature and nuclei were stained with DAPI (Molecular Probes NucBlue Live ReadyProbes Reagent R37605) as manufacturer protocol, then washed and mounted on slides with Mowiol 20%. The cells were examined by confocal microscopy (Leica SP5) and analyzed with Volocity 6.3 software (Perkin Elmer). Lysotracker experiments were performed at least three times, in triplicate.
The RNA seq data are available at www.ebi.ac.uk/arrayexpress with accession number ID: E-MTAB-5089.
|
10.1371/journal.ppat.1000980 | Bid Regulates the Pathogenesis of Neurotropic Reovirus | Reovirus infection leads to apoptosis in both cultured cells and the murine central nervous system (CNS). NF-κB-driven transcription of proapoptotic cellular genes is required for the effector phase of the apoptotic response. Although both extrinsic death-receptor signaling pathways and intrinsic pathways involving mitochondrial injury are implicated in reovirus-induced apoptosis, mechanisms by which either of these pathways are activated and their relationship to NF-κB signaling following reovirus infection are unknown. The proapoptotic Bcl-2 family member, Bid, is activated by proteolytic cleavage following reovirus infection. To understand how reovirus integrates host signaling circuits to induce apoptosis, we examined proapoptotic signaling following infection of Bid-deficient cells. Although reovirus growth was not affected by the absence of Bid, cells lacking Bid failed to undergo apoptosis. Furthermore, we found that NF-κB activation is required for Bid cleavage and subsequent proapoptotic signaling. To examine the functional significance of Bid-dependent apoptosis in reovirus disease, we monitored fatal encephalitis caused by reovirus in the presence and absence of Bid. Survival of Bid-deficient mice was significantly enhanced in comparison to wild-type mice following either peroral or intracranial inoculation of reovirus. Decreased reovirus virulence in Bid-null mice was accompanied by a reduction in viral yield. These findings define a role for NF-κB-dependent cleavage of Bid in the cell death program initiated by viral infection and link Bid to viral virulence.
| Viruses injure host tissues by activating signaling pathways that trigger cell death by a process called apoptosis. Hence, blockade of apoptosis may serve as a useful strategy to dampen the severity of viral disease. However, deployment of such a strategy requires identification of host signaling networks that control cell death and a detailed molecular blueprint of how these pathways are activated by a virus. In this study, we used mammalian reovirus, an important experimental model for studies of viral encephalitis, to elucidate how cell death pathways are activated following viral infection and whether these signaling cascades influence the capacity of a virus to produce lethal CNS disease. We found that Bid, a host regulator of cell death, influences apoptosis induction by reovirus. Moreover, Bid is required for efficient reovirus replication in the CNS and modulates reovirus neurological disease. These findings highlight Bid as a critical regulator of viral pathogenesis and illuminate a potential new target for development of antiviral therapeutics.
| Tissue injury in response to infections by many viruses occurs as a consequence of apoptosis. Multiple studies using animal models of viral disease demonstrate a correlation between apoptotic potential and disease severity [1], [2], [3], [4]. These observations highlight proapoptotic signaling following virus infection as an attractive target for antiviral therapy. However, despite its central importance in viral pathogenesis, gaps in knowledge about the identity of death signaling pathways that modulate virus-induced apoptosis in vivo, along with an incomplete understanding of how these signaling cascades are activated during virus infection, have hampered the deployment of this strategy for treatment of viral disease.
Mammalian reoviruses injure infected cells via apoptosis both in culture and in tissues of infected animals. As such, studies of these viruses have contributed to an understanding of how virus infection culminates in apoptotic cell death. Unlike other viruses in which virulence correlates with cell-death capacity, the identity of viral and cellular factors that regulate reovirus-induced apoptosis in cell culture are for the most part known [4], [5], [6], [7], [8], [9], [10]. Moreover, many of these intermediaries also modulate reovirus-induced apoptosis in vivo [4], [7], [11], [12]. Studies using reassortant reoviruses [5], [6], ectopically expressed proteins [13], and genetically engineered reovirus mutants [4], [7] highlight a critical role for reovirus outer-capsid protein µ1 in apoptosis induction. Collectively, these studies indicate that prodeath signaling evoked by µ1 occurs subsequent to membrane penetration but prior to synthesis of viral RNA or protein [4], [7], [14], [15].
Classical death-receptor-mediated extrinsic apoptotic pathways stimulated by reovirus infection execute the death response [16]. Treatment of cells with soluble TRAIL receptors or expression of a dominant-negative form of Fas-associated death domain (FADD) protein blocks apoptosis, demonstrating that signaling via death receptors is required for execution of the apoptotic program [17]. In keeping with the function of extrinsic apoptotic signaling in reovirus infection, caspase-8 activation [18] and Bid cleavage [19] are observed in cells infected with reovirus [16]. Reovirus infection also stimulates intrinsic apoptotic pathways, as evidenced by release of cytochrome c and Smac/DIABLO from the mitochondria and activation of caspase-9 [16], [20], [21], [22], [23]. Concordantly, reovirus-induced apoptosis is dampened by over-expression of Bcl-2 [24], which inhibits mitochondrial apoptotic pathway activation [25].
Bid is a proapoptotic BH3-only member of the Bcl-2 family that functions to link the extrinsic apoptotic pathway and the mitochondrial amplification loop of the intrinsic pathway. Following death-receptor signaling, cytoplasmically resident Bid is cleaved by activated caspase-8 to generate a truncated form of Bid known as tBid [26]. tBid translocates to the mitochondria and triggers the release of cytochrome c and activation of the core mitochondrial apoptotic machinery [19], [27]. It is not known whether Bid plays a functional role in apoptosis induction by reovirus. Moreover, the relationship between apoptosis effector pathways and early events in viral replication are not understood.
In addition to these classical apoptotic pathways, the innate immune response transcription factor, NF-κB, is activated following reovirus infection [8]. NF-κB activation by reovirus depends on the viral µ1 protein and can be accomplished by genome-deficient reovirus particles [4], [7], [8]. Blockade of NF-κB signaling using chemical inhibitors or cell lines genetically deficient in NF-κB p50, NF-κB p65/RelA, IκB kinase (IKK)-α, or IKK adaptor IKKγ/NEMO significantly diminishes reovirus-induced apoptosis [8], [28]. Consistent with these findings, activation of NF-κB occurs within the first few hours of reovirus infection and precedes the biochemical and morphological hallmarks of apoptotic cell death [8], [28]. These observations suggest that NF-κB couples µ1-mediated events to the cellular apoptotic machinery. Although regulation and function of NF-κB has been extensively studied, the precise relationship between NF-κB and the cell-death machinery remains undefined.
In this study, we examined the function of cellular apoptosis regulator Bid using genetically deficient murine embryo fibroblasts (MEFs) and mice. We found that while Bid is dispensable for reovirus replication in cell culture, its function is required for reovirus-induced apoptosis. Blockade of NF-κB signaling, which diminishes apoptosis induction by reovirus [8], [28], prevents cleavage of Bid. In comparison to wild-type mice, Bid-deficient mice display diminished susceptibility to reovirus-induced CNS disease following either peroral (PO) or intracranial (IC) inoculation. Attenuated reovirus virulence in the absence of Bid is associated with decreased reovirus replication in the murine CNS. These results define an important role for Bid in virus-induced apoptosis and disease and illuminate Bid-dependent prodeath signaling as a viable target for antiviral therapy.
Reovirus infection of HEK293 epithelial cells leads to a biphasic loss of full-length (FL) Bid [16]. Since a mitochondrial amplification loop through Bid is required for apoptosis only in some cell types, such as hepatocytes [29], [30], it is not known if Bid is cleaved in all cell types infected by reovirus. In addition, although calpains [31], [32], caspases [26], [33], [34], [35], and cathepsins [36], [37], [38], [39] can mediate Bid cleavage and have been implicated in apoptosis induction by reovirus [15], [16], [40], the precise identity of the protease that generates tBid following reovirus infection is not known. To determine whether tBid is generated following reovirus infection of fibroblasts, and to define the mechanism of Bid cleavage following reovirus infection, we infected murine L929 fibroblasts with reovirus strain type 3 Dearing (T3D) and monitored levels of FL Bid and tBid over 48 h (Figure 1A). While levels of FL Bid remained unchanged in mock-infected cells, we observed loss of FL Bid between 24 and 48 h post-infection. Decreased levels of FL Bid correlated with a corresponding increase in levels of tBid. To determine whether the generation of tBid results in activation of the mitochondrial loop of the intrinsic apoptotic pathway, we assessed levels of procaspase-9 as a surrogate for the formation of the caspase-9-containing apoptosome (Figure 1A). In a time frame consistent with cleavage-induced generation of tBid, we observed a decrease in procaspase-9 levels in reovirus-infected cells. These findings suggest that following reovirus infection of murine fibroblasts, Bid serves to activate the mitochondrial apoptotic pathway.
The adaptor molecule FADD is required for cleavage of Bid following reovirus infection of HEK293 cells [16]. Based on these data, we hypothesized that caspase-8 activity as a consequence of extrinsic prodeath signaling, is required for cleavage and activation of Bid. To test this hypothesis, we assessed the capacity of reovirus to mediate Bid cleavage in L929 cells treated with caspase-8 inhibitor Z-IETD-FMK (Figure 1B). As anticipated, Bid cleavage was not observed in mock-infected cells or mock-infected cells treated with Z-IETD-FMK (data not shown). Although tBid was generated at 36–48 h following reovirus infection of vehicle-treated cells, reovirus failed to efficiently induce activation of Bid in Z-IETD-FMK-treated cells until 48 h post-infection, providing evidence that reovirus evokes cleavage of Bid via caspase-8. In response to a variety of death agonists, Bid amplifies death signaling by linking the extrinsic (caspase-8) and intrinsic (caspase-9) apoptotic pathways [30]. Since our findings with reovirus parallel this pattern, our results suggest that Bid functions similarly following reovirus infection by linking the death-receptor and mitochondrial apoptotic pathways.
Signaling via the intrinsic pathway is essential for reovirus-induced apoptosis [16]. This observation, along with the dependence of mitochondrial apoptotic signaling on cleavage of Bid, suggests that Bid serves an essential function in reovirus-induced apoptosis. To directly test whether Bid is required for apoptosis induction following reovirus infection, we compared reovirus-induced apoptosis in wild-type and Bid-deficient MEFs. For these experiments, MEFs were infected with T3D, and apoptosis was assessed by chemiluminescent measurement of the activity of caspase-3 and caspase-7, which serve as effector caspases for both the extrinsic and intrinsic apoptotic pathways (Figure 2A). In comparison to mock-infected cells, infection of wild-type cells resulted in a significant increase in caspase-3/7 activity at 24 h post-infection. Since MEFs are poorly permissive for reovirus infection [41], staining of infected cells by indirect immunofluorescence indicated that adsorption with 100 PFU/cell of T3D resulted in infection of only ∼8% of cells at 20 h post infection (data not shown). Despite a low frequency of infection, this MOI resulted in an ∼3-fold increase in caspase-3/7 activity. When infection was initiated at 1000 PFU/cell, ∼20% cells were infected (data not shown), and caspase-3/7 activity increased ∼5-fold. In contrast, infection of Bid-deficient cells resulted in minimal caspase-3/7 activity following infection at either MOI. Increase in caspase-3/7 activity following treatment of each cell type with a broad-spectrum protein kinase inhibitor, staurosporine, was equivalent (∼5-fold), demonstrating that although Bid-deficient cells possess functional death-signaling pathways, they resist apoptosis induction by reovirus.
As an alternative means to quantify apoptosis, we compared wild-type and Bid-deficient MEFs for the onset of morphological characteristics of apoptosis following reovirus infection using an acridine orange (AO) staining assay (Figure 2B). Infection of wild-type cells resulted in a significant increase in the fraction of apoptotic cells at 48 h post-infection with 40% and 100% of the cells exhibiting apoptotic features at MOIs of 100 and 1000 PFU per cell, respectively. In contrast, Bid-deficient cells infected with T3D at either MOI displayed levels of apoptosis equivalent to mock-infected cells, ∼10%. Similar results were obtained following infection with another apoptosis-proficient reovirus strain, T3SA+ (data not shown). These data indicate that Bid is required for apoptosis induction following reovirus infection.
To determine whether decreased apoptosis in Bid-deficient cells is attributable to alterations in reovirus infection in the absence of Bid, we compared reovirus infectivity in wild-type and Bid-deficient cells using an indirect immunofluorescence staining assay (Figure 2C). An equivalent proportion of reovirus antigen-positive cells was detected at 20 h post-adsorption of wild-type and Bid-deficient cells. These data indicate that reovirus is capable of initiating infection in Bid-deficient cells. To determine whether reovirus completes a full infectious cycle in Bid-deficient cells, wild-type and Bid-deficient cells were adsorbed with T3D, and viral titers were determined by plaque assay at 0, 12, 24, and 48 h after infection (Figure 2D). Reovirus replicated with similar kinetics and produced equivalent yields in wild-type and Bid-deficient cells. Thus, the failure of Bid-deficient cells to undergo apoptosis in response to reovirus is not a consequence of diminished reovirus infection of these cells. We conclude that Bid is a key regulator of reovirus-induced apoptotic cell death.
The identification of an essential role for Bid in apoptosis induction following reovirus infection allowed us to examine the relationship between NF-κB activation and Bid cleavage. To determine whether Bid is required for activation of NF-κB following reovirus infection, we compared reovirus-induced NF-κB activation in wild-type and Bid-deficient cells using a reporter assay. Wild-type and Bid-deficient MEFs were transfected with an NF-κB-luciferase reporter plasmid and infected with reovirus. Analogous to treatment with TNFα, a control NF-κB agonist, reovirus infection resulted in equivalent (∼2- to 3-fold) activation of NF-κB-driven gene expression in wild-type and Bid-deficient cells (Figure 3A). These results indicate that Bid is dispensable for NF-κB activation following reovirus infection and suggest that either reovirus-induced NF-κB activation occurs prior to Bid cleavage or that NF-κB activation and Bid cleavage occur in parallel but independent pathways that both function in apoptosis induction by reovirus.
To determine whether cleavage-induced Bid activation is dependent on NF-κB, we examined Bid cleavage in cells lacking p65/RelA, an NF-κB subunit required for apoptosis induction following reovirus infection [8]. Infection of wild-type MEFs with reovirus results in generation of tBid at 36–48 h after infection (Figure 3B). In contrast, infection of p65/RelA-deficient MEFs with reovirus did not lead to tBid generation even though efficient viral replication is observed in these cells [8]. Treatment of both wild-type and p65/RelA-deficient MEFs with apoptotic agonists TNFα and cycloheximide resulted in efficient cleavage of Bid, indicating that cell-death pathways leading to Bid cleavage are intact in both cell types. These findings suggest that cleavage and activation of Bid following reovirus infection requires NF-κB and place Bid cleavage subsequent to NF-κB signaling in response to reovirus infection. Moreover, since Bid amplifies death responses from the extrinsic apoptosis pathway by activating the mitochondrial loop, these findings suggest that death-receptor signaling during reovirus infection occurs in an NF-κB-dependent manner.
Apoptosis-signaling pathways involving death receptors DR4 and DR5 and death ligand TRAIL, as well as Fas and FasL, have been implicated in apoptosis induction by reovirus [17], [42], [43]. However, it is not known which of these pathways mediates cleavage-induced activation of Bid. It is also not understood whether NF-κB regulates the activation of these pathways. Since upregulation of Fas following reovirus infection is dependent on prodeath signaling via c-Jun N terminal kinase (JNK) [43], and because JNK is activated via a mechanism distinct from NF-κB following reovirus infection [44], we focused our efforts on assessing the regulation and function of death-receptor signaling via TRAIL following reovirus infection. For these studies, we assessed the capacity of reovirus to induce apoptosis in MEFs lacking TRAIL-R, the only known receptor for TRAIL on murine cells [45], [46], [47] (Figure 4). In comparison to mock infection, T3D infection of wild-type cells resulted in an MOI-dependent ∼5- to 20-fold increase in caspase-3/7 activity at 24 h post-infection (Figure 4A). Although T3D infection of TRAIL-R-deficient cells also resulted in an increase in caspase-3/7 activity in comparison to mock-infection, the magnitude of this increase was only ∼2- to 6-fold. Assessment of apoptosis in wild-type and TRAIL-R-deficient MEFs using AO staining also showed an increase in apoptosis both in wild-type and TRAIL-R-deficient cells in comparison to mock-infected cells (Figure 4B). However, a substantially greater fraction of wild-type cells showed morphologic features of apoptosis in comparison to TRAIL-R-deficient cells infected at equivalent MOI, suggesting that efficient induction of apoptosis by reovirus requires TRAIL-R. T3D displayed comparable replication kinetics and produced equivalent yields in wild-type and TRAIL-R-deficient cells (Figure 4C). Thus, differences in the apoptotic potential of reovirus in wild-type and TRAIL-R-deficient cells are not associated with differences in reovirus growth in these cells.
To determine whether reovirus-induced cleavage of Bid is dependent on signaling via TRAIL-R, we monitored Bid cleavage following infection of TRAIL-R-deficient cells (Figure 4D). At 48 h post-infection of wild-type cells with T3D, FL Bid was cleaved to generate tBid. In contrast, FL Bid was not cleaved in T3D-infected TRAIL-R-deficient cells. While the apparent difference in the levels of FL Bid in wild-type and TRAIL-R-deficient cells was not reproducible, we consistently observed that levels of FL Bid remained unchanged in TRAIL-R-deficient cells following reovirus infection. These data indicate that TRAIL-R contributes to the induction of apoptosis by reovirus and suggest that cleavage of Bid following reovirus infection is dependent on TRAIL-R signaling.
Reovirus virulence correlates with its capacity to cause apoptosis [4], [7], [11], [12], [48], [49]. Given the central role of Bid in apoptosis induction by reovirus in cell culture, we hypothesized that reovirus apoptosis and virulence would be diminished in the absence of Bid. To test this hypothesis, we inoculated two-day-old wild-type and Bid-deficient mice perorally with a highly virulent, enteric, neurotropic reovirus strain, T3SA+ [50], and monitored infected animals for signs of neurological disease and infection-induced morbidity over a period of 21 days (Figure 5A). Following inoculation with 104 PFU of T3SA+, most wild-type mice developed paralysis and respiratory distress. In contrast, the majority of Bid-deficient mice were asymptomatic. Consistent with this observation, ∼91% of wild-type mice succumbed to reovirus infection with a median survival time of 11 days, whereas only ∼30% of Bid-deficient mice died. Due to the relative resistance of Bid-deficient mice to reovirus-induced encephalitis, a median survival time could not be determined. Thus, the cellular apoptotic regulator Bid modulates reovirus-induced encephalitis.
To determine whether the enhanced survival of Bid-deficient mice in comparison to wild-type mice following T3SA+ infection results from reduced reovirus replication, we compared titers of reovirus at sites of primary and secondary replication at 4, 8, and 12 d post-inoculation (Figure 5B–E). Peak titers of reovirus were comparable or slightly higher (∼5- to 10-fold) in the intestine, liver, and heart of wild-type mice in comparison to Bid-deficient animals. In contrast, substantially greater differences in peak reovirus titers were observed in the brain, with wild-type animals showing ∼25- to 100-fold higher titers in comparison to those in Bid-deficient mice at 8 d post-inoculation. However, by 12 d post-inoculation, titers of reovirus in wild-type and Bid-deficient mouse brains were equivalent. These findings suggest that reovirus infection is inefficient in the absence of Bid, especially in the CNS.
Although titers of reovirus in the CNS were decreased in Bid-null mice following PO inoculation, it was not clear whether reduced reovirus titer in the CNS was a consequence of diminished reovirus dissemination to the CNS or diminished reovirus replication at that site. To distinguish between these possibilities, we inoculated wild-type and Bid-deficient mice intracranially with 100 PFU of T3SA+ and monitored infected animals for signs of CNS disease and mortality for 21 days (Figure 6A). At this dose of T3SA+, most wild-type and Bid-deficient mice displayed symptoms of neurological disease. Concordantly, both strains of mice succumbed to reovirus-induced disease with equivalent frequency and a median survival time of 13 days. Reovirus titers in the brains of wild-type and Bid-deficient mice also were comparable at 4, 8, and 10 d post-inoculation (Figure 6B). These results indicate that following a high-dose inoculation, Bid is dispensable for reovirus growth in the murine CNS and attendant encephalitis.
Peak titers of reovirus in the brains of intracranially-inoculated wild-type mice were ∼1000-fold higher than those in perorally-inoculated wild-type animals (compare Figures 5E and 6B). We thought it possible that this difference in viral load might contribute to the dramatic difference in the requirement for Bid in the pathogenesis of reovirus-induced CNS disease following PO and IC inoculation. To test this hypothesis, we inoculated wild-type and Bid-deficient animals intracranially with a considerably lower but still lethal dose of T3SA+, 5 PFU, and monitored infected animals for signs of reovirus encephalitis (Figure 6C). In comparison to wild-type mice in which ∼95% succumbed to disease, ∼70% of Bid-deficient mice developed lethal encephalitis. Moreover, the median survival time of wild-type mice infected with T3SA+ was significantly less (13 days) than that of Bid-deficient mice (15 days). To determine whether this difference in survival correlates with the efficiency of reovirus replication in the CNS, we compared titers of reovirus in brains resected from infected mice at 4, 8, and 12 d post-inoculation (Figure 6D). Titers of reovirus in brains of wild-type mice were substantially higher (∼10- to 100-fold) at each interval in comparison to those in Bid-deficient mice, with those at 4 and 12 d post-inoculation reaching statistical significance. These findings indicate that at a lower viral inoculum, Bid promotes efficient replication of reovirus in the CNS. Collectively, these data suggest that Bid influences reovirus virulence by regulating the growth of reovirus in the brain.
To assess the capacity of T3SA+ to produce neurological injury in the presence and absence of Bid, we examined hematoxylin and eosin (H&E)-stained coronal brain sections prepared from wild-type and Bid-deficient mice euthanized 10 d following IC inoculation with 5 PFU of T3SA+ (Figure 7). This time point was chosen to coincide with the presence of maximal viral titers following inoculation by this route. Since the inoculum used for these experiments was at least ∼10-fold lower than that used for most other studies of reovirus CNS pathogenesis [4], [7], [11], [12], [43], [48], [51], the extent of injury following infection of wild-type mice was not as extensive. Nonetheless, inoculation of wild-type mice with T3SA+ resulted in neuronal death in the cerebral cortex, hippocampus, thalamus, and hypothalamus, consistent with previous reports [4], [7], [11], [12], [43], [48], [51]. While the majority of infected wild-type mouse brains showed signs of injury, tissue damage was minimal in all of the brains examined from similarly infected Bid-deficient animals. Examination of the hippocampal region of a representative wild-type mouse brain at higher magnification showed damage to the CA3 region, with the pyramidal cells showing condensed nuclei characteristic of apoptosis (Figure 7B). In contrast, little damage was detected in an equivalent region of a Bid-deficient mouse brain (Figure 7B). These findings indicate that Bid is required for neurological injury produced by reovirus in mice.
To determine whether these differences in neurological injury are attributable to alterations in tropism of reovirus in the absence of Bid, sections of mouse brain were stained for reovirus antigen. Reovirus displayed similar tissue distribution in wild-type and Bid-deficient mouse brains, indicating that Bid expression does not influence reovirus tropism (data not shown). The CA3 region of a wild-type mouse brain showed reovirus antigen in areas coincident with extensive neuronal damage (Figure 7B). Regions positive for reovirus antigen also stained with an antibody for activated caspase-3. Although similar regions of Bid-deficient mouse brains contained reovirus antigen, staining was of diminished intensity and frequency (Figure 7B), consistent with the decreased efficiency of reovirus replication in the CNS (Figure 6B). Accordingly, few cells showing intense caspase-3 staining were observed in regions that contained reovirus. These data suggest that neuronal apoptosis following reovirus infection is diminished in the absence of Bid. Thus, Bid links reovirus replication and apoptosis induction in the production of fatal encephalitis.
Early steps in reovirus replication elicit apoptosis via a signaling pathway dependent on NF-κB [4], [7], [8], [14], [15]. It is not understood how virus-induced NF-κB activation leads to cell death. In this study, we evaluated the function and regulation of Bid in apoptosis caused by reovirus. We found that although Bid is dispensable for reovirus replication in cell culture, it is required for the induction of apoptotic cell death following reovirus infection. In this context, Bid is converted to its active, proapoptotic form, tBid, in an NF-κB-dependent manner. Generation of tBid in reovirus-infected cells requires signaling via TRAIL-R and caspase-8. These findings indicate that NF-κB signaling following reovirus infection results in activation of the extrinsic apoptotic pathway. In turn, the extrinsic apoptotic pathway evokes the mitochondrial apoptotic cascade via cleavage-induced activation of Bid. Together, these events culminate in the induction of apoptotic cell death.
Many viruses induce apoptosis via activation of host-encoded apoptosis-regulating factors. For example, the VSV M protein induces apoptosis by inhibiting the transcription of antiapoptotic factors such as Bcl-xl [52]. In other cases, virus-encoded polypeptides insert into mitochondrial membranes and trigger cytochrome c release, leading to activation of the mitochondrial apoptotic pathway. For example, influenza A virus PB1-F2 is thought to directly activate proapoptotic signaling by interaction with the mitochondrial membrane-associated factors ANT3 and VDAC1 [53]. Although one model for apoptosis induction by reovirus suggests that the φ fragment of reovirus µ1 protein induces apoptosis by directly targeting mitochondria analogous to PB1-F2 [13], our studies using apoptosis-defective reovirus mutants [4], [7], coupled with data presented here, support the idea that φ-mediated NF-κB signaling activates the mitochondrial apoptotic pathway indirectly via death-receptor signaling and Bid cleavage. This indirect mechanism of mitochondrial pathway activation by a proximal signal transducer also explains the timing of prodeath signaling in the reovirus replication cycle. We think that events associated with viral entry into cells, which are mediated by the µ1 φ fragment subsequent to membrane penetration, activate NF-κB within 1 h of infection [7], [28]. Unlike other NF-κB agonists such as TNFα which rapidly and transiently activate NF-κB, activation of NF-κB following reovirus infection is gradual and sustained and occurs maximally at 6–8 h post infection [28]. Activated NF-κB complexes lead to expression of genes that promote cleavage-induced activation of Bid at 24–36 h post infection and elicit characteristic features of apoptosis, including effector caspase activation and DNA fragmentation. These changes occur subsequent to completion of viral replication, and, therefore, apoptosis appears to have little detectable effect on viral growth in cell culture [4], [7], [8], [28]. Although unusual, NF-κB-dependent apoptotic pathways are also utilized by other viruses such as Dengue virus [54], HIV [55], infectious bursal disease virus [56], and Sindbis virus [57]. Thus, our studies may have uncovered a potentially conserved signaling pathway utilized by viruses to induce apoptosis via NF-κB.
It is not known how activation of NF-κB by reovirus culminates in cell death. Three previous studies have attempted to identify proapoptotic host genes that serve as effectors of the death response following reovirus infection. In the first, gene-expression profiles following infection with reovirus strains type 1 Lang (T1L) and type 3 Abney (T3A), which differ in the capacity to induce apoptosis [58], were compared by microarray analysis [59]. These experiments did not demonstrate differences in expression of death ligands or their respective receptors following infection by either strain. Thus, it was concluded that expression of these death mediators by reovirus is unlikely to contribute to apoptosis induction by reovirus. However, some differences were observed in expression of regulators of death-receptor signaling [59]. But since T1L and T3A display significant genetic diversity and vary in the modulation of multiple signaling pathways [9], [60], the contribution of NF-κB to the expression of prodeath genes could not be established. In the second study, gene-expression profiles following T3D infection in the presence and absence of functional NF-κB were compared [61]. Although this study identified several NF-κB-dependent genes that coordinate the cellular antiviral immune response, including numerous interferon-stimulated genes (ISGs), no classical components of death receptor-mediated signaling pathways or proapoptotic Bcl-2 family members were significantly upregulated in response to reovirus infection. In the third study, gene-expression profiles of reovirus strains that differ in the capacity to elicit translational shutoff were compared [62]. This study also demonstrated an increase in ISG expression but did not identify obvious NF-κB-dependent candidates that could serve to activate death receptor signaling.
We think there are three possibilities to explain why apoptosis-regulating, transcriptional targets of NF-κB, such as death ligands (e.g., FasL and TRAIL) [63], [64], death receptors (e.g., Fas and DR) [65], [66], and death-signaling regulators (e.g., Bax and Bcl-Xs,) [67], were not identified in these studies. First, changes in the expression of prodeath genes activated by reovirus infection may be too transient to have been detected in the intervals selected for analysis. Second, the transformed nature of the cell lines used in these studies may not have been amenable to detection of alterations in gene expression induced by reovirus infection. Third, NF-κB activation following reovirus infection may regulate death signaling at a post-transcriptional level by an as yet unknown mechanism. In support of this idea, levels of DR5 protein increase following reovirus infection [17] but not its mRNA [59]. Additional studies using primary, non-transformed cell lines and genetically engineered viruses that differ only in the capacity to activate NF-κB are required to define how reovirus activates extrinsic apoptotic pathways to evoke cell death.
In addition to enhancing an understanding of mechanisms by which virus-induced signaling leads to activation of Bid, our studies highlight a critical role for Bid in controlling the pathogenesis of a viral disease. We found that Bid-deficient mice are less susceptible to lethal encephalitis produced by a neurotropic reovirus strain following either PO or IC inoculation. Reovirus replicates with slower kinetics in the absence of Bid, and virus-induced apoptosis and CNS injury are diminished in Bid-deficient animals. Although Bid contributes significantly to reovirus pathogenesis, our data do not allow us to determine whether diminished reovirus virulence in Bid-deficient animals is attributable to reduced capacity of reovirus to replicate in the CNS, diminished capacity of reovirus to injure neurons by apoptosis, or both effects. It is also not clear whether the decreased capacity of reovirus to evoke apoptosis in the CNS is a cause or effect of the lower viral titers at that site.
Because Bid serves to amplify the death response, it is not universally required for apoptosis induction. In some cell types, known as type I cells, caspase-8 activation results in direct, Bid-independent activation of the apoptosis effectors, caspase-3 and caspase-7 [29]. In others, known as type II cells, apoptosis requires amplification of death signals through stimulation of the mitochondrial pathway. In these cases, Bid serves to link the extrinsic and intrinsic apoptotic pathways [30]. Since the requirement for Bid in reovirus virulence is dependent on viral dose, we think that the role of Bid as an apoptosis regulator contributes to viral replication and consequent neurovirulence. Thus, we hypothesize that neurons function like type I cells when infected at a higher dose of virus and do not require the amplification of the mitochondrial apoptotic pathway via Bid to undergo apoptotic cell death. However, at lower infectious doses, neurons function like type II cells and require Bid-driven activation of the intrinsic mitochondrial apoptotic cascade to elicit cell death. This model also may explain why primary neuronal cultures infected with reovirus at a high MOI do not appear to require cytochrome c release and caspase-9 activation for apoptosis induction [42].
It is not known how Bid controls the efficiency of reovirus replication in the CNS. One possibility is that Bid-regulated apoptosis is required for efficient release of virus from neurons. Therefore, cell-to-cell spread of reovirus within the CNS may be inefficient in the absence of Bid. As an extension of this idea, blockade of apoptosis by other means also should cause a delay in reovirus replication. However, although symptoms of encephalitis are alleviated in NF-κB p50-deficient mice or in wild-type mice treated with a JNK inhibitor due to a reduction in virus-induced neuronal apoptosis, reovirus replication kinetics are not substantially diminished [11], [12]. Consistent with these findings, diminished virulence of apoptosis-defective reovirus mutants is not accompanied by significant decreases in reovirus replication efficiency in the CNS [4], [7]. Apoptosis following reovirus infection can occur in absence of p50, albeit at low efficiency [8]. Similarly, apoptosis-defective reovirus mutants retain some capacity to induce apoptotic cell death [4], [7]. Therefore, it is possible that in comparison to reovirus infection of Bid-deficient mice, CNS apoptosis was incompletely blocked in these other studies. Such a difference in the efficiency of apoptosis inhibition could explain the observed discrepancy in the requirement for Bid and other host or viral modulators of apoptosis for efficient replication of reovirus. A second possibility is that a Bid function not related to its capacity to regulate apoptosis contributes to reovirus replication in the CNS.
Analogous to its role in reovirus-induced cell death, Bid is implicated in apoptosis caused by many viruses [68], [69], [70], [71], [72], [73], [74], [75], [76], [77], [78], [79]. However, prior to our study, it was not known whether Bid modulates the pathogenesis of viral disease. The function of Bid in viral pathogenesis has been examined in a previous study, which found that the BH3-only protein, Puma, but not Bid, contributes to apoptosis-mediated elimination of antigen-specific T cells following acute infection with herpes simplex virus-1 [80]. Here, we demonstrate a pathogenic function for Bid in viral infection. Should Bid similarly modulate disease outcomes following infection by other virulent viruses, antiapoptotic compounds targeting Bid [81], [82], [83] may serve as useful antiviral therapeutics.
Murine L929 cells were maintained in Joklik's minimal essential medium supplemented to contain 10% fetal bovine serum (FBS), 2 mM L-glutamine, 100 U/ml penicillin, 100 µg/ml streptomycin, and 25 ng/ml amphotericin B (Invitrogen). Wild-type and Bid-deficient MEFs were maintained in Dulbecco's minimal essential medium (DMEM) supplemented to contain 10% FBS, 2 mM L-glutamine, 100 U/ml penicillin, 100 µg/ml streptomycin, and 25 ng/ml amphotericin B. TRAIL-R-deficient MEFs, prepared from D13 embryos, were maintained in DMEM supplemented to contain 10% FBS, 2 mM L-glutamine, 1× MEM nonessential amino acids, 0.1 mM 2-mercaptoethanol, 20 mM HEPES, 100 U/ml penicillin, 100 µg/ml streptomycin, and 25 ng/ml amphotericin B. Reovirus strain T3D is a laboratory stock. T3SA+ was generated by reassortment of reovirus strains T1L and type 3 clone 44-MA as described [84]. Purified reovirus virions were generated from second- or third-passage L-cell lysate stocks of twice-plaque-purified reovirus [85]. Viral particles were Freon-extracted from infected cell lysates, layered onto 1.2- to 1.4-g/cm3 CsCl gradients, and centrifuged at 62,000×g for 18 h. Bands corresponding to virions (1.36 g/cm3) were collected and dialyzed in virion-storage buffer (150 mM NaCl, 15 mM MgCl2, 10 mM Tris-HCl [pH 7.4]) [86].
Rabbit antisera raised against T1L and T3D have been described [87]. Rabbit antiserum specific for procaspase-9 was purchased from Cell Signaling. Goat antiserum specific for Bid was purchased from R & D systems, and goat antiserum specific for actin was purchased from Santa Cruz Biotechnology. HRP-conjugated anti-rabbit and anti-goat secondary antibodies were purchased from Amersham GE Biosciences. Alexa Fluor-conjugated anti-mouse immunoglobulin (Ig) G, anti-rabbit IgG, and anti-goat IgG secondary antibodies were purchased from Invitrogen.
Plasmids pRenilla-Luc and pNF-κB-Luc [88] were obtained from Dr. Dean Ballard (Vanderbilt University).
L929 cells or wild-type, Bid-deficient, or TRAIL-R-deficient MEFs were either adsorbed with reovirus at an MOI of 100 PFU/cell or mock-infected in serum-free medium at 4°C for 1 h, followed by incubation in serum-containing medium at 37°C for various intervals. Whole cell lysates were prepared by washing cells in phosphate-buffered saline (PBS) followed by lysis using 1× RIPA buffer (50 mM Tris [pH 7.5], 50 mM NaCl, 1% TX-100, 1% DOC, 0.1% SDS, and 1 mM EDTA) containing a protease inhibitor cocktail (Roche). Following centrifugation at 15,000×g to remove debris, the lysates were resolved by electrophoresis in polyacrylamide gels and transferred to nitrocellulose membranes. Membranes were blocked for at least 1 h in blocking buffer (PBS containing 5% milk or 2.5% BSA) and incubated with antisera against Bid (1∶1000), actin (1∶2000), or procaspase-9 (1∶1000) either at room temperature for 1 h or 4°C overnight. Membranes were washed three times for 10 min each with washing buffer (PBS containing 0.1% Tween-20) and incubated with1∶2000 dilution of horseradish peroxidase (HRP)-conjugated or Alexa Fluor-conjugated goat anti-rabbit Ig (for procaspase-9) or donkey anti-goat Ig (for Bid and actin) in blocking buffer. Following three washes, membranes were incubated for 1 min with chemiluminescent peroxidase substrate (Amersham Biosciences) and either exposed to film (for HRP-conjugated secondary antibodies) or scanned using an Odyssey Infrared Imager (LiCor).
Wild-type, Bid-deficient, or TRAIL-R-deficient MEFs (104) were seeded into black clear-bottom 96-well plates (Costar) and adsorbed with reovirus in serum-free medium at room temperature for 1 h. Following incubation of cells at 37°C for 24 h, caspase-3/7 activity was quantified using the Caspase-Glo-3/7 assay (Promega).
Wild-type, Bid-deficient, or TRAIL-R-deficient MEFs (5×104) were grown in 24-well plates (Costar) and adsorbed with reovirus at room temperature for 1 h. The percentage of apoptotic cells after 48 h incubation was determined using AO staining as described [6]. For each experiment, >200 cells were counted, and the percentage of cells exhibiting condensed chromatin was determined by epi-illumination fluorescence microscopy using a fluorescein filter set (Zeiss Photomicroscope III; Thornwood, NY).
Wild-type or Bid-deficient cells (2×105) were grown in 24-well plates and adsorbed with reovirus at room temperature for 1 h. Following removal of the inoculum, cells were washed with PBS and incubated in complete medium at 37°C for 18 h. Monolayers were fixed with methanol, washed twice with PBS, blocked with 2.5% Ig-free bovine serum albumin (Sigma-Aldrich) in PBS, and incubated successively for 1 h with polyclonal rabbit anti-reovirus serum at a 1∶1000 dilution and for 1 h with Alexa Fluor 546-labeled anti-rabbit IgG at a 1∶1000 dilution. Monolayers were washed with PBS, and infected cells were visualized by indirect immunofluorescence using a Zeiss Axiovert 200 fluorescence microscope. Reovirus antigen-positive cells were quantified by counting fluorescent cells in at least two random fields of view in triplicate wells at a magnification of 20×.
Wild-type, Bid-deficient, or TRAIL-R-deficient MEFs (2×105) in 24-well plates were adsorbed with reovirus at room temperature for 1 h in serum-free medium, washed once with PBS, and incubated in serum-containing medium for various intervals. Cells were frozen and thawed twice prior to determination of viral titer by plaque assay using L929 cells [89].
Wild-type and Bid-deficient cells in 24-well plates were transfected with 0.72 µg/well of an NF-κB reporter plasmid, which expresses firefly luciferase under NF-κB control (pNF-κB-Luc), and 0.08 µg/well of control plasmid pRenilla-Luc, which expresses Renilla luciferase constitutively, using Fugene6 (Roche). After incubation for 24 h, transfected cells were adsorbed with reovirus in serum-free medium at room temperature for 1 h and incubated at 37°C in serum-containing medium for 24 h. Luciferase activity in the cultures was quantified using the Dual-Luciferase Assay Kit (Promega) according to the manufacturer's instructions.
Wild-type C57BL/6J mice were obtained from Jackson Laboratory. Bid-deficient mice backcrossed on to a C57BL/6J background for at least 8 generations have been previously described [30]. Two-day-old mice were inoculated either perorally or intracranially with purified virus diluted in PBS. PO inoculations were delivered in a volume of 50 µl by passage of a polyethylene catheter 0.61 mm in diameter (BD) through the esophagus and into the stomach [90]. The inoculum contained 0.3% (vol/vol) green food coloring to allow the accuracy of delivery to be judged. IC inoculations were delivered to the left cerebral hemisphere in a volume of 5 µl using a Hamilton syringe and a 30-gauge needle (BD Biosciences) [91]. For analysis of viral virulence, mice were monitored for weight loss and symptoms of disease for 21 days. For survival experiments, mice were euthanized when found to be moribund (defined by rapid or shallow breathing, lethargy, or paralysis). For determination of viral titer and immunohistochemical staining, mice were euthanized at various intervals following inoculation and organs were resected. For analysis of virus growth, organs were collected into 1 ml of PBS and homogenized by freezing, thawing, and sonication. Viral titers in organ homogenates were determined by plaque assay using L929 cells [89]. For immunohistochemical staining, organs were fixed overnight in 10% formalin, followed by incubation in 70% ethanol. Fixed organs were embedded in paraffin, and 5-µm histological sections were prepared. Consecutive sections were stained with H&E for evaluation of histopathologic changes or processed for immunohistochemical detection of reovirus antigens or activated caspase-3 [11]. Animal husbandry and experimental procedures were performed in accordance with Public Health Service policy and the recommendations of the Association for Assessment and Accreditation of Laboratory Animal Care and approved by the Vanderbilt University School of Medicine Institutional Animal Care and Use Committee.
|
10.1371/journal.pgen.1007085 | A dominant-negative mutant inhibits multiple prion variants through a common mechanism | Prions adopt alternative, self-replicating protein conformations and thereby determine novel phenotypes that are often irreversible. Nevertheless, dominant-negative prion mutants can revert phenotypes associated with some conformations. These observations suggest that, while intervention is possible, distinct inhibitors must be developed to overcome the conformational plasticity of prions. To understand the basis of this specificity, we determined the impact of the G58D mutant of the Sup35 prion on three of its conformational variants, which form amyloids in S. cerevisiae. G58D had been previously proposed to have unique effects on these variants, but our studies suggest a common mechanism. All variants, including those reported to be resistant, are inhibited by G58D but at distinct doses. G58D lowers the kinetic stability of the associated amyloid, enhancing its fragmentation by molecular chaperones, promoting Sup35 resolubilization, and leading to amyloid clearance particularly in daughter cells. Reducing the availability or activity of the chaperone Hsp104, even transiently, reverses curing. Thus, the specificity of inhibition is determined by the sensitivity of variants to the mutant dosage rather than mode of action, challenging the view that a unique inhibitor must be developed to combat each variant.
| Prion proteins adopt alternative conformations and assemble into amyloid fibers, which have been associated with human disease. These fibers are highly stable and self-replicate, leading to their persistence and resulting in a set of progressive and often fatal disorders. Inhibitors have been shown to interfere with some conformations but not others, suggesting that distinct strategies must be developed to target each. However, we show here that a single dominant-negative mutant can inhibit multiple conformations of the same prion protein through the same pathway but at distinct doses. Thus, the basis of this specificity is sensitivity rather than resistance to the mechanism of inhibition, suggesting that common strategies may be used to target a range of prion conformations.
| Alternative, self-replicating protein conformations have emerged as bona fide parallel protein-folding trajectories with significant biological consequences [1]. In most cases, these alternative conformations are β-sheet-rich and self-assembling, forming linear amyloid aggregates [2]. These amyloids replicate the conformation of their constituent monomers by acting as templates to direct the refolding of other conformers of the same protein as they are bound by and incorporated into the growing aggregate. In so doing, the majority of the protein is converted to the alternative conformation, changing protein activity and thereby inducing new phenotypes, such as neurodegenerative diseases (i.e., Transmissible Spongiform Encephalopathies or prion diseases, Alzheimer’s and Huntington’s diseases) and organelle biogenesis in mammals and gene expression regulation in single-celled organisms [1,3]. The high efficiency of this process, when combined with the high kinetic stability of the aggregates [2], contributes to the recalcitrance of amyloids to clearance by protein quality control pathways [4]. As a result, the associated phenotypes are frequently difficult—if not impossible—to reverse, especially in the clinic [5].
One notable exception to the persistence of amyloid-associated phenotypes is their reversal or “curing” by dominant-negative mutants of prion proteins. These sequence variants were first identified by their ability to confer resistance to scrapie in sheep (Q171R or R154H in the mammalian prion protein PrP), sporadic Creutzfeldt-Jakob disease (sCJD) in humans (E219K in PrP), and translation termination infidelity in yeast (G58D in Sup35) [6–19]. Subsequently, these mutants were shown to interfere with the assembly of amyloid by the wildtype prion protein in vitro and to reduce or clear existing amyloid composed of the wildtype prion protein when delivered to tissue culture cells, mice, or yeast [15,19–31]. Given this unique curing ability, elucidating the mechanism(s) by which dominant-negative prion mutants act may reveal potential strategies for reversing amyloid persistence more generally.
Despite the promise of this line of investigation, the inhibition achieved by dominant-negative mutants appears to be conformation-specific. For example, the resistance to sCJD conferred by the E219K PrP mutant in humans is not extended to the conformations, known as variants, responsible for genetic and iatrogenic forms of the disease [14,15,17,32–35]. Similarly, resistance to classical scrapie is not observed for bovine spongiform encephalopathy (BSE) or atypical scrapie variants in sheep with Q171R or R154H mutations in PrP [10,36–43] [44–52]. Finally, the G58D mutation of Sup35 cures the [PSI+]Strong and [PSI+]Sc4 variants (n.b. [PSI+] denotes the transmissible amyloid state of Sup35) to different extents in different genetic backgrounds but is unable to cure the [PSI+]Sc37 and [PSI+]Weak variants in yeast [53,54].
What is the molecular basis of this differential inhibition? One possibility is that the distinct recognition surfaces and/or rate-limiting steps in the self-replication process characteristic of the variants make them susceptible to only certain mechanisms of inhibition [55–61]. Alternatively, the conformational differences may confer distinct sensitivities to the same mechanism of inhibition. Given the conformational plasticity of amyloidogenic proteins [62,63], understanding the forces limiting the efficacy of inhibitors can mean the difference between developing an infinite number of individual interventions for each variant or simply different dosing regimes for the same inhibitor.
Here, we exploit the yeast prion Sup35 to gain this insight. We explored the sensitivity of three variants of Sup35 (i.e., [PSI+]Sc4, [PSI+]Weak, and [PSI+]Sc37) to expression of G58D and the impact of this dominant-negative mutant on the self-replication of each variant. Our studies indicate that “resistance” to G58D can be partially overcome at higher dosage of the mutant, revealing differential sensitivity to the inhibition. G58D reduces the kinetic stabilities of the amyloids associated with the variants, which determines their efficiencies of fragmentation by chaperones [60]. Consistent with this correlation, G58D inhibition of the three variants was dependent on the chaperone Hsp104, as was the case for the previously studied [PSI+]Strong variant [64]. In the presence of G58D, Sup35 amyloid was fragmented by Hsp104 with higher efficiency. This increase led to amyloid clearance in daughter cells, which could be reversed by transient inhibition of Hsp104 specifically in this population. Thus, G58D dominant-negative inhibition targets distinct conformational variants through the same mechanism with differing efficacy, suggesting that the observed “resistance” is relative rather than absolute.
To determine if the specificity of G58D on [PSI+] variants occurs through distinct mechanisms or through distinct sensitivities to the same mechanism of inhibition, we generated diploid [PSI+]Sc4, [PSI+]Weak and [PSI+]Sc37 yeast strains expressing wildtype Sup35 at different ratios relative to G58D (2:1, 1:1, 1:2; S1 Fig). Inhibition of [PSI+] propagation can be monitored functionally because the formation of amyloid by Sup35 partially compromises its activity and leads to a defect in translation termination [65,66]. [PSI+] strains carrying the ade1-14 allele form white colonies on rich medium due to read-through of a premature stop codon in the ADE1 open reading frame. However, strains with defective prion propagation, or those that have lost the prion state (known as [psi-]), form red colonies on rich medium as a result of the accumulation of active Sup35 [67].
Expression of G58D at any ratio in a [PSI+]Sc4 strain promoted the accumulation of red pigment on rich medium, indicating reversal of the prion phenotype (Fig 1A). By colony color, the severity of this effect increased with G58D dosage, with a 1:2 ratio of wildtype to G58D leading to a colony phenotype for [PSI+]Sc4 that was indistinguishable from [psi-] (Fig 1A). For the [PSI+]Sc37 and [PSI+]Weak variants, which were previously reported to be compatible with G58D expression [53,54], efficient prion propagation was also dependent on the ratio of wildtype to G58D, but the critical threshold for phenotypic reversal was distinct in each case. The [PSI+]Sc37 variant formed colonies that were more pink on rich medium at a 1:1 ratio of wildtype to G58D relative to a wildtype strain and that were indistinguishable from [psi-] at a 1:2 ratio of wildtype to G58D (Fig 1B), mirroring our observations for [PSI+]Sc4 (Fig 1A). In contrast, the [PSI+]Weak variant phenotype was only partially reversed at the highest ratio of wildtype to G58D tested (1:2), where the pinker colonies on rich medium relative to the wildtype [PSI+]Weak strain indicated a mild inhibition by G58D (Fig 1C). Thus, the three [PSI+] variants are each dominantly inhibited by G58D expression in a dose-dependent manner, but the dose required for inhibition of [PSI+]Sc4 and [PSI+]Sc37 is lower than that of [PSI+]Weak.
To assess whether reversal of the [PSI+] phenotype upon G58D expression reflected prion loss (i.e., curing), we determined the frequencies of [psi-] appearance during mitotic division for each strain. [PSI+] propagation was largely stable at the 2:1 (~0% curing) and 1:1 (~1% curing) ratios of wildtype to G58D for both [PSI+]Sc4 and [PSI+]Sc37, where the colony phenotype was only mildly reversed (Fig 1A, 1B, 1D and 1E). At a 1:2 ratio of wildtype to G58D, both [PSI+]Sc4 (~9% curing, Fig 1D) and [PSI+]Sc37 (~8% curing, Fig 1E) were more unstable, consistent with the stronger reversal of their prion phenotypes at this ratio (Fig 1A and 1B). For the [PSI+]Weak variant, which is less sensitive to G58D inhibition (Fig 1C), [PSI+] propagation was stable at all wildtype:G58D ratios tested (Fig 1F). Thus, [PSI+] curing in diploids expressing G58D parallels the severity of the phenotypic reversal in all three variants and, for the most sensitive strains (i.e., [PSI+]Sc4 and [PSI+]Sc37), arises in a dose-dependent manner. Together, these observations indicate that the previously described “resistance” of [PSI+]Sc37 and [PSI+]Weak to curing by G58D expression reflected their higher threshold for sensitivity rather than their absolute recalcitrance to inhibition by this mutant.
Although the three [PSI+] variants studied here, in addition to the previously studied [PSI+]Strong variant [64], differ in their sensitivities to G58D inhibition (Fig 1), the dose dependence of this inhibition suggests a common underlying mechanism [64,68]. We previously linked G58D inhibition to a reduction in the kinetic stability of Sup35 aggregates and a resulting increase in their fragmentation by the chaperone Hsp104, which led to their disassembly [64]. In this model, the distinct effective inhibitory ratios of G58D on [PSI+] variants may reflect the impact that this mutant has on the kinetic stability of each. While it has been well-established that Sup35 aggregates in the [PSI+]Sc4 conformation are of lower stability than those in the [PSI+]Sc37 conformation, the relative stabilities of the four variants have not been previously reported [60,69,70].
To gain this insight, we first determined the kinetic stabilities of Sup35 aggregates, in the absence of G58D, by their sensitivity to disruption with 2% SDS at different temperatures as a baseline comparison [71]. Solubilized protein is then quantified by entry into a SDS-polyacrylamide gel and immunoblotting [64]. For wildtype strains, Sup35 was efficiently released from aggregates between 65°C and 75°C in lysates from strains propagating the [PSI+]Strong and [PSI+]Sc4 variants (Fig 2A) or between 70°C and 90°C in lysates from strains propagating the [PSI+]Weak and [PSI+]Sc37 variants (Fig 2B). The higher kinetic stability of the latter variants is consistent with their lower efficiency of fragmentation, which leads to a larger steady-state size for their associated amyloids as assessed by semi-denaturing agarose gel electrophoresis (SDD-AGE) and immunoblotting for Sup35 (S2 Fig) [60,72].
To sensitize the assay in an attempt to reveal biochemical differences between the variants in each group, we deleted the NATA N-terminal acetyltransferase, which reduces the kinetic stability of Sup35 amyloid in [PSI+] strains [73,74]. In this genetic background, the fraction of soluble Sup35 released from amyloid of the [PSI+]Strong variant in the presence of SDS was significantly increased relative to that from the [PSI+]Sc4 variant over the same temperature range (Fig 2C), indicating that the aggregates are less kinetically stable in the [PSI+]Strong than the [PSI+]Sc4 variant. Similarly, a significantly larger fraction of Sup35 was released from amyloid in the presence of SDS from the [PSI+]Sc37 variant than from the [PSI+]Weak variant (Fig 2D), indicating that the aggregates are less kinetically stable in the [PSI+]Sc37 than the [PSI+]Weak variant. Thus, the kinetic stability of Sup35 aggregates in [PSI+] variants increases in the order [PSI+]Strong, [PSI+]Sc4, [PSI+]Sc37, [PSI+]Weak.
If G58D inhibits these variants through a common mechanism, we would expect the kinetic stabilities of each of the variants to decrease in the presence of the mutant. To test this possibility, we assessed the sensitivity of Sup35 aggregates, isolated from diploid strains expressing a 1:1 ratio of wildtype to G58D, to disruption with 2% SDS at different temperatures. Soluble protein was then quantified by entry into an SDS-polyacrylamide gel and immunoblotting for Sup35. For the [PSI+]Sc4 strain, G58D expression increased the amount of soluble Sup35 released from aggregates at all temperatures assayed (65°C, 70°C and 75°C) in comparison with a wildtype strain (Fig 3A). G58D similarly promoted Sup35 release from aggregates isolated from the [PSI+]Sc37 (Fig 3B) and [PSI+]Weak (Fig 3C) strains at 80°C and 85°C, but the magnitude of this effect was greater for the former. Thus, G58D incorporation destabilizes Sup35 aggregates from [PSI+] variants in a manner that correlates directly with the severity of their phenotypic inhibition (Fig 1). These observations are consistent with the idea that G58D acts through a similar mechanism to inhibit the [PSI+] variants.
A decrease in the kinetic stability of amyloid should increase its efficiency of fragmentation and potentially lead to its clearance. To begin to determine the effects of G58D on the fragmentation of Sup35 amyloid associated with these [PSI+] variants, we first assessed the steady-state size distributions of these complexes by SDD-AGE and immunoblotting for Sup35. As we have previously observed for [PSI+]Strong [64], expression of G58D at any ratio relative to wildtype Sup35 in a [PSI+]Sc4 strain led to a decrease in the accumulation of slowly migrating aggregates in comparison to the same dose of wildtype protein alone (Fig 3D). For [PSI+]Sc37, similar decreases were observed (Fig 3E), but for [PSI+]Weak, Sup35 aggregates were only shifted to smaller complexes at the lowest wildtype to G58D ratio tested (1:2, Fig 3F). Together, these observations suggest that the kinetic destabilization of Sup35 aggregates by G58D results in a higher efficiency of fragmentation in vivo, and these effects correlate directly with the severity of their phenotypic inhibition (Fig 1).
To determine how the kinetic destabilization of Sup35 aggregates by G58D impacts the number of heritable prion units (propagons) in [PSI+]Sc4, [PSI+]Sc37 and [PSI+]Weak strains, we used a genetic assay [75]. Specifically, diploid strains expressing either two copies of wildtype SUP35 or one copy each of wildtype SUP35 and G58D were treated with guanidine HCl (GdnHCl), a potent inhibitor of the fragmentation catalyst Hsp104 [67,76–81], allowed to dilute existing aggregates through cell division, and plated in the absence of the inhibitor to quantify the number of cells inheriting an aggregate. As we have previously observed in a [PSI+]Strong strain [64], G58D expression in either [PSI+]Sc4 and [PSI+]Sc37 diploids reduced propagon number by factors of ~2 and ~4, respectively (Fig 3G and 3H), consistent with the reversal of the [PSI+] phenotype and the loss of [PSI+] that we observed in these strains (Fig 1A, 1B, 1D and 1E). In contrast, G58D expression in [PSI+]Weak increased propagon number by a factor of ~2.5 (Fig 3I). Although we did not detect any changes in the severity or stability of the [PSI+]Weak phenotype at this ratio (Fig 1C and 1F), this increase in propagon count provides an explanation for the previously reported strengthening of the [PSI+]Weak phenotype upon G58D expression to much higher levels [53]. Phenotypic strengthening is associated with a decrease in soluble Sup35, which would result from an increase in amyloid templates, detected as propagons in this assay, through enhanced fragmentation [60]. Thus, the phenotypic consequences of G58D expression, both inhibition and enhancement, can be directly explained by changes in the steady-state accumulation of Sup35 propagons. Given the distinct kinetic stabilities of Sup35 amyloid in the [PSI+] variants studied here (Fig 2), the specificity of G58D inhibition and enhancement likely reflect thresholds for fragmentation activity that result in changes in the steady-state accumulation of Sup35 forms in vivo.
If enhanced fragmentation is indeed the mechanism underlying G58D effects, these changes should be Hsp104-dependent. To determine if this is the case, we constructed heterozygous disruptions of HSP104 in diploid strains expressing G58D at different ratios (S3 Fig). In strains expressing only wildtype Sup35, heterozygous disruption of HSP104 significantly decreased the number of propagons in the [PSI+]Sc4 and [PSI+]Sc37 variants tested (Fig 3G and 3H, compare lanes 1 and 3), consistent with its catalytic role in fragmentation [78,79] and the size threshold for Sup35 aggregate transmission [72]. In contrast, heterozygous disruption of HSP104 in [PSI+] variant strains expressing both wildtype and G58D Sup35 increased the number of propagons (Fig 3G and 3I, compare lanes 2 and 4). Thus, the reduction in propagon number associated with G58D is suppressed by lowering the dosage of HSP104 and thereby fragmentation activity, suggesting that enhanced fragmentation is the underlying mechanism.
Next, we determined if these changes in propagon number upon heterozygous disruption of HSP104 impacted the severity and stability of the [PSI+] phenotype. Heterozygous disruption of HSP104 restored the [PSI+] phenotype (Fig 1A) and efficiently suppressed [PSI+] loss (Fig 1D) in the [PSI+]Sc4 strains expressing any ratio of G58D. For the [PSI+]Sc37 and [PSI+]Weak variants, similar although attenuated trends were apparent. Heterozygous disruption of Hsp104 partially reversed the pinker colony color on rich medium for both [PSI+]Sc37 and [PSI+]Weak (Fig 1B and 1C). For [PSI+]Sc37, heterozygous disruption of Hsp104 increased [PSI+] loss in all strains, indicating that wildtype fragmentation levels must be close to the threshold required for efficient propagation of the amyloid state (Fig 1E). Nonetheless, in the strain expressing the 1:2 ratio of wildtype to G58D, the frequency of [PSI+] loss was suppressed by heterozygous disruption of Hsp104 (Fig 1E). Thus, reduction of Hsp104 reverses the G58D-induced inhibition of the [PSI+] phenotype. Together, these observations are consistent with the idea that the downstream effect of G58D is identical for all [PSI+] variants: an enhancement of the fragmentation efficiencies of their Sup35 amyloid.
The enhanced efficiency of fragmentation of Sup35 aggregates in the presence of G58D (Fig 3D and 3E) and the reduction in propagon levels (Fig 3G and 3H) suggests that Sup35 aggregates are being destroyed in strains propagating the [PSI+]Sc4 and [PSI+]Sc37 variants. For [PSI+]Strong, we previously detected this disassembly by monitoring the soluble pool of Sup35 in response to cycloheximide treatment to follow the fate of existing protein [64]. However, [PSI+]Strong is more sensitive to G58D expression than [PSI+]Sc4, [PSI+]Sc37 and [PSI+]Weak (Fig 1) [64], suggesting that release of soluble Sup35 from aggregates by enhanced fragmentation may be less readily detected in the latter variants. Specifically, the individual steps in prion propagation in vivo (e.g. conversion, fragmentation, and transmission) are variant-specific and difficult to monitor in isolation in a living system [60,78]. Moreover, the accumulation of soluble Sup35 is impacted not only by the inherent rate of conversion on fibers ends but also by the cumulative effect of each of the steps of prion propagation on the number of those ends [60,72]. Because the cumulative effects of each event on soluble Sup35 levels are not intuitive to qualitatively predict from those rates, we developed a deterministic model of Sup35 dynamics to deconstruct this complexity and gain additional mechanistic insight into the differential effects of G58D on the variants. This model uses a range of conversion and fragmentation rates that support [PSI+] maintenance to capture different variants (see S1 Text). In addition, we have incorporated the concept of nucleation, which specifies a minimum size for a thermodynamically stable aggregate and has been previously established as a key event in Sup35 aggregation in vitro [82–84].
The steady-state size and number of Sup35 aggregates reflects a balance between conversion, which depends on continuous synthesis of Sup35, and fragmentation [72]; when Sup35 synthesis is halted, aggregates are predicted to increase in number (Fig 4A) and decrease in size (Fig 4B) because fragmentation is proposed to exert a greater influence on the equilibrium state [72]. In line with this observation, our model predicts that cycloheximide treatment will decrease soluble Sup35 levels for prion variants that are stably propagating [PSI+] (Fig 4C) because additional templates have been created (Fig 4A).
Intriguingly, the extent of this decrease is predicted in our mathematical model to correspond inversely with the rate of fragmentation: that is, the slowest rate of fragmentation induces the largest decrease in soluble Sup35 (Fig 4B, black), relative to the steady-state levels prior to the manipulation. If fragmentation produces more templates, which in turn promotes Sup35 conversion to the amyloid state, why would we predict a lower rate of fragmentation to have the most significant effect on soluble Sup35 levels? The reason is, as we have previously demonstrated under heat shock conditions [85], fragmentation resolubilizes Sup35 in addition to creating new templates. Thus, high rates of fragmentation will push the balance between conversion and fragmentation toward the latter, causing a shift from aggregated to soluble Sup35. Consistent with this logic, our model predicts an increase in aggregate number that corresponds inversely with fragmentation rate (i.e. the largest increase in aggregate number corresponds to the slowest fragmentation rate; Fig 4A, black). This correlation can be explained directly by changes in the rate of Sup35 resolubilization from aggregates: the slowest fragmentation rate leads to the slowest rate of resolubilization (Fig 4D, black) and thereby the largest increase in aggregate number (Fig 4A, black).
These predictions correlate with our observations of the [PSI+]Sc4, [PSI+]Sc37, and [PSI+]Weak variants upon treatment with cycloheximide. For strains where wildtype Sup35 was the only form present, the average size of Sup35 aggregates decreased (Fig 5A–5C). In addition, the level of soluble Sup35 decreased upon cycloheximide treatment for the [PSI+]Weak and [PSI+]Sc37 variants, but no significant decrease was observed for [PSI+]Sc4 variant (Fig 5D–5F, lane 1). According to our model, these observations are consistent with a nucleation-dependent aggregation process, which permits resolubilization of aggregates that are fragmented below the minimum thermodynamically stable size (Fig 4D, compare solid and dashed lines), and a higher rate of fragmentation for [PSI+]Sc4, which would release more aggregated Sup35 into the soluble pool (Fig 4D, red). In the presence of G58D, soluble Sup35 levels in [PSI+]Sc37 and [PSI+]Weak are no longer reduced (Fig 5E and 5F, compare lanes 1 and 3), suggesting that G58D expression promotes aggregate fragmentation and thereby resolubilization. Consistent with this idea, treatment of the variants with both cycloheximide and guanidine HCl led to an increase in aggregate size (Fig 5A–5C) and a decrease in soluble Sup35 levels in the presence of G58D (Fig 5D–5F, compare lanes 3 and 4), indicating that Hsp104-catalyzed fragmentation promotes Sup35 resolubilization.
The ability of our mathematical model to capture the behavior of Sup35 in response to these manipulations strongly supports the idea that G58D destabilizes Sup35 aggregates to promote their increased fragmentation by Hsp104 and thereby their resolubilization. However, a more nuanced evaluation indicates that the threshold for inhibition cannot be explained by fragmentation efficiency alone. For example, [PSI+]Sc37 has a similar phenotypic sensitivity to G58D dosage as the [PSI+]Sc4 variant (Fig 1) but a kinetic stability, size, and likely fragmentation efficiency closer to the [PSI+]Weak variant (Fig 2 and S2 Fig). A bulk shift in Sup35 from aggregate to soluble requires that the resolubilized Sup35 does not efficiently reconvert to the aggregated state; thus, conversion efficiencies will also impact the outcome of the G58D effects on aggregate kinetic stability, fragmentation and resolubilization. Sup35 aggregates in the [PSI+]Sc37 conformation direct conversion at a higher rate than those in the [PSI+]Sc4 conformation [60], but the relative rates of conversion for [PSI+]Sc37 and [PSI+]Weak have not been reported. To compare these variants, we transiently treated strains with GdnHCl in liquid culture to reduce propagon number and then monitored propagon recovery as a function of time after removal of GdnHCl by plating cells and assessing their colony-color phenotype. The [PSI+]Weak variant amplified its propagons at a faster rate than the [PSI+]Sc37 variant (S4 Fig). This recovery rate is a function of the product of the conversion and fragmentation rates [60]. Because Sup35 aggregates in the [PSI+]Sc37 conformation are less kinetically stable than those in the [PSI+]Weak conformation (Fig 2B and 2D) and thereby likely fragmented at a higher rate, this observation suggests that the conversion rate of [PSI+]Sc37 is much lower than that of [PSI+]Weak. As a result, resolubilized Sup35 would be less likely to reconvert to the aggregated state in the [PSI+]Sc37 variant than in the [PSI+]Weak variant. Thus, the higher rate of resolution and the lower rate of conversion combine to increase the sensitivity of [PSI+]Sc37 to G58D inhibition relative to [PSI+]Weak.
Together, our studies are consistent with the ideas that resolubilization of aggregated Sup35 is the mechanism of G58D inhibition and that the variant-specific rates of conversion and fragmentation dictate the threshold for phenotypic reversal. However, Weissman and colleagues previously reported that loss of [PSI+]Sc4 propagated by G58D alone was associated with propagon loss from daughter but not mother cells [54]. This observation was interpreted as a G58D-dependent defect in Sup35 aggregate transmission to daughter cells, but using a direct fluorescence-based microscopy assay for Sup35-GFP transmission, we were unable to detect a transmission defect in [PSI+]Strong strains expressing wildtype and G58D Sup35 [64]. The appearance of daughter cells without propagons could also arise if Sup35 aggregates were transmitted but subsequently disassembled by Hsp104 in this compartment. If this scenario is correct, inhibition of Hsp104 will lead to an increase in [PSI+] propagons in daughter cells. To test this hypothesis, we constructed [PSI+]Sc4 diploid strains expressing only G58D Sup35 and compared prion propagation in wildtype and HSP104 heterozygous disruption versions of this strain by plating on rich medium and observing colony-color phenotype. Consistent with previous observations [54], [PSI+]Sc4 propagation is unstable in a wildtype strain (~50% prion loss), but we found that this instability is strongly suppressed by heterozygous disruption of HSP104 (~5% prion loss; Fig 6A).
Propagons are normally distributed between mother and daughter cells in a 2:1 ratio [75]. However, analysis of propagon numbers in mother and daughter cells showed an even stronger bias in the distribution of propagons toward the mothers in the presence of G58D (Fig 6B, black diamonds), including a population of pairs in which the mother but not the daughter retained a large number of propagons (Fig 6B, red diamonds). By contrast, heterozygous disruption of HSP104 reduced the stronger mother bias associated with G58D expression, and more propagons were detected in daughter cells (Fig 6B, white triangles). Notably, daughter cells lacking propagons were not isolated from the HSP104 heterozygous disruption strain, indicating that the suppression of prion loss (Fig 6A) correlated with an increase in propagons in daughter cells (Fig 6B).
Given the suppression of these phenotypes by heterozygous disruption of Hsp104, we next directly determined if Hsp104 inhibition specifically in daughter cells is sufficient to suppress [PSI+] loss. To do so, we isolated daughter cells from [PSI+]Sc4 diploids expressing one copy each of wildtype and G58D SUP35 by FACS, based on the staining of bud scars with Alexa-647 WGA. The absence of bud scars in cells with the lowest fluorescence intensity indicates that this fraction contains the newborn population, in contrast to a mixed population before sorting (Fig 6C and S5 Fig). The isolated daughters were then incubated on rich medium in the presence or absence of GdnHCl for three hours to transiently inhibit Hsp104 activity and then plated to determine the frequency of prion loss. Strikingly, GdnHCl treatment of daughter cells suppressed the frequency of prion loss (Fig 6D). Because daughter cells were biochemically isolated before treatment, the GdnHCl suppression of prion loss cannot be explained by an increased transmission of Sup35 aggregates to daughter cells upon Hsp104 inhibition. Rather, Sup35 aggregates must have already been present, with the transient inhibition of Hsp104 blocking their resolubilization after transfer, consistent with the idea that G58D inhibits the propagon of all [PSI+] variants through the same mechanism.
Together, our studies indicate that a single inhibitor, the dominant-negative G58D mutant of Sup35, can perturb the propagation of four different variants of the [PSI+] prion, [PSI+]Strong, [PSI+]Sc4, [PSI+]Sc37, and [PSI+]Weak, through the same mechanism. The effects of G58D expression are most easily detected at the protein level as kinetic destabilization of Sup35 amyloid (Fig 3A–3C) and related reductions in the size of their SDS-resistant core polymers (Fig 3D–3F). These changes only become apparent at the phenotypic and inheritance levels when the impact on Sup35 amyloid rises above a threshold dictated by the rates of conversion and fragmentation for the variants, allowing disassembly to dominate over reassembly.
The G58D mutation lies in the second oligopeptide repeat of Sup35, a region of the protein that is essential for prion propagation [86–88] and that impacts the ability of the Hsp104 chaperone to thread monomers through its central pore during the fragmentation process [89]. Position 58 is located within the amyloid core of Sup35 in the [PSI+]Sc37 variant but is more accessible in the [PSI+]Sc4 variant [69]. Nonetheless, the kinetic destabilization of the four variants by G58D (Fig 3A–3C) [64] suggests this region contributes directly to associations within each of the aggregates. Structural studies on the isolated second repeat revealed that the G58D substitution introduced a turn into the otherwise extended conformation of the wildtype repeat, suggesting that packing and thereby amyloid kinetic stability could be altered by this conformational change [90].
Previous studies on the [PSI+]Strong and [PSI+]Sc4 conformational variants suggested two different mechanisms for G58D-induced curing. For [PSI+]Strong, curing depended not only on the dosage of G58D but also of HSP104, suggesting that prion propagation was inhibited by amyloid disassembly. Indeed, in the presence of G58D, previously aggregated Sup35 transitioned to the soluble fraction [64]. For [PSI+]Sc4, curing correlated with the loss of heritable aggregates in daughter cells, interpreted as a G58D-induced defect in amyloid transmission [54]. These distinct models for inhibition are consistent with the idea that different conformational variants must be cured through different molecular mechanisms [62,63]. However, our studies resolve this controversy: G58D inhibits both variants by promoting amyloid disassembly in daughter cells. This model is supported by both the Hsp104-dependence of the curing of both variants (Fig 1D) [64] and of the reduction in propagons (Fig 3H) [64]. In addition, overexpression of Hsp104 cures [PSI+]Sc4 propagated by G58D but not wildtype Sup35, suggesting the former is more sensitive to higher fragmentation rates than the latter [54]. Consistent with this interpretation, overexpression of an N-terminally truncated Hsp104 mutant [54], which is deficient in substrate processing [91], is unable to cure [PSI+]Sc4 propagated by G58D.
We have previously drawn parallels between the dominant-negative inhibition of [PSI+] propagation by Sup35 G58D and that of protease-resistant PrP by hamster Q219K (corresponding to E219K in humans and Q218K in mouse). In both cases, the mutant is incorporated into wildtype aggregates but capable of destabilizing the amyloid state only when present in excess to wildtype protein, and the efficacy of dominant-negative inhibition is greater for less kinetically stable conformational variants [15,20,21,64,92]. Given the likelihood that the mechanisms of inhibition are similar between the yeast and mammalian dominant-negative mutants, the “resistance” of sCJD to E219K in humans and of 22L to Q219K in mice may be possible to overcome by increasing the dosage of the mutant, as we have demonstrated here for G58D and [PSI+]Sc37 (Fig 1B and 1E). For G58D, inhibition occurs at a dosage far below that at which the prion state is induced to appear [53], indicating that the threshold between curing and induction is wide enough to accommodate switches in one direction or the other specifically. A similar analysis in mammals would be prudent before pursuing increased dosage of dominant-negative mutants as a therapeutic strategy.
How can the absence of heritable aggregates in some daughter cells be reconciled with amyloid disassembly as a common mechanism of inhibition for G58D? Our previous studies have revealed that increasing chaperone levels by heat shock, leads to amyloid disassembly in a [PSI+]Weak strain [85], suggesting that the ratio of chaperones:amyloid is a key contributor to the balance between amyloid assembly and disassembly. A similar skew in this ratio likely occurs during G58D curing but through a distinct mechanism. Our previous studies uncovered a size threshold for amyloid transmission during yeast cell division: larger aggregates were preferentially retained in mother cells [72]. This asymmetry created an age-dependent difference in aggregate load, with newborn daughters taking several generations to return to the steady-state level of propagons observed in mother cells [72]. This observation suggests that the chaperone:substrate ratio could be skewed toward the former in daughter cells. This altered ratio, when combined with the decrease in the kinetic stability of Sup35 amyloid induced by G58D (Fig 3A–3C), likely creates a niche where amyloid disassembly dominates. Indeed, the normally resistant [PSI+]Strong variant is cured by transient heat shock when G58D is expressed [85]. Consistent with the idea that G58D cures [PSI+] by promoting amyloid disassembly, curing is reduced (Fig 5A), and propagon numbers increase in daughters (Fig 5B) when Hsp104 levels are reduced. Most importantly, transiently blocking Hsp104 activity in daughter cells after division also greatly reduces prion loss (Fig 5D). Thus, G58D–containing Sup35 amyloid is transmitted to daughter cells, but, once there, these aggregates are at greater risk of clearance by Hsp104-mediated disassembly.
Beyond dominant-negative mutants, conformational variants of PrP and Sup35 also differ in their sensitivities to small molecule inhibitors [62,63]. Unfortunately, even sensitive conformational variants can develop resistance to these compounds, further complicating attempts to develop therapeutic interventions for these diseases. For example, treatment of prion-infected mice or tissue culture cells with quinacrine or swainsonine reduced the kinetic stability of protease-resistant PrP and altered its tropism in cell lines, but these properties were reversed when treatment was removed [93–95]. Although it remains unclear whether the emerging conformational variants were minor components that were selected or newly induced by the treatment, this conformational plasticity creates a moving target that is impossible to manage if a unique inhibitor must be developed in each case. Our studies suggest that as prion conformational variants evolve, adapt or mutate, changes in dosing regimes could be effective countermeasures, although the range of possible doses is likely to be restricted because overexpression of even a dominant-negative mutant can lead to prion appearance [53]. Nevertheless, quinacrine can eliminate the RML conformational variant of PrP from CAD5 cells at a 5-fold lower dosage than is required to eliminate an IND24-resistant variant [96].
Much research is focused on the appearance and self-replicating amplification of amyloid, yet these processes are clearly counteracted by disassembly pathways in vivo. This balance between assembly and disassembly contributes strongly to prion persistence, even in mammals. For example, inhibition of PrP expression can reverse accumulation of protease-resistant PrP, pathological changes and clinical progression of prion disease in mice, presumably by allowing clearance pathways to dominate, if initiated before extensive damage arises [97]. While mammals lack an Hsp104 homolog, a chaperone system, composed of mammalian Hsp70, Hsp110, and class A and B J-proteins, possesses strong disaggregase activity [98], capable of directing amyloid disassembly, although this activity has yet to be tested against protease-resistant PrP [99]. Nevertheless, natural variations in the accumulation of prion and chaperone proteins may also serve as a new framework in which to consider phenotypic differences among variants. For example, tropism and clinical progression are likely to be impacted by the balance between assembly and disassembly pathways, as we have observed for mitotic stability and heat shock-induced prion curing in yeast [72,85]. Moreover, the steady-state ratio of chaperones:amyloid may be a key consideration in screening potential therapeutics and in their ultimate efficacy in vivo, particularly for small molecules proteostasis regulators that perturb the assembly/disassembly balance.
All plasmids used in this study are listed in S1 Table. pRS306-PADH contains PADH-Multiple Cloning Site-TCYC1 as a KpnI-SacI fragment from pSM556 (a gift from F.U. Hartl) in a similarly digested pRS306. The SUP35(G58D) ORF was then subcloned into pRS306-PADH as a BamHI-EcoRI fragment isolated from pRS306-SUP35(G58D) to create pRS306-PADHSUP35(G58D) (SB468).
Oligonucleotides used in this study are listed in S2 Table.
All strains are derivatives of 74-D694 and are listed in S3 Table. [PSI+]Sc4 (SY2085) and [PSI+]Sc37 (SY2086) haploid wildtype strains were gifts from J. Weissman. Yeast strains expressing ectopic copies of SUP35 or G58D from URA3 (pRS306) or TRP1 (pRS304)-marked plasmids were constructed by transforming yeast strains with plasmids that were linearized with BstBI or Bsu361, respectively, and by selecting for transformants on the appropriate minimal medium. In all cases, expression was confirmed by quantitative immunoblotting for Sup35. Disruptions of SUP35 (FP35, FP36) were generated by transformation of PCR-generated cassettes using pFA6aKanMX4 as a template with the indicated oligonucleotide primers (S2 Table) and selection on rich medium supplemented with G418. HSP104 disruptions were generated by transformation with a PvuI-BamHI fragment of pYABL5 (a gift from S. Lindquist) and selection on minimal medium lacking leucine. Disruptions of NAT1 (FP29, FP30) were generated by transformation of PCR-generated cassettes using pFA6a-hphMX4 as a template with the indicated primers (S2 Table) and selection on complete medium supplemented with hygromycin. All the disruptions were verified by PCR and 2:2 segregation of the appropriate marker.
Exponentially growing cultures of the indicated strain were plated on YPD for single colonies, and the frequency of [PSI+] loss was determined by the number of red colonies arising.
Semidenaturing detergent agarose gel electrophoresis (SDD-AGE), SDS-PAGE, quantitative immunoblotting and SDS-sensitivity experiments were performed as previously described [73]. To analyze the fate of aggregated Sup35, cultures were grown to midlog phase and treated with cycloheximide (CHX) or both CHX and guanidine HCl (GdnHCl) for 1.7 hours. Yeast lysates were collected before and after treatment and incubated at 53°C and 100°C in the presence of 2% SDS before analysis by SDS-PAGE. Lysates were also prepared from the same cultures and analyzed by SDD-AGE.
The number of propagons per cell was determined using a previously described in vivo dilution, colony-based method [75]. For propagon counting in mothers and daughters, a pair of mother and daughter cells was separated by micromanipulation onto minimal medium (SD-complete with 2.5mM adenine and 4% dextrose) with 3mM GdnHCl. After growing at 30°C for about 48 h, whole colonies were isolated using a cut pipette tip, resuspended in a small volume of water and plated onto YPD plates. The number of white colonies was then counted.
Daughters were separated by FACS based on bud-scar labeling. Yeast cells were incubated for 1 h at room temperature in 1μg/ml Alexa-647 wheat germ agglutinin (WGA) in PBS. After washing twice in PBS, cells with the lowest fluorescence intensity (5%) were sorted as newborn daughter cells, and a sample of this fraction was viewed by fluorescence microscopy to confirm bud scar number. This fraction was also moved to rich medium (1/4 YPD) for color development. For Hsp104 inhibition, sorted fractions were first moved to a minimal medium with 3mM GdnHCl for three hours before being transferred to rich medium. In each case only completely red colonies were counted as [psi-].
Fluorescence microscopy was performed on a DeltaVision deconvolution microscope equipped with a 100x objective. WGA Alexa-647 fluorescence was collected using 650nm excitation and 668nm emission wavelengths and with an exposure of 50ms. Images were processed in ImageJ software.
Cultures were grown in YPAD medium to an OD600 of 0.1 at 30°C. GdnHCl was added to 3mM, and the culture was returned to 30°C for 5 hours to decrease the propagon number. Cultures were then collected by centrifugation, washed and transferred to YPAD medium without GdnHCl for recovery. Samples were plated on YPD, and the number of propagons per cell was counted at the indicated timepoints.
|
10.1371/journal.pgen.1004554 | Response Regulator Heterodimer Formation Controls a Key Stage in Streptomyces Development | The orphan, atypical response regulators BldM and WhiI each play critical roles in Streptomyces differentiation. BldM is required for the formation of aerial hyphae, and WhiI is required for the differentiation of these reproductive structures into mature spores. To gain insight into BldM function, we defined the genome-wide BldM regulon using ChIP-Seq and transcriptional profiling. BldM target genes clustered into two groups based on their whi gene dependency. Expression of Group I genes depended on bldM but was independent of all the whi genes, and biochemical experiments showed that Group I promoters were controlled by a BldM homodimer. In contrast, Group II genes were expressed later than Group I genes and their expression depended not only on bldM but also on whiI and whiG (encoding the sigma factor that activates whiI). Additional ChIP-Seq analysis showed that BldM Group II genes were also direct targets of WhiI and that in vivo binding of WhiI to these promoters depended on BldM and vice versa. We go on to demonstrate that BldM and WhiI form a functional heterodimer that controls Group II promoters, serving to integrate signals from two distinct developmental pathways. The BldM-WhiI system thus exemplifies the potential of response regulator heterodimer formation as a mechanism to expand the signaling capabilities of bacterial cells.
| Two-component signal transduction systems are a primary means of regulating gene expression in bacteria. Recognizing the diversity of mechanisms associated with these systems is therefore critical to understanding the full signaling potential of bacterial cells. We have analyzed the behavior of two orphan, atypical response regulators that play key roles in controlling morphological differentiation in the filamentous bacteria Streptomyces-BldM and WhiI. We demonstrate that BldM activates its Group I target promoters as a homodimer, but that it subsequently activates its Group II target promoters by forming a functional heterodimer with WhiI. BldM-WhiI heterodimer formation thus represents an unusual mechanism for the coactivation of target genes and the integration of regulatory signals at promoters, enhancing the known repertoire of signaling capabilities associated with two-component systems.
| Two-component signal transduction systems are of central importance in regulating gene expression in bacteria. Canonically they consist of a response regulator, which functions as a homodimer, and a cognate sensor histidine kinase (which may also function as a cognate phosphatase). The activity of the kinase/phosphatase is modulated in response to a perceived stimulus. The sensor kinase autophosphorylates on a conserved histidine residue, and the phosphoryl group is then transferred to a conserved aspartate in the response regulator. The addition of the phosphoryl group stabilizes a conformation of the response regulator that drives an output response, most often the activation of gene expression. However, the intrinsic modularity of these systems has allowed bacteria to evolve variations on this basic theme, including more complex multicomponent phosphorelays, and changes in the nature of the response regulator effector domain such that the output can be, for example, an enzymatic activity rather than DNA binding [1]–[3]. Recognizing the diversity of mechanisms associated with these systems is therefore critical to understanding the full potential of the signaling capabilities of bacterial cells. This study concerns the behavior of two response regulators required for morphological development in the filamentous bacteria Streptomyces.
When streptomycete spores germinate, one or two germ tubes emerge and grow by tip extension and branching to form an extensive, multicellular vegetative mycelium [4]–[6]. Streptomycetes differentiate by forming specialized reproductive structures called aerial hyphae, which emerge from the colony surface into the air. The formation of aerial hyphae requires the activity of a class of developmental master regulators encoded by the bld (bald) genes [4]–[6]. Subsequently, in the most dramatic event of the lifecycle, each multigenomic aerial hypha arrests tip growth and undergoes a massive, synchronous septation event, giving rise to ∼50–100 unigenomic prespore compartments that ultimately develop into mature, pigmented exospores [4]–[6]. The differentiation of aerial hyphae into mature spores is coordinated by the activity of a second class of developmental master regulators encoded by the whi (white) genes. The focus of this work is the interaction between two of these global regulators, BldM and WhiI.
BldM and WhiI are both atypical response regulators (ARRs). In canonical response regulators, the aspartate residue that is subject to phosphorylation sits in a highly conserved pocket within the N-terminal receiver domain. ARRs usually lack essential residues within this phosphorylation pocket, suggesting that their activity is not controlled by phosphorylation [7]–[11]. WhiI has a degenerate phosphorylation pocket, lacking a universally conserved lysine and one of a pair of adjacent aspartate residues essential for binding Mg2+ [12], [13]. Although BldM does have a conserved phosphorylation pocket, a bldM allele carrying a D54A substitution at the putative site of phosporylation fully complements a bldM null mutant [14]. Moreover, BldM could not be phosphorylated in vitro [14]. Finally, BldM and WhiI are both ‘orphan’ response regulators – their genes are not adjacent to a sensor kinase gene, as is most often the case for canonical response regulators. Taken together, these observations strongly suggest that BldM and WhiI are not controlled by phosphorylation as part of conventional two-component systems. BldM and WhiI both belong to the NarL/FixJ subfamily of response regulators. bldM and whiI are found in all sequenced streptomycete genomes and their chromosomal context is conserved throughout. Strikingly, the amino acid sequence of BldM is 100% identical across all sequenced streptomycetes (the only response regulator that is 100% identical across all streptomycetes). WhiI is at least 93% identical, with amino acid variations found mainly in the linker region between the degenerate receiver domain and the DNA-binding domain. All sequenced WhiIs have a degenerate phosphorylation pocket.
During differentiation, the whiI and bldM genes are activated by two cognate, development-specific sigma factors, σWhiG and σBldN, respectively. whiI expression is activated from a single σWhiG target promoter, and thus whiI is not expressed in a whiG mutant [12]. σBldN directs transcription of the p1 promoter of bldM (the other promoter, bldMp2, is σBldN-independent, and so transcription of bldM is developmentally activated from only one of its two promoters in a bldN mutant [15], [16]). In addition to bldM, the other key targets of σBldN are the genes encoding the chaplins and rodlins, the major proteins of the hydrophobic sheath that coats the aerial hyphae and spores in Streptomyces [16]–[19].
Streptomyces venezuelae has recently emerged as an attractive new model system for the analysis of Streptomyces development because it sporulates in liquid culture [16], [20]. Here we take advantage of the S. venezuelae system to apply global microarray transcriptional profiling and ChIP-Seq to characterize the BldM and WhiI regulons. Through this route we go on to show that a key stage in Streptomyces development is controlled by response regulator heterodimer formation between BldM and WhiI, and to greatly expand our understanding of the regulatory network that controls morphological differentiation in these multicellular bacteria. The BldM-WhiI system thus exemplifies the potential of response regulator heterodimer formation as a mechanism to expand the signaling capabilities of bacterial cells.
Having established conditions in which S. venezuelae sporulates abundantly in liquid culture [16], [20], immunoblotting of samples taken at 14, 15 and 16 h of growth in MYM liquid sporulation medium showed that BldM was abundant at each of these time points (Figure S1). To gain greater insight into BldM function, we defined the genome-wide BldM regulon using ChIP-Seq. As described in Materials and Methods, wild-type S. venezuelae was subjected to formaldehyde cross-linking, lysis and sonication after 16 h of growth. After immunoprecipitation using a BldM-specific polyclonal antibody, the resulting DNA was subjected to deep sequencing. As a negative control, a ChIP-Seq experiment was performed on the congenic bldM null mutant. Several well-characterized developmental loci, including ssgR, rshA, smeA-sffA, whiB and whiE, were among the direct BldM targets identified.
Next, in order to determine how BldM influences the expression of its target genes, wild-type S. venezuelae and the congenic ΔbldM mutant were subjected to time-resolved, genome-wide transcriptional profiling during vegetative growth and sporulation. Strains were grown under the same conditions used for the ChIP-Seq experiments. RNA samples were prepared at 2-hour intervals from 8 to 20 hours, by which time sporulation was nearing completion, and following cDNA synthesis and labeling, samples were hybridized to Affymetrix DNA microarrays. Three independent biological replicates were performed for each strain, and analysis of the resulting data showed that the expression of 131 direct BldM targets was significantly down-regulated in the ΔbldM mutant (p<0.01) in comparison to the wild type. In contrast, only six genes were up-regulated in the ΔbldM mutant (p<0.01) in comparison to the wild type. These results suggest that BldM functions mainly as a transcriptional activator.
We next determined the time-resolved transcriptional profiles of the BldM target genes in seven constructed white mutants: ΔwhiA, ΔwhiB, ΔwhiD, ΔwhiG, ΔwhiH and ΔwhiI. Strikingly, many of the BldM target genes clustered into two well-defined groups according to their dependencies on the whi genes. Group I genes consisted of developmentally induced genes that depend on bldM, but were activated normally in all the whi mutants (Figure 1A and Table S1). Group II BldM target genes were also developmentally induced, but depended not only on bldM, but also on whiG and whiI (Figure 1B and Table S2).
To gain further insight into Group I binding sites, we fed the sequences of Group I promoter regions into the MEME algorithm [21] to search for over-represented sequences, using as input the entire intergenic region in each case. This analysis revealed a well-conserved copy of a 16 bp palindromic sequence, 5′-TCACcCgnncGgGTGA-3′, for which the sequence logo is shown in Figure 2A. The palindromic nature of this sequence would be consistent with BldM binding as a homodimer to Group I promoters.
To test the validity of the MEME output, and to confirm and extend the ChIP-Seq analysis, we overexpressed and purified BldM from E. coli as a soluble, N-terminally His6-tagged protein. The resulting BldM protein was used in DNase I footprinting analysis on the intergenic regions upstream of two Group I BldM targets, sven1998 and sven4150. In both cases, BldM protected a region containing a well-conserved copy of the palindromic, MEME-predicted binding site (Figure 2B), consistent with this sequence serving as a high-affinity binding site for a BldM homodimer.
Group II BldM target genes were expressed later than Group I genes (see insets in Figure 1). Further, and in contrast to Group I, the expression of Group II genes depended not only on bldM but also on whiG and whiI (Figure 1 and Table S2). It is straightforward to account for the dependence of Group II gene expression on whiG. In S. coelicolor, whiI expression is activated from a single σWhiG target promoter, and thus whiI is not expressed in a whiG mutant [12]. This σWhiG target promoter appears well conserved at the sequence level in S. venezuelae (Figure S2A) and whiI is not expressed in an S. venezuelae whiG mutant (Figure S2B). Thus all genes that depend on whiI must necessarily also depend on whiG. Expression of whiI was not significantly affected in the ΔbldM mutant and vice versa (p<0.01) (Figure S2B), implying independent σWhiG-WhiI and σBldN-BldM regulatory pathways. The challenge then was to determine why Group II BldM target genes depend on whiI.
To determine if the dependence of Group II BldM target genes on WhiI was direct or indirect, we characterized the in vivo WhiI binding sites across the S. venezuelae genome using ChIP-Seq. Wild-type S. venezuelae was harvested at 16 h of growth and treated as described for the BldM ChIP-Seq, except that a WhiI-specific polyclonal antibody was used. As a control, a ChIP-Seq experiment was performed using the congenic whiI null mutant. The data showed that all of the BldM Group II targets were also direct targets of WhiI (Figure 3). As an independent confirmation, we repeated the ChIP-Seq experiment using a FLAG-tagged WhiI protein. An N-terminally 3xFLAG-tagged allele of whiI (TF-WhiI) was constructed such that it was expressed from its native promoter and cloned into the single-copy vector pMS82, which integrates site-specifically into the chromosome at the phage ΦBT1 attB site [22]. This construct fully complemented the phenotype of the whiI null mutant (Figure S3), and the complemented strain was used for the ChIP-Seq experiment, now using wild-type S. venezuelae as the negative control. The ChIP-Seq results seen using FLAG immunoprecipitation were almost identical to those obtained using WhiI polyclonal antibodies, confirming that Group II genes are directly regulated by both BldM and WhiI (Table S2).
Our data showed that expression of Group II genes depends on bldM and whiI and that both BldM and WhiI bind directly to the promoters of these genes. One possible model consistent with these observations would be that BldM and WhiI co-activate Group II promoters by binding as two separate homodimers. An alternative model would be that these two proteins activate Group II promoters by binding as a functional BldM-WhiI heterodimer. To begin to differentiate between these models, we performed BldM ChIP-Seq in a ΔwhiI mutant and WhiI ChIP-Seq in a ΔbldM mutant, using the same conditions described above. In a ΔwhiI mutant BldM binding was still observed at all Group I promoters (which depend solely on bldM), but no BldM binding was seen at Group II promoters (Figure 3). Equally, no WhiI binding to Group II promoters was observed in a ΔbldM mutant (Figure 3). Thus, in vivo, BldM and WhiI show mutual dependence for binding to Group II promoters.
To explore the possibility of BldM-WhiI heterodimer formation, we tested BldM and WhiI for direct interaction in E. coli using a bacterial two-hybrid (BACTH) system [23]. bldM was fused to the gene encoding the T18 fragment of adenylate cyclase in the vector pUT18 such that BldM was at the N-terminus of the fusion protein, and was also fused to the gene encoding the T25 fragment of adenylate cyclase in the vector pKT25 such that BldM was at the C-terminus of the fusion protein. Parallel pUT18 and pKT25 constructs were made carrying fusions to WhiI.
Interacting pairs of proteins were screened initially by transforming E. coli BTH101 with the appropriate plasmids and monitoring restoration of adenylate cyclase activity on X-gal indicator plates; clones of each pair were then assayed for β-galactosidase activity (Figure 4). Interaction of BldM with itself was readily observed (∼750 Miller units), but a stronger interaction (∼3000 Miller units), was observed between BldM and WhiI, regardless of which protein was fused to the T18 fragment of adenylate cyclase and which was fused to the T25 fragment (Figure 4). WhiI showed no detectable interaction with itself (Figure 4).
To confirm and extend the BACTH analysis, we tested the interaction of BldM and WhiI in Streptomyces by coimmunoprecipitation. A C-terminally 3xFLAG-tagged allele of bldM (BldM-TF) was constructed such that it was expressed from its native promoter and cloned into the integrative vector pMS82 [22]. This construct complemented the phenotype of the ΔbldM null mutant to wild-type levels of sporulation (Figure S3). This strain was used in conjunction with the ΔwhiI mutant complemented with the N-terminally 3xFLAG-tagged allele of whiI (TF-WhiI) described above. The BldM-TF and TF-WhiI proteins were immunoprecipitated directly from 16 h MYM liquid cultures using the M2 anti-FLAG monoclonal antibody. WhiI coimmunoprecipitated with BldM-TF and BldM coimmunoprecipitated with TF-WhiI, but neither was detected in the negative controls when the wild-type strain was used (Figure S4). Thus BldM and WhiI interact in Streptomyces in vivo.
An early frustration in the in vitro analysis of WhiI function was that it overexpressed in an insoluble form in E. coli under all conditions tested. The realization that WhiI might act as part of a functional BldM-WhiI heterodimer led us to try an alternative approach. Where proteins form a complex, it is often observed that individual components are insoluble when expressed in isolation but become soluble when expressed with their cognate partner protein. Two examples are the α and β subunits of lambda integrase [24], and Streptomyces σBldN and its cognate anti-sigma factor, RsbN [16]. Accordingly, BldM and WhiI were co-expressed in E. coli using the pETDuet-1 system (Novagen). Initially, BldM was N-terminally His6-tagged and WhiI was left untagged. Co-expression of BldM was found to solubilize WhiI completely. Further, when His6-BldM was purified on a HisTrap Ni column, WhiI copurified with His6-BldM in approximately equal amounts, despite the fact that WhiI was untagged (Figure 5A) (the identity of the two proteins was confirmed by tryptic mass fingerprinting). Thus BldM rescues WhiI from inclusion bodies and the two proteins copurify in approximately stoichiometric amounts via a His6-tag present on BldM only.
This approach was extended by coexpressing BldM and WhiI carrying two compatible affinity tags. N-terminally His6-tagged BldM and N-terminally StrepII-tagged WhiI were co-expressed from the pETDuet-1 vector. As before, both BldM and WhiI were found in the soluble fraction. The BldM-WhiI complex was then purified over consecutive HisTrap Ni and StrepTrap HP affinity columns and the BldM and WhiI proteins were found to be present in stoichiometric amounts in the resulting preparation (Figure 5B).
To further understand Group II binding sites, we searched for over-represented sequences in Group II promoter regions using MEME [21], again using as input the entire intergenic region in each case. This analysis revealed a well-conserved 16 bp non-palindromic sequence, 5′-TGnnCCGnnCGGGTGA-3′, for which the sequence logo is shown in Figure 6A. Strikingly, the 3′ half of the Group II logo was equivalent to a half-site of the Group I palindrome, but the other half was different in sequence, potentially consistent with a BldM-WhiI heterodimer binding to Group II targets.
To directly test the model that Group II promoters are controlled by a functional BldM-WhiI heterodimer, and to validate the MEME-predicted binding motif for these promoters, the doubly-tagged BldM-WhiI that had been purified over consecutive HisTrap Ni and StrepTrap HP affinity columns was used in DNase I footprinting analysis on the intergenic regions upstream of two Group II targets, sven1263 and murA2 (sven5810). BldM-WhiI footprinted on both promoters and in each case the protection region contained a copy of the non-palindromic, MEME-predicted binding site (Figure 6B), consistent with this sequence serving as a high-affinity binding site for a BldM-WhiI heterodimer. In contrast, neither BldM alone nor WhiI (produced as a soluble GST fusion), footprinted on either promoter (Figure 6B).
The focus of this study is to elucidate the mechanism underlying the direct co-activation by BldM and WhiI of the Group II genes, required for the late stages of development. Our data show that BldM activates transcription of these Group II genes as a BldM-WhiI heterodimer, while activating transcription of the Group I genes required for the early stages of development as a BldM homodimer. This work also significantly expands our knowledge of the regulatory network that controls morphological differentiation in Streptomyces (Figure 7), an advance made possible by exploiting S. venezuelae as a new model species for the genus.
BldM homodimer activates several genes known to play key roles in the differentiation of aerial hyphae into spores, including whiB and ssgR. In addition to their positive regulation by BldM homodimer, these genes are also subject to repression during vegetative growth by the master regulator BldD, as is bldM itself (Figure 7) [25].
SsgA and SsgB are homologous proteins directly involved in the positive control of cell division in Streptomyces [26]. Sporogenic aerial hyphae undergo a synchronous round of cell division, initiated by the polymerization of a ladder of 50 or more FtsZ rings. SsgA and SsgB function in the recruitment and accurate positioning of these FtsZ rings, and ΔssgA and ΔssgB mutants of S. coelicolor lack sporulation septa [26]–[28]. ssgR encodes an IclR-family transcriptional regulator that directly activates the expression of ssgA in S. coelicolor [29]. Here we show that ssgR is directly activated by, and is completely dependent on BldM in S. venezuelae. ssgB is developmentally induced in S. venezuelae, but despite being a direct BldM target, its expression is only weakly affected in the ΔbldM mutant, suggesting complex regulation of this gene.
BldM homodimer also activates expression of whiB, which plays a vital role in developmentally controlled cell division. whiB null mutants fail to arrest aerial tip growth, the normal prelude to sporulation, and are completely blocked in the initiation of sporulation sepatation, producing abnormally long, undivided aerial hyphae [30]. WhiB is the founding member of a family of proteins confined to the actinomycetes, and several of these WhiB-like (Wbl) proteins have been shown to play key roles in the biology of streptomycetes and mycobacteria. Wbl proteins carry a [4Fe–4S] iron-sulfur cluster coordinated by four invariant cysteines in a C(X29)C(X2)C(X5)C motif [31]–[35], and although the biochemical role of these unusual proteins has been controversial [36], it seems increasingly certain that they function as transcription factors [33], [37]–[38].
Among the Group II targets controlled by the BldM-WhiI heterodimer are two loci with well characterized roles in sporulation: smeA-sffA and whiE. The smeA-sffA operon encodes a DNA translocase (SffA) involved in chromosome segregation into spores that is specifically targeted to sporulation septa by the small membrane protein SmeA [39]. Deletion of smeA-sffA in S. coelicolor results in a defect in spore chromosome segregation and has pleiotropic effects on spore maturation [39]. Like the Group I targets ssgR and whiB, expression of the smeA-sffA operon is also repressed by the master regulator BldD during vegetative growth [25]. whiE is a complex locus that specifies the spore pigment. The structure of the spore pigment has not been determined in any Streptomyces species but its polyketide nature was first predicted from the sequence of the whiE locus in S. coelicolor, because it encodes proteins that closely resemble the components of type II polyketide synthases involved in the synthesis of aromatic antibiotics [40]–[42]. Based on their coordinate regulation and proposed functions, we predict the whiE locus of S. venezuelae consists of an operon of seven genes (sven6798-6792) and the divergently transcribed gene sven6799. Two distinct ChIP-Seq peaks were seen in the intergenic region separating sven6799 and the sven6798-6792 operon, and all eight genes fail to be expressed in the bldM and whiI mutants, implying the BldM-WhiI heterodimer controls expression of the entire locus.
The work presented here suggests that there is no set of genes regulated by a WhiI homodimer and that WhiI functions as an auxiliary protein to modulate BldM binding specificity through heterodimerization. With no exceptions, all the genes down regulated in a ΔwhiI mutant were also down regulated in a ΔbldM mutant. Although some promoters were exclusively enriched as peaks in the WhiI ChIP-Seq experiment, without exception the transcriptional profile of such targets was unaffected in a ΔwhiI mutant, showing that WhiI has no regulatory influence on these genes. Further, in a bldM null mutant, some WhiI peaks are seen in ChIP-Seq, but these sites, often internal to ORFs, show no correlation with the wild-type WhiI regulon and the targets lacked a consensus binding motif. These results suggest that, in the absence of BldM, any DNA binding by WhiI is aberrant and unrelated to its behavior in the wild type. In contrast, in a ΔwhiI mutant, BldM fails to bind to Group II promoters but binds normally to its Group I promoters.
In a recent study, evidence was presented that the DNA-binding domain of S. coelicolor WhiI (in the absence of the receiver domain) can bind in vitro to the promoter of the sco3900-sco3899 operon encoding a transcriptional regulator (InoR) and an inositol 1-phosphate synthase (InoA), respectively [43]. Our ChIP-Seq data show that WhiI does not regulate these genes in S. venezuelae. Further, in wild-type S. venezuelae both genes are actively expressed during vegetative growth but are down-regulated during development, and this expression pattern is unaffected in a whiI mutant.
Transcription factor heterodimerization can coordinate responses to different cues by integrating signals from distinct regulatory pathways. Although heterodimerization is prevalent as a regulatory mechanism in eukaryotes [44], it is rare in bacteria. Prior to the work described here, the only response regulator reported to heterodimerize with an auxiliary regulator was RcsB.
In Escherichia coli, the typical response regulator RcsB plays a central role in the regulation of capsule synthesis. Once phosphorylated by the histidine kinase RcsD, RcsB directly activates target genes including rprA, osmC, osmB and ftsZ, functioning as a homodimer [45], [46]. It also activates exopolysaccharide synthesis genes, required for capsule formation, as a heterodimer with RcsA, which is distantly related to response regulators (like RcsB, RcsA has a typical LuxR-type C-terminal DNA-binding domain, but its N-terminal domain is not related to typical response-regulator receiver domains). Like WhiI, RcsA appears to lack the capacity to activate genes by itself, and therefore functions solely as a modulator of RcsB binding specificity. RcsA is actively degraded by the Lon protease and in lon mutants capsule genes are highly upregulated causing a mucoid phenotype, due to enhanced activation by the stabilized RcsB-RcsA heterodimer [45], [46]. The capacity of RcsB to complex with other transcription factors is not limited to its interaction with RcsA, since RcsB also forms functional heterodimers with BglJ [47] and with GadE [48]. Using in vitro FRET analysis, weak hetero-pair interactions were detected between several members of the OmpR sub-family of response regulators from E. coli [49]. While these in vitro data may suggest potential crosstalk between distinct signaling pathways, their physiological significance has yet to be demonstrated.
The activities of numerous bacterial promoters respond to multiple cues, and there are many examples of promoters that depend on two activators for their activity. Several different regulatory mechanisms underpinning such codependence have been identified [50], [51]. The most widely documented is found at promoters where both activators bind independently, and both activators make independent contacts with RNA polymerase. However, there are rare examples of coactivators that exhibit cooperative binding, such as MelR and CRP at the E. coli melAB promoter [52]. There are also examples in which DNA binding by a secondary activator leads to the repositioning of the primary activator from a site where it cannot activate transcription to a site where it can, such as the repositioning of MalT by CRP at the malK promoter [53]. Response regulator heterodimer formation provides a new model for coactivation of target genes and the integration of regulatory signals at promoters. BldM-WhiI heterodimer formation serves to integrate signals from two independent pathways (σWhiG-WhiI and σBldN-BldM) and it may also function as a timing device, since Group II genes are activated later than Group I genes. Thus the BldM-WhiI system exemplifies the potential of response regulator heterodimer formation as a mechanism to expand the signaling capabilities of bacterial cells.
Bacterial strains and plasmids are listed in Table S3 and the oligonucleotide primers with corresponding restriction sites used in cloning are listed in Table S4. For microarray and ChIP-Seq experiments, S. venezuelae strains were grown at 30°C in MYM liquid sporulation medium [16] made with 50% tap water and supplemented with 200 µl trace element solution [54] per 100 ml. The phenotypes of mutants and complemented strains were scored after 3–4 days growth on MYM-agar at 30°C. bldM and whiI deletion mutants were constructed by ‘Redirect’ PCR targeting [55] and their chromosomal structures were confirmed by PCR analysis and by Southern hybridization using the parental cosmids as probes.
For each strain, two flasks containing 35 ml of MYM were inoculated with spores (or mycelium in case of the ΔbldM mutant) to give an OD600 ∼0.35 after 8 h of growth. The crosslinking reagent formaldehyde was added to a final concentration of 1% (v/v) to the cultures at 16 h of growth and incubated at 30°C with shaking for 30 min before glycine was added to a final concentration of 125 mM to quench the crosslinking reaction. The samples were incubated at room temperature for 5 min and washed twice in PBS buffer pH 7.4 (Sigma). Mycelial pellets were resuspended in 0.5 ml lysis buffer (10 mM Tris-HCl pH 8, 50 mM NaCl, 15 mg/ml lysozyme, 1x protease inhibitor) and incubated at 37°C for 20 min. The lysate was resuspended in 0.5 ml IP buffer (100 mM Tris- HCl pH 8, 250 mM NaCl, 0.1% Triton-X-100, 1x protease inhibitor) and the lysate was kept on ice for 2 min before sonication. The samples were subjected to seven cycles of sonication, 15 s each, at 10 microns, to shear the chromosome into fragments ranging in size from 300–1000 bp. The sample was then centrifuged twice at top speed, 4°C for 15 minutes to clear cell extracts. To pre-clear non-specific binding, 90 µl protein A sepharose (Sigma) was added to cell lysate (about 900 µl) and incubated for 1 h at 4°C with mixing. The beads were cleared by centrifugation at top speed for 15 min. 100 µl BldM or WhiI antibodies were added to the corresponding cell lysates overnight at 4°C with mixing. 100 µl Protein A Sepharose 1∶1 suspension was added to immunoprecipitate antibody-BldM or WhiI chromatin complexes and incubated for 4 h at 4°C with mixing. The samples were centrifuged at 3500 rpm for 30 s and the beads were washed four times with IP buffer. The pellets were eluted in 150 µl IP elution buffer (50 mM Tris-HCl pH 7.6, 10 mM EDTA, 1% SDS) overnight at 65°C to reverse crosslink. The samples were centrifuged at top speed for 5 min to remove the beads and the pellets were re-extracted with 50 µl TE buffer (10 mM Tris-HCl pH 7.4, 1 mM EDTA). The supernatants were combined and incubated with 3 µl 10 mg/ml proteinase K (Roche) for 2 h at 55°C. The samples were extracted twice with phenol-chloroform to remove protein followed by chloroform extraction to remove traces of phenol and purified with Qiaquick columns (Qiagen). The IP DNA was eluted in 50 µl EB buffer (Qiagen). Sequencing libraries were generated and the IP DNA was sequenced as described previously [20]. The BayesPeak package was used to identify significantly enriched regions and the default parameters were applied [56].
Microarray transcriptional profiling experiments were carried out as described previously [16], [20]. Multi-experiment viewer software (MeV 4.8) was used for viewing and statistical analysis [57]. The non-parametric tool ‘Rank Products’ [58] was used in MeV to assign ‘down regulated’, ‘up regulated’ and ‘not significant’ genes based on expression at 16, 18 and 20 h of growth. Group I genes were defined as direct BldM ChIP-Seq targets that were significantly down regulated in ΔbldM and not significantly changed in all Δwhi mutants (Table S1). Group II genes were defined as direct ChIP-Seq targets of both BldM and WhiI that were significantly down regulated in ΔbldM, ΔwhiG and ΔwhiI (Table S2).
The open reading frame of interest was PCR-amplified using Expand High-Fidelity DNA polymerase (Roche). Plasmids containing the correct inserts were confirmed by sequencing and introduced into electrocompetent E. coli BL21(DE3)/pLysS. The transformed cells were spread on LB-carbenicillin/chloramphenicol and one colony was used for inoculation. Proteins were expressed in two 2.5 litre volumetric flasks each containing 400 ml LB culture and expression was induced with 0.25 mM IPTG. The optimised temperature for expression varied with the protein: His6-BldM was expressed at 25°C for 5 h; GST-WhiI was expressed at 15°C overnight; His6-BldM/WhiI or SII-WhiI were co-expressed at 30°C for 5 h. The pellets were lysed in a buffer containing 50 mM Tris-HCl pH 8, 250 mM NaCl, 10% glycerol, 0.1% Triton X100, protease inhibitor (complete mini, EDTA-free, Roche) and incubated at room temperature for 20 min. HisTrap HP Ni and StrepTrap HP affinity columns (GE Healthcare) were used to purify the His6- and SII-tagged proteins in a tandem manner. The Gst-WhiI was purified with 1 ml GSTrap FF column (GE Healthcare).
Singly 32P end-labelled probes (Table S4) were generated by PCR and purified using Qiaquick columns (Qiagen). Transcription factors were incubated with probe DNA (∼150,000 cpm) for 30 min at room temperature in 40 µl reaction buffer [50 mM Tris-HCl pH 7.5, 100 mM NaCl, 10% glycerol, 10 mM MgCl2, 2 mM dithiothreitol and 1 µg/reaction poly(dI-dC)], prior to treatment with 1 U DNase I (Promega) for 30–50 s in the case of group-II promoters and 3 U DNase I for 15–20 s in the case of group-I promoters. Reactions were terminated with 140 µl of stop buffer (192 mM sodium acetate, 32 mM EDTA, 0.14% SDS, 70 µg/ml yeast tRNA) and samples were extracted with phenol-chloroform prior to ethanol precipitation. Footprinting samples were loaded on 6% polyacrylamide sequencing gels, next to a G+A ladders prepared according to the Sure Track footprinting kit (Amersham Pharmacia Biotech).
C-terminally 3×FLAG-tagged bldM expressed from the native promoter was used to complement the ΔbldM mutant and N-terminal 3×FLAG-tagged whiI carrying native promoter was used to complement the ΔwhiI mutant. 50 µl dense spore suspension was used to inoculate 300 ml MYM in 2 litre flasks with spring baffles. After 17 h growth, cultures were harvested by centrifugation at 6000 rpm for 15 min at 4°C. Pellets were lysed in a buffer containing 50 mM Tris-HCl pH 8, 250 mM NaCl, 10% glycerol, 0.1% Triton X100, protease inhibitor (complete mini, EDTA-free, Roche) and sonicated for 5 cycles at 15-micron amplitude for 20 s. FLAG-tagged proteins were immunoprecipitiated using M2 beads [anti-FLAG antibodies covalently attached to agarose beads (Sigma)], the beads were washed using TBS buffer containing protease inhibitor (Roche) and 0.05% Triton X100, and the protein was eluted using FLAG peptides as recommended by the manufacturer.
S. venezuelae strains were grown in MYM medium. 10 ml samples were taken at 14, 15, and 16 hours of growth. Strains were harvested by centrifugation at 3000 rpm for 1 min, washed in 5 ml ice-cold washing buffer (20 mM Tris pH 8.0, 5 mM EDTA) and resuspended in 0.4 ml of ice-cold sonication buffer [20 mM Tris pH 8.0, 5 mM EDTA, 1x protease inhibitor (Roche)]. Samples were sonicated immediately for 5 cycles, 20 s at 10 microns with 1 min intervals of ice incubation, then centrifuged at 13000 rpm at 4°C for 15 min to remove cell debris. Protein concentrations of the supernatant crude cell extracts were measured by Bradford assay and samples (10 µg protein) were separated on a 12.5% SDS-PA gel and blotted onto nitrocellulose membrane. The membrane was incubated in blocking solution [10% dried milk powder in TBS (0.05 M Tris, 0.9% NaCl, pH 7.6, 0.1% Tween)] overnight and then incubated for 1 h at room temperature with the 1/2500 dilution of anti-BldM antiserum in blocking solution. The membrane was rinsed (twice for 10 min) in TBS and then incubated for 1 h with 1/5,000 dilutions of horseradish peroxidase-linked goat anti-rabbit immunoglobulin G antibody (GE Healthcare). Blots were developed using the ECL enhanced chemiluminescence system from GE Healthcare and were typically exposed to X- ray film for between 30 s and 5 min.
|
10.1371/journal.ppat.1000309 | Export of a Toxoplasma gondii Rhoptry Neck Protein Complex at the Host Cell Membrane to Form the Moving Junction during Invasion | One of the most conserved features of the invasion process in Apicomplexa parasites is the formation of a moving junction (MJ) between the apex of the parasite and the host cell membrane that moves along the parasite and serves as support to propel it inside the host cell. The MJ was, up to a recent period, completely unknown at the molecular level. Recently, proteins originated from two distinct post-Golgi specialised secretory organelles, the micronemes (for AMA1) and the neck of the rhoptries (for RON2/RON4/RON5 proteins), have been shown to form a complex. AMA1 and RON4 in particular, have been localised to the MJ during invasion. Using biochemical approaches, we have identified RON8 as an additional member of the complex. We also demonstrated that all RON proteins are present at the MJ during invasion. Using metabolic labelling and immunoprecipitation, we showed that RON2 and AMA1 were able to interact in the absence of the other members. We also discovered that all MJ proteins are subjected to proteolytic maturation during trafficking to their respective organelles and that they could associate as non-mature forms in vitro. Finally, whereas AMA1 has previously been shown to be inserted into the parasite membrane upon secretion, we demonstrated, using differential permeabilization and loading of RON-specific antibodies into the host cell, that the RON complex is targeted to the host cell membrane, where RON4/5/8 remain associated with the cytoplasmic face. Globally, these results point toward a model of MJ organization where the parasite would be secreting and inserting interacting components on either side of the MJ, both at the host and at its own plasma membranes.
| A unique feature of apicomplexan parasites is the formation of an intimate contact between the apex of the parasite and the host cell membrane called the moving junction that moves along the parasite during invasion. Proteins originated from two distinct secretory organelles, the microneme for AMA1 and the rhoptry neck for RON2/4/5 proteins, are associated to form the junction. Here, we have furthered the characterization of the MJ complex by describing RON8, an additional protein component. AMA1 has previously been shown to be inserted into the parasite membrane upon secretion. Our study demonstrates that all the RON proteins are translocated into the host cell, where RON4/5/8 remain associated with the cytoplasmic face of the host cell plasma membrane. Furthermore, we identified a privileged interaction between transmembrane MJ proteins AMA1 and RON2 in vitro. Overall, this led us to propose the first model describing the putative MJ organisation at the interface between the host cell and Toxoplasma. In this original concept, the parasite would export its own receptor (RON2) and ligand (AMA1) on either side of the MJ.
| Invasion by Apicomplexa is an essential step of the pathologies associated with these protozoan parasites that include Plasmodium spp., the causative agents of malaria, as well as Toxoplasma gondii, responsible for human and animal toxoplasmosis. The invasive stages of these parasites share a highly conserved architecture, including a cytoskeleton-associated original pellicular complex, and two types of vesicular apical organelles (micronemes and rhoptries) that participate to the invasion process through the exocytosis of their contents in a sequential manner [1]. Host cell invasion has been well described at the ultrastructural level, but the precise molecular interactions and the specific role of the exocytosed parasite proteins are still poorly understood. Proteins located on the surface of the parasite probably mediate the initial interaction with the target cell. This is followed by an intimate contact between the apical tip of the parasite and the host cell membrane, called the moving junction (MJ) [2]. This singular structure, likely linked to the subpellicular cytoskeleton motor of the parasite, might serve as a support to propel the parasite into the parasitophorous vacuole (PV) that forms inside the host cell. To do so, the MJ rapidly turns into a ring that is moved backward along the parasite during invasion and ends up at the posterior end of the invaded parasite at the end of the process. Despite a number of investigations having led to the discovery of a variety of putative parasite adhesive molecules secreted from micronemes, and of an original acto-myosin based motor for gliding motility [3], the process of invasion itself (i.e. MJ-dependent host cell entry), remains a major conundrum. Indeed, although the morphological features of the process have been described 30 years ago [2], the MJ was, up to a recent period, completely unknown at the molecular level. The major reason for this was its transient nature, since host cell invasion is a very rapid process (a few seconds), and therefore isolating the structure was not possible.
Rhoptries are elongated organelles composed of a bulbous body that tapers into a thin duct-like neck. Rhoptries empty their contents apically during the invasion process, after microneme exocytosis, and their contribution to invasion was considered mostly as providing building material for the developing PV, since proteins of the bulb of the rhoptry (ROPs) were found associated with the nascent vacuole membrane (for a review see [4]). Recently, an unexpected function of the rhoptries in MJ formation arose from the discovery that one rhoptry neck protein (RON4) was associated to the MJ [5],[6]. It was proposed that MJ formation would derive from a cooperation between i) newly discovered RONs located in the rhoptry neck and ii) the micronemal protein AMA1 [5]. Numerous lines of evidence suggest that the conserved AMA1 protein plays a central role during invasion of Apicomplexa. For instance, AMA1 has been shown to be essential for Plasmodium merozoites and Toxoplasma tachyzoites [7],[8]. In T. gondii, AMA1 and RON4 have been found to be associated in a complex in vitro and they localize precisely at the MJ during cell invasion, although a direct association of the two proteins has not been demonstrated in vivo [5],[6].
The isolation of RON4 from parasite extracts by affinity purification led us to the simultaneous purification of the rhoptry neck protein RON2 and protein TwinScan_4705 (annotated also 583.m00636) [6], which was later shown to be also a RON (RON5, P. Bradley personal communication). Like AMA1, RONs are conserved throughout the Apicomplexa including Plasmodium spp., and they are not found outside this phylum. AMA1 and RONs are stored in two distinct compartments that release their content sequentially during invasion. Cross-linking experiments on invading parasites showed that the interaction of AMA1 with RONs takes place during invasion and is not the result of non-specific or indirect binding occurring in the parasite lysate during IP [5]. One intriguing question is how the micronemal protein AMA1 and the complex of rhoptry neck proteins RON2/RON4/RON5 avoid interacting in the secretory pathway. Another important question is how these proteins are organized at the MJ. The microneme protein AMA1 has been characterized structurally and appears to be translocated as a type-1 transmembrane (TM) protein in the tachyzoite plasma membrane [9],[10]. On the contrary, the topology of the RONs at the MJ is still obscure and several important questions remain unanswered. Are RONs directly or indirectly linked to the parasite surface? Could they be binding to a host cell receptor or, as we speculated previously [6], are they directly inserted into the host cell membrane to serve as a receptor for AMA1?
Here, we describe an additional partner of the previously characterized AMA1/RON2/4/5 complex named RON8. We also show that the complex may be assembled as pro-proteins but that a distinct timing of biosynthesis between MICs and RONs precludes the association of RONs with AMA1 before secretion. Furthermore, we demonstrate that RONs are exported to the host cell membrane, RON/4/5/8 being exposed to the host cell cytosol and RON2 being probably an integral membrane protein that displays a privileged interaction with AMA1. These results provide an important clue to understand how such a crucial structure for the invasive and developmental processes of the parasite is built and organized.
In order to further refine the molecular characterization the MJ complex of T. gondii, we searched for additional proteins co-immuno-purified (IP) by the anti-RON4 antibody matrix, as previously described [6]. The RON4-associated proteins were subjected to mass spectrometric analysis. As in our first analysis, we detected two principal bands at ∼120 kDa and ∼100 kDa, which corresponded to RON2, RON4 and RON5 (Figure 1A). In addition to the proteins of the MJ already known to be associated with each other (RON2, RON4, RON5 and AMA1), mass spectrometry analysis identified peptides from proteins that are described in Table S1. Peptides from proteins originated from the secretory organelles involved in invasion (microneme and rhoptry) have retained our attention. First, peptides from two microneme proteins MIC1 [11] and MIC3 [12] were detected. However, Western blot and reverse IP analysis using anti-MIC3 or anti-MIC1 antibodies did not confirm a specific interaction of these proteins with the MJ complex proteins (data not shown). Second, we found peptides from TwinScan_0092 (80.m02161) and TwinScan_2001 (541.m00141), two predicted Toxoplasma proteins that had also been detected in the proteomic analysis of the rhoptries [13]. TwinScan_0092 predicts a protein of 49 kDa that is not localising at the MJ but was instead found to be a new dense granule protein [14]. Concerning TwinScan_2001, a previous study using an antibody raised against a specific peptide had localised it to the apicoplast by IFA [13], although it does not possess any bona fide plastid-targeting element in its amino acid sequence. We then decided to reassess its subcellular localization by generating a specific polyclonal antiserum directed against a recombinant TwinScan_2001 protein corresponding to the central part of the protein (Figure 1D). This antibody (anti-Tw2001) reacted on Western blot with a major band of about to 250 kDa and several minor bands of lower molecular mass (Figure 1C), that were also detected with an additional serum raised against another region of the protein (not described here), but were absent when probed with the pre-immune serum (Figure 1C, Figure S1A). By IFA, the anti-Tw2001 serum recognized an antigen co-localized with RON4 in intracellular parasites (Figure 1B), suggesting that TwinScan_2001 was a new rhoptry neck protein that we named RON8.
To further verify that RON8 is associated with the AMA1/RON2/4/5 complex, we performed an IP using the anti-Tw2001 serum (referred to as anti-RON8 throughout the manuscript), as described previously for RON4 [6] and showed co-purification of RON8, AMA1 and RON4 (Figure 1C). The formation of a stable complex in 1% NP40 and 1 M NaCl conditions, containing RON2/4/5/8 and AMA1, was further confirmed by co-IP of all members after affinity chromatography using either of the specific anti-RONs (data not shown).
The complete coding sequence of RON8 was determined (GenBank accession number ACK57540) and showed that it coded potentially for a 2979 amino acids-long protein, with a theoretical molecular mass of 329 kDa. A putative signal peptide was found at position 1–29. PROSITE search yielded no obvious sequence motifs. A search of the GenBank non-redundant database and ApiDB showed that RON8 is unique to Toxoplasma and Neospora among Apicomplexa (in contrast to other MJ proteins) and is not found in other organisms.
We have previously shown that RON4 is associated with the MJ during invasion [6], here we examined if RON2, RON5 and RON8 would also follow the MJ. We first generated antisera specific of RON2, RON4 and RON5. For RON2, two sera were prepared against different regions of the protein produced as recombinant proteins named RON2n and RON2c (see Figure 1D). An anti-serum against the N terminal part of RON4 (RON4n) was also produced.
The specificity of the sera was first analyzed by Western blot on whole tachyzoite lysates (Figure S1A). All sera recognized in non-reduced condition a band at about the predicted size (RON2: 155 kDa,, RON5: 179 kDa, RON8: 329 kDa). No detection was observed with pre-immune sera. In reduced condition, the anti-RON2c, anti-RON2n, anti-RON4n, and anti-RON5 recognized proteins that migrated faster, indicating that, as previously shown for RON4 ([6] and Figure S1A), RON2 and RON5 are also sensitive to reduction of disulfide bonds (discussed later). The sera were also analyzed by IFA on intracellular parasites. As shown in Figure 2A, all the anti-RONs labelled the neck of the tachyzoites rhoptries, as indicated by co-localisation with RON4 and RON8. Throughout the study, the anti-RON4 T5 4H1 [15] and the anti-AMA1 CL22 mAbs were used systematically, except when specified.
On invading parasites, in permeabilization conditions optimized to detect only the material secreted by the parasite [1], we showed that anti-RON5 and anti-RON8 recognized exclusively the characteristic ring-shaped MJ (Figure 2B). In contrast, in these conditions both anti-RON2 antibodies failed to react (Figure 2B, lower panel). We have shown previously that when the PVM has pinched off the host cell, the MJ can still be detected at the posterior pole of the parasite for a few hours and is characterized by a dot-like signal with anti-RON4 mAb [6]. Again, we showed using specific antibodies that RON5 and RON8 could be found together at the same location, but not RON2 (data not shown). Since cytochalasin D (Cyt-D, an inhibitor of actin polymerization)-treated parasites form a “static” junction that is labelled by anti-RON4 [6] but not translocated to the posterior end of the tachyzoites [16],[17], we tested if all the RONs could be immunolocalized at the junction in these conditions. We found that after Cyt-D treatment, in addition to RON4 [6], all the proteins of the complex, this time including RON2 (yet only with the anti-RON2c antibody), could be detected at the same location (Figure 2C). The detection of RON2 in these conditions could be explained by the fact that the Cyt-D treatment had improved the accessibility of the protein to the antibody, either because it destabilised some link of RON2 with Cyt-D sensitive structures of the host or, more simply, that it blocked the junction in an early stage where the protein is more accessible. Overall, this is strengthening the idea that RON2, RON4, RON5 and RON8 are present together in the MJ complex during invasion.
The generation of antisera against the individual members of the MJ complex allowed us to analyse more precisely the RONs and their interactions by IP using different conditions for solubilization of the parasite. After lysing the parasites in 1% NP40 (the condition used to immunopurify the complex [6]), all members of the complex were recovered using each of the anti-sera available, as exemplified in Figure 3A with an IP using the anti-RON2n serum followed by Western blot analysis of each member of the MJ complex. We then tested the stability of the complex upon tachyzoite lysis in 0.6% SDS followed by heat denaturation. In these conditions, only the interaction between AMA1 and RON2 was maintained after IP with either anti-RON2n or anti-AMA1 (Figure 3A) and no interaction between the others RONs was observed (ie using anti-RON4n, anti-RON5 and anti-RON8, Figure S1B). These results were confirmed by comparing the profiles obtained after metabolic labelling of intracellular parasites with [35S]-methionine/[35S]-cysteine, and lysis in either 0.6% SDS or 1% NP40, followed by IP with anti-RONs antibodies (Figure 3B). A similar profile in which all members of the complex were detected was obtained after IP in 1% NP40 whatever the antibody used (left panel). In contrast in 0.6% SDS, while the anti-RON4, anti-RON5 and anti-RON8 immunopurified only the corresponding protein, the anti-RON2n and anti-AMA1 immunopurified both AMA1 and RON2 (right panel). These results clearly indicated that the whole complex was not maintained in 0.6% SDS, but that AMA1 and RON2 proteins interact together particularly strongly, independently of the other MJ proteins.
Most T. gondii rhoptry bulb proteins described so far are synthesized as pro-proteins that are subjected to removal of their N-terminal pro-region by proteolytic cleavage during traffic to the organelle. To determine whether the RONs are also processed, we studied their biosynthesis and maturation by pulse-chase metabolic labeling with [35S]-methionine/[35S]-cysteine followed by IP with anti-RONs antibodies (Figure 4A). The infected cells were lysed and boiled in 0.6% SDS to avoid co-purification of the whole complex.
For RON2, after 20 minutes of pulse, a protein of ∼150 kDa (reduced) was immunoprecipitated, which is consistent with the predicted size of RON2 after the removal of the signal peptide; a minor band was also found at ∼120 kDa. After one hour of chase, the 150 kDa disappeared and the 120 kDa band was the major one detected. A 65 kDa band after chase and a slightly slower migrating one in the pulse corresponded to AMA1 (as described above) and proAMA1 respectively (see below). In non-reduced condition, the 150 kDa band was detected both in pulse and chase fractions, indicating that RON2 is processed and that the two fragments might be linked by internal disulfide bonds (several cysteines are present in both fragments). For RON5, after 20 minutes of pulse, a major protein of ∼180 kDa (unreduced) was immunoprecipitated, which is consistent with the predicted size of the protein. After one hour of chase, the 180 kDa product almost disappeared and a ∼150 kDa band was detected instead. In reduced condition, the 180 kDa form was also detected in pulse, while a ∼110 kDa form was immunoprecipitated after one hour of chase. A band of ∼30 kDa was also detected by Western blot on whole tachyzoites in reduced condition (data not shown) and was recovered by IP (Figure S2). These results indicated that RON5 is cleaved at least at two sites, one processing event resulting in removal of a pro-sequence (as for many ROP proteins), and another processing event yielding two polypeptides possibly bound by a disulfide bond (as for RON2). Concerning RON8, a processing event was also detected in reduced and non-reduced conditions, indicating that RON8 was also subjected to removal of a pro-sequence. Pulse-chase experiment for RON4 also showed that it is expressed as a pro-protein (∼120 kDa reduced condition and ∼145 kDa unreduced) that is cleaved to yield a mature protein of ∼110 kDa (reduced) or ∼120 kDa (unreduced). One additional minor band of lower molecular mass was also sometimes present. The persistence of the immature form of RON4 after one hour of chase indicated that RON4 was only partially matured. This could be linked to the fact that, as shown before [6],[13], part of RON4 is secreted in the PV (arrow in Figure 2A) and therefore avoids the rhoptry-specific processing compartment. Serendipitously, the generation of a transgenic parasite cell line expressing a Ty-tagged version of RON4 (see Text S1) that was, for unknown reasons, entirely secreted in the vacuolar space (Figure S3A) and remained entirely unprocessed (Figure S3B), strengthened this hypothesis.
In order to determine in which compartment the RONs were processed, we then generated antibodies directed against the RON8 pro-peptide. As for all rhoptry proteins described so far, this latter was assumed to be located N-term and cleaved by the protease TgSUB2 [18]. Three putative TgSUB2 cleavage sites were found in RON8, two in RON5 and one in RON2 (Figure 1D). We therefore raised antibodies against a peptide spanning RON8 AA 1-91, located before the first SFVE motif of the RON8 sequence (Figure 1D). IP using anti-proRON8 demonstrated the specificity of the anti-proRON8 for the immature form of RON8 (Figure S4A), whereas IFA showed reactivity restricted to the characteristic pre-rhoptry compartment (Figure S4B), which corresponds to the nascent rhoptries of daughter parasites during endodyogeny. The cleavage of all RONs beyond the ER was showed by pulse/chase analysis in the presence of the Golgi transport-inhibiting drug brefeldin A (BFA), pro-RONs remaining the only forms of the proteins at the end of the chase (data not shown and Figure 4B).
Since RONs undergo a proteolytic maturation, we analyzed if they could bind as immature proteins or if processing was required for this binding. To this end, we analyzed the MJ complex by pulse-chase experiments, followed by IP in lysis conditions known to preserve the association of the complex (using 1% NP40). As shown in Figure 4B, after a 15 min pulse, the immature forms of RON2, RON8 and RON5 could be recovered after IP with the anti-RON4 monoclonal. Similarly, and as a complementary approach, IP of the pro-forms of the other MJ partners was achieved using anti-RON2n, anti-RON5 or anti-RON8 (data not shown). To confirm these data, we checked for the association of the complex in BFA-treated cells that would express only immature radiolabeled RONs. As expected, in presence of the drug, proRON2 (150 kDa), proRON8 (329 kDa), proRON5 (180 kDa) and proRON4 (120 kDa, reduced) were the only species found at the end of the chase and co-precipitated together (Figure 4B). Pulse-chase with anti-AMA1 serum confirmed that AMA1 was also processed during traffic [19], and that proAMA1 could also associate to pro-RONs (Figure 4B, right panel), a result which was also observed in pulse chase experiment using anti-RON2n (Figure 4A, left panel).
Overall, these results show that all known members of the AMA1-RONs complex could associate together as pro-proteins in vitro.
Since AMA1 and RON2/4/5/8 could interact together as pro-proteins in vitro and would follow the same secretory pathway (i.e. rough ER and Golgi apparatus) before being packaged in their respective compartments in the parasite, we raised the question of how the micronemal protein AMA1 and the complex of rhoptry neck proteins RON2/4/5/8 could avoid interacting before secretion. One possibility would be that they are synthesized sequentially and never coexist in the same compartment. We checked this hypothesis by IFA. Since maturation of both MICs and RONs occurs with rapid kinetics (MICs mature within 15-60 min, [20] and Figure 4A) and the pro-sequence is only transiently detected by IFA, detection using anti-propeptide antibodies faithfully reflects the timing of their synthesis. We therefore performed double IFA with the rabbit polyclonal anti-pro-RON8 and the antiserum raised by Hehl et al. [9] against a peptide corresponding to the AMA1 pro-sequence (data not shown). Unfortunately, this anti-proAMA1 serum gave a very low signal/noise ratio and no significant data could be obtained with this probe. Since we had previously studied the fate of other microneme prodomains and showed for example that 80% of the parasites co-expressed both pro-forms of microneme proteins M2AP and MIC3 simultaneously [21], we reasoned that AMA1 could follow the same scheme and therefore decided to use the mouse anti-proMIC3 and the rabbit anti-proM2AP sera instead. Interestingly, no colocalization of proRON8 and proMIC3 was ever found and both markers were only observed simultaneously in ∼7% of the parasites (±2%, mean±SEM of 3 independent experiments; usually in very large parasites in mid-stages of endodyogeny) (Figure 4C, upper panels). We then took advantage that mAb anti-RON4 did not label the mature rhoptries but only the pre-rhoptries of dividing parasites upon formaldehyde fixation and triton permeabilization [6], to compare the timing of synthesis of RON4 with that of M2AP using rabbit anti-proM2AP. Dual staining using rabbit anti-proM2AP serum and anti-RON4 showed that pre-rhoptry RON4 staining was almost never associated with a proM2AP staining in the same parasite (8.1%±3%, mean±SEM of 3 independent experiments) and apparently not in the same compartment, while RON2, RON5 and RON8 were systematically detected simultaneously with RON4 in pre-rhoptries in the same conditions (data not shown), confirming that RONs and MICs biosynthesis are asynchronous. This would allow MICs and RONs to reach their correct destination without interacting before secretion.
As reported above, RON 2, 4, 5, and 8 are found at the MJ by IFA, but their precise location respectively to the parasite or host cell membrane is not known. We thus sought to determine which membrane these RONs were associated with.
First, we observed that in IFA on parasites invading cells in the presence of Cyt-D, the rhoptry protein ROP1 was sometimes detected on cells in the absence of any surrounding parasite (Figure 5A), suggesting that it would either correspond to an abortive invasion after secretion of the rhoptry content or that the parasite has been mechanically removed by the washes during the experimental procedure. Abortive invasion has been previously documented in a recent mathematical model showing that approximately 55% of the parasites detach within 5 min of initial attachment, but this paper did not conclude on whether the moving junction was built or not before detachment [22]. We thus assessed the presence of RONs in this particular situation. Dual IFA showed that all RONs could usually be detected as a punctuate signal, near the point of contact where the rhoptry content had apparently been initially discharged, as detected by anti-ROP1 staining in the host cell permeabilized with saponin (Figure 5A). RON2 was only detected with anti-RON2c, as reported above. In contrast, no signal was obtained with anti-MIC2 and neither with anti-SAG1 (directed against the major surface antigen of T. gondii), indicating that the signal obtained with anti-RONs was not due to the presence of residual membrane fragments of the parasite (data not shown). Instead, this appeared to reflect a specific association of the RONs with the host cell membrane. It is also to note that we could only rarely detect an AMA1 signal in these conditions. Indeed, quantitative analysis on three independent experiments showed that 89%±2% of the ROP1 evacuoles observed without parasites were RONs-positive while only 8.5%±2% of the ROP1 evacuoles observed without parasites were positive using the anti-AMA1 ectodomain B3.90 mAb [9] (Figure 5A), strongly suggesting that the RONs and AMA1 associate with different membranes.
Second, IFA of HFF cells pulse-invaded for 15 min showed the presence of the PVM marker ROP1 on empty vacuoles (Figure 5B). The same labelling was observed with anti-ROP2 that labelled another associated PVM rhoptry protein (data not shown). Empty vacuoles represented 6%±2% (mean±SEM of 7 independent experiments) of the total vacuole numbers when invasion was synchronized using a K+ buffer shift [23] and corresponded mainly to early egress of the parasite. We then checked the presence of the MJ scar on the PVM of these empty vacuoles. The association of the RONs with the PVM was systematically observed by the immunodetection of RON4/5/8 (but not RON2) on empty PVs labelled with the PVM marker ROP1 but devoid of any parasite (Figure 5B and Figure S6). Since the PV derives from the host cell membrane, this also shows that RONs are associated with the host cell membrane during invasion and probably maintained together as a complex, even after the PVM has pinched off from the host cell membrane.
We sought to address the topology of the RONs at the MJ. To this end, we first analyzed by IFA if the characteristic ring-like pattern of the RONs on invading parasites could be detected in the absence of any permeabilization (which was verified by the absence of labelling of the PVM with anti-ROP2 or anti-ROP1 sera). Since we had observed that the use of formaldehyde to fix parasite during invasion could result in partial permeabilization of the host cell membrane (data not shown), we stopped the invasion process on ice instead and performed the IFA on unfixed cells at 4°C. In these conditions, the MJ complex could not be detected unambiguously with any of the anti-RONs sera. The lack of detection of the epitopes by the antibodies could suggest either a lack of accessibility within the junction, or spatial and conformational constraints or, finally, a localization of these epitopes on the cytoplasmic side of the host cell membrane.
We thus addressed the possible association of RONs with the cytoplasmic face of the host plasma membrane. To this end, we examined the topology of the RONs at the MJ remnant in fully invaded parasites by differential permeabilization. Note that since RON2 was not detected at this residual junction, this approach did not allow defining the topology of RON2 in the host cell membrane. In streptolysin-O (SLO)-treated infected cells, the host cell plasma membrane was selectively permeabilized without affecting the PVM (Beckers et al., 1994), allowing the selective detection of exposed cytosolic domains of PVM-associated protein (Figure 6A). These experiments were carried out with the transgenic GRA5-HA strain [24], where the HA tagged C-terminal end of the PVM marker GRA5, is exposed to the vacuolar space: hence, the absence of C-terminal labelling of GRA5 was used as control of the integrity of the PVM (Figure 6A). In addition, anti-SAG1 antibodies were used as additional control of the integrity of the PVM and to distinguish intracellular parasites (SAG1-negative) from the extracellular ones that remained attached to the cells (SAG1-positive). We first controlled that the RON scar was not detectable in the absence of streptolysin showing that it is inside the cell and not on the surface (data not shown). On samples SLO-permeabilized 15 min after invasion, RON4,5,8 (RON2 could not be detected) were found to be exposed toward the host cell cytoplasm (Figure 6A). This was also confirmed with the detection of RON4, RON5 and RON8 at the surface of isolated intact parasite-containing vacuoles (Text S1, Figure S5).
Since these two approaches allowed the detection of RONs when the parasites had fully invaded, we could not exclude that the topology of the RONs had changed during the closure of the MJ. Thus, we decided to assess the topology of these proteins during the course of invasion. To this end, we pre-loaded the host cells with anti-RONs antibodies by mechanical glass beads loading and subsequently infected the cells with Toxoplasma tachyzoites. In these conditions, RONs would only be detected by the pre-loaded antibodies if they were secreted into the host cell cytoplasm. The cells were then fixed during invasion, permeabilized and subjected to fluorescent secondary antibody detection. The results clearly showed the detection of the ring-shaped MJ with anti-RON4,5,8 (Figure 6B), whereas no signal was detected in the absence of permeabilization (data not shown). When cells were loaded using this technique with anti-RON2 or anti AMA1 antibodies, these proteins could not be detected at the MJ. It is to note that this approach could not be used to assess the inhibitory effect of the anti-RONs on the invasion, as the amount of antibody loaded in the cells is variable and cannot be quantified.
Taken together, our results show that the parasites can secrete the RONs complex directly into the host cell cytoplasm, RON4, RON5 and RON8 remaining associated with the cytoplasmic side of the plasma membrane/PVM during invasion, after which they persist there for a few hours as a residual structure.
The invasion process in Apicomplexa is unique among eukaryotic pathogens in that it involves a MJ structure that is used by the parasite to propel itself inside the cell using its gliding motion. Described morphologically more than 30 years ago, the composition of the structure is still poorly understood at the molecular level. A complex of four proteins AMA1/RON2/RON4/RON5 has been recently described, and two of these proteins (AMA1 and RON4) had been detected at the MJ, AMA1 being associated with the parasite membrane. We have completed the biochemical characterization of the complex and demonstrated that all the RONs are exported to the host cell membrane and, more surprisingly, three are exposed to the cytosolic face of the host cell membrane. The data we have obtained have led us to propose the first tentative model for the molecular organisation of the MJ (Figure 7).
The export of parasite material to the host membrane described here, particularly facing the cytoplasm of the host cell, is perfectly compatible with the thickening of the inner leaflet of the host cell membrane bilayer that has been observed at the MJ by electron microscopy [2]. It is unclear how RONs are exported into the host cell and how they could insert into, or bind to, the host cell membrane, but a secretion of rhoptry material through a transient pore in the host cell membrane has been proposed [4]. Proteins from the bulb of the rhoptry (ROPs) are known to be secreted in association with vesicles (e-vacuoles) into the host cell cytoplasm [16], but this is likely to occur after junction formation and RONs are not found in e-vacuoles. RONs must therefore be translocated at a very early stage, likely corresponding to the transient spike of conductance detected by patch clamp study of T. gondii invasion [25]. An association with lipids might also facilitate membrane insertion. RON4 and RON8, which are not predicted to possess a TM domain, are exposed to the cytoplasmic face of the host cell. RON5, which contains only one putative TM domain in its N-terminal end is, at least partially, exposed on this cytoplasmic face. Although we showed unambiguously that RON2 (which possesses three putative TM domains) is also a component of the MJ associated with the host cell, its precise topology at the membrane will need further studies. Indeed, whereas RON4/5/8 proteins are easily detected at the MJ, RON2 is only observed at a very early stage of junction formation or when the actin cytoskeleton is destabilized with Cyt-D, using a serum directed against a very short sequence located between the last two TM domains. This may reflect a direct or indirect interaction of this domain with the sub-plasmalemmal cytoskeleton, although invasion has been shown to critically depend on actin filaments of the parasite but not of the host cell [26]. Interestingly, only the last two TM domains of RON2 are unambiguously recognised by the prediction programs we used and the loop between these TM domains is particularly well conserved between all Apicomplexa RON2 orthologs, suggesting a conserved function.
T. gondii develops within a vacuole that derives from the host cell membrane. A fascinating phenomenon in Apicomplexa invasion is the selective restricted access of host proteins to the forming vacuole in which the parasite develops. This molecular sieving takes place at the MJ [2],[27]. Indeed, the presence of RONs at the cytoplasmic face of the host cell could also be involved in the exclusion of host plasma membrane proteins from the PV membrane; they would constitute a selective and protective physical barrier that would prevent protein candidates, which could mediate the fusion of the PV with the endo-lysosomal system, to be incorporated and therefore creating a non-fusigenic compartment in which the parasite could develop. These results call for functional studies to assess the respective roles of the RONs in mediating a successful invasion.
A firm attachment of the parasite to the host cell membrane is necessary to propel it inside the PV. To do so, several scenarios might be envisaged. First, parasite ligands might be binding to host cell receptors. Another possibility is that parasite proteins, which would be directly inserted into the host cell membrane, could serve as receptors for the parasite through their extracellular domain. Our data fit perfectly with the latter scenario. Indeed, the parasite targets proteins on both membranes of the MJ; on the parasite surface for AMA1 via its TM domain [10] and on the host cell membrane counterpart, for RON2/4/5/8 (this study). In addition, we have shown that AMA1 can interact directly with RON2 in vitro. The domains interacting together still remain to be mapped. The precise function of AMA1 is not yet known, but several previous works showed a role of the protein in establishing close contact with the host cell suggesting that AMA1 could be involved in a receptor-ligand interaction [7],[28]. However direct binding of Plasmodium or T. gondii AMA1 to the target cell has not been proven unambiguously. Here we propose that the interaction of AMA1 with the host cell could be mediated by a RON2 receptor inserted into the host cell membrane (Figure 7). This model may account for the conserved mechanism of invasion by Apicomplexa. This type of secretion by a pathogen of a receptor for its own invasion machinery is reminiscent of the translocated intimin receptor (TIR) exported by enteropathogenic Escherichia coli [29], but it would be the first one to be characterised for a eukaryotic pathogen.
One crucial question is which protein is dragging the MJ backward during invasion? AMA1 does not possess the critical tryptophan that is necessary for interaction of its C-terminus with the sub-membranous motor [30]. In addition, during invasion, AMA-1 is present at the MJ but the majority of AMA1 is clearly on the parasite surface, on both sides of this adhesion zone [5],[19], implying that at least part of the AMA1 pool is not translocated posteriorly as opposed to other microneme proteins. Consistent with this notion, it is possible that part of the AMA1 pool serves as the anchor for the junction, but that another microneme TM protein interacts with the complex once assembled to propel it backwards in a glideosome-dependent motion.
Apicomplexan parasites show a wide range of host cell specificities that may depend on the expression of the MIC repertoire that differs greatly between parasites or stages of the same parasite; we hypothesize that the conserved process of invasion itself (i.e. MJ-dependent host cell entry) would be rather mediated by the specific protein complex described here, which is mostly conserved among Apicomplexa. However in this study we have characterized a new member of the MJ complex, RON8, which, in contrast to other MJ members, is specific to T. gondii and N. caninum. This highlights the fact that the MJ complex could have a different composition in several Apicomplexa and would suggest that some MJ partners could also account for driving the specificity to the host cell type.
The invasion by apicomplexan parasites is a well-orchestrated mechanism involving the targeting of interacting proteins from two distinct compartments. One intriguing question is how the micronemal protein AMA1 and the RONs complex, which move through a conventional eukaryotic secretory pathway involving the rough ER, the Golgi apparatus and endosome-like structures, avoid interacting before secretion. Here we showed that even if they could physically interact as pro-proteins in cell extracts, it is probably not the case in these intermediate compartments because of a distinct timing of biosynthesis between MICs and RONs. Indeed, all the RONs (in addition to the ROPs, data not shown) are present at the same time in the pre-rhoptries in dividing parasites, whereas newly-synthesized MIC3 and M2AP (their immature forms) are not yet detected in these parasites. They are synthesized later, when the rhoptry compartment has been fully loaded.
This is the first study showing unambiguously that MICs and RONs are not expressed at the same time, which indicates that the biogenesis of rhoptries and micronemes is asynchronous in T. gondii, as previously suggested by ultrastructural analysis of other Apicomplexa such as P. berghei [31]. This distinct timing of biogenesis for proteins destined to two secretory organelles could be a general mechanism of segregation used by the parasite for interacting proteins, which would allow interaction only after secretion and during invasion.
In summary, this study extends significantly our understanding of the MJ formation and composition. The finely-tuned rhoptry and micronemal protein biosynthesis, the cooperation of these proteins originating from two different secretory organelles and the secretion of MJ components directly into the host cell, highlight the sophisticated strategies driving the active invasion of the Apicomplexa.
The primers, antibodies and recombinant proteins generated in this study are described in Text S1.
All T. gondii tachyzoites were grown in human foreskin fibroblasts (HFF) or Vero cells grown in standard condition. Tachyzoites of the RH hxgprt- strain of T. gondii deleted for hypoxanthine guanine phosphoribosyl transferase (ΔHX strain) [32] and GRA5-HA [24] were used throughout the study.
Total RNA was isolated using RNAqueous (Ambion), according to the manufacturer's instructions. cDNA was synthetized from RH hxgprt- parasites using random hexamers and SuperScript II (Invitrogen) or using the SMART RACE cDNA Amplification Kit (Clontech Laboratories, Inc). cDNA fragments of TwinScan_2001 were amplified using a set of primers ML208/ML162, ML209/ML165, ML211/ML210, ML212/ML213 and ML214/ML215, and cloned into the pCR-Blunt II-TOPO vector or into the pCR-2.1-TOPO vector (Invitrogen). After sequencing, the complete open reading frame of RON8 was reconstituted from the overlapping cDNA sequences.
For IFA of intracellular parasites, confluent HFF monolayers were infected with RH tachyzoites for 24 h, then fixed for 30 min in 4% paraformaldehyde (PAF) in PBS. For methanol fixation, monolayers were immersed in methanol for 6 min at −20°C before IFA. After three washes in PBS, cells were permeabilized with 0.2% Triton X-100 in PBS for 10 min, blocked with 10% fetal bovine serum in PBS (PFBS) for 30 min. The cells were stained with primary antibody diluted in 10% PFBS for one hour, washed and then incubated with secondary antibody coupled to Alexa 594 or Alexa 488 (Sigma).
IFA of invading parasites were obtained by synchronisation of invasion at 4°C [6] or using a K+ buffer shift [23]. Invasion was carried out for 2 min30 and was stopped and fixed by adding an excess volume of 4% PAF in PBS. After three washes in PBS, cells were permeabilized with 0.05% saponin (w/v, Sigma) in PBS for 10 min, then IFA was performed as described above. Alternatively, invasion was stopped on ice, and live cells were incubated for 1 h on ice with primary antibodies, before fixation in 4% PAF, saponin permeabilization and incubation with conjugates. When needed, invasion was blocked with Cyt-D treatment by incubation of extracellular parasites with 1 µM of drug for 20 min at 37°C before invasion and then incubation of parasites for 20 min at 37°C in the presence of the drug. SLO permeabilization was conducted as described previously [33].
The coverslips were mounted onto microscope slides using Immunomount (Calbiochem). Observations were performed on a Leica DMRA2 microscope equipped for epifluorescence and images were recorded with a COOLSNAP CCD camera (Photometrics, Tucson, AZ) driven by Metaview (Universal Imaging Co., Downington, PA). Image acquisition was performed on workstations of Montpellier RIO imaging facility.
Loading of antibodies into the host cell was done as described previously [34]. Acid-washed 150–212 µm glass beads (Sigma) were washed 3 times with distilled water. 0.1 mg of beads were then resuspended in 300 µl of the appropriate medium containing the antibody of interest (ie. hybridoma culture supernatant, or antiserum diluted 1/30). HFF cultures growing on coverslips in a 24 wells-plate were washed twice with Hanks' Balanced Salt Solution (HBSS) before the antibodies-beads solution was put into each well. The beads were rolled onto the coverslip by tilting the plate ∼10 times, until evenly distributed over its surface. The coverslip was then transferred to another well where it was washed 3 times with HBSS and returned to DMEM culture medium and allowed to recover at 37°C and 5% CO2 for 30 minutes. Invasion assays were then carried out by allowing T. gondii tachyzoites to sediment on the HFF for 20 minutes at 4°C and subsequently warming them during 2–5 min at 37°C to trigger invasion. Invasion was stopped and cells were fixed by adding an excess volume of 4% PAF in PBS. The extracellular portion of the tachyzoites was labelled with mAb T3 1E5 specific for the surface protein SAG1. Parasites and cells were then permeabilized with saponin and incubated with anti-RONs or anti-AMA1antibodies.
Heavily infected HFF monolayers were incubated in methionine and cysteine-free DMEM (Invitrogen) containing 4% dialyzed FCS for 30 min at 37°C in a 5% CO2 incubator prior to the addition of 50 µCi/ml [35S]-methionine/[35S]-cysteine (700 Ci/mM; Perkin Elmer) with or without BFA (5 µg/ml). The infected monolayers were then labelled for 15 or 20 min, rinsed with complete DMEM containing 10% FCS, and either processed or incubated in this medium complemented or not with BFA (5 µg/ml) for 1 h chase prior to IP. Parasite solubilization in 1% NP40 or in 0.6% SDS and immunosoption procedures were done as described previously [6],[35]. Elution was performed during 5 min at 95°C with electrophoresis sample buffer. After SDS-PAGE, the gel was impregnated with Amplify (Amersham), dried, and exposed to Biomax film (Kodak) at −80°C.
Individual bands from Coomassie stained SDS-PAGE gels were excised, treated with trypsin, and extracted for analysis by nanoflow HPLC-nano-electrospray ionization on a Bruker Esquire 3000+ ion trap mass spectrometer coupled with a LC-Packings HPLC as described previously [6].
|
10.1371/journal.pgen.1002928 | Genomics of Adaptation during Experimental Evolution of the Opportunistic Pathogen Pseudomonas aeruginosa | Adaptation is likely to be an important determinant of the success of many pathogens, for example when colonizing a new host species, when challenged by antibiotic treatment, or in governing the establishment and progress of long-term chronic infection. Yet, the genomic basis of adaptation is poorly understood in general, and for pathogens in particular. We investigated the genetics of adaptation to cystic fibrosis-like culture conditions in the presence and absence of fluoroquinolone antibiotics using the opportunistic pathogen Pseudomonas aeruginosa. Whole-genome sequencing of experimentally evolved isolates revealed parallel evolution at a handful of known antibiotic resistance genes. While the level of antibiotic resistance was largely determined by these known resistance genes, the costs of resistance were instead attributable to a number of mutations that were specific to individual experimental isolates. Notably, stereotypical quinolone resistance mutations in DNA gyrase often co-occurred with other mutations that, together, conferred high levels of resistance but no consistent cost of resistance. This result may explain why these mutations are so prevalent in clinical quinolone-resistant isolates. In addition, genes involved in cyclic-di-GMP signalling were repeatedly mutated in populations evolved in viscous culture media, suggesting a shared mechanism of adaptation to this CF–like growth environment. Experimental evolutionary approaches to understanding pathogen adaptation should provide an important complement to studies of the evolution of clinical isolates.
| Pathogens face a hostile and often novel environment when infecting a new host, and adaptation to this environment can be critical to a pathogen's survival. The genetic basis of pathogen adaptation is in turn important for treatment, since the consistency with which therapies succeed may depend on the extent to which a pathogen adapts via the same routes in different patients. In this study, we investigate adaptation of the bacterium Pseudomonas aeruginosa to laboratory conditions that resemble the lungs of cystic fibrosis patients and to quinolone antibiotics. We find that a handful of genes and genetic pathways are repeatedly involved in adaptation to each condition. Nonetheless, other, less common mutations can play important roles in determining fitness, complicating strategies aimed at reducing the prevalence of antibiotic resistance.
| In the mid-1800's, Louis Pasteur advised microbiologists to think of the human body as a “culture vessel” for microbes, in the context of understanding immunity [1]. Pasteur's approach has been revised and updated several times [2], [3], with a recent review encouraging researchers to be attentive to the effects of different in vivo carbon sources on bacterial metabolism and physiology [2]. Pasteur's advice is particularly relevant for an understanding of the evolution of disease-causing microbes. Natural selection may be imposed by the particular nutritional and metabolic resources available in a given tissue, the innate and adaptive immune systems, and, in the past 80 or so years, by antibiotics or anti-virals. Many pathogens – particularly opportunistic pathogens, emerging pathogens, and microbes causing chronic disease – are faced with a novel and hostile growth environment to which they must adapt or face extinction. Colonization and establishment of an infection in a new host or host species can thus be interpreted as a specific instance of a more general process of adaptation to a novel environment.
Understanding adaptive processes in pathogen populations, and in particular characterizing the variety of genetic routes to adaptation, is important for developing effective treatment strategies. Take as an example the management of antibiotic resistance. Resistance is often thought to be costly, in the sense that resistant strains should be less fit than susceptible strains in the absence of antibiotic. If so, then attempts to reduce the frequency of resistance in patient populations by stopping the use of an antibiotic should afford sensitive strains an advantage, and so prolong the utility of an antibiotic for treatment. Antibiotic cessation has met with mixed success (e.g., [4]–[6]), however, either because some resistance mutations actually pay little or no cost, or because second site mutations that restore fitness without compromising resistance are common. The management of antibiotic resistance in patient populations depends crucially on which of these two mechanisms is more often responsible for the persistence of resistance.
The last 15 years have seen a number of studies of in vivo genome evolution in select pathogens, primarily viruses (e.g., [7], [8]) and bacteria (e.g., [9], [10]), that shed vital insight onto the genetic changes that occur during epidemics or chronic infections. The importance of these changes for pathogen fitness in a host can be difficult to ascertain, however, because it is rarely possible to establish with certainty that the observed mutations are adaptive, since some neutral or deleterious mutations may accumulate through drift or by hitchhiking with adaptive mutations. Moreover, it can be difficult to obtain sufficient in vivo samples to ask questions about the repeatability of in vivo evolution – that is, how often pathogens take the same adaptive routes in independent patients or populations.
For these reasons we have turned to a complementary approach, laboratory selection experiments, to provide an understanding of the broad patterns and principles of pathogen evolution. In a typical microbial experimental evolution protocol, many populations are founded from a single genotype, and are propagated serially or in a chemostat for tens, hundreds, or thousands of generations (reviewed in [11]). By maintaining multiple replicate populations in each of two or more environments (e.g., antibiotic treated vs. not antibiotic treated), the effects of a treatment can be systematically investigated in a manner that is often inaccessible with in vivo samples. Experimental evolution has by now a rich history in studying basic evolutionary processes (e.g., [11]–[14] for reviews), as well as more applied topics such as the evolution of antibiotic resistance [15], [16] and of virulence [7]. In addition, experimental evolution has significant potential as an investigative tool for elucidating basic biological processes [17], [18]. With the development of technologies that allow the rapid and affordable sequencing of entire bacterial genomes, an increasing number of studies have sought to describe the genomic basis of laboratory adaptation (reviewed in [19]).
Here we use a combination of experimental evolution and whole-genome sequencing (WGS) to investigate the initial stages of pathogen adaptation using the bacterium Pseudomonas aeruginosa. This gram-negative bacterium is widely distributed in nature [20], and is an important opportunistic pathogen. P. aeruginosa can cause acute infections of wounds, burns and of lungs, and is frequently implicated in nosocomial infections. Moreover, P. aeruginosa is an important pathogen of individuals with cystic fibrosis (CF), with approximately 60–70% of Canadian adults with CF harbouring this bacterium [21]. P. aeruginosa chronically infects the CF lung, and once the infection is established, it is virtually impossible to eradicate: Intensive antibiotic regimens are effective at reducing symptoms, but almost never succeed in clearing the infection entirely.
P. aeruginosa populations that have persisted for long periods of time in the lungs of individuals with CF show characteristic signatures of adaptation to this novel culture environment. Recent studies have documented patterns of parallel evolution at the level of phenotype, gene expression, and genotype [10], [22]–[25], indicating repeatable patterns of long-term adaptation to the CF lung. For example, CF lung sputum is highly viscous, and P. aeruginosa typically grows as an unattached biofilm, or microcolony, in this environment [26]. While environmental isolates of P. aeruginosa are motile, long-term CF colonists show evidence of adaptation to the sessile lifestyle of the microcolony, including reduced motility, and a morphological shift to small colony variants (SCVs) on agar plates [27], [28]. Increased intracellular levels of cyclic di-GMP are thought to be important for this adaptive shift [27]–[29], but the causative mutations have yet to be fully elucidated. Other characteristic changes include mutations associated with reduced virulence, presumably to avoid detection by the host immune system, and increased small molecule efflux that can afford resistance to antibiotics commonly used with CF patients [10].
Given evidence of long-term adaptation during chronic infection in P. aeruginosa, we have examined the genomic basis of adaptation to CF-like culture conditions and to fluoroquinolone antibiotics through WGS of experimentally evolved P. aeruginosa isolates. Our primary aim is to describe the genetic changes underlying adaptation to this novel environment, and to ask how repeatable these changes are. In addition, we also investigate the genetic architecture of the costs of resistance: When antibiotic resistance evolves, how often is it costly, and what mutations underlie those costs? Our data allow us to quantify the nature and extent of parallel genomic evolution and, in so doing, provide a unique view of the variety of genetic routes taken during adaptation to a medically relevant novel environment.
In our selection experiment, we manipulated the bacterial growth environment so as to resemble the CF lung with respect to nutrition, viscosity, and antibiotic treatment. Populations of P. aeruginosa were evolved in synthetic cystic fibrosis sputum (scfm; [30]) for 8 days in the presence or absence of ciprofloxacin (Cip) and/or mucin. Scfm is a defined medium resembling the nutritional environment of the CF lung [30]. Ciprofloxacin was added at a concentration comparable to that found in the sputum of CF patients (1 ug/ml; [31]). Mucin increases the viscosity of the culture medium, and is meant to mimic the high viscosity of CF sputum [32], [33]. In vivo, viscous sputum is thought to support the growth of P. aeruginosa in unattached biofilms, called microcolonies [26], [34], and similar structures have been observed in mucin-supplemented media (e.g., [32], [33]). Mucin was added at 10 g/L. Mucin may also act as a source of nutrients. The selection experiment comprised a fully factorial design giving four selection environments: scfm alone, scfm+Cip, scfm+mucin, and scfm+mucin+Cip; 12 replicate populations were propagated in each environment. Populations were maintained in a 37°C shaking incubator in 1.5 ml of medium, with serial transfer at a 1∶61 dilution every 24 hours, with approximately 5.9 generations of growth per day (47.5 generations in total).
Since the CF lung – and by extension laboratory media designed to mimic aspects of the CF lung – is a unique growth environment for bacteria, our evolved P. aeruginosa populations are expected to adapt to this novel habitat. Adaptation is also expected to occur in response to ciprofloxacin through the selection of mutations conferring resistance. Our experimental design allows us to disentangle these two effects, with fitness in the absence of antibiotic serving as a measure of adaptation to the growth medium, and changes in resistance to ciprofloxacin indicating adaptation to the presence of this antibiotic. Since populations may harbour extensive genetic and phenotypic variation, we measured resistance and fitness for evolved populations, as well as for a single genotype isolated from each population.
As expected, antibiotic resistance evolved in the presence of ciprofloxacin at both the population and genotype levels (Figure 1). Populations evolved in the presence of Cip showed a 32-fold to 192-fold increase in minimal inhibitory concentration (MIC) over the ancestral genotype Pa14, whereas those evolved in the absence of Cip increased MIC by no more than 2-fold. Single genotypes isolated from each population gave similar results: genotypes evolved in Cip had MICs ranging from 32-fold to 192-fold greater than the ancestor.
To detect adaptation to the growth medium we assayed the fitness of evolved populations and genotypes in the absence of antibiotic using direct, head-to-head competitions against Pa14 (see Materials and Methods). We interpret the population-level assays as a measure of the extent of adaptation achieved, since these reflect the average increase in fitness of all genotypes present at the end of the experiment. The single-genotype assays provide a measure of adaptation for the same genotypes we have sequenced (see below). Note that there will be a close correspondence between measures of fitness at the population and single-genotype levels only if the population is genetically uniform, as expected under a model of periodic strong selection. If, however, the population is genetically polymorphic, perhaps because mutation supply rates (the product of population size, N, and mutation rate, u) are high or distinct genotypes are maintained by negative frequency dependent selection, then adaptation detected at the level of the population may not be accurately predicted by assays of fitness from single genotypes.
Our results are shown in Figure 2, where the dark bars represent the extent of adaptation by entire populations and the light bars adaptation by single genotypes. Evolved populations adapted to the growth medium without antibiotic only when mucin was present in the medium. In the absence of mucin, there was either no response to selection (scfm) or a significant cost to adaptation to Cip (scfm+Cip; ANOVA: P = 2.9×10−5; Table 1). Thus, the presence of mucin in the environment affords a greater opportunity for rapid adaptation.
The single genotype fitness data are more mixed and do not correspond well with the population-level fitness assays (Figure S1), suggesting the presence of substantial amounts of genetic diversity within populations. We saw no consistent effect of mucin or of antibiotic on adaptation to the growth medium, as indicated by a lack of main effect for either of these factors by ANOVA (Table 1). There was, however, a significant interaction between medium and antibiotic (ANOVA: P = 0.013; Table 1), reflecting the observation that scfm+Cip-evolved genotypes were on average more fit than the ancestor (mean relative fitness w = 1.09/generation), whereas the scfm+mucin+Cip-evolved genotypes were on average less fit than the ancestor (mean w = 0.86/generation). This interpretation is reinforced by a lack of correlation between genotypes and populations for MIC, for which there was little correspondence between the level of resistance (Figure S2).
Taken together, these results suggest two important conclusions about short-term adaptation to a CF lung-like environment: (1) adaptation does occur, and it is driven primarily by the presence of mucin; and (2) substantial genetic diversity is likely to be present in evolving populations shortly after colonization, a result consistent with the observation that P. aeruginosa isolates from CF patients can often be highly diverse [10], [35], [36].
In order to gain insight into the genetic causes of adaptation, we sequenced the genomes of the pure genotypes assayed above, with one genotype sampled from each of the 48 evolved populations (that is, a single genotype from each population evolved in scfm alone, scfm+ciprofloxacin, scfm+mucin, and scfm+mucin+ciprofloxacin), as well as of our laboratory's isolate of the ancestral strain Pa14. We obtained a median coverage of ∼56-fold per genotype (mean = 55.5; range 31.8–85.4) on the Illumina platform, using 75-bp paired-end reads. Given that a previous study suggested that 15–20-fold coverage is sufficient for identifying a modest number of mutations in laboratory selected microbial strains [37], the depth of coverage we achieved should allow us to identify all SNPs and small indels throughout most of the genome. In addition, the sequenced genomes were surveyed for large insertion/deletion events, such as mobile element insertions or excisions. We were unable to survey ∼0.53% of the genome in each strain due to low coverage (defined as less than five reads covering a given nucleotide).
Across all 48 evolved strains, we identified 98 SNPs and small indels (mean 2.04/strain) not present in the ancestor (Table S1 lists all mutations and their predicted functional consequences). These mutations represented 77 unique changes, affecting a total of 44 genes and 4 putatively intergenic regions. No large insertion/deletion events were found using BRESEQ [38]. Two genotypes, both isolated from the scfm+mucin+Cip treatment, bore lesions in mutS and were thus likely mutator strains, an inference supported by the relatively high number of mutations found in these strains (one carried 30 mutations, and the other carried 4, representing the 1st and 3rd ranked genotypes in terms of number of mutations), as well as by an extreme transition∶transversion bias amongst point mutations (all 26 point mutations found in these two strains were transitions), which is characteristic of mutS mutants [39]. If these putative mutator strains are omitted, we found 64 mutations (44 unique changes) affecting 20 genes and 1 intergenic region (Figure S3). These mutations included 41 point mutations and 23 insertion/deletions (indels).
Genotypes evolved in the presence of ciprofloxacin or mucin carried more mutations on average than genotypes not evolved with antibiotic (Figure 3). Interestingly, genotypes from the most complex environment, containing both ciprofloxacin and mucin, carried more mutations than any other environment, on average. This result is broadly consistent with the idea that the number of mutations involved in adaptation increases with the number of distinct niche dimensions in the environment, an interpretation supported by both antibiotic and presence/absence of mucin being significant predictors of the number of mutations identified (ANOVA, mutators excluded; medium: F = 8.6, P = 0.005; antibiotic: F = 111.8, P = 2×10−13).
Previous studies of the genomic basis of adaptation in experimentally evolved bacterial populations have detected, on average, 1.07 mutations/100 generations (range: 0.09–3.94; [40]). The numbers of mutations observed after ∼48 generations in our antibiotic-evolved genotypes (mean 2.1 and 2.6 in scfm and scfm+mucin, respectively) are thus substantially higher than observed in previous studies. This difference probably reflects the strong selection imposed by antibiotic treatment, as opposed to the weaker selection commonly observed in resource-adaptation experiments, combined with sufficiently large population sizes to ensure the availability of multiple beneficial mutations in the same population or even the same genome [41]. Notably, the rate of accumulation of adaptive mutations observed here is consistent with theoretical models of substitution under strong selection that show expected fixation times of 50 generations or less for mutations with large selection coefficients (see Figure S4 from [42]). At the opposite end of the spectrum, very few mutations were detected in our scfm populations, with 10 genotypes bearing no mutations, and 2 genotypes carrying a single mutation each. This result is consistent with the lack of fitness response observed above (Figure 2) and is broadly consistent with the theoretical expectation under neutrality, whereby the expected fraction of 6.5 Mb genomes with zero mutations after 48 generations should be 0.73–0.97, depending on the per base pair mutation rate (taken as 1×10−10 to 1×10−9 for these estimates; [43]).
Broad patterns of nucleotide variability suggest that natural selection has played an important role in shaping the observed spectrum of mutations. Amongst the 41 point mutations observed in the non-mutator strains, 39 were nonsynonymous, 1 was synonymous, and 1 was putatively non-coding. Since approximately 1/3 of random coding changes are expected to be synonymous, the lack of synonymous mutations is consistent with natural selection favouring a substantial fraction of the observed mutations in the non-mutators. Using a randomization approach (see Materials and Methods), we find that both the excess of non-synonymous mutations, and the paucity of synonymous mutations, are highly significant (Figure S4; P<0.0005). By contrast, many more synonymous mutations were observed in the putative mutator strains, with 15 non-synonymous, 8 synonymous, 8 genic frame-shifts, and 3 intergenic mutations identified in the 2 putative mutators. The observed counts of non-synonymous and synonymous mutations in these mutators are not significantly different than expected by chance (non-synonymous: P = 0.30; synonymous: P = 0.43), suggesting that many more mutations are neutral and that these strains show a general and unbiased increase in mutation rate. The observed number of intergenic mutations (3) in the mutator strains is significantly higher than expected by chance, however (P = 0.011), suggesting that at least one of these mutations has been driven by selection.
Observed changes in ciprofloxacin MIC and in fitness are attributable to some or all of the mutations identified by WGS. For example, in the ciprofloxacin-evolved strains, we observed multiple mutations in the known fluoroquinolone-resistance genes gyrA, gyrB, and nfxB. Amongst 24 genotypes from populations evolved in the presence of ciprofloxacin, 20 bore mutations in nfxB, 9 carried mutations in gyrB, and 4 genotypes bore gyrA mutations. Each of the gyrA mutations is a known resistance mutation affecting its quinolone-resistance determining region (QRDR; [44], [45]), with one strain carrying a T83I mutation, two with D87G, and one with a D87N mutation. The gyrB mutations were dispersed throughout this gene, with 6 different lesions amongst the 9 strains (Figure 4). In nfxB, loss of function mutations would be expected to be prevalent, since inactivation of this transcriptional repressor results in up-regulation of the MexCD-OprJ efflux pump (e.g., [46]). Concordant with this expectation, 8 distinct mutations were found in nfxB among the 20 genotypes bearing mutations (Figure 4). Interestingly, three sites were mutated in multiple strains (T39P in 3 strains, in a predicted helix-turn-helix DNA-binding domain; E146K in 5 strains; G180S in 8 strains), providing further evidence that these mutations are adaptive.
Additionally, 7 ciprofloxacin-resistant genotypes carried mutations in the gene orfN, 6 being isolated from populations evolved in scfm+Cip. orfN encodes a predicted glycosyl transferase, and is necessary for the glycosylation of type A flagellins [47]. 6 of the orfN mutants carried a single base pair deletion in a poly-G repeat, leading to the introduction of a premature stop codon. The predicted mutant protein is truncated after 53 amino acid residues (vs 338 for the wild-type protein). The seventh orfN mutant carries a single base-pair deletion in a poly-T repeat, leading to a truncated protein of 133 residues. The predicted mutant proteins are truncated before or in the glycosyl transferase domain, suggesting that the orfN mutations are likely to be loss-of-function mutations (Figure 4). While this gene has not previously been associated with fluoroquinolone resistance, this observation of extensive parallel evolution strongly suggests that orfN mutants have increased fitness in the presence of ciprofloxacin.
To obtain further evidence for an effect of orfN and other putative novel resistance mutations on Cip resistance, we surveyed isolates from evolving populations from early time points and assayed their MICs in the genetic backgrounds in which they arose. This approach allows us to sample candidate genes relatively quickly in the context in which they evolved. For orfN mutants, we sampled single colony isolates from early time points (days 3–5 of the evolution experiment) from populations where an orfN mutation was observed at day 8. Early time-point isolates were sequenced at all genes bearing a SNP at day 8, and clones bearing only an orfN mutation were selected. In this way, we identified several apparent single orfN mutants: 2 from population scfm-A5 at day 3, and 1 from population scfm-D6 at day 5. As expected each of these putative single mutants showed a 32-fold elevation in ciprofloxacin MIC in comparison to the ancestral Pa14 genotype, suggesting that orfN is a novel resistance gene.
While the observation of parallel evolution at nfxB, gyrA, gyrB, and orfN is indicative of natural selection acting on these genes, 12 of the mutations identified in the non-mutator strains appeared in only a single isolate each (Figure S5). Such mutations may represent adaptive mutations of minor effect, or they may be neutral mutations that are either segregating due to drift or have hitchhiked alongside other strongly adaptive mutations. In several cases, MIC analyses suggest a benefit to these mutations arising through increased levels of antibiotic resistance. Genotypes containing a single mutation in Pa14_32420 (encoding a putative oxidoreductase) isolated from an early time point (day 3) showed a 4-fold increase in ciprofloxacin MIC and a SNP in Pa14_46110 (encoding a predicted sodium∶solute symporter), which was the third mutation to arise in the population, had an 8-fold higher MIC than did genotypes carrying only the first two mutations (which occurred in nfxB and Pa14_23430). Thus, the evolution of quinolone resistance appears to have involved both highly parallel changes, as well as mutations specific to individual experimental populations.
Previously, Breidenstein et al. [48] conducted a screen of transposable-element insertions for novel ciprofloxacin resistance determinants. Interestingly, there is almost no overlap between between the 114 genes identified by Breidenstein et al. and the 44 genes bearing SNPs in this study. nfxB and mutS mutants were isolated in both experiments, but no other gene was found as a potential resistance factor in both studies. In addition, Breidenstein et al. identified a number of phage-related or phage-derived genes as resistance modifiers, and we found a non-coding mutation in a different cluster of phage-related genes (at position 1927375 of the Pa14 genome). The difference between these two studies is likely due to the different mechanisms that lead to resistance mutations in the two studies: transposon insertions were used by Breidenstein et al. paper, and spontaneous point mutations and indels in the current study. Importantly, the lack of overlap between the two studies is an indication that many genes potentially contribute to fluoroquinolone resistance in P. aeruginosa, and suggests that in general multiple approaches should be taken in the identification of genes underlying phenotypes of interest. Experimental evolutionary approaches, such as the one adopted here, differ from traditional mutational studies in that selection acts as an extra sieve that will weed out slow-growing mutants that, while they confer resistance, are out-competed on the way to fixation by other mutations conferring higher fitness (see [17], [18] for discussions of the use of experimental evolution as a tool for mutation discovery). While this effect of natural selection will likely eliminate some mutations of interest (especially for understanding underlying biological pathways), mutations observed under selection may be more clinically relevant due to their relatively high fitness.
Surveys of clinical samples of Pseudomonas aeruginosa often uncover a handful of genes with major effects on fluoroquinolone resistance. Most commonly, these genes are gyrA and gyrB, which encode the subunits of the fluoroquinolone target DNA gyrase, and the efflux pump regulators nfxB and mexR (e.g., [44], [46], [49], [50]). Given that mutational surveys have revealed many other genes that can confer resistance to fluoroquinolones, why is it that these four genes are repeatedly recovered from clinical samples?
One possibility is that these genes enjoy a large fitness advantage in the presence of antibiotic because they confer large increases in MIC. To test this prediction, we asked to what extent the presence or absence of mutations in classical resistance genes is a predictor of the level of ciprofloxacin resistance. As described above, many of the ciprofloxacin-evolved strains in this study bore mutations in one or several of gyrA, gyrB, and nfxB, although no mexR mutants were isolated. A linear model including selection medium (scfm+Cip or scfm+mucin+Cip) and presence or absence of mutations in gyrA, gyrB, and nfxB explains ∼87% of variation in MIC between genotypes (Table 2, Figure 5A, 5B). Under this model, mutations in nfxB, gyrA, and gyrB are associated with average MIC increases of 25.3, 3.2, and 10.9-fold, respectively. Thus, a substantial fraction of variation in the level of resistance is attributable to mutations in classical resistance genes. It should be noted that the genotypes indicated in Figure 5 are not exhaustive – for example, a given nfxB mutant on Figure 5 will also carry at least one additional mutation. Thus, variation within a genotype class (for example, the nfxB mutants) is attributable to these additional mutations.
An alternative, and not mutually exclusive, possibility is that these mutations pay little cost of resistance in the absence of antibiotic. Cost-free resistance may arise because the mutations themselves are not costly or because second-site mutations rapidly evolve that compensate for whatever cost they do incur. We tested this prediction by examining the fitness of strains bearing (or not) mutations in classical resistance genes in the absence of ciprofloxacin and found little relationship between genotype and fitness (Table 3, Figure 5C, 5D; see also Figure S6). Notably, only strains carrying nfxB mutations from the scfm+Cip environment show an increase in fitness in the absence of antibiotic (Table 3) and none of the gyrA, gyrB, or nfxB mutants from the scfm+mucin+Cip environment were significantly different from the ancestor. This result may be surprising, given that single mutations in gyrA and nfxB are typically costly [16], [51], [52] but we note that none of the strains examined here carried only a gyrA or nfxB mutation; all were at least double mutants. This result suggests that fitness in the absence of antibiotic appears to be determined or modulated by mutations in genes other than nfxB, gyrA, and gyrB. Thus cost-free resistance probably arises through second-site mutations that compensate for the costs incurred by these classical resistance genes, consistent with the results of previous studies [13], [53]–[56]. It is notable that these compensatory mutations would have to have arisen very quickly alongside or soon after resistance had evolved for them to be observed in the short time frame of our experiment.
What sorts of second-site mutations might be involved in compensating for the fitness costs of nfxB, gyrA, or gyrB resistance mutations? Our genome-wide survey of mutations provides some insight. We have found a wide range of mutations amongst the Cip-resistant genotypes sequenced in this study. These include mutations in the gene nusA encoding an elongation factor, a putative kinase encoding gene Pa14_28895, and ate1, which encodes an arginyl-trNA-protein transferase (see Table S2 for a full list).
While genotype at classical resistance genes predicts MIC (but generally not fitness), we find no evidence that the raw number of mutations present in a lineage predicts either MIC or fitness in the absence of antibiotic (data not shown). These data are consistent with a model in which classical resistance genes make particularly large contributions to MIC that can mask the smaller effects of other resistance mutations, even if these latter mutations occur first or provide additional increases to MIC or fitness.
Taken together, these results suggest that the prevalence of classical fluoroquinolone resistance mutations such as those in gyrA and nfxB in clinical isolates is due to the combination of high levels of resistance and apparent lack of costs due to second site mutations. These results are of clinical importance because they suggest that attempts to combat resistance in patient populations by stopping the use of the offending antibiotic in the hopes that drug sensitive types will replace resistant ones will often fail (e.g., [57]). Epidemiological evidence on the effectiveness of this strategy at controlling resistance is both limited and mixed [6], [58]: reducing the use of antibiotics often leads to a reduction in the frequency of resistant strains, but it rarely succeeds in eliminating them altogether [4], [5]. Our results suggest that the mechanistic reason for this failure is not that resistance mutations are cost-free but, rather, that their costs are rapidly compensated for by a diverse array of mutations elsewhere in the genome.
Our genomic analysis also sheds light on the genetic pathways to adaptation in CF-like conditions. Strains evolved in the most CF-like environment, scfm+mucin, often contained mutations in genes implicated in cyclic-di-GMP signalling. Elevated levels of intracellular cyclic-di-GMP are thought to induce a shift from a motile, planktonic lifestyle to a non-motile biofilm state in a variety of bacteria [27]–[29]. We suspect that increases in diguanylate cyclase activity may be adaptive in the presence of mucin, which encourages biofilm growth. Consistent with this hypothesis, three genes with putative roles in diguanylate cyclase signalling were repeatedly found mutated in the evolved strains. 9 of 24 populations (8 without ciprofloxacin, 1 with ciprofloxacin) contained isolates bearing mutations in the morA gene (Figure 4). morA encodes a predicted membrane-localized diguanylate cyclase, and serves as a negative regulator of flagellum formation [59]. In P. aeruginosa, expression of morA is required for the switch from wild-type colony morphology to the small-colony variant morphology [27], which is associated with biofilm formation in CF infections [25]. 7 distinct morA mutations – all missense point mutations - were identified in our evolved strains (Figure 4). Two scfm+mucin-evolved strains bore mutations in wspF, which encodes a regulator of the diguanylate cyclase WspR, with wspF loss-of-function mutants showing increased biofilm formation [28] and wrinkly colony morphologies in Pseudomonas fluorescens [60]. One of the wspF alleles recovered in this study is likely a loss-of-function mutation, since it encodes an early frame-shift. The second allele is a single in-frame codon deletion whose effects we cannot predict. Finally, the gene Pa14_56280, encoding another predicted diguanylate cyclase, was found to be mutated in two further scfm+mucin adapted strains.
In light of the role of cyclic-di-GMP signalling in biofilm formation [27]–[29], we predicted that our putative cyclic-di-GMP signalling mutants should show increased aggregation and biofilm formation. To test this prediction, we examined colony morphology on Coomassie blue/Congo red agar plates, which is a sensitive indicator of aggregation (e.g., [61]–[63]). Isolates bearing mutations in morA, wspF, or Pa14_56280 showed wrinkly, red morphologies in comparison to the ancestral Pa14 strain (Figure 6A–6E), consistent with increased aggregation and biofilm formation. Genotypes bearing mutations in different genes, and even different mutations in the same gene (e.g. for morA, compare Figure 6B and 6C), showed different colony morphologies, suggestive of different effects on the level, timing, and/or localization of aggregation signals, presumably cyclic-di-GMP.
The frequency with which cyclic-di-GMP signalling genes are mutated in our mucin evolved strains – with apparent consequences for aggregation and biofilm formation – strongly suggests a shared mode of adaptation towards a novel in vitro environment. This finding parallels data from clinical isolates of P. aeruginosa: Long-term adaptation of P. aeruginosa to the CF lung is characterized in part by extensive biofilm formation (e.g., [26], [64]) and the switch to a largely non-motile lifestyle is likely mediated by cyclic-di-GMP signalling (e.g., [25], [65]). Notably, wspF mutations have previously been documented in CF isolates (e.g., [10]); the current data suggest several other possible mediators of biofilm formation in clinical isolates.
Unexpectedly, all strains bearing mutations in the quinolone-resistance gene nfxB showed smooth colony morphologies (Figure 6F), a phenotype typically associated with impaired biofilm production (e.g., [61]–[63]). This observation suggests an effect of nfxB on biofilm formation and/or extracellular matrix production, which to our knowledge has not been previously reported.
The extent of parallel evolution during adaptation is of interest for a variety of reasons; evolution is in principle predictable (or not) to the extent that independent populations adapt to similar environments via the same (or different) mutations. The observation of substantial parallel evolution is also used as an indicator of strong positive selection. Previous experimental evolution studies have documented varying degrees of parallel evolution at both the phenotypic and genotypic levels [66]–[73]. We have already noted parallel evolution in response to ciprofloxacin and to mucin in our study, with multiple lineages bearing mutations in the quinolone resistance genes gyrA, gyrB, nfxB, orfN, and in the apparently mucin-adaptive genes morA, Pa14_56280, and wspF. These observations provide strong evidence that these mutations are beneficial.
How prevalent is parallel evolution in our study? To answer this we used the Jaccard index (J) to quantify the extent of within- and between-environment genic parallel evolution. For a given pair of evolved genotypes, J ranges from 0 to 1, with 0 indicating no parallel evolution and 1 indicating identity (see Materials and Methods for further details). We calculated the average Jaccard index for within- and between-environment comparisons, excluding genotypes with no SNPs, as well as mutS mutator strains (Figure S7). Within environments, was highest for the scfm+mucin genotypes, due to the high frequency of morA mutations in this environment. was intermediate for the scfm+Cip and scfm+mucin+Cip genotypes, reflecting parallel evolution at a handful of genes combined with a number of lineage-specific mutations. Between-environments, was 0, except for between the two ciprofloxacin treatments, indicative of some shared mechanisms of resistance. We rarely saw the exact same mutation evolving in parallel selection lines, suggesting that the bulk of parallel evolution in our experiment is through de novo mutations rather than the selection of rare, pre-existing variants. For the few cases where the same mutation was observed in multiple lineages, however, we note that the current study design cannot formally distinguish between these two alternatives since our experimental populations were started from a common founding culture.
We suspect that several different factors contribute to differences in the propensity for parallel evolution at different genes. Chevin et al. [74], analyzing an explicitly genomic model of trait evolution, show that the probability of parallel evolution at a given locus can depend on the locus specific mutation rate, the probability of a mutation being beneficial, and the probability of a mutation going to fixation. For some loci, e.g. nfxB, loss-of-function mutations are likely to be beneficial, and so the probability of a mutation being beneficial will be quite high (see [73] for a similar example). For other loci, such as gyrA, the probability of fixation for beneficial mutations may be high due to their large effects on MIC. Finally, in the case of orfN, where a slippage mutational mechanism is implicated by the observation of single base deletions in repeat regions, both the mutation rate and the probability of a mutation being beneficial are likely to be elevated. Thus, different genes may undergo parallel evolution for rather different reasons.
We have studied the genomic basis of adaptation to CF-like culture conditions and to ciprofloxacin in experimentally evolved isolates of the opportunistic pathogen P. aeruginosa. Adaptation did occur to the most CF-like conditions and to the presence of ciprofloxacin, although our evolving populations are likely highly polymorphic. We observed parallel evolution at a handful of antibiotic resistance genes (gyrA, gyrB, nfxB, and orfN), as well as at putative cyclic-di-GMP signalling genes in the mucin environment. While the level of antibiotic resistance was determined largely by known resistance genes, fitness in the absence of antibiotic was not, such that there was no overall relationship between resistance and its associated costs.
These findings have several implications for understanding antibiotic resistance and pathogen evolution. First, we have identified a suite of novel ciprofloxacin resistance mutations. Our evolved antibiotic resistant isolates harbour mutations in 12 genes not previously implicated in fluoroquinolone resistance, and initial assays are consistent with effects on ciprofloxacin MIC for 3 of these genes (orfN, Pa14_46110, and Pa14_32420). Thus, experimental evolution, coupled with WGS, represents a powerful approach to identifying novel genes of interest.
Second, we find that the costs of resistance are not systematically determined by the same mutations that account for most of the variation in level of resistance (i.e., mutations in gyrA, gyrB, and nfxB). This finding suggests that whatever costs are associated with single resistance mutations are easily remediated by mutations at other loci. Moreover, these results suggest that the prevalence of these resistance mutations in clinical isolates are likely the result both of the high levels of resistance they confer and the rapid compensation of costs by second-site mutations.
Third, the finding of multiple cyclic-di-GMP mutations in the mucin environment underscores the importance of GMP-mediated biofilm formation in viscous environments, such as the CF lung.
Finally, our findings suggest that pathogen evolution has a partially repeatable genomic basis, insofar as some genes are repeatedly mutated in multiple replicate populations, while others are not. This observation has important implications for understanding pathogen evolution. Those genes that show highly parallel evolution may be particularly important in their influence on key adaptive traits governing infection or resistance to antibiotics. However, genes that are mutated only rarely are not necessarily unimportant: they often appear to have important phenotypic consequences, such as compensating for costs of resistance, and so cannot be ignored. In designing novel medical interventions, therefore, our results suggest that we would do well to focus attention first on these common targets of adaptation to the lung environment, while not losing sight of the potential importance of rare and sometimes idiosyncratic mutations that nevertheless play a major role in determining the overall fitness of the pathogen.
A single colony of P. aeruginosa strain Pa14 was grown overnight in minimal medium (NH4Cl 1 g/L, KH2PO4 3 g/L, NaCl 0.5 g/L, Na2HPO4 6.8 g/L; supplemented with CaCl2 15 mg/L, MgSO4 0.5 g/L; 0.8% dextrose as a carbon source). Forty-eight populations were founded from this progenitor by adding 25 µL overnight culture to 1.5 mL of fresh medium (media described below). An aliquot of progenitor was frozen at −80°C in glycerol. Populations were grown on an orbital shaker (150 rpm) at 37°C for 24 hours in 24-well plates. After 24 hours, each population was serially propagated by transferring 25 µL of overnight culture to 1.5 mL of fresh medium. Overnight cultures were frozen at −80°C in glycerol. Seven such transfers were conducted in total, such that approximately 50 generations of evolution occurred (∼5.9/day for 8 days).
Four selection environments were used, consisting of two different media with or without antibiotic. The media were chosen so as to examine the effects of CF sputum nutrition and viscosity on the evolution of antibiotic resistance in P. aeruginosa. Synthetic CF sputum (scfm) was prepared as described by [30]). In order to manipulate viscosity, we added 10 g/L porcine mucin (Sigma) to synthetic CF sputum (scfm+mucin)[32], [33]. For antibiotic treated populations, we used 1 µg/mL ciprofloxacin to mimic the concentration typically found in the sputum of CF patients [31].
For each evolved population, or for pure genotypes isolated from each population, level of resistance was assayed as the minimal inhibitory concentration (MIC) of ciprofloxacin. Overnight cultures were grown in Mueller-Hinton broth (MHB; Sigma), of which 5 µL was inoculated into 195 µL of fresh MHB with varying concentrations of ciprofloxacin in 96 well plates. MIC of the ancestor, i.e., the concentration at which growth was inhibited by 90%, was 0.05 µg/mL. For each evolved strain, we assayed growth at 0x, 0.5x, 1x, 2x, 4x, 8x, 16x, 32x, 64x, 128x, 192x, and 256x the ancestral MIC.
Fitness of each evolved population or genotype was assayed using a competitive fitness assay against a lacZ marked ancestral strain. Independent assays verified that the lacZ-marked strain did not bear a fitness cost in competitions with unmarked Pa14. Both competitors were grown for 24 hours in the competition medium. At time 0, 12.5 µL of marked ancestor and 12.5 µL of evolved strain were inoculated into 1.5 mL of fresh medium in a 24-well plate, and an aliquot was frozen at −80°C in glycerol. Following 24 hours of growth at 37°C at 150 rpm, a final aliquot was frozen at −80°C in glycerol. Serial dilutions of initial and final aliquots were grown on solid minimal media+X-gal, allowing us to determine the numbers of blue (ancestral) and white (evolved) individuals at the beginning and end of the competition. The selection coefficient s was calculated as:Relative fitness w was then calculcated as 1+s, where the units for both w and s are in per generation.
Colony morphology was assayed according to [61]. Briefly, 10 µL of culture grown overnight in LB were spotted in triplicate onto tryptone plates (10 g/L) supplemented with 20 µg/ml Coomassie blue and 40 µg/ml Congo red. Plates were grown for 4 days at room temperature, after which digital photos were taken.
For whole-genome sequencing, a single colony was picked from each evolved population, as well as for the ancestral Pa14 genotype. For each genotype, genomic DNA was extracted from an overnight culture using the Promega Wizard Genomic DNA Purification kit. 75-bp paired-end Illumina sequencing was performed by the Michael Smith Genome Sciences Centre, using DNA barcodes to sequence 10–12 isolates per lane. Mean coverage across all 49 genotypes was 55.5-fold at a quality score of 20 (range: 31.8–85.4).
We performed a pair-end mapping of reads on the Pa14 reference genome number NC_008463.1 using novoalign (http://novocraft.com/main/index.php). We used samtools [75] to call snps/indels, and filtered the resulting calls using the provided samtools.pl script, changing the window size for snps around indels at 5 base pairs, removing the limit on number of reads spanning a snp/indel position, and leaving the remaining parameters at their default values. We further filtered calls with quality scores below 60 for indels, and 20 for snps. To annotate the remaining snps/indels with respect to the reference genome, we used snpEff (http://snpeff.sourceforge.net/). We found results to be robust to performing a pre-mapping clipping of reads based on quality across cycles using FastQC (http://www.bioinformatics.bbsrc.ac.uk/projects/fastqc/), and to performing local multiple sequence re-alignment around indels using the Genome Analysis ToolKit [76]. We also used the BRESEQ [38] pipeline as a further validation, and for its insertion/deletions detection capabilities.
Following removal of common assembly errors using custom perl scripts, a subset of SNPs was verified by Sanger sequencing of polymerase chain reaction (PCR) amplicons. For each of 31 mutations (out of 98 mutations identified in the 48 evolved strains), we amplified a 500–700 bp PCR product containing the putative SNP, and directly sequenced the PCR products (Genome Quebec, Montreal). All 31 mutations that we interrogated were successfully verified.
We used a randomization approach to determine the probability of observing by chance the distribution of non-synonymous, synonymous, and intergenic point mutations. This analysis was performed separately for putative mutator strains (two mutS mutants) and for putative non-mutator strains (the remaining 46 strains). 10 000 sets of point mutations were generated at random from the Pa14 genome sequence, maintaining the observed numbers of transitions and transversions (mutators: 30 transitions and 11 transversions; non-muators: 26 transitions and 0 transversions), and SNP effects were predicted using snpEff. Mean numbers of non-synonymous, synonymous, and intergenic mutations, as well as the 2.5% and 97.5% quantiles of the random distribution, were calculated in R [77].
The extent of parallel evolution was quantified using the Jaccard Index J. Given two sets G1 and G2 that list mutation-bearing genes found in genotypes 1 and 2, respectively,That is, J is the number of genes mutated in both strains divided by the total number of genes mutated in genotype 1 or in genotype 2. J ranges from 0 to 1, with 1 indicating identical genotypes and 0 indicating no shared mutations.
J was calculated for all possible pairs of different genotypes amongst the 46 non-mutator strains. The average Jaccard Index was calculated within a treatment group as the mean J for all pairs of strains, where both strains were evolved under the same treatment. Similarly, was calculated between treatments A and B as the mean J for all pairs of strains, where one strain was evolved under treatment A and the second strain was evolved under treatment B.
|
10.1371/journal.ppat.1006604 | Divergent LysM effectors contribute to the virulence of Beauveria bassiana by evasion of insect immune defenses | The lysin motif (LysM) containing proteins can bind chitin and are ubiquitous in various organisms including fungi. In plant pathogenic fungi, a few LysM proteins have been characterized as effectors to suppress chitin-induced immunity in plant hosts and therefore contribute to fungal virulence. The effector mechanism is still questioned in fungus-animal interactions. In this study, we found that LysM proteins are also present in animal pathogenic fungi and have evolved divergently. The genome of the insect pathogen Beauveria bassiana encodes 12 LysM proteins, and the genes were differentially transcribed by the fungus when grown in different conditions. Deletion of six genes that were expressed by the fungus growing in insects revealed that two, Blys2 and Blys5, were required for full fungal virulence. Both proteins could bind chitin and Blys5 (containing two LysM domains) could additionally bind chitosan and cellulose. Truncation analysis of Blys2 (containing five LysM domains) indicated that the combination of LysM domains could determine protein-binding affinity and specificity for different carbohydrates. Relative to the wild-type strain, loss of Blys2 or Blys5 could impair fungal propagation in insect hemocoels and lead to the upregulation of antifungal gene in insects. Interestingly, the virulence defects of ΔBlys2 and ΔBlys5 could be fully restored by complementation with the Slp1 effector from the rice blast fungus Magnaporthe oryzae. In contrast to Slp1 and Blys2, Blys5 could potentially protect fungal hyphae against chitinase hydrolysis. The results of this study not only advance the understanding of LysM protein evolution but also establish the effector mechanism of fungus-animal interactions.
| Insect pathogenic fungi are of importance for both applied and basic research. Relative to the advances in understanding fungus-plant interactions, the mechanisms of the molecular pathogenesis of entomopathogenic fungi are rather limitedly understood. In particular, the machinery of effector-mediated inhibition of host immunity has not been well established in fungus-insect interactions. LysM effectors have been characterized as virulence factors in plant pathogens to suppress chitin-triggered immunity in plants. We found that the divergent LysM proteins are also present in animal pathogens. By using the insect pathogen Beauveria bassiana as a model, we revealed that two of 12 encoded LysM protein genes Blys2 and Blys5 that were transcribed by the fungus growing in insects are required for full fungal virulence against insect hosts. Interestingly, the virulence defects of ΔBys2 and ΔBys5 could be fully restored by complementation with the divergent Slp1 effector from the plant pathogen Magnaporthe oryzae. Both Blys2 and Blys5 can deregulate insect immune responses, and the latter can additionally protect fungal cells from chitinase hydrolysis. The findings of this study establish the contribution of LysM effectors to fungal virulence against insect hosts.
| Insect pathogenic fungi such as the ascomycete species Beauveria bassiana and Metarhizium robertsii have been developed as promising biocontrol agents [1, 2]. Fungal species such as B. bassiana, M. robertsii and M. acridum have also been investigated as genetically tractable systems to unravel the mechanisms of fungus-insect interactions [3]. Similar to plant pathogenic fungi, various strategies ranging from cell wall remodeling to the secretion of immune suppressors have been employed by insect pathogens to evade insect immune responses [3, 4]. For example, the coat protein Mcl1 can be highly expressed by M. robertsii during fungal growth in the insect hemocoel (body cavity) to camouflage fungal cells from hemocyte recognition and encapsulation [5]. The blastospores of B. bassiana isolated from insect hemolymph have also been found with shielded carbohydrate epitopes to counteract insect immune defenses [6, 7]. It has also been reported that small molecules such as the cyclopeptide destruxins produced by Metarhizium species and the red pigment oosporein produced by B. bassiana can facilitate fungal infection of insect hosts by inhibiting host immunities [8, 9]. Relative to the well-understood mechanisms of fungus-plant interactions [10], the effector-mediated perturbation of host immunities has not been fully elucidated in animal pathogens including both insect and mammalian pathogenic fungi.
The chitin in cell walls is a well-characterized pathogen-associated molecular pattern (PAMP) in plant pathogenic fungi that can trigger host immune responses [11]. The lysin motif (LysM)-containing receptor kinases have been investigated in plants for mediating recognition of PAMP chitin during microbe-plant interactions [12, 13]. The diverse LysM proteins are also widely distributed in the fungal kingdom and have been characterized as chitin-binding effectors in plant pathogens to deregulate host immunity [14]. Of these, Ecp6 is the first characterized LysM effector that contributes to the virulence of the tomato leaf mold Cladosporum fulvum [15, 16]. It was later found that LysM proteins can also function as virulence factors in other plant pathogens, such as Slp1 in the rice blast fungus Magnaporthe oryzae [17], Mg1LysM and Mg3LysM in the wheat pathogen Mycosphaerella graminicola [18], ChElp1 and ChElp2 in the plant anthracnose fungus Collectotrichum higginsianum [19], and Vd2LysM in the soil borne wilt disease Verticillium dahliae [20]. The virulence effect of LysM proteins in animal pathogenic fungi is still unclear.
Functional studies of Ecp6 and Slp1 revealed that these LysM effectors could sequester the chitin oligosaccharides released from the cell walls of invading fungal hyphae to prevent the activation of plant immunity and/or outcompete the host immune receptor for chitin binding [16, 17, 21]. The chitin in fungal cell walls can be targeted and degraded by plant chitinases [12]. LysM effectors, such as the Mg1LysM and Mg3LysM of M. graminicola, can protect fungal hyphae against the hydrolytic activity of plant-derived chitinases [22]. In contrast, both Slp1 and Ecp6 cannot protect fungal cells from chitinase hydrolysis [16, 17]. LysM proteins are also widely distributed in animal fungal pathogens [23, 24], and chitin can trigger immune responses in both mammals and insects [25]. These data would suggest the existence of a LysM effector machinery in fungus-animal interactions, which, however, has been questioned due to the consideration of the non-intimate relationships between fungi and animals compared to that which exist for plasmalemma-enveloped fungus and plant cells [11, 26]. Molecular evidence is still required to substantiate these arguments.
Our previous genomic analysis of insect pathogenic fungi identified an array of plant pathogen-like effectors, including LysM proteins [24, 27]. In this study, we characterized the function of LysM proteins encoded in the genome of the insect pathogenic fungus B. bassiana and found that divergent proteins can bind chitin polymers and fungal cell walls to deregulate insect immune defenses. Of particular interest, we found that the Slp1 effector from the rice blast fungus could restore the virulence defect of the gene deletion mutants against insect hosts.
We first performed bioinformatic analyses of fungal LysM proteins by including those from insect and mammalian pathogenic fungi. Consistent with previous observations [14, 28], we found that the number of LysM proteins varied highly among fungal species (S1 and S2 Tables). For example, 12 LysM proteins are present in the genome of B. bassiana (termed Blys1-Blys12, Fig 1) whereas 13 in Metarhizium robertsii, eight are present in M. oryzae and 18 in C. higginsianum. Fungal LysM proteins are usually cysteine-rich and vary in length [14]. By plotting protein length versus the cysteine ratio for each protein, we found that the 282 examined fungal proteins could be divided into two groups: one group containing proteins of 90–850 aa and 2–7% cysteine residues (g1 group), and the other group containing proteins > 850 aa and 2.5–4% cysteine (g2 group) (S1 Fig). In the g2 group, the proteins from the plant pathogens are highly underrepresented (5/96) compared to those from the insect (41/133) and mammalian (10/53) pathogens. In addition, we found that most LysM proteins contain a signal peptide (78% of 133 examined proteins from insect pathogens) (S1 Table), which is considerably higher than the genome-wide average (ca. 15%) for secretable proteins [29].
As evident in 12 LysM proteins from B. bassiana (Fig 1), varied numbers (1–7) of LysM domains (termed as LysMs for abbreviation) are present in various fungal proteins, which may or may not contain the additional ChtBD1 type of chitin-binding domain, the Glyco_18 type of chitinase domain and/or the Hce2 effector domain (Pfam: PF14856) (S2 Fig). Phylogenetic analysis indicated that the examined proteins could be grouped into different lineages partially through association with protein structures. For example, most of the large LysM proteins that contain the ChtBD1, Glyco_18 and/or Hce2 domains are clustered together (Cluster I), as are the intracellular proteins with a single LysM domain (Cluster II, including Blys4) (S2 Fig). For Blys1-12 of B. bassiana, Blys10 and Blys11, and Blys9 and Blys12 are close to each other, whereas the rest of the Blys proteins fall into different lineages. Individual LysM sequences were also retrieved from each protein for phylogenetic analysis, and the results indicated that more than five clustering patterns could be obtained (S3 Fig). Even most of LysMs were clustered together independent of their sequential positions within the parental proteins (C4 cluster), the LysM1 (C2 cluster) and LysM2 (C3 cluster) from those proteins containing additional ChtBD1 and Glyco_18 domains could be grouped into respective lineages (S3 Fig), i.e., the relationships with protein structures.
The LysMs of the effectors Ecp6, Slp1 and Mg3LysM are more similar to the counterparts from bacteria with zero or one cysteine residue [14]. Consistently, the LysM domains from these proteins were clustered together into a basal lineage (S3 Fig). In addition, further analysis of LysM sequence consensus indicated that the LysMs from plant pathogens could be divided into two types, i.e., the pattern similar to those from bacteria (S4A and S4B Fig), and the pattern containing four cysteine residues (S4C Fig). The LysMs from insect pathogens contain four cysteine residues (S4D Fig), i.e., the typical fungal-specific LysMs that can putatively form two disulfide bridges within each domain [14]. The divergently evolved LysM proteins suggest functional diversities of these proteins in fungal biology.
As indicated above, 12 LysM proteins are encoded by B. bassiana, and these proteins vary in length and contain a different number of LysM and/or other domains (Fig 1). Except for Blys4, all other proteins each contain a signal peptide. We performed gene expression analysis by growing the fungus in various conditions, including the in vivo infection stages within the insect hemocoels. The results indicated that these genes were differentially expressed by the fungus (Fig 2A). For example, Blys7 and Blys8 were highly transcribed in the conidia. Relative to growth on solid medium, fewer genes were expressed by the fungus growing in an artificial liquid medium. In particular, six genes (i.e., Blys2, Blys4, Blys5, Blys6, Blys7 and Blys8) were differentially transcribed by the fungus during in vivo infection of insect hosts. The closely related Blys10 and Blys11, and Blys9 and Blys12 remained largely silent in B. bassiana under the examined conditions.
The LysM effectors expressed by plant pathogens during colonization of hosts are required for fungal virulence [11]. To examine the virulence contribution of LysM proteins in B. bassiana, we performed homologous recombination-mediated deletions of the six genes upregulated by the fungus growing in insects. Different null mutants were obtained and verified by reverse-transcription PCR (RT-PCR) (S5A Fig). Deletion of these genes had no obvious negative effect on fungal growth on potato dextrose agar (PDA) and PDA amended with Calcofluor White. However, ΔBlys4 and ΔBlys5 became relatively tolerant against H2O2-induced oxidative stress when compared to the wild type (WT) (S5B Fig). Both the WT and mutant cultures could not grow at 37°C. We conducted both injection and topical infection bioassays using the last instar larvae of the wax moth Galleria mellonella to compare the virulence difference between the WT and null mutants of B. bassiana. The estimation and statistical comparison of the median lethal time (LT50) values indicated that the deletions of Blys2 and Blys5, but not the other genes, significantly (P < 0.01) impaired fungal virulence in both types of bioassays (Table 1; S6 Fig). For example, during the injection assays, the LT50 values of ΔBlys2 (3.133 ± 0.095 d) and ΔBlys5 (2.944 ± 0.108 d) was significantly (P < 0.01) extended compared to that of the WT (2.600 ± 0.097 d). Similar to the virulence contributions of the LysM effectors in plant pathogens [11, 28], both Blys2 and Blys5 are therefore required for the full virulence of B. bassiana to infect insect hosts. Unfortunately, the trials to delete both the Blys2 and Blys5 genes were not successful for reasons that remain unclear.
Additional mutants were generated to overexpress Blys2 in the WT strain, whereby Blys2 expression was made under the control of the constitutive promoter of either the GpdA or laccase gene of B. bassiana [30]. We also tried to rescue the Blys2 and Blys5 deletion mutants using the M. oryzae Slp1 gene under the control of the GpdA promoter for transformation. Both Blys2 and Slp1 could be successfully transcribed by the obtained mutants (Fig 2B). The injection and topical infection bioassays indicated that the LT50 values of Blys2 overexpression mutants had no obvious difference (P > 0.1) from those of the WT. Interestingly, complementation of ΔBlys2 with Slp1 fully restored the virulence defect of ΔBlys2 (Table 1; S7A and S7B Fig). Likewise, the statistical difference between WT and ΔBlys5::Slp1 was non-significant (χ2 = 2.081; P = 0.149) (S7C Fig). Nevertheless, the full sequence and two LysM domains of Slp1 exhibited no obvious conservation and phylogenetic relatedness with those of Blys2 and Blys5 (S8 Fig).
It has been found that the LysM domains determine the protein-binding affinity to chitin or specificity to other type of carbohydrates [14]. Blys2 contains five LysMs whereas Blys5 has two (Fig 1). To determine their binding abilities to different carbohydrates, the cDNAs of both genes were cloned into the vector for expression in Escherichia coli. In addition, different truncated forms of Blys2 with a reduced number of LysM domains were also expressed. The purified proteins were all soluble (Fig 3A). The saturated affinity binding assays were performed using the chitin polymers extracted from the conidial cell wall of B. bassiana, chitin beads, chitosan and cellulose. The results indicated that, similar to Ecp6 [16] and Slp1 [17], both Blys2 and Blys5 could bind fungal cell wall chitin and chitin beads. However, in contrast to Blys2, Blys5 could also target chitosan and cellulose (Fig 3B). The truncated forms of Blys2, i.e., Blys2D1-2 (Blys2 only contains first two LysM domains), Blys2D1-4, Blys2D2-5, Blys2D3-5 and Blys2D4-5, could also bind fungal chitin and chitin beads. Unlike Blys2D1-2 and Blys2D1-4, the other forms also slightly bound chitosan and cellulose. In addition, relative to the full protein and other truncated forms, the binding ability of the forms Blys2D3-5 and Blys2D4-5 was reduced, as these truncated proteins were also detected in the supernatant samples (Fig 3B).
Different LysM proteins play a distinct role in protecting fungal cells from chitinase hydrolysis [17, 22]. To determine the protection potential of Blys2 and Blys5 against chitinase, the WT spores were germinated in a liquid medium for 16 hrs. The germlings were incubated with each protein and then treated with chitinase cocktails to compare the difference in formation of protoplasts. The results indicated that a similar ratio (P = 0.1618) of protoplasts was released from the Blys2-treated sample whereas a significantly fewer number of protoplasts was formed from the Blys5-incubated germlings (P < 0.0001) when compared to the mock control (the WT germlings without protein incubation). Consistent with the previous report that Slp1 cannot protect fungal hyphae from chitinase degradation [16], non-significant difference (P = 0.0919) was observed between the mock and WT::Slp1 samples (Fig 3C). Thus, Blys5 but not Blys2 can protect fungal hyphae against the hydrolytic activity of chitinase.
A recent report indicated that Blys2 was identified as a cell wall protein from both the blastospores and hyphal bodies of B. bassiana [7]. To determine the secretion and localization feature of Blys2, a green fluorescent protein (GFP) was fused in frame at the C-terminus of Blys2. The cleaved form of Blys2 without the signal peptide (Blys2-SP) was also generated, and both cassettes were placed under the control of the GpdA gene promoter for transformations of the WT strain of B. bassiana. Both types of proteins could be successfully expressed and detected in the mycelial-protein samples of B. bassiana by Western blot analysis using an antibody against the GFP protein. However, in contrast to the WT Blys2, Blys2-SP could not be detected in the culture medium. The result confirmed that Blys2 is an extracellular protein (Fig 2C).
To examine the localization of Blys2, fungal cells were stained with the fluorochrome Calcofluor White for chitin labeling. We found that the Blys2-GFP signal could be detected in the chitin-stained cell walls of various cell types (labelled as WT::gp-Blys2-GFP), including conidial spores, hyphae, blastospores and hyphal bodies (Fig 4A). This result was consistent with the observation that Blys2 can bind fungal cell wall chitin (Fig 3B). By contrast, smeared GFP signals were observed in mutant cells (labelled as WT::gp-Blys2-SP-GFP) expressing Blys2-SP-GFP (Fig 4B). This finding was similar to those obtained from the cells (labelled as WT::gp-GFP) that were only transformed with a GFP gene (Fig 4C). In addition, consistent with the chitin-binding nature of Slp1 [16], we found that GFP-fused Slp1 could also be detected in the mutant (labelled as WT::gp-Slp1-GFP) cell walls of B. bassiana (Fig 4D). In comparison to other types of cells, chitin-staining, Blys2-GFP and Slp1-GFP signals were more weakly detected in hyphal bodies that were harvested from insect hemocoels. This result could be due to the occurrence of cell wall re-modification with reduced contents of chitin and β-glucans during fungal growth in insect body cavities [31].
Having established that Blys2 and Blys5 are required for full fungal virulence and can bind fungal cell wall chitin, we performed further experiments to investigate the mechanism of protein virulence contribution. Thus, the spores of the WT, ΔBlys2, ΔBlys5 and Slp1-rescued mutants were injected into the last instar of wax moth larvae, and the insects were bled at various times to examine fungal developments. We found that the typical cellular immune responses, i.e., hemocyte encapsulation and melanization [5], similarly occurred in insects against both WT and mutant spores up to 24 hrs post injection (Fig 5). However, in contrast to ΔBlys2, the WT, Blys2 overexpression (i.e., the mutants WT::gp-Blys2 and WT::lp-Blys2) and Slp1-rescued mutants (ΔBlys2::Slp1) began to produce hyphal bodies 36 hrs post treatments. Subsequently, significantly fewer (P < 0.0001) free-living cells were produced by ΔBlys2 when compared to the WT 48 hrs post injection. We also found that, relative to the WT, the propagation of ΔBlys5 was considerably (P < 0.0001) impaired in insect hemocoels as well. However, no obvious differences (P > 0.05) were observed between the WT and other mutants (Fig 6A). In addition, a quantitative real-time PCR (qRT-PCR) analysis indicated that the expression of the antifungal gallerimycin gene could be more highly (P < 0.05) induced in insects infected by ΔBlys2 than by the WT (Fig 6B). In comparison to the WT, deletion of Blys5 could also lead to a higher level (P < 0.05) of induction of the antifungal gene expression in insects (Fig 6C). However, no significant difference was observed between the WT and other mutants including the Slp1-rescued mutants. Thus, deletion of either Blys2 or Blys5 could impair fungal propagations and ability in suppressing immune responses in insects, and the mutant defects could be complemented with the Magnaporthe Slp1 gene.
To perturb chitin detection by host cells, plant pathogens have evolved various strategies to deregulate host immune responses [10, 11]. It has been posited that animal pathogens may not require any effector due to the non-intimate relationships between pathogens and animal cells [11, 26]. In this study, we report that the genomes of animal pathogenic fungi encode different numbers of LysM domain-containing proteins. Our functional studies revealed that two of 12 LysM proteins, i.e., Blys2 and Blys5, are required for the full virulence of the insect pathogen B. bassiana. These proteins can be secreted and function as bona fide effectors by targeting fungal cell wall chitin to deregulate insect host immunities. Of particular interest, the gene-rescue experiment with the Slp1 effector from the rice blast fungus M. oryzae could restore the virulence defect of ΔBlys2 and ΔBlys5 against insect hosts. The results of this study confirm that animal pathogens can employ a similar strategy to that used by plant pathogenic fungi for the effector-mediated evasion of host immune defenses to facilitate fungal infection.
LysM proteins are widely distributed in different organisms, from bacteria to fungi to plants [14]. Protein length diversity, domain number and sequence variations of LysM proteins are observed in different fungi, even between closely related fungal species. For example, 12 LysM proteins are present in the genome of B. bassiana but 14 in B. brongniartii. The inter- and intra-specific functional diversities of LysM proteins are still unknown. We found that the 12 Blys genes were differentially regulated by B. bassiana, and only the Blys2 and Blys5 genes that were activated in insect hemocoels were virulence factors. A recent proteomic investigation of the cell wall proteins of B. bassiana indicated that Blys8 could be identified from the hyphal bodies isolated from insect hemocoels [7]. We found that deletion of Blys8 did not impair fungal virulence. Considering that B. bassiana is also a plant endophyte [32], other Blys proteins may be involved in fungal interactions with plants that remain to be determined.
Variations in LysM protein numbers were also observed among different strains of the plant pathogen V. dahliae. Moreover, a lineage-specific LysM effector but not the core LysM proteins (those present in all strains) was found to contribute to the virulence of the strain to alternative plant hosts [20]. This finding suggests that the LysM effector may play a role in influencing fungal host ranges. For insect pathogenic Metarhizium species, 13 LysM proteins are present in the genome of the generalist pathogen M. robertsii, whereas only four are present in the specialist species M. album and M. acridum [27]. This variation further suggests that the LysM protein may be connected with fungal lifestyles, including host specificity. In addition to suppress chitin-triggered immunity, the LysM proteins ChElp1 and ChElp2 of C. higginsianum are also required for the appressorium-mediated penetration of plant cells [19]. Interestingly, the LysM protein Tal6 from the mycoparasite Trichoderma atroviride could specifically inhibit the germination of the spores of Trichoderma species but not other fungi [33]. Thus, the exact function(s) of LysM protein remains to be determined in a species-specific manner.
Varied numbers of LysM domains (from 1–7) are present in different proteins. Except for the lineage-specific clustering pattern of the LysM1 and LysM2 domains retrieved from those proteins containing additional ChtBD1 and Glyco_18 domains, most LysM domains are clustered independent of their positions within the parental proteins (S3 Fig). Functional diversities of these variations are unclear. Structural analysis of the effector Ecp6 (containing three LysM domains) indicated that LysM1 and LysM3 can tightly form an intrachain dimer to mediate chitin binding with ultra-high affinity whereas the remaining LysM2 binds chitin with low affinity [21]. The effectors Slp1, ChElp1, ChElp2 and Vd2LysM each have two LysM domains, whereas Mg1LysM only contains a single LysM [17, 19, 20]. The formation of LysM dimers between the compartmented domains is therefore not applicable to these proteins that all can bind chitin but not chitosan, xylan and cellulose. We performed truncation studies of Blys2 and found that the five truncated forms retained the ability to bind chitin. However, the Blys2D2-5 form could additionally bind chitosan and cellulose with low affinity, whereas the losses of LysMs 1–2 (i.e., Blys2D3-5) or LysMs 1–3 (i.e., Blys2D4-5) reduced the mutant proteins’ chitin-binding affinity (Fig 3B). These results would suggest that the LysM1 of Blys2 might affect the protein binding specificity whereas the LysM2 and LysM3 domains might determine the protein’s chitin-binding affinity. Unlike Blys2, Blys5 with two LysM domains could additionally bind chitosan and cellulose. It has been reported that the Tal6 of Trichoderma, with seven LysM domains, could bind chitosan but not the chitin and cell walls of various fungal species. However, the truncated form of Tal6 that contains the last four LysM domains could bind colloidal chitin but not chitin powder and chitin flakes [33]. Thus, both the sequence and combination of LysM domains jointly determine the carbohydrate-binding specificity and/or affinity of LysM proteins.
To suppress the chitin-induced immune responses in plants, the LysM effectors upregulated by fungal pathogens can function as a competitive inhibitor of plant chitin receptors, a scavenger of chitin oligomers and/or a protective coat to shield fungal cells from the hydrolytic activity of plant chitinases [11, 12]. In contrast to the identification of LysM kinase receptors in plants [34], chitin receptor has not been identified in insects. Nevertheless, chitin oligomers could activate the expression of antimicrobial peptides in insects [35]. It has also been reported that the alternative chitinases encoded in insects might play a role in immune defense against chitin-containing pathogens [36]. In this respect, it is not surprising that entomopathogenic fungi have evolved similar strategies to suppress chitin-induced immunity in insects. Similar to the finding that extracellular ChElp2 is localized in fungal cell walls of C. higginsianum at the biotrophic interface [19], secreted Blys2 can target the cell walls of B. bassiana. In addition, we found that the loss of Blys2 could lead to the delay of fungal cell escape from hemocyte encapsulation and the upregulation of antifungal peptide gene in insects. Deletion of Blys5 could also result in a higher level of activation of antifungal gene expression in insects when compared to the infection by the WT strain of B. bassiana. In contrast to Blys2, Blys5 can additionally protect fungal hyphae against chitinase degradation. Taken together, these results suggest therefore that the Blys2 effector could coat and protect the cell walls of insect pathogens from host cell recognition while Blys5 could additionally shield fungal cells from the hydrolysis of insect chitinases, the non-redundant functions of these two proteins in evasion of insect immune defenses.
The study of Slp1 in M. oryzae revealed that the effector is a competitive inhibitor of the rice receptor CEPiB to suppress chitin-induced immune responses in rice cells [17]. Since the virulence defects of ΔBlys2 and ΔBlys5 could be heterologously restored by Slp1, it cannot be ruled out that both Blys2 and Blys5 may also be able to outcompete the chitin receptor of insects, if any, to deregulate host immunities. Cell wall re-modifications occur during fungal invasion of the insect body cavity [3]. Consistent with a previous observation [31], we found that the chitin content of cell walls was reduced when the fungus was growing in insect hemocoels (Fig 4). This result could help explain why the overexpression of Blys2 and Slp1 did not lead to an obvious increase of fungal virulence that might be due to the saturation of the chitin substrate. As indicated above, Blys5 can additionally bind chitosan, a biopolymer that is rich in insects and has an immediate antifungal activity [37]. Thus, besides its shield effect against chitinases, Blys5 may additionally contribute to the detoxification of insect chitosan to facilitate fungal growth in insects.
In conclusion, we report the presence of diverse LysM proteins in animal pathogenic fungi and reveal that, similar to the LysM effectors of plant pathogenic fungi, the extracellular LysM proteins are virulence factors of the insect pathogen B. bassiana. Alternative proteins can bind chitin, coat fungal cell walls, deregulate insect immunities and/or protect fungal cells from host chitinase damage to facilitate fungal infection. Intriguingly, we found that the highly divergent plant pathogen effector Slp1 could restore the virulence defect of the Blys2 and Blys5 deletion mutants. The results of this study not only expand our knowledge of LysM protein evolution and functional diversification/similarity but also establish that the employment of effectors to evade host immunities similarly occurs during fungal interactions with animal hosts.
The WT and mutants of the B. bassiana strain ARSEF 2860 were routinely cultured on PDA (BD Difco) at 25°C for two weeks for conidial spore isolation. The rice blast fungus M. oryzae strain 70–15 was used to amplify the Slp1 gene. For liquid incubation, fungi were grown in Sabouraud dextrose broth (SDB, BD Difco) at 25°C in a rotatory shaker. Conidium suspensions were prepared in 0.05% (v/v) Tween-20 and filtered through four layers of sterile lens-cleaning tissues to remove hyphal fragments for different experiments. The WT and mutant strains were also grown on PDA amended with H2O2 (3 mM) and Calcofluor White (200 μg/ml) at 25°C for different times [38]. The cultures were additionally incubated at 37°C to compare the stress responses between the WT and mutant strains.
The proteins containing LysM domain(s) were retrieved from the genomes of 13 insect pathogenic fungi (S1 Table), 13 plant pathogenic fungi, and nine mammalian pathogenic fungi (S2 Table) using the program HMMER 3.1 (http://hmmer.org/). The obtained sequences were manually curated for further analysis of signal peptide, cysteine-residue ratio, transmembrane feature and length of the LysM domain proteins. A phylogenetic tree was constructed for the LysM domain-containing proteins retrieved from the genomes of the 13 insect pathogenic fungi mentioned above and the three plant pathogenic fungi that have been investigated and genome sequenced: M. oryzae, F. oxysporum and Zymoseptoria tritici. Thus, the full sequences of each protein were aligned using the program MUSCLE 3.8.31 [39], and a maximum likelihood tree was constructed using RAxML (ver. 3.1) using a WAG model and a bootstrap test of 1,000 replicates. The conserved sites at LysM domains extracted from the Pfam PF01476 of proteins from bacteria (8,738 proteins), proteins from insect pathogens (133) and plant pathogens (96) were searched and plotted using the program GLAM2 [40]. The LysM domain sequences extracted from the selected proteins (S1 and S2 Tables) were also used for neighbor-joining of phylogenetic analysis using MEGA (ver. 7.0) [41] with a bootstrap test of 1,000 replicates.
Twelve LysM domain-containing protein genes are encoded by B. bassiana [30]. To examine the expression of these genes, total RNA was extracted from the mycelia or blastospores harvested from SDB and the conidial spores or hyphae from the PDA plates. To determine gene expression during fungal in vivo infection, the last instar larvae of wax moth (G. mellonella) were individually injected with 10 μl of spore suspension (107 spores/ml) for 60 hrs. Insect hemolymph was collected on ice, and fungal hyphal bodies were harvested by gradient centrifugation using Centricoll (Sigma-Aldrich) [30]. Each RNA sample was converted to cDNA using an AffinityScript multiple-temperature cDNA synthesis kit (Toyobo). qRT-PCR analysis was performed using a SYBR Premix Ex Taq kit (Takara) containing the primer pairs for different genes (S3 Table) on an ABI Prism 7000 system (Applied Biosystems). The β-tubulin gene (BBA_07018) of B. bassiana was amplified as an internal control. To determine insect antifungal gallerimycin gene expression, the last instar wax moth larvae were individually injected with the spore suspensions of WT and mutants for 36 hrs. Insect fat bodies were then dissected on ice and collected for RNA extraction to quantify the expression of the antifungal gene [9].
Based on the RT-PCR analysis, the genes Blys2, Blys4, Blys5, Blys6, Blys7 and Blys8 were found to be transcribed by the fungus during growth in insect hemocoels. Targeted gene deletion of these sixe genes was individually performed by homologous recombination via Agrobacterium-mediated fungal transformation as previously described [42]. Briefly, the 5′- and 3′-flanking sequences were amplified using genomic DNA as a template with the appropriate primer pairs (S3 Table). The products were purified, digested with restriction enzymes and then inserted into the corresponding sites of the binary vector pDHt-bar (conferring resistance against ammonium glufosinate) to generate different plasmids (S4 Table) for transformation of the WT strain. In addition, the LysM effector Slp1 gene of M. oryzae [17] was amplified with the primers Slp1F/Slp1R and placed under the control of a constitutive GpdA gene (BBA_05480) promoter, and the cassette was integrated into the binary vector pDHt-ben (conferring resistance against benomyl) to transform the deletion mutants ΔBlys2 and ΔBlys5 for heterologous complementation. The Slp1-gene cassette was also cloned into the plasmid pDHt-bar to transform the WT strain for constitutive expression. A laccase gene (BBA_08183) promoter was also used to control Blys2 to transform the WT because this laccase gene was highly transcribed by the fungus during growth in insect hemocoels [30]. The obtained mutants (S4 Table) were verified by PCR and RT-PCR analyses using the corresponding primer pairs (S3 Table).
To determine protein localization, Blys2 and Slp1 were individually fused in frame at the C-termini with a GFP protein, and the cassettes were placed under the control of the GpdA gene promoter. The binary vectors were used to transform the WT strain of B. bassiana. The truncated Blys2 without signal peptide region (Blys2-SP) was also fused with the GFP protein. The WT and obtained mutants were grown under various nutrient conditions, and the fungal cells were harvested for microscopic observations. For chitin staining, fungal cells were incubated with 0.01% Calcofluor White (Sigma-Aldrich) buffered in 10% KOH for 1 min and rinsed twice with phosphate buffered saline (PBS, pH 8.0) before examination using a confocal microscopy (TCS SP8, Leica). To determine the secretion of Blys2, the obtained mutants engineered with GFP-fused Blys2 with and without SP regions were grown in SDB medium in a rotatory shaker at 220 rpm for three days. The cultures were filtered through filter paper, and the filtrates were further treated with a syringe filter unit (GP/SLGP033RS, Millipore) to remove fungal cells. Extracellular proteins were concentrated by precipitation with ammonium sulfate. The samples were centrifuged at 15,000× g for five minutes, and the proteins were reconstituted in Tris-HCl buffer (pH 8.0) and kept at 4°C. Mycelial samples were washed twice with sterile water and homogenized for total protein extraction. Extracellular and intracellular protein samples were separated using a 12% SDS-PAGE gel, and the Western blotting analysis was performed using the anti-GFP antibody (Abcam, China).
To determine the binding feature of Blys2 and Blys5 to cell wall chitins, conidia of B. bassiana were harvested from the two-week old of PDA plates, and suspended in 1 ml of 0.05% (v/v) Tween-20. The suspension was filtered through two layers of sterile filter paper to remove fungal hyphae. The conidia were collected by centrifugation, washed twice with distilled water and then resuspended in 5% (w/v) KOH for boiling for 30 min. After cooling down, cell wall chitin sample was collected by centrifugation at 15,000× g for 3 min, and the pellets were then washed with distilled water for three times, and resuspended in the solution of 40% H2O2 and glacial acetic acid (1:1) for boiling in water for 45 min [43]. Chitin was collected by low speed of centrifugation. The obtained pellets were washed for three times and resuspended in PBS (pH 8.0) for experiments.
To determine the effect of LysM domain number on chitin binding, we performed Blys2 protein truncations and polysaccharide binding assays. Thus, the primer pairs BF1/BR2, BF1/BR4, BF2/BR5, BF3/BR5 and BF4/BR5 (S3 Table) were used to amplify the regions containing the LysM domains 1–2, 1–4, 2–5, 3–5 and 4–5 of Blys2, respectively. The cDNA of Blys5 was amplified with the primers B5F and B5R. The purified products were integrated into the expression vector pET28b containing the His ×6 tag at the C-terminus using the Gateway clone system (Invitrogen). The plasmids were individually transformed into the BL21 (DE3) strain of E. coli for expression. For binding assays, the purified proteins (at a final concentration of 20 μg/ml each) were individually incubated with 3 μg of fungal chitin isolated above, chitin beads (Bioleaf, China), chitosan (Sigma-Aldrich) and cellulose (Sigma-Aldrich) in a total volume of 800 μl of water for 30 min with gentle rotating at room temperature. The samples were centrifuged at 15,000× g for 5 min [15]. The supernatants were collected, and the pellets were washed with PBS for three times. A SDS-PAGE analysis was conducted to detect the proteins in the supernatants and those bound to polysaccharides.
To determine the potential effect of Blys2 and Blys5 on protecting fungal cell walls against chitinase, the WT and WT::Slp1 spores were germinated in SDB (at a final concentration of ca. 1 × 105 spores /ml) for 16 hrs. The germlings were harvested by centrifugation and washed twice with the 0.1 M potassium phosphate buffer containing 0.7 M KCl and 5 mM MgSO4. The WT germlings were pre-incubated with either Blys2 or Blys5 (10 μg) for 2 hrs, and additional WT (mock control) and the WT::Slp1 germlings were treated with the buffer for 2 hrs. The samples were washed twice with the buffer and then treated with the buffered chitinase (Sigma-Aldrich; 0.2%, w/v) mixtures containing β-glucuronidase (Sigma-Aldrich; 0.4%, w/v), cellulase (Sigma-Aldrich; 0.4%, w/v) and lysozyme (Yeasen; 0.4%, w/v) [5] for 2 hrs at 30°C under gentle shaking. The rate of protoplast formation was estimated for each sample by examining 50 microscopic fields.
Insect bioassays were conducted against the newly emerged last instar larvae of wax moth G. mellonella. Conidial suspensions were prepared for both topical infection (1 × 107 conidia/ml) and injection (1 × 106 conidia/ml) assays. Each treatment had three replicates with 15 insects each, and the experiments were repeated twice. For injection assays, spore suspensions (10 μl each) were injected into the base of the second proleg of the insects. Additional insects were injected and bled at various times to examine insect cellular immune response and fungal developments within the insect hemocoels. Insect mortality was recorded every 12 h after the treatments, and the LT50 values were calculated for each strain by Kaplan-Meier analysis. The differences were estimated between the WT and each mutant by the Log-rank test with the program SPSS (ver. 19) [44].
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.